diff --git a/.github/workflows/fastvideo-kernel-publish.yml b/.github/workflows/fastvideo-kernel-publish.yml new file mode 100644 index 000000000..60d9d9376 --- /dev/null +++ b/.github/workflows/fastvideo-kernel-publish.yml @@ -0,0 +1,222 @@ +name: Publish FastVideo Kernel to PyPI on Version Change + +on: + push: + branches: + - main + paths: + - "fastvideo-kernel/pyproject.toml" + workflow_dispatch: + +jobs: + check-version-change: + runs-on: ubuntu-latest + outputs: + version-changed: ${{ steps.check-version.outputs.changed }} + new-version: ${{ steps.check-version.outputs.new-version }} + steps: + - name: Checkout code + uses: actions/checkout@v4 + with: + fetch-depth: 2 + + - name: Check if version changed + id: check-version + run: | + cd fastvideo-kernel + # Get current commit's version from pyproject.toml + # Use ^ to match start of line to avoid matching minimum-version + NEW_VERSION=$(grep -oP '^version\s*=\s*"\K[^"]+' pyproject.toml) + echo "New version: $NEW_VERSION" + + # Get previous version from git history + # Note: git show expects path relative to repo root + OLD_VERSION=$(git show HEAD~1:fastvideo-kernel/pyproject.toml | grep -oP '^version\s*=\s*"\K[^"]+' || echo "0.0.0") + echo "Old version: $OLD_VERSION" + + if [ "$NEW_VERSION" != "$OLD_VERSION" ]; then + echo "Version changed from $OLD_VERSION to $NEW_VERSION" + echo "changed=true" >> $GITHUB_OUTPUT + echo "new-version=$NEW_VERSION" >> $GITHUB_OUTPUT + else + echo "Version did not change" + echo "changed=false" >> $GITHUB_OUTPUT + fi + + build_wheels: + name: Build Wheel + needs: check-version-change + if: ${{ needs.check-version-change.outputs.version-changed == 'true' || github.event_name == 'workflow_dispatch' }} + runs-on: ${{ matrix.os }} + + strategy: + fail-fast: false + matrix: + os: [ubuntu-22.04] + python-version: ['3.10', '3.11', '3.12'] + torch-cuda: + # - torch-version: '2.5.1' + # cuda-version: '12.4.1' + # torch-cuda-short: 'cu124' + # - torch-version: '2.6.0' + # cuda-version: '12.6.3' + # torch-cuda-short: 'cu126' + # - torch-version: '2.7.1' + # cuda-version: '12.8.0' + # torch-cuda-short: 'cu128' + - torch-version: '2.9.1' + cuda-version: '12.8.0' + torch-cuda-short: 'cu128' + + steps: + - name: Free up disk space + run: | + echo "Initial disk space:" + df -h + + # Remove large directories + sudo rm -rf /usr/share/dotnet + sudo rm -rf /usr/local/lib/android + sudo rm -rf /opt/ghc + sudo rm -rf /usr/local/share/boost + sudo rm -rf /usr/share/swift + sudo rm -rf /usr/local/lib/node_modules + sudo rm -rf /usr/local/share/powershell + sudo rm -rf /usr/share/rust + sudo rm -rf /usr/local/.ghcup + + # Remove cached files + sudo rm -rf /var/lib/apt/lists/* + sudo rm -rf /var/cache/apt/archives/* + + echo "Disk space after cleanup:" + df -h + + - name: Checkout + uses: actions/checkout@v4 + + - name: Set up Python + uses: actions/setup-python@v5 + with: + python-version: ${{ matrix.python-version }} + + - name: Install CUDA ${{ matrix.torch-cuda.cuda-version }} + uses: Jimver/cuda-toolkit@v0.2.21 + id: cuda-toolkit + with: + cuda: ${{ matrix.torch-cuda.cuda-version }} + linux-local-args: '["--toolkit"]' + method: 'network' + + - name: Install dependencies (GCC, Clang, CUDA Paths, Git) + run: | + sudo apt update + sudo apt install -y git patchelf gcc-11 g++-11 clang-11 + sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11 + + # Allow Git to Access Safe Directory + git config --global --add safe.directory /__w/FastVideo/FastVideo + + # Set CUDA environment variables + export CUDA_HOME=/usr/local/cuda-${{ matrix.torch-cuda.cuda-version }} + export PATH=${CUDA_HOME}/bin:${PATH} + export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH + + # Verify installation + gcc --version + g++ --version + clang-11 --version + nvcc --version + + - name: Install PyTorch ${{ matrix.torch-cuda.torch-version }}+cu${{ matrix.torch-cuda.cuda-version }} + run: | + pip install --upgrade pip + pip install typing-extensions==4.12.2 + pip install --no-cache-dir torch==${{ matrix.torch-cuda.torch-version }} --index-url https://download.pytorch.org/whl/${{matrix.torch-cuda.torch-cuda-short}} + nvcc --version + python --version + python -c "import torch; print('PyTorch:', torch.__version__)" + python -c "import torch; print('CUDA:', torch.version.cuda)" + python -c "from torch.utils import cpp_extension; print (cpp_extension.CUDA_HOME)" + + - name: Build wheel + run: | + export PYTHONPATH=$GITHUB_WORKSPACE:$PYTHONPATH + + pip install setuptools ninja packaging wheel triton scikit-build-core cmake build + + cd fastvideo-kernel + git submodule update --init --recursive # Ensure ThunderKittens submodule is initialized + # Release builds are produced on GPU-less runners, so force-enable TK and target Hopper. + export TORCH_CUDA_ARCH_LIST="9.0a" + export CMAKE_ARGS="${CMAKE_ARGS:-} -DFASTVIDEO_KERNEL_BUILD_TK=ON -DCMAKE_CUDA_ARCHITECTURES=90a" + + # Build standard wheel (no local version suffix) for PyPI + python -m build --wheel --outdir dist + + # Fix the wheel to be manylinux compliant + pip install auditwheel + # Point auditwheel at torch libs, but do not vendor them into the wheel. + TORCH_LIB_DIR=$(python - <<'PY' + import os + import torch + + print(os.path.join(os.path.dirname(torch.__file__), "lib")) + PY + ) + export LD_LIBRARY_PATH="${TORCH_LIB_DIR}:${LD_LIBRARY_PATH}" + # Target manylinux_2_35 (Ubuntu 22.04 native) + auditwheel repair dist/*.whl --plat manylinux_2_35_x86_64 -w fixed_dist \ + --exclude libtorch_cuda.so \ + --exclude libtorch_cpu.so \ + --exclude libtorch.so \ + --exclude libc10.so \ + --exclude libc10_cuda.so \ + --exclude libtorch_python.so + # Move fixed wheels back to dist for upload consistency + rm dist/*.whl + mv fixed_dist/*.whl dist/ + + - name: Upload wheel artifact + # Only upload if it's the "main" CUDA version we want on PyPI + # We upload all to artifacts for inspection/GH releases, but give them distinct artifact names + uses: actions/upload-artifact@v4 + with: + name: fastvideo_kernel-py${{ matrix.python-version }}-${{ matrix.torch-cuda.torch-cuda-short }}-torch${{ matrix.torch-cuda.torch-version }} + path: fastvideo-kernel/dist/*.whl + retention-days: 90 + + publish_package: + name: Publish package + needs: [build_wheels, check-version-change] + if: ${{ needs.check-version-change.outputs.version-changed == 'true' || github.event_name == 'workflow_dispatch' }} + runs-on: ubuntu-22.04 + permissions: + id-token: write # Needed for OIDC Trusted Publishing + + steps: + - uses: actions/checkout@v4 + + - uses: actions/setup-python@v5 + with: + python-version: '3.10' + + - name: Download PyPI wheels + uses: actions/download-artifact@v4 + with: + path: fastvideo-kernel/dist/ + pattern: 'fastvideo_kernel-py*' + merge-multiple: true + + - name: Build source distribution + run: | + pip install build scikit-build-core cmake ninja + + cd fastvideo-kernel + # We don't need full CUDA/Torch to just package the source (sdist) + python -m build --sdist --outdir dist + + - name: Publish release distributions to PyPI + uses: pypa/gh-action-pypi-publish@release/v1 + with: + packages-dir: fastvideo-kernel/dist/ diff --git a/docs/inference/inference_quick_start.md b/docs/inference/inference_quick_start.md index a5e0db2f9..4a647ab48 100644 --- a/docs/inference/inference_quick_start.md +++ b/docs/inference/inference_quick_start.md @@ -107,6 +107,8 @@ If you encounter CUDA out of memory errors: (single GPU) or `use_fsdp_inference=True` (multi-GPU) - Try a smaller model or use distilled versions - Use `num_gpus` > 1 if multiple GPUs are available +- Try enabling FSDP inference with `use_fsdp_inference=True` (may slow down generation) +- Try enabling DiT layerwise offload with `dit_layerwise_offload=True` (now only a few models support this, but may introduce less overhead than FSDP) ### Slow Generation diff --git a/download.sh b/download.sh new file mode 100644 index 000000000..cebb6c13a --- /dev/null +++ b/download.sh @@ -0,0 +1 @@ +python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "data2/Wan-Syn_77x448x832_600k" --repo_type "dataset" \ No newline at end of file diff --git a/examples/dataset/vidprom/download_dataset.sh b/examples/dataset/vidprom/download_dataset.sh index 7ce4e3c30..adaeba606 100644 --- a/examples/dataset/vidprom/download_dataset.sh +++ b/examples/dataset/vidprom/download_dataset.sh @@ -1,3 +1,3 @@ #! /bin/bash -huggingface-cli download gdhe17/Self-Forcing vidprom_filtered_extended.txt --local-dir prompts +huggingface-cli download FastVideo/vidprom_16k_umt5_text_embed --repo-type dataset --local-dir data/vidprom_16k_umt5_text_embed \ No newline at end of file diff --git a/examples/dataset/vidprom/download_dataset_ode_init.sh b/examples/dataset/vidprom/download_dataset_ode_init.sh new file mode 100644 index 000000000..7ce4e3c30 --- /dev/null +++ b/examples/dataset/vidprom/download_dataset_ode_init.sh @@ -0,0 +1,3 @@ +#! /bin/bash + +huggingface-cli download gdhe17/Self-Forcing vidprom_filtered_extended.txt --local-dir prompts diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/README.md b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/README.md index cce5a097f..b43e87fdf 100644 --- a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/README.md +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/README.md @@ -9,7 +9,7 @@ pip install vsa ### 1. Download dataset: ```bash -bash examples/distill/Wan-Syn-480P/download_dataset.sh +bash examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh ``` ### 2. Configure and run distillation: diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B.sh new file mode 100644 index 000000000..b3ccd78d6 --- /dev/null +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B.sh @@ -0,0 +1,124 @@ +#!/usr/bin/env bash +set -euo pipefail + +############################################ +# Single node, 6 GPUs +############################################ +NUM_GPUS=1 +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +# export CUDA_VISIBLE_DEVICES=0,1,2,3 + +export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_MODE="online" +export WANDB_BASE_URL="https://api.wandb.ai" +export TOKENIZERS_PARALLELISM=false +export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN # or TORCH_SDPA +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +export TRITON_CACHE_DIR=/tmp/triton_cache + +echo "NUM_GPUS: $NUM_GPUS" +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" + +############################################ +# Configs +############################################ +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +OUTPUT_DIR="checkpoints/base_distill" + +# Training arguments +training_args=( + --tracker_project_name wan_t2v_distill_dmd_qat + --output_dir "$OUTPUT_DIR" + --max_train_steps 4000 + --train_batch_size 64 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 12 + --num_height 480 + --num_width 832 + --num_frames 45 + --enable_gradient_checkpointing_type "full" +) + +# Parallel arguments (adjusted to 6 GPUs) +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path "$MODEL_PATH" + --pretrained_model_name_or_path "$MODEL_PATH" + --real_score_model_path "$REAL_SCORE_MODEL_PATH" + --fake_score_model_path "$FAKE_SCORE_MODEL_PATH" +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 1 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file "$VALIDATION_DATASET_FILE" + --validation_steps 100 + --validation_sampling_steps "3" + --validation_guidance_scale "6.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 2e-6 + --mixed_precision "bf16" + --training_state_checkpointing_steps 2000 + --weight_only_checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.0 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 8 + --seed 1000 +) + +# DMD arguments +dmd_args=( + --dmd_denoising_steps '1000,757,522' + --min_timestep_ratio 0.02 + --max_timestep_ratio 0.98 + --generator_update_interval 5 + --real_score_guidance_scale 3.5 +) + +############################################ +# Launch (no rendezvous flags needed) +############################################ +set -x +torchrun \ + --nproc_per_node="$NUM_GPUS" \ + fastvideo/training/wan_distillation_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ + "${dmd_args[@]}" diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_linear_fp4.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_linear_fp4.sh new file mode 100644 index 000000000..4ba2c5571 --- /dev/null +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_linear_fp4.sh @@ -0,0 +1,120 @@ +#!/usr/bin/env bash +set -euo pipefail + +############################################ +# Single node, 6 GPUs +############################################ +NUM_GPUS=6 +export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 + + +export WANDB_MODE="online" +export WANDB_BASE_URL="https://api.wandb.ai" +export TOKENIZERS_PARALLELISM=false +export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN # or TORCH_SDPA +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +export TRITON_CACHE_DIR=/tmp/triton_cache + +echo "NUM_GPUS: $NUM_GPUS" +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" + +############################################ +# Configs +############################################ +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k/train +VALIDATION_DATASET_FILE=data/Wan-Syn_77x448x832_600k/validation_6.json +OUTPUT_DIR="checkpoints/linear_fp4_distill" + +# Training arguments +training_args=( + --tracker_project_name wan_t2v_distill_dmd + --output_dir "$OUTPUT_DIR" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 21 + --num_height 480 + --num_width 832 + --num_frames 81 + --enable_gradient_checkpointing_type "full" + --generator_4bit_linear True +) + +# Parallel arguments (adjusted to 6 GPUs) +parallel_args=( + --num_gpus 6 + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim 6 +) + +# Model arguments +model_args=( + --model_path "$MODEL_PATH" + --pretrained_model_name_or_path "$MODEL_PATH" +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file "$VALIDATION_DATASET_FILE" + --validation_steps 100 + --validation_sampling_steps "3" + --validation_guidance_scale "6.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 2e-6 + --mixed_precision "bf16" + --training_state_checkpointing_steps 2000 + --weight_only_checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.0 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 8 + --seed 1000 +) + +# DMD arguments +dmd_args=( + --dmd_denoising_steps '1000,757,522' + --min_timestep_ratio 0.02 + --max_timestep_ratio 0.98 + --generator_update_interval 5 + --real_score_guidance_scale 3.5 +) + +############################################ +# Launch (no rendezvous flags needed) +############################################ +set -x +torchrun \ + --nproc_per_node="$NUM_GPUS" \ + fastvideo/training/wan_distillation_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ + "${dmd_args[@]}" diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_qat.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_qat.sh new file mode 100644 index 000000000..ce3f13d5a --- /dev/null +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_qat.sh @@ -0,0 +1,122 @@ +#!/usr/bin/env bash +set -euo pipefail + +############################################ +# Single node, 6 GPUs +############################################ +NUM_GPUS=1 +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +export WANDB_MODE="online" +export WANDB_BASE_URL="https://api.wandb.ai" +export TOKENIZERS_PARALLELISM=false +export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN # or TORCH_SDPA +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +export TRITON_CACHE_DIR=/tmp/triton_cache + +echo "NUM_GPUS: $NUM_GPUS" +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" + +############################################ +# Configs +############################################ +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +OUTPUT_DIR="checkpoints/wan_1.3B_t2v_distill_dmd_qat" +REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + +# Training arguments +training_args=( + --tracker_project_name wan_t2v_distill_dmd_qat + --output_dir $OUTPUT_DIR + --max_train_steps 4000 + --train_batch_size 64 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 21 + --num_height 480 + --num_width 832 + --num_frames 81 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments (adjusted to 6 GPUs) +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH + --real_score_model_path $REAL_SCORE_MODEL_PATH + --fake_score_model_path $FAKE_SCORE_MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file "$VALIDATION_DATASET_FILE" + --validation_steps 200 + --validation_sampling_steps "3" + --validation_guidance_scale "6.0" # not used for dmd inference +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 2e-6 + --mixed_precision "bf16" + --training_state_checkpointing_steps 500 + --weight_only_checkpointing_steps 500 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.0 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 8 + --seed 1000 +) + +# DMD arguments +dmd_args=( + --dmd_denoising_steps '1000,757,522' + --min_timestep_ratio 0.02 + --max_timestep_ratio 0.98 + --generator_update_interval 5 + --real_score_guidance_scale 3.5 +) + +############################################ +# Launch (no rendezvous flags needed) +############################################ +set -x +torchrun \ + --nproc_per_node="$NUM_GPUS" \ + fastvideo/training/wan_distillation_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ + "${dmd_args[@]}" diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_sage3.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_sage3.sh new file mode 100644 index 000000000..528c1219c --- /dev/null +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_sage3.sh @@ -0,0 +1,137 @@ +#!/usr/bin/env bash +set -euo pipefail + +############################################ +# Single node, 6 GPUs +############################################ +<<<<<<< HEAD +NUM_GPUS=4 +======= +NUM_GPUS=1 +>>>>>>> 98ce47c7 (checkpoint (qat attn in progress)) +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +# export CUDA_VISIBLE_DEVICES=0,1,2,3 + +export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_MODE="online" +export WANDB_BASE_URL="https://api.wandb.ai" +export TOKENIZERS_PARALLELISM=false +export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN # or TORCH_SDPA +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +export TRITON_CACHE_DIR=/tmp/triton_cache + +echo "NUM_GPUS: $NUM_GPUS" +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" + +############################################ +# Configs +############################################ +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +<<<<<<< HEAD +DATA_DIR=data/Wan-Syn_77x448x832_600k +======= +DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/" +>>>>>>> 98ce47c7 (checkpoint (qat attn in progress)) +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +OUTPUT_DIR="checkpoints/sage3_distill" +REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + +# Training arguments +training_args=( + --tracker_project_name wan_t2v_distill_dmd_qat + --output_dir "$OUTPUT_DIR" + --max_train_steps 4000 + --train_batch_size 16 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 12 + --num_height 480 + --num_width 832 + --num_frames 45 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments (adjusted to 6 GPUs) +parallel_args=( + --num_gpus $NUM_GPUS +<<<<<<< HEAD + --sp_size 4 +======= + --sp_size 1 +>>>>>>> 98ce47c7 (checkpoint (qat attn in progress)) + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path "$MODEL_PATH" + --pretrained_model_name_or_path "$MODEL_PATH" + --real_score_model_path "$REAL_SCORE_MODEL_PATH" + --fake_score_model_path "$FAKE_SCORE_MODEL_PATH" +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 1 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file "$VALIDATION_DATASET_FILE" + --validation_steps 100 + --validation_sampling_steps "3" + --validation_guidance_scale "6.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 2e-6 + --mixed_precision "bf16" + --training_state_checkpointing_steps 2000 + --weight_only_checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.0 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 8 + --seed 1000 +) + +# DMD arguments +dmd_args=( + --dmd_denoising_steps '1000,757,522' + --min_timestep_ratio 0.02 + --max_timestep_ratio 0.98 + --generator_update_interval 5 + --real_score_guidance_scale 3.5 +) + +############################################ +# Launch (no rendezvous flags needed) +############################################ +set -x +torchrun \ + --nproc_per_node="$NUM_GPUS" \ + fastvideo/training/wan_distillation_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ + "${dmd_args[@]}" diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom.sh new file mode 100644 index 000000000..bc258a031 --- /dev/null +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom.sh @@ -0,0 +1,122 @@ +#!/usr/bin/env bash +set -euo pipefail + +############################################ +# Single node, 6 GPUs +############################################ +NUM_GPUS=1 +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +export WANDB_MODE="online" +export WANDB_BASE_URL="https://api.wandb.ai" +export TOKENIZERS_PARALLELISM=false +export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN # or TORCH_SDPA +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +export TRITON_CACHE_DIR=/tmp/triton_cache + +echo "NUM_GPUS: $NUM_GPUS" +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" + +############################################ +# Configs +############################################ +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/vidprom_16k_umt5_text_embed +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +OUTPUT_DIR="checkpoints/wan_1.3B_t2v_distill_dmd_vidprom" +REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + +# Training arguments +training_args=( + --tracker_project_name wan_t2v_distill_dmd_qat + --output_dir $OUTPUT_DIR + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 21 + --num_height 480 + --num_width 832 + --num_frames 81 + --enable_gradient_checkpointing_type "full" +) + +# Parallel arguments (adjusted to 6 GPUs) +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH + --real_score_model_path $REAL_SCORE_MODEL_PATH + --fake_score_model_path $FAKE_SCORE_MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file "$VALIDATION_DATASET_FILE" + --validation_steps 200 + --validation_sampling_steps "3" + --validation_guidance_scale "6.0" # not used for dmd inference +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 2e-6 + --mixed_precision "bf16" + --training_state_checkpointing_steps 500 + --weight_only_checkpointing_steps 500 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.0 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 8 + --seed 1000 +) + +# DMD arguments +dmd_args=( + --dmd_denoising_steps '1000,757,522' + --min_timestep_ratio 0.02 + --max_timestep_ratio 0.98 + --generator_update_interval 5 + --real_score_guidance_scale 3.5 + --simulate_generator_forward +) + +############################################ +# Launch (no rendezvous flags needed) +############################################ +set -x +torchrun \ + --nproc_per_node="$NUM_GPUS" \ + fastvideo/training/wan_distillation_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ + "${dmd_args[@]}" \ No newline at end of file diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_e2e_fp4.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_e2e_fp4.sh new file mode 100644 index 000000000..fb30044b2 --- /dev/null +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_e2e_fp4.sh @@ -0,0 +1,124 @@ +#!/usr/bin/env bash +set -euo pipefail + +############################################ +# Single node, 6 GPUs +############################################ +NUM_GPUS=1 +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +export WANDB_MODE="online" +export WANDB_BASE_URL="https://api.wandb.ai" +export TOKENIZERS_PARALLELISM=false +export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN # or TORCH_SDPA +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +export TRITON_CACHE_DIR=/tmp/triton_cache + +echo "NUM_GPUS: $NUM_GPUS" +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" + +############################################ +# Configs +############################################ +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/vidprom_16k_umt5_text_embed +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +OUTPUT_DIR="checkpoints/wan_1.3B_t2v_distill_dmd_e2e_fp4_vidprom" +REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + +# Training arguments +training_args=( + --tracker_project_name wan_t2v_distill_dmd_qat + --output_dir $OUTPUT_DIR + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 21 + --num_height 480 + --num_width 832 + --num_frames 81 + --enable_gradient_checkpointing_type "full" + --generator_4bit_linear True + --generator_4bit_attn True +) + +# Parallel arguments (adjusted to 6 GPUs) +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH + --real_score_model_path $REAL_SCORE_MODEL_PATH + --fake_score_model_path $FAKE_SCORE_MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file "$VALIDATION_DATASET_FILE" + --validation_steps 200 + --validation_sampling_steps "3" + --validation_guidance_scale "6.0" # not used for dmd inference +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 2e-6 + --mixed_precision "bf16" + --training_state_checkpointing_steps 500 + --weight_only_checkpointing_steps 500 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.0 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 8 + --seed 1000 +) + +# DMD arguments +dmd_args=( + --dmd_denoising_steps '1000,757,522' + --min_timestep_ratio 0.02 + --max_timestep_ratio 0.98 + --generator_update_interval 5 + --real_score_guidance_scale 3.5 + --simulate_generator_forward +) + +############################################ +# Launch (no rendezvous flags needed) +############################################ +set -x +torchrun \ + --nproc_per_node="$NUM_GPUS" \ + fastvideo/training/wan_distillation_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ + "${dmd_args[@]}" \ No newline at end of file diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_linear_fp4.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_linear_fp4.sh new file mode 100644 index 000000000..c3aa922bb --- /dev/null +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_linear_fp4.sh @@ -0,0 +1,123 @@ +#!/usr/bin/env bash +set -euo pipefail + +############################################ +# Single node, 6 GPUs +############################################ +NUM_GPUS=1 +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +export WANDB_MODE="online" +export WANDB_BASE_URL="https://api.wandb.ai" +export TOKENIZERS_PARALLELISM=false +export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN # or TORCH_SDPA +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +export TRITON_CACHE_DIR=/tmp/triton_cache + +echo "NUM_GPUS: $NUM_GPUS" +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" + +############################################ +# Configs +############################################ +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/vidprom_16k_umt5_text_embed +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +OUTPUT_DIR="checkpoints/wan_1.3B_t2v_distill_dmd_linear_fp4_vidprom" +REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + +# Training arguments +training_args=( + --tracker_project_name wan_t2v_distill_dmd_qat + --output_dir $OUTPUT_DIR + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 21 + --num_height 480 + --num_width 832 + --num_frames 81 + --enable_gradient_checkpointing_type "full" + --generator_4bit_linear True +) + +# Parallel arguments (adjusted to 6 GPUs) +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH + --real_score_model_path $REAL_SCORE_MODEL_PATH + --fake_score_model_path $FAKE_SCORE_MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file "$VALIDATION_DATASET_FILE" + --validation_steps 200 + --validation_sampling_steps "3" + --validation_guidance_scale "6.0" # not used for dmd inference +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 2e-6 + --mixed_precision "bf16" + --training_state_checkpointing_steps 500 + --weight_only_checkpointing_steps 500 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.0 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 8 + --seed 1000 +) + +# DMD arguments +dmd_args=( + --dmd_denoising_steps '1000,757,522' + --min_timestep_ratio 0.02 + --max_timestep_ratio 0.98 + --generator_update_interval 5 + --real_score_guidance_scale 3.5 + --simulate_generator_forward +) + +############################################ +# Launch (no rendezvous flags needed) +############################################ +set -x +torchrun \ + --nproc_per_node="$NUM_GPUS" \ + fastvideo/training/wan_distillation_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ + "${dmd_args[@]}" \ No newline at end of file diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_qat.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_qat.sh new file mode 100644 index 000000000..88e84cde0 --- /dev/null +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B_vidprom_qat.sh @@ -0,0 +1,123 @@ +#!/usr/bin/env bash +set -euo pipefail + +############################################ +# Single node, 6 GPUs +############################################ +NUM_GPUS=1 +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +export WANDB_MODE="online" +export WANDB_BASE_URL="https://api.wandb.ai" +export TOKENIZERS_PARALLELISM=false +export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN # or TORCH_SDPA +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +export TRITON_CACHE_DIR=/tmp/triton_cache + +echo "NUM_GPUS: $NUM_GPUS" +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" + +############################################ +# Configs +############################################ +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/vidprom_16k_umt5_text_embed +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +OUTPUT_DIR="checkpoints/wan_1.3B_t2v_distill_dmd_qat_vidprom" +REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + +# Training arguments +training_args=( + --tracker_project_name wan_t2v_distill_dmd_qat + --output_dir $OUTPUT_DIR + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 21 + --num_height 480 + --num_width 832 + --num_frames 81 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments (adjusted to 6 GPUs) +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH + --real_score_model_path $REAL_SCORE_MODEL_PATH + --fake_score_model_path $FAKE_SCORE_MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file "$VALIDATION_DATASET_FILE" + --validation_steps 200 + --validation_sampling_steps "3" + --validation_guidance_scale "6.0" # not used for dmd inference +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 2e-6 + --mixed_precision "bf16" + --training_state_checkpointing_steps 500 + --weight_only_checkpointing_steps 500 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.0 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 8 + --seed 1000 +) + +# DMD arguments +dmd_args=( + --dmd_denoising_steps '1000,757,522' + --min_timestep_ratio 0.02 + --max_timestep_ratio 0.98 + --generator_update_interval 5 + --real_score_guidance_scale 3.5 + --simulate_generator_forward +) + +############################################ +# Launch (no rendezvous flags needed) +############################################ +set -x +torchrun \ + --nproc_per_node="$NUM_GPUS" \ + fastvideo/training/wan_distillation_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ + "${dmd_args[@]}" \ No newline at end of file diff --git a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh index 50e072a4a..390027143 100644 --- a/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh +++ b/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh @@ -1,3 +1,3 @@ #!/bin/bash - -python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset" +mkdir -p data +python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "data/Wan-Syn_77x448x832_600k" --repo_type "dataset" diff --git a/examples/inference/batch_generate.py b/examples/inference/batch_generate.py new file mode 100644 index 000000000..d62986424 --- /dev/null +++ b/examples/inference/batch_generate.py @@ -0,0 +1,386 @@ +#!/usr/bin/env python3 +""" +Batch video generation script for FastVideo. + +Takes a text file with one prompt per line and generates videos for each prompt. + +Usage: + python batch_generate.py --prompts prompts.txt --output_dir outputs/ + python batch_generate.py --prompts prompts.txt --model Wan-AI/Wan2.1-T2V-14B-Diffusers --num_gpus 2 + python batch_generate.py --prompts prompts.txt --attention_backend SAGE_ATTN_THREE +""" + +import argparse +import os +import re +import time +from pathlib import Path + +# Note: VideoGenerator import is deferred to main() so we can set +# FASTVIDEO_ATTENTION_BACKEND environment variable first + + +def sanitize_filename(prompt: str, max_length: int = 50) -> str: + """ + Create a safe filename from a prompt. + + Args: + prompt: The prompt text + max_length: Maximum length of the filename (excluding extension) + + Returns: + A sanitized filename string + """ + # Take first part of prompt and clean it + sanitized = re.sub(r'[^\w\s-]', '', prompt.lower()) + sanitized = re.sub(r'[-\s]+', '_', sanitized) + sanitized = sanitized.strip('_') + + # Truncate to max length + if len(sanitized) > max_length: + sanitized = sanitized[:max_length].rsplit('_', 1)[0] + + return sanitized or "video" + + +def load_prompts(prompt_file: str) -> list[str]: + """ + Load prompts from a text file. + + Args: + prompt_file: Path to the text file with prompts (one per line) + + Returns: + List of prompts + """ + with open(prompt_file, 'r', encoding='utf-8') as f: + prompts = [line.strip() for line in f if line.strip()] + return prompts + + +def main(): + parser = argparse.ArgumentParser( + description="Batch video generation with FastVideo", + formatter_class=argparse.RawDescriptionHelpFormatter, + epilog=""" +Examples: + # Basic usage with default model (Wan2.1-T2V-1.3B) + python batch_generate.py --prompts prompts.txt + + # Use 14B model with 2 GPUs + python batch_generate.py --prompts prompts.txt --model Wan-AI/Wan2.1-T2V-14B-Diffusers --num_gpus 2 + + # Custom resolution and frame count + python batch_generate.py --prompts prompts.txt --height 720 --width 1280 --num_frames 125 + + # Use CPU offloading for low VRAM + python batch_generate.py --prompts prompts.txt --dit_cpu_offload --vae_cpu_offload + + # Use custom fine-tuned weights + python batch_generate.py --prompts prompts.txt --init_weights_from_safetensors checkpoints/my_model/transformer + + # Use SageAttention3 backend (RTX 5090 only) + python batch_generate.py --prompts prompts.txt --attention_backend SAGE_ATTN_THREE + + # Use Flash Attention backend + python batch_generate.py --prompts prompts.txt --attention_backend FLASH_ATTN + """ + ) + + # Attention backend argument (must be parsed early) + parser.add_argument( + "--attention_backend", + type=str, + default=None, + choices=[ + "TORCH_SDPA", + "FLASH_ATTN", + "SLIDING_TILE_ATTN", + "VIDEO_SPARSE_ATTN", + "SAGE_ATTN", + "SAGE_ATTN_THREE", + ], + help="Attention backend to use (default: auto-detect). SAGE_ATTN_THREE requires RTX 5090." + ) + + # Input/Output arguments + parser.add_argument( + "--prompts", "-p", + type=str, + required=True, + help="Path to text file with prompts (one per line)" + ) + parser.add_argument( + "--output_dir", "-o", + type=str, + default="batch_outputs", + help="Directory to save generated videos (default: batch_outputs)" + ) + + # Model arguments + parser.add_argument( + "--model", "-m", + type=str, + default="Wan-AI/Wan2.1-T2V-1.3B-Diffusers", + help="Model path or HuggingFace model ID (default: Wan-AI/Wan2.1-T2V-1.3B-Diffusers)" + ) + parser.add_argument( + "--num_gpus", + type=int, + default=1, + help="Number of GPUs to use (default: 1)" + ) + + # Video parameters + parser.add_argument( + "--height", + type=int, + default=None, + help="Video height (default: model default)" + ) + parser.add_argument( + "--width", + type=int, + default=None, + help="Video width (default: model default)" + ) + parser.add_argument( + "--num_frames", + type=int, + default=None, + help="Number of frames to generate (default: model default)" + ) + parser.add_argument( + "--fps", + type=int, + default=None, + help="Frames per second (default: model default)" + ) + + # Sampling parameters + parser.add_argument( + "--num_inference_steps", + type=int, + default=None, + help="Number of inference steps (default: model default)" + ) + parser.add_argument( + "--guidance_scale", + type=float, + default=None, + help="Guidance scale (default: model default)" + ) + parser.add_argument( + "--seed", + type=int, + default=None, + help="Random seed (default: None, uses different seed per video)" + ) + parser.add_argument( + "--negative_prompt", + type=str, + default=None, + help="Negative prompt (default: model default)" + ) + + # Memory optimization arguments + parser.add_argument( + "--use_fsdp_inference", + action="store_true", + help="Use FSDP for inference (set if GPU OOM)" + ) + parser.add_argument( + "--dit_cpu_offload", + action="store_true", + help="Offload DiT to CPU when not in use" + ) + parser.add_argument( + "--vae_cpu_offload", + action="store_true", + help="Offload VAE to CPU when not in use" + ) + parser.add_argument( + "--text_encoder_cpu_offload", + action="store_true", + default=True, + help="Offload text encoder to CPU when not in use (default: True)" + ) + parser.add_argument( + "--no_text_encoder_cpu_offload", + action="store_true", + help="Disable text encoder CPU offloading" + ) + parser.add_argument( + "--pin_cpu_memory", + action="store_true", + default=True, + help="Pin CPU memory for faster transfers (default: True)" + ) + parser.add_argument( + "--no_pin_cpu_memory", + action="store_true", + help="Disable CPU memory pinning" + ) + parser.add_argument( + "--init_weights_from_safetensors", + type=str, + default=None, + help="Path to custom safetensors weights to load (e.g., checkpoint/transformer directory)" + ) + + # Processing arguments + parser.add_argument( + "--start_idx", + type=int, + default=0, + help="Start processing from this prompt index (0-based)" + ) + parser.add_argument( + "--end_idx", + type=int, + default=None, + help="Stop processing at this prompt index (exclusive)" + ) + parser.add_argument( + "--filename_prefix", + type=str, + default="", + help="Prefix for output filenames" + ) + parser.add_argument( + "--use_index_only", + action="store_true", + help="Use only index numbers for filenames instead of prompt text" + ) + + args = parser.parse_args() + + # Set attention backend BEFORE importing VideoGenerator + if args.attention_backend: + os.environ["FASTVIDEO_ATTENTION_BACKEND"] = args.attention_backend + print(f"Using attention backend: {args.attention_backend}") + + # Now import VideoGenerator (after setting env var) + from fastvideo import VideoGenerator + + # Validate inputs + if not os.path.exists(args.prompts): + raise FileNotFoundError(f"Prompts file not found: {args.prompts}") + + # Load prompts + prompts = load_prompts(args.prompts) + print(f"Loaded {len(prompts)} prompts from {args.prompts}") + + # Apply start/end indices + end_idx = args.end_idx if args.end_idx is not None else len(prompts) + prompts_to_process = prompts[args.start_idx:end_idx] + print(f"Processing prompts {args.start_idx} to {end_idx - 1} ({len(prompts_to_process)} videos)") + + # Create output directory + os.makedirs(args.output_dir, exist_ok=True) + + # Handle boolean flag inversions + text_encoder_cpu_offload = not args.no_text_encoder_cpu_offload + pin_cpu_memory = not args.no_pin_cpu_memory + + # Initialize generator + print(f"\nInitializing VideoGenerator with model: {args.model}") + if args.init_weights_from_safetensors: + print(f"Loading custom weights from: {args.init_weights_from_safetensors}") + generator = VideoGenerator.from_pretrained( + args.model, + num_gpus=args.num_gpus, + use_fsdp_inference=args.use_fsdp_inference, + dit_cpu_offload=args.dit_cpu_offload, + vae_cpu_offload=args.vae_cpu_offload, + text_encoder_cpu_offload=text_encoder_cpu_offload, + pin_cpu_memory=pin_cpu_memory, + init_weights_from_safetensors=args.init_weights_from_safetensors, + ) + print("Generator initialized successfully!\n") + + # Build optional kwargs for generate_video + generate_kwargs = {} + if args.height is not None: + generate_kwargs["height"] = args.height + if args.width is not None: + generate_kwargs["width"] = args.width + if args.num_frames is not None: + generate_kwargs["num_frames"] = args.num_frames + if args.fps is not None: + generate_kwargs["fps"] = args.fps + if args.num_inference_steps is not None: + generate_kwargs["num_inference_steps"] = args.num_inference_steps + if args.guidance_scale is not None: + generate_kwargs["guidance_scale"] = args.guidance_scale + if args.negative_prompt is not None: + generate_kwargs["negative_prompt"] = args.negative_prompt + + # Process each prompt + total_start_time = time.time() + successful = 0 + failed = 0 + + for i, prompt in enumerate(prompts_to_process): + global_idx = args.start_idx + i + + # Generate filename + if args.use_index_only: + filename = f"{args.filename_prefix}{global_idx:04d}" + else: + prompt_part = sanitize_filename(prompt) + filename = f"{args.filename_prefix}{global_idx:04d}_{prompt_part}" + + output_path = os.path.join(args.output_dir, f"{filename}.mp4") + + print(f"\n{'='*60}") + print(f"[{i+1}/{len(prompts_to_process)}] Generating video {global_idx}") + print(f"Prompt: {prompt[:100]}{'...' if len(prompt) > 100 else ''}") + print(f"Output: {output_path}") + print(f"{'='*60}") + + try: + video_start_time = time.time() + + # Set seed for this video if specified + kwargs = generate_kwargs.copy() + if args.seed is not None: + kwargs["seed"] = args.seed + i # Different seed per video but reproducible + + # Generate video + result = generator.generate_video( + prompt=prompt, + output_path=output_path, + save_video=True, + **kwargs + ) + + video_time = time.time() - video_start_time + print(f"✓ Video generated in {video_time:.2f}s") + successful += 1 + + except Exception as e: + print(f"✗ Failed to generate video: {e}") + failed += 1 + continue + + # Summary + total_time = time.time() - total_start_time + print(f"\n{'='*60}") + print("BATCH GENERATION COMPLETE") + print(f"{'='*60}") + print(f"Total videos: {len(prompts_to_process)}") + print(f"Successful: {successful}") + print(f"Failed: {failed}") + print(f"Total time: {total_time:.2f}s") + if successful > 0: + print(f"Average time per video: {total_time/successful:.2f}s") + print(f"Output directory: {args.output_dir}") + + # Shutdown generator + generator.shutdown() + + +if __name__ == "__main__": + main() + diff --git a/examples/inference/optimizations/quantization_example.py b/examples/inference/optimizations/quantization_example.py new file mode 100644 index 000000000..171c3f6ce --- /dev/null +++ b/examples/inference/optimizations/quantization_example.py @@ -0,0 +1,94 @@ +import torch +from fastvideo import VideoGenerator +from fastvideo.configs.pipelines.base import PipelineConfig +from fastvideo.layers.quantization.fp4_config import FP4Config +# from fastvideo.layers.quantization.fp8_config import FP8Config +from fastvideo.layers.linear import ReplicatedLinear + +OUTPUT_PATH = "video_samples" + + +def main(): + print("=== FP4 Quantization Video Generation Example ===") + + if not torch.cuda.is_available(): + print("Warning: CUDA not available. FP4 quantization requires GPU.") + return + + gpu_capability = torch.cuda.get_device_capability() + if gpu_capability[0] < 9: # H100 and newer + print(f"Warning: GPU capability {gpu_capability} may not support FP4. Recommended: 9.0+") + + print(f"GPU: {torch.cuda.get_device_name()}") + print(f"GPU Capability: {gpu_capability}") + + model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + # model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers" + pipeline_config = PipelineConfig.from_pretrained(model_id) + pipeline_config.dit_precision = "bf16" + pipeline_config.dit_config.quant_config = FP4Config() + # pipeline_config.dit_config.quant_config = FP8Config() + + # ReplicatedLinear.print_shape_summary() + + print("\nLoading model with FP4 quantization...") + generator = VideoGenerator.from_pretrained( + model_id, + pipeline_config=pipeline_config, + num_gpus=1, + use_fsdp_inference=True, + dit_cpu_offload=False, + vae_cpu_offload=False, + text_encoder_cpu_offload=True, + pin_cpu_memory=False, + ) + + print("FP4 configuration applied. Generating videos...") + + print("\n=== Generating Video with FP4 Quantization ===") + + prompt1 = ( + "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " + "wide with interest. The playful yet serene atmosphere is complemented by soft " + "natural light filtering through the petals. Mid-shot, warm and cheerful tones." + ) + + print(f"Prompt: {prompt1}") + print("Generating video...") + + try: + video1 = generator.generate_video( + prompt1, + output_path=OUTPUT_PATH, + save_video=True, + ) + print("✓ First video generated successfully with FP4 quantization!") + + # # Generate a second video to show the model can be reused + prompt2 = ( + "A majestic lion strides across the golden savanna, its powerful frame " + "glistening under the warm afternoon sun. The tall grass ripples gently in " + "the breeze, enhancing the lion's commanding presence. The tone is vibrant, " + "embodying the raw energy of the wild. Low angle, steady tracking shot, " + "cinematic." + ) + + print(f"\nGenerating second video...") + print(f"Prompt: {prompt2}") + + video2 = generator.generate_video( + prompt2, + output_path=OUTPUT_PATH, + save_video=True, + ) + print("✓ Second video generated successfully with FP4 quantization!") + + except Exception as e: + print(f"Error during video generation: {e}") + return + + print(f"Videos saved to: {OUTPUT_PATH}") + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/examples/inference/optimizations/sageattn_inference_example_custom_weights_5090_16.py b/examples/inference/optimizations/sageattn_inference_example_custom_weights_5090_16.py new file mode 100644 index 000000000..ab16e1e27 --- /dev/null +++ b/examples/inference/optimizations/sageattn_inference_example_custom_weights_5090_16.py @@ -0,0 +1,54 @@ +from fastvideo import VideoGenerator +import os +from pathlib import Path +# from fastvideo.configs.sample import SamplingParam + +OUTPUT_PATH = "video_samples" +os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SAGE_ATTN_THREE" + +CHECKPOINT_PATH = Path(__file__).parent.parent.parent + +def main(): + # FastVideo will automatically use the optimal default arguments for the + # model. + # If a local path is provided, FastVideo will make a best effort + # attempt to identify the optimal arguments. + generator = VideoGenerator.from_pretrained( + "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", + # FastVideo will automatically handle distributed setup + num_gpus=1, + use_fsdp_inference=True, + dit_cpu_offload=False, + vae_cpu_offload=False, + text_encoder_cpu_offload=True, + pin_cpu_memory=False, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # image_encoder_cpu_offload=False, + # Load custom weights from checkpoint-5000 + init_weights_from_safetensors="/workspace/FastVideo-Quantization/M2_16" + ) + + # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers") + # sampling_param.num_frames = 45 + # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + # Generate videos with the same simple API, regardless of GPU count + prompt = ( + "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " + "wide with interest. The playful yet serene atmosphere is complemented by soft " + "natural light filtering through the petals. Mid-shot, warm and cheerful tones." + ) + video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True) + # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/") + + # Generate another video with a different prompt, without reloading the + # model! + prompt2 = ( + "A majestic lion strides across the golden savanna, its powerful frame " + "glistening under the warm afternoon sun. The tall grass ripples gently in " + "the breeze, enhancing the lion's commanding presence. The tone is vibrant, " + "embodying the raw energy of the wild. Low angle, steady tracking shot, " + "cinematic.") + video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True) + + +if __name__ == "__main__": + main() diff --git a/examples/inference/optimizations/sageattn_inference_example_custom_weights_5090_32.py b/examples/inference/optimizations/sageattn_inference_example_custom_weights_5090_32.py new file mode 100644 index 000000000..f69ee031a --- /dev/null +++ b/examples/inference/optimizations/sageattn_inference_example_custom_weights_5090_32.py @@ -0,0 +1,54 @@ +from fastvideo import VideoGenerator +import os +from pathlib import Path +# from fastvideo.configs.sample import SamplingParam + +OUTPUT_PATH = "video_samples" +os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SAGE_ATTN_THREE" + +CHECKPOINT_PATH = Path(__file__).parent.parent.parent + +def main(): + # FastVideo will automatically use the optimal default arguments for the + # model. + # If a local path is provided, FastVideo will make a best effort + # attempt to identify the optimal arguments. + generator = VideoGenerator.from_pretrained( + "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", + # FastVideo will automatically handle distributed setup + num_gpus=1, + use_fsdp_inference=True, + dit_cpu_offload=False, + vae_cpu_offload=False, + text_encoder_cpu_offload=True, + pin_cpu_memory=False, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # image_encoder_cpu_offload=False, + # Load custom weights from checkpoint-5000 + init_weights_from_safetensors="/workspace/FastVideo-Quantization/M2_32" + ) + + # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers") + # sampling_param.num_frames = 45 + # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + # Generate videos with the same simple API, regardless of GPU count + prompt = ( + "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " + "wide with interest. The playful yet serene atmosphere is complemented by soft " + "natural light filtering through the petals. Mid-shot, warm and cheerful tones." + ) + video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True) + # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/") + + # Generate another video with a different prompt, without reloading the + # model! + prompt2 = ( + "A majestic lion strides across the golden savanna, its powerful frame " + "glistening under the warm afternoon sun. The tall grass ripples gently in " + "the breeze, enhancing the lion's commanding presence. The tone is vibrant, " + "embodying the raw energy of the wild. Low angle, steady tracking shot, " + "cinematic.") + video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True) + + +if __name__ == "__main__": + main() diff --git a/examples/inference/sageattn+fp4linear.py b/examples/inference/sageattn+fp4linear.py new file mode 100644 index 000000000..dceb6a3c2 --- /dev/null +++ b/examples/inference/sageattn+fp4linear.py @@ -0,0 +1,49 @@ +from fastvideo import VideoGenerator +import os +# from fastvideo.configs.sample import SamplingParam + +OUTPUT_PATH = "video_samples" +os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SAGE_ATTN_THREE" + +def main(): + # FastVideo will automatically use the optimal default arguments for the + # model. + # If a local path is provided, FastVideo will make a best effort + # attempt to identify the optimal arguments. + generator = VideoGenerator.from_pretrained( + "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", + # FastVideo will automatically handle distributed setup + num_gpus=1, + use_fsdp_inference=True, + dit_cpu_offload=False, + vae_cpu_offload=False, + text_encoder_cpu_offload=True, + pin_cpu_memory=False, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # image_encoder_cpu_offload=False, + ) + + # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers") + # sampling_param.num_frames = 45 + # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + # Generate videos with the same simple API, regardless of GPU count + prompt = ( + "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " + "wide with interest. The playful yet serene atmosphere is complemented by soft " + "natural light filtering through the petals. Mid-shot, warm and cheerful tones." + ) + video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True) + # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/") + + # Generate another video with a different prompt, without reloading the + # model! + prompt2 = ( + "A majestic lion strides across the golden savanna, its powerful frame " + "glistening under the warm afternoon sun. The tall grass ripples gently in " + "the breeze, enhancing the lion's commanding presence. The tone is vibrant, " + "embodying the raw energy of the wild. Low angle, steady tracking shot, " + "cinematic.") + video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True) + + +if __name__ == "__main__": + main() diff --git a/examples/inference/sageattn_inference_example.py b/examples/inference/sageattn_inference_example.py new file mode 100644 index 000000000..dceb6a3c2 --- /dev/null +++ b/examples/inference/sageattn_inference_example.py @@ -0,0 +1,49 @@ +from fastvideo import VideoGenerator +import os +# from fastvideo.configs.sample import SamplingParam + +OUTPUT_PATH = "video_samples" +os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SAGE_ATTN_THREE" + +def main(): + # FastVideo will automatically use the optimal default arguments for the + # model. + # If a local path is provided, FastVideo will make a best effort + # attempt to identify the optimal arguments. + generator = VideoGenerator.from_pretrained( + "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", + # FastVideo will automatically handle distributed setup + num_gpus=1, + use_fsdp_inference=True, + dit_cpu_offload=False, + vae_cpu_offload=False, + text_encoder_cpu_offload=True, + pin_cpu_memory=False, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # image_encoder_cpu_offload=False, + ) + + # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers") + # sampling_param.num_frames = 45 + # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + # Generate videos with the same simple API, regardless of GPU count + prompt = ( + "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " + "wide with interest. The playful yet serene atmosphere is complemented by soft " + "natural light filtering through the petals. Mid-shot, warm and cheerful tones." + ) + video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True) + # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/") + + # Generate another video with a different prompt, without reloading the + # model! + prompt2 = ( + "A majestic lion strides across the golden savanna, its powerful frame " + "glistening under the warm afternoon sun. The tall grass ripples gently in " + "the breeze, enhancing the lion's commanding presence. The tone is vibrant, " + "embodying the raw energy of the wild. Low angle, steady tracking shot, " + "cinematic.") + video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True) + + +if __name__ == "__main__": + main() diff --git a/examples/inference/sageattn_inference_example_custom_weights.py b/examples/inference/sageattn_inference_example_custom_weights.py new file mode 100644 index 000000000..5457a7325 --- /dev/null +++ b/examples/inference/sageattn_inference_example_custom_weights.py @@ -0,0 +1,54 @@ +from fastvideo import VideoGenerator +import os +from pathlib import Path +# from fastvideo.configs.sample import SamplingParam + +OUTPUT_PATH = "video_samples" +os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SAGE_ATTN_THREE" + +CHECKPOINT_PATH = Path(__file__).parent.parent.parent / "checkpoints" / "wan_t2v_finetune_qat" / "checkpoint-5000" / "transformer" + +def main(): + # FastVideo will automatically use the optimal default arguments for the + # model. + # If a local path is provided, FastVideo will make a best effort + # attempt to identify the optimal arguments. + generator = VideoGenerator.from_pretrained( + "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", + # FastVideo will automatically handle distributed setup + num_gpus=1, + use_fsdp_inference=True, + dit_cpu_offload=False, + vae_cpu_offload=False, + text_encoder_cpu_offload=True, + pin_cpu_memory=False, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # image_encoder_cpu_offload=False, + # Load custom weights from checkpoint-5000 + init_weights_from_safetensors=str(CHECKPOINT_PATH), + ) + + # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers") + # sampling_param.num_frames = 45 + # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + # Generate videos with the same simple API, regardless of GPU count + prompt = ( + "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " + "wide with interest. The playful yet serene atmosphere is complemented by soft " + "natural light filtering through the petals. Mid-shot, warm and cheerful tones." + ) + video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True) + # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/") + + # Generate another video with a different prompt, without reloading the + # model! + prompt2 = ( + "A majestic lion strides across the golden savanna, its powerful frame " + "glistening under the warm afternoon sun. The tall grass ripples gently in " + "the breeze, enhancing the lion's commanding presence. The tone is vibrant, " + "embodying the raw energy of the wild. Low angle, steady tracking shot, " + "cinematic.") + video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True) + + +if __name__ == "__main__": + main() diff --git a/examples/inference/sageattn_inference_example_custom_weights_5090.py b/examples/inference/sageattn_inference_example_custom_weights_5090.py new file mode 100644 index 000000000..ecb51a732 --- /dev/null +++ b/examples/inference/sageattn_inference_example_custom_weights_5090.py @@ -0,0 +1,54 @@ +from fastvideo import VideoGenerator +import os +from pathlib import Path +# from fastvideo.configs.sample import SamplingParam + +OUTPUT_PATH = "video_samples" +os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SAGE_ATTN_THREE" + +CHECKPOINT_PATH = Path(__file__).parent.parent.parent + +def main(): + # FastVideo will automatically use the optimal default arguments for the + # model. + # If a local path is provided, FastVideo will make a best effort + # attempt to identify the optimal arguments. + generator = VideoGenerator.from_pretrained( + "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", + # FastVideo will automatically handle distributed setup + num_gpus=1, + use_fsdp_inference=True, + dit_cpu_offload=False, + vae_cpu_offload=False, + text_encoder_cpu_offload=True, + pin_cpu_memory=False, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # image_encoder_cpu_offload=False, + # Load custom weights from checkpoint-5000 + init_weights_from_safetensors="/workspace/FastVideo-Quantization/M2_8" + ) + + # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers") + # sampling_param.num_frames = 45 + # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + # Generate videos with the same simple API, regardless of GPU count + prompt = ( + "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " + "wide with interest. The playful yet serene atmosphere is complemented by soft " + "natural light filtering through the petals. Mid-shot, warm and cheerful tones." + ) + video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True) + # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/") + + # Generate another video with a different prompt, without reloading the + # model! + prompt2 = ( + "A majestic lion strides across the golden savanna, its powerful frame " + "glistening under the warm afternoon sun. The tall grass ripples gently in " + "the breeze, enhancing the lion's commanding presence. The tone is vibrant, " + "embodying the raw energy of the wild. Low angle, steady tracking shot, " + "cinematic.") + video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True) + + +if __name__ == "__main__": + main() diff --git a/examples/inference/sageattn_inference_example_hy15.py b/examples/inference/sageattn_inference_example_hy15.py new file mode 100644 index 000000000..02ee98cc8 --- /dev/null +++ b/examples/inference/sageattn_inference_example_hy15.py @@ -0,0 +1,49 @@ +from fastvideo import VideoGenerator +import os +# from fastvideo.configs.sample import SamplingParam + +OUTPUT_PATH = "video_samples" +os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SAGE_ATTN_THREE" + +def main(): + # FastVideo will automatically use the optimal default arguments for the + # model. + # If a local path is provided, FastVideo will make a best effort + # attempt to identify the optimal arguments. + generator = VideoGenerator.from_pretrained( + "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v", + # FastVideo will automatically handle distributed setup + num_gpus=1, + use_fsdp_inference=True, + dit_cpu_offload=False, + vae_cpu_offload=False, + text_encoder_cpu_offload=True, + pin_cpu_memory=False, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # image_encoder_cpu_offload=False, + ) + + # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers") + # sampling_param.num_frames = 45 + # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + # Generate videos with the same simple API, regardless of GPU count + prompt = ( + "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " + "wide with interest. The playful yet serene atmosphere is complemented by soft " + "natural light filtering through the petals. Mid-shot, warm and cheerful tones." + ) + video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True) + # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/") + + # Generate another video with a different prompt, without reloading the + # model! + prompt2 = ( + "A majestic lion strides across the golden savanna, its powerful frame " + "glistening under the warm afternoon sun. The tall grass ripples gently in " + "the breeze, enhancing the lion's commanding presence. The tone is vibrant, " + "embodying the raw energy of the wild. Low angle, steady tracking shot, " + "cinematic.") + video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True) + + +if __name__ == "__main__": + main() diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh index 3dd2f77ec..3ba86ac51 100644 --- a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh @@ -1,27 +1,47 @@ #!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune +#SBATCH --partition=all +#SBATCH --nodes=1 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=1 +#SBATCH --output=logs/wan_t2v_1.3B_finetune.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune.err + +source .venv/bin/activate + export WANDB_BASE_URL="https://api.wandb.ai" export WANDB_MODE=online export TOKENIZERS_PARALLELISM=false # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" -DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/" -VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json" -NUM_GPUS=4 +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=1 # export CUDA_VISIBLE_DEVICES=4,5 +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# Launch ONE torchrun per node (via srun) and let torchrun spawn 4 workers per node. +MASTER_ADDR="$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)" +MASTER_PORT="${MASTER_PORT:-29500}" +export MASTER_ADDR MASTER_PORT + # Training arguments training_args=( - --tracker_project_name "wan_t2v_finetune" - --output_dir "checkpoints/wan_t2v_finetune" - --max_train_steps 5000 + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_1.3B_77" + --max_train_steps 4000 --train_batch_size 1 --train_sp_batch_size 1 - --gradient_accumulation_steps 8 + --gradient_accumulation_steps 1 --num_latent_t 20 - --num_height 480 + --num_height 448 --num_width 832 --num_frames 77 --enable_gradient_checkpointing_type "full" @@ -30,7 +50,7 @@ training_args=( # Parallel arguments parallel_args=( --num_gpus $NUM_GPUS - --sp_size $NUM_GPUS + --sp_size 1 --tp_size 1 --hsdp_replicate_dim 1 --hsdp_shard_dim $NUM_GPUS @@ -45,7 +65,7 @@ model_args=( # Dataset arguments dataset_args=( --data_path $DATA_DIR - --dataloader_num_workers 1 + --dataloader_num_workers 4 ) # Validation arguments @@ -54,16 +74,16 @@ validation_args=( --validation_dataset_file $VALIDATION_DATASET_FILE --validation_steps 200 --validation_sampling_steps "50" - --validation_guidance_scale "3.0" + --validation_guidance_scale "5.0" ) # Optimizer arguments optimizer_args=( - --learning_rate 5e-5 + --learning_rate 1e-6 --mixed_precision "bf16" --weight_only_checkpointing_steps 1000 --training_state_checkpointing_steps 1000 - --weight_decay 1e-4 + --weight_decay 0.01 --max_grad_norm 1.0 ) @@ -72,23 +92,24 @@ miscellaneous_args=( --inference_mode False --checkpoints_total_limit 3 --training_cfg_rate 0.1 - --multi_phased_distill_schedule "4000-1" - --not_apply_cfg_solver --dit_precision "fp32" - --num_euler_timesteps 50 --ema_start_step 0 - --enable_gradient_checkpointing_type "full" - # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500" + --flow_shift 5 + --seed 1000 ) -torchrun \ - --nnodes 1 \ - --nproc_per_node $NUM_GPUS \ - fastvideo/training/wan_training_pipeline.py \ - "${parallel_args[@]}" \ - "${model_args[@]}" \ - "${dataset_args[@]}" \ - "${training_args[@]}" \ - "${optimizer_args[@]}" \ - "${validation_args[@]}" \ - "${miscellaneous_args[@]}" +srun --nodes="$SLURM_NNODES" --ntasks="$SLURM_NNODES" --ntasks-per-node=1 \ + torchrun \ + --nnodes "$SLURM_NNODES" \ + --nproc_per_node 4 \ + --rdzv_backend c10d \ + --rdzv_endpoint "${MASTER_ADDR}:${MASTER_PORT}" \ + --rdzv_id "$SLURM_JOB_ID" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_fp4_e2e.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_fp4_e2e.sh new file mode 100644 index 000000000..9bc207f98 --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_fp4_e2e.sh @@ -0,0 +1,98 @@ +#!/bin/bash + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=1 +# export CUDA_VISIBLE_DEVICES=4,5 + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_fp4_e2e_77" + --max_train_steps 5000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" + --generator_4bit_linear True + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size $NUM_GPUS + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 100 + --validation_sampling_steps "50" + --validation_guidance_scale "6.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 500 + --training_state_checkpointing_steps 500 + --weight_decay 1e-4 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --multi_phased_distill_schedule "4000-1" + --not_apply_cfg_solver + --dit_precision "fp32" + --num_euler_timesteps 50 + --ema_start_step 0 + --enable_gradient_checkpointing_type "full" + # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500" +) + +torchrun \ + --nnodes 1 \ + --nproc_per_node $NUM_GPUS \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_fp4_linear.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_fp4_linear.sh new file mode 100644 index 000000000..c86825555 --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_fp4_linear.sh @@ -0,0 +1,97 @@ +#!/bin/bash + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=1 +# export CUDA_VISIBLE_DEVICES=4,5 + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_fp4_linear_77" + --max_train_steps 5000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" + --generator_4bit_linear True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size $NUM_GPUS + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 100 + --validation_sampling_steps "50" + --validation_guidance_scale "6.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 500 + --training_state_checkpointing_steps 500 + --weight_decay 1e-4 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --multi_phased_distill_schedule "4000-1" + --not_apply_cfg_solver + --dit_precision "fp32" + --num_euler_timesteps 50 + --ema_start_step 0 + --enable_gradient_checkpointing_type "full" + # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500" +) + +torchrun \ + --nnodes 1 \ + --nproc_per_node $NUM_GPUS \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_nvfp4_5090.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_nvfp4_5090.sh new file mode 100644 index 000000000..56c7da9fd --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_nvfp4_5090.sh @@ -0,0 +1,94 @@ +#!/bin/bash + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR="data/Wan-Syn_77x448x832_600k/train/" +VALIDATION_DATASET_FILE="data/Wan-Syn_77x448x832_600k/validation_6.json" +NUM_GPUS=6 +# export CUDA_VISIBLE_DEVICES=4,5 + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune" + --output_dir "checkpoints/wan_t2v_finetune" + --max_train_steps 5000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 6 + --num_height 480 + --num_width 832 + --num_frames 21 + --enable_gradient_checkpointing_type "full" + --generator_4bit_linear True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 1 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 50 + --validation_sampling_steps "50" + --validation_guidance_scale "6.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --checkpointing_steps 2000 + --weight_decay 1e-4 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --multi_phased_distill_schedule "4000-1" + --not_apply_cfg_solver + --dit_precision "fp32" + --num_euler_timesteps 50 + --ema_start_step 0 + --enable_gradient_checkpointing_type "full" + # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500" +) + +torchrun \ + --nnodes 1 \ + --nproc_per_node $NUM_GPUS \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_16.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_16.sh new file mode 100644 index 000000000..814000544 --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_16.sh @@ -0,0 +1,125 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_16 +#SBATCH --partition=all +#SBATCH --nodes=4 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=1 +#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_16.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_16.err + +source .venv/bin/activate + + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +# Use node-local Triton cache to avoid stale file handle errors on shared filesystems +export TRITON_CACHE_DIR="/tmp/triton_cache_${SLURM_JOB_ID}_${SLURM_NODEID}" +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data2/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=16 +# export CUDA_VISIBLE_DEVICES=4,5 + +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# 1. Get the hostname of the first node (Master) +nodes=( $( scontrol show hostnames $SLURM_JOB_NODELIST ) ) +nodes_array=($nodes) +head_node=${nodes_array[0]} +MASTER_ADDR=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address) +MASTER_PORT=29500 + +# 2. Get the node count automatically +NNODES=$SLURM_NNODES +GPUS_PER_NODE=$SLURM_GPUS_ON_NODE +NUM_GPUS=$((NNODES * GPUS_PER_NODE)) + +echo "MASTER_ADDR=$MASTER_ADDR MASTER_PORT=$MASTER_PORT NNODES=$NNODES" + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat_16_gsf_qkv" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" # if OOM enable this + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim $NUM_GPUS + --hsdp_shard_dim 1 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 200 + --training_state_checkpointing_steps 200 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 1 + --seed 1000 +) + +srun torchrun \ + --nnodes $NNODES \ + --nproc_per_node $GPUS_PER_NODE \ + --node_rank $SLURM_PROCID \ + --rdzv_backend c10d \ + --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" \ No newline at end of file diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_16_M2.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_16_M2.sh new file mode 100644 index 000000000..ffc7220e3 --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_16_M2.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_16 +#SBATCH --partition=main +#SBATCH --nodes=2 +#SBATCH --ntasks-per-node=1 +#SBATCH --gres=gpu:8 +#SBATCH --cpus-per-task=128 +#SBATCH --mem=1440G +#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_16.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_16.err +#SBATCH --exclusive + +source ~/conda/miniconda/bin/activate +conda activate matthew-fv + +# Basic Info +export WANDB_MODE="online" +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +# different cache dir for different processes +export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID} +export MASTER_PORT=29500 +export NODE_RANK=$SLURM_PROCID +nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) ) +export MASTER_ADDR=${nodes[0]} +export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID +export TOKENIZERS_PARALLELISM=false +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +echo "MASTER_ADDR: $MASTER_ADDR" +echo "NODE_RANK: $NODE_RANK" + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=/mnt/weka/home/hao.zhang/wl/sharefs/Vchitect-2M/Wan-Syn-upload/latents_i2v/train +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS_PER_NODE=8 +TOTAL_GPUS=$((NUM_GPUS_PER_NODE * SLURM_JOB_NUM_NODES)) +# export CUDA_VISIBLE_DEVICES=4,5 + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat_16" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" # if OOM enable this + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $TOTAL_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim $TOTAL_GPUS + --hsdp_shard_dim 1 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 200 + --training_state_checkpointing_steps 200 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 1 + --seed 1000 +) + +srun torchrun \ +--nnodes $SLURM_JOB_NUM_NODES \ +--nproc_per_node $NUM_GPUS_PER_NODE \ +--node_rank $SLURM_PROCID \ +--rdzv_backend=c10d \ +--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_32.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_32.sh new file mode 100644 index 000000000..f894354a1 --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_32.sh @@ -0,0 +1,116 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_32 +#SBATCH --partition=all +#SBATCH --nodes=8 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=1 +#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_32.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_32.err + +source .venv/bin/activate + + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=32 +# export CUDA_VISIBLE_DEVICES=4,5 + +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# Launch ONE torchrun per node (via srun) and let torchrun spawn 4 workers per node. +MASTER_ADDR="$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)" +MASTER_PORT="${MASTER_PORT:-29500}" +export MASTER_ADDR MASTER_PORT + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat_32" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" # if OOM enable this + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim $NUM_GPUS + --hsdp_shard_dim 1 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 200 + --training_state_checkpointing_steps 200 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 1 + --seed 1000 +) + +srun --nodes="$SLURM_NNODES" --ntasks="$SLURM_NNODES" --ntasks-per-node=1 \ + torchrun \ + --nnodes "$SLURM_NNODES" \ + --nproc_per_node 4 \ + --rdzv_backend c10d \ + --rdzv_endpoint "${MASTER_ADDR}:${MASTER_PORT}" \ + --rdzv_id "$SLURM_JOB_ID" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_32_M2.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_32_M2.sh new file mode 100644 index 000000000..b16aaa17d --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_32_M2.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_32 +#SBATCH --partition=main +#SBATCH --nodes=4 +#SBATCH --ntasks-per-node=1 +#SBATCH --gres=gpu:8 +#SBATCH --cpus-per-task=128 +#SBATCH --mem=1440G +#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_32.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_32.err +#SBATCH --exclusive + +source ~/conda/miniconda/bin/activate +conda activate matthew-fv + +# Basic Info +export WANDB_MODE="online" +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +# different cache dir for different processes +export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID} +export MASTER_PORT=29500 +export NODE_RANK=$SLURM_PROCID +nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) ) +export MASTER_ADDR=${nodes[0]} +export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID +export TOKENIZERS_PARALLELISM=false +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +echo "MASTER_ADDR: $MASTER_ADDR" +echo "NODE_RANK: $NODE_RANK" + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=/mnt/weka/home/hao.zhang/wl/sharefs/Vchitect-2M/Wan-Syn-upload/latents_i2v/train +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS_PER_NODE=8 +TOTAL_GPUS=$((NUM_GPUS_PER_NODE * SLURM_JOB_NUM_NODES)) +# export CUDA_VISIBLE_DEVICES=4,5 + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat_32" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" # if OOM enable this + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $TOTAL_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim $TOTAL_GPUS + --hsdp_shard_dim 1 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 200 + --training_state_checkpointing_steps 200 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 1 + --seed 1000 +) + +srun torchrun \ +--nnodes $SLURM_JOB_NUM_NODES \ +--nproc_per_node $NUM_GPUS_PER_NODE \ +--node_rank $SLURM_PROCID \ +--rdzv_backend=c10d \ +--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_5090.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_5090.sh new file mode 100644 index 000000000..3ff8498bb --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_5090.sh @@ -0,0 +1,97 @@ +#!/bin/bash + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=4 +# export CUDA_VISIBLE_DEVICES=4,5 + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat" + --max_train_steps 5000 + --train_batch_size 8 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 12 + --num_height 480 + --num_width 832 + --num_frames 45 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size $NUM_GPUS + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 100 + --validation_sampling_steps "50" + --validation_guidance_scale "6.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 500 + --training_state_checkpointing_steps 500 + --weight_decay 1e-4 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --multi_phased_distill_schedule "4000-1" + --not_apply_cfg_solver + --dit_precision "fp32" + --num_euler_timesteps 50 + --ema_start_step 0 + --enable_gradient_checkpointing_type "full" + # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500" +) + +torchrun \ + --nnodes 1 \ + --nproc_per_node $NUM_GPUS \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8.sh new file mode 100644 index 000000000..21008bade --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8.sh @@ -0,0 +1,119 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_8 +#SBATCH --partition=all +#SBATCH --nodes=2 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=1 +#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_8.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_8.err + +source .venv/bin/activate + + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# Use node-local Triton cache to avoid stale file handle errors on shared filesystems +export TRITON_CACHE_DIR="/tmp/triton_cache_${SLURM_JOB_ID}_${SLURM_NODEID}" + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=8 +# export CUDA_VISIBLE_DEVICES=4,5 + +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# Launch ONE torchrun per node (via srun) and let torchrun spawn 4 workers per node. +MASTER_ADDR="$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)" +MASTER_PORT="${MASTER_PORT:-29500}" +export MASTER_ADDR MASTER_PORT + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat_8" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" # if OOM enable this + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim $NUM_GPUS + --hsdp_shard_dim 1 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 200 + --training_state_checkpointing_steps 200 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 1 + --seed 1000 +) + +srun --nodes="$SLURM_NNODES" --ntasks="$SLURM_NNODES" --ntasks-per-node=1 \ + torchrun \ + --nnodes "$SLURM_NNODES" \ + --nproc_per_node 4 \ + --rdzv_backend c10d \ + --rdzv_endpoint "${MASTER_ADDR}:${MASTER_PORT}" \ + --rdzv_id "$SLURM_JOB_ID" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8_M2.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8_M2.sh new file mode 100644 index 000000000..13dd80f3a --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8_M2.sh @@ -0,0 +1,123 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_8 +#SBATCH --partition=main +#SBATCH --nodes=1 +#SBATCH --ntasks-per-node=1 +#SBATCH --gres=gpu:8 +#SBATCH --cpus-per-task=128 +#SBATCH --mem=1440G +#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_8.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_8.err +#SBATCH --exclusive + +source ~/conda/miniconda/bin/activate +conda activate matthew-fv + +# Basic Info +export WANDB_MODE="online" +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +# different cache dir for different processes +export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID} +export MASTER_PORT=29500 +export NODE_RANK=$SLURM_PROCID +nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) ) +export MASTER_ADDR=${nodes[0]} +export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID +export TOKENIZERS_PARALLELISM=false +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +echo "MASTER_ADDR: $MASTER_ADDR" +echo "NODE_RANK: $NODE_RANK" + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=/mnt/weka/home/hao.zhang/wl/sharefs/Vchitect-2M/Wan-Syn-upload/latents_i2v/train +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=8 +# export CUDA_VISIBLE_DEVICES=4,5 + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat_8" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" # if OOM enable this + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim $NUM_GPUS + --hsdp_shard_dim 1 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 200 + --training_state_checkpointing_steps 200 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 1 + --seed 1000 +) + +srun torchrun \ +--nnodes $SLURM_JOB_NUM_NODES \ +--nproc_per_node $NUM_GPUS \ +--node_rank $SLURM_PROCID \ +--rdzv_backend=c10d \ +--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8_NVL72.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8_NVL72.sh new file mode 100644 index 000000000..7eb2db8bd --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_8_NVL72.sh @@ -0,0 +1,121 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_8_NVL72 +#SBATCH --nodes=4 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=1 +#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_8_NVL72.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_8_NVL72.err + +source .venv/bin/activate + + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +# export CUDA_VISIBLE_DEVICES=4,5 + +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# 1. Get the hostname of the first node (Master) +nodes=( $( scontrol show hostnames $SLURM_JOB_NODELIST ) ) +nodes_array=($nodes) +head_node=${nodes_array[0]} +MASTER_ADDR=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address) +MASTER_PORT=29500 + +# 2. Get the node count automatically +NNODES=$SLURM_NNODES +GPUS_PER_NODE=$SLURM_GPUS_ON_NODE +NUM_GPUS=$((NNODES * GPUS_PER_NODE)) + +echo "MASTER_ADDR=$MASTER_ADDR MASTER_PORT=$MASTER_PORT NNODES=$NNODES" + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_14B" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 768 + --num_width 1280 + --num_frames 77 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 4 + --tp_size 1 + --hsdp_replicate_dim 4 + --hsdp_shard_dim 4 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + # --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 1000 + --training_state_checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 5 + --seed 1000 +) + +srun torchrun \ + --nnodes $NNODES \ + --nproc_per_node $GPUS_PER_NODE \ + --node_rank $SLURM_PROCID \ + --rdzv_backend c10d \ + --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_sage3_5090.sh b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_sage3_5090.sh new file mode 100644 index 000000000..bcfd77cd4 --- /dev/null +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_sage3_5090.sh @@ -0,0 +1,94 @@ +#!/bin/bash + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=4 +# export CUDA_VISIBLE_DEVICES=4,5 + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_sage3" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 16 + --num_latent_t 20 + --num_height 448 + --num_width 832 + --num_frames 77 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 4 + --tp_size 1 + --hsdp_replicate_dim $NUM_GPUS + --hsdp_shard_dim 1 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 200 + --training_state_checkpointing_steps 200 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 1 + --seed 1000 +) + +torchrun \ + --nnodes 1 \ + --nproc_per_node $NUM_GPUS \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json b/examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json index 674c3cb62..d324f5ee6 100644 --- a/examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json +++ b/examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json @@ -1,31 +1,131 @@ { "data": [ { - "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.", - "image_path": null, - "video_path": null, - "num_inference_steps": 50, - "height": 480, + "caption": "In the video, a woman is elegantly showcasing her earrings, bringing attention to their intricate design with a gentle touch of her fingers. She is bathed in ambient purple and pink lighting, which casts a soft glow on her delicate features and enhances the vivid tones of her lipstick and eye makeup. Her hair is styled to frame her face smoothly, emphasizing the contours of her jawline and cheekbones. The background features a blurred neon light, adding an artistic and modern touch to the overall aesthetic.", + "video_path": "Fashion/mixkit-face-of-an-elegant-and-captivating-woman-41914_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, "width": 832, "num_frames": 77 }, { - "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.", - "image_path": null, - "video_path": null, - "num_inference_steps": 50, - "height": 480, + "caption": "In the video, a lone rider guides a majestic horse across an expansive, open field as the sun sets in the background. The rider, dressed in a classic blue shirt and wide-brimmed hat, sits confidently in the saddle, silhouetted against the warm glow of the evening sky. The horse moves gracefully, its mane and tail flowing with each step, creating a sense of harmony between horse and rider. Surrounding the pair, towering trees form a natural border, their leaves gently rustling in the breeze. The shadows lengthen on the ground, accentuating the serene and timeless feel of the scene. The distant hills and wooden fences frame the horizon, adding depth to the tranquil landscape. A few horses graze peacefully in the background, blending into the pastoral setting. The overall ambiance evokes a sense of calmness and quietude, capturing a perfect moment in the golden light of dusk.", + "video_path": "Man/mixkit-a-rancher-riding-a-horse-at-sunset-1143_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, "width": 832, "num_frames": 77 }, { - "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.", - "image_path": null, - "video_path": null, - "num_inference_steps": 50, - "height": 480, + "caption": "In a dimly lit, eerie setting, a mysterious pink bottle labeled \"Authentic 100% organic POISON\" sits prominently in the foreground, casting a menacing aura. The bottle is accentuated by green fog, which swirls lightly around it, enhancing its sinister allure. Behind it, a shadowy golden bottle adorned with a spider emblem subtly emerges, adding an extra layer of mystery to the scene. Dim candles provide faint, flickering light, which complements the dark atmosphere, making the setting ideal for an illusion of hidden dangers.", + "video_path": "smoke/mixkit-poison-in-halloween-ritual-33879_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, "width": 832, "num_frames": 77 - } - ] + }, + { + "caption": "The video opens with a tranquil scene in the heart of a dense forest, emphasizing two large, textured tree trunks in the foreground framing the view. Sunlight filters through the canopy above, casting intricate patterns of light and shadow on the trees and the ground. Between the tree trunks, a clear view of a calm, muddy river unfolds, its surface shimmering under the gentle sunlight. The riverbank is decorated with a variety of small bushes and vibrant foliage, subtly transitioning into the deep greens of tall, leafy plants. In the background, the dense forest looms, filled with dark, towering trees, their branches intertwining to form an intricate canopy. The scene is bathed in the soft glow of the sun, creating a serene and picturesque setting. Occasional sunbeams pierce through the foliage, adding a magical aura to the landscape. The vibrant reds and oranges of the smaller plants add contrast, bringing warmth to the earthy tones of the scenery. Overall, this harmonious blend of natural elements creates a peaceful and idyllic forest setting.", + "video_path": "forest/mixkit-view-of-a-river-between-two-old-trees-560_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "In the video, a martial artist dressed in a traditional white uniform with a black belt demonstrates a series of precise movements against a stark black background. The individual gracefully transitions between stances, embodying a sense of focused discipline and control. Each motion is executed with a deliberate pace, showcasing the fluidity of martial arts techniques. The soft lighting creates subtle highlights on the uniform, adding depth to the figure as it moves. The practitioner begins with an open-hand pose, feet firmly grounded, gradually shifting to a powerful forward punch. The fluidity of the sequence displays a mastery of balance and poise. Every trajectory of the limbs is precise and deliberate, capturing the elegance and strength of martial arts. The serene, isolated setting enhances the intensity and concentration of the practitioner. This visual presentation is an elegant interplay of motion and stillness, displaying the art form's discipline and grace.", + "video_path": "Man/mixkit-a-young-man-practicing-his-karate-moves-49635_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "A tranquil coastal scene unfolds with a drone's aerial view capturing a serene beach landscape. The camera glides over a quiet stretch of sandy shoreline, where gentle waves kiss the shore under a clear blue sky. Nestled amidst lush palm trees are a series of traditional thatched-roof huts, their earthy tones blending harmoniously with the natural surroundings. The sandy beach stretches endlessly, bordered by the rhythmic dance of ocean waves on one side and verdant greenery on the other. A pair of white umbrellas is set up on the sand, suggesting a place to relax and enjoy the sun. In the distance, two small human figures can be seen walking leisurely along the water's edge, leaving faint footprints behind them. The scene exudes a calm and inviting atmosphere, with the soft rustle of palm leaves and the whisper of the ocean breeze almost audible. The overall composition is a captivating blend of nature's tranquility and architectural simplicity. This picturesque setting invites viewers to imagine themselves steps away from this idyllic coastal escape.", + "video_path": "beach/mixkit-sunny-beach-in-a-dynamic-shot-from-a-drone-44383_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "A lone figure stands on a large, moss-covered rock, surrounded by the soft rush of a nearby stream. The figure is wearing white sneakers and shorts, with a plaid shirt that hangs loosely in the breeze. The lighting creates dramatic shadows, enhancing the textures of the rock and the subtle movement of the water below. In the background, a waterfall cascades into the stream, completing this tranquil and serene nature scene.", + "video_path": "forest/mixkit-woman-standing-in-front-of-waterfall-559_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "In an industrial setting, a person leans casually against a railing, exuding a sense of confidence and composure. They are wearing a striking outfit, consisting of a vibrant, patterned jacket over a simple white crop top, creating a bold contrast. The atmosphere is infused with warm, ambient lighting that casts soft shadows on the concrete walls and metallic surfaces. Intricate wiring and pipes form an intricate backdrop, enhancing the urban aesthetic. Their relaxed posture and direct, engaging gaze suggest a sense of ease in this industrial environment. This scene encapsulates a blend of modern fashion and gritty, urban architecture, creating a visually compelling narrative.", + "video_path": "Fashion/mixkit-portrait-of-a-hipster-woman-walking-down-a-stairs-1297_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "A man is energetically stretching in an open-air setting, surrounded by rows of vibrant red seats that suggest an amphitheater or outdoor venue. He wears a sleeveless black shirt layered with a hooded vest, emphasizing his athletic build as he engages in a warm-up routine. Behind him, the striking modern architecture of the building features geometric panels, with large sections of glass and overlapping metallic beams creating a dynamic backdrop. The scene captures the contrast between his focused movements and the static, bold design of the structure, while the surrounding greenery adds a touch of nature to the environment. The overall atmosphere is one of preparation and anticipation, with the man appearing determined and ready for an upcoming event or performance.", + "video_path": "Sport/mixkit-man-doing-arm-stretches-595_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "A young woman is seated on the floor in front of a plush, beige tufted couch, fully engrossed in sorting through a stack of papers. Her dark hair falls loosely past her shoulders, and she wears a green plaid shirt, contributing to the casual yet focused atmosphere. She gently places the papers onto a small round white table, occasionally lifting individual sheets to examine them more closely. Her expression shifts subtly, reflecting concentration and contemplation as she processes the information on the pages. Two small, round nested tables hold her documents, along with a small plant in a gray pot, adding a touch of greenery to the scene. The background features a dark paneled wall, creating a contrasting backdrop for the light-colored furniture. The setting is tranquil and organized, the couch and tables arranged symmetrically, conveying a sense of harmony. A calculator rests on the smaller table, hinting at a task involving calculations or budgeting.", + "video_path": "Woman/mixkit-frustrated-woman-throws-paperwork-on-the-floor-4526_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "A heavily rusted metal gate stands firmly locked, with two vertical bars joined by a thick, old chain that loops elegantly around them. The chain's texture is coarse and rugged, its surface reflecting varying shades of orange and brown, indicative of years exposed to the elements. At the heart of the chain, a black iron padlock, slightly worn yet imposing, secures the gate, its curves and edges smooth against the aged links. The gate's metalwork is outlined by a backdrop of soft, blurred greenery, suggesting a serene and isolated location beyond the barrier. Tall trees rise in the distance, their trunks and leaves creating a lush, forest-like setting that contrasts with the gate's severe rust. A pathway leads away from the gate, its surface uneven with patches of moss and weathered stone visible in the soft focus, inviting yet inaccessible. The ambiance is quiet and mysterious, with a sense of abandonment hanging subtly in the air, evoking curiosity about what lies beyond. Shadows play across the gate, cast by branches swaying gently in the breeze, adding to the dynamic interaction of light and texture. This scene, rich in detail and atmosphere, captures the viewer's imagination, evoking both the allure of the forbidden and the beauty of decay.", + "video_path": "forest/mixkit-rusty-fence-with-a-chain-of-a-property-in-nature-5294_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "In a serene and softly lit yoga studio, three individuals engage in a yoga session, each performing an upward-facing stretch. The central figure is a woman with shoulder-length brown hair, dressed in a light cropped top and green leggings, her posture reflecting grace and concentration. To her right, another participant, a woman in a purple outfit, mirrors the pose with equal poise. On her left, a person with a bun focuses intently, supported slightly by yoga blocks beneath their hands. The warm-colored wooden floor contrasts soothingly with the soft pastel mural on the back wall, featuring an abstract design and partial visage of a serene face. Natural light floods the space from a large window on the right, where lush greens peek through, adding an element of tranquility. In the corner of the room, a collection of meditation instruments, including a gong and a Buddha statue, subtly frame the peaceful setting. The mood is calm yet focused, as all three participants are deeply engaged in their practice. The scene combines elements of balance, harmony, and a shared journey towards mindfulness. This depiction captures the essence of a yoga session that blends personal growth with collective experience.", + "video_path": "People/mixkit-small-group-of-people-doing-yoga-together-43730_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "In the deep blue expanse of the ocean, two dolphins glide effortlessly, their sleek bodies reflecting the sunlight filtering through the water. The prominent shadows and caustics create a shimmering effect on their skin, capturing the beauty of their natural habitat. Each dolphin moves with a fluid grace, occasionally interacting with gentle nudges, showcasing their playful and social nature. The scene is vibrant and dynamic, with the clear blue background accentuating the dolphins' movements, making it an ideal subject for AI recreation.", + "video_path": "sea/mixkit-dolphins-underwater-4133_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "In the video, a young woman stands against a vibrant graffiti-covered wall, deeply engrossed in her smartphone. Her expression reflects a mix of focus and subtle satisfaction as she interacts with the screen. She wears a black floral-patterned top, which contrasts with the bright, abstract shapes and bold colors of the mural behind her. As she continues to engage with her phone, a series of like count notifications appear on the screen, indicating a growing online appreciation. The wall behind her features a striking mix of geometric and organic shapes, including swirls of teal, orange, and black, with large humanoid figures in a pop-art style. Her long, light-brown hair frames her face, adding a calm, composed aura amidst the lively backdrop. The video captures a blend of contemporary digital interaction and expressive urban art, creating a dynamic yet harmonious scene.", + "video_path": "Girl/mixkit-girl-looking-at-the-likes-in-her-post-4914_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "A young mother and her baby sit comfortably on a bed, surrounded by an inviting, cozy atmosphere. The woman, wearing a sleeveless top and jeans, is gently engaging with the baby, who is dressed in an adorable animal-print onesie. The child is seated on the bed with colorful toys scattered around, including a plush toy and a board book. The warm glow from a hanging lamp casts a soft light on them, enhancing the serene environment. Pillows are propped up against the headboard, providing a cushioned backdrop as the mother leans slightly over to interact with the baby. A small bottle is visible beside her, suggesting a nurturing setting. Her hand gestures animatedly as she holds up a soft, white cushion with red and blue accents, likely stimulating the baby\u2019s curiosity. Their shared moment is filled with affection and joy, a perfect snapshot of familial bonding.", + "video_path": "Baby/mixkit-loving-mother-and-her-baby-playing-with-soft-toys-49966_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + }, + { + "caption": "A young girl with long brown hair sits at a round wooden table, engrossed in working on her laptop. The laptop screen is a vivid green, suggesting a green screen effect is in use. To her left, a doll dressed in a yellow and white outfit is casually laid on top of some books, adding a playful and innocent touch to the scene. The setting is cozy, with sheer curtains in the background allowing soft natural light to spill into the room. The girl's posture and focused attention on the laptop suggest she is either playing a game or learning something new. This serene and domestic atmosphere is complemented by the slight blur of a dark couch in the foreground, framing the focused activity of the child.", + "video_path": "Girl/mixkit-little-girl-doing-homework-on-a-laptop-4757_clip_1.mp4", + "num_inference_steps": 50, + "height": 448, + "width": 832, + "num_frames": 77 + } ] } diff --git a/examples/training/finetune/wan_t2v_14B/finetune_t2v_B200.sh b/examples/training/finetune/wan_t2v_14B/finetune_t2v_B200.sh new file mode 100644 index 000000000..01d612c9b --- /dev/null +++ b/examples/training/finetune/wan_t2v_14B/finetune_t2v_B200.sh @@ -0,0 +1,115 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_14B_finetune_B200 +#SBATCH --partition=all +#SBATCH --nodes=16 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=1 +#SBATCH --output=logs/wan_t2v_14B_finetune_B200.out +#SBATCH --error=logs/wan_t2v_14B_finetune_B200.err + +source .venv/bin/activate + + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers" +DATA_DIR=data/Wan-Syn_77x768x1280_250k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_14B/validation.json" +NUM_GPUS=64 +# export CUDA_VISIBLE_DEVICES=4,5 + +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# Launch ONE torchrun per node (via srun) and let torchrun spawn 4 workers per node. +MASTER_ADDR="$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)" +MASTER_PORT="${MASTER_PORT:-29500}" +export MASTER_ADDR MASTER_PORT + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_14B" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 768 + --num_width 1280 + --num_frames 77 + --enable_gradient_checkpointing_type "full" +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + # --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 1000 + --training_state_checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 5 + --seed 1000 +) + +srun --nodes="$SLURM_NNODES" --ntasks="$SLURM_NNODES" --ntasks-per-node=1 \ + torchrun \ + --nnodes "$SLURM_NNODES" \ + --nproc_per_node 4 \ + --rdzv_backend c10d \ + --rdzv_endpoint "${MASTER_ADDR}:${MASTER_PORT}" \ + --rdzv_id "$SLURM_JOB_ID" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_14B/finetune_t2v_B200_qat_debug.sh b/examples/training/finetune/wan_t2v_14B/finetune_t2v_B200_qat_debug.sh new file mode 100644 index 000000000..2876fabb6 --- /dev/null +++ b/examples/training/finetune/wan_t2v_14B/finetune_t2v_B200_qat_debug.sh @@ -0,0 +1,125 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_14B_finetune_B200 +#SBATCH --nodes=1 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=1 +#SBATCH --output=logs/wan_t2v_14B_finetune_B200.out +#SBATCH --error=logs/wan_t2v_14B_finetune_B200.err + +source .venv/bin/activate + +export NCCL_P2P_DISABLE=1 + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# Use node-local Triton cache to avoid stale file handle errors on shared filesystems +export TRITON_CACHE_DIR="/tmp/triton_cache_${SLURM_JOB_ID}_${SLURM_NODEID}" + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +# export CUDA_VISIBLE_DEVICES=4,5 + +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# 1. Get the hostname of the first node (Master) +nodes=( $( scontrol show hostnames $SLURM_JOB_NODELIST ) ) +nodes_array=($nodes) +head_node=${nodes_array[0]} +MASTER_ADDR=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address) +MASTER_PORT=29500 + +# 2. Get the node count automatically +NNODES=$SLURM_NNODES +GPUS_PER_NODE=$SLURM_GPUS_ON_NODE +NUM_GPUS=$((NNODES * GPUS_PER_NODE)) + +echo "MASTER_ADDR=$MASTER_ADDR MASTER_PORT=$MASTER_PORT NNODES=$NNODES" + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_1.3B_debug" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 768 + --num_width 1280 + --num_frames 77 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim 4 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + # --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 1000 + --training_state_checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 5 + --seed 1000 +) + +srun torchrun \ + --nnodes $NNODES \ + --nproc_per_node $GPUS_PER_NODE \ + --node_rank $SLURM_PROCID \ + --rdzv_backend c10d \ + --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat.sh b/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat.sh new file mode 100644 index 000000000..7d1082a42 --- /dev/null +++ b/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat.sh @@ -0,0 +1,93 @@ +#!/bin/bash + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=4 +# export CUDA_VISIBLE_DEVICES=4,5 + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 21 + --num_height 720 + --num_width 1280 + --num_frames 81 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-5 + --mixed_precision "bf16" + --checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 5 + --seed 1000 +) + +torchrun \ + --nnodes 1 \ + --nproc_per_node $NUM_GPUS \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_32_14B_M2.sh b/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_32_14B_M2.sh new file mode 100644 index 000000000..b9b155ce3 --- /dev/null +++ b/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_32_14B_M2.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_16 +#SBATCH --partition=main +#SBATCH --nodes=4 +#SBATCH --ntasks-per-node=1 +#SBATCH --gres=gpu:8 +#SBATCH --cpus-per-task=128 +#SBATCH --mem=1440G +#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_16.out +#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_16.err +#SBATCH --exclusive + +source ~/conda/miniconda/bin/activate +conda activate matthew-fv + +# Basic Info +export WANDB_MODE="online" +export NCCL_P2P_DISABLE=1 +export TORCH_NCCL_ENABLE_MONITORING=0 +# different cache dir for different processes +export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID} +export MASTER_PORT=29500 +export NODE_RANK=$SLURM_PROCID +nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) ) +export MASTER_ADDR=${nodes[0]} +export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID +export TOKENIZERS_PARALLELISM=false +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +echo "MASTER_ADDR: $MASTER_ADDR" +echo "NODE_RANK: $NODE_RANK" + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers" +DATA_DIR=/mnt/weka/home/hao.zhang/wl/sharefs/Vchitect-2M/Wan-Syn_77x768x1280/latents_i2v/train +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS_PER_NODE=8 +TOTAL_GPUS=$((NUM_GPUS_PER_NODE * SLURM_JOB_NUM_NODES)) +# export CUDA_VISIBLE_DEVICES=4,5 + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_qat_16_14B_M2" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 768 + --num_width 1280 + --num_frames 77 + --enable_gradient_checkpointing_type "full" # if OOM enable this + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $TOTAL_GPUS + --sp_size 4 + --tp_size 1 + --hsdp_replicate_dim 4 + --hsdp_shard_dim 8 +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path "$DATA_DIR" + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + # --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 200 + --training_state_checkpointing_steps 200 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 5 + --seed 1000 +) + +srun torchrun \ +--nnodes $SLURM_JOB_NUM_NODES \ +--nproc_per_node $NUM_GPUS_PER_NODE \ +--node_rank $SLURM_PROCID \ +--rdzv_backend=c10d \ +--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_B200.sh b/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_B200.sh new file mode 100644 index 000000000..99a693d0a --- /dev/null +++ b/examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_B200.sh @@ -0,0 +1,116 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_14B_finetune_B200 +#SBATCH --partition=all +#SBATCH --nodes=16 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=1 +#SBATCH --output=logs/wan_t2v_14B_finetune_B200.out +#SBATCH --error=logs/wan_t2v_14B_finetune_B200.err + +source .venv/bin/activate + + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers" +DATA_DIR=data/Wan-Syn_77x768x1280_250k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_14B/validation.json" +NUM_GPUS=64 +# export CUDA_VISIBLE_DEVICES=4,5 + +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# Launch ONE torchrun per node (via srun) and let torchrun spawn 4 workers per node. +MASTER_ADDR="$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)" +MASTER_PORT="${MASTER_PORT:-29500}" +export MASTER_ADDR MASTER_PORT + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_14B" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 20 + --num_height 768 + --num_width 1280 + --num_frames 77 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 1 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + # --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-6 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 1000 + --training_state_checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 5 + --seed 1000 +) + +srun --nodes="$SLURM_NNODES" --ntasks="$SLURM_NNODES" --ntasks-per-node=1 \ + torchrun \ + --nnodes "$SLURM_NNODES" \ + --nproc_per_node 4 \ + --rdzv_backend c10d \ + --rdzv_endpoint "${MASTER_ADDR}:${MASTER_PORT}" \ + --rdzv_id "$SLURM_JOB_ID" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_14B/finetune_t2v_sage3.sh b/examples/training/finetune/wan_t2v_14B/finetune_t2v_sage3.sh new file mode 100644 index 000000000..8849c7246 --- /dev/null +++ b/examples/training/finetune/wan_t2v_14B/finetune_t2v_sage3.sh @@ -0,0 +1,113 @@ +#!/bin/bash +#SBATCH --job-name=wan_t2v_14B_finetune_sage3 +#SBATCH --partition=all +#SBATCH --nodes=2 +#SBATCH --gres=gpu:4 +#SBATCH --ntasks-per-node=4 +#SBATCH --output=logs/wan_t2v_14B_finetune_sage3_%j.out +#SBATCH --error=logs/wan_t2v_14B_finetune_sage3_%j.err + +export WANDB_BASE_URL="https://api.wandb.ai" +export WANDB_MODE=online +export TOKENIZERS_PARALLELISM=false +# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA + +# export TRITON_PRINT_AUTOTUNING=1 # to print the best config +export WANDB_API_KEY=2f25ad37933894dbf0966c838c0b8494987f9f2f +MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers" +DATA_DIR=data/Wan-Syn_77x448x832_600k +VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json" +NUM_GPUS=8 +# export CUDA_VISIBLE_DEVICES=4,5 + +set -euo pipefail + +# ---- torchrun rendezvous (multi-node) ---- +# Launch ONE torchrun per node (via srun) and let torchrun spawn 4 workers per node. +MASTER_ADDR="$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)" +MASTER_PORT="${MASTER_PORT:-29500}" +export MASTER_ADDR MASTER_PORT + + +# Training arguments +training_args=( + --tracker_project_name "wan_t2v_finetune_qat" + --output_dir "checkpoints/wan_t2v_finetune_sage3_14B" + --max_train_steps 4000 + --train_batch_size 1 + --train_sp_batch_size 1 + --gradient_accumulation_steps 1 + --num_latent_t 32 + --num_height 720 + --num_width 1280 + --num_frames 125 + --enable_gradient_checkpointing_type "full" + --generator_4bit_attn True +) + +# Parallel arguments +parallel_args=( + --num_gpus $NUM_GPUS + --sp_size 8 + --tp_size 1 + --hsdp_replicate_dim 1 + --hsdp_shard_dim $NUM_GPUS +) + +# Model arguments +model_args=( + --model_path $MODEL_PATH + --pretrained_model_name_or_path $MODEL_PATH +) + +# Dataset arguments +dataset_args=( + --data_path $DATA_DIR + --dataloader_num_workers 4 +) + +# Validation arguments +validation_args=( + --log_validation + --validation_dataset_file $VALIDATION_DATASET_FILE + --validation_steps 200 + --validation_sampling_steps "50" + --validation_guidance_scale "5.0" +) + +# Optimizer arguments +optimizer_args=( + --learning_rate 1e-5 + --mixed_precision "bf16" + --weight_only_checkpointing_steps 1000 + --training_state_checkpointing_steps 1000 + --weight_decay 0.01 + --max_grad_norm 1.0 +) + +# Miscellaneous arguments +miscellaneous_args=( + --inference_mode False + --checkpoints_total_limit 3 + --training_cfg_rate 0.1 + --dit_precision "fp32" + --ema_start_step 0 + --flow_shift 5 + --seed 1000 +) + +srun --nodes="$SLURM_NNODES" --ntasks="$SLURM_NNODES" --ntasks-per-node=1 \ + torchrun \ + --nnodes "$SLURM_NNODES" \ + --nproc_per_node 4 \ + --rdzv_backend c10d \ + --rdzv_endpoint "${MASTER_ADDR}:${MASTER_PORT}" \ + --node_rank "$SLURM_NODEID" \ + fastvideo/training/wan_training_pipeline.py \ + "${parallel_args[@]}" \ + "${model_args[@]}" \ + "${dataset_args[@]}" \ + "${training_args[@]}" \ + "${optimizer_args[@]}" \ + "${validation_args[@]}" \ + "${miscellaneous_args[@]}" diff --git a/examples/training/finetune/wan_t2v_14B/validation.json b/examples/training/finetune/wan_t2v_14B/validation.json new file mode 100644 index 000000000..1757fb7be --- /dev/null +++ b/examples/training/finetune/wan_t2v_14B/validation.json @@ -0,0 +1,131 @@ +{ + "data": [ + { + "caption": "In the video, a woman is elegantly showcasing her earrings, bringing attention to their intricate design with a gentle touch of her fingers. She is bathed in ambient purple and pink lighting, which casts a soft glow on her delicate features and enhances the vivid tones of her lipstick and eye makeup. Her hair is styled to frame her face smoothly, emphasizing the contours of her jawline and cheekbones. The background features a blurred neon light, adding an artistic and modern touch to the overall aesthetic.", + "video_path": "Fashion/mixkit-face-of-an-elegant-and-captivating-woman-41914_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "In the video, a lone rider guides a majestic horse across an expansive, open field as the sun sets in the background. The rider, dressed in a classic blue shirt and wide-brimmed hat, sits confidently in the saddle, silhouetted against the warm glow of the evening sky. The horse moves gracefully, its mane and tail flowing with each step, creating a sense of harmony between horse and rider. Surrounding the pair, towering trees form a natural border, their leaves gently rustling in the breeze. The shadows lengthen on the ground, accentuating the serene and timeless feel of the scene. The distant hills and wooden fences frame the horizon, adding depth to the tranquil landscape. A few horses graze peacefully in the background, blending into the pastoral setting. The overall ambiance evokes a sense of calmness and quietude, capturing a perfect moment in the golden light of dusk.", + "video_path": "Man/mixkit-a-rancher-riding-a-horse-at-sunset-1143_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "In a dimly lit, eerie setting, a mysterious pink bottle labeled \"Authentic 100% organic POISON\" sits prominently in the foreground, casting a menacing aura. The bottle is accentuated by green fog, which swirls lightly around it, enhancing its sinister allure. Behind it, a shadowy golden bottle adorned with a spider emblem subtly emerges, adding an extra layer of mystery to the scene. Dim candles provide faint, flickering light, which complements the dark atmosphere, making the setting ideal for an illusion of hidden dangers.", + "video_path": "smoke/mixkit-poison-in-halloween-ritual-33879_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "The video opens with a tranquil scene in the heart of a dense forest, emphasizing two large, textured tree trunks in the foreground framing the view. Sunlight filters through the canopy above, casting intricate patterns of light and shadow on the trees and the ground. Between the tree trunks, a clear view of a calm, muddy river unfolds, its surface shimmering under the gentle sunlight. The riverbank is decorated with a variety of small bushes and vibrant foliage, subtly transitioning into the deep greens of tall, leafy plants. In the background, the dense forest looms, filled with dark, towering trees, their branches intertwining to form an intricate canopy. The scene is bathed in the soft glow of the sun, creating a serene and picturesque setting. Occasional sunbeams pierce through the foliage, adding a magical aura to the landscape. The vibrant reds and oranges of the smaller plants add contrast, bringing warmth to the earthy tones of the scenery. Overall, this harmonious blend of natural elements creates a peaceful and idyllic forest setting.", + "video_path": "forest/mixkit-view-of-a-river-between-two-old-trees-560_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "In the video, a martial artist dressed in a traditional white uniform with a black belt demonstrates a series of precise movements against a stark black background. The individual gracefully transitions between stances, embodying a sense of focused discipline and control. Each motion is executed with a deliberate pace, showcasing the fluidity of martial arts techniques. The soft lighting creates subtle highlights on the uniform, adding depth to the figure as it moves. The practitioner begins with an open-hand pose, feet firmly grounded, gradually shifting to a powerful forward punch. The fluidity of the sequence displays a mastery of balance and poise. Every trajectory of the limbs is precise and deliberate, capturing the elegance and strength of martial arts. The serene, isolated setting enhances the intensity and concentration of the practitioner. This visual presentation is an elegant interplay of motion and stillness, displaying the art form's discipline and grace.", + "video_path": "Man/mixkit-a-young-man-practicing-his-karate-moves-49635_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "A tranquil coastal scene unfolds with a drone's aerial view capturing a serene beach landscape. The camera glides over a quiet stretch of sandy shoreline, where gentle waves kiss the shore under a clear blue sky. Nestled amidst lush palm trees are a series of traditional thatched-roof huts, their earthy tones blending harmoniously with the natural surroundings. The sandy beach stretches endlessly, bordered by the rhythmic dance of ocean waves on one side and verdant greenery on the other. A pair of white umbrellas is set up on the sand, suggesting a place to relax and enjoy the sun. In the distance, two small human figures can be seen walking leisurely along the water's edge, leaving faint footprints behind them. The scene exudes a calm and inviting atmosphere, with the soft rustle of palm leaves and the whisper of the ocean breeze almost audible. The overall composition is a captivating blend of nature's tranquility and architectural simplicity. This picturesque setting invites viewers to imagine themselves steps away from this idyllic coastal escape.", + "video_path": "beach/mixkit-sunny-beach-in-a-dynamic-shot-from-a-drone-44383_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "A lone figure stands on a large, moss-covered rock, surrounded by the soft rush of a nearby stream. The figure is wearing white sneakers and shorts, with a plaid shirt that hangs loosely in the breeze. The lighting creates dramatic shadows, enhancing the textures of the rock and the subtle movement of the water below. In the background, a waterfall cascades into the stream, completing this tranquil and serene nature scene.", + "video_path": "forest/mixkit-woman-standing-in-front-of-waterfall-559_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "In an industrial setting, a person leans casually against a railing, exuding a sense of confidence and composure. They are wearing a striking outfit, consisting of a vibrant, patterned jacket over a simple white crop top, creating a bold contrast. The atmosphere is infused with warm, ambient lighting that casts soft shadows on the concrete walls and metallic surfaces. Intricate wiring and pipes form an intricate backdrop, enhancing the urban aesthetic. Their relaxed posture and direct, engaging gaze suggest a sense of ease in this industrial environment. This scene encapsulates a blend of modern fashion and gritty, urban architecture, creating a visually compelling narrative.", + "video_path": "Fashion/mixkit-portrait-of-a-hipster-woman-walking-down-a-stairs-1297_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "A man is energetically stretching in an open-air setting, surrounded by rows of vibrant red seats that suggest an amphitheater or outdoor venue. He wears a sleeveless black shirt layered with a hooded vest, emphasizing his athletic build as he engages in a warm-up routine. Behind him, the striking modern architecture of the building features geometric panels, with large sections of glass and overlapping metallic beams creating a dynamic backdrop. The scene captures the contrast between his focused movements and the static, bold design of the structure, while the surrounding greenery adds a touch of nature to the environment. The overall atmosphere is one of preparation and anticipation, with the man appearing determined and ready for an upcoming event or performance.", + "video_path": "Sport/mixkit-man-doing-arm-stretches-595_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "A young woman is seated on the floor in front of a plush, beige tufted couch, fully engrossed in sorting through a stack of papers. Her dark hair falls loosely past her shoulders, and she wears a green plaid shirt, contributing to the casual yet focused atmosphere. She gently places the papers onto a small round white table, occasionally lifting individual sheets to examine them more closely. Her expression shifts subtly, reflecting concentration and contemplation as she processes the information on the pages. Two small, round nested tables hold her documents, along with a small plant in a gray pot, adding a touch of greenery to the scene. The background features a dark paneled wall, creating a contrasting backdrop for the light-colored furniture. The setting is tranquil and organized, the couch and tables arranged symmetrically, conveying a sense of harmony. A calculator rests on the smaller table, hinting at a task involving calculations or budgeting.", + "video_path": "Woman/mixkit-frustrated-woman-throws-paperwork-on-the-floor-4526_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "A heavily rusted metal gate stands firmly locked, with two vertical bars joined by a thick, old chain that loops elegantly around them. The chain's texture is coarse and rugged, its surface reflecting varying shades of orange and brown, indicative of years exposed to the elements. At the heart of the chain, a black iron padlock, slightly worn yet imposing, secures the gate, its curves and edges smooth against the aged links. The gate's metalwork is outlined by a backdrop of soft, blurred greenery, suggesting a serene and isolated location beyond the barrier. Tall trees rise in the distance, their trunks and leaves creating a lush, forest-like setting that contrasts with the gate's severe rust. A pathway leads away from the gate, its surface uneven with patches of moss and weathered stone visible in the soft focus, inviting yet inaccessible. The ambiance is quiet and mysterious, with a sense of abandonment hanging subtly in the air, evoking curiosity about what lies beyond. Shadows play across the gate, cast by branches swaying gently in the breeze, adding to the dynamic interaction of light and texture. This scene, rich in detail and atmosphere, captures the viewer's imagination, evoking both the allure of the forbidden and the beauty of decay.", + "video_path": "forest/mixkit-rusty-fence-with-a-chain-of-a-property-in-nature-5294_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "In a serene and softly lit yoga studio, three individuals engage in a yoga session, each performing an upward-facing stretch. The central figure is a woman with shoulder-length brown hair, dressed in a light cropped top and green leggings, her posture reflecting grace and concentration. To her right, another participant, a woman in a purple outfit, mirrors the pose with equal poise. On her left, a person with a bun focuses intently, supported slightly by yoga blocks beneath their hands. The warm-colored wooden floor contrasts soothingly with the soft pastel mural on the back wall, featuring an abstract design and partial visage of a serene face. Natural light floods the space from a large window on the right, where lush greens peek through, adding an element of tranquility. In the corner of the room, a collection of meditation instruments, including a gong and a Buddha statue, subtly frame the peaceful setting. The mood is calm yet focused, as all three participants are deeply engaged in their practice. The scene combines elements of balance, harmony, and a shared journey towards mindfulness. This depiction captures the essence of a yoga session that blends personal growth with collective experience.", + "video_path": "People/mixkit-small-group-of-people-doing-yoga-together-43730_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "In the deep blue expanse of the ocean, two dolphins glide effortlessly, their sleek bodies reflecting the sunlight filtering through the water. The prominent shadows and caustics create a shimmering effect on their skin, capturing the beauty of their natural habitat. Each dolphin moves with a fluid grace, occasionally interacting with gentle nudges, showcasing their playful and social nature. The scene is vibrant and dynamic, with the clear blue background accentuating the dolphins' movements, making it an ideal subject for AI recreation.", + "video_path": "sea/mixkit-dolphins-underwater-4133_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "In the video, a young woman stands against a vibrant graffiti-covered wall, deeply engrossed in her smartphone. Her expression reflects a mix of focus and subtle satisfaction as she interacts with the screen. She wears a black floral-patterned top, which contrasts with the bright, abstract shapes and bold colors of the mural behind her. As she continues to engage with her phone, a series of like count notifications appear on the screen, indicating a growing online appreciation. The wall behind her features a striking mix of geometric and organic shapes, including swirls of teal, orange, and black, with large humanoid figures in a pop-art style. Her long, light-brown hair frames her face, adding a calm, composed aura amidst the lively backdrop. The video captures a blend of contemporary digital interaction and expressive urban art, creating a dynamic yet harmonious scene.", + "video_path": "Girl/mixkit-girl-looking-at-the-likes-in-her-post-4914_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "A young mother and her baby sit comfortably on a bed, surrounded by an inviting, cozy atmosphere. The woman, wearing a sleeveless top and jeans, is gently engaging with the baby, who is dressed in an adorable animal-print onesie. The child is seated on the bed with colorful toys scattered around, including a plush toy and a board book. The warm glow from a hanging lamp casts a soft light on them, enhancing the serene environment. Pillows are propped up against the headboard, providing a cushioned backdrop as the mother leans slightly over to interact with the baby. A small bottle is visible beside her, suggesting a nurturing setting. Her hand gestures animatedly as she holds up a soft, white cushion with red and blue accents, likely stimulating the baby\u2019s curiosity. Their shared moment is filled with affection and joy, a perfect snapshot of familial bonding.", + "video_path": "Baby/mixkit-loving-mother-and-her-baby-playing-with-soft-toys-49966_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + }, + { + "caption": "A young girl with long brown hair sits at a round wooden table, engrossed in working on her laptop. The laptop screen is a vivid green, suggesting a green screen effect is in use. To her left, a doll dressed in a yellow and white outfit is casually laid on top of some books, adding a playful and innocent touch to the scene. The setting is cozy, with sheer curtains in the background allowing soft natural light to spill into the room. The girl's posture and focused attention on the laptop suggest she is either playing a game or learning something new. This serene and domestic atmosphere is complemented by the slight blur of a dark couch in the foreground, framing the focused activity of the child.", + "video_path": "Girl/mixkit-little-girl-doing-homework-on-a-laptop-4757_clip_1.mp4", + "num_inference_steps": 50, + "height": 768, + "width": 1280, + "num_frames": 77 + } ] +} diff --git a/fastvideo/attention/backends/benchmark_scripts/benchmark_blockscaled_fp4_attn.py b/fastvideo/attention/backends/benchmark_scripts/benchmark_blockscaled_fp4_attn.py new file mode 100644 index 000000000..2f49d1f26 --- /dev/null +++ b/fastvideo/attention/backends/benchmark_scripts/benchmark_blockscaled_fp4_attn.py @@ -0,0 +1,284 @@ +import torch +import sys +import os +import traceback + +from fastvideo.attention.backends.sageattn.api import ( + blockscaled_fp4_attn, + preprocess_qkv, + scale_and_quant_fp4, + scale_and_quant_fp4_permute, + scale_and_quant_fp4_transpose, +) +from fastvideo.attention.backends.sageattn.quantization.bench.bench_utils import bench + + +def calculate_attention_flops(batch_size, num_heads, seq_len_q, seq_len_k, head_dim, is_causal=False): + """Calculate FLOPs for attention (FlashAttention standard - matmuls only).""" + f = 4 * batch_size * num_heads * seq_len_q * seq_len_k * head_dim + if is_causal: + f = f // 2 + return f + + +def benchmark_blockscaled_fp4_attn(batch_size, num_heads, seq_len, head_dim, + is_causal=False, dtype=torch.bfloat16, + per_block_mean=True, single_level_p_quant=False, + num_warmups=100, num_tests=1000): + """ + Benchmark blockscaled_fp4_attn function (excluding quantization overhead). + + This benchmarks ONLY the core FP4 attention kernel, with pre-quantized inputs. + The quantization step is performed once before benchmarking. + + Args: + batch_size: Batch size + num_heads: Number of attention heads + seq_len: Sequence length (same for Q, K, V) + head_dim: Head dimension + is_causal: Whether to use causal masking + dtype: Data type (torch.bfloat16 or torch.float16) + per_block_mean: Whether to use per-block mean for Q smoothing + single_level_p_quant: If True, use single-level quantization for P matrix + num_warmups: Number of warmup iterations + num_tests: Number of test iterations + + Returns: + dict with performance metrics + """ + device = 'cuda' + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is not available. This benchmark requires a CUDA device.") + + # Create input tensors + q = torch.randn(batch_size, num_heads, seq_len, head_dim, + device=device, dtype=dtype) + k = torch.randn(batch_size, num_heads, seq_len, head_dim, + device=device, dtype=dtype) + v = torch.randn(batch_size, num_heads, seq_len, head_dim, + device=device, dtype=dtype) + + # Pre-process and quantize inputs (done once, not included in benchmark) + is_bf16 = dtype == torch.bfloat16 + KL = k.size(2) + q_processed, k_processed, v_processed, delta_s = preprocess_qkv(q, k, v, per_block_mean) + qlist = scale_and_quant_fp4(q_processed) + klist = scale_and_quant_fp4_permute(k_processed) + vlist = scale_and_quant_fp4_transpose(v_processed) + + # Synchronize to ensure quantization is complete + torch.cuda.synchronize() + + # Create closure for benchmarking (only the attention kernel) + def run_attention(): + return blockscaled_fp4_attn( + qlist, klist, vlist, + delta_s, + KL, + is_causal=is_causal, + per_block_mean=per_block_mean, + is_bf16=is_bf16, + single_level_p_quant=single_level_p_quant + ) + + # Benchmark using the bench utility (handles warmup and timing) + avg_time_ms = bench(run_attention, num_warmups=num_warmups, num_tests=num_tests) + avg_time_s = avg_time_ms / 1000.0 + + # Calculate FLOPs (use processed sequence length after padding) + processed_seq_len = q_processed.size(2) + total_flops = calculate_attention_flops( + batch_size, num_heads, processed_seq_len, processed_seq_len, head_dim, is_causal + ) + + # Calculate TFLOPs + tflops = total_flops / (avg_time_s * 1e12) + + # Calculate throughput (tokens/sec) - use original seq_len for meaningful metric + tokens_per_second = (batch_size * seq_len) / avg_time_s + + return { + 'batch_size': batch_size, + 'num_heads': num_heads, + 'seq_len': seq_len, + 'processed_seq_len': processed_seq_len, + 'head_dim': head_dim, + 'is_causal': is_causal, + 'dtype': str(dtype), + 'per_block_mean': per_block_mean, + 'single_level_p_quant': single_level_p_quant, + 'avg_time_ms': avg_time_ms, + 'avg_time_s': avg_time_s, + 'total_flops': total_flops, + 'tflops': tflops, + 'tokens_per_second': tokens_per_second, + } + + +def print_results(results): + """Print benchmark results in a formatted table.""" + print("\n" + "="*100) + print("blockscaled_fp4_attn Benchmark Results (Kernel Only, No Quantization)") + print("="*100) + print(f"Configuration:") + print(f" Batch Size: {results['batch_size']}") + print(f" Num Heads: {results['num_heads']}") + print(f" Sequence Length: {results['seq_len']}") + print(f" Processed Seq Length: {results['processed_seq_len']} (after padding)") + print(f" Head Dimension: {results['head_dim']}") + print(f" Causal: {results['is_causal']}") + print(f" Data Type: {results['dtype']}") + print(f" Per Block Mean: {results['per_block_mean']}") + print(f" Single Level P Quant: {results['single_level_p_quant']}") + print(f"\nPerformance:") + print(f" Average Time: {results['avg_time_ms']:.3f} ms") + print(f" Total FLOPs: {results['total_flops']/1e12:.4f} TFLOPs (theoretical)") + print(f" Throughput: {results['tflops']:.4f} TFLOPs/s") + print(f" Tokens/sec: {results['tokens_per_second']:,.0f}") + print("="*100 + "\n") + sys.stdout.flush() + + +def run_benchmark_suite(): + """Run a comprehensive benchmark suite with various configurations.""" + print("Starting blockscaled_fp4_attn Benchmark Suite (Kernel Only)...") + print(f"CUDA Device: {torch.cuda.get_device_name(0)}") + print(f"CUDA Version: {torch.version.cuda}") + print(f"PyTorch Version: {torch.__version__}\n") + print("Note: This benchmark measures only the FP4 attention kernel,") + print(" excluding quantization overhead.\n") + sys.stdout.flush() + + # Default configurations to test + # (batch_size, num_heads, seq_len, head_dim, is_causal, dtype) + configs = [ + (1, 16, 512, 64, False, torch.bfloat16), + (1, 16, 1024, 64, False, torch.bfloat16), + (1, 16, 2048, 64, False, torch.bfloat16), + (1, 16, 4096, 64, False, torch.bfloat16), + (1, 16, 8192, 64, False, torch.bfloat16), + (1, 16, 16384, 64, False, torch.bfloat16), + + (1, 16, 512, 128, False, torch.bfloat16), + (1, 16, 1024, 128, False, torch.bfloat16), + (1, 16, 2048, 128, False, torch.bfloat16), + (1, 16, 4096, 128, False, torch.bfloat16), + (1, 16, 8192, 128, False, torch.bfloat16), + (1, 16, 16384, 128, False, torch.bfloat16), + + (1, 32, 512, 64, False, torch.bfloat16), + (1, 32, 1024, 64, False, torch.bfloat16), + (1, 32, 2048, 64, False, torch.bfloat16), + (1, 32, 4096, 64, False, torch.bfloat16), + (1, 32, 8192, 64, False, torch.bfloat16), + (1, 32, 16384, 64, False, torch.bfloat16), + + (1, 32, 512, 128, False, torch.bfloat16), + (1, 32, 1024, 128, False, torch.bfloat16), + (1, 32, 2048, 128, False, torch.bfloat16), + (1, 32, 4096, 128, False, torch.bfloat16), + (1, 32, 8192, 128, False, torch.bfloat16), + (1, 32, 16384, 128, False, torch.bfloat16), + ] + + all_results = [] + + for config in configs: + batch_size, num_heads, seq_len, head_dim, is_causal, dtype = config + + print(f"\nBenchmarking: B={batch_size}, H={num_heads}, L={seq_len}, D={head_dim}, " + f"Causal={is_causal}, dtype={dtype}...") + sys.stdout.flush() + + try: + results = benchmark_blockscaled_fp4_attn( + batch_size=batch_size, + num_heads=num_heads, + seq_len=seq_len, + head_dim=head_dim, + is_causal=is_causal, + dtype=dtype, + num_warmups=10, + num_tests=50 + ) + + print_results(results) + all_results.append(results) + + except Exception as e: + print(f"Error benchmarking configuration {config}:") + print(f" Exception: {e}") + traceback.print_exc() + sys.stdout.flush() + continue + + # Print summary table + print("\n" + "="*120) + print("Summary Table (blockscaled_fp4_attn Kernel Only)") + print("="*120) + print(f"{'B':<4} {'H':<4} {'L':<6} {'D':<4} {'Causal':<7} {'Time (ms)':<12} {'TFLOPs/s':<12} {'Tokens/s':<15}") + print("-"*120) + + for r in all_results: + print(f"{r['batch_size']:<4} {r['num_heads']:<4} {r['seq_len']:<6} {r['head_dim']:<4} " + f"{str(r['is_causal']):<7} {r['avg_time_ms']:<12.3f} {r['tflops']:<12.4f} " + f"{r['tokens_per_second']:<15,.0f}") + + print("="*120 + "\n") + sys.stdout.flush() + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser(description='Benchmark blockscaled_fp4_attn kernel (excluding quantization)') + parser.add_argument('--batch-size', type=int, default=None, help='Batch size') + parser.add_argument('--num-heads', type=int, default=None, help='Number of attention heads') + parser.add_argument('--seq-len', type=int, default=None, help='Sequence length') + parser.add_argument('--head-dim', type=int, default=None, help='Head dimension') + parser.add_argument('--causal', action='store_true', help='Use causal attention') + parser.add_argument('--dtype', type=str, default='bfloat16', choices=['bfloat16', 'float16'], + help='Data type') + parser.add_argument('--per-block-mean', action='store_true', default=True, + help='Use per-block mean for Q smoothing (default: True)') + parser.add_argument('--no-per-block-mean', action='store_false', dest='per_block_mean', + help='Disable per-block mean for Q smoothing') + parser.add_argument('--single-level-p-quant', action='store_true', default=False, + help='Use single-level P quantization (default: False)') + parser.add_argument('--two-level-p-quant', action='store_false', dest='single_level_p_quant', + help='Use two-level P quantization') + parser.add_argument('--num-warmups', type=int, default=10, help='Number of warmup iterations') + parser.add_argument('--num-tests', type=int, default=50, help='Number of test iterations') + parser.add_argument('--suite', action='store_true', help='Run full benchmark suite') + + args = parser.parse_args() + + dtype_map = { + 'bfloat16': torch.bfloat16, + 'float16': torch.float16 + } + + if args.suite: + run_benchmark_suite() + elif args.batch_size and args.num_heads and args.seq_len and args.head_dim: + results = benchmark_blockscaled_fp4_attn( + batch_size=args.batch_size, + num_heads=args.num_heads, + seq_len=args.seq_len, + head_dim=args.head_dim, + is_causal=args.causal, + dtype=dtype_map[args.dtype], + per_block_mean=args.per_block_mean, + single_level_p_quant=args.single_level_p_quant, + num_warmups=args.num_warmups, + num_tests=args.num_tests + ) + print_results(results) + else: + print("Running default benchmark suite. Use --suite for full suite or provide all parameters.") + print("Example: python benchmark_blockscaled_fp4_attn.py --batch-size 1 --num-heads 16 --seq-len 4096 --head-dim 128") + print("\nNote: This benchmark measures only the FP4 attention kernel,") + print(" excluding quantization overhead.") + sys.stdout.flush() + run_benchmark_suite() + diff --git a/fastvideo/attention/backends/benchmark_scripts/benchmark_combined.py b/fastvideo/attention/backends/benchmark_scripts/benchmark_combined.py new file mode 100644 index 000000000..c0c60d743 --- /dev/null +++ b/fastvideo/attention/backends/benchmark_scripts/benchmark_combined.py @@ -0,0 +1,380 @@ +import torch +import sys +import traceback +import argparse +import matplotlib.pyplot as plt +import matplotlib +matplotlib.use('Agg') # Use non-interactive backend for server environments +import numpy as np +from typing import Dict, List, Tuple, Optional + +from flash_attn import flash_attn_func +from fastvideo.attention.backends.sageattn.quantization.bench.bench_utils import bench + +# Import SageAttn components for direct control +from fastvideo.attention.backends.sageattn.api import ( + preprocess_qkv, + scale_and_quant_fp4, + scale_and_quant_fp4_permute, + scale_and_quant_fp4_transpose, + blockscaled_fp4_attn +) + + +def calculate_attention_flops(batch_size, num_heads, seq_len_q, seq_len_k, head_dim, is_causal=False): + """Calculate FLOPs for attention (FlashAttention standard - matmuls only).""" + f = 4 * batch_size * num_heads * seq_len_q * seq_len_k * head_dim + if is_causal: + f = f // 2 + return f + + +def sageattn_blackwell_configurable(q, k, v, is_causal=False, per_block_mean=True, + single_level_p_quant=True, + enable_smoothing_q=False, enable_smoothing_k=False): + """ + Configurable SageAttention3 Blackwell kernel with explicit smoothing control. + + Args: + q: Query tensor [B, H, L, D] + k: Key tensor [B, H, L, D] + v: Value tensor [B, H, L, D] + is_causal: Whether to use causal masking + per_block_mean: Whether to use per-block mean for Q smoothing + single_level_p_quant: If True, use single-level quantization for P matrix + enable_smoothing_q: Enable Q smoothing + enable_smoothing_k: Enable K smoothing + + Returns: + Output tensor [B, H, L, D] + """ + QL = q.size(2) + KL = k.size(2) + is_bf16 = q.dtype == torch.bfloat16 + + # Preprocess with explicit smoothing control + q, k, v, delta_s = preprocess_qkv(q, k, v, per_block_mean, enable_smoothing_q, enable_smoothing_k) + + qlist_from_cuda = scale_and_quant_fp4(q) + klist_from_cuda = scale_and_quant_fp4_permute(k) + vlist_from_cuda = scale_and_quant_fp4_transpose(v) + + o_fp4 = blockscaled_fp4_attn( + qlist_from_cuda, + klist_from_cuda, + vlist_from_cuda, + delta_s, + KL, + is_causal, + per_block_mean, + is_bf16, + single_level_p_quant + )[0][:, :, :QL, :].contiguous() + + return o_fp4 + + +def benchmark_flashattn2(batch_size, num_heads, seq_len, head_dim, + is_causal=False, dtype=torch.bfloat16, + num_warmups=10, num_tests=50): + """Benchmark FlashAttention2.""" + device = 'cuda' + + # FlashAttention2 expects (batch, seq_len, num_heads, head_dim) + q = torch.randn(batch_size, seq_len, num_heads, head_dim, device=device, dtype=dtype) + k = torch.randn(batch_size, seq_len, num_heads, head_dim, device=device, dtype=dtype) + v = torch.randn(batch_size, seq_len, num_heads, head_dim, device=device, dtype=dtype) + + def run_attention(): + return flash_attn_func(q, k, v, causal=is_causal) + + avg_time_ms = bench(run_attention, num_warmups=num_warmups, num_tests=num_tests) + return avg_time_ms + + +def benchmark_sageattn3(batch_size, num_heads, seq_len, head_dim, + is_causal=False, dtype=torch.bfloat16, + per_block_mean=True, single_level_p_quant=False, + enable_smoothing_q=True, enable_smoothing_k=True, + num_warmups=10, num_tests=50): + """Benchmark SageAttention3 with configurable smoothing.""" + device = 'cuda' + + # SageAttn expects (batch, num_heads, seq_len, head_dim) + q = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + k = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + v = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + + def run_attention(): + return sageattn_blackwell_configurable( + q, k, v, + is_causal=is_causal, + per_block_mean=per_block_mean, + single_level_p_quant=single_level_p_quant, + enable_smoothing_q=enable_smoothing_q, + enable_smoothing_k=enable_smoothing_k + ) + + avg_time_ms = bench(run_attention, num_warmups=num_warmups, num_tests=num_tests) + return avg_time_ms + + +def time_to_tflops(time_ms, batch_size, num_heads, seq_len, head_dim, is_causal=False): + """Convert time to TFLOPs (Tera FLOPs per Second).""" + total_flops = calculate_attention_flops(batch_size, num_heads, seq_len, seq_len, head_dim, is_causal) + time_s = time_ms / 1000.0 + tflops = total_flops / (time_s * 1e12) + return tflops + + +def run_benchmark_suite(head_dim=64, is_causal=False, num_heads=12, batch_size=1, + num_warmups=10, num_tests=50, + seq_lens=None, output_file="benchmark_attention.png"): + """ + Run comprehensive benchmark suite and generate plot. + + Args: + head_dim: Head dimension (64 or 128) + is_causal: Whether to use causal attention + num_heads: Number of attention heads + batch_size: Batch size + num_warmups: Number of warmup iterations + num_tests: Number of test iterations + seq_lens: List of sequence lengths to test + output_file: Output plot filename + """ + if seq_lens is None: + seq_lens = [1024, 2048, 4096, 8192, 16384, 32768] + + device_name = torch.cuda.get_device_name(0) + # Extract short name (e.g., "RTX5090" from full name) + short_name = device_name.split()[-1] if 'RTX' in device_name or 'A100' in device_name else device_name[:20] + + print(f"Starting Combined Attention Benchmark Suite...") + print(f"CUDA Device: {device_name}") + print(f"CUDA Version: {torch.version.cuda}") + print(f"PyTorch Version: {torch.__version__}") + print(f"Head Dim: {head_dim}, Causal: {is_causal}, Num Heads: {num_heads}, Batch Size: {batch_size}") + print("="*80) + sys.stdout.flush() + + # Results storage: {method_name: {seq_len: tflops}} + results: Dict[str, Dict[int, Optional[float]]] = { + 'FlashAttn': {}, + 'SageAttn3': {}, + 'FP4': {}, + } + + dtype = torch.bfloat16 + + for seq_len in seq_lens: + print(f"\n--- Sequence Length: {seq_len} ---") + sys.stdout.flush() + + # FlashAttention2 + print(f" Benchmarking FlashAttn2...", end=" ") + sys.stdout.flush() + try: + time_ms = benchmark_flashattn2( + batch_size, num_heads, seq_len, head_dim, + is_causal=is_causal, dtype=dtype, + num_warmups=num_warmups, num_tests=num_tests + ) + tflops = time_to_tflops(time_ms, batch_size, num_heads, seq_len, head_dim, is_causal) + results['FlashAttn'][seq_len] = tflops + print(f"{tflops:.0f} TFLOPs ({time_ms:.3f} ms)") + except Exception as e: + print(f"OOM or Error: {e}") + results['FlashAttn'][seq_len] = None + sys.stdout.flush() + + # SageAttn3 (with smoothing: single_level_p_quant=False, enable_smoothing_q=True, enable_smoothing_k=True) + print(f" Benchmarking SageAttn3 (smoothing ON)...", end=" ") + sys.stdout.flush() + try: + time_ms = benchmark_sageattn3( + batch_size, num_heads, seq_len, head_dim, + is_causal=is_causal, dtype=dtype, + per_block_mean=True, + single_level_p_quant=False, + enable_smoothing_q=True, + enable_smoothing_k=True, + num_warmups=num_warmups, num_tests=num_tests + ) + tflops = time_to_tflops(time_ms, batch_size, num_heads, seq_len, head_dim, is_causal) + results['SageAttn3'][seq_len] = tflops + print(f"{tflops:.0f} TFLOPs ({time_ms:.3f} ms)") + except Exception as e: + print(f"OOM or Error: {e}") + results['SageAttn3'][seq_len] = None + sys.stdout.flush() + + # FP4 (no smoothing: single_level_p_quant=True, enable_smoothing_q=False, enable_smoothing_k=False) + print(f" Benchmarking FP4 (smoothing OFF)...", end=" ") + sys.stdout.flush() + try: + time_ms = benchmark_sageattn3( + batch_size, num_heads, seq_len, head_dim, + is_causal=is_causal, dtype=dtype, + per_block_mean=True, + single_level_p_quant=True, + enable_smoothing_q=False, + enable_smoothing_k=False, + num_warmups=num_warmups, num_tests=num_tests + ) + tflops = time_to_tflops(time_ms, batch_size, num_heads, seq_len, head_dim, is_causal) + results['FP4'][seq_len] = tflops + print(f"{tflops:.0f} TFLOPs ({time_ms:.3f} ms)") + except Exception as e: + print(f"OOM or Error: {e}") + results['FP4'][seq_len] = None + sys.stdout.flush() + + # Print summary table + print("\n" + "="*100) + print("Summary Table (TFLOPs)") + print("="*100) + header = f"{'SeqLen':<10}" + for method in results.keys(): + header += f"{method:<15}" + print(header) + print("-"*100) + + for seq_len in seq_lens: + row = f"{seq_len:<10}" + for method in results.keys(): + val = results[method].get(seq_len) + if val is not None: + row += f"{val:<15.0f}" + else: + row += f"{'OOM':<15}" + print(row) + print("="*100) + sys.stdout.flush() + + # Generate plot + generate_plot(results, seq_lens, head_dim, is_causal, short_name, output_file) + + return results + + +def generate_plot(results: Dict[str, Dict[int, Optional[float]]], + seq_lens: List[int], + head_dim: int, + is_causal: bool, + device_name: str, + output_file: str): + """Generate bar plot comparing attention implementations.""" + + # Prepare data + methods = list(results.keys()) + x_labels = [f"{sl//1024}K" for sl in seq_lens] + + # Colors for each method (red, blue, green scheme) + colors = { + 'FlashAttn': '#1E90FF', # Blue (Dodger Blue) + 'SageAttn3': '#228B22', # Green (Forest Green) + 'FP4': '#DC143C', # Red (Crimson) + } + + # Number of methods and positions + n_methods = len(methods) + n_positions = len(seq_lens) + + # Bar width and positions + bar_width = 0.25 + x = np.arange(n_positions) + + # Create figure + fig, ax = plt.subplots(figsize=(12, 6)) + + # Plot bars for each method + for i, method in enumerate(methods): + values = [] + for seq_len in seq_lens: + val = results[method].get(seq_len) + values.append(val if val is not None else 0) + + offset = (i - n_methods/2 + 0.5) * bar_width + bars = ax.bar(x + offset, values, bar_width, + label=method, color=colors.get(method, f'C{i}'), + edgecolor='black', linewidth=0.5) + + # Add value labels on top of bars + for bar, val, seq_len in zip(bars, values, seq_lens): + if results[method].get(seq_len) is None: + label = 'OOM' + else: + label = f'{int(val)}' + + height = bar.get_height() + ax.annotate(label, + xy=(bar.get_x() + bar.get_width() / 2, height), + xytext=(0, 3), # 3 points vertical offset + textcoords="offset points", + ha='center', va='bottom', + fontsize=8, rotation=0) + + # Customize plot + ax.set_xlabel('Sequence Length', fontsize=12, fontweight='bold') + ax.set_ylabel('Speed (TFLOPs)', fontsize=12, fontweight='bold') + ax.set_title(f'{device_name}, (Head dim = {head_dim}, causal = {is_causal})', fontsize=14, fontweight='bold') + ax.set_xticks(x) + ax.set_xticklabels(x_labels) + legend = ax.legend(loc='upper left', ncol=len(methods), fontsize=10) + # Make legend text bold + for text in legend.get_texts(): + text.set_fontweight('bold') + + # Set y-axis to start from 0 + ax.set_ylim(bottom=0) + + # Add grid for readability + ax.yaxis.grid(True, linestyle='--', alpha=0.7) + ax.set_axisbelow(True) + + # Tight layout + plt.tight_layout() + + # Save plot + plt.savefig(output_file, dpi=150, bbox_inches='tight') + print(f"\nPlot saved to: {output_file}") + + # Also save as PDF for high quality + pdf_file = output_file.rsplit('.', 1)[0] + '.pdf' + plt.savefig(pdf_file, bbox_inches='tight') + print(f"PDF saved to: {pdf_file}") + + plt.close() + + +if __name__ == "__main__": + parser = argparse.ArgumentParser(description='Combined Attention Benchmark (FlashAttn2 vs SageAttn3 vs FP4)') + parser.add_argument('--batch-size', type=int, default=1, help='Batch size') + parser.add_argument('--num-heads', type=int, default=16, help='Number of attention heads') + parser.add_argument('--head-dim', type=int, default=64, choices=[64, 128], help='Head dimension') + parser.add_argument('--causal', action='store_true', help='Use causal attention') + parser.add_argument('--num-warmups', type=int, default=10, help='Number of warmup iterations') + parser.add_argument('--num-tests', type=int, default=50, help='Number of test iterations') + parser.add_argument('--seq-lens', type=int, nargs='+', + default=[1024, 2048, 4096, 8192, 16384, 32768], + help='Sequence lengths to benchmark') + parser.add_argument('--output', type=str, default='benchmark_attention.png', + help='Output plot filename') + + args = parser.parse_args() + + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is not available. This benchmark requires a CUDA device.") + + run_benchmark_suite( + head_dim=args.head_dim, + is_causal=args.causal, + num_heads=args.num_heads, + batch_size=args.batch_size, + num_warmups=args.num_warmups, + num_tests=args.num_tests, + seq_lens=args.seq_lens, + output_file=args.output + ) + diff --git a/fastvideo/attention/backends/benchmark_scripts/benchmark_flashattn2.py b/fastvideo/attention/backends/benchmark_scripts/benchmark_flashattn2.py new file mode 100644 index 000000000..a3a42e06e --- /dev/null +++ b/fastvideo/attention/backends/benchmark_scripts/benchmark_flashattn2.py @@ -0,0 +1,231 @@ +import torch +import sys +import os +import traceback + +from flash_attn import flash_attn_func +from fastvideo.attention.backends.sageattn.quantization.bench.bench_utils import bench + + +def calculate_attention_flops(batch_size, num_heads, seq_len_q, seq_len_k, head_dim, is_causal=False): + """Calculate FLOPs for attention (FlashAttention standard - matmuls only).""" + f = 4 * batch_size * num_heads * seq_len_q * seq_len_k * head_dim + if is_causal: + f = f // 2 + return f + + +def benchmark_flashattn2(batch_size, num_heads, seq_len, head_dim, + is_causal=False, dtype=torch.bfloat16, + num_warmups=100, num_tests=1000): + """ + Benchmark FlashAttention2 and return performance metrics. + + Args: + batch_size: Batch size + num_heads: Number of attention heads + seq_len: Sequence length (same for Q, K, V) + head_dim: Head dimension + is_causal: Whether to use causal masking + dtype: Data type (torch.bfloat16 or torch.float16) + num_warmups: Number of warmup iterations + num_tests: Number of test iterations + + Returns: + dict with performance metrics + """ + device = 'cuda' + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is not available. This benchmark requires a CUDA device.") + + # Create input tensors - FlashAttention2 expects (batch, seq_len, num_heads, head_dim) + q = torch.randn(batch_size, seq_len, num_heads, head_dim, + device=device, dtype=dtype) + k = torch.randn(batch_size, seq_len, num_heads, head_dim, + device=device, dtype=dtype) + v = torch.randn(batch_size, seq_len, num_heads, head_dim, + device=device, dtype=dtype) + + # Create closure for benchmarking + def run_attention(): + return flash_attn_func( + q, k, v, + causal=is_causal, + ) + + # Benchmark using the bench utility (handles warmup and timing) + avg_time_ms = bench(run_attention, num_warmups=num_warmups, num_tests=num_tests) + avg_time_s = avg_time_ms / 1000.0 + + # Calculate FLOPs + total_flops = calculate_attention_flops( + batch_size, num_heads, seq_len, seq_len, head_dim, is_causal + ) + + # Calculate TFLOPs + tflops = total_flops / (avg_time_s * 1e12) + + # Calculate throughput (tokens/sec) + tokens_per_second = (batch_size * seq_len) / avg_time_s + + return { + 'batch_size': batch_size, + 'num_heads': num_heads, + 'seq_len': seq_len, + 'head_dim': head_dim, + 'is_causal': is_causal, + 'dtype': str(dtype), + 'avg_time_ms': avg_time_ms, + 'avg_time_s': avg_time_s, + 'total_flops': total_flops, + 'tflops': tflops, + 'tokens_per_second': tokens_per_second, + } + + +def print_results(results): + """Print benchmark results in a formatted table.""" + print("\n" + "="*100) + print("FlashAttention2 Benchmark Results") + print("="*100) + print(f"Configuration:") + print(f" Batch Size: {results['batch_size']}") + print(f" Num Heads: {results['num_heads']}") + print(f" Sequence Length: {results['seq_len']}") + print(f" Head Dimension: {results['head_dim']}") + print(f" Causal: {results['is_causal']}") + print(f" Data Type: {results['dtype']}") + print(f"\nPerformance:") + print(f" Average Time: {results['avg_time_ms']:.3f} ms") + print(f" Total FLOPs: {results['total_flops']/1e12:.4f} TFLOPs (theoretical)") + print(f" Throughput: {results['tflops']:.4f} TFLOPs/s") + print(f" Tokens/sec: {results['tokens_per_second']:,.0f}") + print("="*100 + "\n") + sys.stdout.flush() + + +def run_benchmark_suite(): + """Run a comprehensive benchmark suite with various configurations.""" + print("Starting FlashAttention2 Benchmark Suite...") + print(f"CUDA Device: {torch.cuda.get_device_name(0)}") + print(f"CUDA Version: {torch.version.cuda}") + print(f"PyTorch Version: {torch.__version__}\n") + sys.stdout.flush() + + # Default configurations to test + # (batch_size, num_heads, seq_len, head_dim, is_causal, dtype) + configs = [ + (1, 16, 1024, 64, False, torch.bfloat16), + (1, 16, 2048, 64, False, torch.bfloat16), + (1, 16, 4096, 64, False, torch.bfloat16), + (1, 16, 8192, 64, False, torch.bfloat16), + (1, 16, 16384, 64, False, torch.bfloat16), + + (1, 16, 1024, 128, False, torch.bfloat16), + (1, 16, 2048, 128, False, torch.bfloat16), + (1, 16, 4096, 128, False, torch.bfloat16), + (1, 16, 8192, 128, False, torch.bfloat16), + (1, 16, 16384, 128, False, torch.bfloat16), + + (1, 32, 1024, 64, False, torch.bfloat16), + (1, 32, 2048, 64, False, torch.bfloat16), + (1, 32, 4096, 64, False, torch.bfloat16), + (1, 32, 8192, 64, False, torch.bfloat16), + (1, 32, 16384, 64, False, torch.bfloat16), + + (1, 32, 1024, 128, False, torch.bfloat16), + (1, 32, 2048, 128, False, torch.bfloat16), + (1, 32, 4096, 128, False, torch.bfloat16), + (1, 32, 8192, 128, False, torch.bfloat16), + (1, 32, 16384, 128, False, torch.bfloat16), + ] + + all_results = [] + + for config in configs: + batch_size, num_heads, seq_len, head_dim, is_causal, dtype = config + + print(f"\nBenchmarking: B={batch_size}, H={num_heads}, L={seq_len}, D={head_dim}, " + f"Causal={is_causal}, dtype={dtype}...") + sys.stdout.flush() + + try: + results = benchmark_flashattn2( + batch_size=batch_size, + num_heads=num_heads, + seq_len=seq_len, + head_dim=head_dim, + is_causal=is_causal, + dtype=dtype, + num_warmups=10, + num_tests=50 + ) + + print_results(results) + all_results.append(results) + + except Exception as e: + print(f"Error benchmarking configuration {config}:") + print(f" Exception: {e}") + traceback.print_exc() + sys.stdout.flush() + continue + + # Print summary table + print("\n" + "="*120) + print("Summary Table") + print("="*120) + print(f"{'B':<4} {'H':<4} {'L':<6} {'D':<4} {'Causal':<7} {'Time (ms)':<12} {'TFLOPs/s':<12} {'Tokens/s':<15}") + print("-"*120) + + for r in all_results: + print(f"{r['batch_size']:<4} {r['num_heads']:<4} {r['seq_len']:<6} {r['head_dim']:<4} " + f"{str(r['is_causal']):<7} {r['avg_time_ms']:<12.3f} {r['tflops']:<12.4f} " + f"{r['tokens_per_second']:<15,.0f}") + + print("="*120 + "\n") + sys.stdout.flush() + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser(description='Benchmark FlashAttention2 in TFLOPs') + parser.add_argument('--batch-size', type=int, default=None, help='Batch size') + parser.add_argument('--num-heads', type=int, default=None, help='Number of attention heads') + parser.add_argument('--seq-len', type=int, default=None, help='Sequence length') + parser.add_argument('--head-dim', type=int, default=None, help='Head dimension') + parser.add_argument('--causal', action='store_true', help='Use causal attention') + parser.add_argument('--dtype', type=str, default='bfloat16', choices=['bfloat16', 'float16'], + help='Data type') + parser.add_argument('--num-warmups', type=int, default=10, help='Number of warmup iterations') + parser.add_argument('--num-tests', type=int, default=50, help='Number of test iterations') + parser.add_argument('--suite', action='store_true', help='Run full benchmark suite') + + args = parser.parse_args() + + dtype_map = { + 'bfloat16': torch.bfloat16, + 'float16': torch.float16 + } + + if args.suite: + run_benchmark_suite() + elif args.batch_size and args.num_heads and args.seq_len and args.head_dim: + results = benchmark_flashattn2( + batch_size=args.batch_size, + num_heads=args.num_heads, + seq_len=args.seq_len, + head_dim=args.head_dim, + is_causal=args.causal, + dtype=dtype_map[args.dtype], + num_warmups=args.num_warmups, + num_tests=args.num_tests + ) + print_results(results) + else: + print("Running default benchmark suite. Use --suite for full suite or provide all parameters.") + print("Example: python benchmark_flashattn2.py --batch-size 1 --num-heads 16 --seq-len 4096 --head-dim 128") + sys.stdout.flush() + run_benchmark_suite() + diff --git a/fastvideo/attention/backends/benchmark_scripts/benchmark_sageattn3.py b/fastvideo/attention/backends/benchmark_scripts/benchmark_sageattn3.py new file mode 100644 index 000000000..776c6e695 --- /dev/null +++ b/fastvideo/attention/backends/benchmark_scripts/benchmark_sageattn3.py @@ -0,0 +1,250 @@ +import torch +import sys +import os +import traceback + +from fastvideo.attention.backends.sageattn.api import sageattn_blackwell +from fastvideo.attention.backends.sageattn.quantization.bench.bench_utils import bench + + +def calculate_attention_flops(batch_size, num_heads, seq_len_q, seq_len_k, head_dim, is_causal=False): + """Calculate FLOPs for attention (FlashAttention standard - matmuls only).""" + f = 4 * batch_size * num_heads * seq_len_q * seq_len_k * head_dim + if is_causal: + f = f // 2 + return f + + +def benchmark_sageattn3(batch_size, num_heads, seq_len, head_dim, + is_causal=False, dtype=torch.bfloat16, + per_block_mean=True, single_level_p_quant=False, + num_warmups=100, num_tests=1000): + """ + Benchmark SageAttention3 and return performance metrics. + + Args: + batch_size: Batch size + num_heads: Number of attention heads + seq_len: Sequence length (same for Q, K, V) + head_dim: Head dimension + is_causal: Whether to use causal masking + dtype: Data type (torch.bfloat16 or torch.float16) + per_block_mean: Whether to use per-block mean for Q smoothing + single_level_p_quant: If True, use single-level quantization for P matrix + num_warmups: Number of warmup iterations + num_tests: Number of test iterations + + Returns: + dict with performance metrics + """ + device = 'cuda' + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is not available. This benchmark requires a CUDA device.") + + # Create input tensors + q = torch.randn(batch_size, num_heads, seq_len, head_dim, + device=device, dtype=dtype) + k = torch.randn(batch_size, num_heads, seq_len, head_dim, + device=device, dtype=dtype) + v = torch.randn(batch_size, num_heads, seq_len, head_dim, + device=device, dtype=dtype) + + # Create closure for benchmarking (no extra stream needed - bench handles synchronization) + def run_attention(): + return sageattn_blackwell( + q, k, v, + is_causal=is_causal, + per_block_mean=per_block_mean, + single_level_p_quant=single_level_p_quant + ) + + # Benchmark using the bench utility (handles warmup and timing) + avg_time_ms = bench(run_attention, num_warmups=num_warmups, num_tests=num_tests) + avg_time_s = avg_time_ms / 1000.0 + + # Calculate FLOPs + total_flops = calculate_attention_flops( + batch_size, num_heads, seq_len, seq_len, head_dim, is_causal + ) + + # Calculate TFLOPs + tflops = total_flops / (avg_time_s * 1e12) + + # Calculate throughput (tokens/sec) + tokens_per_second = (batch_size * seq_len) / avg_time_s + + return { + 'batch_size': batch_size, + 'num_heads': num_heads, + 'seq_len': seq_len, + 'head_dim': head_dim, + 'is_causal': is_causal, + 'dtype': str(dtype), + 'per_block_mean': per_block_mean, + 'single_level_p_quant': single_level_p_quant, + 'avg_time_ms': avg_time_ms, + 'avg_time_s': avg_time_s, + 'total_flops': total_flops, + 'tflops': tflops, + 'tokens_per_second': tokens_per_second, + } + + +def print_results(results): + """Print benchmark results in a formatted table.""" + print("\n" + "="*100) + print("SageAttention3 Benchmark Results") + print("="*100) + print(f"Configuration:") + print(f" Batch Size: {results['batch_size']}") + print(f" Num Heads: {results['num_heads']}") + print(f" Sequence Length: {results['seq_len']}") + print(f" Head Dimension: {results['head_dim']}") + print(f" Causal: {results['is_causal']}") + print(f" Data Type: {results['dtype']}") + print(f" Per Block Mean: {results['per_block_mean']}") + print(f" Single Level P Quant: {results['single_level_p_quant']}") + print(f"\nPerformance:") + print(f" Average Time: {results['avg_time_ms']:.3f} ms") + print(f" Total FLOPs: {results['total_flops']/1e12:.4f} TFLOPs (theoretical)") + print(f" Throughput: {results['tflops']:.4f} TFLOPs/s") + print(f" Tokens/sec: {results['tokens_per_second']:,.0f}") + print("="*100 + "\n") + sys.stdout.flush() + + +def run_benchmark_suite(): + """Run a comprehensive benchmark suite with various configurations.""" + print("Starting SageAttention3 Benchmark Suite...") + print(f"CUDA Device: {torch.cuda.get_device_name(0)}") + print(f"CUDA Version: {torch.version.cuda}") + print(f"PyTorch Version: {torch.__version__}\n") + sys.stdout.flush() + + # Default configurations to test + # (batch_size, num_heads, seq_len, head_dim, is_causal, dtype) + configs = [ + (1, 16, 1024, 64, False, torch.bfloat16), + (1, 16, 2048, 64, False, torch.bfloat16), + (1, 16, 4096, 64, False, torch.bfloat16), + (1, 16, 8192, 64, False, torch.bfloat16), + (1, 16, 16384, 64, False, torch.bfloat16), + + (1, 16, 1024, 128, False, torch.bfloat16), + (1, 16, 2048, 128, False, torch.bfloat16), + (1, 16, 4096, 128, False, torch.bfloat16), + (1, 16, 8192, 128, False, torch.bfloat16), + (1, 16, 16384, 128, False, torch.bfloat16), + + (1, 32, 1024, 64, False, torch.bfloat16), + (1, 32, 2048, 64, False, torch.bfloat16), + (1, 32, 4096, 64, False, torch.bfloat16), + (1, 32, 8192, 64, False, torch.bfloat16), + (1, 32, 16384, 64, False, torch.bfloat16), + + (1, 32, 1024, 128, False, torch.bfloat16), + (1, 32, 2048, 128, False, torch.bfloat16), + (1, 32, 4096, 128, False, torch.bfloat16), + (1, 32, 8192, 128, False, torch.bfloat16), + (1, 32, 16384, 128, False, torch.bfloat16), + ] + + all_results = [] + + for config in configs: + batch_size, num_heads, seq_len, head_dim, is_causal, dtype = config + + print(f"\nBenchmarking: B={batch_size}, H={num_heads}, L={seq_len}, D={head_dim}, " + f"Causal={is_causal}, dtype={dtype}...") + sys.stdout.flush() + + try: + results = benchmark_sageattn3( + batch_size=batch_size, + num_heads=num_heads, + seq_len=seq_len, + head_dim=head_dim, + is_causal=is_causal, + dtype=dtype, + num_warmups=10, + num_tests=50 + ) + + print_results(results) + all_results.append(results) + + except Exception as e: + print(f"Error benchmarking configuration {config}:") + print(f" Exception: {e}") + traceback.print_exc() + sys.stdout.flush() + continue + + # Print summary table + print("\n" + "="*120) + print("Summary Table") + print("="*120) + print(f"{'B':<4} {'H':<4} {'L':<6} {'D':<4} {'Causal':<7} {'Time (ms)':<12} {'TFLOPs/s':<12} {'Tokens/s':<15}") + print("-"*120) + + for r in all_results: + print(f"{r['batch_size']:<4} {r['num_heads']:<4} {r['seq_len']:<6} {r['head_dim']:<4} " + f"{str(r['is_causal']):<7} {r['avg_time_ms']:<12.3f} {r['tflops']:<12.4f} " + f"{r['tokens_per_second']:<15,.0f}") + + print("="*120 + "\n") + sys.stdout.flush() + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser(description='Benchmark SageAttention3 in TFLOPs') + parser.add_argument('--batch-size', type=int, default=None, help='Batch size') + parser.add_argument('--num-heads', type=int, default=None, help='Number of attention heads') + parser.add_argument('--seq-len', type=int, default=None, help='Sequence length') + parser.add_argument('--head-dim', type=int, default=None, help='Head dimension') + parser.add_argument('--causal', action='store_true', help='Use causal attention') + parser.add_argument('--dtype', type=str, default='bfloat16', choices=['bfloat16', 'float16'], + help='Data type') + parser.add_argument('--per-block-mean', action='store_true', default=True, + help='Use per-block mean for Q smoothing (default: True)') + parser.add_argument('--no-per-block-mean', action='store_false', dest='per_block_mean', + help='Disable per-block mean for Q smoothing') + parser.add_argument('--single-level-p-quant', action='store_true', default=False, + help='Use single-level P quantization (default: True)') + parser.add_argument('--two-level-p-quant', action='store_false', dest='single_level_p_quant', + help='Use two-level P quantization') + parser.add_argument('--num-warmups', type=int, default=10, help='Number of warmup iterations') + parser.add_argument('--num-tests', type=int, default=50, help='Number of test iterations') + parser.add_argument('--suite', action='store_true', help='Run full benchmark suite') + + args = parser.parse_args() + + dtype_map = { + 'bfloat16': torch.bfloat16, + 'float16': torch.float16 + } + + if args.suite: + run_benchmark_suite() + elif args.batch_size and args.num_heads and args.seq_len and args.head_dim: + results = benchmark_sageattn3( + batch_size=args.batch_size, + num_heads=args.num_heads, + seq_len=args.seq_len, + head_dim=args.head_dim, + is_causal=args.causal, + dtype=dtype_map[args.dtype], + per_block_mean=args.per_block_mean, + single_level_p_quant=args.single_level_p_quant, + num_warmups=args.num_warmups, + num_tests=args.num_tests + ) + print_results(results) + else: + print("Running default benchmark suite. Use --suite for full suite or provide all parameters.") + print("Example: python benchmark_sageattn3.py --batch-size 1 --num-heads 16 --seq-len 4096 --head-dim 128") + sys.stdout.flush() + run_benchmark_suite() + diff --git a/fastvideo/attention/backends/flash_attn.py b/fastvideo/attention/backends/flash_attn.py index 389a0f5aa..72ba38435 100644 --- a/fastvideo/attention/backends/flash_attn.py +++ b/fastvideo/attention/backends/flash_attn.py @@ -2,6 +2,7 @@ import torch import torch.nn.functional as F +import torch.nn.functional as F from flash_attn import flash_attn_func as flash_attn_2_func from dataclasses import dataclass diff --git a/fastvideo/attention/backends/sage_attn3.py b/fastvideo/attention/backends/sage_attn3.py index b2d3379bb..2b4aede2e 100644 --- a/fastvideo/attention/backends/sage_attn3.py +++ b/fastvideo/attention/backends/sage_attn3.py @@ -1,13 +1,480 @@ # SPDX-License-Identifier: Apache-2.0 +import os +import sys +from pathlib import Path + +# Add project root to sys.path to allow importing qat_attn +_project_root = Path(__file__).resolve().parent.parent.parent.parent +if str(_project_root) not in sys.path: + sys.path.insert(0, str(_project_root)) + import torch -from sageattn3 import sageattn3_blackwell +import triton +try: + # Use local version first (avoids issues with installed package) + from fastvideo.attention.backends.sageattn import sageattn_blackwell +except ImportError: + # Fall back to installed package if local version not available + try: + from sageattn import sageattn_blackwell + except ImportError: + sageattn_blackwell = None from fastvideo.attention.backends.abstract import (AttentionBackend, AttentionImpl, AttentionMetadata, AttentionMetadataBuilder) from fastvideo.logger import init_logger +from test_quantization_precision import get_fake_quant +from qat_attn import attention +from qat_attn import _attn_bwd_preprocess, _attn_bwd_dq_cross, _attn_bwd_dkdv_cross, _attn_bwd +# from fastvideo.attention.backends.sageattn.api import triton_group_mean +logger = init_logger(__name__) + +from math import sqrt +try: + from flash_attn.flash_attn_interface import _wrapped_flash_attn_forward, _wrapped_flash_attn_backward +except ImportError: + _wrapped_flash_attn_forward = None + _wrapped_flash_attn_backward = None logger = init_logger(__name__) +from flashinfer import SfLayout, mm_fp4, nvfp4_quantize +from quant_utils import fake_quantize_q, fake_quantize_kv + +def fp_16_qk(q, k): + qk = torch.matmul(q, k) / torch.sqrt(torch.tensor(q.shape[-1], device=q.device, dtype=q.dtype)) + p = torch.softmax(qk, dim=-1) + return p + + +def matmul_3d_4bit(a, b): + H, M, K1 = a.shape + H, N, K2 = b.shape + assert K1 == K2 + padded_K = (K2 + 127) // 128 * 128 + a = torch.nn.functional.pad(a, (0, padded_K - K2)) + b = torch.nn.functional.pad(b, (0, padded_K - K2)) + out = torch.empty((H, M, N), device=a.device, dtype=a.dtype) + one = torch.ones(1, device=a.device, dtype=torch.float32) + for h in range(H): + q_fp4, q_inv_s = nvfp4_quantize(a[h], one, sfLayout=SfLayout.layout_128x4, do_shuffle=False) + k_fp4, k_inv_s = nvfp4_quantize(b[h], one, sfLayout=SfLayout.layout_128x4, do_shuffle=False) + mm_fp4( + q_fp4, k_fp4.T, q_inv_s, k_inv_s.T, one, + a.dtype, out[h], + block_size=16, + use_8x4_sf_layout=False, + backend="cutlass", + ) + return out + + +class _Matmul3d4bitFWD16bitBWD(torch.autograd.Function): + @staticmethod + def forward(ctx, a, b, scale=None): + ctx.save_for_backward(a, b) + if scale is not None: + a = a / scale + return matmul_3d_4bit(a, b) * scale + else: + return matmul_3d_4bit(a, b) + + @staticmethod + def backward(ctx, grad_out): + a, b = ctx.saved_tensors + grad_a = grad_out.matmul(b) + grad_b = grad_out.transpose(1, 2).matmul(a) + # None for the three extra forward args + return grad_a, grad_b, None + + +class _Matmul3d4bitFWD16bitBWD_with_smoothing(torch.autograd.Function): + @staticmethod + def forward(ctx, q, k): + ctx.save_for_backward(q, k) + k = k - k.mean(dim=-2, keepdim=True) # [H, L, D] + qm = q.mean(dim=-2, keepdim=True) # [H, 1, D] + q = q - qm # [B, H, L, D] + delta_s = torch.matmul(qm, k.transpose(-2, -1)).contiguous() # [H, 1, L] + return matmul_3d_4bit(q, k) + delta_s + + @staticmethod + def backward(ctx, grad_out): + a, b = ctx.saved_tensors + grad_a = grad_out.matmul(b) + grad_b = grad_out.transpose(1, 2).matmul(a) + # None for the three extra forward args + return grad_a, grad_b, None + +def attn_forward_4bit_fwd_16bit_bwd(q, k, v, is_causal=False, per_block_mean=False): + assert is_causal == False + assert per_block_mean == False + q, k, v = q.transpose(1, 2).contiguous(), k.transpose(1, 2).contiguous(), v.transpose(1, 2).contiguous() + p = sage_attn_qk_torch(q, k).to(q.dtype) + row_max_p = p.max(dim=-1, keepdim=True).values + scale_row = row_max_p / (448 * 6) + out = _Matmul3d4bitFWD16bitBWD.apply(p.squeeze(0), v.squeeze(0).transpose(1, 2).contiguous(), scale_row.squeeze(0)).unsqueeze(0) + return out.transpose(1, 2).contiguous() + +def preprocess(q, k, v): + return get_fake_quant(q), get_fake_quant(k), get_fake_quant(v) + +def qat_attn_qk_torch(fq_q, fq_k): + """ + Inputs: q/k/v in [B, H, L, D] + Returns: out in [B, H, L, D] + """ + B, H, qL, D = fq_q.shape + B, H, kL, D = fq_k.shape + assert B == 1 + q, k = fq_q.squeeze(0), fq_k.squeeze(0) + qk = torch.matmul(q, k.transpose(-2, -1)) / torch.sqrt(torch.tensor(q.shape[-1], device=q.device, dtype=q.dtype)) + qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) + p = torch.softmax(qk, dim=-1) + return p.unsqueeze(0) + +def qat_attn_backward_16bit(fq_q_BLHD, fq_k_BLHD, fq_v_BLHD, high_prec_o_BLHD, grad_out_BLHD, is_causal=False): + assert is_causal == False + fq_q, fq_k, fq_v, high_prec_o, do = fq_q_BLHD.transpose(1 ,2).contiguous(), fq_k_BLHD.transpose(1 ,2).contiguous(), fq_v_BLHD.transpose(1 ,2).contiguous(), high_prec_o_BLHD.transpose(1 ,2).contiguous(), grad_out_BLHD.transpose(1 ,2).contiguous() + B, H, _, D = fq_q.shape + p = qat_attn_qk_torch(fq_q, fq_k).to(do.dtype) + fq_p = get_fake_quant(p).unsqueeze(0) + dV = fq_p.transpose(-2, -1) @ do # [B,H,L,D] + + dP = do @ fq_v.transpose(-2, -1) # [B,H,L,L] + + do_dot_o = (do * high_prec_o).sum(dim=-1, keepdim=True) # [B,H,L,1] + dS = p * (dP - do_dot_o) # [B,H,L,L] + + + dQ = (dS @ fq_k) / torch.sqrt(torch.tensor(D, device=fq_q.device, dtype=fq_q.dtype)) # [B,H,L,D] + dK = (dS.transpose(-2, -1) @ fq_q) / torch.sqrt(torch.tensor(D, device=fq_q.device, dtype=fq_q.dtype)) # [B,H,L,D] + + # grads w.r.t inputs, None for non-tensor inputs + return dQ.transpose(1, 2).contiguous(), dK.transpose(1, 2).contiguous(), dV.transpose(1, 2).contiguous(), None, None, None + +USE_SMOOTHING = False +def sage_attn_qk_torch(q, k, per_block_mean=False): + """ + Inputs: q/k/v in [B, H, L, D] + Returns: out in [B, H, L, D] + """ + B, H, qL, D = q.shape + B, H, kL, D = k.shape + assert B == 1 + q, k = q.squeeze(0), k.squeeze(0) + if USE_SMOOTHING: + qk = _Matmul3d4bitFWD16bitBWD_with_smoothing.apply(q, k) + else: + qk = _Matmul3d4bitFWD16bitBWD.apply(q, k) + qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) + p = torch.softmax(qk, dim=-1) + return p.unsqueeze(0) + +def attn_backward_16bit(q_BLHD, k_BLHD, v_BLHD, out_BLHD, grad_out_BLHD, is_causal=False): + assert is_causal == False + q, k, v, o, do = q_BLHD.transpose(1 ,2).contiguous(), k_BLHD.transpose(1 ,2).contiguous(), v_BLHD.transpose(1 ,2).contiguous(), out_BLHD.transpose(1 ,2).contiguous(), grad_out_BLHD.transpose(1 ,2).contiguous() + B, H, _, D = q.shape + p = sage_attn_qk_torch(q, k, per_block_mean=False).to(do.dtype) + dV = p.transpose(-2, -1) @ do # [B,H,L,D] + + dP = do @ v.transpose(-2, -1) # [B,H,L,L] + + # do_dot_o = (do.float() * o.float()).sum(dim=-1, keepdim=True).to(dP.dtype) # [B,H,L,1] + # do_dot_o = (do * o).sum(dim=-1, keepdim=True) # [B,H,L,1] + dp_dot_p = (dP * p).sum(dim=-1, keepdim=True) # [B,H,L,1] + dS = p * (dP - dp_dot_p) # [B,H,L,L] + + + dQ = (dS @ k) / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) # [B,H,L,D] + dK = (dS.transpose(-2, -1) @ q) / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) # [B,H,L,D] + + # grads w.r.t inputs, None for non-tensor inputs + return dQ.transpose(1, 2).contiguous(), dK.transpose(1, 2).contiguous(), dV.transpose(1, 2).contiguous(), None, None, None + +def attn_forward_4bit(q_BHLD, k_BHLD, v_BHLD, is_causal=False): + assert is_causal == False + p_bf16 = sage_attn_qk_torch(q_BHLD, k_BHLD, per_block_mean=False).to(q_BHLD.dtype) + row_max_p = p_bf16.max(dim=-1, keepdim=True).values + scale_row = row_max_p / (448 * 6) + p = p_bf16 / scale_row + out = matmul_3d_4bit(p.squeeze(0), v_BHLD.squeeze(0).transpose(1, 2).contiguous()).unsqueeze(0) + out = out * scale_row + return out, p_bf16 + + +def flash_attn_backward(q_BLHD, k_BLHD, v_BLHD, out_BLHD, grad_out_BLHD, is_causal=False): + D = q_BLHD.shape[-1] + softmax_scale = 1.0 / sqrt(D) + + # FA forward only to get softmax_lse (rng_state is also returned) + _, softmax_lse, _S_dmask, rng_state = _wrapped_flash_attn_forward( + q_BLHD, k_BLHD, v_BLHD, + 0.0, # dropout_p + softmax_scale, # softmax_scale + is_causal, # causal + -1, -1, # window_size_left/right + 0.0, # softcap + None, # alibi_slopes + False # return_softmax + ) + + # Allocate grads and call FA backward using the SAGE output + dq_BLHD = torch.empty_like(q_BLHD) + dk_BLHD = torch.empty_like(k_BLHD) + dv_BLHD = torch.empty_like(v_BLHD) + + _wrapped_flash_attn_backward( + grad_out_BLHD, # dout (BLHD, 16-bit) + q_BLHD, k_BLHD, v_BLHD, + out_BLHD, # use SAGE forward output here + softmax_lse, + dq_BLHD, dk_BLHD, dv_BLHD, + 0.0, # dropout_p + softmax_scale, # softmax_scale + is_causal, # causal + -1, -1, # window_size_left/right + 0.0, # softcap + None, # alibi_slopes + False, # deterministic + rng_state=rng_state, + ) + return dq_BLHD, dk_BLHD, dv_BLHD, None, None + +USE_TORCH_4BIT_FWD = False +USE_TORCH_16BIT_BWD = False +USE_QAT_ATTN = False +class _SageAttnBlackwellWith16bitBwd(torch.autograd.Function): + @staticmethod + def forward(ctx, q_BLHD, k_BLHD, v_BLHD, is_causal=False, per_block_mean=True): + """ + Inputs: q/k/v in [B, L, H, D] + Returns: out in [B, L, H, D] + """ + # Save originals for backward (we'll recompute FA in 16-bit there) + ctx.is_causal = bool(is_causal) + + # Convert to BHLD for your sageattn forward + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + + with torch.no_grad(): + out_BHLD = sageattn_blackwell( + q_BHLD, k_BHLD, v_BHLD, + attn_mask=None, + is_causal=is_causal, + per_block_mean=per_block_mean, + ) + # Convert to BLHD for saving + out_BLHD = out_BHLD.permute(0, 2, 1, 3).contiguous() + ctx.save_for_backward(q_BLHD, k_BLHD, v_BLHD, out_BLHD) + + return out_BLHD + + @staticmethod + def backward(ctx, grad_out_BLHD): + is_causal = ctx.is_causal + q_BLHD, k_BLHD, v_BLHD, out_BLHD = ctx.saved_tensors + return flash_attn_backward(q_BLHD, k_BLHD, v_BLHD, out_BLHD, grad_out_BLHD, is_causal) + + + +def sageattn_blackwell_with_16bit_bwd(q_BLHD, k_BLHD, v_BLHD, is_causal=False, per_block_mean=True): + """ + Forward: uses sageattn_blackwell under the hood. + Backward: recomputes FlashAttention fwd+bwd in 16bit directly. + """ + return _SageAttnBlackwellWith16bitBwd.apply(q_BLHD, k_BLHD, v_BLHD, is_causal, per_block_mean) + +class _SageAttnBlackwellWithTritonBwd(torch.autograd.Function): + @staticmethod + def forward(ctx, q_BHLD, k_BHLD, v_BHLD, softmax_scale, softmax_lse, is_causal=False, per_block_mean=True, use_global_sf=True): + k_mean = k_BHLD.mean(dim=(0, 1, 2), keepdim=True).view(-1) + + out_BHLD = sageattn_blackwell( + q_BHLD, k_BHLD, v_BHLD, + attn_mask=None, + is_causal=is_causal, + per_block_mean=per_block_mean, + ) + + ctx.HEAD_DIM = q_BHLD.shape[-1] + ctx.sm_scale = softmax_scale # Use sm_scale to match backward access + ctx.causal = is_causal + ctx.IS_QAT = True + ctx.k_mean = k_mean + ctx.use_global_sf = use_global_sf + ctx.save_for_backward(q_BHLD, k_BHLD, v_BHLD, out_BHLD, softmax_lse) + return out_BHLD + + @staticmethod + def backward(ctx, do): + # note this assumes that we do everything in algorithm 1 from SageAttention3 except adding delta s + q_BHLD, k_BHLD, v_BHLD, o_for_bwd_BHLD, M_BHLD = ctx.saved_tensors + do = do.contiguous() + dq_BHLD = torch.empty_like(q_BHLD) + dk_BHLD = torch.empty_like(k_BHLD) + dv_BHLD = torch.empty_like(v_BHLD) + BATCH, N_HEAD, N_CTX_Q = q_BHLD.shape[:3] + N_CTX_KV = k_BHLD.shape[2] + assert k_BHLD.shape[2] == v_BHLD.shape[2], "k and v must have the same sequence length" + PRE_BLOCK = 128 + NUM_STAGES = 1 + NUM_WARPS = 4 + BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 64, 64, 64, 64 + BLK_SLICE_FACTOR = 1 + pre_grid = ((N_CTX_Q + PRE_BLOCK - 1) // PRE_BLOCK, BATCH * N_HEAD) + delta = torch.empty_like(M_BHLD) + _attn_bwd_preprocess[pre_grid]( + o_for_bwd_BHLD, do, + delta, + BATCH, N_HEAD, N_CTX_Q, + BLOCK_M=PRE_BLOCK, HEAD_DIM=ctx.HEAD_DIM + ) + + if ctx.IS_QAT: + fake_q = torch.empty_like(q_BHLD) + fake_k = torch.empty_like(k_BHLD) + fake_v = torch.empty_like(v_BHLD) + + grid_1 = (triton.cdiv(q_BHLD.shape[2], BLOCK_M1), q_BHLD.shape[0] * q_BHLD.shape[1], 1) + grid_2 = (triton.cdiv(k_BHLD.shape[2], BLOCK_N2), q_BHLD.shape[0] * q_BHLD.shape[1], 1) + + fake_quantize_q[grid_1]( + q_BHLD, fake_q, + q_BHLD.stride(0), q_BHLD.stride(1), + q_BHLD.stride(2), q_BHLD.stride(3), + fake_q.stride(0), fake_q.stride(1), + fake_q.stride(2), fake_q.stride(3), + N_HEAD, N_CTX_Q, + BLOCK_M=BLOCK_M1, HEAD_DIM=ctx.HEAD_DIM, use_global_sf=ctx.use_global_sf + ) + + fake_quantize_kv[grid_2]( + k_BHLD, v_BHLD, fake_k, fake_v, + k_BHLD.stride(0), k_BHLD.stride(1), + k_BHLD.stride(2), k_BHLD.stride(3), + fake_k.stride(0), fake_k.stride(1), + fake_k.stride(2), fake_k.stride(3), + N_HEAD, N_CTX_KV, + BLOCK_N=BLOCK_N2, HEAD_DIM=ctx.HEAD_DIM, use_global_sf=ctx.use_global_sf + ) + + # Use fake quantized tensors for gradient computation (QAT) + # NOTE: The reassignment below only changes local variable references. + # The gradients (dq_BHLD, dk_BHLD, dv_BHLD) computed using fake_q/fake_k/fake_v + # will still correctly backpropagate to the original q_BHLD/k_BHLD/v_BHLD tensors + # from ctx.saved_tensors, because those original tensors remain in the computation graph. + q_BHLD = fake_q + k_BHLD = fake_k + v_BHLD = fake_v + + # NOTE: K is NOT pre-scaled here - scaling is applied AFTER qk dot product in kernels + # This improves precision by avoiding rounding errors in K before the dot product + arg_k = k_BHLD + + if N_CTX_Q == N_CTX_KV: + # Use existing kernel for self-attention (same sequence lengths) + grid = ((N_CTX_KV + BLOCK_N1 - 1) // BLOCK_N1, 1, BATCH * N_HEAD) + # Q_MEAN is required by the kernel signature but only used if SMOOTH_Q=True + # Since we're not using SMOOTH_Q, create a dummy tensor with the same shape as q_BHLD + q_mean = torch.empty_like(q_BHLD) + _attn_bwd[grid]( + q_BHLD, arg_k, v_BHLD, ctx.sm_scale, do, dq_BHLD, dk_BHLD, dv_BHLD, + M_BHLD, delta, q_mean, + q_BHLD.stride(0), q_BHLD.stride(1), q_BHLD.stride(2), q_BHLD.stride(3), + N_HEAD, N_CTX_KV, + ctx.k_mean, + BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1, + BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2, + BLK_SLICE_FACTOR=BLK_SLICE_FACTOR, + HEAD_DIM=ctx.HEAD_DIM, + CAUSAL=ctx.causal, + IS_QAT=ctx.IS_QAT, + SMOOTH_K=True, + two_level_quant_P=True, + fake_quant_P=True, + use_global_sf=ctx.use_global_sf, + num_warps=NUM_WARPS, + num_stages=NUM_STAGES + ) + else: + # Use separate kernels for cross-attention (different sequence lengths) + grid_dq = ((N_CTX_Q + BLOCK_M2 - 1) // BLOCK_M2, 1, BATCH * N_HEAD) + _attn_bwd_dq_cross[grid_dq]( + q_BHLD, arg_k, v_BHLD, ctx.sm_scale, do, dq_BHLD, M_BHLD, delta, + q_BHLD.stride(0), k_BHLD.stride(0), q_BHLD.stride(1), k_BHLD.stride(1), q_BHLD.stride(2), k_BHLD.stride(2), q_BHLD.stride(3), k_BHLD.stride(3), + N_HEAD, N_CTX_Q, N_CTX_KV, + ctx.k_mean, + BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2, + HEAD_DIM=ctx.HEAD_DIM, + num_warps=NUM_WARPS, + num_stages=NUM_STAGES, + SMOOTH_K=True, + ) + grid_dkdv = ((N_CTX_KV + BLOCK_N1 - 1) // BLOCK_N1, 1, BATCH * N_HEAD) + # Q_MEAN is required by the kernel signature but only used if SMOOTH_Q=True + # Since we're not using SMOOTH_Q, create a dummy tensor with the same shape as q_BHLD + q_mean = torch.empty_like(q_BHLD) + _attn_bwd_dkdv_cross[grid_dkdv]( + q_BHLD, arg_k, v_BHLD, ctx.sm_scale, do, dk_BHLD, dv_BHLD, M_BHLD, delta, q_mean, + q_BHLD.stride(0), k_BHLD.stride(0), q_BHLD.stride(1), k_BHLD.stride(1), q_BHLD.stride(2), k_BHLD.stride(2), q_BHLD.stride(3), k_BHLD.stride(3), + N_HEAD, N_CTX_Q, N_CTX_KV, + BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1, + HEAD_DIM=ctx.HEAD_DIM, + IS_QAT=ctx.IS_QAT, + two_level_quant_P=True, + fake_quant_P=True, + use_global_sf=ctx.use_global_sf, + num_warps=NUM_WARPS, + num_stages=NUM_STAGES + ) + + return dq_BHLD, dk_BHLD, dv_BHLD, None, None, None, None, None + +def sageattn_blackwell_with_triton_bwd(q_BLHD, k_BLHD, v_BLHD, is_causal=False, per_block_mean=True, use_global_sf=False): + softmax_scale = 1.0 / sqrt(q_BLHD.shape[-1]) + # FA forward only to get softmax_lse (which is shape BHL) + # Use no_grad to prevent autograd from tracking this call + with torch.no_grad(): + _, softmax_lse, _S_dmask, rng_state = _wrapped_flash_attn_forward( + q_BLHD, k_BLHD, v_BLHD, + 0.0, # dropout_p + softmax_scale, # softmax_scale + is_causal, # causal + -1, -1, # window_size_left/right + 0.0, # softcap + None, # alibi_slopes + False # return_softmax + ) + + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_BHLD = _SageAttnBlackwellWithTritonBwd.apply(q_BHLD, k_BHLD, v_BHLD, softmax_scale, softmax_lse, is_causal, per_block_mean, use_global_sf) + return out_BHLD.permute(0, 2, 1, 3).contiguous() + +def qat_attn(q_BLHD, k_BLHD, v_BLHD, is_causal=False): + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + + use_qat_qkv_backward = True + smooth_k = False + warp_specialize = True + IS_QAT = True + two_level_quant_P_sage3 = False + fake_quant_P_bwd = True + use_high_prec_o = True + smooth_q = False + sm_scale = 1.0 / sqrt(q_BHLD.shape[-1]) + use_global_sf_qkv = False + use_global_sf_p = False + o_BHLD = attention(q_BHLD, k_BHLD, v_BHLD, is_causal, sm_scale, + use_qat_qkv_backward, smooth_k, warp_specialize, IS_QAT, + two_level_quant_P_sage3, fake_quant_P_bwd, use_high_prec_o, smooth_q, use_global_sf_p, use_global_sf_qkv) + return o_BHLD.permute(0, 2, 1, 3).contiguous() class SageAttention3Backend(AttentionBackend): @@ -16,7 +483,7 @@ class SageAttention3Backend(AttentionBackend): @staticmethod def get_supported_head_sizes() -> list[int]: - return [64, 128, 256] + return [64, 128] @staticmethod def get_name() -> str: @@ -62,9 +529,7 @@ class SageAttention3Impl(AttentionImpl): value: torch.Tensor, attn_metadata: AttentionMetadata, ) -> torch.Tensor: - query = query.transpose(1, 2) - key = key.transpose(1, 2) - value = value.transpose(1, 2) - output = sageattn3_blackwell(query, key, value, is_causal=self.causal) - output = output.transpose(1, 2) + output = sageattn_blackwell_with_16bit_bwd(query, key, value, is_causal=self.causal) + # output = sageattn_blackwell_with_triton_bwd(query, key, value, is_causal=self.causal) + # output = qat_attn(query, key, value, is_causal=self.causal) return output diff --git a/fastvideo/attention/backends/sageattn/__init__.py b/fastvideo/attention/backends/sageattn/__init__.py new file mode 100644 index 000000000..9b2ee4e37 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/__init__.py @@ -0,0 +1,16 @@ +""" +Copyright (c) 2025 by SageAttention team. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +""" +from .api import sageattn_blackwell \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/api.py b/fastvideo/attention/backends/sageattn/api.py new file mode 100644 index 000000000..5adab6f11 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/api.py @@ -0,0 +1,184 @@ +""" +Copyright (c) 2025 by SageAttention team. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +""" +import torch +import triton +import triton.language as tl +import torch.nn.functional as F +from typing import Tuple +from torch.nn.functional import scaled_dot_product_attention as sdpa +import fp4attn_cuda +import fp4quant_cuda + +# Centralized block size configuration for sageattn_blackwell kernels +# These should match the values in fastvideo/attention/backends/sageattn/blackwell/block_config.h +BLOCK_M = 128 # Block size for M dimension (query sequence length) +BLOCK_N = 128 # Block size for N dimension (key/value sequence length) + + +@triton.jit +def group_mean_kernel( + q_ptr, + q_out_ptr, + qm_out_ptr, + B, H, L, D: tl.constexpr, + stride_qb, stride_qh, stride_ql, stride_qd, + stride_qmb, stride_qmh, stride_qml, stride_qmd, + GROUP_SIZE: tl.constexpr +): + pid_b = tl.program_id(0) + pid_h = tl.program_id(1) + pid_group = tl.program_id(2) + + group_start = pid_group * GROUP_SIZE + offsets = group_start + tl.arange(0, GROUP_SIZE) + + q_offsets = pid_b * stride_qb + pid_h * stride_qh + offsets[:, None] * stride_ql + tl.arange(0, D)[None, :] * stride_qd + q_group = tl.load(q_ptr + q_offsets) + + qm_group = tl.sum(q_group, axis=0) / GROUP_SIZE + + q_group = q_group - qm_group + tl.store(q_out_ptr + q_offsets, q_group) + + qm_offset = pid_b * stride_qmb + pid_h * stride_qmh + pid_group * stride_qml + tl.arange(0, D) * stride_qmd + tl.store(qm_out_ptr + qm_offset, qm_group) + + +def triton_group_mean(q: torch.Tensor): + B, H, L, D = q.shape + GROUP_SIZE = BLOCK_M + num_groups = L // GROUP_SIZE + + q_out = torch.empty_like(q) # [B, H, L, D] + qm = torch.empty(B, H, num_groups, D, device=q.device, dtype=q.dtype) + + grid = (B, H, num_groups) + + group_mean_kernel[grid]( + q, q_out, qm, + B, H, L, D, + q.stride(0), q.stride(1), q.stride(2), q.stride(3), + qm.stride(0), qm.stride(1), qm.stride(2), qm.stride(3), + GROUP_SIZE=GROUP_SIZE + ) + return q_out, qm + + +def preprocess_qkv(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, per_block_mean: bool = True, enable_smoothing_q: bool = False, enable_smoothing_k: bool = False): + + def pad_to_block_size(x): + L = x.size(2) + pad_len = (BLOCK_M - L % BLOCK_M) % BLOCK_M + if pad_len == 0: + return x.contiguous() + return F.pad(x, (0, 0, 0, pad_len), value=0).contiguous() + + if enable_smoothing_k: + k -= k.mean(dim=-2, keepdim=True) + q, k, v = map(lambda x: pad_to_block_size(x), [q, k, v]) + if per_block_mean and enable_smoothing_q: + q, qm = triton_group_mean(q) + elif enable_smoothing_q: + qm = q.mean(dim=-2, keepdim=True) + q = q - qm + if enable_smoothing_q: + delta_s = torch.matmul(qm, k.transpose(-2, -1)).to(torch.float32).contiguous() + else: # used to disable q smoothing + B, H, L, D = q.shape + delta_s = torch.zeros((B, H, L // BLOCK_M, k.shape[2]), device=q.device, dtype=torch.float32) + + return q, k, v, delta_s + +def scale_and_quant_fp4(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + assert x.ndim == 4 + B, H, N, D = x.shape + packed_fp4 = torch.empty((B, H, N, D // 2), device=x.device, dtype=torch.uint8) + fp8_scale = torch.empty((B, H, N, D // 16), device=x.device, dtype=torch.float8_e4m3fn) + fp4quant_cuda.scaled_fp4_quant(x, packed_fp4, fp8_scale, 1) + return packed_fp4, fp8_scale + +def scale_and_quant_fp4_permute(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + assert x.ndim == 4 + B, H, N, D = x.shape + packed_fp4 = torch.empty((B, H, N, D // 2), device=x.device, dtype=torch.uint8) + fp8_scale = torch.empty((B, H, N, D // 16), device=x.device, dtype=torch.float8_e4m3fn) + fp4quant_cuda.scaled_fp4_quant_permute(x, packed_fp4, fp8_scale, 1) + return packed_fp4, fp8_scale + +def scale_and_quant_fp4_transpose(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + assert x.ndim == 4 + B, H, N, D = x.shape + packed_fp4 = torch.empty((B, H, D, N // 2), device=x.device, dtype=torch.uint8) + fp8_scale = torch.empty((B, H, D, N // 16), device=x.device, dtype=torch.float8_e4m3fn) + fp4quant_cuda.scaled_fp4_quant_trans(x, packed_fp4, fp8_scale, 1) + return packed_fp4, fp8_scale + +def blockscaled_fp4_attn(qlist: Tuple, + klist: Tuple, + vlist: Tuple, + delta_s: torch.Tensor, + KL: int, + is_causal: bool = False, + per_block_mean: bool = True, + is_bf16: bool = True, + single_level_p_quant: bool = False + ): + softmax_scale = (qlist[0].shape[-1] * 2) ** (-0.5) + return fp4attn_cuda.fwd(qlist[0], klist[0], vlist[0], qlist[1], klist[1], vlist[1], delta_s, KL, None, softmax_scale, is_causal, per_block_mean, is_bf16, single_level_p_quant) + + +def sageattn_blackwell(q, k, v, attn_mask = None, is_causal = False, per_block_mean = True, single_level_p_quant = True, **kwargs): + """ + SageAttention3 Blackwell kernel for FP4 attention. + + Args: + q: Query tensor [B, H, L, D] + k: Key tensor [B, H, L, D] + v: Value tensor [B, H, L, D] + attn_mask: Attention mask (not used) + is_causal: Whether to use causal masking + per_block_mean: Whether to use per-block mean for Q smoothing + single_level_p_quant: If True, use single-level quantization: s_P2, P̂_2 = φ(P̃) directly + (standard per-block FP4 quantization like V, no s_P1). + If False (default), use two-level quantization: + s_P1 = rowmax(P̃)/(448×6), then s_P2, P̂_2 = φ(P̃/s_P1). + **kwargs: Additional arguments (ignored) + + Returns: + Output tensor [B, H, L, D] + """ + if q.size(-1) >= 256: + print(f"Unsupported Headdim {q.size(-1)}") + return sdpa(q, k, v, is_causal = is_causal) + QL = q.size(2) + KL = k.size(2) + is_bf16 = q.dtype == torch.bfloat16 + q, k, v, delta_s = preprocess_qkv(q, k, v, per_block_mean) + qlist_from_cuda = scale_and_quant_fp4(q) + klist_from_cuda = scale_and_quant_fp4_permute(k) + vlist_from_cuda = scale_and_quant_fp4_transpose(v) + o_fp4 = blockscaled_fp4_attn( + qlist_from_cuda, + klist_from_cuda, + vlist_from_cuda, + delta_s, + KL, + is_causal, + per_block_mean, + is_bf16, + single_level_p_quant + )[0][:, :, :QL, :].contiguous() + return o_fp4 \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/blackwell/__init__.py b/fastvideo/attention/backends/sageattn/blackwell/__init__.py new file mode 100644 index 000000000..2e33087c5 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/__init__.py @@ -0,0 +1 @@ +__version__ = "3.0.0.b1" diff --git a/fastvideo/attention/backends/sageattn/blackwell/api.cu b/fastvideo/attention/backends/sageattn/blackwell/api.cu new file mode 100644 index 000000000..8bb6ee3a6 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/api.cu @@ -0,0 +1,345 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +// Include these 2 headers instead of torch/extension.h since we don't need all of the torch headers. +#include +#include +#include +#include + +#include + +#include "params.h" +#include "launch.h" +#include "static_switch.h" +#include "block_config.h" + +#define CHECK_DEVICE(x) TORCH_CHECK(x.is_cuda(), #x " must be on CUDA") +#define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")") +#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") + + +void set_params_fprop(Flash_fwd_params ¶ms, + // sizes + const size_t b, + const size_t seqlen_q, + const size_t seqlen_k, + const size_t unpadded_seqlen_k, + const size_t seqlen_q_rounded, + const size_t seqlen_k_rounded, + const size_t h, + const size_t h_k, + const size_t d, + const size_t d_rounded, + // device pointers + const at::Tensor q, + const at::Tensor k, + const at::Tensor v, + const at::Tensor delta_s, + at::Tensor out, + const at::Tensor sfq, + const at::Tensor sfk, + const at::Tensor sfv, + void *cu_seqlens_q_d, + void *cu_seqlens_k_d, + void *seqused_k, + void *p_d, + void *softmax_lse_d, + float p_dropout, + float softmax_scale, + int window_size_left, + int window_size_right, + bool per_block_mean, + bool is_bf16, + bool single_level_p_quant=false, + bool seqlenq_ngroups_swapped=false) { + + // Reset the parameters + params = {}; + // Set the pointers and strides. + params.q_ptr = q.data_ptr(); + params.k_ptr = k.data_ptr(); + params.v_ptr = v.data_ptr(); + params.delta_s_ptr = delta_s.data_ptr(); + params.sfq_ptr = sfq.data_ptr(); + params.sfk_ptr = sfk.data_ptr(); + params.sfv_ptr = sfv.data_ptr(); + + // All stride are in elements, not bytes. + params.q_row_stride = q.stride(-2) * 2; + params.k_row_stride = k.stride(-2) * 2; + params.v_row_stride = v.stride(-2) * 2;; + params.q_head_stride = q.stride(-3) * 2; + params.k_head_stride = k.stride(-3) * 2; + params.v_head_stride = v.stride(-3) * 2; // for packed q k v + + params.ds_row_stride = delta_s.stride(-2); + params.ds_head_stride = delta_s.stride(-3); + + params.sfq_row_stride = sfq.stride(-2); + params.sfk_row_stride = sfk.stride(-2); + params.sfv_row_stride = sfv.stride(-2); + params.sfq_head_stride = sfq.stride(-3); + params.sfk_head_stride = sfk.stride(-3); + params.sfv_head_stride = sfv.stride(-3); + params.o_ptr = out.data_ptr(); + params.o_row_stride = out.stride(-2); + params.o_head_stride = out.stride(-3); + + if (cu_seqlens_q_d == nullptr) { + params.q_batch_stride = q.stride(0) * 2; + params.k_batch_stride = k.stride(0) * 2; + params.v_batch_stride = v.stride(0) * 2; + params.ds_batch_stride = delta_s.stride(0); + params.sfq_batch_stride = sfq.stride(0); + params.sfk_batch_stride = sfk.stride(0); + params.sfv_batch_stride = sfv.stride(0); + params.o_batch_stride = out.stride(0); + if (seqlenq_ngroups_swapped) { + params.q_batch_stride *= seqlen_q; + params.o_batch_stride *= seqlen_q; + } + } + + params.cu_seqlens_q = static_cast(cu_seqlens_q_d); + params.cu_seqlens_k = static_cast(cu_seqlens_k_d); + params.seqused_k = static_cast(seqused_k); + + // P = softmax(QK^T) + params.p_ptr = p_d; + + // Softmax sum + params.softmax_lse_ptr = softmax_lse_d; + + // Set the dimensions. + params.b = b; + params.h = h; + params.h_k = h_k; + params.h_h_k_ratio = h / h_k; + params.seqlen_q = seqlen_q; + params.seqlen_k = seqlen_k; + params.unpadded_seqlen_k = unpadded_seqlen_k; + params.seqlen_q_rounded = seqlen_q_rounded; + params.seqlen_k_rounded = seqlen_k_rounded; + params.d = d; + params.d_rounded = d_rounded; + + params.head_divmod = cutlass::FastDivmod(int(h)); + + // Set the different scale values. + params.scale_softmax = softmax_scale; + params.scale_softmax_log2 = softmax_scale * M_LOG2E; + __half scale_softmax_log2_half = __float2half(params.scale_softmax_log2); + __half2 scale_softmax_log2_half2 = __half2(scale_softmax_log2_half, scale_softmax_log2_half); + params.scale_softmax_log2_half2 = reinterpret_cast(scale_softmax_log2_half2); + + // Set this to probability of keeping an element to simplify things. + params.p_dropout = 1.f - p_dropout; + // Convert p from float to int so we don't have to convert the random uint to float to compare. + // [Minor] We want to round down since when we do the comparison we use <= instead of < + // params.p_dropout_in_uint = uint32_t(std::floor(params.p_dropout * 4294967295.0)); + // params.p_dropout_in_uint16_t = uint16_t(std::floor(params.p_dropout * 65535.0)); + params.p_dropout_in_uint8_t = uint8_t(std::floor(params.p_dropout * 255.0)); + params.rp_dropout = 1.f / params.p_dropout; + params.scale_softmax_rp_dropout = params.rp_dropout * params.scale_softmax; + TORCH_CHECK(p_dropout < 1.f); + #ifdef FLASHATTENTION_DISABLE_DROPOUT + TORCH_CHECK(p_dropout == 0.0f, "This flash attention build does not support dropout."); + #endif + + // Causal is the special case where window_size_right == 0 and window_size_left < 0. + // Local is the more general case where window_size_right >= 0 or window_size_left >= 0. + params.is_causal = window_size_left < 0 && window_size_right == 0; + params.per_block_mean = per_block_mean; + if (per_block_mean) { + params.seqlen_s = seqlen_q; + } else { + params.seqlen_s = flash::BLOCK_M; // size of BLOCK_M + } + if (window_size_left < 0 && window_size_right >= 0) { window_size_left = seqlen_k; } + if (window_size_left >= 0 && window_size_right < 0) { window_size_right = seqlen_k; } + params.window_size_left = window_size_left; + params.window_size_right = window_size_right; + + #ifdef FLASHATTENTION_DISABLE_LOCAL + TORCH_CHECK(params.is_causal || (window_size_left < 0 && window_size_right < 0), + "This flash attention build does not support local attention."); + #endif + + params.is_seqlens_k_cumulative = true; + params.is_bf16 = is_bf16; + params.single_level_p_quant = single_level_p_quant; + #ifdef FLASHATTENTION_DISABLE_UNEVEN_K + TORCH_CHECK(d == d_rounded, "This flash attention build does not support headdim not being a multiple of 32."); + #endif +} + +template +void run_mha_fwd_dispatch_dtype(Flash_fwd_params ¶ms, cudaStream_t stream) { + using OType = std::conditional_t; + if (params.d == 64) { + run_mha_fwd_, 64, OType>(params, stream); + } else if (params.d == 128) { + run_mha_fwd_, 128, OType>(params, stream); + } +} + +void run_mha_fwd(Flash_fwd_params ¶ms, cudaStream_t stream, bool force_split_kernel = false) { + BOOL_SWITCH(params.is_bf16, IsBF16, ([&] { + run_mha_fwd_dispatch_dtype(params, stream); + })); +} + +std::vector +mha_fwd(at::Tensor &q, // batch_size x seqlen_q x num_heads x (head_size // 2) + const at::Tensor &k, // batch_size x seqlen_k x num_heads_k x (head_size // 2) + const at::Tensor &v, // batch_size x seqlen_k x num_heads_k x (head_size // 2) + const at::Tensor &sfq, + const at::Tensor &sfk, + const at::Tensor &sfv, + const at::Tensor &delta_s, + int unpadded_k, + c10::optional &out_, // batch_size x seqlen_q x num_heads x head_size + const float softmax_scale, + bool is_causal, + bool per_block_mean, + bool is_bf16, + bool single_level_p_quant=false // If true, use only per-row scale s_P2 (no per-block s_P1) + ) { + + auto dprops = at::cuda::getCurrentDeviceProperties(); + bool is_sm120 = dprops->major == 12 && dprops->minor == 0; + TORCH_CHECK(is_sm120, "only supports Blackwell GPUs or newer."); + + auto q_dtype = q.dtype(); + auto sfq_dtype = sfq.dtype(); + TORCH_CHECK(q_dtype == torch::kUInt8, "q dtype must be uint8"); + TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype"); + TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype"); + CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v); + + TORCH_CHECK(sfq_dtype == torch::kFloat8_e4m3fn, "q dtype must be uint8"); + TORCH_CHECK(sfk.dtype() == sfq_dtype, "query and key must have the same dtype"); + TORCH_CHECK(sfv.dtype() == sfq_dtype, "query and value must have the same dtype"); + CHECK_DEVICE(sfq); CHECK_DEVICE(sfk); CHECK_DEVICE(sfv); + + TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension"); + TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension"); + TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension"); + TORCH_CHECK(delta_s.stride(-1) == 1, "Input tensor must have contiguous last dimension"); + + TORCH_CHECK(q.is_contiguous(), "Input tensor must be contiguous"); + TORCH_CHECK(k.is_contiguous(), "Input tensor must be contiguous"); + TORCH_CHECK(v.is_contiguous(), "Input tensor must be contiguous"); + + const auto sizes = q.sizes(); + auto opts = q.options(); + const int batch_size = sizes[0]; + int seqlen_q = sizes[2]; + int num_heads = sizes[1]; + const int head_size_og = sizes[3]; + const int unpacked_head_size = head_size_og * 2; + const int seqlen_k = k.size(2); + const int num_heads_k = k.size(1); + + TORCH_CHECK(batch_size > 0, "batch size must be postive"); + TORCH_CHECK(unpacked_head_size <= 256, "FlashAttention forward only supports head dimension at most 256"); + TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query"); + TORCH_CHECK(num_heads == num_heads_k, "We do not support MQA/GQA yet"); + + TORCH_CHECK(unpacked_head_size == 64 || unpacked_head_size == 128 || unpacked_head_size == 256, "Only support head size 64, 128, and 256 for now"); + + CHECK_SHAPE(q, batch_size, num_heads, seqlen_q, head_size_og); + CHECK_SHAPE(k, batch_size, num_heads_k, seqlen_k, head_size_og); + CHECK_SHAPE(v, batch_size, num_heads_k, unpacked_head_size, seqlen_k/2); + // CHECK_SHAPE(delta_s, batch_size, num_heads, seqlen_q / 128, seqlen_k); + // CHECK_SHAPE(sfq, batch_size, seqlen_q, num_heads, unpacked_head_size); + // CHECK_SHAPE(sfk, batch_size, seqlen_k, num_heads_k, unpacked_head_size); + // CHECK_SHAPE(sfv, batch_size, unpacked_head_size, num_heads_k, seqlen_k); + TORCH_CHECK(unpacked_head_size % 8 == 0, "head_size must be a multiple of 8"); + + auto dtype = is_bf16 ? at::ScalarType::BFloat16 : at::ScalarType::Half; + at::Tensor out = torch::empty({batch_size, num_heads, seqlen_q, unpacked_head_size}, opts.dtype(dtype)); + + auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; }; + // const int head_size = round_multiple(head_size_og, 8); + // const int head_size_rounded = round_multiple(head_size, 32); + const int seqlen_q_rounded = round_multiple(seqlen_q, flash::BLOCK_M); + const int seqlen_k_rounded = round_multiple(seqlen_k, flash::BLOCK_N); + + // Otherwise the kernel will be launched from cuda:0 device + // Cast to char to avoid compiler warning about narrowing + at::cuda::CUDAGuard device_guard{(char)q.get_device()}; + + + + auto softmax_lse = torch::empty({batch_size, num_heads, seqlen_q}, opts.dtype(at::kFloat)); + at::Tensor p; + + Flash_fwd_params params; + set_params_fprop(params, + batch_size, + seqlen_q, seqlen_k, unpadded_k, + seqlen_q_rounded, seqlen_k_rounded, + num_heads, num_heads_k, + unpacked_head_size, unpacked_head_size, + q, k, v, delta_s, out, + sfq, sfk, sfv, + /*cu_seqlens_q_d=*/nullptr, + /*cu_seqlens_k_d=*/nullptr, + /*seqused_k=*/nullptr, + nullptr, + softmax_lse.data_ptr(), + /*p_dropout=*/0.f, + softmax_scale, + /*window_size_left=*/-1, + /*window_size_right=*/is_causal ? 0 : -1, + per_block_mean, + is_bf16, + single_level_p_quant + ); + // TODO: 132 sm count? + auto tile_count_semaphore = is_causal ? torch::full({1}, 132, opts.dtype(torch::kInt32)) : torch::empty({1}, opts.dtype(torch::kInt32)); + params.tile_count_semaphore = tile_count_semaphore.data_ptr(); + + if (seqlen_k > 0) { + auto stream = at::cuda::getCurrentCUDAStream().stream(); + run_mha_fwd(params, stream); + } else { + // If seqlen_k == 0, then we have an empty tensor. We need to set the output to 0. + out.zero_(); + softmax_lse.fill_(std::numeric_limits::infinity()); + } + + // at::Tensor out_padded = out; + // if (head_size_og % 8 != 0) { + // out = out.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)}); + // if (out_.has_value()) { out_.value().copy_(out); } + // } + + // return {out, q_padded, k_padded, v_padded, out_padded, softmax_lse, p}; + // cudaDeviceSynchronize(); + // auto err = cudaGetLastError(); + // printf("%s\n", cudaGetErrorString(err)); + return {out, softmax_lse}; +} + + + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.doc() = "FlashAttention"; + m.def("fwd", &mha_fwd, "Forward pass"); +} \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/blackwell/block_config.h b/fastvideo/attention/backends/sageattn/blackwell/block_config.h new file mode 100644 index 000000000..e4ecfda77 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/block_config.h @@ -0,0 +1,28 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +// Centralized block size configuration for sageattn_blackwell kernels +// Block sizes for M and N dimensions +namespace flash { + // Block size for M dimension (query sequence length) + static constexpr int BLOCK_M = 128; + + // Block size for N dimension (key/value sequence length) + static constexpr int BLOCK_N = 128; +} + diff --git a/fastvideo/attention/backends/sageattn/blackwell/block_info.h b/fastvideo/attention/backends/sageattn/blackwell/block_info.h new file mode 100644 index 000000000..bd0d45488 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/block_info.h @@ -0,0 +1,60 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * This code is based on code from FlashAttention3, https://github.com/Dao-AILab/flash-attention + * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +namespace flash { + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +template +struct BlockInfo { + + template + __device__ BlockInfo(const Params ¶ms, const int bidb) + : sum_s_q(!Varlen || params.cu_seqlens_q == nullptr ? -1 : params.cu_seqlens_q[bidb]) + , sum_s_k(!Varlen || params.cu_seqlens_k == nullptr || !params.is_seqlens_k_cumulative ? -1 : params.cu_seqlens_k[bidb]) + , actual_seqlen_q(!Varlen || params.cu_seqlens_q == nullptr ? params.seqlen_q : params.cu_seqlens_q[bidb + 1] - sum_s_q) + // If is_seqlens_k_cumulative, then seqlen_k is cu_seqlens_k[bidb + 1] - cu_seqlens_k[bidb]. + // Otherwise it's cu_seqlens_k[bidb], i.e., we use cu_seqlens_k to store the sequence lengths of K. + , seqlen_k_cache(!Varlen || params.cu_seqlens_k == nullptr ? params.seqlen_k : (params.is_seqlens_k_cumulative ? params.cu_seqlens_k[bidb + 1] - sum_s_k : params.cu_seqlens_k[bidb])) + , actual_seqlen_k(params.seqused_k ? params.seqused_k[bidb] : seqlen_k_cache + (params.knew_ptr == nullptr ? 0 : params.seqlen_knew)) + { + } + + template + __forceinline__ __device__ index_t q_offset(const index_t batch_stride, const index_t row_stride, const int bidb) const { + return sum_s_q == -1 ? bidb * batch_stride : uint32_t(sum_s_q) * row_stride; + } + + template + __forceinline__ __device__ index_t k_offset(const index_t batch_stride, const index_t row_stride, const int bidb) const { + return sum_s_k == -1 ? bidb * batch_stride : uint32_t(sum_s_k) * row_stride; + } + + const int sum_s_q; + const int sum_s_k; + const int actual_seqlen_q; + // We have to have seqlen_k_cache declared before actual_seqlen_k, otherwise actual_seqlen_k is set to 0. + const int seqlen_k_cache; + const int actual_seqlen_k; +}; + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +} // namespace flash diff --git a/fastvideo/attention/backends/sageattn/blackwell/blockscaled_layout.h b/fastvideo/attention/backends/sageattn/blackwell/blockscaled_layout.h new file mode 100644 index 000000000..6c36f3cce --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/blockscaled_layout.h @@ -0,0 +1,149 @@ +/*************************************************************************************************** + * Copyright (c) 2023 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. + * SPDX-License-Identifier: BSD-3-Clause + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are met: + * + * 1. Redistributions of source code must retain the above copyright notice, this + * list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright notice, + * this list of conditions and the following disclaimer in the documentation + * and/or other materials provided with the distribution. + * + * 3. Neither the name of the copyright holder nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" + * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR + * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER + * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, + * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + **************************************************************************************************/ + +/*! \file + \brief Blocked Scale configs specific for SM100 BlockScaled MMA +*/ + +#pragma once + +#include "cutlass/layout/matrix.h" + +#include "cute/int_tuple.hpp" +#include "cute/atom/mma_traits_sm100.hpp" + +namespace flash { + +///////////////////////////////////////////////////////////////////////////////////////////////// +using namespace cute; + +template +struct BlockScaledBasicChunk { + + using Blk_MN = _64; + using Blk_SF = _4; + + using SfAtom = Layout< Shape< Shape<_16,_4>, Shape, _4>>, + Stride, Stride< _0, _1>>>; +}; + +template +struct BlockScaledConfig { + // We are creating the SFA and SFB tensors' layouts in the collective since they always have the same layout. + // k-major order + static constexpr int SFVecSize = SFVecSize_; + static constexpr int MMA_NSF = 4; // SFVecSize, MMA_NSF + using BlkScaledChunk = BlockScaledBasicChunk; + using Blk_MN = _64; + using Blk_SF = _4; + using mnBasicBlockShape = Shape<_16,_4>; + using mnBasicBlockStride = Stride<_16,_4>; + using kBasicBlockShape = Shape, Int>; // SFVecSize, MMA_NSF + using kBasicBlockStride = Stride<_0, _1>; + using SfAtom = Layout< Shape< mnBasicBlockShape, kBasicBlockShape>, + Stride>; + + using LayoutSF = decltype(blocked_product(SfAtom{}, + make_layout( + make_shape(int32_t(0), int32_t(0), int32_t(0), int32_t(0)), + make_stride(int32_t(0), _1{}, int32_t(0), int32_t(0))))); + // A single indivisible block will hold 4 scale factors of 64 rows/columns (A/B matrix). + // 4 is chosen to make consecutive 32bits of data to have scale factors for only a single row (col). 32bits corresponds to the TMEM word size + using Blk_Elems = decltype(Blk_MN{} * Blk_SF{}); + using sSF_strideMN = decltype(prepend(Blk_Elems{}, mnBasicBlockStride{})); + + + // The following function is provided for user fill dynamic problem size to the layout_SFA. + template < class ProblemShape> + CUTE_HOST_DEVICE + static constexpr auto + tile_atom_to_shape_SFQKV(ProblemShape problem_shape) { + auto [Seqlen, Dim, HeadNum, Batch] = problem_shape; + return tile_to_shape(SfAtom{}, make_shape(Seqlen, Dim, HeadNum, Batch), Step<_2,_1,_3,_4>{}); + } + + // The following function is provided for user fill dynamic problem size to the layout_SFB. + template + CUTE_HOST_DEVICE + static constexpr auto + tile_atom_to_shape_SFVt(ProblemShape problem_shape) { + auto [Dim, Seqlen, HeadNum, Batch] = problem_shape; + return tile_to_shape(SfAtom{}, make_shape(Dim, Seqlen, HeadNum, Batch), Step<_2,_1,_3,_4>{}); + } + + template + CUTE_HOST_DEVICE + static constexpr auto + deduce_smem_layoutSFQ(TiledMma tiled_mma, TileShape_MNK tileshape_mnk) { + + using sSFQ_shapeK = decltype(prepend(make_shape(Blk_SF{}/Int{}, size<2>(TileShape_MNK{}) / Int{} / Blk_SF{}), kBasicBlockShape{})); + using sSFQ_shapeM = decltype(prepend(size<0>(TileShape_MNK{}) / Blk_MN{}, mnBasicBlockShape{})); + using sSFQ_strideM = sSF_strideMN; + using sSFQ_strideK = decltype(prepend(make_stride(Int{}, size<0>(TileShape_MNK{}) / Blk_MN{} * Blk_Elems{}), kBasicBlockStride{})); + using sSFQ_shape = decltype(make_shape(sSFQ_shapeM{}, sSFQ_shapeK{})); + using sSFQ_stride = decltype(make_stride(sSFQ_strideM{}, sSFQ_strideK{})); + using SmemLayoutAtomSFQ = decltype(make_layout(sSFQ_shape{}, sSFQ_stride{})); + return SmemLayoutAtomSFQ{}; + } + + template + CUTE_HOST_DEVICE + static constexpr auto + deduce_smem_layoutSFKV(TiledMma tiled_mma, TileShape_MNK tileshape_mnk) { + + using sSFK_shapeK = decltype(prepend(make_shape(Blk_SF{}/Int{}, size<2>(TileShape_MNK{}) / Int{} / Blk_SF{}), kBasicBlockShape{})); + using sSFK_shapeN = decltype(prepend(size<1>(TileShape_MNK{}) / Blk_MN{}, mnBasicBlockShape{})); + using sSFK_strideN = sSF_strideMN; + using sSFK_strideK = decltype(prepend(make_stride(Int{}, size<1>(TileShape_MNK{}) / Blk_MN{} * Blk_Elems{}), kBasicBlockStride{})); + using sSFK_shape = decltype(make_shape(sSFK_shapeN{}, sSFK_shapeK{})); + using sSFK_stride = decltype(make_stride(sSFK_strideN{}, sSFK_strideK{})); + using SmemLayoutAtomSFK = decltype(make_layout(sSFK_shape{}, sSFK_stride{})); + return SmemLayoutAtomSFK{}; + } + + template + CUTE_HOST_DEVICE + static constexpr auto + deduce_smem_layoutSFVt(TiledMma tiled_mma, TileShape_MNK tileshape_mnk) { + + using sSFVt_shapeK = decltype(prepend(make_shape(Blk_SF{}/Int{}, size<2>(TileShape_MNK{}) / Int{} / Blk_SF{}), kBasicBlockShape{})); + using sSFVt_shapeN = decltype(prepend(size<1>(TileShape_MNK{}) / Blk_MN{}, mnBasicBlockShape{})); + using sSFVt_strideN = sSF_strideMN; + using sSFVt_strideK = decltype(prepend(make_stride(Int{}, size<1>(TileShape_MNK{}) / Blk_MN{} * Blk_Elems{}), kBasicBlockStride{})); + using sSFVt_shape = decltype(make_shape(sSFVt_shapeN{}, sSFVt_shapeK{})); + using sSFVt_stride = decltype(make_stride(sSFVt_strideN{}, sSFVt_strideK{})); + using SmemLayoutAtomSFVt = decltype(make_layout(sSFVt_shape{}, sSFVt_stride{})); + return SmemLayoutAtomSFVt{}; + } +}; + + +} // namespace flash diff --git a/fastvideo/attention/backends/sageattn/blackwell/cute_extension.h b/fastvideo/attention/backends/sageattn/blackwell/cute_extension.h new file mode 100644 index 000000000..bcd7cf3b5 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/cute_extension.h @@ -0,0 +1,327 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include "cute/arch/mma_sm120.hpp" +#include "cute/atom/mma_traits_sm120.hpp" +#include "cute/atom/mma_atom.hpp" +#include "cutlass/cutlass.h" +#include "cutlass/float8.h" +#include "cutlass/float_subbyte.h" + +namespace cute::SM120::BLOCKSCALED { + +using cutlass::float_e2m1_t; +using cutlass::float_ue4m3_t; + +// MMA.SF 16x32x64 TN E2M1 x E2M1 with SF E4M3 +struct SM120_16x32x64_TN_VS_NVFP4 { + using DRegisters = float[16]; + using ARegisters = uint32_t[4]; + using BRegisters = uint32_t[8]; + using CRegisters = float[16]; + + static constexpr int SFBits = 32; + using RegTypeSF = cute::uint_bit_t; + + using SFARegisters = RegTypeSF[1]; + using SFBRegisters = RegTypeSF[1]; + + CUTE_HOST_DEVICE static void + fma(float & d0 , float & d1 , float & d2 , float & d3 , + float & d4 , float & d5 , float & d6 , float & d7 , + float & d8 , float & d9 , float & d10, float & d11, + float & d12, float & d13, float & d14, float & d15, + uint32_t const& a0 , uint32_t const& a1 , uint32_t const& a2 , uint32_t const& a3 , + uint32_t const& b0 , uint32_t const& b1 , uint32_t const& b2 , uint32_t const& b3 , + uint32_t const& b4 , uint32_t const& b5 , uint32_t const& b6 , uint32_t const& b7 , + float const & c0 , float const & c1 , float const & c2 , float const & c3 , + float const & c4 , float const & c5 , float const & c6 , float const & c7 , + float const & c8 , float const & c9 , float const & c10 , float const & c11, + float const & c12, float const & c13, float const & c14, float const & c15, + RegTypeSF const& sfa0, + RegTypeSF const& sfb0) + { + static constexpr uint16_t tidA = 0; + static constexpr uint16_t bidA = 0; + static constexpr uint16_t bidB = 0; + static constexpr uint16_t tidB0 = 0; + static constexpr uint16_t tidB1 = 1; + static constexpr uint16_t tidB2 = 2; + static constexpr uint16_t tidB3 = 3; + +#if defined(CUTE_ARCH_MXF4NVF4_4X_UE4M3_MMA_ENABLED) + asm volatile( + "mma.sync.aligned.kind::mxf4nvf4.block_scale.scale_vec::4X.m16n8k64.row.col.f32.e2m1.e2m1.f32.ue4m3 " + "{%0, %1, %2, %3}," + "{%4, %5, %6, %7}," + "{%8, %9}," + "{%10, %11, %12, %13}," + "{%14}," + "{%15, %16}," + "{%17}," + "{%18, %19};\n" + : "=f"(d0), "=f"(d1), "=f"(d8), "=f"(d9) + : "r"(a0), "r"(a1), "r"(a2), "r"(a3), + "r"(b0), "r"(b1), + "f"(c0), "f"(c1), "f"(c8), "f"(c9), + "r"(uint32_t(sfa0)) , "h"(bidA), "h"(tidA), + "r"(uint32_t(sfb0)) , "h"(bidB), "h"(tidB0)); + + asm volatile( + "mma.sync.aligned.kind::mxf4nvf4.block_scale.scale_vec::4X.m16n8k64.row.col.f32.e2m1.e2m1.f32.ue4m3 " + "{%0, %1, %2, %3}," + "{%4, %5, %6, %7}," + "{%8, %9}," + "{%10, %11, %12, %13}," + "{%14}," + "{%15, %16}," + "{%17}," + "{%18, %19};\n" + : "=f"(d2), "=f"(d3), "=f"(d10), "=f"(d11) + : "r"(a0), "r"(a1), "r"(a2), "r"(a3), + "r"(b2), "r"(b3), + "f"(c2), "f"(c3), "f"(c10), "f"(c11), + "r"(uint32_t(sfa0)) , "h"(bidA), "h"(tidA), + "r"(uint32_t(sfb0)) , "h"(bidB), "h"(tidB1)); + + asm volatile( + "mma.sync.aligned.kind::mxf4nvf4.block_scale.scale_vec::4X.m16n8k64.row.col.f32.e2m1.e2m1.f32.ue4m3 " + "{%0, %1, %2, %3}," + "{%4, %5, %6, %7}," + "{%8, %9}," + "{%10, %11, %12, %13}," + "{%14}," + "{%15, %16}," + "{%17}," + "{%18, %19};\n" + : "=f"(d4), "=f"(d5), "=f"(d12), "=f"(d13) + : "r"(a0), "r"(a1), "r"(a2), "r"(a3), + "r"(b4), "r"(b5), + "f"(c4), "f"(c5), "f"(c12), "f"(c13), + "r"(uint32_t(sfa0)) , "h"(bidA), "h"(tidA), + "r"(uint32_t(sfb0)) , "h"(bidB), "h"(tidB2)); + + asm volatile( + "mma.sync.aligned.kind::mxf4nvf4.block_scale.scale_vec::4X.m16n8k64.row.col.f32.e2m1.e2m1.f32.ue4m3 " + "{%0, %1, %2, %3}," + "{%4, %5, %6, %7}," + "{%8, %9}," + "{%10, %11, %12, %13}," + "{%14}," + "{%15, %16}," + "{%17}," + "{%18, %19};\n" + : "=f"(d6), "=f"(d7), "=f"(d14), "=f"(d15) + : "r"(a0), "r"(a1), "r"(a2), "r"(a3), + "r"(b6), "r"(b7), + "f"(c6), "f"(c7), "f"(c14), "f"(c15), + "r"(uint32_t(sfa0)) , "h"(bidA), "h"(tidA), + "r"(uint32_t(sfb0)) , "h"(bidB), "h"(tidB3)); +#else + CUTE_INVALID_CONTROL_PATH("Attempting to use SM120::BLOCKSCALED::SM120_16x8x64_TN_VS without CUTE_ARCH_MXF4NVF4_4X_UE4M3_MMA_ENABLED"); +#endif + + } +}; + +} // namespace cute::SM120::BLOCKSCALED + +namespace cute { + +// MMA NVFP4 16x32x64 TN +template <> +struct MMA_Traits +{ + // The MMA accepts 4-bit inputs regardless of the types for A and B + using ValTypeA = uint4_t; + using ValTypeB = uint4_t; + + using ValTypeD = float; + using ValTypeC = float; + + using ValTypeSF = cutlass::float_ue4m3_t; + constexpr static int SFVecSize = 16; + + using Shape_MNK = Shape<_16,_32,_64>; + using ThrID = Layout<_32>; + + // (T32,V32) -> (M16,K64) + using ALayout = Layout,Shape < _8,_2, _2>>, + Stride,Stride<_16,_8,_512>>>; + // (T32,V64) -> (N32,K64) + using BLayout = Layout,Shape <_8, _2, _4>>, + Stride,Stride<_32,_1024, _8>>>; + // (T32,V64) -> (M16,K64) + using SFALayout = Layout,_64>, + Stride,_16>>; + // (T32,V64) -> (N32,K64) + using SFBLayout = Layout,_64>, + Stride, _32>>; + // (T32,V16) -> (M16,N32) + using CLayout = Layout,Shape < Shape<_2, _4>,_2>>, + Stride,Stride,_8>>>; +}; + + +template +CUTE_HOST_DEVICE constexpr +auto +thrfrg_SFA(SFATensor&& sfatensor, TiledMMA& mma) +{ + CUTE_STATIC_ASSERT_V(rank(sfatensor) >= Int<2>{}); + + using AtomShape_MNK = typename Atom::Shape_MNK; + using AtomLayoutSFA_TV = typename Atom::Traits::SFALayout; + + auto permutation_mnk = TiledPerm{}; + auto thr_layout_vmnk = mma.get_thr_layout_vmnk(); + + // Reorder the tensor for the TiledAtom + auto t_tile = make_tile(get<0>(permutation_mnk), + get<2>(permutation_mnk)); + auto t_tensor = logical_divide(sfatensor, t_tile); // (PermM,PermK) + + // Tile the tensor for the Atom + auto a_tile = make_tile(make_layout(size<0>(AtomShape_MNK{})), + make_layout(size<2>(AtomShape_MNK{}))); + auto a_tensor = zipped_divide(t_tensor, a_tile); // ((AtomM,AtomK),(RestM,RestK)) + + // Transform the Atom mode from (M,K) to (Thr,Val) + auto tv_tensor = a_tensor.compose(AtomLayoutSFA_TV{},_); // ((ThrV,FrgV),(RestM,RestK)) + + // Tile the tensor for the Thread + auto thr_tile = make_tile(_, + make_tile(make_layout(size<1>(thr_layout_vmnk)), + make_layout(size<3>(thr_layout_vmnk)))); + auto thr_tensor = zipped_divide(tv_tensor, thr_tile); // ((ThrV,(ThrM,ThrK)),(FrgV,(RestM,RestK))) + + return thr_tensor; +} + +template +CUTE_HOST_DEVICE constexpr +auto +thrfrg_SFB(SFBTensor&& sfbtensor, TiledMMA& mma) +{ + CUTE_STATIC_ASSERT_V(rank(sfbtensor) >= Int<2>{}); + + using AtomShape_MNK = typename Atom::Shape_MNK; + using AtomLayoutSFB_TV = typename Atom::Traits::SFBLayout; + + auto permutation_mnk = TiledPerm{}; + auto thr_layout_vmnk = mma.get_thr_layout_vmnk(); + + // Reorder the tensor for the TiledAtom + auto t_tile = make_tile(get<1>(permutation_mnk), + get<2>(permutation_mnk)); + auto t_tensor = logical_divide(sfbtensor, t_tile); // (PermN,PermK) + + // Tile the tensor for the Atom + auto a_tile = make_tile(make_layout(size<1>(AtomShape_MNK{})), + make_layout(size<2>(AtomShape_MNK{}))); + auto a_tensor = zipped_divide(t_tensor, a_tile); // ((AtomN,AtomK),(RestN,RestK)) + + // Transform the Atom mode from (M,K) to (Thr,Val) + auto tv_tensor = a_tensor.compose(AtomLayoutSFB_TV{},_); // ((ThrV,FrgV),(RestN,RestK)) + + // Tile the tensor for the Thread + auto thr_tile = make_tile(_, + make_tile(make_layout(size<2>(thr_layout_vmnk)), + make_layout(size<3>(thr_layout_vmnk)))); + auto thr_tensor = zipped_divide(tv_tensor, thr_tile); // ((ThrV,(ThrN,ThrK)),(FrgV,(RestN,RestK))) + return thr_tensor; +} + +template +CUTE_HOST_DEVICE constexpr +auto +partition_SFA(SFATensor&& sfatensor, ThrMma& thread_mma) { + auto thr_tensor = make_tensor(static_cast(sfatensor).data(), thrfrg_SFA(sfatensor.layout(),thread_mma)); + auto thr_vmnk = thread_mma.thr_vmnk_; + auto thr_vmk = make_coord(get<0>(thr_vmnk), make_coord(get<1>(thr_vmnk), get<3>(thr_vmnk))); + return thr_tensor(thr_vmk, make_coord(_, repeat(thr_tensor)>(_))); +} + +template +CUTE_HOST_DEVICE constexpr +auto +partition_fragment_SFA(SFATensor&& sfatensor, ThrMma& thread_mma) { + using ValTypeSF = typename ThrMma::Atom::Traits::ValTypeSF; + return make_fragment_like(partition_SFA(sfatensor, thread_mma)); +} + +template +CUTE_HOST_DEVICE constexpr +auto +partition_SFB(SFBTensor&& sfbtensor, ThrMma& thread_mma) { + auto thr_tensor = make_tensor(static_cast(sfbtensor).data(), thrfrg_SFB(sfbtensor.layout(),thread_mma)); + auto thr_vmnk = thread_mma.thr_vmnk_; + auto thr_vnk = make_coord(get<0>(thr_vmnk), make_coord(get<2>(thr_vmnk), get<3>(thr_vmnk))); + return thr_tensor(thr_vnk, make_coord(_, repeat(thr_tensor)>(_))); +} + +template +CUTE_HOST_DEVICE constexpr +auto +partition_fragment_SFB(SFBTensor&& sfbtensor, ThrMma& thread_mma) { + using ValTypeSF = typename ThrMma::Atom::Traits::ValTypeSF; + return make_fragment_like(partition_SFB(sfbtensor, thread_mma)); +} + +template +CUTE_HOST_DEVICE constexpr +auto +get_layoutSFA_TV(TiledMma& mma) +{ + // (M,K) -> (M,K) + auto tile_shape_mnk = tile_shape(mma); + auto ref_A = make_layout(make_shape(size<0>(tile_shape_mnk), size<2>(tile_shape_mnk))); + auto thr_layout_vmnk = mma.get_thr_layout_vmnk(); + + // (ThrV,(ThrM,ThrK)) -> (ThrV,(ThrM,ThrN,ThrK)) + auto atile = make_tile(_, + make_tile(make_layout(make_shape (size<1>(thr_layout_vmnk), size<2>(thr_layout_vmnk)), + make_stride( Int<1>{} , Int<0>{} )), + _)); + + // thr_idx -> (ThrV,ThrM,ThrN,ThrK) + auto thridx_2_thrid = right_inverse(thr_layout_vmnk); + // (thr_idx,val) -> (M,K) + return thrfrg_SFA(ref_A, mma).compose(atile, _).compose(thridx_2_thrid, _); +} + +template +CUTE_HOST_DEVICE constexpr +auto +get_layoutSFB_TV(TiledMma& mma) +{ + // (N,K) -> (N,K) + auto tile_shape_mnk = tile_shape(mma); + auto ref_B = make_layout(make_shape(size<1>(tile_shape_mnk), size<2>(tile_shape_mnk))); + auto thr_layout_vmnk = mma.get_thr_layout_vmnk(); + + // (ThrV,(ThrM,ThrK)) -> (ThrV,(ThrM,ThrN,ThrK)) + auto btile = make_tile(_, + make_tile(make_layout(make_shape (size<1>(thr_layout_vmnk), size<2>(thr_layout_vmnk)), + make_stride( Int<0>{} , Int<1>{} )), + _)); + + // thr_idx -> (ThrV,ThrM,ThrN,ThrK) + auto thridx_2_thrid = right_inverse(thr_layout_vmnk); + // (thr_idx,val) -> (M,K) + return thrfrg_SFB(ref_B, mma).compose(btile, _).compose(thridx_2_thrid, _); +} + +} // namespace cute \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/blackwell/epilogue_tma_ws.h b/fastvideo/attention/backends/sageattn/blackwell/epilogue_tma_ws.h new file mode 100644 index 000000000..85e82a053 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/epilogue_tma_ws.h @@ -0,0 +1,222 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#include +#include "cute/tensor.hpp" + +#include "cutlass/gemm/collective/collective_builder.hpp" +#include "named_barrier.h" +#include "utils.h" + +namespace flash { + +using namespace cute; + +template +struct CollectiveEpilogueFwd{ + + using Element = typename Ktraits::ElementOut; + static constexpr int kBlockM = Ktraits::kBlockM; + static constexpr int kBlockN = Ktraits::kBlockN; + static constexpr int kHeadDim = Ktraits::kHeadDim; + using TileShape_MNK = Shape, Int, Int>; + static constexpr int kNWarps = Ktraits::kNWarps; + static constexpr int kNThreads = kNWarps * cutlass::NumThreadsPerWarp; + static constexpr int NumMmaThreads = kNThreads - cutlass::NumThreadsPerWarpGroup; + + using GmemTiledCopyOTMA = cute::SM90_TMA_STORE; + + // These are for storing the output tensor without TMA (e.g., for setting output to zero) + static constexpr int kGmemElemsPerLoad = sizeof(cute::uint128_t) / sizeof(Element); + static_assert(kHeadDim % kGmemElemsPerLoad == 0, "kHeadDim must be a multiple of kGmemElemsPerLoad"); + static constexpr int kGmemThreadsPerRow = kHeadDim / kGmemElemsPerLoad; + static_assert(NumMmaThreads % kGmemThreadsPerRow == 0, "NumMmaThreads must be a multiple of kGmemThreadsPerRow"); + using GmemLayoutAtom = Layout, Int>, + Stride, _1>>; + using GmemTiledCopyO = decltype( + make_tiled_copy(Copy_Atom{}, + GmemLayoutAtom{}, + Layout>>{})); // Val layout, 8 or 16 vals per store + + using SmemLayoutO = typename Ktraits::SmemLayoutO; + + using SmemCopyAtomO = Copy_Atom; + using SharedStorage = cute::array_aligned>; + + using ShapeO = cute::Shape; // (seqlen_q, d, head, batch) + using StrideO = cute::Stride; + using StrideLSE = cute::Stride<_1, int64_t, int64_t>; // (seqlen_q, head, batch) + + using TMA_O = decltype(make_tma_copy( + GmemTiledCopyOTMA{}, + make_tensor(make_gmem_ptr(static_cast(nullptr)), repeat_like(StrideO{}, int32_t(0)), StrideO{}), + SmemLayoutO{}, + select<0, 2>(TileShape_MNK{}), + _1{})); // no mcast for O + + // Host side kernel arguments + struct Arguments { + Element* ptr_O; + ShapeO const shape_O; + StrideO const stride_O; + float* ptr_LSE; + StrideLSE const stride_LSE; + }; + + // Device side kernel params + struct Params { + Element* ptr_O; + ShapeO const shape_O; + StrideO const stride_O; + float* ptr_LSE; + StrideLSE const stride_LSE; + TMA_O tma_store_O; + }; + + static Params + to_underlying_arguments(Arguments const& args) { + Tensor mO = make_tensor(make_gmem_ptr(args.ptr_O), args.shape_O, args.stride_O); + TMA_O tma_store_O = make_tma_copy( + GmemTiledCopyOTMA{}, + mO, + SmemLayoutO{}, + select<0, 2>(TileShape_MNK{}), + _1{}); // no mcast for O + return {args.ptr_O, args.shape_O, args.stride_O, args.ptr_LSE, args.stride_LSE, tma_store_O}; + } + + /// Issue Tma Descriptor Prefetch -- ideally from a single thread for best performance + CUTLASS_DEVICE + static void prefetch_tma_descriptors(Params const& epilogue_params) { + cute::prefetch_tma_descriptor(epilogue_params.tma_store_O.get_tma_descriptor()); + } + + template + CUTLASS_DEVICE void + mma_store( + SharedStorage& shared_storage, + TiledMma tiled_mma, + FrgTensorO const& tOrO, + int thread_idx + ){ + Tensor sO = cute::as_position_independent_swizzle_tensor(make_tensor(make_smem_ptr(shared_storage.smem_o.begin()), SmemLayoutO{})); + auto smem_tiled_copy_O = make_tiled_copy_C(SmemCopyAtomO{}, tiled_mma); + auto smem_thr_copy_O = smem_tiled_copy_O.get_thread_slice(thread_idx); + constexpr int numel = decltype(size(tOrO))::value; + cutlass::NumericArrayConverter convert_op; + // HACK: this requires tensor to be "contiguous" + auto frag = convert_op(*reinterpret_cast *>(tOrO.data())); + auto tOrO_out = make_tensor(make_rmem_ptr(&frag), tOrO.layout()); + Tensor taccOrO = smem_thr_copy_O.retile_S(tOrO_out); // ((Atom,AtomNum), MMA_M, MMA_N) + Tensor taccOsO = smem_thr_copy_O.partition_D(sO); // ((Atom,AtomNum),PIPE_M,PIPE_N) + cute::copy(smem_tiled_copy_O, taccOrO, taccOsO); + cutlass::arch::fence_view_async_shared(); // ensure smem writes are visible to TMA + } + + template + CUTLASS_DEVICE void + tma_store( + SharedStorage& shared_storage, + Params const& epilogue_params, + WorkTileInfo work_tile_info, + SchedulerParams const& scheduler_params, + int thread_idx + ) { + auto [m_block, bidh, bidb] = work_tile_info.get_block_coord(scheduler_params); + Tensor sO = cute::as_position_independent_swizzle_tensor(make_tensor(make_smem_ptr(shared_storage.smem_o.begin()), SmemLayoutO{})); + Tensor mO = epilogue_params.tma_store_O.get_tma_tensor(epilogue_params.shape_O); + Tensor gO = local_tile(mO(_, _, bidh, bidb), select<0, 2>(TileShape_MNK{}), make_coord(m_block, _0{})); // (M, K) + auto block_tma_O = epilogue_params.tma_store_O.get_slice(_0{}); + Tensor tOgO = block_tma_O.partition_D(gO); // (TMA, TMA_M, TMA_K) + Tensor tOsO = block_tma_O.partition_S(sO); // (TMA, TMA_M, TMA_K) + + // auto shape_LSE = select<0, 2, 3>(epilogue_params.shape_O); + // Tensor mLSE = make_tensor(make_gmem_ptr(epilogue_params.ptr_LSE), shape_LSE, epilogue_params.stride_LSE); + // Tensor gLSE = local_tile(mLSE(_, bidh, bidb), Shape>{}, make_coord(m_block)); + + // Tensor caccO = cute::make_identity_tensor(select<0, 2>(TileShape_MNK{})); + // auto thread_mma = tiled_mma.get_thread_slice(thread_idx); + // Tensor taccOcO = thread_mma.partition_C(caccO); // (MMA,MMA_M,MMA_K) + // static_assert(decltype(size<0, 0>(taccOcO))::value == 2); + // static_assert(decltype(size<0, 1>(taccOcO))::value == 2); + // // // // taccOcO has shape ((2, 2, V), MMA_M, MMA_K), we only take only the row indices. + // Tensor taccOcO_row = taccOcO(make_coord(_0{}, _), _, _0{}); + // CUTE_STATIC_ASSERT_V(size(lse) == size(taccOcO_row)); // MMA_M + // if (get<1>(taccOcO_row(_0{})) == 0) { + // #pragma unroll + // for (int mi = 0; mi < size(lse); ++mi) { + // const int row = get<0>(taccOcO_row(mi)); + // if (row < get<0>(shape_LSE) - m_block * kBlockM) { gLSE(row) = lse(mi); } + // } + // } + + // if (cutlass::canonical_warp_idx_sync() == kNWarps - 1) { + // cutlass::arch::NamedBarrier::sync(NumMmaThreads + cutlass::NumThreadsPerWarp, + // static_cast(FP4NamedBarriers::EpilogueBarrier)); + // int const lane_predicate = cute::elect_one_sync(); + // if (lane_predicate) { + // cute::copy(epilogue_params.tma_store_O, tOsO, tOgO); + // tma_store_arrive(); + // } + // } + cute::copy(epilogue_params.tma_store_O, tOsO, tOgO); + tma_store_arrive(); + } + + CUTLASS_DEVICE void + store_tail() { + tma_store_wait<0>(); + } + + // Write 0 to output and -inf to LSE + CUTLASS_DEVICE void + store_zero( + Params const& epilogue_params, + int thread_idx, + cute::tuple const& block_coord + ) { + auto [m_block, bidh, bidb] = block_coord; + Tensor mO = make_tensor(make_gmem_ptr(epilogue_params.ptr_O), epilogue_params.shape_O, epilogue_params.stride_O); + Tensor gO = local_tile(mO(_, _, bidh, bidb), select<0, 2>(TileShape_MNK{}), make_coord(m_block, _0{})); // (M, K) + auto shape_LSE = select<0, 2, 3>(epilogue_params.shape_O); + Tensor mLSE = make_tensor(make_gmem_ptr(epilogue_params.ptr_LSE), shape_LSE, epilogue_params.stride_LSE); + Tensor gLSE = local_tile(mLSE(_, bidh, bidb), Shape>{}, make_coord(m_block)); + + GmemTiledCopyO gmem_tiled_copy_O; + auto gmem_thr_copy_O = gmem_tiled_copy_O.get_thread_slice(thread_idx); + Tensor tOgO = gmem_thr_copy_O.partition_D(gO); + Tensor tOrO = make_fragment_like(tOgO); + clear(tOrO); + // Construct identity layout for sO + Tensor cO = cute::make_identity_tensor(select<0, 2>(TileShape_MNK{})); // (BLK_M,BLK_K) -> (blk_m,blk_k) + // Repeat the partitioning with identity layouts + Tensor tOcO = gmem_thr_copy_O.partition_D(cO); + Tensor tOpO = make_tensor(make_shape(size<2>(tOgO))); + #pragma unroll + for (int k = 0; k < size(tOpO); ++k) { tOpO(k) = get<1>(tOcO(_0{}, _0{}, k)) < get<1>(epilogue_params.shape_O); } + // Clear_OOB_K must be false since we don't want to write zeros to gmem + flash::copy( + gmem_tiled_copy_O, tOrO, tOgO, tOcO, tOpO, get<0>(epilogue_params.shape_O) - m_block * kBlockM + ); + static_assert(kBlockM <= NumMmaThreads); + if (thread_idx < get<0>(shape_LSE) - m_block * kBlockM) { gLSE(thread_idx) = INFINITY; } + } + +}; + +} // namespace flash diff --git a/fastvideo/attention/backends/sageattn/blackwell/kernel_traits.h b/fastvideo/attention/backends/sageattn/blackwell/kernel_traits.h new file mode 100644 index 000000000..5035f4b58 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/kernel_traits.h @@ -0,0 +1,202 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#include "cute/algorithm/copy.hpp" +#include "cute/atom/mma_atom.hpp" +#include "cutlass/gemm/collective/collective_builder.hpp" +#include "cute/tensor.hpp" +#include "cutlass/cutlass.h" +#include "cutlass/layout/layout.h" +#include "cutlass/numeric_types.h" +#include "cutlass/pipeline/pipeline.hpp" + +#include "blockscaled_layout.h" +#include "cute_extension.h" +#include "named_barrier.h" +using namespace cute; + +template < + int kStages, + int EpiStages, + typename Element, + typename ElementSF, + typename OutputType, + typename SmemLayoutQ, + typename SmemLayoutK, + typename SmemLayoutV, + typename SmemLayoutDS, + typename SmemLayoutO, + typename SmemLayoutSFQ, + typename SmemLayoutSFK, + typename SmemLayoutSFV +> +struct SharedStorageQKVOwithSF : cute::aligned_struct<128, _0>{ + + alignas(1024) cute::ArrayEngine> smem_q; + alignas(1024) cute::ArrayEngine> smem_k; + cute::ArrayEngine> smem_SFQ; + cute::ArrayEngine> smem_SFK; + cute::ArrayEngine> smem_SFV; + alignas(1024) cute::ArrayEngine> smem_ds; + alignas(1024) cute::ArrayEngine> smem_v; + alignas(1024) cute::ArrayEngine> smem_o; + + struct { + alignas(16) typename cutlass::PipelineTmaAsync<1>::SharedStorage pipeline_q; + alignas(16) typename cutlass::PipelineTmaAsync::SharedStorage pipeline_k; + alignas(16) typename cutlass::PipelineTmaAsync::SharedStorage pipeline_v; + alignas(16) typename flash::OrderedSequenceBarrierVarGroupSize::SharedStorage barrier_o; + int tile_count_semaphore; + }; + }; + +template < + int kHeadDim_, + int kBlockM_, + int kBlockN_, + int kStages_, + int kClusterM_, + bool BlockMean_, + typename ElementPairType_ = cutlass::nv_float4_t, + typename ElementOut_ = cutlass::bfloat16_t +> +struct Flash_fwd_kernel_traits { + static constexpr int kBlockM = kBlockM_; + static constexpr int kBlockN = kBlockN_; + static constexpr int kHeadDim = kHeadDim_; + static constexpr bool BlockMean = BlockMean_; + static constexpr bool SmoothQ = true; + static_assert(kHeadDim % 32 == 0); + static_assert(kBlockM == 64 || kBlockM == 128); + static constexpr int kNWarps = kBlockM == 128 ? 12 : 8; + static constexpr int kNThreads = kNWarps * cutlass::NumThreadsPerWarp; + static constexpr int kClusterM = kClusterM_; + static constexpr int kStages = kStages_; + static constexpr int EpiStages = 1; + static constexpr int NumSFQK = kHeadDim / 16; + static constexpr int NumSFPV = kBlockN / 16; + using ElementSF = cutlass::float_ue4m3_t; + using Element = cutlass::float_e2m1_t; + using ElementAccum = float; + using ElementOut = ElementOut_; + using index_t = int64_t; + static constexpr auto SFVectorSize = 16; + using TileShape_MNK = Shape, Int, Int>; + using ClusterShape_MNK = Shape<_1, _1, _1>; + using PermTileM = decltype(cute::min(size<0>(TileShape_MNK{}), _128{})); + using PermTileN = _32; + using PermTileK = Int; + + using ElementQMma = decltype(cutlass::gemm::collective::detail::sm1xx_kernel_input_element_to_mma_input_element()); + using ElementKMma = decltype(cutlass::gemm::collective::detail::sm1xx_kernel_input_element_to_mma_input_element()); + + using AtomLayoutMNK = std::conditional_t>, + Layout> + >; + using TiledMmaQK = decltype(cute::make_tiled_mma( + cute::SM120::BLOCKSCALED::SM120_16x32x64_TN_VS_NVFP4{}, + AtomLayoutMNK{}, + Tile{} + )); + + using TiledMmaPV = decltype(cute::make_tiled_mma( + cute::SM120::BLOCKSCALED::SM120_16x32x64_TN_VS_NVFP4{}, + AtomLayoutMNK{}, + Tile{} + )); + + static constexpr int MMA_NSF = size<2>(typename TiledMmaQK::AtomShape_MNK{}) / SFVectorSize; + + using GmemTiledCopy = SM90_TMA_LOAD; + using GmemTiledCopySF = SM90_TMA_LOAD; + + using SmemLayoutAtomQ = decltype(cutlass::gemm::collective::detail::sm120_rr_smem_selector(TileShape_MNK{}))>()); + using SmemLayoutAtomK = decltype(cutlass::gemm::collective::detail::sm120_rr_smem_selector(TileShape_MNK{}))>()); + using SmemLayoutAtomV = decltype(cutlass::gemm::collective::detail::sm120_rr_smem_selector(TileShape_MNK{}))>()); + using SmemLayoutAtomVt = decltype(cutlass::gemm::collective::detail::sm120_rr_smem_selector(TileShape_MNK{}))>()); + using SmemLayoutQ = decltype(tile_to_shape(SmemLayoutAtomQ{}, select<0, 2>(TileShape_MNK{}))); + using SmemLayoutK = + decltype(tile_to_shape(SmemLayoutAtomK{}, + make_shape(shape<1>(TileShape_MNK{}), shape<2>(TileShape_MNK{}), Int{}))); + using SmemLayoutV = + decltype(tile_to_shape(SmemLayoutAtomV{}, + make_shape(shape<1>(TileShape_MNK{}), shape<2>(TileShape_MNK{}), Int{}))); + using SmemLayoutVt = + decltype(tile_to_shape(SmemLayoutAtomVt{}, + make_shape(shape<2>(TileShape_MNK{}), shape<1>(TileShape_MNK{}), Int{}))); + using SmemLayoutAtomDS = Layout, Int>, Stride<_0, _1>>; + using SmemLayoutDS = + decltype(tile_to_shape(SmemLayoutAtomDS{}, + make_shape(shape<0>(TileShape_MNK{}), shape<1>(TileShape_MNK{}), Int{}))); + + using SmemCopyAtomQ = Copy_Atom; + using SmemCopyAtomKV = Copy_Atom; + using SmemCopyAtomSF = Copy_Atom, ElementSF>; + using SmemCopyAtomDS = Copy_Atom, float>; + + using BlkScaledConfig = flash::BlockScaledConfig; + using LayoutSF = typename BlkScaledConfig::LayoutSF; + using SfAtom = typename BlkScaledConfig::SfAtom; + using SmemLayoutAtomSFQ = decltype(BlkScaledConfig::deduce_smem_layoutSFQ(TiledMmaQK{}, TileShape_MNK{})); + using SmemLayoutAtomSFK = decltype(BlkScaledConfig::deduce_smem_layoutSFKV(TiledMmaQK{}, TileShape_MNK{})); + using SmemLayoutAtomSFV = decltype(BlkScaledConfig::deduce_smem_layoutSFKV(TiledMmaPV{}, TileShape_MNK{})); + using SmemLayoutAtomSFVt = decltype(BlkScaledConfig::deduce_smem_layoutSFVt(TiledMmaPV{}, Shape, Int, Int>{})); + using LayoutSFP = decltype( + make_layout( + make_shape(make_shape(_16{}, _4{}), _1{}, Int{}), + make_stride(make_stride(_0{}, _1{}), _0{}, _4{}) + ) + ); + using LayoutP = decltype( + make_layout( + make_shape(make_shape(_8{}, _2{}, _2{}), _1{}, Int{}), + make_stride(make_stride(_1{}, _8{}, _16{}), _0{}, _32{}) + ) + ); + using SmemLayoutSFQ = decltype(make_layout( + shape(SmemLayoutAtomSFQ{}), + stride(SmemLayoutAtomSFQ{}) + )); + using SmemLayoutSFK = decltype(make_layout( + append(shape(SmemLayoutAtomSFK{}), Int{}), + append(stride(SmemLayoutAtomSFK{}), size(filter_zeros(SmemLayoutAtomSFK{}))) + )); + using SmemLayoutSFV = decltype(make_layout( + append(shape(SmemLayoutAtomSFV{}), Int{}), + append(stride(SmemLayoutAtomSFV{}), size(filter_zeros(SmemLayoutAtomSFV{}))) + )); + using SmemLayoutSFVt = decltype(make_layout( + append(shape(SmemLayoutAtomSFVt{}), Int{}), + append(stride(SmemLayoutAtomSFVt{}), size(filter_zeros(SmemLayoutAtomSFVt{}))) + )); + + using SmemLayoutAtomO = decltype(cutlass::gemm::collective::detail::ss_smem_selector(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>()); + using SmemLayoutO = decltype(tile_to_shape(SmemLayoutAtomO{}, select<0, 2>(TileShape_MNK{}), Step<_1, _2>{})); + using SharedStorage = SharedStorageQKVOwithSF; + using MainloopPipeline = typename cutlass::PipelineTmaAsync; + using PipelineState = typename cutlass::PipelineState; + using MainloopPipelineQ = cutlass::PipelineTmaAsync<1>; + using PipelineParamsQ = typename MainloopPipelineQ::Params; + using PipelineStateQ = typename cutlass::PipelineState<1>; + using EpilogueBarrier = typename flash::OrderedSequenceBarrierVarGroupSize; +}; + diff --git a/fastvideo/attention/backends/sageattn/blackwell/kernel_ws.h b/fastvideo/attention/backends/sageattn/blackwell/kernel_ws.h new file mode 100644 index 000000000..e4c8d2f8c --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/kernel_ws.h @@ -0,0 +1,203 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#include "cute/tensor.hpp" + +#include +#include +#include +#include +#include +#include "cutlass/pipeline/pipeline.hpp" + +#include "params.h" +#include "utils.h" +#include "tile_scheduler.h" +#include "mainloop_tma_ws.h" +#include "epilogue_tma_ws.h" +#include "named_barrier.h" +#include "softmax_fused.h" + +namespace flash { + +using namespace cute; + +template +__global__ void __launch_bounds__(Ktraits::kNWarps * cutlass::NumThreadsPerWarp, 1) + compute_attn_ws(CUTE_GRID_CONSTANT Flash_fwd_params const params, + CUTE_GRID_CONSTANT typename CollectiveMainloopFwd::Params const mainloop_params, + CUTE_GRID_CONSTANT typename CollectiveEpilogueFwd::Params const epilogue_params, + CUTE_GRID_CONSTANT typename TileScheduler::Params const scheduler_params + ) { + + using Element = typename Ktraits::Element; + using ElementAccum = typename Ktraits::ElementAccum; + using SoftType = ElementAccum; + using TileShape_MNK = typename Ktraits::TileShape_MNK; + using ClusterShape = typename Ktraits::ClusterShape_MNK; + + static constexpr int NumMmaThreads = size(typename Ktraits::TiledMmaQK{}); + static constexpr int NumCopyThreads = cutlass::NumThreadsPerWarpGroup; + static constexpr int kBlockM = Ktraits::kBlockM; + + using CollectiveMainloop = CollectiveMainloopFwd; + using CollectiveEpilogue = CollectiveEpilogueFwd; + + using MainloopPipeline = typename Ktraits::MainloopPipeline; + using PipelineParams = typename MainloopPipeline::Params; + using PipelineState = typename MainloopPipeline::PipelineState; + using MainloopPipelineQ = typename Ktraits::MainloopPipelineQ; + using PipelineParamsQ = typename Ktraits::PipelineParamsQ; + using PipelineStateQ = typename Ktraits::PipelineStateQ; + using EpilogueBarrier = typename Ktraits::EpilogueBarrier; + + + enum class WarpGroupRole { + Producer = 0, + Consumer0 = 1, + Consumer1 = 2 + }; + enum class ProducerWarpRole { + Mainloop = 0, + Epilogue = 1, + Warp2 = 2, + Warp3 = 3 + }; + + extern __shared__ char shared_memory[]; + auto &shared_storage = *reinterpret_cast(shared_memory); + + int const lane_predicate = cute::elect_one_sync(); + int const warp_idx = cutlass::canonical_warp_idx_sync(); + int warp_group_idx = cutlass::canonical_warp_group_idx(); + int const warp_group_thread_idx = threadIdx.x % cutlass::NumThreadsPerWarpGroup; + int warp_idx_in_warp_group = warp_idx % cutlass::NumWarpsPerWarpGroup; + auto warp_group_role = WarpGroupRole(warp_group_idx); + auto producer_warp_role = ProducerWarpRole(warp_idx_in_warp_group); + + // Issue Tma Descriptor Prefetch from a single thread + if (warp_idx == 0 && lane_predicate) { + CollectiveMainloop::prefetch_tma_descriptors(mainloop_params); + CollectiveEpilogue::prefetch_tma_descriptors(epilogue_params); + } + + // Obtain warp index + + PipelineParams pipeline_params_v; + pipeline_params_v.transaction_bytes = CollectiveMainloop::TmaTransactionBytesV; + pipeline_params_v.role = warp_group_role == WarpGroupRole::Producer + ? MainloopPipeline::ThreadCategory::Producer + : MainloopPipeline::ThreadCategory::Consumer; + pipeline_params_v.is_leader = warp_group_thread_idx == 0; + pipeline_params_v.num_consumers = NumMmaThreads; + + PipelineParams pipeline_params_k; + pipeline_params_k.transaction_bytes = CollectiveMainloop::TmaTransactionBytesK; + pipeline_params_k.role = warp_group_role == WarpGroupRole::Producer + ? MainloopPipeline::ThreadCategory::Producer + : MainloopPipeline::ThreadCategory::Consumer; + pipeline_params_k.is_leader = warp_group_thread_idx == 0; + pipeline_params_k.num_consumers = NumMmaThreads; + + PipelineParamsQ pipeline_params_q; + pipeline_params_q.transaction_bytes = CollectiveMainloop::TmaTransactionBytesQ; + pipeline_params_q.role = warp_group_role == WarpGroupRole::Producer + ? MainloopPipelineQ::ThreadCategory::Producer + : MainloopPipelineQ::ThreadCategory::Consumer; + pipeline_params_q.is_leader = warp_group_thread_idx == 0; + pipeline_params_q.num_consumers = NumMmaThreads; + + // We're counting on pipeline_k to call cutlass::arch::fence_barrier_init(); + MainloopPipelineQ pipeline_q(shared_storage.pipeline_q, pipeline_params_q, ClusterShape{}); + MainloopPipeline pipeline_k(shared_storage.pipeline_k, pipeline_params_k, ClusterShape{}); + MainloopPipeline pipeline_v(shared_storage.pipeline_v, pipeline_params_v, ClusterShape{}); + + uint32_t epilogue_barrier_group_size_list[2] = {cutlass::NumThreadsPerWarp, NumMmaThreads}; + typename EpilogueBarrier::Params params_epilogue_barrier; + params_epilogue_barrier.group_id = (warp_group_role == WarpGroupRole::Producer); + params_epilogue_barrier.group_size_list = epilogue_barrier_group_size_list; + EpilogueBarrier barrier_o(shared_storage.barrier_o, params_epilogue_barrier); + + CollectiveMainloop collective_mainloop; + CollectiveEpilogue collective_epilogue; + __syncthreads(); + + if (warp_group_role == WarpGroupRole::Producer) { + cutlass::arch::warpgroup_reg_dealloc<24>(); + TileScheduler scheduler; + + if (producer_warp_role == ProducerWarpRole::Mainloop) { // Load Q, K, V + PipelineStateQ smem_pipe_write_q = cutlass::make_producer_start_state(); + PipelineState smem_pipe_write_k = cutlass::make_producer_start_state(); + PipelineState smem_pipe_write_v = cutlass::make_producer_start_state(); + + int work_idx = 0; + for (auto work_tile_info = scheduler.get_initial_work(); work_tile_info.is_valid(scheduler_params); work_tile_info = scheduler.get_next_work(scheduler_params, work_tile_info)) { + int tile_count_semaphore = 0; + collective_mainloop.load(mainloop_params, scheduler_params, + pipeline_q, pipeline_k, pipeline_v, + smem_pipe_write_q, smem_pipe_write_k, smem_pipe_write_v, + shared_storage, work_tile_info, work_idx, tile_count_semaphore); + } + collective_mainloop.load_tail(pipeline_q, pipeline_k, pipeline_v, + smem_pipe_write_q, smem_pipe_write_k, smem_pipe_write_v); + } else if (producer_warp_role == ProducerWarpRole::Epilogue) { + for (auto work_tile_info = scheduler.get_initial_work(); work_tile_info.is_valid(scheduler_params); work_tile_info = scheduler.get_next_work(scheduler_params, work_tile_info)) { + barrier_o.wait(); + collective_epilogue.tma_store(shared_storage, epilogue_params, work_tile_info, scheduler_params, threadIdx.x); + collective_epilogue.store_tail(); + barrier_o.arrive(); + } + + } + } else if (warp_group_role == WarpGroupRole::Consumer0 || warp_group_role == WarpGroupRole::Consumer1) { + cutlass::arch::warpgroup_reg_alloc<232>(); + typename Ktraits::TiledMmaPV tiled_mma_pv; + TileScheduler scheduler{}; + PipelineState smem_pipe_read_k, smem_pipe_read_v; + PipelineStateQ smem_pipe_read_q; + + int work_idx = 0; + + CUTLASS_PRAGMA_NO_UNROLL + for (auto work_tile_info = scheduler.get_initial_work(); work_tile_info.is_valid(scheduler_params); work_tile_info = scheduler.get_next_work(scheduler_params, work_tile_info)) { + // Attention output (GEMM-II) accumulator. + Tensor tOrO = partition_fragment_C(tiled_mma_pv, select<0, 2>(TileShape_MNK{})); + // flash::Softmax<2 * (2 * kBlockM / NumMmaThreads)> softmax; + // Pass single_level_p_quant flag to control P quantization mode + flash::SoftmaxFused<2 * (2 * kBlockM / NumMmaThreads)> softmax_fused(params.single_level_p_quant); + auto block_coord = work_tile_info.get_block_coord(scheduler_params); + auto [m_block, bidh, bidb] = block_coord; + + int n_block_max = collective_mainloop.get_n_block_max(mainloop_params, m_block); + if (Is_causal && n_block_max <= 0) { // We exit early and write 0 to gO and -inf to gLSE. + collective_epilogue.store_zero(epilogue_params, threadIdx.x - NumCopyThreads, block_coord); + continue; + } + + collective_mainloop.mma(mainloop_params, pipeline_q, pipeline_k, pipeline_v, smem_pipe_read_q, smem_pipe_read_k, smem_pipe_read_v, + tOrO, softmax_fused, n_block_max, threadIdx.x - NumCopyThreads, work_idx, m_block, shared_storage); + barrier_o.wait(); + collective_epilogue.mma_store(shared_storage, tiled_mma_pv, tOrO, threadIdx.x - NumCopyThreads); + barrier_o.arrive(); + ++work_idx; + } + } +} + +} // namespace flash diff --git a/fastvideo/attention/backends/sageattn/blackwell/launch.h b/fastvideo/attention/backends/sageattn/blackwell/launch.h new file mode 100644 index 000000000..91398c5f4 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/launch.h @@ -0,0 +1,114 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#include + +#include "cute/tensor.hpp" + +#include "cutlass/cluster_launch.hpp" + +#include "static_switch.h" +#include "params.h" +#include "tile_scheduler.h" +#include "kernel_ws.h" +#include "kernel_traits.h" +#include "block_config.h" + + +template +void run_flash_fwd(Flash_fwd_params ¶ms, cudaStream_t stream) { + using Element = typename Kernel_traits::Element; + using ElementSF = typename Kernel_traits::ElementSF; + using ElementOut = typename Kernel_traits::ElementOut; + using TileShape_MNK = typename Kernel_traits::TileShape_MNK; + using ClusterShape = typename Kernel_traits::ClusterShape_MNK; + using CollectiveMainloop = flash::CollectiveMainloopFwd; + using CollectiveEpilogue = flash::CollectiveEpilogueFwd; + // using Scheduler = flash::SingleTileScheduler; + using Scheduler = flash::StaticPersistentTileScheduler; + typename CollectiveMainloop::Params mainloop_params = + CollectiveMainloop::to_underlying_arguments({ + static_cast(params.q_ptr), + {params.seqlen_q, params.d, params.h, params.b}, // shape_Q + {params.q_row_stride, _1{}, params.q_head_stride, params.q_batch_stride}, // stride_Q + static_cast(params.k_ptr), + {params.seqlen_k, params.d, params.h_k, params.b}, // shape_K + {params.k_row_stride, _1{}, params.k_head_stride, params.k_batch_stride}, // stride_K + {params.unpadded_seqlen_k, params.d, params.h_k, params.b}, // shape_K + static_cast(params.v_ptr), + {params.d, params.seqlen_k, params.h_k, params.b}, // shape_Vt + {params.v_row_stride, _1{}, params.v_head_stride, params.v_batch_stride}, // stride_Vt + static_cast(params.sfq_ptr), + {params.seqlen_q, params.d, params.h, params.b}, // shape_SFQ + static_cast(params.sfk_ptr), + {params.seqlen_k, params.d, params.h_k, params.b}, // shape_SFK + static_cast(params.sfv_ptr), + {params.d, params.seqlen_k, params.h_k, params.b}, // shape_SFVt + static_cast(params.delta_s_ptr), + {params.seqlen_s, params.seqlen_k, params.h_k, params.b}, + {params.ds_row_stride, _1{}, params.ds_head_stride, params.ds_batch_stride}, + params.scale_softmax_log2 + }); + typename CollectiveEpilogue::Params epilogue_params = + CollectiveEpilogue::to_underlying_arguments({ + static_cast(params.o_ptr), + {params.seqlen_q, params.d, params.h, params.b}, // shape_O + {params.o_row_stride, _1{}, params.o_head_stride, params.o_batch_stride}, // stride_O + static_cast(params.softmax_lse_ptr), + {_1{}, params.seqlen_q, params.h * params.seqlen_q}, // stride_LSE + }); + + int num_blocks_m = cutlass::ceil_div(params.seqlen_q, Kernel_traits::kBlockM); + num_blocks_m = cutlass::ceil_div(num_blocks_m, size<0>(ClusterShape{})) * size<0>(ClusterShape{}); + typename Scheduler::Arguments scheduler_args = {num_blocks_m, params.h, params.b}; + typename Scheduler::Params scheduler_params = Scheduler::to_underlying_arguments(scheduler_args); + // Get the ptr to kernel function. + void *kernel; + kernel = (void *)flash::compute_attn_ws; + int smem_size = sizeof(typename Kernel_traits::SharedStorage); + if (smem_size >= 48 * 1024) { + C10_CUDA_CHECK(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size)); + } + static constexpr int ctaSize = Kernel_traits::kNWarps * 32; + params.m_block_divmod = cutlass::FastDivmod(num_blocks_m); + params.total_blocks = num_blocks_m * params.h * params.b; + dim3 grid_dims = Scheduler::get_grid_dim(scheduler_args, 170); + dim3 block_dims(ctaSize); + dim3 cluster_dims(size<0>(ClusterShape{}), size<1>(ClusterShape{}), size<2>(ClusterShape{})); + cutlass::ClusterLaunchParams launch_params{grid_dims, block_dims, cluster_dims, smem_size, stream}; + cutlass::launch_kernel_on_cluster(launch_params, kernel, params, mainloop_params, epilogue_params, scheduler_params); + + C10_CUDA_KERNEL_LAUNCH_CHECK(); +} + + +template +void run_mha_fwd_(Flash_fwd_params ¶ms, cudaStream_t stream) { + BOOL_SWITCH(params.is_causal, Is_causal, [&] { + BOOL_SWITCH(params.per_block_mean, per_block, [&] { + if constexpr (Headdim == 64 || Headdim == 128) { + run_flash_fwd< + Flash_fwd_kernel_traits, + Is_causal + >(params, stream); + } else { + static_assert(Headdim == 64 || Headdim == 128, "Unsupported Headdim"); + } + }); + }); +} diff --git a/fastvideo/attention/backends/sageattn/blackwell/mainloop_tma_ws.h b/fastvideo/attention/backends/sageattn/blackwell/mainloop_tma_ws.h new file mode 100644 index 000000000..d6ef18f4a --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/mainloop_tma_ws.h @@ -0,0 +1,908 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#include +#include +#include +#include +#include "cutlass/pipeline/pipeline.hpp" + +#include "cute/tensor.hpp" + +#include "cutlass/gemm/collective/collective_builder.hpp" + +#include "utils.h" +#include "named_barrier.h" +namespace flash { + +using namespace cute; + +template +struct CollectiveMainloopFwd { + + using Element = typename Ktraits::Element; + using ElementSF = typename Ktraits::ElementSF; + // using TMAElement = Element; + // using TMAElementSF = typename Ktraits::ElementSF; + using TileShape_MNK = typename Ktraits::TileShape_MNK; + using ClusterShape = typename Ktraits::ClusterShape_MNK; + + static constexpr int kStages = Ktraits::kStages; + static constexpr int kHeadDim = Ktraits::kHeadDim; + static constexpr int BlockMean = Ktraits::BlockMean; + using GmemTiledCopy = typename Ktraits::GmemTiledCopy; + using SmemLayoutQ = typename Ktraits::SmemLayoutQ; + using SmemLayoutK = typename Ktraits::SmemLayoutK; + using SmemLayoutV = typename Ktraits::SmemLayoutV; + using SmemLayoutVt = typename Ktraits::SmemLayoutVt; + using SmemLayoutDS = typename Ktraits::SmemLayoutDS; + using SmemLayoutAtomDS = typename Ktraits::SmemLayoutAtomDS; + using LayoutDS = decltype( + blocked_product( + SmemLayoutAtomDS{}, + make_layout( + make_shape(int32_t(0), int32_t(0), int32_t(0), int32_t(0)), + make_stride(int32_t(0), _1{}, int32_t(0), int32_t(0))) + ) + ); + using ShapeQKV = cute::Shape; // (seqlen, d, head, batch) + using StrideQKV = cute::Stride; + using ShapeSF = cute::Shape; // (seqlen, d // 16, head, batch) + using LayoutSF = typename Ktraits::LayoutSF; + using LayoutP = typename Ktraits::LayoutP; + using LayoutSFP = typename Ktraits::LayoutSFP; + using SfAtom = typename Ktraits::SfAtom; + using TMA_Q = decltype(make_tma_copy( + GmemTiledCopy{}, + make_tensor(make_gmem_ptr(static_cast(nullptr)), repeat_like(StrideQKV{}, int32_t(0)), StrideQKV{}), + SmemLayoutQ{}, + select<0, 2>(TileShape_MNK{}), + _1{})); + + using TMA_KV = decltype(make_tma_copy( + GmemTiledCopy{}, + make_tensor(make_gmem_ptr(static_cast(nullptr)), repeat_like(StrideQKV{}, int32_t(0)), StrideQKV{}), + take<0, 2>(SmemLayoutK{}), + select<1, 2>(TileShape_MNK{}), + _1{})); + + using TMA_Vt = decltype(make_tma_copy( + GmemTiledCopy{}, + make_tensor(make_gmem_ptr(static_cast(nullptr)), repeat_like(StrideQKV{}, int32_t(0)), StrideQKV{}), + take<0, 2>(SmemLayoutVt{}), + make_shape(shape<2>(TileShape_MNK{}), shape<1>(TileShape_MNK{})), + _1{})); + + using TMA_DS = decltype(make_tma_copy( + GmemTiledCopy{}, + make_tensor(make_gmem_ptr(static_cast(nullptr)), LayoutDS{}), + take<0, 2>(SmemLayoutDS{}), + make_shape(shape<0>(TileShape_MNK{}), shape<1>(TileShape_MNK{})), + _1{})); + + using BlkScaledConfig = typename Ktraits::BlkScaledConfig; + using GmemTiledCopySF = typename Ktraits::GmemTiledCopySF; + using SmemLayoutSFQ = typename Ktraits::SmemLayoutSFQ; + using SmemLayoutSFK = typename Ktraits::SmemLayoutSFK; + using SmemLayoutSFV = typename Ktraits::SmemLayoutSFV; + using SmemLayoutSFVt = typename Ktraits::SmemLayoutSFVt; + + using TMA_SFQ = decltype(make_tma_copy( + GmemTiledCopySF{}, + make_tensor(static_cast(nullptr), LayoutSF{}), + SmemLayoutSFQ{}, + make_shape(shape<0>(TileShape_MNK{}), shape<2>(TileShape_MNK{})), + _1{})); // No programmatic multicast + + + using TMA_SFKV = decltype(make_tma_copy( + GmemTiledCopySF{}, + make_tensor(static_cast(nullptr), LayoutSF{}), + SmemLayoutSFK{}(_,_,cute::Int<0>{}), + make_shape(shape<1>(TileShape_MNK{}), shape<2>(TileShape_MNK{})), + _1{})); + + using TMA_SFVt = decltype(make_tma_copy( + GmemTiledCopySF{}, + make_tensor(static_cast(nullptr), LayoutSF{}), + SmemLayoutSFVt{}(_,_,cute::Int<0>{}), + make_shape(shape<2>(TileShape_MNK{}), shape<1>(TileShape_MNK{})), + _1{})); + + using SmemCopyAtomQ = typename Ktraits::SmemCopyAtomQ; + using SmemCopyAtomKV = typename Ktraits::SmemCopyAtomKV; + using SmemCopyAtomSF = typename Ktraits::SmemCopyAtomSF; + using TiledMmaQK = typename Ktraits::TiledMmaQK; + using TiledMmaPV = typename Ktraits::TiledMmaPV; + static constexpr int NumMmaThreads = size(TiledMmaQK{}); + using MainloopPipeline = typename Ktraits::MainloopPipeline; + using PipelineParams = typename MainloopPipeline::Params; + using PipelineState = typename MainloopPipeline::PipelineState; + using MainloopPipelineQ = typename Ktraits::MainloopPipelineQ; + using PipelineParamsQ = typename Ktraits::PipelineParamsQ; + using PipelineStateQ = typename Ktraits::PipelineStateQ; + using EpilogueBarrier = typename Ktraits::EpilogueBarrier; + + // Set the bytes transferred in this TMA transaction (may involve multiple issues) + static constexpr uint32_t TmaTransactionBytesQ = static_cast( + cutlass::bits_to_bytes(cosize((SmemLayoutSFQ{})) * cute::sizeof_bits_v) + + cutlass::bits_to_bytes(size((SmemLayoutQ{})) * sizeof_bits::value)); + + static constexpr uint32_t TmaTransactionBytesK = static_cast( + cutlass::bits_to_bytes(cosize(take<0,2>(SmemLayoutSFK{})) * cute::sizeof_bits_v) + + cutlass::bits_to_bytes(cosize(take<0,2>(SmemLayoutDS{})) * cute::sizeof_bits_v) + + cutlass::bits_to_bytes(size(take<0,2>(SmemLayoutK{})) * sizeof_bits::value)); + + static constexpr uint32_t TmaTransactionBytesV = static_cast( + cutlass::bits_to_bytes(cosize(take<0,2>(SmemLayoutSFVt{})) * cute::sizeof_bits_v) + + cutlass::bits_to_bytes(size(take<0,2>(SmemLayoutVt{})) * sizeof_bits::value)); + + // Host side kernel arguments + struct Arguments { + Element const* ptr_Q; + ShapeQKV const shape_Q; + StrideQKV const stride_Q; + Element const* ptr_K; + ShapeQKV const shape_K; + StrideQKV const stride_K; + ShapeQKV const unpadded_shape_K; + Element const* ptr_Vt; + ShapeQKV const shape_Vt; + StrideQKV const stride_Vt; + ElementSF const* ptr_SFQ{nullptr}; + ShapeSF const shape_SFQ{}; + ElementSF const* ptr_SFK{nullptr}; + ShapeSF const shape_SFK{}; + ElementSF const* ptr_SFVt{nullptr}; + ShapeSF const shape_SFVt{}; + float const* ptr_ds; + ShapeQKV const shape_ds; + StrideQKV const stride_ds; + float const softmax_scale_log2; + }; + + // Device side kernel params + struct Params { + ShapeQKV const shape_Q; + LayoutSF const layout_SFQ; + ShapeQKV const shape_K; + ShapeQKV const unpadded_shape_K; + LayoutSF const layout_SFK; + ShapeQKV const shape_Vt; + LayoutSF const layout_SFVt; + LayoutDS const layout_DS; + TMA_Q tma_load_Q; + TMA_SFQ tma_load_SFQ; + TMA_KV tma_load_K; + TMA_SFKV tma_load_SFK; + TMA_Vt tma_load_Vt; + TMA_SFVt tma_load_SFVt; + TMA_DS tma_load_DS; + float const softmax_scale_log2; + }; + + + static Params + to_underlying_arguments(Arguments const& args) { + Tensor mQ = make_tensor(make_gmem_ptr(args.ptr_Q), args.shape_Q, args.stride_Q); + TMA_Q tma_load_Q = make_tma_copy( + GmemTiledCopy{}, + mQ, + SmemLayoutQ{}, + select<0, 2>(TileShape_MNK{}), + _1{}); // no mcast for Q + Tensor mK = make_tensor(make_gmem_ptr(args.ptr_K), args.shape_K, args.stride_K); + TMA_KV tma_load_K = make_tma_copy( + GmemTiledCopy{}, + mK, + SmemLayoutK{}(_, _, _0{}), + select<1, 2>(TileShape_MNK{}), + _1{}); // mcast along M mode for this N load, if any + Tensor mVt = make_tensor(make_gmem_ptr(args.ptr_Vt), args.shape_Vt, args.stride_Vt); + TMA_Vt tma_load_Vt = make_tma_copy( + GmemTiledCopy{}, + mVt, + SmemLayoutVt{}(_, _, _0{}), + make_shape(shape<2>(TileShape_MNK{}), shape<1>(TileShape_MNK{})), + _1{}); // mcast along M mode for this N load, if any + auto [Seqlen_Q, Seqlen_K, HeadNum, Batch] = args.shape_ds; + LayoutDS layout_ds = tile_to_shape(SmemLayoutAtomDS{}, make_shape(Seqlen_Q, Seqlen_K, HeadNum, Batch), Step<_2,_1,_3,_4>{}); + Tensor mDS = make_tensor(make_gmem_ptr(args.ptr_ds), layout_ds); + TMA_DS tma_load_ds = make_tma_copy ( + GmemTiledCopy{}, + mDS, + SmemLayoutDS{}(_, _, _0{}), + make_shape(shape<0>(TileShape_MNK{}), shape<1>(TileShape_MNK{})), + _1{}); + LayoutSF layout_sfq = BlkScaledConfig::tile_atom_to_shape_SFQKV(args.shape_SFQ); + Tensor mSFQ = make_tensor(make_gmem_ptr(args.ptr_SFQ), layout_sfq); + TMA_SFQ tma_load_sfq = make_tma_copy( + GmemTiledCopySF{}, + mSFQ, + SmemLayoutSFQ{}, + make_shape(shape<0>(TileShape_MNK{}), shape<2>(TileShape_MNK{})), + _1{}); + LayoutSF layout_sfk = BlkScaledConfig::tile_atom_to_shape_SFQKV(args.shape_SFK); + Tensor mSFK = make_tensor(make_gmem_ptr(args.ptr_SFK), layout_sfk); + TMA_SFKV tma_load_sfk = make_tma_copy( + GmemTiledCopySF{}, + mSFK, + SmemLayoutSFK{}(_, _, _0{}), + make_shape(shape<1>(TileShape_MNK{}), shape<2>(TileShape_MNK{})), + _1{}); + LayoutSF layout_sfvt = BlkScaledConfig::tile_atom_to_shape_SFVt(args.shape_SFVt); + Tensor mSFVt = make_tensor(make_gmem_ptr(args.ptr_SFVt), layout_sfvt); + TMA_SFVt tma_load_sfvt = make_tma_copy( + GmemTiledCopySF{}, + mSFVt, + SmemLayoutSFVt{}(_, _, _0{}), + make_shape(shape<2>(TileShape_MNK{}), shape<1>(TileShape_MNK{})), + _1{}); + return {args.shape_Q, layout_sfq, + args.shape_K, args.unpadded_shape_K, layout_sfk, + args.shape_Vt, layout_sfvt, + layout_ds, + tma_load_Q, tma_load_sfq, + tma_load_K, tma_load_sfk, + tma_load_Vt, tma_load_sfvt, + tma_load_ds, + args.softmax_scale_log2}; + } + + /// Issue Tma Descriptor Prefetch -- ideally from a single thread for best performance + CUTLASS_DEVICE + static void prefetch_tma_descriptors(Params const& mainloop_params) { + cute::prefetch_tma_descriptor(mainloop_params.tma_load_Q.get_tma_descriptor()); + cute::prefetch_tma_descriptor(mainloop_params.tma_load_K.get_tma_descriptor()); + cute::prefetch_tma_descriptor(mainloop_params.tma_load_Vt.get_tma_descriptor()); + cute::prefetch_tma_descriptor(mainloop_params.tma_load_SFQ.get_tma_descriptor()); + cute::prefetch_tma_descriptor(mainloop_params.tma_load_SFK.get_tma_descriptor()); + cute::prefetch_tma_descriptor(mainloop_params.tma_load_SFVt.get_tma_descriptor()); + cute::prefetch_tma_descriptor(mainloop_params.tma_load_DS.get_tma_descriptor()); + } + + CUTLASS_DEVICE + int get_n_block_max(Params const& mainloop_params, int m_block) { + static constexpr int kBlockM = get<0>(TileShape_MNK{}); + static constexpr int kBlockN = get<1>(TileShape_MNK{}); + int const seqlen_q = get<0>(mainloop_params.shape_Q); + int const seqlen_k = get<0>(mainloop_params.shape_K); + int n_block_max = cute::ceil_div(seqlen_k, kBlockN); + if constexpr (Is_causal) { + n_block_max = std::min(n_block_max, + cute::ceil_div((m_block + 1) * kBlockM + seqlen_k - seqlen_q, kBlockN)); + } + return n_block_max; + } + + template + CUTE_HOST_DEVICE constexpr + auto + thrfrg_SFA(SFATensor&& sfatensor, TiledMMA& mma) + { + CUTE_STATIC_ASSERT_V(rank(sfatensor) >= Int<2>{}); + + using AtomShape_MNK = typename Atom::Shape_MNK; + using AtomLayoutSFA_TV = typename Atom::Traits::SFALayout; + + auto permutation_mnk = TiledPerm{}; + auto thr_layout_vmnk = mma.get_thr_layout_vmnk(); + + // Reorder the tensor for the TiledAtom + auto t_tile = make_tile(get<0>(permutation_mnk), + get<2>(permutation_mnk)); + auto t_tensor = logical_divide(sfatensor, t_tile); // (PermM,PermK) + + // Tile the tensor for the Atom + auto a_tile = make_tile(make_layout(size<0>(AtomShape_MNK{})), + make_layout(size<2>(AtomShape_MNK{}))); + auto a_tensor = zipped_divide(t_tensor, a_tile); // ((AtomM,AtomK),(RestM,RestK)) + + // Transform the Atom mode from (M,K) to (Thr,Val) + auto tv_tensor = a_tensor.compose(AtomLayoutSFA_TV{},_); // ((ThrV,FrgV),(RestM,RestK)) + + // Tile the tensor for the Thread + auto thr_tile = make_tile(_, + make_tile(make_layout(size<1>(thr_layout_vmnk)), + make_layout(size<3>(thr_layout_vmnk)))); + auto thr_tensor = zipped_divide(tv_tensor, thr_tile); // ((ThrV,(ThrM,ThrK)),(FrgV,(RestM,RestK))) + + return thr_tensor; + } + + template + CUTE_HOST_DEVICE constexpr + auto + thrfrg_SFB(SFBTensor&& sfbtensor, TiledMMA& mma) + { + CUTE_STATIC_ASSERT_V(rank(sfbtensor) >= Int<2>{}); + + using AtomShape_MNK = typename Atom::Shape_MNK; + using AtomLayoutSFB_TV = typename Atom::Traits::SFBLayout; + + auto permutation_mnk = TiledPerm{}; + auto thr_layout_vmnk = mma.get_thr_layout_vmnk(); + + // Reorder the tensor for the TiledAtom + auto t_tile = make_tile(get<1>(permutation_mnk), + get<2>(permutation_mnk)); + auto t_tensor = logical_divide(sfbtensor, t_tile); // (PermN,PermK) + + // Tile the tensor for the Atom + auto a_tile = make_tile(make_layout(size<1>(AtomShape_MNK{})), + make_layout(size<2>(AtomShape_MNK{}))); + auto a_tensor = zipped_divide(t_tensor, a_tile); // ((AtomN,AtomK),(RestN,RestK)) + + // Transform the Atom mode from (M,K) to (Thr,Val) + auto tv_tensor = a_tensor.compose(AtomLayoutSFB_TV{},_); // ((ThrV,FrgV),(RestN,RestK)) + + // Tile the tensor for the Thread + auto thr_tile = make_tile(_, + make_tile(make_layout(size<2>(thr_layout_vmnk)), + make_layout(size<3>(thr_layout_vmnk)))); + auto thr_tensor = zipped_divide(tv_tensor, thr_tile); // ((ThrV,(ThrN,ThrK)),(FrgV,(RestN,RestK))) + return thr_tensor; + } + + template + CUTE_HOST_DEVICE constexpr + auto + partition_fragment_SFA(SFATensor&& sfatensor, ThrMma& thread_mma) + { + using ValTypeSF = typename ThrMma::Atom::Traits::ValTypeSF; + auto thr_tensor = make_tensor(static_cast(sfatensor).data(), thrfrg_SFA(sfatensor.layout(),thread_mma)); + auto thr_vmnk = thread_mma.thr_vmnk_; + auto thr_vmk = make_coord(get<0>(thr_vmnk), make_coord(get<1>(thr_vmnk), get<3>(thr_vmnk))); + auto partition_SFA = thr_tensor(thr_vmk, make_coord(_, repeat(thr_tensor)>(_))); + return make_fragment_like(partition_SFA); + } + + template + CUTE_HOST_DEVICE constexpr + auto + partition_fragment_SFB(SFBTensor&& sfbtensor, ThrMma& thread_mma) + { + using ValTypeSF = typename ThrMma::Atom::Traits::ValTypeSF; + auto thr_tensor = make_tensor(static_cast(sfbtensor).data(), thrfrg_SFB(sfbtensor.layout(),thread_mma)); + auto thr_vmnk = thread_mma.thr_vmnk_; + auto thr_vnk = make_coord(get<0>(thr_vmnk), make_coord(get<2>(thr_vmnk), get<3>(thr_vmnk))); + auto partition_SFB = thr_tensor(thr_vnk, make_coord(_, repeat(thr_tensor)>(_))); + return make_fragment_like(partition_SFB); + } + + template + CUTE_HOST_DEVICE constexpr + auto + get_layoutSFA_TV(TiledMma& mma) + { + // (M,K) -> (M,K) + auto tile_shape_mnk = tile_shape(mma); + auto ref_A = make_layout(make_shape(size<0>(tile_shape_mnk), size<2>(tile_shape_mnk))); + auto thr_layout_vmnk = mma.get_thr_layout_vmnk(); + + // (ThrV,(ThrM,ThrK)) -> (ThrV,(ThrM,ThrN,ThrK)) + auto atile = make_tile(_, + make_tile(make_layout(make_shape (size<1>(thr_layout_vmnk), size<2>(thr_layout_vmnk)), + make_stride( Int<1>{} , Int<0>{} )), + _)); + + // thr_idx -> (ThrV,ThrM,ThrN,ThrK) + auto thridx_2_thrid = right_inverse(thr_layout_vmnk); + // (thr_idx,val) -> (M,K) + return thrfrg_SFA(ref_A, mma).compose(atile, _).compose(thridx_2_thrid, _); + } + + template + CUTE_HOST_DEVICE constexpr + auto + get_layoutSFB_TV(TiledMma& mma) + { + // (N,K) -> (N,K) + auto tile_shape_mnk = tile_shape(mma); + auto ref_B = make_layout(make_shape(size<1>(tile_shape_mnk), size<2>(tile_shape_mnk))); + auto thr_layout_vmnk = mma.get_thr_layout_vmnk(); + + // (ThrV,(ThrM,ThrK)) -> (ThrV,(ThrM,ThrN,ThrK)) + auto btile = make_tile(_, + make_tile(make_layout(make_shape (size<1>(thr_layout_vmnk), size<2>(thr_layout_vmnk)), + make_stride( Int<0>{} , Int<1>{} )), + _)); + + // thr_idx -> (ThrV,ThrM,ThrN,ThrK) + auto thridx_2_thrid = right_inverse(thr_layout_vmnk); + // (thr_idx,val) -> (M,K) + return thrfrg_SFB(ref_B, mma).compose(btile, _).compose(thridx_2_thrid, _); + } + + template + CUTLASS_DEVICE void + load(Params const& mainloop_params, + SchedulerParams const& scheduler_params, + MainloopPipelineQ pipeline_q, + MainloopPipeline pipeline_k, + MainloopPipeline pipeline_v, + PipelineStateQ& smem_pipe_write_q, + PipelineState& smem_pipe_write_k, + PipelineState& smem_pipe_write_v, + SharedStorage &shared_storage, + WorkTileInfo work_tile_info, + int& work_idx, + int& tile_count_semaphore + ) { + + static constexpr int kBlockM = get<0>(TileShape_MNK{}); + static constexpr int kBlockN = get<1>(TileShape_MNK{}); + + auto [m_block, bidh, bidb] = work_tile_info.get_block_coord(scheduler_params); + + int n_block_max = get_n_block_max(mainloop_params, m_block); + + Tensor sQ = make_tensor(make_smem_ptr(shared_storage.smem_q.begin()), SmemLayoutQ{}); + Tensor sK = make_tensor(make_smem_ptr(shared_storage.smem_k.begin()), SmemLayoutK{}); + Tensor sVt = make_tensor(make_smem_ptr(shared_storage.smem_v.begin()), SmemLayoutVt{}); + Tensor sSFQ = make_tensor(make_smem_ptr(shared_storage.smem_SFQ.begin()), SmemLayoutSFQ{}); + Tensor sSFK = make_tensor(make_smem_ptr(shared_storage.smem_SFK.begin()), SmemLayoutSFK{}); + Tensor sSFVt = make_tensor(make_smem_ptr(shared_storage.smem_SFV.begin()), SmemLayoutSFVt{}); + Tensor sDS = make_tensor(make_smem_ptr(shared_storage.smem_ds.begin()), SmemLayoutDS{}); + + Tensor mQ = mainloop_params.tma_load_Q.get_tma_tensor(mainloop_params.shape_Q); + Tensor mK = mainloop_params.tma_load_K.get_tma_tensor(mainloop_params.shape_K); + Tensor mVt = mainloop_params.tma_load_Vt.get_tma_tensor(mainloop_params.shape_Vt); + Tensor mDS = mainloop_params.tma_load_DS.get_tma_tensor(shape(mainloop_params.layout_DS)); + Tensor mSFQ = mainloop_params.tma_load_SFQ.get_tma_tensor(shape(mainloop_params.layout_SFQ)); + Tensor mSFK = mainloop_params.tma_load_SFK.get_tma_tensor(shape(mainloop_params.layout_SFK)); + Tensor mSFVt = mainloop_params.tma_load_SFVt.get_tma_tensor(shape(mainloop_params.layout_SFVt)); + uint32_t block_rank_in_cluster = cute::block_rank_in_cluster(); + constexpr uint32_t cluster_shape_x = get<0>(ClusterShape()); + uint2 cluster_local_block_id = {block_rank_in_cluster % cluster_shape_x, block_rank_in_cluster / cluster_shape_x}; + Tensor gQ = local_tile(mQ(_, _, bidh, bidb), select<0, 2>(TileShape_MNK{}), make_coord(m_block, _0{})); // (M, K) + Tensor gK = local_tile(mK(_, _, bidh, bidb), select<1, 2>(TileShape_MNK{}), make_coord(_, _0{})); // (N, K, _) + Tensor gVt = local_tile(mVt(_, _, bidh, bidb), make_shape(shape<2>(TileShape_MNK{}), shape<1>(TileShape_MNK{})), make_coord(_0{}, _)); // (N, K, _) + Tensor gDS = [&] { + if constexpr (BlockMean) { + return local_tile(mDS(_, _, bidh, bidb), select<0, 1>(TileShape_MNK{}), make_coord(m_block, _)); + } else { + return local_tile(mDS(_, _, bidh, bidb), select<0, 1>(TileShape_MNK{}), make_coord(_0{}, _)); + } + }(); + Tensor gSFQ = local_tile(mSFQ(_, _, bidh, bidb), select<0, 2>(TileShape_MNK{}), make_coord(m_block, _0{})); + Tensor gSFK = local_tile(mSFK(_, _, bidh, bidb), select<1, 2>(TileShape_MNK{}), make_coord(_, _0{})); + Tensor gSFVt = local_tile(mSFVt(_, _, bidh, bidb), make_shape(shape<2>(TileShape_MNK{}), shape<1>(TileShape_MNK{})), make_coord(_0{}, _)); + auto block_tma_q = mainloop_params.tma_load_Q.get_slice(_0{}); + Tensor tQgQ = block_tma_q.partition_S(gQ); + Tensor tQsQ = block_tma_q.partition_D(sQ); + auto block_tma_sfq = mainloop_params.tma_load_SFQ.get_slice(_0{}); + Tensor tQgSFQ = block_tma_sfq.partition_S(gSFQ); + Tensor tQsSFQ = block_tma_sfq.partition_D(sSFQ); + auto block_tma_k = mainloop_params.tma_load_K.get_slice(cluster_local_block_id.x); + Tensor tKgK = group_modes<0, 3>(block_tma_k.partition_S(gK)); + Tensor tKsK = group_modes<0, 3>(block_tma_k.partition_D(sK)); + auto block_tma_sfk = mainloop_params.tma_load_SFK.get_slice(cluster_local_block_id.x); + Tensor tKgSFK = group_modes<0, 3>(block_tma_sfk.partition_S(gSFK)); + Tensor tKsSFK = group_modes<0, 3>(block_tma_sfk.partition_D(sSFK)); + auto block_tma_vt = mainloop_params.tma_load_Vt.get_slice(cluster_local_block_id.x); + Tensor tVgVt = group_modes<0, 3>(block_tma_vt.partition_S(gVt)); + Tensor tVsVt = group_modes<0, 3>(block_tma_vt.partition_D(sVt)); + auto block_tma_sfvt = mainloop_params.tma_load_SFVt.get_slice(cluster_local_block_id.x); + Tensor tVgSFVt = group_modes<0, 3>(block_tma_sfvt.partition_S(gSFVt)); + Tensor tVsSFVt = group_modes<0, 3>(block_tma_sfvt.partition_D(sSFVt)); + auto block_tma_ds = mainloop_params.tma_load_DS.get_slice(cluster_local_block_id.x); + Tensor tDSgDS = group_modes<0, 3>(block_tma_ds.partition_S(gDS)); + Tensor tDSsDS = group_modes<0, 3>(block_tma_ds.partition_D(sDS)); + uint16_t mcast_mask_kv = 0; + + int n_block = n_block_max - 1; + int lane_predicate = cute::elect_one_sync(); + if (lane_predicate) { + pipeline_q.producer_acquire(smem_pipe_write_q); + copy(mainloop_params.tma_load_Q.with(*pipeline_q.producer_get_barrier(smem_pipe_write_q), 0), tQgQ, tQsQ); + copy(mainloop_params.tma_load_SFQ.with(*pipeline_q.producer_get_barrier(smem_pipe_write_q), 0), tQgSFQ, tQsSFQ); + ++smem_pipe_write_q; + pipeline_k.producer_acquire(smem_pipe_write_k); + copy(mainloop_params.tma_load_K.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv), + tKgK(_, n_block), tKsK(_, smem_pipe_write_k.index())); + copy(mainloop_params.tma_load_SFK.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv), + tKgSFK(_, n_block), tKsSFK(_, smem_pipe_write_k.index())); + copy(mainloop_params.tma_load_DS.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv), + tDSgDS(_, n_block), tDSsDS(_, smem_pipe_write_k.index())); + ++smem_pipe_write_k; + pipeline_v.producer_acquire(smem_pipe_write_v); + copy(mainloop_params.tma_load_Vt.with(*pipeline_v.producer_get_barrier(smem_pipe_write_v), mcast_mask_kv), + tVgVt(_, n_block), tVsVt(_, smem_pipe_write_v.index())); + copy(mainloop_params.tma_load_SFVt.with(*pipeline_v.producer_get_barrier(smem_pipe_write_v), mcast_mask_kv), + tVgSFVt(_, n_block), tVsSFVt(_, smem_pipe_write_v.index())); + ++smem_pipe_write_v; + } + + n_block--; + if (lane_predicate) { + // CUTLASS_PRAGMA_NO_UNROLL + #pragma unroll 2 + for (; n_block >= 0; --n_block) { + pipeline_k.producer_acquire(smem_pipe_write_k); + copy(mainloop_params.tma_load_K.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv), + tKgK(_, n_block), tKsK(_, smem_pipe_write_k.index())); + copy(mainloop_params.tma_load_SFK.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv), + tKgSFK(_, n_block), tKsSFK(_, smem_pipe_write_k.index())); + copy(mainloop_params.tma_load_DS.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv), + tDSgDS(_, n_block), tDSsDS(_, smem_pipe_write_k.index())); + ++smem_pipe_write_k; + pipeline_v.producer_acquire(smem_pipe_write_v); + copy(mainloop_params.tma_load_Vt.with(*pipeline_v.producer_get_barrier(smem_pipe_write_v), mcast_mask_kv), + tVgVt(_, n_block), tVsVt(_, smem_pipe_write_v.index())); + copy(mainloop_params.tma_load_SFVt.with(*pipeline_v.producer_get_barrier(smem_pipe_write_v), mcast_mask_kv), + tVgSFVt(_, n_block), tVsSFVt(_, smem_pipe_write_v.index())); + ++smem_pipe_write_v; + } + } + ++work_idx; + } + + /// Perform a Producer Epilogue to prevent early exit of blocks in a Cluster + CUTLASS_DEVICE void + load_tail(MainloopPipelineQ pipeline_q, + MainloopPipeline pipeline_k, + MainloopPipeline pipeline_v, + PipelineStateQ& smem_pipe_write_q, + PipelineState& smem_pipe_write_k, + PipelineState& smem_pipe_write_v) { + int lane_predicate = cute::elect_one_sync(); + // Issue the epilogue waits + if (lane_predicate) { + pipeline_q.producer_tail(smem_pipe_write_q); + pipeline_k.producer_tail(smem_pipe_write_k); + pipeline_v.producer_tail(smem_pipe_write_v); + } + } + + template + CUTLASS_DEVICE void + mma(Params const& mainloop_params, + MainloopPipelineQ pipeline_q, + MainloopPipeline pipeline_k, + MainloopPipeline pipeline_v, + PipelineStateQ& smem_pipe_read_q, + PipelineState& smem_pipe_read_k, + PipelineState& smem_pipe_read_v, + FrgTensorO& tOrO_store, + SoftmaxFused& softmax_fused, + int n_block_count, + int thread_idx, + int work_idx, + int m_block, + SharedStorage& shared_storage + ) { + + static_assert(is_rmem::value, "O tensor must be rmem resident."); + + static constexpr int kBlockM = get<0>(TileShape_MNK{}); + static constexpr int kBlockN = get<1>(TileShape_MNK{}); + static constexpr int kBlockK = get<2>(TileShape_MNK{}); + Tensor sQ = make_tensor(make_smem_ptr(shared_storage.smem_q.begin()), SmemLayoutQ{}); + Tensor sK = make_tensor(make_smem_ptr(shared_storage.smem_k.begin()), SmemLayoutK{}); + Tensor sVt = make_tensor(make_smem_ptr(shared_storage.smem_v.begin()), SmemLayoutVt{}); + Tensor sDS = make_tensor(make_smem_ptr(shared_storage.smem_ds.begin()), SmemLayoutDS{}); + Tensor sSFQ = make_tensor(make_smem_ptr(shared_storage.smem_SFQ.begin()), SmemLayoutSFQ{}); + Tensor sSFK = make_tensor(make_smem_ptr(shared_storage.smem_SFK.begin()), SmemLayoutSFK{}); + Tensor sSFVt = make_tensor(make_smem_ptr(shared_storage.smem_SFV.begin()), SmemLayoutSFVt{}); + + Tensor cQ = make_identity_tensor(make_shape(size<0>(sQ), size<1>(sQ))); + Tensor cKV = make_identity_tensor(make_shape(size<0>(sK), size<1>(sK))); + TiledMmaQK tiled_mma_qk; + TiledMmaPV tiled_mma_pv; + auto thread_mma_qk = tiled_mma_qk.get_thread_slice(thread_idx); + auto thread_mma_pv = tiled_mma_pv.get_thread_slice(thread_idx); + + Tensor tSrQ = thread_mma_qk.partition_fragment_A(sQ); + Tensor tSrK = thread_mma_qk.partition_fragment_B(sK(_,_,Int<0>{})); + Tensor tOrVt = thread_mma_pv.partition_fragment_B(sVt(_,_,Int<0>{})); + Tensor tOrP = make_tensor_like(LayoutP{}); + Tensor tSrSFQ = partition_fragment_SFA(sSFQ, thread_mma_qk); + Tensor tSrSFK = partition_fragment_SFB(sSFK(_,_,Int<0>{}), thread_mma_qk); + Tensor tOrSFVt = partition_fragment_SFB(sSFVt(_,_,Int<0>{}), thread_mma_pv); + Tensor tOrSFP = make_tensor(LayoutSFP{}); + Tensor tOrSFP_flt = filter_zeros(tOrSFP); + Tensor tSrDS = make_tensor(make_shape(_8{}, _4{}), make_stride(_1{}, _8{})); + // copy qk and sf from smem to rmem + auto smem_tiled_copy_Q = make_tiled_copy_A(SmemCopyAtomQ{}, tiled_mma_qk); + auto smem_thr_copy_Q = smem_tiled_copy_Q.get_thread_slice(thread_idx); + Tensor tSsQ = smem_thr_copy_Q.partition_S(as_position_independent_swizzle_tensor(sQ)); + Tensor tSrQ_copy_view = smem_thr_copy_Q.retile_D(tSrQ); + + auto smem_tiled_copy_K = make_tiled_copy_B(SmemCopyAtomKV{}, tiled_mma_qk); + auto smem_thr_copy_K = smem_tiled_copy_K.get_thread_slice(thread_idx); + Tensor tSsK = smem_thr_copy_K.partition_S(as_position_independent_swizzle_tensor(sK)); + Tensor tSrK_copy_view = smem_thr_copy_K.retile_D(tSrK); + + auto smem_tiled_copy_V = make_tiled_copy_B(SmemCopyAtomKV{}, tiled_mma_pv); + auto smem_thr_copy_V = smem_tiled_copy_V.get_thread_slice(thread_idx); + Tensor tOsVt = smem_thr_copy_V.partition_S(as_position_independent_swizzle_tensor(sVt)); + Tensor tOrVt_copy_view = smem_thr_copy_V.retile_D(tOrVt); + + auto tile_shape_mnk = tile_shape(tiled_mma_qk); + auto smem_tiled_copy_SFQ = make_tiled_copy_impl(SmemCopyAtomSF{}, + get_layoutSFA_TV(tiled_mma_qk), + make_shape(size<0>(tile_shape_mnk), size<2>(tile_shape_mnk)) + ); + auto smem_thr_copy_SFQ = smem_tiled_copy_SFQ.get_thread_slice(thread_idx); + Tensor tSsSFQ = smem_thr_copy_SFQ.partition_S(as_position_independent_swizzle_tensor(sSFQ)); + Tensor tSrSFQ_copy_view = smem_thr_copy_SFQ.retile_D(tSrSFQ); + + auto smem_tiled_copy_SFK = make_tiled_copy_impl(SmemCopyAtomSF{}, + get_layoutSFB_TV(tiled_mma_qk), + make_shape(size<1>(tile_shape_mnk), size<2>(tile_shape_mnk)) + ); + auto smem_thr_copy_SFK = smem_tiled_copy_SFK.get_thread_slice(thread_idx); + Tensor tSsSFK = smem_thr_copy_SFK.partition_S(as_position_independent_swizzle_tensor(sSFK)); + Tensor tSrSFK_copy_view = smem_thr_copy_SFK.retile_D(tSrSFK); + + auto smem_tiled_copy_SFV = make_tiled_copy_impl(SmemCopyAtomSF{}, + get_layoutSFB_TV(tiled_mma_pv), + make_shape(size<1>(tile_shape_mnk), size<2>(tile_shape_mnk)) + ); + auto smem_thr_copy_SFV = smem_tiled_copy_SFV.get_thread_slice(thread_idx); + Tensor tOsSFVt = smem_thr_copy_SFV.partition_S(as_position_independent_swizzle_tensor(sSFVt)); + Tensor tOrSFVt_copy_view = smem_thr_copy_SFV.retile_D(tOrSFVt); + + auto consumer_wait = [](auto& pipeline, auto& smem_pipe_read) { + auto barrier_token = pipeline.consumer_try_wait(smem_pipe_read); + pipeline.consumer_wait(smem_pipe_read, barrier_token); + }; + + int const seqlen_q = get<0>(mainloop_params.shape_Q); + int const seqlen_k = get<0>(mainloop_params.shape_K); + int const unpadded_seqlen_k = get<0>(mainloop_params.unpadded_shape_K); + int n_block = n_block_count - 1; + + auto copy_k_block = [&](auto block_id) { + auto tSsK_stage = tSsK(_, _, _, smem_pipe_read_k.index()); + auto tSsSFK_stage = tSsSFK(_, _, _, smem_pipe_read_k.index()); + copy(smem_tiled_copy_K, tSsK_stage(_, _, block_id), tSrK_copy_view(_, _, block_id)); + copy(smem_tiled_copy_SFK, tSsSFK_stage(_, _, block_id), tSrSFK_copy_view(_, _, block_id)); + }; + + auto copy_v_block = [&](auto block_id) { + auto tOsVt_stage = tOsVt(_, _, _, smem_pipe_read_v.index()); + auto tOsSFVt_stage = tOsSFVt(_, _, _, smem_pipe_read_v.index()); + copy(smem_tiled_copy_V, tOsVt_stage(_, _, block_id), tOrVt_copy_view(_, _, block_id)); + copy(smem_tiled_copy_SFV, tOsSFVt_stage(_, _, block_id), tOrSFVt_copy_view(_, _, block_id)); + }; + // auto gemm_qk = [&](auto block_id) { + // cute::gemm(tiled_mma_qk, make_zip_tensor(tSrQ(_, _, block_id), tSrSFQ(_, _, block_id)), make_zip_tensor(tSrK(_, _, block_id), tSrSFK(_, _, block_id)), tSrS); + // }; + // auto gemm_pv = [&](auto block_id) { + // cute::gemm(tiled_mma_pv, make_zip_tensor(tOrP(_, _, block_id), tOrSFP(_, _, block_id)), make_zip_tensor(tOrVt(_, _, block_id), tOrSFVt(_, _, block_id)), tOrO); + // }; + auto add_delta_s = [&](auto& acc) { + auto tSsDS_stage = recast(sDS(_, _, smem_pipe_read_k.index())); + auto acc_float4 = recast(acc); + int quad_id = (threadIdx.x % 4) * 2; + for (int i = 0; i < 4; i++) { + auto num = quad_id + i * 8; + float4 delta_s_0 = tSsDS_stage(make_coord(_0{}, _0{}), make_coord(num, _0{})); + float4 delta_s_1 = tSsDS_stage(make_coord(_0{}, _0{}), make_coord(num + 1, _0{})); + acc_float4(make_coord(make_coord(_0{}, _0{}), _0{}), _0{}, i) = delta_s_0; + acc_float4(make_coord(make_coord(_0{}, _0{}), _1{}), _0{}, i) = delta_s_0; + acc_float4(make_coord(make_coord(_0{}, _1{}), _0{}), _0{}, i) = delta_s_1; + acc_float4(make_coord(make_coord(_0{}, _1{}), _1{}), _0{}, i) = delta_s_1; + } + }; + consumer_wait(pipeline_q, smem_pipe_read_q); + copy(smem_tiled_copy_Q, tSsQ, tSrQ_copy_view); + copy(smem_tiled_copy_SFQ, tSsSFQ, tSrSFQ_copy_view); + pipeline_q.consumer_release(smem_pipe_read_q); + ++smem_pipe_read_q; + + Tensor tSrS = partition_fragment_C(tiled_mma_qk, select<0, 1>(TileShape_MNK{})); + Tensor tSrS_converion_view = make_tensor(tSrS.data(), flash::convert_to_conversion_layout(tSrS.layout())); + Tensor AbsMaxP = make_tensor_like( + make_layout(shape(group<1, 4>(flatten(tSrS_converion_view.layout()(make_coord(_0{}, _), _, _))))) + ); + consumer_wait(pipeline_k, smem_pipe_read_k); + copy_k_block(_0{}); + add_delta_s(tSrS); + CUTLASS_PRAGMA_UNROLL + for (int k_block = 0; k_block < size<2>(tSrQ); ++k_block) { + cute::gemm(tiled_mma_qk, make_zip_tensor(tSrQ(_, _, k_block), tSrSFQ(_, _, k_block)), + make_zip_tensor(tSrK(_, _, k_block), tSrSFK(_, _, k_block)), tSrS); + if (k_block < size<2>(tSrQ) - 1) { + copy_k_block(k_block + 1); + } else { + pipeline_k.consumer_release(smem_pipe_read_k); + ++smem_pipe_read_k; + } + } + + + auto col_limit_causal = [&](int row, int n_block) { + return row + 1 + seqlen_k - n_block * kBlockN - seqlen_q + m_block * kBlockM; + }; + { + Tensor cS = cute::make_identity_tensor(select<0, 1>(TileShape_MNK{})); + Tensor tScS = thread_mma_qk.partition_C(cS); + CUTLASS_PRAGMA_UNROLL + for (int i = 0; i < size(tSrS); ++i) { + if constexpr (!Is_causal) { // Just masking based on col + if (int(get<1>(tScS(i))) >= int(unpadded_seqlen_k - n_block * kBlockN)) { tSrS(i) = -INFINITY; } + } else { + if (int(get<1>(tScS(i))) >= std::min(seqlen_k - n_block * kBlockN, + col_limit_causal(int(get<0>(tScS(i))), n_block))) { + tSrS(i) = -INFINITY; + } + } + } + } + auto quantize = [&](auto mma_k, auto acc_conversion_view) { + Tensor AbsMaxP_stagek = AbsMaxP(_, make_coord(_, _, mma_k)); + Tensor acc_conversion_stagek = acc_conversion_view(_, _, mma_k); + Tensor SFP = make_tensor_like(AbsMaxP_stagek.layout()); + Tensor SFP_uint32_view = recast(SFP); + CUTLASS_PRAGMA_UNROLL + for (int i = 0; i < size(AbsMaxP_stagek); i += 4) { + uint32_t& tmp = SFP_uint32_view(i / 4); + flash::packed_float_to_ue4m3( + AbsMaxP_stagek(i), + AbsMaxP_stagek(i + 1), + AbsMaxP_stagek(i + 2), + AbsMaxP_stagek(i + 3), + tmp + ); + } + int const quad_id = threadIdx.x & 3; + uint32_t MASK = (0xFF00FF) << ((quad_id & 1) * 8); + Tensor tOrSFP_uint32_view = recast(tOrSFP(_, _, mma_k)); + Tensor tOrP_uint32_view = recast(tOrP(_, _, mma_k)); + + CUTLASS_PRAGMA_UNROLL + for (int mma_m = 0; mma_m < size<1>(tOrP); ++mma_m) { + CUTLASS_PRAGMA_UNROLL + for (int i = 0; i < 4; ++i) { + flash::packed_float_to_e2m1( + acc_conversion_stagek(make_coord(_0{}, i), mma_m), + acc_conversion_stagek(make_coord(_1{}, i), mma_m), + acc_conversion_stagek(make_coord(_2{}, i), mma_m), + acc_conversion_stagek(make_coord(_3{}, i), mma_m), + acc_conversion_stagek(make_coord(_4{}, i), mma_m), + acc_conversion_stagek(make_coord(_5{}, i), mma_m), + acc_conversion_stagek(make_coord(_6{}, i), mma_m), + acc_conversion_stagek(make_coord(_7{}, i), mma_m), + tOrP_uint32_view(i, mma_m) + ); + } + uint32_t local_sfp = SFP_uint32_view(_0{}, _0{}, mma_m); + uint32_t peer_sfp = __shfl_xor_sync(int32_t(-1), local_sfp, 2); + if ((quad_id & 1) == 0) { + uint32_t sfp = (local_sfp & MASK) | ((peer_sfp & MASK) << 8); + tOrSFP_uint32_view(_0{}, mma_m) = sfp; + } else { + uint32_t sfp = (peer_sfp & MASK) | ((local_sfp & MASK) >> 8); + tOrSFP_uint32_view(_0{}, mma_m) = sfp; + } + } + }; + + softmax_fused.template online_softmax_with_quant(tSrS, AbsMaxP, mainloop_params.softmax_scale_log2); + + consumer_wait(pipeline_v, smem_pipe_read_v); + copy_v_block(_0{}); + quantize(_0{}, tSrS_converion_view); + CUTLASS_PRAGMA_UNROLL + for (int v_block = 0; v_block < size<2>(tOrP); ++v_block) { + cute::gemm(tiled_mma_pv, make_zip_tensor(tOrP(_, _, v_block), tOrSFP(_, _, v_block)), + make_zip_tensor(tOrVt(_, _, v_block), tOrSFVt(_, _, v_block)), tOrO_store); + if (v_block < size<2>(tOrP) - 1) { + copy_v_block(v_block + 1); + quantize(v_block + 1, tSrS_converion_view); + } else { + pipeline_v.consumer_release(smem_pipe_read_v); + ++smem_pipe_read_v; + } + } + + n_block--; + constexpr int n_masking_steps = !Is_causal ? 1 : cute::ceil_div(kBlockM, kBlockN) + 1; + // // Only go through these if Is_causal, since n_masking_steps = 1 when !Is_causal + CUTLASS_PRAGMA_UNROLL + for (int masking_step = 0; masking_step < n_masking_steps - 1 && n_block >= 0; ++masking_step, --n_block) { + Tensor tSrS = partition_fragment_C(tiled_mma_qk, select<0, 1>(TileShape_MNK{})); + Tensor tSrS_converion_view = make_tensor(tSrS.data(), flash::convert_to_conversion_layout(tSrS.layout())); + consumer_wait(pipeline_k, smem_pipe_read_k); + copy_k_block(_0{}); + add_delta_s(tSrS); + CUTLASS_PRAGMA_UNROLL + for (int k_block = 0; k_block < size<2>(tSrQ); ++k_block) { + cute::gemm(tiled_mma_qk, make_zip_tensor(tSrQ(_, _, k_block), tSrSFQ(_, _, k_block)), + make_zip_tensor(tSrK(_, _, k_block), tSrSFK(_, _, k_block)), tSrS); + if (k_block < size<2>(tSrQ) - 1) { + copy_k_block(k_block + 1); + } + } + pipeline_k.consumer_release(smem_pipe_read_k); // release K + ++smem_pipe_read_k; + Tensor cS = cute::make_identity_tensor(select<0, 1>(TileShape_MNK{})); + Tensor tScS = thread_mma_qk.partition_C(cS); + #pragma unroll + for (int i = 0; i < size(tSrS); ++i) { + if (int(get<1>(tScS(i))) >= col_limit_causal(int(get<0>(tScS(i))), n_block - 1)) { + tSrS(i) = -INFINITY; + } + } + softmax_fused.template online_softmax_with_quant(tSrS, AbsMaxP, mainloop_params.softmax_scale_log2); + Tensor tOrO = make_fragment_like(tOrO_store); + consumer_wait(pipeline_v, smem_pipe_read_v); + copy_v_block(_0{}); + quantize(_0{}, tSrS_converion_view); + CUTLASS_PRAGMA_UNROLL + for (int v_block = 0; v_block < size<2>(tOrP); ++v_block) { + cute::gemm(tiled_mma_pv, make_zip_tensor(tOrP(_, _, v_block), tOrSFP(_, _, v_block)), + make_zip_tensor(tOrVt(_, _, v_block), tOrSFVt(_, _, v_block)), tOrO); + if (v_block < size<2>(tOrP) - 1) { + copy_v_block(v_block + 1); + quantize(v_block + 1, tSrS_converion_view); + } + } + pipeline_v.consumer_release(smem_pipe_read_v); + ++smem_pipe_read_v; + if (masking_step > 0) { softmax_fused.rescale_o(tOrO_store, tOrO); } + } + + #pragma unroll 1 + for (; n_block >= 0; --n_block) { + Tensor tSrS = partition_fragment_C(tiled_mma_qk, select<0, 1>(TileShape_MNK{})); + Tensor tSrS_converion_view = make_tensor(tSrS.data(), flash::convert_to_conversion_layout(tSrS.layout())); + consumer_wait(pipeline_k, smem_pipe_read_k); + copy_k_block(_0{}); + add_delta_s(tSrS); + CUTLASS_PRAGMA_UNROLL + for (int k_block = 0; k_block < size<2>(tSrQ); ++k_block) { + cute::gemm(tiled_mma_qk, make_zip_tensor(tSrQ(_, _, k_block), tSrSFQ(_, _, k_block)), + make_zip_tensor(tSrK(_, _, k_block), tSrSFK(_, _, k_block)), tSrS); + if (k_block < size<2>(tSrQ) - 1) { + copy_k_block(k_block + 1); + } else { + pipeline_k.consumer_release(smem_pipe_read_k); + ++smem_pipe_read_k; + } + } + softmax_fused.template online_softmax_with_quant(tSrS, AbsMaxP, mainloop_params.softmax_scale_log2); + Tensor tOrO = make_fragment_like(tOrO_store); + consumer_wait(pipeline_v, smem_pipe_read_v); + copy_v_block(_0{}); + quantize(_0{}, tSrS_converion_view); + CUTLASS_PRAGMA_UNROLL + for (int v_block = 0; v_block < size<2>(tOrP); ++v_block) { + cute::gemm(tiled_mma_pv, make_zip_tensor(tOrP(_, _, v_block), tOrSFP(_, _, v_block)), + make_zip_tensor(tOrVt(_, _, v_block), tOrSFVt(_, _, v_block)), tOrO); + if (v_block < size<2>(tOrP) - 1) { + copy_v_block(v_block + 1); + quantize(v_block + 1, tSrS_converion_view); + } else { + pipeline_v.consumer_release(smem_pipe_read_v); + ++smem_pipe_read_v; + } + } + softmax_fused.rescale_o(tOrO_store, tOrO); + } + softmax_fused.finalize(tOrO_store); + return; + } + +}; + +} // namespace flash + diff --git a/fastvideo/attention/backends/sageattn/blackwell/named_barrier.h b/fastvideo/attention/backends/sageattn/blackwell/named_barrier.h new file mode 100644 index 000000000..bd8aefb6a --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/named_barrier.h @@ -0,0 +1,119 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#pragma once + +#include "cutlass/arch/barrier.h" +#include "cutlass/pipeline/sm90_pipeline.hpp" + +namespace flash { + +enum class FP4NamedBarriers { + QueryEmpty = 1, + WarpSpecializedConsumer = 2, + WarpSpecializedPingPongConsumer1 = 3, + WarpSpecializedPingPongConsumer2 = 4, + ProducerEnd = 5, + ConsumerEnd = 6, + EpilogueBarrier = 7 +}; + +template +struct OrderedSequenceBarrierVarGroupSizeSharedStorage { + using Barrier = cutlass::arch::ClusterBarrier; + Barrier barrier_[SequenceDepth][SequenceLength]; +}; + + template +class OrderedSequenceBarrierVarGroupSize { +public: + static constexpr int SequenceDepth = SequenceDepth_; + static constexpr int SequenceLength = SequenceLength_; + using Barrier = cutlass::arch::ClusterBarrier; + using SharedStorage = flash::OrderedSequenceBarrierVarGroupSizeSharedStorage; + + + struct Params { + uint32_t group_id; + uint32_t* group_size_list; + }; + +private : + // In future this Params object can be replaced easily with a CG object + Params params_; + Barrier *barrier_ptr_; + cutlass::PipelineState stage_; + + static constexpr int Depth = SequenceDepth; + static constexpr int Length = SequenceLength; + +public: + OrderedSequenceBarrierVarGroupSize() = delete; + OrderedSequenceBarrierVarGroupSize(const OrderedSequenceBarrierVarGroupSize&) = delete; + OrderedSequenceBarrierVarGroupSize(OrderedSequenceBarrierVarGroupSize&&) = delete; + OrderedSequenceBarrierVarGroupSize& operator=(const OrderedSequenceBarrierVarGroupSize&) = delete; + OrderedSequenceBarrierVarGroupSize& operator=(OrderedSequenceBarrierVarGroupSize&&) = delete; + ~OrderedSequenceBarrierVarGroupSize() = default; + + CUTLASS_DEVICE + OrderedSequenceBarrierVarGroupSize(SharedStorage& storage, Params const& params) : + params_(params), + barrier_ptr_(&storage.barrier_[0][0]), + // Group 0 - starts with an opposite phase + stage_({0, params.group_id == 0, 0}) { + int warp_idx = cutlass::canonical_warp_idx_sync(); + int lane_predicate = cute::elect_one_sync(); + + // Barrier FULL, EMPTY init + // Init is done only by the one elected thread of the block + if (warp_idx == 0 && lane_predicate) { + for (int d = 0; d < Depth; ++d) { + for (int l = 0; l < Length; ++l) { + barrier_ptr_[d * Length + l].init(*(params.group_size_list + l)); + } + } + } + cutlass::arch::fence_barrier_init(); + } + + // Wait on a stage to be unlocked + CUTLASS_DEVICE + void wait() { + get_barrier_for_current_stage(params_.group_id).wait(stage_.phase()); + } + + // Signal completion of Stage and move to the next stage + // (group_id) signals to (group_id+1) + CUTLASS_DEVICE + void arrive() { + int signalling_id = (params_.group_id + 1) % Length; + get_barrier_for_current_stage(signalling_id).arrive(); + ++stage_; + } + + CUTLASS_DEVICE + void advance() { + ++stage_; + } + +private: + + CUTLASS_DEVICE + Barrier& get_barrier_for_current_stage(int group_id) { + return barrier_ptr_[stage_.index() * Length + group_id]; + } +}; + + } // flash \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/blackwell/params.h b/fastvideo/attention/backends/sageattn/blackwell/params.h new file mode 100644 index 000000000..c87b7db08 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/params.h @@ -0,0 +1,179 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + #pragma once + + #include + #include + + #ifdef OLD_GENERATOR_PATH + #include + #else + #include + #endif + + #include // For at::cuda::philox::unpack + + #include "cutlass/fast_math.h" // For cutlass::FastDivmod + + //////////////////////////////////////////////////////////////////////////////////////////////////// + + struct Qkv_params { + using index_t = int64_t; + // The QKV matrices. + void *__restrict__ q_ptr; + void *__restrict__ k_ptr; + void *__restrict__ v_ptr; + void *__restrict__ delta_s_ptr; + // The QKV scale factor matrices. + void *__restrict__ sfq_ptr; + void *__restrict__ sfk_ptr; + void *__restrict__ sfv_ptr; + // The stride between rows of the Q, K and V matrices. + index_t q_batch_stride; + index_t k_batch_stride; + index_t v_batch_stride; + index_t q_row_stride; + index_t k_row_stride; + index_t v_row_stride; + index_t q_head_stride; + index_t k_head_stride; + index_t v_head_stride; + index_t ds_batch_stride; + index_t ds_row_stride; + index_t ds_head_stride; + // The stride of the Q, K and V scale factor matrices. + index_t sfq_batch_stride; + index_t sfk_batch_stride; + index_t sfv_batch_stride; + index_t sfq_row_stride; + index_t sfk_row_stride; + index_t sfv_row_stride; + index_t sfq_head_stride; + index_t sfk_head_stride; + index_t sfv_head_stride; + + // The number of heads. + int h, h_k; + // In the case of multi-query and grouped-query attention (MQA/GQA), nheads_k could be + // different from nheads (query). + int h_h_k_ratio; // precompute h / h_k, + }; + + //////////////////////////////////////////////////////////////////////////////////////////////////// + + struct Flash_fwd_params : public Qkv_params { + + // The O matrix (output). + void * __restrict__ o_ptr; + void * __restrict__ oaccum_ptr; + void * __restrict__ s_ptr; + + // The stride between rows of O. + index_t o_batch_stride; + index_t o_row_stride; + index_t o_head_stride; + + // The pointer to the P matrix. + void * __restrict__ p_ptr; + + // The pointer to the softmax sum. + void * __restrict__ softmax_lse_ptr; + void * __restrict__ softmax_lseaccum_ptr; + + // The dimensions. + int b, seqlen_q, seqlen_k, seqlen_knew, d, seqlen_q_rounded, seqlen_k_rounded, d_rounded, rotary_dim, unpadded_seqlen_k; + cutlass::FastDivmod head_divmod, m_block_divmod; + int total_blocks; + int seqlen_s; + + // The scaling factors for the kernel. + float scale_softmax; + float scale_softmax_log2; + uint32_t scale_softmax_log2_half2; + + // array of length b+1 holding starting offset of each sequence. + int * __restrict__ cu_seqlens_q; + int * __restrict__ cu_seqlens_k; + + // If provided, the actual length of each k sequence. + int * __restrict__ seqused_k; + + int *__restrict__ blockmask; + + // The K_new and V_new matrices. + void * __restrict__ knew_ptr; + void * __restrict__ vnew_ptr; + + // The stride between rows of the Q, K and V matrices. + index_t knew_batch_stride; + index_t vnew_batch_stride; + index_t knew_row_stride; + index_t vnew_row_stride; + index_t knew_head_stride; + index_t vnew_head_stride; + + // The cos and sin matrices for rotary embedding. + void * __restrict__ rotary_cos_ptr; + void * __restrict__ rotary_sin_ptr; + + // The indices to index into the KV cache. + int * __restrict__ cache_batch_idx; + + // Paged KV cache + int * __restrict__ block_table; + index_t block_table_batch_stride; + int page_block_size; + + // The dropout probability (probability of keeping an activation). + float p_dropout; + // uint32_t p_dropout_in_uint; + // uint16_t p_dropout_in_uint16_t; + uint8_t p_dropout_in_uint8_t; + + // Scale factor of 1 / (1 - p_dropout). + float rp_dropout; + float scale_softmax_rp_dropout; + + // Local window size + int window_size_left, window_size_right; + + // Random state. + at::PhiloxCudaState philox_args; + + // Pointer to the RNG seed (idx 0) and offset (idx 1). + uint64_t * rng_state; + + bool is_bf16; + bool is_e4m3; + bool is_causal; + bool per_block_mean; + bool single_level_p_quant; // If true, use single-level 1x16 block scale quantization for P (like V), instead of two-level quantization + // If is_seqlens_k_cumulative, then seqlen_k is cu_seqlens_k[bidb + 1] - cu_seqlens_k[bidb]. + // Otherwise it's cu_seqlens_k[bidb], i.e., we use cu_seqlens_k to store the sequence lengths of K. + bool is_seqlens_k_cumulative; + + bool is_rotary_interleaved; + + int num_splits; // For split-KV version + + void * __restrict__ alibi_slopes_ptr; + index_t alibi_slopes_batch_stride; + + int * __restrict__ tile_count_semaphore; + }; + + //////////////////////////////////////////////////////////////////////////////////////////////////// \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/blackwell/softmax_fused.h b/fastvideo/attention/backends/sageattn/blackwell/softmax_fused.h new file mode 100644 index 000000000..3619aef38 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/softmax_fused.h @@ -0,0 +1,189 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + #pragma once + + #include + #include "cute/tensor.hpp" + #include "cutlass/numeric_types.h" + #include "utils.h" + + namespace flash { + + using namespace cute; + + template + struct SoftmaxFused{ + + using TensorT = decltype(make_fragment_like(Shape>{})); + TensorT row_sum, row_max, scores_scale; + static constexpr float fp8_scalexfp4_scale = 1.f / (448 * 6); + static constexpr float fp8_scalexfp4_scale_log2 = -11.392317422778762f; //log2f(fp8_scalexfp4_scale) + static constexpr float fp4_scale_log2 = -2.584962500721156f; // log2f(fp4_scale) + static constexpr int RowReductionThr = 4; + + // If true, use single-level quantization: s_P2, P̂_2 = φ(P̃) directly (standard per-block FP4 quantization like V) + // If false (default), use two-level quantization: s_P1 = rowmax(P̃)/(448×6), then s_P2, P̂_2 = φ(P̃/s_P1) + bool single_level_p_quant; + + CUTLASS_DEVICE SoftmaxFused(bool single_level = false) : single_level_p_quant(single_level) {}; + + template + CUTLASS_DEVICE auto online_softmax_with_quant( + TensorAcc& acc, + TensorMax& AbsMaxP, + const float softmax_scale_log2 + ) { + Tensor acc_reduction_view = make_tensor(acc.data(), flash::convert_to_reduction_layout(acc.layout())); + Tensor acc_conversion_view = make_tensor(acc.data(), flash::convert_to_conversion_layout(acc.layout())); + Tensor acc_conversion_flatten = group_modes<1, 5>(group_modes<0, 2>(flatten(acc_conversion_view))); + + if constexpr (FirstTile) { + fill(row_max, -INFINITY); + clear(row_sum); + fill(scores_scale, 1.f); + + CUTLASS_PRAGMA_UNROLL + for (int mi = 0; mi < size<0>(acc_reduction_view); mi++) { + CUTLASS_PRAGMA_UNROLL + for (int ni = 0; ni < size<1, 1>(acc_reduction_view); ni++) { + CUTLASS_PRAGMA_UNROLL + for (int ei = 0; ei < size<1, 0>(acc_reduction_view); ei++) { + AbsMaxP(mi, ni) = fmaxf(AbsMaxP(mi, ni), acc_reduction_view(mi, make_coord(ei, ni))); + } + float max_recv = __shfl_xor_sync(int32_t(-1), AbsMaxP(mi, ni), 1); // exchange max with neighbour thread of 8 elements + AbsMaxP(mi, ni) = fmaxf(AbsMaxP(mi, ni), max_recv); + row_max(mi) = fmaxf(row_max(mi), AbsMaxP(mi, ni)); + } + + float max_recv = __shfl_xor_sync(int32_t(-1), row_max(mi), 2); // exchange max in a quad in a row + row_max(mi) = fmaxf(row_max(mi), max_recv); + + // Two-level P quantization (default): s_P1 = rowmax(P̃)/(448×6), then s_P2,P̂_2 = φ(P̃/s_P1) + // - Pre-scales P to [0, 448×6] range before φ, output scaled by s_P1 + // Single-level P quantization: s_P2, P̂_2 = φ(P̃) directly (like V quantization) + // - No s_P1, just standard per-block FP4 quantization φ + const float s_P1_offset = single_level_p_quant ? 0.f : fp8_scalexfp4_scale_log2; + const float max_scaled = InfCheck + ? (row_max(mi) == -INFINITY ? 0.f : (row_max(mi) * softmax_scale_log2 + s_P1_offset)) + : (row_max(mi) * softmax_scale_log2 + s_P1_offset); + CUTLASS_PRAGMA_UNROLL + for (int ni = 0; ni < size<1>(acc_reduction_view); ni++) { + acc_reduction_view(mi, ni) = flash::ptx_exp2(acc_reduction_view(mi, ni) * softmax_scale_log2 - max_scaled); + } + // s_P2 = max(P_block)/6 — per-block scale factor from φ function (same formula for both modes) + // The difference is in max_scaled: two-level includes 448×6 pre-scaling, single-level doesn't + CUTLASS_PRAGMA_UNROLL + for (int sfi = 0; sfi < size<1>(AbsMaxP); sfi++) { + AbsMaxP(mi, sfi) = flash::ptx_exp2(AbsMaxP(mi, sfi) * softmax_scale_log2 - max_scaled + fp4_scale_log2); + } + } + CUTLASS_PRAGMA_UNROLL + for (int mi = 0; mi < size<0>(acc_reduction_view); mi++) { + CUTLASS_PRAGMA_UNROLL + for (int ni = 0; ni < size<1>(acc_reduction_view); ni++) { + row_sum(mi) += acc_reduction_view(mi, ni); + } + } + } + else { + Tensor scores_max_prev = make_fragment_like(row_max); + cute::copy(row_max, scores_max_prev); + CUTLASS_PRAGMA_UNROLL + for (int mi = 0; mi < size<0>(acc_reduction_view); mi++) { + CUTLASS_PRAGMA_UNROLL + for (int ni = 0; ni < size<1, 1>(acc_reduction_view); ni++) { + float local_max = -INFINITY; + CUTLASS_PRAGMA_UNROLL + for (int ei = 0; ei < size<1, 0>(acc_reduction_view); ei++) { + local_max = fmaxf(local_max, acc_reduction_view(mi, make_coord(ei, ni))); + } + float max_recv = __shfl_xor_sync(int32_t(-1), local_max, 1); // exchange max with neighbour thread of 8 elements + AbsMaxP(mi, ni) = fmaxf(local_max, max_recv); + row_max(mi) = fmaxf(row_max(mi), AbsMaxP(mi, ni)); + } + + float max_recv = __shfl_xor_sync(int32_t(-1), row_max(mi), 2); // exchange max in a quad in a row + row_max(mi) = fmaxf(row_max(mi), max_recv); + + float scores_max_cur = !InfCheck + ? row_max(mi) + : (row_max(mi) == -INFINITY ? 0.0f : row_max(mi)); + scores_scale(mi) = flash::ptx_exp2((scores_max_prev(mi) - scores_max_cur) * softmax_scale_log2); + + // Two-level P quantization (default): s_P1 = rowmax(P̃)/(448×6), then s_P2,P̂_2 = φ(P̃/s_P1) + // Single-level P quantization: s_P2, P̂_2 = φ(P̃) directly (like V quantization) + const float s_P1_offset = single_level_p_quant ? 0.f : fp8_scalexfp4_scale_log2; + const float max_scaled = InfCheck + ? (row_max(mi) == -INFINITY ? 0.f : (row_max(mi) * softmax_scale_log2 + s_P1_offset)) + : (row_max(mi) * softmax_scale_log2 + s_P1_offset); + row_sum(mi) = row_sum(mi) * scores_scale(mi); + CUTLASS_PRAGMA_UNROLL + for (int ni = 0; ni < size<1>(acc_reduction_view); ni++) { + acc_reduction_view(mi, ni) = flash::ptx_exp2(acc_reduction_view(mi, ni) * softmax_scale_log2 - max_scaled); + row_sum(mi) += acc_reduction_view(mi, ni); + } + // s_P2 = max(P_block)/6 — per-block scale factor from φ function + CUTLASS_PRAGMA_UNROLL + for (int sfi = 0; sfi < size<1>(AbsMaxP); sfi++) { + AbsMaxP(mi, sfi) = flash::ptx_exp2(AbsMaxP(mi, sfi) * softmax_scale_log2 - max_scaled + fp4_scale_log2); + } + // scores_scale(mi) = max_scaled; + } + } + CUTLASS_PRAGMA_UNROLL + for (int i = 0; i < size(AbsMaxP); ++i) { + CUTLASS_PRAGMA_UNROLL + for (int j = 0; j < size<0>(acc_conversion_flatten); ++j) + acc_conversion_flatten(j, i) /= AbsMaxP(i); + } + } + + template + CUTLASS_DEVICE void finalize(TensorAcc& o_store) { + Tensor o_store_reduction_view = make_tensor(o_store.data(), flash::convert_to_reduction_layout(o_store.layout())); + CUTLASS_PRAGMA_UNROLL + for (int mi = 0; mi < size(row_max); ++mi) { + CUTLASS_PRAGMA_UNROLL + for (int i = 1; i < RowReductionThr; i <<= 1) { + float sum_recv = __shfl_xor_sync(int32_t(-1), row_sum(mi), i); + row_sum(mi) += sum_recv; + } + float sum = row_sum(mi); + float inv_sum = (sum == 0.f || sum != sum) ? 0.f : 1 / sum; + CUTLASS_PRAGMA_UNROLL + for (int ni = 0; ni < size<1>(o_store_reduction_view); ++ni) { + o_store_reduction_view(mi, ni) *= inv_sum; + } + } + } + + template + CUTLASS_DEVICE void rescale_o(TensorAcc& o_store, TensorAcc const& o_tmp) { + Tensor o_store_reduction_view = make_tensor(o_store.data(), flash::convert_to_reduction_layout(o_store.layout())); + Tensor o_tmp_reduction_view = make_tensor(o_tmp.data(), flash::convert_to_reduction_layout(o_tmp.layout())); + CUTLASS_PRAGMA_UNROLL + for (int mi = 0; mi < size(row_max); ++mi) { + CUTLASS_PRAGMA_UNROLL + for (int ni = 0; ni < size<1>(o_store_reduction_view); ++ni) { + o_store_reduction_view(mi, ni) = o_store_reduction_view(mi, ni) * scores_scale(mi) + o_tmp_reduction_view(mi, ni); + } + } + + } + + + }; + } // namespace flash \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/blackwell/static_switch.h b/fastvideo/attention/backends/sageattn/blackwell/static_switch.h new file mode 100644 index 000000000..e870643e7 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/static_switch.h @@ -0,0 +1,83 @@ +// Inspired by +// https://github.com/NVIDIA/DALI/blob/main/include/dali/core/static_switch.h +// and https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/Dispatch.h + +#pragma once + +/// @param COND - a boolean expression to switch by +/// @param CONST_NAME - a name given for the constexpr bool variable. +/// @param ... - code to execute for true and false +/// +/// Usage: +/// ``` +/// BOOL_SWITCH(flag, BoolConst, [&] { +/// some_function(...); +/// }); +/// ``` +// + +#define BOOL_SWITCH(COND, CONST_NAME, ...) \ + [&] { \ + if (COND) { \ + constexpr static bool CONST_NAME = true; \ + return __VA_ARGS__(); \ + } else { \ + constexpr static bool CONST_NAME = false; \ + return __VA_ARGS__(); \ + } \ + }() + +#define PREC_SWITCH(PRECTYPE, ...) \ + [&] { \ + if (PRECTYPE == 1) { \ + using kPrecType = cutlass::half_t; \ + constexpr static bool kSoftFp16 = false; \ + constexpr static bool kHybrid = false; \ + return __VA_ARGS__(); \ + } else if (PRECTYPE == 2) { \ + using kPrecType = cutlass::float_e4m3_t; \ + constexpr static bool kSoftFp16 = false; \ + constexpr static bool kHybrid = false; \ + return __VA_ARGS__(); \ + } else if (PRECTYPE == 3) { \ + using kPrecType = cutlass::float_e4m3_t; \ + constexpr static bool kSoftFp16 = false; \ + constexpr static bool kHybrid = true; \ + return __VA_ARGS__(); \ + } else if (PRECTYPE == 4) { \ + using kPrecType = cutlass::float_e4m3_t; \ + constexpr static bool kSoftFp16 = true; \ + constexpr static bool kHybrid = false; \ + return __VA_ARGS__(); \ + } \ + }() + +#define HEADDIM_SWITCH(HEADDIM, ...) \ + [&] { \ + if (HEADDIM == 64) { \ + constexpr static int kHeadSize = 64; \ + return __VA_ARGS__(); \ + } else if (HEADDIM == 128) { \ + constexpr static int kHeadSize = 128; \ + return __VA_ARGS__(); \ + } else if (HEADDIM == 256) { \ + constexpr static int kHeadSize = 256; \ + return __VA_ARGS__(); \ + } \ + }() + +#define SEQLEN_SWITCH(USE_VAR_SEQ_LEN, SEQ_LEN_OUT_OF_BOUND_CHECK, ...) \ + [&] { \ + if (!USE_VAR_SEQ_LEN) { \ + if (SEQ_LEN_OUT_OF_BOUND_CHECK) { \ + using kSeqLenTraitsType = FixedSeqLenTraits; \ + return __VA_ARGS__(); \ + } else { \ + using kSeqLenTraitsType = FixedSeqLenTraits; \ + return __VA_ARGS__(); \ + } \ + } else { \ + using kSeqLenTraitsType = VarSeqLenTraits; \ + return __VA_ARGS__(); \ + } \ + }() diff --git a/fastvideo/attention/backends/sageattn/blackwell/tile_scheduler.h b/fastvideo/attention/backends/sageattn/blackwell/tile_scheduler.h new file mode 100644 index 000000000..a559bef07 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/tile_scheduler.h @@ -0,0 +1,304 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * This code is based on code from FlashAttention3, https://github.com/Dao-AILab/flash-attention + * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#include "cutlass/fast_math.h" + +namespace flash { + +/////////////////////////////////////////////////////////////////////////////// + +class StaticPersistentTileSchedulerOld { + // + // Data members + // + +private: + int current_work_linear_idx_; + cutlass::FastDivmod const &m_block_divmod, &head_divmod; + int const total_blocks; + +public: + struct WorkTileInfo { + int M_idx = 0; + int H_idx = 0; + int B_idx = 0; + bool is_valid_tile = false; + + CUTLASS_HOST_DEVICE + bool + is_valid() const { + return is_valid_tile; + } + + CUTLASS_HOST_DEVICE + static WorkTileInfo + invalid_work_tile() { + return {-1, -1, -1, false}; + } + + }; + +public: + + CUTLASS_DEVICE explicit StaticPersistentTileSchedulerOld(cutlass::FastDivmod const &m_block_divmod_, + cutlass::FastDivmod const &head_divmod_, + int const total_blocks_) : + m_block_divmod(m_block_divmod_), head_divmod(head_divmod_), total_blocks(total_blocks_) { + + // MSVC requires protecting use of CUDA-specific nonstandard syntax, + // like blockIdx and gridDim, with __CUDA_ARCH__. +#if defined(__CUDA_ARCH__) + // current_work_linear_idx_ = blockIdx.x + blockIdx.y * gridDim.x + blockIdx.z * gridDim.x * gridDim.y; + current_work_linear_idx_ = blockIdx.x; +#else + CUTLASS_ASSERT(false && "This line should never be reached"); +#endif + } + + CUTLASS_DEVICE + WorkTileInfo + get_current_work() const { + return get_current_work_for_linear_idx(current_work_linear_idx_); + } + + CUTLASS_DEVICE + WorkTileInfo + get_current_work_for_linear_idx(int linear_idx) const { + if (linear_idx >= total_blocks) { + return WorkTileInfo::invalid_work_tile(); + } + + // Map worker's linear index into the CTA tiled problem shape to the corresponding MHB indices + int M_idx, H_idx, B_idx; + int quotient = m_block_divmod.divmod(M_idx, linear_idx); + B_idx = head_divmod.divmod(H_idx, quotient); + return {M_idx, H_idx, B_idx, true}; + } + + CUTLASS_DEVICE + void + // advance_to_next_work(int advance_count = 1) { + advance_to_next_work() { + // current_work_linear_idx_ += int(gridDim.x * gridDim.y * gridDim.z); + current_work_linear_idx_ += int(gridDim.x); + } + + CUTLASS_DEVICE + WorkTileInfo + fetch_next_work() { + WorkTileInfo new_work_tile_info; + advance_to_next_work(); + new_work_tile_info = get_current_work(); + return new_work_tile_info; + } + +}; + +/////////////////////////////////////////////////////////////////////////////// + +class SingleTileScheduler { + +public: + + // Host side kernel arguments + struct Arguments { + int const num_blocks_m, num_head, num_batch; + int const* tile_count_semaphore = nullptr; + }; + + // Device side kernel params + struct Params {}; + + static Params + to_underlying_arguments(Arguments const& args) { + return {}; + } + + static dim3 + get_grid_dim(Arguments const& args, int num_sm) { + return {uint32_t(args.num_blocks_m), uint32_t(args.num_head), uint32_t(args.num_batch)}; + } + + struct WorkTileInfo { + int M_idx = 0; + int H_idx = 0; + int B_idx = 0; + bool is_valid_tile = false; + + CUTLASS_DEVICE + bool + is_valid(Params const& params) const { + return is_valid_tile; + } + + CUTLASS_DEVICE + cute::tuple + get_block_coord(Params const& params) const { + return {M_idx, H_idx, B_idx}; + } + + CUTLASS_DEVICE + WorkTileInfo + get_next_work(Params const& params) const { + return {-1, -1, -1, false}; + } + + }; + + CUTLASS_DEVICE + WorkTileInfo + get_initial_work() const { + return {int(blockIdx.x), int(blockIdx.y), int(blockIdx.z), true}; + } + + CUTLASS_DEVICE + WorkTileInfo + get_next_work(Params const& params, WorkTileInfo const& current_work) const { + return {-1, -1, -1, false}; + } + +}; + +/////////////////////////////////////////////////////////////////////////////// + +class StaticPersistentTileScheduler { + +public: + + // Host side kernel arguments + struct Arguments { + int const num_blocks_m, num_head, num_batch; + int const* tile_count_semaphore = nullptr; + }; + + // Device side kernel params + struct Params { + int total_blocks; + cutlass::FastDivmod m_block_divmod, head_divmod; + }; + + static Params + to_underlying_arguments(Arguments const& args) { + return {args.num_blocks_m * args.num_head * args.num_batch, + cutlass::FastDivmod(args.num_blocks_m), cutlass::FastDivmod(args.num_head)}; + } + + static dim3 + get_grid_dim(Arguments const& args, int num_sm) { + return {uint32_t(num_sm)}; + } + + struct WorkTileInfo { + int tile_idx; + + CUTLASS_DEVICE + bool + is_valid(Params const& params) const { + return tile_idx < params.total_blocks; + } + + CUTLASS_DEVICE + cute::tuple + get_block_coord(Params const& params) const { + int m_block, bidh, bidb; + bidb = params.head_divmod.divmod(bidh, params.m_block_divmod.divmod(m_block, tile_idx)); + return {m_block, bidh, bidb}; + } + + }; + + CUTLASS_DEVICE + WorkTileInfo + get_initial_work() const { + return {int(blockIdx.x)}; + } + + CUTLASS_DEVICE + WorkTileInfo + get_next_work(Params const& params, WorkTileInfo const& current_work) const { + return {current_work.tile_idx + int(gridDim.x)}; + } + +}; + +class DynamicPersistentTileScheduler { + +public: + + // Host side kernel arguments + struct Arguments { + int const num_blocks_m, num_head, num_batch; + int const* tile_count_semaphore; + }; + + // Device side kernel params + struct Params { + int const total_blocks; + cutlass::FastDivmod const m_block_divmod, head_divmod; + int const* tile_count_semaphore; + }; + + static Params + to_underlying_arguments(Arguments const& args) { + return {args.num_blocks_m * args.num_head * args.num_batch, + cutlass::FastDivmod(args.num_blocks_m), cutlass::FastDivmod(args.num_head), + args.tile_count_semaphore}; + } + + static dim3 + get_grid_dim(Arguments const& args, int num_sm) { + return {uint32_t(num_sm)}; + } + + using WorkTileInfo = StaticPersistentTileScheduler::WorkTileInfo; + // struct WorkTileInfo { + // int tile_idx; + + // CUTLASS_DEVICE + // bool + // is_valid(Params const& params) const { + // return tile_idx < params.total_blocks; + // } + + // CUTLASS_DEVICE + // cute::tuple + // get_block_coord(Params const& params) const { + // int m_block, bidh, bidb; + // bidb = params.head_divmod.divmod(bidh, params.m_block_divmod.divmod(m_block, tile_idx)); + // return {m_block, bidh, bidb}; + // } + + // }; + + CUTLASS_DEVICE + WorkTileInfo + get_initial_work() const { + return {int(blockIdx.x)}; + } + + CUTLASS_DEVICE + WorkTileInfo + get_next_work(Params const& params, WorkTileInfo const& current_work) const { + return {current_work.tile_idx + int(gridDim.x)}; + } + +}; + +} // flash diff --git a/fastvideo/attention/backends/sageattn/blackwell/utils.h b/fastvideo/attention/backends/sageattn/blackwell/utils.h new file mode 100644 index 000000000..793270196 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/blackwell/utils.h @@ -0,0 +1,408 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#include +#include +#include + +#include + +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800 +#include +#endif + +#include + +#include +#include +#include +#include + +namespace flash { + +using namespace cute; + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +template +struct MaxOp { +__device__ __forceinline__ T operator()(T const & x, T const & y) { return x > y ? x : y; } +}; + +template <> +struct MaxOp { +// This is slightly faster +__device__ __forceinline__ float operator()(float const &x, float const &y) { return max(x, y); } +}; + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +template +struct SumOp { +__device__ __forceinline__ T operator()(T const & x, T const & y) { return x + y; } +}; + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +template +struct Allreduce { + static_assert(THREADS == 32 || THREADS == 16 || THREADS == 8 || THREADS == 4); + template + static __device__ __forceinline__ T run(T x, Operator &op) { + constexpr int OFFSET = THREADS / 2; + x = op(x, __shfl_xor_sync(uint32_t(-1), x, OFFSET)); + return Allreduce::run(x, op); + } +}; + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +template<> +struct Allreduce<2> { +template +static __device__ __forceinline__ T run(T x, Operator &op) { + x = op(x, __shfl_xor_sync(uint32_t(-1), x, 1)); + return x; +} +}; + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +template +__device__ __forceinline__ void thread_reduce_(Tensor const &tensor, Tensor &summary, Operator &op) { + static_assert(Layout0::rank == 2, "Only support 2D Tensor"); + static_assert(Layout1::rank == 1, "Only support 1D Tensor"); + CUTE_STATIC_ASSERT_V(size<0>(summary) == size<0>(tensor)); + #pragma unroll + for (int mi = 0; mi < size<0>(tensor); mi++) { + summary(mi) = zero_init ? tensor(mi, 0) : op(summary(mi), tensor(mi, 0)); + #pragma unroll + for (int ni = 1; ni < size<1>(tensor); ni++) { + summary(mi) = op(summary(mi), tensor(mi, ni)); + } + } +} + +template +__device__ __forceinline__ void quad_allreduce_(Tensor &dst, Tensor &src, Operator &op) { + CUTE_STATIC_ASSERT_V(size(dst) == size(src)); + #pragma unroll + for (int i = 0; i < size(dst); i++){ + dst(i) = Allreduce<4>::run(src(i), op); + } +} + +template +__device__ __forceinline__ void reduce_(Tensor const& tensor, Tensor &summary, Operator &op) { + thread_reduce_(tensor, summary, op); + quad_allreduce_(summary, summary, op); +} + +template +__device__ __forceinline__ void reduce_max(Tensor const& tensor, Tensor &max){ + MaxOp max_op; + reduce_(tensor, max, max_op); +} + +template +__device__ __forceinline__ void reduce_sum(Tensor const& tensor, Tensor &sum){ + SumOp sum_op; + thread_reduce_(tensor, sum, sum_op); + if constexpr (warp_reduce) { quad_allreduce_(sum, sum, sum_op); } +} + +__forceinline__ __device__ __half2 half_exp(__half2 x) { + uint32_t tmp_out, tmp_in; + tmp_in = reinterpret_cast(x); + asm ("ex2.approx.f16x2 %0, %1;\n" + : "=r"(tmp_out) + : "r"(tmp_in)); + __half2 out = reinterpret_cast<__half2&>(tmp_out); + return out; +} + +// Apply the exp to all the elements. +template +__forceinline__ __device__ void max_scale_exp2_sum(Tensor &tensor, Tensor &max, Tensor &sum, const float scale) { + static_assert(Layout0::rank == 2, "Only support 2D Tensor"); static_assert(Layout1::rank == 1, "Only support 1D Tensor"); CUTE_STATIC_ASSERT_V(size<0>(max) == size<0>(tensor)); + #pragma unroll + for (int mi = 0; mi < size<0>(tensor); ++mi) { + MaxOp max_op; + max(mi) = zero_init ? tensor(mi, 0) : max_op(max(mi), tensor(mi, 0)); + #pragma unroll + for (int ni = 1; ni < size<1>(tensor); ni++) { + max(mi) = max_op(max(mi), tensor(mi, ni)); + } + max(mi) = Allreduce<4>::run(max(mi), max_op); + // If max is -inf, then all elements must have been -inf (possibly due to masking). + // We don't want (-inf - (-inf)) since that would give NaN. + const float max_scaled = max(mi) == -INFINITY ? 0.f : max(mi) * scale; + sum(mi) = 0; + #pragma unroll + for (int ni = 0; ni < size<1>(tensor); ++ni) { + // Instead of computing exp(x - max), we compute exp2(x * log_2(e) - + // max * log_2(e)) This allows the compiler to use the ffma + // instruction instead of fadd and fmul separately. + tensor(mi, ni) = exp2f(tensor(mi, ni) * scale - max_scaled); + sum(mi) += tensor(mi, ni); + } + } +} + +// Apply the exp to all the elements. +template +__forceinline__ __device__ void scale_apply_exp2(Tensor &tensor, Tensor const &max, const float scale) { + static_assert(Layout0::rank == 2, "Only support 2D Tensor"); + static_assert(Layout1::rank == 1, "Only support 1D Tensor"); + CUTE_STATIC_ASSERT_V(size<0>(max) == size<0>(tensor)); + #pragma unroll + for (int mi = 0; mi < size<0>(tensor); ++mi) { + // If max is -inf, then all elements must have been -inf (possibly due to masking). + // We don't want (-inf - (-inf)) since that would give NaN. + // If we don't have float around M_LOG2E the multiplication is done in fp64. + const float max_scaled = Check_inf + ? (max(mi) == -INFINITY ? 0.f : (max(mi) * (Scale_max ? scale : float(M_LOG2E)))) + : (max(mi) * (Scale_max ? scale : float(M_LOG2E))); + #pragma unroll + for (int ni = 0; ni < size<1>(tensor); ++ni) { + // Instead of computing exp(x - max), we compute exp2(x * log_2(e) - + // max * log_2(e)) This allows the compiler to use the ffma + // instruction instead of fadd and fmul separately. + tensor(mi, ni) = exp2f(tensor(mi, ni) * scale - max_scaled); + } + } +} + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +__forceinline__ __device__ float ptx_exp2(float x) { + float y; + asm volatile("ex2.approx.ftz.f32 %0, %1;" : "=f"(y) : "f"(x)); + return y; +} + +CUTLASS_DEVICE void +packed_float_to_ue4m3( + float const &f0, float const &f1, float const &f2, float const &f3, + uint32_t &out +) { + asm volatile( \ + "{\n" \ + ".reg .b16 lo;\n" \ + ".reg .b16 hi;\n" \ + "cvt.rn.satfinite.e4m3x2.f32 lo, %2, %1;\n" \ + "cvt.rn.satfinite.e4m3x2.f32 hi, %4, %3;\n" \ + "mov.b32 %0, {lo, hi};\n" \ + "}" \ + : "=r"(out) : "f"(f0), "f"(f1), "f"(f2), "f"(f3)); +} + +CUTLASS_DEVICE void +packed_float_to_e2m1( + float const &f0, float const &f1, float const &f2, float const& f3, + float const &f4, float const &f5, float const &f6, float const& f7, + uint32_t &out +) { + + asm volatile( \ + "{\n" \ + ".reg .b8 byte0;\n" \ + ".reg .b8 byte1;\n" \ + ".reg .b8 byte2;\n" \ + ".reg .b8 byte3;\n" \ + "cvt.rn.satfinite.e2m1x2.f32 byte0, %2, %1;\n" \ + "cvt.rn.satfinite.e2m1x2.f32 byte1, %4, %3;\n" \ + "cvt.rn.satfinite.e2m1x2.f32 byte2, %6, %5;\n" \ + "cvt.rn.satfinite.e2m1x2.f32 byte3, %8, %7;\n" \ + "mov.b32 %0, {byte0, byte1, byte2, byte3};\n" \ + "}" \ + : "=r"(out) : "f"(f0), "f"(f1), "f"(f2), "f"(f3), + "f"(f4), "f"(f5), "f"(f6), "f"(f7)); + +} + +CUTLASS_DEVICE void +add(float2 & c, + float2 const& a, + float2 const& b) +{ +asm volatile("add.f32x2 %0, %1, %2;\n" + : "=l"(reinterpret_cast(c)) + : "l"(reinterpret_cast(a)), + "l"(reinterpret_cast(b))); +} + +CUTLASS_DEVICE void +add_inplace(float2 &a, + float2 const& b) +{ + asm volatile("add.f32x2 %0, %0, %1;\n" + : "+l"(reinterpret_cast(a)) // a: input/output + : "l"(reinterpret_cast(b)) // b: input + ); +} + + +CUTLASS_DEVICE void +sub(float2 & c, + float2 const& a, + float2 const& b) +{ +asm volatile("sub.f32x2 %0, %1, %2;\n" + : "=l"(reinterpret_cast(c)) + : "l"(reinterpret_cast(a)), + "l"(reinterpret_cast(b))); +} + +CUTLASS_DEVICE void +sub_inplace(float2 &a, + float2 const& b) +{ + asm volatile("sub.f32x2 %0, %0, %1;\n" + : "+l"(reinterpret_cast(a)) // a: input/output + : "l"(reinterpret_cast(b)) // b: input + ); +} + + +CUTLASS_DEVICE void +mul(float2 & c, + float2 const& a, + float2 const& b) +{ + asm volatile("mul.f32x2 %0, %1, %2;\n" + : "=l"(reinterpret_cast(c)) + : "l"(reinterpret_cast(a)), + "l"(reinterpret_cast(b))); +} + +CUTLASS_DEVICE void +fma(float2 & d, + float2 const& a, + float2 const& b, + float2 const& c) +{ + asm volatile("fma.rn.f32x2 %0, %1, %2, %3;\n" + : "=l"(reinterpret_cast(d)) + : "l"(reinterpret_cast(a)), + "l"(reinterpret_cast(b)), + "l"(reinterpret_cast(c))); +} + +CUTLASS_DEVICE void +fma_inplace(float2 &a, + float2 const& b, + float2 const& c) +{ + asm volatile("fma.rn.f32x2 %0, %0, %1, %2;\n" + : "+l"(reinterpret_cast(a)) + : "l"(reinterpret_cast(b)), + "l"(reinterpret_cast(c))); +} + +//////////////////////////////////////////////////////////////////////////////////////////////////// + +template < + class Layout +> +CUTLASS_DEVICE constexpr +auto convert_to_reduction_layout(Layout mma_layout) { + static_assert(rank(mma_layout) == 3, "Mma Layout should be (MmaAtom, MmaM, MmaN)"); + static_assert(rank(get<0>(shape(mma_layout))) == 2, "MmaAtom should be (AtomN, AtomM)"); + + return make_layout( + make_layout(get<0,1>(mma_layout), get<1>(mma_layout)), + make_layout(get<0,0>(mma_layout), get<2>(mma_layout)) + ); +} + +template < + class Tensor +> +CUTLASS_DEVICE constexpr +auto convert_to_reduction_tensor(Tensor mma_tensor) { + return make_tensor(mma_tensor.data(), convert_to_reduction_layout(mma_tensor.layout())); +} + + +template < + class Layout +> +CUTLASS_DEVICE constexpr +auto convert_to_conversion_layout(Layout mma_layout) { + static_assert(rank(mma_layout) == 3, "Mma Layout should be (MmaAtom, MmaM, MmaN)"); + static_assert(rank(get<0>(shape(mma_layout))) == 2, "MmaAtom should be (AtomN, AtomM)"); + + constexpr int MmaAtomN = size<0, 0>(mma_layout); + constexpr int MmaAtomM = size<0, 1>(mma_layout); + constexpr int MmaM = size<1>(mma_layout); + constexpr int MmaN = size<2>(mma_layout); + + static_assert(MmaAtomN == 8, "MmaAtomN should be 8."); + static_assert(MmaAtomM == 2, "MmaAtomM should be 2."); + static_assert(MmaN % 2 == 0, "MmaN should be multiple of 2."); + + auto mma_n_division = zipped_divide( + layout<2>(mma_layout), make_tile(_2{}) + ); + return make_layout( + make_layout(layout<0,0>(mma_layout), make_layout(layout<0,1>(mma_layout), layout<0>(mma_n_division))), + layout<1>(mma_layout), layout<1>(mma_n_division) + ); +} + +template < + class Tensor +> +CUTLASS_DEVICE constexpr +auto convert_to_conversion_tensor(Tensor mma_tensor) { + return make_tensor(mma_tensor.data(), convert_to_conversion_layout(mma_tensor.layout())); +} +//////////////////////////////////////////////////////////////////////////////////////////////////// + +template +CUTLASS_DEVICE void copy(TiledCopy tiled_copy, Tensor const &S, + Tensor &D, Tensor const &identity_MN, + Tensor const &predicate_K, const int max_MN=0) { + CUTE_STATIC_ASSERT_V(rank(S) == Int<3>{}); + CUTE_STATIC_ASSERT_V(rank(D) == Int<3>{}); + CUTE_STATIC_ASSERT_V(size<0>(S) == size<0>(D)); // MMA + CUTE_STATIC_ASSERT_V(size<1>(S) == size<1>(D)); // MMA_M + CUTE_STATIC_ASSERT_V(size<2>(S) == size<2>(D)); // MMA_K + // There's no case where !Clear_OOB_K && Clear_OOB_MN + static_assert(!(Clear_OOB_MN && !Clear_OOB_K)); + #pragma unroll + for (int m = 0; m < size<1>(S); ++m) { + if (Is_even_MN || get<0>(identity_MN(0, m, 0)) < max_MN) { + #pragma unroll + for (int k = 0; k < size<2>(S); ++k) { + if (Is_even_K || predicate_K(k)) { + cute::copy(tiled_copy, S(_, m, k), D(_, m, k)); + } else if (Clear_OOB_K) { + cute::clear(D(_, m, k)); + } + } + } else if (Clear_OOB_MN) { + cute::clear(D(_, m, _)); + } + } +} + +} // namespace flash diff --git a/fastvideo/attention/backends/sageattn/quantization/__init__.py b/fastvideo/attention/backends/sageattn/quantization/__init__.py new file mode 100644 index 000000000..5d46677a1 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/quantization/__init__.py @@ -0,0 +1 @@ +__version__ = "3.0.0.b1" \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_k.py b/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_k.py new file mode 100644 index 000000000..94c10681c --- /dev/null +++ b/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_k.py @@ -0,0 +1,90 @@ +""" +Copyright (c) 2025 by SageAttention team. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +""" +import torch +import fp4quant +from triton.tools.mxfp import MXFP4Tensor + +from bench_utils import bench_kineto +b = 1 +h = 32 +n = 16384 +d = 128 + +def test(): + q = torch.randn((b, h, n, d), device="cuda", dtype=torch.float16) + o = torch.empty((b, h, n, d // 2), device="cuda", dtype=torch.uint8) + o_s = torch.empty((b, h, n, d // 16), device="cuda", dtype=torch.float8_e4m3fn) + fp4quant.scaled_fp4_quant_permute(q, o, o_s, 1) + +test() + +t = bench_kineto(test, "scaled_fp4_quant_kernel", suppress_kineto_output=True) + +IO = b * h * n * d * 2 + b * h * n * d * 0.5 + b * h * n * d // 16 * 1 +throughput = IO / t * 1e-9 + +print(f"Throughput: {throughput:.2f} GB/s") + +def scale_and_fp4_tensor(x: torch.Tensor, packed_dim: int = 3, all_ones: bool = False, permuted: bool = False): + assert x.is_contiguous() and x.ndim == 4 and x.shape[-1] % 16 == 0 + B, H, M, N = x.shape + x = x.view(B, H, M, N // 16, 16) + scales = (x.abs().amax(dim=-1, keepdim=True) / 6).to(torch.float32) + if all_ones: + scales = torch.ones_like(scales) + x_scaled = x / scales + packed_fp4 = MXFP4Tensor(x_scaled.flatten(start_dim=-2)).to_packed_tensor(dim=packed_dim) + dequant_x = (MXFP4Tensor(x_scaled).to(torch.float32) * scales.to(torch.float8_e4m3fn).to(torch.float32)).flatten(start_dim=-2) + fp8_scale = scales.flatten(start_dim=-2).to(torch.float8_e4m3fn) + permuted_fp8_scale = None + if permuted: + scales = scales.view(B, H // 64, 4, 16, M, N // 16).permute(0, 1, 3, 2, 4, 5).reshape(B, H, M, N // 16) + permuted_fp8_scale = scales.view(B, H // 64, 64, M, N // 64, 4).permute(0, 1, 4, 3, 2, 5).reshape(B, H, M, N // 16).to(torch.float8_e4m3fn) + return fp8_scale, packed_fp4, dequant_x, permuted_fp8_scale + +b = 2 +h = 4 +n = 251 +n_padded = (n + 127) // 128 * 128 +d = 128 + +q = torch.randn(b, h, n, d, dtype=torch.float16, device='cuda') +o = torch.empty((b, h, n, d // 2), dtype=torch.uint8, device='cuda') +o_s = torch.empty((b, h, n, d // 16), dtype=torch.float8_e4m3fn, device='cuda') + +fp4quant.scaled_fp4_quant(q, o, o_s, 1) + +k_permute = [0, 1, 8, 9, 16, 17, 24, 25, 2, 3, 10, 11, 18, 19, 26, 27, 4, 5, 12, 13, 20, 21, 28, 29, 6, 7, 14, 15, 22, 23, 30, 31] +o_permuted = torch.empty((b, h, n_padded, d // 2), dtype=torch.uint8, device='cuda') +o_s_permuted = torch.empty((b, h, n_padded, d // 16), dtype=torch.float8_e4m3fn, device='cuda') +fp4quant.scaled_fp4_quant_permute(q, o_permuted, o_s_permuted, 1) + +# padding +if n % 128 != 0: + o_permuted_gt = torch.cat([o, torch.zeros((b, h, n_padded - n, d // 2), dtype=torch.uint8, device='cuda')], dim=2) + o_s_permuted_gt = torch.cat([o_s, torch.zeros((b, h, n_padded - n, d // 16), dtype=torch.float8_e4m3fn, device='cuda')], dim=2) +else: + o_permuted_gt = o + o_s_permuted_gt = o_s + +# use scale_and_fp4_tensor + torch permutation to get the ground truth +o_permuted_gt = o_permuted_gt.reshape(b, h, n_padded // 32, 32, d // 2)[:, :, :, k_permute, :].reshape(b, h, n_padded, d // 2) +o_s_permuted_gt = o_s_permuted_gt.reshape(b, h, n_padded // 32, 32, d // 16)[:, :, :, k_permute, :].reshape(b, h, n_padded, d // 16) + +assert((o_permuted - o_permuted_gt).abs().max() == 0) +assert((o_s_permuted.float() - o_s_permuted_gt.float()).abs().max() == 0) + +print("All tests passed!") \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_q.py b/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_q.py new file mode 100644 index 000000000..ffbab0a7b --- /dev/null +++ b/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_q.py @@ -0,0 +1,86 @@ +""" +Copyright (c) 2025 by SageAttention team. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +""" +import torch +import fp4quant +from triton.tools.mxfp import MXFP4Tensor + +from bench_utils import bench_kineto +b = 1 +h = 32 +n = 16384 +d = 128 + +def test(): + q = torch.randn((b, h, n, d), device="cuda", dtype=torch.float16) + o = torch.empty((b, h, n, d // 2), device="cuda", dtype=torch.uint8) + o_s = torch.empty((b, h, n, d // 16), device="cuda", dtype=torch.float8_e4m3fn) + fp4quant.scaled_fp4_quant(q, o, o_s, 1) + +test() + +t = bench_kineto(test, "scaled_fp4_quant_kernel", suppress_kineto_output=True) + +IO = b * h * n * d * 2 + b * h * n * d * 0.5 + b * h * n * d // 16 * 1 +throughput = IO / t * 1e-9 + +print(f"Throughput: {throughput:.2f} GB/s") + +def scale_and_fp4_tensor(x: torch.Tensor, packed_dim: int = 3, all_ones: bool = False, permuted: bool = False): + assert x.is_contiguous() and x.ndim == 4 and x.shape[-1] % 16 == 0 + B, H, M, N = x.shape + x = x.view(B, H, M, N // 16, 16) + scales = (x.abs().amax(dim=-1, keepdim=True) / 6).to(torch.float32) + if all_ones: + scales = torch.ones_like(scales) + x_scaled = x / scales + packed_fp4 = MXFP4Tensor(x_scaled.flatten(start_dim=-2)).to_packed_tensor(dim=packed_dim) + dequant_x = (MXFP4Tensor(x_scaled).to(torch.float32) * scales.to(torch.float8_e4m3fn).to(torch.float32)).flatten(start_dim=-2) + fp8_scale = scales.flatten(start_dim=-2).to(torch.float8_e4m3fn) + permuted_fp8_scale = None + if permuted: + scales = scales.view(B, H // 64, 4, 16, M, N // 16).permute(0, 1, 3, 2, 4, 5).reshape(B, H, M, N // 16) + permuted_fp8_scale = scales.view(B, H // 64, 64, M, N // 64, 4).permute(0, 1, 4, 3, 2, 5).reshape(B, H, M, N // 16).to(torch.float8_e4m3fn) + return fp8_scale, packed_fp4, dequant_x, permuted_fp8_scale + +b = 2 +h = 4 +n = 251 +d = 128 + +q = torch.randn(b, h, n, d, dtype=torch.float16, device='cuda') +o = torch.empty((b, h, n, d // 2), dtype=torch.uint8, device='cuda') +o_s = torch.empty((b, h, n, d // 16), dtype=torch.float8_e4m3fn, device='cuda') + +fp4quant.scaled_fp4_quant(q, o, o_s, 1) + +fp8_scale, packed_fp4, dequant_x, permuted_fp8_scale = scale_and_fp4_tensor(q, packed_dim=3) + +assert((fp8_scale.float() - o_s.float()).abs().max() == 0) + +o_binary = [ + (int(bin_str[:4], 2), int(bin_str[4:], 2)) + for bin_str in [format(x.item(), '08b') for x in o.view(-1)] +] +o_binary_gt = [ + (int(bin_str[:4], 2), int(bin_str[4:], 2)) + for bin_str in [format(x.item(), '08b') for x in packed_fp4.view(-1)] +] +for i in range(len(o_binary)): + # check contiguous 4 bits. Difference should be at most one + assert(abs(o_binary[i][0] - o_binary_gt[i][0]) <= 1) + assert(abs(o_binary[i][1] - o_binary_gt[i][1]) <= 1) + +print("All tests passed!") \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_v.py b/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_v.py new file mode 100644 index 000000000..24b87b700 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/quantization/bench/bench_quant_v.py @@ -0,0 +1,86 @@ +""" +Copyright (c) 2025 by SageAttention team. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +""" +import torch +import fp4quant +from triton.tools.mxfp import MXFP4Tensor + +from bench_utils import bench_kineto +b = 1 +h = 32 +n = 16384 +d = 128 + +def test(): + q = torch.randn((b, h, n, d), device="cuda", dtype=torch.float16) + o = torch.empty((b, h, d, n // 2), device="cuda", dtype=torch.uint8) + o_s = torch.empty((b, h, d, n // 16), device="cuda", dtype=torch.float8_e4m3fn) + fp4quant.scaled_fp4_quant_trans(q, o, o_s, 1) + +test() + +t = bench_kineto(test, "scaled_fp4_quant_trans_kernel", suppress_kineto_output=True) + +IO = b * h * n * d * 2 + b * h * n * d * 0.5 + b * h * n * d // 16 * 1 +throughput = IO / t * 1e-9 + +print(f"Throughput: {throughput:.2f} GB/s") + +def scale_and_fp4_tensor(x: torch.Tensor, packed_dim: int = 3, all_ones: bool = False, permuted: bool = False): + assert x.is_contiguous() and x.ndim == 4 and x.shape[-1] % 16 == 0 + B, H, M, N = x.shape + x = x.view(B, H, M, N // 16, 16) + scales = (x.abs().amax(dim=-1, keepdim=True) / 6).to(torch.float32) + if all_ones: + scales = torch.ones_like(scales) + x_scaled = x / scales + packed_fp4 = MXFP4Tensor(x_scaled.flatten(start_dim=-2)).to_packed_tensor(dim=packed_dim) + dequant_x = (MXFP4Tensor(x_scaled).to(torch.float32) * scales.to(torch.float8_e4m3fn).to(torch.float32)).flatten(start_dim=-2) + fp8_scale = scales.flatten(start_dim=-2).to(torch.float8_e4m3fn) + permuted_fp8_scale = None + if permuted: + scales = scales.view(B, H // 64, 4, 16, M, N // 16).permute(0, 1, 3, 2, 4, 5).reshape(B, H, M, N // 16) + permuted_fp8_scale = scales.view(B, H // 64, 64, M, N // 64, 4).permute(0, 1, 4, 3, 2, 5).reshape(B, H, M, N // 16).to(torch.float8_e4m3fn) + return fp8_scale, packed_fp4, dequant_x, permuted_fp8_scale + +b = 2 +h = 4 +n = 491 +n_padded = (n + 127) // 128 * 128 +d = 128 + +q = torch.randn(b, h, n, d, dtype=torch.float16, device='cuda') +o = torch.empty((b, h, d, n_padded // 2), dtype=torch.uint8, device='cuda') +o_s = torch.empty((b, h, d, n_padded // 16), dtype=torch.float8_e4m3fn, device='cuda') + +fp4quant.scaled_fp4_quant_trans(q, o, o_s, 1) + +if n % 128 != 0: + q_padded = torch.cat([q, torch.zeros((b, h, n_padded - n, d), dtype=torch.float16, device='cuda')], dim=2) +else: + q_padded = q + +# use torch transpose + scaled_fp4_quant to get the ground truth +q_padded = q_padded.transpose(2, 3).reshape(b, h, n_padded, d).contiguous() +o_gt = torch.empty((b, h, n_padded, d // 2), dtype=torch.uint8, device='cuda') +o_s_gt = torch.empty((b, h, n_padded, d // 16), dtype=torch.float8_e4m3fn, device='cuda') +fp4quant.scaled_fp4_quant(q_padded, o_gt, o_s_gt, 1) +o_gt = o_gt.reshape(b, h, d, n_padded // 2).contiguous() +o_s_gt = o_s_gt.reshape(b, h, d, n_padded // 16).contiguous() + +assert((o_s_gt.float() - o_s.float()).abs().max() == 0) +assert((o_gt - o).abs().max() == 0) + +print("All tests passed!") \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/quantization/bench/bench_utils.py b/fastvideo/attention/backends/sageattn/quantization/bench/bench_utils.py new file mode 100644 index 000000000..a7046de42 --- /dev/null +++ b/fastvideo/attention/backends/sageattn/quantization/bench/bench_utils.py @@ -0,0 +1,169 @@ +""" +Copyright (c) 2025 by SageAttention team. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +""" +import os +import sys +import torch +import torch.distributed as dist + + +def bench(fn, num_warmups: int = 5, num_tests: int = 10, + high_precision: bool = False): + # Flush L2 cache with 256 MB data + torch.cuda.synchronize() + cache = torch.empty(int(256e6 // 4), dtype=torch.int, device='cuda') + cache.zero_() + + # Warmup + for _ in range(num_warmups): + fn() + + # Add a large kernel to eliminate the CPU launch overhead + if high_precision: + x = torch.randn((8192, 8192), dtype=torch.float, device='cuda') + y = torch.randn((8192, 8192), dtype=torch.float, device='cuda') + x @ y + + # Testing + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + start_event.record() + for i in range(num_tests): + fn() + end_event.record() + torch.cuda.synchronize() + + return start_event.elapsed_time(end_event) / num_tests + + +class empty_suppress: + def __enter__(self): + return self + + def __exit__(self, *_): + pass + + +class suppress_stdout_stderr: + def __enter__(self): + self.outnull_file = open(os.devnull, 'w') + self.errnull_file = open(os.devnull, 'w') + + self.old_stdout_fileno_undup = sys.stdout.fileno() + self.old_stderr_fileno_undup = sys.stderr.fileno() + + self.old_stdout_fileno = os.dup(sys.stdout.fileno()) + self.old_stderr_fileno = os.dup(sys.stderr.fileno()) + + self.old_stdout = sys.stdout + self.old_stderr = sys.stderr + + os.dup2(self.outnull_file.fileno(), self.old_stdout_fileno_undup) + os.dup2(self.errnull_file.fileno(), self.old_stderr_fileno_undup) + + sys.stdout = self.outnull_file + sys.stderr = self.errnull_file + return self + + def __exit__(self, *_): + sys.stdout = self.old_stdout + sys.stderr = self.old_stderr + + os.dup2(self.old_stdout_fileno, self.old_stdout_fileno_undup) + os.dup2(self.old_stderr_fileno, self.old_stderr_fileno_undup) + + os.close(self.old_stdout_fileno) + os.close(self.old_stderr_fileno) + + self.outnull_file.close() + self.errnull_file.close() + + +def bench_kineto(fn, kernel_names, num_tests: int = 30, suppress_kineto_output: bool = False, + trace_path: str = None, barrier_comm_profiling: bool = False, flush_l2: bool = False): + # Conflict with Nsight Systems + using_nsys = os.environ.get('DG_NSYS_PROFILING', False) + + # For some auto-tuning kernels with prints + fn() + + # Profile + suppress = suppress_stdout_stderr if suppress_kineto_output and not using_nsys else empty_suppress + with suppress(): + schedule = torch.profiler.schedule(wait=0, warmup=1, active=1, repeat=1) if not using_nsys else None + profiler = torch.profiler.profile(activities=[torch.profiler.ProfilerActivity.CUDA], schedule=schedule) if not using_nsys else empty_suppress() + with profiler: + for i in range(2): + # NOTES: use a large kernel and a barrier to eliminate the unbalanced CPU launch overhead + if barrier_comm_profiling: + lhs = torch.randn((8192, 8192), dtype=torch.float, device='cuda') + rhs = torch.randn((8192, 8192), dtype=torch.float, device='cuda') + lhs @ rhs + dist.all_reduce(torch.ones(1, dtype=torch.float, device='cuda')) + for _ in range(num_tests): + if flush_l2: + torch.empty(int(256e6 // 4), dtype=torch.int, device='cuda').zero_() + fn() + + if not using_nsys: + profiler.step() + + # Return 1 if using Nsight Systems + if using_nsys: + return 1 + + # Parse the profiling table + assert isinstance(kernel_names, str) or isinstance(kernel_names, tuple) + is_tupled = isinstance(kernel_names, tuple) + prof_lines = profiler.key_averages().table(sort_by='cuda_time_total', max_name_column_width=100).split('\n') + kernel_names = (kernel_names, ) if isinstance(kernel_names, str) else kernel_names + assert all([isinstance(name, str) for name in kernel_names]) + for name in kernel_names: + assert sum([name in line for line in prof_lines]) == 1, f'Errors of the kernel {name} in the profiling table' + + # Save chrome traces + if trace_path is not None: + profiler.export_chrome_trace(trace_path) + + # Return average kernel times + units = {'ms': 1e3, 'us': 1e6} + kernel_times = [] + for name in kernel_names: + for line in prof_lines: + if name in line: + time_str = line.split()[-2] + for unit, scale in units.items(): + if unit in time_str: + kernel_times.append(float(time_str.replace(unit, '')) / scale) + break + break + return tuple(kernel_times) if is_tupled else kernel_times[0] + + +def calc_diff(x, y): + x, y = x.double(), y.double() + denominator = (x * x + y * y).sum() + sim = 2 * (x * y).sum() / denominator + return 1 - sim + + +def count_bytes(tensors): + total = 0 + for t in tensors: + if isinstance(t, tuple): + total += count_bytes(t) + else: + total += t.numel() * t.element_size() + return total \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/quantization/cuda_utils.h b/fastvideo/attention/backends/sageattn/quantization/cuda_utils.h new file mode 100644 index 000000000..89013d69d --- /dev/null +++ b/fastvideo/attention/backends/sageattn/quantization/cuda_utils.h @@ -0,0 +1,52 @@ +#pragma once + +#include + +#if defined(__HIPCC__) + #define HOST_DEVICE_INLINE __host__ __device__ + #define DEVICE_INLINE __device__ + #define HOST_INLINE __host__ +#elif defined(__CUDACC__) || defined(_NVHPC_CUDA) + #define HOST_DEVICE_INLINE __host__ __device__ __forceinline__ + #define DEVICE_INLINE __device__ __forceinline__ + #define HOST_INLINE __host__ __forceinline__ +#else + #define HOST_DEVICE_INLINE inline + #define DEVICE_INLINE inline + #define HOST_INLINE inline +#endif + +#define CUDA_CHECK(cmd) \ + do { \ + cudaError_t e = cmd; \ + if (e != cudaSuccess) { \ + printf("Failed: Cuda error %s:%d '%s'\n", __FILE__, __LINE__, \ + cudaGetErrorString(e)); \ + exit(EXIT_FAILURE); \ + } \ + } while (0) + +int64_t get_device_attribute(int64_t attribute, int64_t device_id) { + static int value = [=]() { + int device = static_cast(device_id); + if (device < 0) { + CUDA_CHECK(cudaGetDevice(&device)); + } + int value; + CUDA_CHECK(cudaDeviceGetAttribute( + &value, static_cast(attribute), device)); + return static_cast(value); + }(); + + return value; +} + +namespace cuda_utils { + +template +HOST_DEVICE_INLINE constexpr std::enable_if_t, T> +ceil_div(T a, T b) { + return (a + b - 1) / b; +} + +}; // namespace cuda_utils \ No newline at end of file diff --git a/fastvideo/attention/backends/sageattn/quantization/fp4_quantization_4d.cu b/fastvideo/attention/backends/sageattn/quantization/fp4_quantization_4d.cu new file mode 100644 index 000000000..536a362ed --- /dev/null +++ b/fastvideo/attention/backends/sageattn/quantization/fp4_quantization_4d.cu @@ -0,0 +1,629 @@ +/* + * Copyright (c) 2025 by SageAttention team. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + #include + #include + #include + #include + #include + #include + #include + + #include + #include + + #include + + #include "cuda_utils.h" + #include "../blackwell/block_config.h" + + #define DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(pytorch_dtype, c_type, ...) \ + if (pytorch_dtype == at::ScalarType::Half) { \ + using c_type = half; \ + __VA_ARGS__ \ + } else if (pytorch_dtype == at::ScalarType::BFloat16) { \ + using c_type = nv_bfloat16; \ + __VA_ARGS__ \ + } else { \ + std::ostringstream oss; \ + oss << __PRETTY_FUNCTION__ << " failed to dispatch data type " << pytorch_dtype; \ + TORCH_CHECK(false, oss.str()); \ + } + + #define DISPATCH_HEAD_DIM(head_dim, HEAD_DIM, ...) \ + if (head_dim == 64) { \ + constexpr int HEAD_DIM = 64; \ + __VA_ARGS__ \ + } else if (head_dim == 128) { \ + constexpr int HEAD_DIM = 128; \ + __VA_ARGS__ \ + } else { \ + std::ostringstream err_msg; \ + err_msg << "Unsupported head dim: " << int(head_dim); \ + throw std::invalid_argument(err_msg.str()); \ + } + + #define CHECK_CUDA(x) \ + TORCH_CHECK(x.is_cuda(), "Tensor " #x " must be on CUDA") + #define CHECK_DTYPE(x, true_dtype) \ + TORCH_CHECK(x.dtype() == true_dtype, \ + "Tensor " #x " must have dtype (" #true_dtype ")") + #define CHECK_DIMS(x, true_dim) \ + TORCH_CHECK(x.dim() == true_dim, \ + "Tensor " #x " must have dimension number (" #true_dim ")") + #define CHECK_SHAPE(x, ...) \ + TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), \ + "Tensor " #x " must have shape (" #__VA_ARGS__ ")") + #define CHECK_CONTIGUOUS(x) \ + TORCH_CHECK(x.is_contiguous(), "Tensor " #x " must be contiguous") + #define CHECK_LASTDIM_CONTIGUOUS(x) \ + TORCH_CHECK(x.stride(-1) == 1, \ + "Tensor " #x " must be contiguous at the last dimension") + + constexpr int CVT_FP4_ELTS_PER_THREAD = 16; + + // Convert 4 float2 values into 8 e2m1 values (represented as one uint32_t). + inline __device__ uint32_t fp32_vec_to_e2m1(float2 *array) { + #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000) + uint32_t val; + asm volatile( + "{\n" + ".reg .b8 byte0;\n" + ".reg .b8 byte1;\n" + ".reg .b8 byte2;\n" + ".reg .b8 byte3;\n" + "cvt.rn.satfinite.e2m1x2.f32 byte0, %2, %1;\n" + "cvt.rn.satfinite.e2m1x2.f32 byte1, %4, %3;\n" + "cvt.rn.satfinite.e2m1x2.f32 byte2, %6, %5;\n" + "cvt.rn.satfinite.e2m1x2.f32 byte3, %8, %7;\n" + "mov.b32 %0, {byte0, byte1, byte2, byte3};\n" + "}" + : "=r"(val) + : "f"(array[0].x), "f"(array[0].y), "f"(array[1].x), "f"(array[1].y), + "f"(array[2].x), "f"(array[2].y), "f"(array[3].x), "f"(array[3].y)); + return val; + #else + return 0; + #endif + } + + // Get type2 from type or vice versa (applied to half and bfloat16) + template + struct TypeConverter { + using Type = half2; + }; // keep for generality + + template <> + struct TypeConverter { + using Type = half; + }; + + template <> + struct TypeConverter { + using Type = half2; + }; + + template <> + struct TypeConverter<__nv_bfloat162> { + using Type = __nv_bfloat16; + }; + + template <> + struct TypeConverter<__nv_bfloat16> { + using Type = __nv_bfloat162; + }; + + // Define a 32 bytes packed data type. + template + struct PackedVec { + typename TypeConverter::Type elts[8]; + }; + + template + __global__ void scaled_fp4_quant_kernel( + const T* input, uint8_t* output, uint8_t* output_sf, + int batch_size, int num_heads, int num_tokens, + int stride_bz_input, int stride_h_input, int stride_seq_input, + int stride_bz_output, int stride_h_output, int stride_seq_output, + int stride_bz_output_sf, int stride_h_output_sf, int stride_seq_output_sf) { + static_assert(std::is_same::value || std::is_same::value, "Only half and bfloat16 input are supported"); + using PackedVec = PackedVec; + + const int batch_id = blockIdx.y; + const int head_id = blockIdx.z; + const int token_block_id = blockIdx.x; + + static_assert(CVT_FP4_ELTS_PER_THREAD == 8 || CVT_FP4_ELTS_PER_THREAD == 16, + "CVT_FP4_ELTS_PER_THREAD must be 8 or 16"); + static_assert(sizeof(PackedVec) == sizeof(T) * CVT_FP4_ELTS_PER_THREAD, + "Vec size is not matched."); + + constexpr uint32_t NUM_THREADS_PER_TOKEN = head_dim / CVT_FP4_ELTS_PER_THREAD; + + // load input + const int token_id = token_block_id * BLOCK_SIZE + threadIdx.x / NUM_THREADS_PER_TOKEN; + + int load_token_id; + if constexpr (!permute) { + load_token_id = token_id; + } else { + int local_token_id = threadIdx.x / NUM_THREADS_PER_TOKEN; + int local_token_id_residue = local_token_id % 32; + // [0, 1, 8, 9, 16, 17, 24, 25, 2, 3, 10, 11, 18, 19, 26, 27, 4, 5, 12, 13, 20, 21, 28, 29, 6, 7, 14, 15, 22, 23, 30, 31] + load_token_id = token_block_id * BLOCK_SIZE + (local_token_id / 32) * 32 + + (local_token_id_residue / 8) * 2 + + ((local_token_id_residue % 8) / 2) * 8 + + (local_token_id_residue % 8) % 2; + } + + PackedVec in_vec; + + #pragma unroll + for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) { + reinterpret_cast(in_vec.elts[i]) = 0; + } + + if (load_token_id < num_tokens) { + in_vec = reinterpret_cast(input + + batch_id * stride_bz_input + // batch dim + head_id * stride_h_input + // head dim + load_token_id * stride_seq_input + // seq dim + (threadIdx.x % NUM_THREADS_PER_TOKEN) * CVT_FP4_ELTS_PER_THREAD)[0]; // feature dim + } + + // calculate max of every consecutive 16 elements + auto localMax = __habs2(in_vec.elts[0]); + #pragma unroll + for (int i = 1; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) { // local max + localMax = __hmax2(localMax, __habs2(in_vec.elts[i])); + } + + if constexpr (CVT_FP4_ELTS_PER_THREAD == 8) { // shuffle across two threads + localMax = __hmax2(__shfl_xor_sync(0xffffffff, localMax, 1, 32), localMax); + } + + float vecMax = float(__hmax(localMax.x, localMax.y)); + + // scaling factor + float SFValue = vecMax / 6.0f; + uint8_t SFValueFP8; + reinterpret_cast<__nv_fp8_e4m3&>(SFValueFP8) = __nv_fp8_e4m3(SFValue); + SFValue = float(reinterpret_cast<__nv_fp8_e4m3&>(SFValueFP8)); + + float SFValueInv = (SFValue == 0.0f) ? 0.0f : 1.0f / SFValue; + + // convert input to float2 and apply scale + float2 fp2Vals[CVT_FP4_ELTS_PER_THREAD / 2]; + + #pragma unroll + for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) { + if constexpr (std::is_same::value) { + fp2Vals[i] = __half22float2(in_vec.elts[i]); + } else { + fp2Vals[i] = __bfloat1622float2(in_vec.elts[i]); + } + fp2Vals[i].x = fp2Vals[i].x * SFValueInv; + fp2Vals[i].y = fp2Vals[i].y * SFValueInv; + } + + // convert to e2m1 + uint32_t e2m1Vals[CVT_FP4_ELTS_PER_THREAD / 8]; + #pragma unroll + for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 8; i++) { + e2m1Vals[i] = fp32_vec_to_e2m1(fp2Vals + i * 4); + } + + // save, do not check range + if constexpr (CVT_FP4_ELTS_PER_THREAD == 8) { + reinterpret_cast(output + + batch_id * stride_bz_output + + head_id * stride_h_output + + token_id * stride_seq_output + + (threadIdx.x % NUM_THREADS_PER_TOKEN) * CVT_FP4_ELTS_PER_THREAD / 2)[0] = e2m1Vals[0]; + } else { + reinterpret_cast(output + + batch_id * stride_bz_output + + head_id * stride_h_output + + token_id * stride_seq_output + + (threadIdx.x % NUM_THREADS_PER_TOKEN) * CVT_FP4_ELTS_PER_THREAD / 2)[0] = reinterpret_cast(e2m1Vals)[0]; + } + + uint8_t* output_sf_save_base = output_sf + batch_id * stride_bz_output_sf + head_id * stride_h_output_sf + (token_id / 64) * 64 * stride_seq_output_sf; + uint32_t token_id_local = token_id % 64; + + if constexpr (CVT_FP4_ELTS_PER_THREAD == 16) { + uint32_t col_id_local = threadIdx.x % NUM_THREADS_PER_TOKEN; + uint32_t offset_local = (col_id_local / 4) * 256 + (col_id_local % 4) + + (token_id_local / 16) * 4 + (token_id_local % 16) * 16; + reinterpret_cast(output_sf_save_base + offset_local)[0] = SFValueFP8; + } else { + if (threadIdx.x % 2 == 0) { + uint32_t col_id_local = (threadIdx.x % NUM_THREADS_PER_TOKEN) / 2; + uint32_t offset_local = (col_id_local / 4) * 256 + (col_id_local % 4) + + (token_id_local / 16) * 4 + (token_id_local % 16) * 16; + reinterpret_cast(output_sf_save_base + offset_local)[0] = SFValueFP8; + } + } + } + + template + __global__ void scaled_fp4_quant_trans_kernel( + const T* input, uint8_t* output, uint8_t* output_sf, + int batch_size, int num_heads, int num_tokens, + int stride_bz_input, int stride_h_input, int stride_seq_input, + int stride_bz_output, int stride_h_output, int stride_d_output, + int stride_bz_output_sf, int stride_h_output_sf, int stride_d_output_sf) { + static_assert(std::is_same::value || std::is_same::value, "Only half and bfloat16 input are supported"); + using PackedVec = PackedVec; + + const int batch_id = blockIdx.y; + const int head_id = blockIdx.z; + const int token_block_id = blockIdx.x; + + static_assert(CVT_FP4_ELTS_PER_THREAD == 8 || CVT_FP4_ELTS_PER_THREAD == 16, + "CVT_FP4_ELTS_PER_THREAD must be 8 or 16"); + static_assert(sizeof(PackedVec) == sizeof(T) * CVT_FP4_ELTS_PER_THREAD, + "Vec size is not matched."); + + constexpr uint32_t NUM_THREADS_PER_TOKEN = head_dim / CVT_FP4_ELTS_PER_THREAD; + constexpr uint32_t NUM_THREADS_PER_SEQ = BLOCK_SIZE / CVT_FP4_ELTS_PER_THREAD; + + // load input + const int token_id = token_block_id * BLOCK_SIZE + threadIdx.x / NUM_THREADS_PER_TOKEN; + + PackedVec in_vec; + + #pragma unroll + for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) { + reinterpret_cast(in_vec.elts[i]) = 0; + } + + if (token_id < num_tokens) { + in_vec = reinterpret_cast(input + + batch_id * stride_bz_input + // batch dim + head_id * stride_h_input + // head dim + token_id * stride_seq_input + // seq dim + (threadIdx.x % NUM_THREADS_PER_TOKEN) * CVT_FP4_ELTS_PER_THREAD)[0]; // feature dim + } + + // transpose + __shared__ T shared_input[BLOCK_SIZE * head_dim]; + reinterpret_cast(shared_input)[threadIdx.x] = in_vec; + __syncthreads(); + #pragma unroll + for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) { + in_vec.elts[i].x = shared_input[(threadIdx.x / NUM_THREADS_PER_SEQ) + ((threadIdx.x % NUM_THREADS_PER_SEQ) * CVT_FP4_ELTS_PER_THREAD + 2 * i) * head_dim]; + in_vec.elts[i].y = shared_input[(threadIdx.x / NUM_THREADS_PER_SEQ) + ((threadIdx.x % NUM_THREADS_PER_SEQ) * CVT_FP4_ELTS_PER_THREAD + 2 * i + 1) * head_dim]; + } + + // calculate max of every consecutive 16 elements + auto localMax = __habs2(in_vec.elts[0]); + #pragma unroll + for (int i = 1; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) { // local max + localMax = __hmax2(localMax, __habs2(in_vec.elts[i])); + } + + if constexpr (CVT_FP4_ELTS_PER_THREAD == 8) { // shuffle across two threads + localMax = __hmax2(__shfl_xor_sync(0xffffffff, localMax, 1, 32), localMax); + } + + float vecMax = float(__hmax(localMax.x, localMax.y)); + + // scaling factor + float SFValue = vecMax / 6.0f; + uint8_t SFValueFP8; + reinterpret_cast<__nv_fp8_e4m3&>(SFValueFP8) = __nv_fp8_e4m3(SFValue); + SFValue = float(reinterpret_cast<__nv_fp8_e4m3&>(SFValueFP8)); + + float SFValueInv = (SFValue == 0.0f) ? 0.0f : 1.0f / SFValue; + + // convert input to float2 and apply scale + float2 fp2Vals[CVT_FP4_ELTS_PER_THREAD / 2]; + + #pragma unroll + for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) { + if constexpr (std::is_same::value) { + fp2Vals[i] = __half22float2(in_vec.elts[i]); + } else { + fp2Vals[i] = __bfloat1622float2(in_vec.elts[i]); + } + fp2Vals[i].x = fp2Vals[i].x * SFValueInv; + fp2Vals[i].y = fp2Vals[i].y * SFValueInv; + } + + // convert to e2m1 + uint32_t e2m1Vals[CVT_FP4_ELTS_PER_THREAD / 8]; + #pragma unroll + for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 8; i++) { + e2m1Vals[i] = fp32_vec_to_e2m1(fp2Vals + i * 4); + } + + // save + if constexpr (CVT_FP4_ELTS_PER_THREAD == 8) { + reinterpret_cast(output + + batch_id * stride_bz_output + + head_id * stride_h_output + + (threadIdx.x / NUM_THREADS_PER_SEQ) * stride_d_output + + (token_block_id * BLOCK_SIZE + (threadIdx.x % NUM_THREADS_PER_SEQ) * CVT_FP4_ELTS_PER_THREAD) / 2)[0] = e2m1Vals[0]; + } else { + reinterpret_cast(output + + batch_id * stride_bz_output + + head_id * stride_h_output + + (threadIdx.x / NUM_THREADS_PER_SEQ) * stride_d_output + + (token_block_id * BLOCK_SIZE + (threadIdx.x % NUM_THREADS_PER_SEQ) * CVT_FP4_ELTS_PER_THREAD) / 2)[0] = reinterpret_cast(e2m1Vals)[0]; + } + + uint8_t *output_sf_save_base = output_sf + + batch_id * stride_bz_output_sf + + head_id * stride_h_output_sf + + (threadIdx.x / NUM_THREADS_PER_SEQ / 64) * 64 * stride_d_output_sf; + uint32_t row_id_local = (threadIdx.x / NUM_THREADS_PER_SEQ) % 64; + + if constexpr (CVT_FP4_ELTS_PER_THREAD == 16) { + uint32_t col_id_local = token_block_id * BLOCK_SIZE / CVT_FP4_ELTS_PER_THREAD + threadIdx.x % NUM_THREADS_PER_SEQ; + uint32_t offset_local = (col_id_local / 4) * 256 + (col_id_local % 4) + + (row_id_local / 16) * 4 + (row_id_local % 16) * 16; + reinterpret_cast(output_sf_save_base + offset_local)[0] = SFValueFP8; + } else { + if (threadIdx.x % 2 == 0) { + uint32_t col_id_local = token_block_id * BLOCK_SIZE / CVT_FP4_ELTS_PER_THREAD + (threadIdx.x % NUM_THREADS_PER_SEQ) / 2; + uint32_t offset_local = (col_id_local / 4) * 256 + (col_id_local % 4) + + (row_id_local / 16) * 4 + (row_id_local % 16) * 16; + reinterpret_cast(output_sf_save_base + offset_local)[0] = SFValueFP8; + } + } + } + + void scaled_fp4_quant(torch::Tensor const& input, + torch::Tensor const& output, + torch::Tensor const& output_sf, + int tensor_layout) { + constexpr int BLOCK_SIZE = flash::BLOCK_M; + + CHECK_CUDA(input); + CHECK_CUDA(output); + CHECK_CUDA(output_sf); + + CHECK_LASTDIM_CONTIGUOUS(input); + CHECK_LASTDIM_CONTIGUOUS(output); + CHECK_LASTDIM_CONTIGUOUS(output_sf); + + CHECK_DTYPE(output, at::ScalarType::Byte); + CHECK_DTYPE(output_sf, at::ScalarType::Float8_e4m3fn); + + CHECK_DIMS(input, 4); + CHECK_DIMS(output, 4); + CHECK_DIMS(output_sf, 4); + + const int batch_size = input.size(0); + const int head_dim = input.size(3); + + const int stride_bz_input = input.stride(0); + const int stride_bz_output = output.stride(0); + const int stride_bz_output_sf = output_sf.stride(0); + + int num_tokens, num_heads; + int stride_seq_input, stride_seq_output, stride_seq_output_sf; + int stride_h_input, stride_h_output, stride_h_output_sf; + if (tensor_layout == 0) { + num_tokens = input.size(1); + num_heads = input.size(2); + stride_seq_input = input.stride(1); + stride_seq_output = output.stride(1); + stride_seq_output_sf = output_sf.stride(1); + stride_h_input = input.stride(2); + stride_h_output = output.stride(2); + stride_h_output_sf = output_sf.stride(2); + + CHECK_SHAPE(output, batch_size, num_tokens, num_heads, head_dim / 2); + CHECK_SHAPE(output_sf, batch_size, num_tokens, num_heads, head_dim / 16); + } else { + num_tokens = input.size(2); + num_heads = input.size(1); + stride_seq_input = input.stride(2); + stride_seq_output = output.stride(2); + stride_seq_output_sf = output_sf.stride(2); + stride_h_input = input.stride(1); + stride_h_output = output.stride(1); + stride_h_output_sf = output_sf.stride(1); + + CHECK_SHAPE(output, batch_size, num_heads, num_tokens, head_dim / 2); + CHECK_SHAPE(output_sf, batch_size, num_heads, num_tokens, head_dim / 16); + } + + auto input_dtype = input.scalar_type(); + auto stream = at::cuda::getCurrentCUDAStream(input.get_device()); + + DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input_dtype, c_type, { + DISPATCH_HEAD_DIM(head_dim, HEAD_DIM, { + dim3 block(BLOCK_SIZE * HEAD_DIM / CVT_FP4_ELTS_PER_THREAD, 1, 1); + dim3 grid((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE, batch_size, num_heads); + + scaled_fp4_quant_kernel + <<>>( + reinterpret_cast(input.data_ptr()), + reinterpret_cast(output.data_ptr()), + reinterpret_cast(output_sf.data_ptr()), + batch_size, num_heads, num_tokens, + stride_bz_input, stride_h_input, stride_seq_input, + stride_bz_output, stride_h_output, stride_seq_output, + stride_bz_output_sf, stride_h_output_sf, stride_seq_output_sf); + }); + }); + } + + void scaled_fp4_quant_permute(torch::Tensor const& input, + torch::Tensor const& output, + torch::Tensor const& output_sf, + int tensor_layout) { + constexpr int BLOCK_SIZE = flash::BLOCK_M; + + CHECK_CUDA(input); + CHECK_CUDA(output); + CHECK_CUDA(output_sf); + + CHECK_LASTDIM_CONTIGUOUS(input); + CHECK_LASTDIM_CONTIGUOUS(output); + CHECK_LASTDIM_CONTIGUOUS(output_sf); + + CHECK_DTYPE(output, at::ScalarType::Byte); + CHECK_DTYPE(output_sf, at::ScalarType::Float8_e4m3fn); + + CHECK_DIMS(input, 4); + CHECK_DIMS(output, 4); + CHECK_DIMS(output_sf, 4); + + const int batch_size = input.size(0); + const int head_dim = input.size(3); + + const int stride_bz_input = input.stride(0); + const int stride_bz_output = output.stride(0); + const int stride_bz_output_sf = output_sf.stride(0); + + int num_tokens, num_heads; + int stride_seq_input, stride_seq_output, stride_seq_output_sf; + int stride_h_input, stride_h_output, stride_h_output_sf; + if (tensor_layout == 0) { + num_tokens = input.size(1); + num_heads = input.size(2); + stride_seq_input = input.stride(1); + stride_seq_output = output.stride(1); + stride_seq_output_sf = output_sf.stride(1); + stride_h_input = input.stride(2); + stride_h_output = output.stride(2); + stride_h_output_sf = output_sf.stride(2); + + CHECK_SHAPE(output, batch_size, ((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE, num_heads, head_dim / 2); + CHECK_SHAPE(output_sf, batch_size, ((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE, num_heads, head_dim / 16); + } else { + num_tokens = input.size(2); + num_heads = input.size(1); + stride_seq_input = input.stride(2); + stride_seq_output = output.stride(2); + stride_seq_output_sf = output_sf.stride(2); + stride_h_input = input.stride(1); + stride_h_output = output.stride(1); + stride_h_output_sf = output_sf.stride(1); + + CHECK_SHAPE(output, batch_size, num_heads, ((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE, head_dim / 2); + CHECK_SHAPE(output_sf, batch_size, num_heads, ((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE, head_dim / 16); + } + + auto input_dtype = input.scalar_type(); + auto stream = at::cuda::getCurrentCUDAStream(input.get_device()); + + DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input_dtype, c_type, { + DISPATCH_HEAD_DIM(head_dim, HEAD_DIM, { + constexpr int BLOCK_SIZE = flash::BLOCK_M; + dim3 block(BLOCK_SIZE * HEAD_DIM / CVT_FP4_ELTS_PER_THREAD, 1, 1); + dim3 grid((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE, batch_size, num_heads); + + scaled_fp4_quant_kernel + <<>>( + reinterpret_cast(input.data_ptr()), + reinterpret_cast(output.data_ptr()), + reinterpret_cast(output_sf.data_ptr()), + batch_size, num_heads, num_tokens, + stride_bz_input, stride_h_input, stride_seq_input, + stride_bz_output, stride_h_output, stride_seq_output, + stride_bz_output_sf, stride_h_output_sf, stride_seq_output_sf); + }); + }); + } + + void scaled_fp4_quant_trans(torch::Tensor const& input, + torch::Tensor const& output, + torch::Tensor const& output_sf, + int tensor_layout) { + constexpr int BLOCK_SIZE = flash::BLOCK_M; + + CHECK_CUDA(input); + CHECK_CUDA(output); + CHECK_CUDA(output_sf); + + CHECK_LASTDIM_CONTIGUOUS(input); + CHECK_LASTDIM_CONTIGUOUS(output); + CHECK_LASTDIM_CONTIGUOUS(output_sf); + + CHECK_DTYPE(output, at::ScalarType::Byte); + CHECK_DTYPE(output_sf, at::ScalarType::Float8_e4m3fn); + + CHECK_DIMS(input, 4); + CHECK_DIMS(output, 4); + CHECK_DIMS(output_sf, 4); + + const int batch_size = input.size(0); + const int head_dim = input.size(3); + + const int stride_bz_input = input.stride(0); + const int stride_bz_output = output.stride(0); + const int stride_bz_output_sf = output_sf.stride(0); + + int num_tokens, num_heads; + int stride_seq_input; + int stride_d_output, stride_d_output_sf; + int stride_h_input, stride_h_output, stride_h_output_sf; + if (tensor_layout == 0) { + num_tokens = input.size(1); + num_heads = input.size(2); + stride_seq_input = input.stride(1); + stride_d_output = output.stride(1); + stride_d_output_sf = output_sf.stride(1); + stride_h_input = input.stride(2); + stride_h_output = output.stride(2); + stride_h_output_sf = output_sf.stride(2); + + CHECK_SHAPE(output, batch_size, head_dim, num_heads, ((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE / 2); + CHECK_SHAPE(output_sf, batch_size, head_dim, num_heads, ((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE / 16); + } else { + num_tokens = input.size(2); + num_heads = input.size(1); + stride_seq_input = input.stride(2); + stride_d_output = output.stride(2); + stride_d_output_sf = output_sf.stride(2); + stride_h_input = input.stride(1); + stride_h_output = output.stride(1); + stride_h_output_sf = output_sf.stride(1); + + CHECK_SHAPE(output, batch_size, num_heads, head_dim, ((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE / 2); + CHECK_SHAPE(output_sf, batch_size, num_heads, head_dim, ((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE / 16); + } + + auto input_dtype = input.scalar_type(); + auto stream = at::cuda::getCurrentCUDAStream(input.get_device()); + + DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input_dtype, c_type, { + DISPATCH_HEAD_DIM(head_dim, HEAD_DIM, { + dim3 block(BLOCK_SIZE * HEAD_DIM / CVT_FP4_ELTS_PER_THREAD, 1, 1); + dim3 grid((num_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE, batch_size, num_heads); + + scaled_fp4_quant_trans_kernel + <<>>( + reinterpret_cast(input.data_ptr()), + reinterpret_cast(output.data_ptr()), + reinterpret_cast(output_sf.data_ptr()), + batch_size, num_heads, num_tokens, + stride_bz_input, stride_h_input, stride_seq_input, + stride_bz_output, stride_h_output, stride_d_output, + stride_bz_output_sf, stride_h_output_sf, stride_d_output_sf); + }); + }); + } + + PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("scaled_fp4_quant", &scaled_fp4_quant); + m.def("scaled_fp4_quant_permute", &scaled_fp4_quant_permute); + m.def("scaled_fp4_quant_trans", &scaled_fp4_quant_trans); + } \ No newline at end of file diff --git a/fastvideo/attention/backends/sdpa.py b/fastvideo/attention/backends/sdpa.py index 5d9b21591..5fc2d5c2e 100644 --- a/fastvideo/attention/backends/sdpa.py +++ b/fastvideo/attention/backends/sdpa.py @@ -2,6 +2,7 @@ import torch from dataclasses import dataclass +from dataclasses import dataclass from fastvideo.attention.backends.abstract import ( # FlashAttentionMetadata, AttentionBackend, AttentionImpl, AttentionMetadata, AttentionMetadataBuilder) from fastvideo.logger import init_logger diff --git a/fastvideo/attention/backends/setup.py b/fastvideo/attention/backends/setup.py new file mode 100644 index 000000000..8a88e0427 --- /dev/null +++ b/fastvideo/attention/backends/setup.py @@ -0,0 +1,178 @@ +import warnings +import os +from pathlib import Path +from packaging.version import parse, Version +from setuptools import setup, find_packages +import subprocess +from wheel.bdist_wheel import bdist_wheel as _bdist_wheel + +import torch +from torch.utils.cpp_extension import BuildExtension, CppExtension, CUDAExtension, CUDA_HOME + +this_dir = os.path.dirname(os.path.abspath(__file__)) + +PACKAGE_NAME = "sageattn3" + +# FORCE_BUILD: Force a fresh build locally, instead of attempting to find prebuilt wheels +# SKIP_CUDA_BUILD: Intended to allow CI to use a simple `python setup.py sdist` run to copy over raw files, without any cuda compilation +FORCE_BUILD = os.getenv("FAHOPPER_FORCE_BUILD", "FALSE") == "TRUE" +SKIP_CUDA_BUILD = os.getenv("FAHOPPER_SKIP_CUDA_BUILD", "FALSE") == "TRUE" +# For CI, we want the option to build with C++11 ABI since the nvcr images use C++11 ABI +FORCE_CXX11_ABI = os.getenv("FAHOPPER_FORCE_CXX11_ABI", "FALSE") == "TRUE" + + + +def get_cuda_bare_metal_version(cuda_dir): + raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True) + output = raw_output.split() + release_idx = output.index("release") + 1 + bare_metal_version = parse(output[release_idx].split(",")[0]) + + return raw_output, bare_metal_version + + +def check_if_cuda_home_none(global_option: str) -> None: + if CUDA_HOME is not None: + return + # warn instead of error because user could be downloading prebuilt wheels, so nvcc won't be necessary + # in that case. + warnings.warn( + f"{global_option} was requested, but nvcc was not found. Are you sure your environment has nvcc available? " + "If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, " + "only images whose names contain 'devel' will provide nvcc." + ) + + +def append_nvcc_threads(nvcc_extra_args): + return nvcc_extra_args + ["--threads", "4"] + + +cmdclass = {} +ext_modules = [] + +if not SKIP_CUDA_BUILD: + print("\n\ntorch.__version__ = {}\n\n".format(torch.__version__)) + TORCH_MAJOR = int(torch.__version__.split(".")[0]) + TORCH_MINOR = int(torch.__version__.split(".")[1]) + + check_if_cuda_home_none("--fahopper") + cc_flag = [] + _, bare_metal_version = get_cuda_bare_metal_version(CUDA_HOME) + if bare_metal_version < Version("12.8"): + raise RuntimeError("Sage3 is only supported on CUDA 12.8 and above") + cc_flag.append("-gencode") + cc_flag.append("arch=compute_120a,code=sm_120a") + + # HACK: The compiler flag -D_GLIBCXX_USE_CXX11_ABI is set to be the same as + # torch._C._GLIBCXX_USE_CXX11_ABI + # https://github.com/pytorch/pytorch/blob/8472c24e3b5b60150096486616d98b7bea01500b/torch/utils/cpp_extension.py#L920 + if FORCE_CXX11_ABI: + torch._C._GLIBCXX_USE_CXX11_ABI = True + repo_dir = Path(this_dir) + cutlass_dir = repo_dir / "csrc" / "cutlass" + (repo_dir / "csrc").mkdir(parents=True, exist_ok=True) + if not cutlass_dir.exists(): + subprocess.run( + ["git", "clone", "--depth", "1", "https://github.com/NVIDIA/cutlass.git", str(cutlass_dir)], + check=True + ) + nvcc_flags = [ + "-O3", + # "-O0", + "-std=c++17", + "-U__CUDA_NO_HALF_OPERATORS__", + "-U__CUDA_NO_HALF_CONVERSIONS__", + "-U__CUDA_NO_BFLOAT16_OPERATORS__", + "-U__CUDA_NO_BFLOAT16_CONVERSIONS__", + "-U__CUDA_NO_BFLOAT162_OPERATORS__", + "-U__CUDA_NO_BFLOAT162_CONVERSIONS__", + "--expt-relaxed-constexpr", + "--expt-extended-lambda", + "--use_fast_math", + # "--ptxas-options=-v", # printing out number of registers + "--ptxas-options=--verbose,--warn-on-local-memory-usage", # printing out number of registers + "-lineinfo", + "-DCUTLASS_DEBUG_TRACE_LEVEL=0", # Can toggle for debugging + "-DNDEBUG", # Important, otherwise performance is severely impacted + "-DQBLKSIZE=128", + "-DKBLKSIZE=128", + "-DCTA256", + "-DDQINRMEM", + ] + include_dirs = [ + repo_dir / "sageattn", + cutlass_dir / "include", + cutlass_dir / "tools" / "util" / "include", + ] + + ext_modules.append( + CUDAExtension( + name="fp4attn_cuda", + sources=["sageattn/blackwell/api.cu"], + extra_compile_args={ + "cxx": ["-O3", "-std=c++17"], + "nvcc": append_nvcc_threads( + nvcc_flags + ["-DEXECMODE=0"] + cc_flag + ), + }, + include_dirs=include_dirs, + # Without this we get and error about cuTensorMapEncodeTiled not defined + libraries=["cuda"] + ) + ) + ext_modules.append( + CUDAExtension( + name="fp4quant_cuda", + sources=["sageattn/quantization/fp4_quantization_4d.cu"], + extra_compile_args={ + "cxx": ["-O3", "-std=c++17"], + "nvcc": append_nvcc_threads( + nvcc_flags + ["-DEXECMODE=0"] + cc_flag + ), + }, + include_dirs=include_dirs, + # Without this we get and error about cuTensorMapEncodeTiled not defined + libraries=["cuda"] + ) + ) + + + +class CachedWheelsCommand(_bdist_wheel): + def run(self): + super().run() + +setup( + name=PACKAGE_NAME, + version="1.0.0", + packages=find_packages( + exclude=( + "build", + "csrc", + "tests", + "dist", + "docs", + "benchmarks", + ) + ), + description="FP4FlashAttention", + long_description_content_type="text/markdown", + classifiers=[ + "Programming Language :: Python :: 3", + "License :: OSI Approved :: Apache Software License", + "Operating System :: Unix", + ], + ext_modules=ext_modules, + cmdclass={"bdist_wheel": CachedWheelsCommand, "build_ext": BuildExtension} + if ext_modules + else { + "bdist_wheel": CachedWheelsCommand, + }, + python_requires=">=3.8", + install_requires=[ + "torch", + "einops", + "packaging", + "ninja", + ], +) diff --git a/fastvideo/attention/layer.py b/fastvideo/attention/layer.py index 84fa3d3bb..b83d7019c 100644 --- a/fastvideo/attention/layer.py +++ b/fastvideo/attention/layer.py @@ -48,6 +48,10 @@ class DistributedAttention(nn.Module): **extra_impl_args) # Register attn_impl as submodule if it has learnable parameters (e.g., SLA's proj_l) # This ensures its parameters are included in state_dict() for saving/loading + if isinstance(self.attn_impl, nn.Module): + self.add_module('attn_impl', self.attn_impl) + # Register attn_impl as submodule if it has learnable parameters (e.g., SLA's proj_l) + # This ensures its parameters are included in state_dict() for saving/loading if isinstance(self.attn_impl, nn.Module): self.add_module('attn_impl', self.attn_impl) self.num_heads = num_heads @@ -158,6 +162,7 @@ class DistributedAttention_VSA(DistributedAttention): replicated_v: torch.Tensor | None = None, gate_compress: torch.Tensor | None = None, freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None, + attention_mask: torch.Tensor | None = None, ) -> tuple[torch.Tensor, torch.Tensor | None]: """Forward pass for distributed attention. @@ -170,6 +175,7 @@ class DistributedAttention_VSA(DistributedAttention): replicated_q (Optional[torch.Tensor]): Replicated query tensor, typically for text tokens replicated_k (Optional[torch.Tensor]): Replicated key tensor replicated_v (Optional[torch.Tensor]): Replicated value tensor + attention_mask (Optional[torch.Tensor]): Attention mask [batch_size, seq_len] Returns: Tuple[torch.Tensor, Optional[torch.Tensor]]: A tuple containing: diff --git a/fastvideo/fastvideo_args.py b/fastvideo/fastvideo_args.py index 10a5fe0b3..55e079a76 100644 --- a/fastvideo/fastvideo_args.py +++ b/fastvideo/fastvideo_args.py @@ -199,6 +199,7 @@ class FastVideoArgs: logger.error("Failed to load V-MoBA config from %s: %s", self.moba_config_path, e) raise self._apply_ltx2_vae_overrides() + self._apply_ltx2_vae_overrides() self.check_fastvideo_args() def _apply_ltx2_vae_overrides(self) -> None: @@ -380,6 +381,100 @@ class FastVideoArgs: help="Path to load/save a precomputed LTX-2 initial latent.", ) + # LoRA parameters (inference-time adapter loading) + parser.add_argument( + "--lora-path", + type=str, + default=FastVideoArgs.lora_path, + help= + "Path to a LoRA adapter (directory or HF repo id). If set, LoRA will be applied at inference.", + ) + parser.add_argument( + "--lora-nickname", + type=str, + default=FastVideoArgs.lora_nickname, + help= + "Nickname to refer to the loaded LoRA adapter (useful for swapping).", + ) + parser.add_argument( + "--lora-target-modules", + nargs="+", + type=str, + default=FastVideoArgs.lora_target_modules, + help= + "Optional list of module name substrings to restrict LoRA injection (e.g. q_proj k_proj v_proj).", + ) + + # BSA runtime control (LongCat) + parser.add_argument( + "--enable-bsa", + action=StoreBoolean, + help= + "Enable Block Sparse Attention (BSA) at runtime (overrides config).", + ) + parser.add_argument( + "--bsa-sparsity", + type=float, + help="BSA sparsity (e.g., 0.9375).", + ) + parser.add_argument( + "--bsa-cdf-threshold", + type=float, + help="BSA CDF threshold (optional).", + ) + parser.add_argument( + "--bsa-chunk-q", + nargs=3, + type=int, + metavar=("T", "H", "W"), + help="BSA chunk_3d_shape_q as three ints, e.g., 4 4 4.", + ) + parser.add_argument( + "--bsa-chunk-k", + nargs=3, + type=int, + metavar=("T", "H", "W"), + help="BSA chunk_3d_shape_k as three ints, e.g., 4 4 4.", + ) + + # LTX-2 VAE tiling overrides + parser.add_argument( + "--ltx2-vae-tiling", + action=StoreBoolean, + default=FastVideoArgs.ltx2_vae_tiling, + help="Enable LTX-2 VAE tiling overrides.", + ) + parser.add_argument( + "--ltx2-vae-spatial-tile-size-in-pixels", + type=int, + default=FastVideoArgs.ltx2_vae_spatial_tile_size_in_pixels, + help="LTX-2 VAE spatial tile size in pixels.", + ) + parser.add_argument( + "--ltx2-vae-spatial-tile-overlap-in-pixels", + type=int, + default=FastVideoArgs.ltx2_vae_spatial_tile_overlap_in_pixels, + help="LTX-2 VAE spatial tile overlap in pixels.", + ) + parser.add_argument( + "--ltx2-vae-temporal-tile-size-in-frames", + type=int, + default=FastVideoArgs.ltx2_vae_temporal_tile_size_in_frames, + help="LTX-2 VAE temporal tile size in frames.", + ) + parser.add_argument( + "--ltx2-vae-temporal-tile-overlap-in-frames", + type=int, + default=FastVideoArgs.ltx2_vae_temporal_tile_overlap_in_frames, + help="LTX-2 VAE temporal tile overlap in frames.", + ) + parser.add_argument( + "--ltx2-initial-latent-path", + type=str, + default=FastVideoArgs.ltx2_initial_latent_path, + help="Path to load/save a precomputed LTX-2 initial latent.", + ) + # LoRA parameters (inference-time adapter loading) parser.add_argument( "--lora-path", @@ -457,12 +552,15 @@ class FastVideoArgs: action=StoreBoolean, help="Enable layerwise CPU offload with async H2D prefetch overlap.", ) + parser.add_argument( + "--dit-layerwise-offload", + action=StoreBoolean, + help="Enable layerwise CPU offload with async H2D prefetch overlap.", + ) parser.add_argument( "--use-fsdp-inference", action=StoreBoolean, - help= - "Use FSDP for inference by sharding the model weights. FSDP helps reduce GPU memory usage but may introduce" - + " weight transfer overhead depending on the specific setup. Enable if run out of memory.", + help="Use FSDP for inference by sharding the model weights. Latency is very low due to prefetch--enable if run out of memory.", ) parser.add_argument( "--text-encoder-cpu-offload", @@ -540,6 +638,12 @@ class FastVideoArgs: default=FastVideoArgs.override_pipeline_cls_name, help="Override pipeline cls name", ) + parser.add_argument( + "--override-pipeline-cls-name", + type=str, + default=FastVideoArgs.override_pipeline_cls_name, + help="Override pipeline cls name", + ) parser.add_argument("--init-weights-from-safetensors", type=str, help="Path to safetensors file for initial weight loading") @@ -855,6 +959,8 @@ class TrainingArgs(FastVideoArgs): # simulate generator forward to match inference simulate_generator_forward: bool = False warp_denoising_step: bool = False + generator_4bit_attn: bool = False + generator_4bit_linear: bool = False # Self-forcing specific arguments num_frame_per_block: int = 3 diff --git a/fastvideo/layers/fp4linear.py b/fastvideo/layers/fp4linear.py new file mode 100644 index 000000000..1ccf14380 --- /dev/null +++ b/fastvideo/layers/fp4linear.py @@ -0,0 +1,84 @@ +import math +import torch +import torch.nn as nn +from flashinfer import SfLayout, mm_fp4, nvfp4_quantize + +class _LinearFWD4BWD16Fn(torch.autograd.Function): + @staticmethod + def forward(ctx, x, weight, bias, backend="cutlass", block_size=16, use_128x4_sf_layout=True): + # assert activation dtype + if x.dtype not in (torch.float16, torch.bfloat16): + x = x.to(dtype=torch.bfloat16) + + # cast params (can be fp32) to activation dtype for quantization + weight_cast = weight.to(dtype=x.dtype) + bias_cast = bias.to(dtype=x.dtype) if bias is not None else None + + # shapes + orig_shape = x.shape + k = weight_cast.shape[1] + n = weight_cast.shape[0] + x2d = x.reshape(-1, k).contiguous() + M = x2d.shape[0] + + out2d = torch.empty((M, n), device=x.device, dtype=x.dtype) + + @torch.compile + def _global_sf(t: torch.Tensor) -> torch.Tensor: + maxabs = t.float().abs().nan_to_num().max() + maxabs = torch.maximum(maxabs, torch.tensor(1e-12, device=t.device, dtype=maxabs.dtype)) + return (448.0 * 6.0) / maxabs + + a_sf_layout = SfLayout.layout_128x4 if use_128x4_sf_layout else SfLayout.layout_8x4 + global_sf_a = _global_sf(x2d) + global_sf_b = _global_sf(weight_cast) + + a_fp4, a_inv_s = nvfp4_quantize(x2d, global_sf_a, sfLayout=a_sf_layout, do_shuffle=False) + b_fp4, b_inv_s = nvfp4_quantize(weight_cast, global_sf_b, sfLayout=SfLayout.layout_128x4, do_shuffle=False) + + alpha = 1.0 / (global_sf_a * global_sf_b) + + mm_fp4( + a_fp4, b_fp4.T, a_inv_s, b_inv_s.T, alpha, + x.dtype, out2d, + block_size=block_size, + use_8x4_sf_layout=(not use_128x4_sf_layout), + backend=backend, + ) + + if bias_cast is not None: + out2d.add_(bias_cast) + + # save tensors for backward (keep original dtypes) + ctx.save_for_backward(x2d, weight, bias) + ctx.k = k + ctx.n = n + ctx.orig_shape = orig_shape + return out2d.reshape(*orig_shape[:-1], n) + + @staticmethod + def backward(ctx, grad_out): + x2d, weight, bias = ctx.saved_tensors + M = x2d.shape[0] + n = ctx.n + + grad_out_2d = grad_out.reshape(M, n).contiguous() + + # cast to grad dtype for matmuls + weight_cast = weight.to(dtype=grad_out.dtype) + x_cast = x2d.to(dtype=grad_out.dtype) + + grad_x = grad_out_2d.matmul(weight_cast).reshape(*ctx.orig_shape) + grad_w = grad_out_2d.t().matmul(x_cast) + grad_b = grad_out_2d.sum(dim=0) if bias is not None else None + + # None for the three extra forward args + return grad_x, grad_w, grad_b, None, None, None + + +def fp4_linear_forward(self, x: torch.Tensor) -> torch.Tensor: + # pass config **positionally**; autograd.Function.apply ignores kwargs + return _LinearFWD4BWD16Fn.apply( + x, self.weight, self.bias, + "cutlass", 16, True + ), None diff --git a/fastvideo/layers/linear.py b/fastvideo/layers/linear.py index e33a738ec..d20810682 100644 --- a/fastvideo/layers/linear.py +++ b/fastvideo/layers/linear.py @@ -150,8 +150,8 @@ class UnquantizedLinearMethod(LinearMethodBase): bias: torch.Tensor | None = None, ) -> torch.Tensor: output = ( - F.linear(x, layer.weight, bias) if torch.cuda.is_available() or bias is None else F.linear( - x, layer.weight, bias.to(x.dtype)) + F.linear(x, layer.weight, bias) if torch.cuda.is_available() + or bias is None else F.linear(x, layer.weight, bias.to(x.dtype)) ) # NOTE: this line assumes that we are using amp when using cuda and is needed to account for the fact that amp isn't supported in mps return output @@ -189,7 +189,8 @@ class LinearBase(torch.nn.Module): self.quant_config = quant_config self.prefix = prefix if quant_config is None: - self.quant_method: QuantizeMethodBase | None = (UnquantizedLinearMethod()) + self.quant_method: QuantizeMethodBase | None = ( + UnquantizedLinearMethod()) else: self.quant_method = quant_config.get_quant_method(self, prefix=prefix) @@ -211,42 +212,38 @@ class ReplicatedLinear(LinearBase): (e.g. model.layers.0.qkv_proj) """ - def __init__( - self, - input_size: int, - output_size: int, - bias: bool = True, - skip_bias_add: bool = False, - params_dtype: torch.dtype | None = None, - quant_config: QuantizationConfig | None = None, - prefix: str = "", - ): - super().__init__( - input_size, - output_size, - skip_bias_add, - params_dtype, - quant_config, - prefix=prefix, - ) + def __init__(self, + input_size: int, + output_size: int, + bias: bool = True, + skip_bias_add: bool = False, + params_dtype: torch.dtype | None = None, + quant_config: QuantizationConfig | None = None, + prefix: str = ""): + super().__init__(input_size, + output_size, + skip_bias_add, + params_dtype, + quant_config, + prefix=prefix) # All the linear layer supports quant method. - assert self.quant_method is not None - self.quant_method.create_weights( - self, - self.input_size, - [self.output_size], - self.input_size, - self.output_size, - self.params_dtype, - weight_loader=self.weight_loader, - ) + if self.quant_method is None: + self.quant_method = UnquantizedLinearMethod() + + self.quant_method.create_weights(self, + self.input_size, [self.output_size], + self.input_size, + self.output_size, + self.params_dtype, + weight_loader=self.weight_loader) if bias: - self.bias = Parameter(torch.empty( - self.output_size, - dtype=self.params_dtype, - )) + self.bias = Parameter( + torch.empty( + self.output_size, + dtype=self.params_dtype, + )) set_weight_attrs( self.bias, { @@ -269,8 +266,25 @@ class ReplicatedLinear(LinearBase): def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, Parameter | None]: bias = self.bias if not self.skip_bias_add else None - assert self.quant_method is not None + if self.quant_method is None: + self.quant_method = UnquantizedLinearMethod() + output = self.quant_method.apply(self, x, bias) + + # Create a shape key for uniqueness checking + # shape_key = (x.shape, output.shape) + + # Track unique shapes and map them to layer types + # if shape_key not in self._unique_shapes: + # self._unique_shapes.add(shape_key) + # self._shape_to_layer_types[shape_key] = [] + # print(f"Layer: {self.prefix} | input shape: {x.shape} --> output shape: {output.shape}, Quant Method: {self.quant_method.__class__.__name__}") + + # Add this layer type to the list for this shape if not already present + # layer_type = self.__class__.__name__ + # if layer_type not in self._shape_to_layer_types[shape_key]: + # self._shape_to_layer_types[shape_key].append(layer_type) + output_bias = self.bias if self.skip_bias_add else None return output, output_bias @@ -279,6 +293,28 @@ class ReplicatedLinear(LinearBase): s += f", output_features={self.output_size}" s += f", bias={self.bias is not None}" return s + + @classmethod + def get_shape_mapping(cls) -> dict: + """Get the mapping from (input_shape, output_shape) to layer types.""" + return cls._shape_to_layer_types.copy() + + @classmethod + def print_shape_summary(cls): + """Print a summary of all unique shapes and their layer types.""" + if not cls._shape_to_layer_types: + print("No shapes have been processed yet.") + return + + print(f"\n=== Matrix Multiplication Shape Summary ===") + print(f"Total unique shapes: {len(cls._shape_to_layer_types)}") + print() + + for i, (shape_key, layer_types) in enumerate(cls._shape_to_layer_types.items(), 1): + input_shape, output_shape = shape_key + print(f"{i}. Input: {input_shape} → Output: {output_shape}") + print(f" Layer types: {', '.join(layer_types)}") + print() class ColumnParallelLinear(LinearBase): @@ -340,7 +376,9 @@ class ColumnParallelLinear(LinearBase): if output_sizes is None: output_sizes = [output_size] - assert self.quant_method is not None + if self.quant_method is None: + self.quant_method = UnquantizedLinearMethod() + self.quant_method.create_weights( layer=self, input_size_per_partition=self.input_size_per_partition, @@ -348,14 +386,16 @@ class ColumnParallelLinear(LinearBase): input_size=self.input_size, output_size=self.output_size, params_dtype=self.params_dtype, - weight_loader=(self.weight_loader_v2 if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED - else self.weight_loader), + weight_loader=( + self.weight_loader_v2 if self.quant_method.__class__.__name__ + in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader), ) if bias: - self.bias = Parameter(torch.empty( - self.output_size_per_partition, - dtype=params_dtype, - )) + self.bias = Parameter( + torch.empty( + self.output_size_per_partition, + dtype=params_dtype, + )) set_weight_attrs( self.bias, { @@ -399,7 +439,8 @@ class ColumnParallelLinear(LinearBase): bias = self.bias if not self.skip_bias_add else None # Matrix multiply. - assert self.quant_method is not None + if self.quant_method is None: + self.quant_method = UnquantizedLinearMethod() output_parallel = self.quant_method.apply(self, input_, bias) # All-gather across the partitions if needed. output = tensor_model_parallel_all_gather(output_parallel) if self.gather_output else output_parallel @@ -556,8 +597,9 @@ class MergedColumnParallelLinear(ColumnParallelLinear): # for the packing. if (isinstance(param, PackedColumnParameter | PackedvLLMParameter) and param.packed_dim == param.output_dim): - shard_size, shard_offset = (param.adjust_shard_indexes_for_packing(shard_size=shard_size, - shard_offset=shard_offset)) + shard_size, shard_offset = ( + param.adjust_shard_indexes_for_packing( + shard_size=shard_size, shard_offset=shard_offset)) loaded_weight_shard = loaded_weight.narrow(param.output_dim, shard_offset, shard_size) self.weight_loader_v2(param, loaded_weight_shard, shard_id) @@ -665,7 +707,8 @@ class QKVParallelLinear(ColumnParallelLinear): self.num_kv_heads = divide(self.total_num_kv_heads, tp_size) self.num_kv_head_replicas = 1 input_size = self.hidden_size - output_size = ((self.num_heads + 2 * self.num_kv_heads) * tp_size * self.head_size) + output_size = ((self.num_heads + 2 * self.num_kv_heads) * tp_size * + self.head_size) self.output_sizes = [ self.num_heads * self.head_size * tp_size, # q_proj self.num_kv_heads * self.head_size * tp_size, # k_proj @@ -720,7 +763,8 @@ class QKVParallelLinear(ColumnParallelLinear): ), ( "v", - (self.total_num_heads + self.total_num_kv_heads) * self.head_size, + (self.total_num_heads + self.total_num_kv_heads) * + self.head_size, self.total_num_kv_heads * self.head_size, ), ] @@ -731,8 +775,9 @@ class QKVParallelLinear(ColumnParallelLinear): # for the packing. if (isinstance(param, PackedColumnParameter | PackedvLLMParameter) and param.packed_dim == param.output_dim): - shard_size, shard_offset = (param.adjust_shard_indexes_for_packing(shard_size=shard_size, - shard_offset=shard_offset)) + shard_size, shard_offset = ( + param.adjust_shard_indexes_for_packing( + shard_size=shard_size, shard_offset=shard_offset)) loaded_weight_shard = loaded_weight.narrow(param.output_dim, shard_offset, shard_size) self.weight_loader_v2(param, loaded_weight_shard, shard_id) @@ -801,13 +846,15 @@ class QKVParallelLinear(ColumnParallelLinear): ), ( "v", - (self.total_num_heads + self.total_num_kv_heads) * self.head_size, + (self.total_num_heads + self.total_num_kv_heads) * + self.head_size, self.total_num_kv_heads * self.head_size, ), ] for shard_id, shard_offset, shard_size in shard_offsets: - loaded_weight_shard = loaded_weight.narrow(output_dim, shard_offset, shard_size) + loaded_weight_shard = loaded_weight.narrow( + output_dim, shard_offset, shard_size) self.weight_loader(param, loaded_weight_shard, shard_id) return @@ -916,7 +963,9 @@ class RowParallelLinear(LinearBase): self.input_is_parallel = input_is_parallel self.reduce_results = reduce_results - assert self.quant_method is not None + if self.quant_method is None: + self.quant_method = UnquantizedLinearMethod() + self.quant_method.create_weights( layer=self, input_size_per_partition=self.input_size_per_partition, @@ -924,15 +973,17 @@ class RowParallelLinear(LinearBase): input_size=self.input_size, output_size=self.output_size, params_dtype=self.params_dtype, - weight_loader=(self.weight_loader_v2 if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED - else self.weight_loader), + weight_loader=( + self.weight_loader_v2 if self.quant_method.__class__.__name__ + in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader), ) if not reduce_results and (bias and not skip_bias_add): raise ValueError("When not reduce the results, adding bias to the " "results can lead to incorrect results") if bias: - self.bias = Parameter(torch.empty(self.output_size, dtype=params_dtype)) + self.bias = Parameter( + torch.empty(self.output_size, dtype=params_dtype)) set_weight_attrs( self.bias, { @@ -965,7 +1016,8 @@ class RowParallelLinear(LinearBase): assert param_data.shape == loaded_weight.shape param_data.copy_(loaded_weight) - def weight_loader_v2(self, param: BasevLLMParameter, loaded_weight: torch.Tensor): + def weight_loader_v2(self, param: BasevLLMParameter, + loaded_weight: torch.Tensor): # Special case for loading scales off disk, which often do not # have a shape (such as in the case of AutoFP8). if len(loaded_weight.shape) == 0: @@ -983,7 +1035,8 @@ class RowParallelLinear(LinearBase): input_parallel = splitted_input[tp_rank].contiguous() # Matrix multiply. - assert self.quant_method is not None + if self.quant_method is None: + self.quant_method = UnquantizedLinearMethod() # Only fuse bias add into GEMM for rank 0 (this ensures that # bias will not get added more than once in TP>1 case) bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias @@ -1004,3 +1057,4 @@ class RowParallelLinear(LinearBase): s += f", tp_size={self.tp_size}" s += f", reduce_results={self.reduce_results}" return s + diff --git a/fastvideo/layers/mlp.py b/fastvideo/layers/mlp.py index 7610a775d..89b89c29e 100644 --- a/fastvideo/layers/mlp.py +++ b/fastvideo/layers/mlp.py @@ -5,6 +5,7 @@ import torch.nn as nn from fastvideo.layers.activation import get_act_fn from fastvideo.layers.linear import ReplicatedLinear +from fastvideo.layers.quantization import QuantizationConfig class MLP(nn.Module): @@ -21,21 +22,38 @@ class MLP(nn.Module): act_type: str = "gelu_pytorch_tanh", dtype: torch.dtype | None = None, prefix: str = "", + quant_config: QuantizationConfig | None = None, ): super().__init__() self.fc_in = ReplicatedLinear( input_dim, mlp_hidden_dim, # For activation func like SiLU that need 2x width bias=bias, - params_dtype=dtype) + params_dtype=dtype, + quant_config=quant_config, + prefix=f"{prefix}.fc_in", + ) + if quant_config is not None: + quant_method = self.fc_in.quant_config.get_quant_method(self.fc_in, f"{prefix}.fc_in") + if quant_method is not None: + quant_method.process_weights_after_loading(self.fc_in) self.act = get_act_fn(act_type) if output_dim is None: output_dim = input_dim - self.fc_out = ReplicatedLinear(mlp_hidden_dim, output_dim, bias=bias, params_dtype=dtype) + self.fc_out = ReplicatedLinear(mlp_hidden_dim, + output_dim, + bias=bias, + params_dtype=dtype, + quant_config=quant_config, + prefix=f"{prefix}.fc_out") + if quant_config is not None: + quant_method = self.fc_out.quant_config.get_quant_method(self.fc_out, f"{prefix}.fc_out") + if quant_method is not None: + quant_method.process_weights_after_loading(self.fc_out) def forward(self, x: torch.Tensor) -> torch.Tensor: x, _ = self.fc_in(x) x = self.act(x) x, _ = self.fc_out(x) - return x + return x \ No newline at end of file diff --git a/fastvideo/layers/quantization/fp4_config.py b/fastvideo/layers/quantization/fp4_config.py new file mode 100644 index 000000000..aa6819c41 --- /dev/null +++ b/fastvideo/layers/quantization/fp4_config.py @@ -0,0 +1,108 @@ +# SPDX-License-Identifier: Apache-2.0 +from fastvideo.layers.quantization.base_config import QuantizationConfig, QuantizeMethodBase +import torch +from torch.nn.parameter import Parameter + +from typing import Any, Tuple +from flashinfer import nvfp4_quantize, mm_fp4, SfLayout +from fastvideo.models.utils import set_weight_attrs +import logging +logger = logging.getLogger(__name__) + +class FP4QuantizeMethod(QuantizeMethodBase): + def __init__(self): + super().__init__() + self.weight_fp4 = None + self.weight_scale = None + + def create_weights(self, layer: torch.nn.Module, + input_size_per_partition: int, + output_partition_sizes: list[int], input_size: int, + output_size: int, params_dtype: torch.dtype, + **extra_weight_attrs): + """Create weights for a linear layer. Note the corrected signature to match LinearMethodBase.""" + weight = Parameter(torch.empty( + sum(output_partition_sizes), + input_size_per_partition, + dtype=params_dtype, + ), + requires_grad=False) + set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0}) + layer.register_parameter("weight", weight) + set_weight_attrs(weight, extra_weight_attrs) + + @torch.compile + def apply(self, layer: torch.nn.Module, x: torch.Tensor, bias: torch.Tensor | None = None) -> torch.Tensor: + """Apply FP4 quantized computation.""" + out_dim = layer.weight.shape[0] + original_shape = x.shape + assert x.dtype == torch.bfloat16 or x.dtype == torch.float16, f"only allow bf16/fp16 inputs to fp4 linear, got {x.dtype}" + + x = x.view(-1, x.shape[-1]) + + x_global_sf = (448 * 6) / x.float().abs().nan_to_num().max(); + x_fp4, x_scale = nvfp4_quantize(x, x_global_sf, sfLayout=SfLayout.layout_128x4, do_shuffle=False) + weight_fp4 = layer._fp4_weight + weight_scale = layer._fp4_weight_scale + weight_global_sf = layer._weight_global_sf + + out = mm_fp4(x_fp4, weight_fp4.T, x_scale, weight_scale.T, 1.0/(x_global_sf * weight_global_sf), torch.bfloat16, None, backend='cutlass') + + if bias is not None: + if bias.device != out.device or bias.dtype != out.dtype: + bias = bias.to(device=out.device, dtype=out.dtype) + out = out + bias + + if len(original_shape) == 3: + out = out.view(original_shape[0], original_shape[1], out_dim) + + return out + + +class FP4Config(QuantizationConfig): + def __init__(self): + super().__init__() + + def get_name(self): + return "fp4" + + def get_supported_act_dtypes(self): + return [torch.bfloat16, torch.float16] + + @classmethod + def get_min_capability(cls): + return 100 + + @staticmethod + def get_config_filenames(): + return [] + + @classmethod + def from_config(cls, config: dict[str, Any]) -> "FP4Config": + return cls() + + def get_quant_method(self, layer: torch.nn.Module, prefix: str): + from fastvideo.layers.linear import LinearBase + fp4_layers = ["ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out"] + if isinstance(layer, LinearBase) and any(layer_name in prefix for layer_name in fp4_layers): + return FP4QuantizeMethod() + return None + +@torch.compile +def convert_model_to_fp4(model: torch.nn.Module): + from torch.distributed.tensor import DTensor # type: ignore + for mod in model.modules(): + qm = getattr(mod, "quant_method", None) + if isinstance(qm, FP4QuantizeMethod): + weight = getattr(mod, "weight", None) + if weight is None: + continue + if isinstance(weight, DTensor): # type: ignore + weight_local = weight.to_local() + else: + weight_local = weight + weight_global_sf = (448 * 6) / weight_local.float().abs().nan_to_num().max(); + fp4_w, fp4_s = nvfp4_quantize(weight_local, weight_global_sf, sfLayout=SfLayout.layout_128x4, do_shuffle=False) + mod.register_buffer("_fp4_weight", fp4_w, persistent=False) + mod.register_buffer("_fp4_weight_scale", fp4_s, persistent=False) + mod.register_buffer("_weight_global_sf", torch.tensor(weight_global_sf, dtype=torch.bfloat16), persistent=False) \ No newline at end of file diff --git a/fastvideo/layers/quantization/fp8_config.py b/fastvideo/layers/quantization/fp8_config.py new file mode 100644 index 000000000..25e938f59 --- /dev/null +++ b/fastvideo/layers/quantization/fp8_config.py @@ -0,0 +1,109 @@ +# # SPDX-License-Identifier: Apache-2.0 +# from fastvideo.layers.quantization.base_config import QuantizationConfig, QuantizeMethodBase +# import torch +# from torch.nn.parameter import Parameter + +# from typing import Any, Tuple +# import deep_gemm +# from deep_gemm import ceil_div, get_mn_major_tma_aligned_tensor +# from deep_gemm.utils import per_token_cast_to_fp8, per_block_cast_to_fp8 +# from fastvideo.models.utils import set_weight_attrs + +# class FP8QuantizeMethod(QuantizeMethodBase): +# def __init__(self): +# super().__init__() +# self.weight_fp8 = None +# self.weight_scale = None + +# def create_weights(self, layer: torch.nn.Module, +# input_size_per_partition: int, +# output_partition_sizes: list[int], input_size: int, +# output_size: int, params_dtype: torch.dtype, +# **extra_weight_attrs): +# """Create weights for a linear layer. Note the corrected signature to match LinearMethodBase.""" +# weight = Parameter(torch.empty( +# sum(output_partition_sizes), +# input_size_per_partition, +# dtype=params_dtype, +# ), +# requires_grad=False) +# set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0}) +# layer.register_parameter("weight", weight) +# set_weight_attrs(weight, extra_weight_attrs) + +# @torch.compile +# def apply(self, layer: torch.nn.Module, x: torch.Tensor, bias: torch.Tensor | None = None) -> torch.Tensor: +# """Apply FP8 quantized computation.""" +# out_dim = layer.weight.shape[0] +# # Need contiguous tensors for collectives. +# assert x.dtype == torch.bfloat16 or x.dtype == torch.float16 or x.dtype == torch.float32, f"only allow bf16/fp16/fp32 inputs to fp8 linear, got {x.dtype}" +# x_fp8, x_scale = per_token_cast_to_fp8(x.view(-1, x.shape[-1]), use_ue8m0=False) +# x_scale = get_mn_major_tma_aligned_tensor(x_scale) +# weight_fp8 = layer._fp8_weight +# weight_scale = layer._fp8_weight_scale + +# original_shape = x.shape +# out = torch.zeros((x_fp8.shape[0], out_dim), device=x.device, dtype=x.dtype) +# deep_gemm.fp8_gemm_nt( +# (x_fp8, x_scale), +# (weight_fp8, weight_scale), +# out, +# disable_ue8m0_cast=False +# ) + +# if bias is not None: +# if bias.device != out.device or bias.dtype != out.dtype: +# bias = bias.to(device=out.device, dtype=out.dtype) +# out = out + bias + +# if len(original_shape) == 3: +# out = out.view(original_shape[0], original_shape[1], out_dim) + +# return out + + +# class FP8Config(QuantizationConfig): +# def __init__(self): +# super().__init__() + +# def get_name(self): +# return "fp8" + +# def get_supported_act_dtypes(self): +# return [torch.bfloat16, torch.float16, torch.float32] + +# @classmethod +# def get_min_capability(cls): +# return 90 + +# @staticmethod +# def get_config_filenames(): +# return [] + +# @classmethod +# def from_config(cls, config: dict[str, Any]) -> "FP8Config": +# return cls() + +# def get_quant_method(self, layer: torch.nn.Module, prefix: str): +# from fastvideo.layers.linear import LinearBase +# fp8_layers = ["ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out"] +# if isinstance(layer, LinearBase) and any(layer_name in prefix for layer_name in fp8_layers): +# return FP8QuantizeMethod() +# return None + +# @torch.compile +# def convert_model_to_fp8(model: torch.nn.Module): +# from torch.distributed.tensor import DTensor # type: ignore +# for mod in model.modules(): +# qm = getattr(mod, "quant_method", None) +# if isinstance(qm, FP8QuantizeMethod): +# weight = getattr(mod, "weight", None) +# if weight is None: +# continue +# if isinstance(weight, DTensor): # type: ignore +# weight_local = weight.to_local() +# else: +# weight_local = weight +# fp8_w, fp8_s = per_block_cast_to_fp8(weight_local, use_ue8m0=False) +# mod.register_buffer("_fp8_weight", fp8_w, persistent=False) +# mod.register_buffer("_fp8_weight_scale", fp8_s, persistent=False) \ No newline at end of file diff --git a/fastvideo/models/dits/hunyuanvideo15.py b/fastvideo/models/dits/hunyuanvideo15.py index b8e5562eb..e69de29bb 100644 --- a/fastvideo/models/dits/hunyuanvideo15.py +++ b/fastvideo/models/dits/hunyuanvideo15.py @@ -1,766 +0,0 @@ -# Copyright 2025 The Hunyuan Team and The HuggingFace Team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from typing import Any, Dict, Optional, List - -import torch -import torch.nn as nn - -from fastvideo.attention import DistributedAttention, LocalAttention -from fastvideo.forward_context import set_forward_context -from fastvideo.distributed.communication_op import ( - sequence_model_parallel_all_gather_with_unpad, - sequence_model_parallel_shard) -from fastvideo.configs.models.dits import HunyuanVideo15Config -from fastvideo.layers.layernorm import (LayerNormScaleShift, ScaleResidual, - ScaleResidualLayerNormScaleShift) -from fastvideo.layers.linear import ReplicatedLinear -# TODO(will-PY-refactor): RMSNorm .... -from fastvideo.layers.mlp import MLP -from fastvideo.layers.rotary_embedding import get_rotary_pos_embed -from fastvideo.layers.visual_embedding import (ModulateProjection, PatchEmbed, - TimestepEmbedder, unpatchify) -from fastvideo.models.dits.base import BaseDiT -from fastvideo.platforms import AttentionBackendEnum -from fastvideo.logger import init_logger - -from fastvideo.distributed.parallel_state import get_sp_world_size - -logger = init_logger(__name__) - -class HunyuanRMSNorm(nn.Module): - - def __init__( - self, - dim: int, - elementwise_affine=True, - eps: float = 1e-6, - device=None, - dtype=None, - ): - """ - Initialize the RMSNorm normalization layer. - - Args: - dim (int): The dimension of the input tensor. - eps (float, optional): A small value added to the denominator for numerical stability. Default is 1e-6. - - Attributes: - eps (float): A small value added to the denominator for numerical stability. - weight (nn.Parameter): Learnable scaling parameter. - - """ - factory_kwargs = {"device": device, "dtype": dtype} - super().__init__() - self.eps = eps - if elementwise_affine: - self.weight = nn.Parameter(torch.ones(dim, **factory_kwargs)) - - def _norm(self, x) -> torch.Tensor: - """ - Apply the RMSNorm normalization to the input tensor. - - Args: - x (torch.Tensor): The input tensor. - - Returns: - torch.Tensor: The normalized tensor. - - """ - return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) - - def forward(self, x): - """ - Forward pass through the RMSNorm layer. - - Args: - x (torch.Tensor): The input tensor. - - Returns: - torch.Tensor: The output tensor after applying RMSNorm. - - """ - output = self._norm(x.float()).type_as(x) - if hasattr(self, "weight"): - output = output * self.weight - return output - - -class HunyuanVideo15TimeEmbedding(nn.Module): - r""" - Time embedding for HunyuanVideo 1.5. - - Supports standard timestep embedding and optional reference timestep embedding for MeanFlow-based super-resolution - models. - - Args: - embedding_dim (`int`): - The dimension of the output embedding. - """ - - def __init__(self, embedding_dim: int, use_meanflow: bool = False): - super().__init__() - - self.timestep_embedder = TimestepEmbedder(hidden_size=embedding_dim) - - self.use_meanflow = use_meanflow - self.time_proj_r = None - self.timestep_embedder_r = None - if use_meanflow: - self.timestep_embedder_r = TimestepEmbedder(hidden_size=embedding_dim) - - def forward( - self, - timestep: torch.Tensor, - timestep_r: Optional[torch.Tensor] = None, - ) -> torch.Tensor: - timesteps_emb = self.timestep_embedder(timestep) - - if timestep_r is not None: - timesteps_emb_r = self.timestep_embedder_r(timestep_r) - timesteps_emb = timesteps_emb + timesteps_emb_r - - return timesteps_emb - - -class HunyuanVideo15ByT5TextProjection(nn.Module): - def __init__(self, in_features: int, hidden_size: int, out_features: int): - super().__init__() - self.norm = nn.LayerNorm(in_features) - self.linear_1 = nn.Linear(in_features, hidden_size) - self.linear_2 = nn.Linear(hidden_size, hidden_size) - self.linear_3 = nn.Linear(hidden_size, out_features) - self.act_fn = nn.GELU() - - def forward(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor: - hidden_states = self.norm(encoder_hidden_states) - hidden_states = self.linear_1(hidden_states) - hidden_states = self.act_fn(hidden_states) - hidden_states = self.linear_2(hidden_states) - hidden_states = self.act_fn(hidden_states) - hidden_states = self.linear_3(hidden_states) - return hidden_states - - -class HunyuanVideo15ImageProjection(nn.Module): - def __init__(self, in_channels: int, hidden_size: int): - super().__init__() - self.norm_in = nn.LayerNorm(in_channels) - self.linear_1 = nn.Linear(in_channels, in_channels) - self.act_fn = nn.GELU() - self.linear_2 = nn.Linear(in_channels, hidden_size) - self.norm_out = nn.LayerNorm(hidden_size) - - def forward(self, image_embeds: torch.Tensor) -> torch.Tensor: - hidden_states = self.norm_in(image_embeds) - hidden_states = self.linear_1(hidden_states) - hidden_states = self.act_fn(hidden_states) - hidden_states = self.linear_2(hidden_states) - hidden_states = self.norm_out(hidden_states) - return hidden_states - - -class MMDoubleStreamBlock(nn.Module): - """ - A multimodal DiT block with separate modulation for text and image/video, - using distributed attention and linear layers. - """ - - def __init__( - self, - hidden_size: int, - num_attention_heads: int, - mlp_ratio: float, - dtype: torch.dtype | None = None, - supported_attention_backends: tuple[AttentionBackendEnum, ...] - | None = None, - prefix: str = "", - ): - super().__init__() - - self.deterministic = False - self.num_attention_heads = num_attention_heads - head_dim = hidden_size // num_attention_heads - mlp_hidden_dim = int(hidden_size * mlp_ratio) - - # Image modulation components - self.img_mod = ModulateProjection( - hidden_size, - factor=6, - act_layer="silu", - dtype=dtype, - prefix=f"{prefix}.img_mod", - ) - - # Fused operations for image stream - self.img_attn_norm = LayerNormScaleShift(hidden_size, - norm_type="layer", - elementwise_affine=False, - dtype=dtype) - self.img_attn_residual_mlp_norm = ScaleResidualLayerNormScaleShift( - hidden_size, - norm_type="layer", - elementwise_affine=False, - dtype=dtype) - self.img_mlp_residual = ScaleResidual() - - # Image attention components - self.img_attn_qkv = ReplicatedLinear(hidden_size, - hidden_size * 3, - bias=True, - params_dtype=dtype, - prefix=f"{prefix}.img_attn_qkv") - - self.img_attn_q_norm = HunyuanRMSNorm(head_dim, eps=1e-6, dtype=dtype) - self.img_attn_k_norm = HunyuanRMSNorm(head_dim, eps=1e-6, dtype=dtype) - - self.img_attn_proj = ReplicatedLinear(hidden_size, - hidden_size, - bias=True, - params_dtype=dtype, - prefix=f"{prefix}.img_attn_proj") - - self.img_mlp = MLP(hidden_size, - mlp_hidden_dim, - bias=True, - dtype=dtype, - prefix=f"{prefix}.img_mlp") - - # Text modulation components - self.txt_mod = ModulateProjection( - hidden_size, - factor=6, - act_layer="silu", - dtype=dtype, - prefix=f"{prefix}.txt_mod", - ) - - # Fused operations for text stream - self.txt_attn_norm = LayerNormScaleShift(hidden_size, - norm_type="layer", - elementwise_affine=False, - dtype=dtype) - self.txt_attn_residual_mlp_norm = ScaleResidualLayerNormScaleShift( - hidden_size, - norm_type="layer", - elementwise_affine=False, - dtype=dtype) - self.txt_mlp_residual = ScaleResidual() - - # Text attention components - self.txt_attn_qkv = ReplicatedLinear(hidden_size, - hidden_size * 3, - bias=True, - params_dtype=dtype) - - # QK norm layers for text - self.txt_attn_q_norm = HunyuanRMSNorm(head_dim, eps=1e-6, dtype=dtype) - self.txt_attn_k_norm = HunyuanRMSNorm(head_dim, eps=1e-6, dtype=dtype) - - self.txt_attn_proj = ReplicatedLinear(hidden_size, - hidden_size, - bias=True, - params_dtype=dtype) - - self.txt_mlp = MLP(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype) - - # Distributed attention - self.attn = DistributedAttention( - num_heads=num_attention_heads, - head_size=head_dim, - causal=False, - supported_attention_backends=supported_attention_backends, - prefix=f"{prefix}.attn") - - def forward( - self, - img: torch.Tensor, - txt: torch.Tensor, - vec: torch.Tensor, - freqs_cis: tuple, - original_seq_len: int, - ) -> tuple[torch.Tensor, torch.Tensor]: - # Process modulation vectors - img_mod_outputs = self.img_mod(vec) - ( - img_attn_shift, - img_attn_scale, - img_attn_gate, - img_mlp_shift, - img_mlp_scale, - img_mlp_gate, - ) = torch.chunk(img_mod_outputs, 6, dim=-1) - - txt_mod_outputs = self.txt_mod(vec) - ( - txt_attn_shift, - txt_attn_scale, - txt_attn_gate, - txt_mlp_shift, - txt_mlp_scale, - txt_mlp_gate, - ) = torch.chunk(txt_mod_outputs, 6, dim=-1) - - # Prepare image for attention using fused operation - img_attn_input = self.img_attn_norm(img, img_attn_shift, img_attn_scale) - # Get QKV for image - img_qkv, _ = self.img_attn_qkv(img_attn_input) - batch_size, image_seq_len = img_qkv.shape[0], img_qkv.shape[1] - - # Split QKV - img_qkv = img_qkv.view(batch_size, image_seq_len, 3, - self.num_attention_heads, -1) - img_q, img_k, img_v = img_qkv[:, :, 0], img_qkv[:, :, 1], img_qkv[:, :, - 2] - - # Apply QK-Norm if needed - img_q = self.img_attn_q_norm(img_q).to(img_v) - img_k = self.img_attn_k_norm(img_k).to(img_v) - - # Prepare text for attention using fused operation - txt_attn_input = self.txt_attn_norm(txt, txt_attn_shift, txt_attn_scale) - - # Get QKV for text - txt_qkv, _ = self.txt_attn_qkv(txt_attn_input) - batch_size, text_seq_len = txt_qkv.shape[0], txt_qkv.shape[1] - - # Split QKV - txt_qkv = txt_qkv.view(batch_size, text_seq_len, 3, - self.num_attention_heads, -1) - txt_q, txt_k, txt_v = txt_qkv[:, :, 0], txt_qkv[:, :, 1], txt_qkv[:, :, - 2] - # Apply QK-Norm if needed - txt_q = self.txt_attn_q_norm(txt_q).to(txt_q.dtype) - txt_k = self.txt_attn_k_norm(txt_k).to(txt_k.dtype) - - - img_attn, txt_attn = self.attn(img_q, img_k, img_v, original_seq_len, txt_q, txt_k, txt_v, freqs_cis=freqs_cis) - - img_attn_out, _ = self.img_attn_proj( - img_attn.view(batch_size, image_seq_len, -1)) - # Use fused operation for residual connection, normalization, and modulation - img_mlp_input, img_residual = self.img_attn_residual_mlp_norm( - img, img_attn_out, img_attn_gate, img_mlp_shift, img_mlp_scale) - - # Process image MLP - img_mlp_out = self.img_mlp(img_mlp_input) - img = self.img_mlp_residual(img_residual, img_mlp_out, img_mlp_gate) - - # Process text attention output - txt_attn_out, _ = self.txt_attn_proj( - txt_attn.reshape(batch_size, text_seq_len, -1)) - - # Use fused operation for residual connection, normalization, and modulation - txt_mlp_input, txt_residual = self.txt_attn_residual_mlp_norm( - txt, txt_attn_out, txt_attn_gate, txt_mlp_shift, txt_mlp_scale) - - # Process text MLP - txt_mlp_out = self.txt_mlp(txt_mlp_input) - txt = self.txt_mlp_residual(txt_residual, txt_mlp_out, txt_mlp_gate) - - return img, txt - - -class HunyuanVideo15Transformer3DModel(BaseDiT): - r""" - A Transformer model for video-like data used in [HunyuanVideo1.5](https://huggingface.co/tencent/HunyuanVideo1.5). - """ - - # shard single stream, double stream blocks, and refiner_blocks - _fsdp_shard_conditions = HunyuanVideo15Config()._fsdp_shard_conditions - _compile_conditions = HunyuanVideo15Config()._compile_conditions - _supported_attention_backends = HunyuanVideo15Config( - )._supported_attention_backends - param_names_mapping = HunyuanVideo15Config().param_names_mapping - reverse_param_names_mapping = HunyuanVideo15Config( - ).reverse_param_names_mapping - lora_param_names_mapping = HunyuanVideo15Config().lora_param_names_mapping - - def __init__( - self, - config: HunyuanVideo15Config, - hf_config: dict[str, Any], - ) -> None: - super().__init__(config=config, hf_config=hf_config) - - self.hidden_size = config.hidden_size - self.num_attention_heads = config.num_attention_heads - self.num_channels_latents = config.num_channels_latents - self.out_channels = config.out_channels or config.in_channels - self.patch_size = (config.patch_size_t, config.patch_size, config.patch_size) - - # 1. Latent and condition embedders - self.img_in = PatchEmbed(self.patch_size, - config.in_channels, - self.hidden_size, - prefix=f"{config.prefix}.img_in") - self.image_embedder = HunyuanVideo15ImageProjection(config.image_embed_dim, self.hidden_size) - - self.txt_in = SingleTokenRefiner(config.text_embed_dim, - self.hidden_size, - config.num_attention_heads, - depth=config.num_refiner_layers, - dtype=None, - prefix=f"{config.prefix}.txt_in") - - self.txt_in_2 = HunyuanVideo15ByT5TextProjection(config.text_embed_2_dim, 2048, self.hidden_size) - - self.time_in = HunyuanVideo15TimeEmbedding(self.hidden_size, use_meanflow=config.use_meanflow) - - self.cond_type_embed = nn.Embedding(3, self.hidden_size) - - # 3. Dual stream transformer blocks - - self.double_blocks = nn.ModuleList( - [ - MMDoubleStreamBlock( - hidden_size=self.hidden_size, - num_attention_heads=config.num_attention_heads, - mlp_ratio=config.mlp_ratio, - dtype=None, - supported_attention_backends=self._supported_attention_backends, - prefix=f"{config.prefix}.double_blocks.{i}" - ) - for i in range(config.num_layers) - ] - ) - - # 5. Output projection - self.final_layer = FinalLayer(self.hidden_size, - self.patch_size, - self.out_channels, - prefix=f"{config.prefix}.final_layer") - - self.gradient_checkpointing = False - - self.__post_init__() - - def forward( - self, - hidden_states: torch.Tensor, - encoder_hidden_states: List[torch.Tensor], - timestep: torch.LongTensor, - encoder_hidden_states_image: List[torch.Tensor], - guidance: Optional[torch.Tensor] = None, - timestep_r: Optional[torch.LongTensor] = None, - attention_kwargs: Optional[Dict[str, Any]] = None, - ): - encoder_hidden_states_image = encoder_hidden_states_image[0] - encoder_hidden_states, encoder_hidden_states_2 = encoder_hidden_states - batch_size, num_channels, num_frames, height, width = hidden_states.shape - p_t, p_h, p_w = self.config.patch_size_t, self.config.patch_size, self.config.patch_size - post_patch_num_frames = num_frames // p_t - post_patch_height = height // p_h - post_patch_width = width // p_w - - # 1. RoPE - # Get rotary embeddings - freqs_cos, freqs_sin = get_rotary_pos_embed( - (post_patch_num_frames, post_patch_height, post_patch_width), self.hidden_size, - self.num_attention_heads, self.config.rope_axes_dim, self.config.rope_theta) - freqs_cos = freqs_cos.to(hidden_states.device) - freqs_sin = freqs_sin.to(hidden_states.device) - freqs_cis = (freqs_cos, freqs_sin) if freqs_cos is not None else None - - # 2. Conditional embeddings - temb = self.time_in(timestep, timestep_r=timestep_r) - - hidden_states = self.img_in(hidden_states) - hidden_states, original_seq_len = sequence_model_parallel_shard(hidden_states, dim=1) - - current_seq_len = hidden_states.shape[1] - sp_world_size = get_sp_world_size() - padded_seq_len = current_seq_len * sp_world_size - - - - encoder_hidden_states = self.txt_in(encoder_hidden_states, timestep) - - encoder_hidden_states_cond_emb = self.cond_type_embed( - torch.zeros_like(encoder_hidden_states[:, :, 0], dtype=torch.long) - ) - encoder_hidden_states = encoder_hidden_states + encoder_hidden_states_cond_emb - - # byt5 text embedding - encoder_hidden_states_2 = self.txt_in_2(encoder_hidden_states_2) - - encoder_hidden_states_2_cond_emb = self.cond_type_embed( - torch.ones_like(encoder_hidden_states_2[:, :, 0], dtype=torch.long) - ) - encoder_hidden_states_2 = encoder_hidden_states_2 + encoder_hidden_states_2_cond_emb - - # image embed - encoder_hidden_states_3 = self.image_embedder(encoder_hidden_states_image) - is_t2v = torch.all(encoder_hidden_states_image == 0) - encoder_hidden_states_3_cond_emb = self.cond_type_embed( - 2 - * torch.ones_like( - encoder_hidden_states_3[:, :, 0], - dtype=torch.long, - ) - ) - encoder_hidden_states_3 = encoder_hidden_states_3 + encoder_hidden_states_3_cond_emb - - encoder_hidden_states = torch.cat([encoder_hidden_states_2, encoder_hidden_states], dim=1) if is_t2v else torch.cat([encoder_hidden_states_3, encoder_hidden_states_2, encoder_hidden_states], dim=1) - - # 4. Transformer blocks - if torch.is_grad_enabled() and self.gradient_checkpointing: - for block in self.double_blocks: - hidden_states, encoder_hidden_states = self._gradient_checkpointing_func( - block, - hidden_states, - encoder_hidden_states, - temb, - freqs_cis, - original_seq_len - ) - - else: - for block in self.double_blocks: - hidden_states, encoder_hidden_states = block( - hidden_states, - encoder_hidden_states, - temb, - freqs_cis, - original_seq_len - ) - - # Final layer processing - if get_sp_world_size() > 1: - hidden_states = hidden_states.contiguous() - hidden_states = sequence_model_parallel_all_gather_with_unpad(hidden_states, original_seq_len, dim=1) - hidden_states = self.final_layer(hidden_states, temb) - # Unpatchify to get original shape - hidden_states = unpatchify(hidden_states, post_patch_num_frames, post_patch_height, post_patch_width, self.patch_size, self.out_channels) - - return hidden_states - -class SingleTokenRefiner(nn.Module): - """ - A token refiner that processes text embeddings with attention to improve - their representation for cross-attention with image features. - """ - - def __init__( - self, - in_channels, - hidden_size, - num_attention_heads, - depth=2, - qkv_bias=True, - dtype=None, - prefix: str = "", - ) -> None: - super().__init__() - - # Input projection - # self.input_embedder = ReplicatedLinear( - # in_channels, - # hidden_size, - # bias=True, - # params_dtype=dtype, - # prefix=f"{prefix}.input_embedder") - self.input_embedder = nn.Linear(in_channels, hidden_size, bias=True) - - # Timestep embedding - self.t_embedder = TimestepEmbedder(hidden_size, - act_layer="silu", - dtype=dtype, - prefix=f"{prefix}.t_embedder") - - # Context embedding - self.c_embedder = MLP(in_channels, - hidden_size, - hidden_size, - act_type="silu", - dtype=dtype, - prefix=f"{prefix}.c_embedder") - - # Refiner blocks - self.refiner_blocks = nn.ModuleList([ - IndividualTokenRefinerBlock( - hidden_size, - num_attention_heads, - qkv_bias=qkv_bias, - dtype=dtype, - prefix=f"{prefix}.refiner_blocks.{i}", - ) for i in range(depth) - ]) - - def forward(self, x, t, mask=None): - # Get timestep embeddings - timestep_aware_representations = self.t_embedder(t) - - original_dtype = x.dtype - if mask is None: - context_aware_representations = x.mean(dim=1) - else: - mask_float = mask.float().unsqueeze(-1) # [B, L, 1] - context_aware_representations = (x * mask_float).sum(dim=1) / mask_float.sum(dim=1) - - context_aware_representations = self.c_embedder( - context_aware_representations) - c = timestep_aware_representations + context_aware_representations - # Project input - x = self.input_embedder(x) - # Process through refiner blocks - for block in self.refiner_blocks: - x = block(x, c, mask=mask) - return x - -class IndividualTokenRefinerBlock(nn.Module): - """ - A transformer block for refining individual tokens with self-attention. - """ - - def __init__( - self, - hidden_size, - num_attention_heads, - mlp_ratio=4.0, - qkv_bias=True, - dtype=None, - prefix: str = "", - ) -> None: - super().__init__() - self.num_attention_heads = num_attention_heads - mlp_hidden_dim = int(hidden_size * mlp_ratio) - - # Normalization and attention - self.norm1 = nn.LayerNorm(hidden_size, - eps=1e-6, - elementwise_affine=True, - dtype=dtype) - - self.self_attn_qkv = ReplicatedLinear(hidden_size, - hidden_size * 3, - bias=qkv_bias, - params_dtype=dtype, - prefix=f"{prefix}.self_attn_qkv") - - self.self_attn_proj = ReplicatedLinear( - hidden_size, - hidden_size, - bias=qkv_bias, - params_dtype=dtype, - prefix=f"{prefix}.self_attn_proj") - - # MLP - self.norm2 = nn.LayerNorm(hidden_size, - eps=1e-6, - elementwise_affine=True, - dtype=dtype) - self.mlp = MLP(hidden_size, - mlp_hidden_dim, - bias=True, - act_type="silu", - dtype=dtype, - prefix=f"{prefix}.mlp") - - # Modulation - self.adaLN_modulation = ModulateProjection( - hidden_size, - factor=2, - act_layer="silu", - dtype=dtype, - prefix=f"{prefix}.adaLN_modulation") - - # Scaled dot product attention - self.attn = LocalAttention( - num_heads=num_attention_heads, - head_size=hidden_size // num_attention_heads, - # TODO: remove hardcode - supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, - AttentionBackendEnum.TORCH_SDPA), - ) - - def forward(self, x, c, mask=None): - - # Get modulation parameters - gate_msa, gate_mlp = self.adaLN_modulation(c).chunk(2, dim=-1) - # Self-attention - norm_x = self.norm1(x) - qkv, _ = self.self_attn_qkv(norm_x) - - batch_size, seq_len = qkv.shape[0], qkv.shape[1] - qkv = qkv.view(batch_size, seq_len, 3, self.num_attention_heads, -1) - q, k, v = qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2] - if mask is None: - attn_output = self.attn(q, k, v) - else: - from fastvideo.attention.backends.flash_attn import FlashAttnMetadataBuilder - attn_metadata = FlashAttnMetadataBuilder().build( - current_timestep=0, - attn_mask=mask, - ) - # Run distributed attention - with set_forward_context(current_timestep=0, attn_metadata=attn_metadata): - attn_output = self.attn(q, k, v) # [B, L, H, D] - attn_output = attn_output.reshape(batch_size, seq_len, - -1) - - # Project and apply residual connection with gating - attn_out, _ = self.self_attn_proj(attn_output) - x = x + attn_out * gate_msa.unsqueeze(1) - - # MLP - mlp_out = self.mlp(self.norm2(x)) - x = x + mlp_out * gate_mlp.unsqueeze(1) - - return x - -class FinalLayer(nn.Module): - """ - The final layer of DiT that projects features to pixel space. - """ - - def __init__(self, - hidden_size, - patch_size, - out_channels, - dtype=None, - prefix: str = "") -> None: - super().__init__() - - # Normalization - self.norm_final = nn.LayerNorm(hidden_size, - eps=1e-6, - elementwise_affine=False, - dtype=dtype) - - output_dim = patch_size[0] * patch_size[1] * patch_size[2] * out_channels - - self.linear = ReplicatedLinear(hidden_size, - output_dim, - bias=True, - params_dtype=dtype, - prefix=f"{prefix}.linear") - - # Modulation - self.adaLN_modulation = ModulateProjection( - hidden_size, - factor=2, - act_layer="silu", - dtype=dtype, - prefix=f"{prefix}.adaLN_modulation") - - def forward(self, x, c): - # What the heck HF? Why you change the scale and shift order here??? - scale, shift = self.adaLN_modulation(c).chunk(2, dim=-1) - x = self.norm_final(x) * (1.0 + scale.unsqueeze(1)) + shift.unsqueeze(1) - x, _ = self.linear(x) - return x - -# Entry point for model registry -EntryClass = HunyuanVideo15Transformer3DModel diff --git a/fastvideo/models/dits/matrix_game/__init__.py b/fastvideo/models/dits/matrix_game/__init__.py new file mode 100644 index 000000000..0da9c954c --- /dev/null +++ b/fastvideo/models/dits/matrix_game/__init__.py @@ -0,0 +1,11 @@ +from .model import MatrixGameWanModel, MatrixGameTransformerBlock +from .causal_model import CausalMatrixGameWanModel, CausalMatrixGameTransformerBlock +from .action_module import ActionModule + +__all__ = [ + "MatrixGameWanModel", + "MatrixGameTransformerBlock", + "CausalMatrixGameWanModel", + "CausalMatrixGameTransformerBlock", + "ActionModule", +] diff --git a/fastvideo/models/dits/matrix_game/action_module.py b/fastvideo/models/dits/matrix_game/action_module.py new file mode 100644 index 000000000..ae4d01b4b --- /dev/null +++ b/fastvideo/models/dits/matrix_game/action_module.py @@ -0,0 +1,567 @@ +# SPDX-License-Identifier: Apache-2.0 +# Adapted from Matrix-Game: https://github.com/SkyworkAI/Matrix-Game/blob/main/Matrix-Game-2/wan/modules/action_module.py + +from einops import rearrange +import torch +import torch.nn as nn +import math +from torch.nn.attention.flex_attention import flex_attention + +from fastvideo.attention import LocalAttention +from fastvideo.layers.linear import ReplicatedLinear +from fastvideo.layers.layernorm import FP32LayerNorm, RMSNorm +from fastvideo.layers.rotary_embedding import ( + get_nd_rotary_pos_embed as _fv_get_nd_rotary_pos_embed, + _apply_rotary_emb, +) +from fastvideo.platforms import AttentionBackendEnum + + +DISABLE_COMPILE = False +flex_attention = torch.compile( + flex_attention, dynamic=False, mode="max-autotune-no-cudagraphs") + + +def _get_nd_rotary_pos_embed_matrixgame( + rope_dim_list, + rope_sizes, + theta: float = 10000.0, + theta_rescale_factor: float = 1.0, +): + cos, sin = _fv_get_nd_rotary_pos_embed( + rope_dim_list, + rope_sizes, + theta=theta, + theta_rescale_factor=theta_rescale_factor, + dtype=torch.float32, + ) + # convert from [S, D/2] to [S, D] format + cos = cos.repeat_interleave(2, dim=1) + sin = sin.repeat_interleave(2, dim=1) + return cos, sin + + +def _apply_rotary_emb_qk( + xq: torch.Tensor, + xk: torch.Tensor, + freqs_cos: torch.Tensor, + freqs_sin: torch.Tensor, + start_offset: int = 0, +) -> tuple[torch.Tensor, torch.Tensor]: + seq_len = xq.shape[1] + + # Slice frequencies based on offset + cos = freqs_cos[start_offset:start_offset + seq_len] # [S, D] + sin = freqs_sin[start_offset:start_offset + seq_len] # [S, D] + + # Move to device + cos = cos.to(xq.device) + sin = sin.to(xq.device) + + # Convert from [S, D] (interleaved) back to [S, D/2] + cos_half = cos[:, ::2] # [S, D/2] + sin_half = sin[:, ::2] # [S, D/2] + + # xq/xk are [B, S, H, D], need to reshape for each batch + B, S, H, D = xq.shape + + xq_out = _apply_rotary_emb(xq, cos_half, sin_half, is_neox_style=False) + xk_out = _apply_rotary_emb(xk, cos_half, sin_half, is_neox_style=False) + + return xq_out, xk_out + +class ActionModule(nn.Module): + """ + action module from https://arxiv.org/pdf/2501.08325 + """ + + def __init__( + self, + mouse_dim_in: int = 2, + keyboard_dim_in: int = 6, + hidden_size: int = 128, + img_hidden_size: int = 1536, + keyboard_hidden_dim: int = 1024, + mouse_hidden_dim: int = 1024, + vae_time_compression_ratio: int = 4, + windows_size: int = 3, + heads_num: int = 16, + patch_size: list | None = None, + qk_norm: bool = True, + qkv_bias: bool = False, + rope_dim_list: list | None = None, + rope_theta = 256, + mouse_qk_dim_list: list | None = None, + enable_mouse = True, + enable_keyboard = True, + local_attn_size = 6, + blocks: list | None = None, + ): + super().__init__() + # Initialize mutable defaults + patch_size = patch_size if patch_size is not None else [1, 2, 2] + rope_dim_list = rope_dim_list if rope_dim_list is not None else [8, 28, 28] + mouse_qk_dim_list = mouse_qk_dim_list if mouse_qk_dim_list is not None else [8, 28, 28] + blocks = blocks if blocks is not None else [] + self.local_attn_size = local_attn_size + self.enable_mouse = enable_mouse + self.enable_keyboard = enable_keyboard + + self.rope_dim_list = rope_dim_list + self.rope_theta = rope_theta + if self.enable_keyboard: + self.keyboard_embed = nn.Sequential( + nn.Linear(keyboard_dim_in, hidden_size, bias=True), + nn.SiLU(), + nn.Linear(hidden_size, hidden_size, bias=True) + ) + + self.mouse_qk_dim_list = mouse_qk_dim_list + self.heads_num = heads_num + if self.enable_mouse: + c = mouse_hidden_dim + self.mouse_mlp = nn.Sequential( + nn.Linear(mouse_dim_in * vae_time_compression_ratio * windows_size + img_hidden_size, c, bias=True), + nn.GELU(approximate="tanh"), + nn.Linear(c, c), + FP32LayerNorm(c, elementwise_affine=True), + ) + + head_dim = c // heads_num + self.t_qkv = ReplicatedLinear(c, c*3, bias=qkv_bias) + self.img_attn_q_norm = ( + RMSNorm(head_dim, eps=1e-6) + if qk_norm + else nn.Identity() + ) + self.img_attn_k_norm = ( + RMSNorm(head_dim, eps=1e-6) + if qk_norm + else nn.Identity() + ) + self.proj_mouse = ReplicatedLinear(c, img_hidden_size, bias=qkv_bias) + + if self.enable_keyboard: + head_dim_key = keyboard_hidden_dim // heads_num + self.key_attn_q_norm = ( + RMSNorm(head_dim_key, eps=1e-6) + if qk_norm + else nn.Identity() + ) + self.key_attn_k_norm = ( + RMSNorm(head_dim_key, eps=1e-6) + if qk_norm + else nn.Identity() + ) + + self.mouse_attn_q = ReplicatedLinear(img_hidden_size, keyboard_hidden_dim, bias=qkv_bias) + self.keyboard_attn_kv = ReplicatedLinear(hidden_size * windows_size * vae_time_compression_ratio, keyboard_hidden_dim * 2, bias=qkv_bias) + self.proj_keyboard = ReplicatedLinear(keyboard_hidden_dim, img_hidden_size, bias=qkv_bias) + + self.mouse_attn_layer = LocalAttention( + num_heads=heads_num, + head_size=mouse_hidden_dim // heads_num, + causal=False, + supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, + AttentionBackendEnum.TORCH_SDPA) + ) if self.enable_mouse else None + + self.keyboard_attn_layer = LocalAttention( + num_heads=heads_num, + head_size=keyboard_hidden_dim // heads_num, + causal=False, + supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, + AttentionBackendEnum.TORCH_SDPA) + ) if self.enable_keyboard else None + + self.vae_time_compression_ratio = vae_time_compression_ratio + self.windows_size = windows_size + self.patch_size = patch_size + # Lazy initialization: freqs will be created on first forward pass + self._freqs_cos = None + self._freqs_sin = None + + def patchify(self, x, patch_size): + """ + x : (N C T H W) + """ + pt, ph, pw = self.patch_size + t, h, w = x.shape[2] // pt, x.shape[3] // ph, x.shape[4] // pw + c = x.shape[1] + x = x.reshape(shape=(x.shape[0], c, t , pt, h , ph, w , pw)) + x = torch.einsum("nctohpwq->nthwcopq", x) + x = x.reshape(shape=(x.shape[0], t*h*w, c*pt*ph*pw)) + return x + + def unpatchify(self, x, t, h, w, patch_size): + """ + x: (N, T, patch_size**2 * C) + imgs: (N, H, W, C) + """ + c = x.shape[2] // patch_size #self.unpatchify_channels + pt, ph, pw = self.patch_size + assert t * h * w == x.shape[1] + + x = x.reshape(shape=(x.shape[0], t, h, w, c, pt, ph, pw)) + x = torch.einsum("nthwcopq->nctohpwq", x) + imgs = x.reshape(shape=(x.shape[0], c, t * pt, h * ph, w * pw)) + + return imgs + + def get_rotary_pos_embed(self, video_length, height, width, head_dim, rope_dim_list = None, start_offset=0): + target_ndim = 3 + ndim = 5 - 2 + + latents_size = [video_length+start_offset, height, width] + + if isinstance(self.patch_size, int): + assert all(s % self.patch_size == 0 for s in latents_size), ( + f"Latent size(last {ndim} dimensions) should be divisible by patch size({self.patch_size}), " + f"but got {latents_size}." + ) + rope_sizes = [s // self.patch_size for s in latents_size] + elif isinstance(self.patch_size, list): + assert all( + s % self.patch_size[idx] == 0 + for idx, s in enumerate(latents_size) + ), ( + f"Latent size(last {ndim} dimensions) should be divisible by patch size({self.patch_size}), " + f"but got {latents_size}." + ) + rope_sizes = [ + s // self.patch_size[idx] for idx, s in enumerate(latents_size) + ] + + if len(rope_sizes) != target_ndim: + rope_sizes = [1] * (target_ndim - len(rope_sizes)) + rope_sizes # time axis + + if rope_dim_list is None: + rope_dim_list = [head_dim // target_ndim for _ in range(target_ndim)] + assert ( + sum(rope_dim_list) == head_dim + ), "sum(rope_dim_list) should equal to head_dim of attention layer" + # Use Matrix-Game wrapper for FastVideo's function + freqs_cos, freqs_sin = _get_nd_rotary_pos_embed_matrixgame( + rope_dim_list, + rope_sizes, + theta=self.rope_theta, + theta_rescale_factor=1, + ) + return freqs_cos[-video_length*rope_sizes[1]*rope_sizes[2]//self.patch_size[0]:], freqs_sin[-video_length*rope_sizes[1]*rope_sizes[2]//self.patch_size[0]:] + + def forward(self, x, tt, th, tw, mouse_condition=None, keyboard_condition=None, block_mask_mouse=None, block_mask_keyboard=None, is_causal=False, kv_cache_mouse=None, kv_cache_keyboard=None, start_frame=0, use_rope_keyboard=True, num_frame_per_block=3): + ''' + hidden_states: B, tt*th*tw, C + mouse_condition: B, N_frames, C1 + keyboard_condition: B, N_frames, C2 + ''' + assert use_rope_keyboard + + B, N_frames, C = keyboard_condition.shape + assert tt*th*tw == x.shape[1] + assert ((N_frames - 1) + self.vae_time_compression_ratio) % self.vae_time_compression_ratio == 0 + N_feats = int((N_frames - 1) / self.vae_time_compression_ratio) + 1 + + # Lazy initialization of freqs on first forward pass + if self._freqs_cos is None or self._freqs_sin is None: + self._freqs_cos, self._freqs_sin = self.get_rotary_pos_embed( + 7500, self.patch_size[1], self.patch_size[2], 64, + self.mouse_qk_dim_list, start_offset=0 + ) + + # Defined freqs_cis early so it's available for both mouse and keyboard + freqs_cis = (self._freqs_cos, self._freqs_sin) + + assert (N_feats == tt and ((is_causal and kv_cache_mouse is None) or not is_causal)) or ((N_frames - 1) // self.vae_time_compression_ratio + 1 == start_frame + num_frame_per_block and is_causal) + + if self.enable_mouse and mouse_condition is not None: + hidden_states = rearrange(x, "B (T S) C -> (B S) T C", T=tt, S=th*tw) # 65*272*480 -> 17*(272//16)*(480//16) -> 8670 + B, N_frames, C = mouse_condition.shape + else: + hidden_states = x + # padding + + pad_t = self.vae_time_compression_ratio * self.windows_size + if self.enable_mouse and mouse_condition is not None: + pad = mouse_condition[:, 0:1, :].expand(-1, pad_t, -1) + mouse_condition = torch.cat([pad, mouse_condition], dim=1) + if is_causal and kv_cache_mouse is not None: + mouse_condition = mouse_condition[:, self.vae_time_compression_ratio*(N_feats - num_frame_per_block - self.windows_size) + pad_t:, :] + group_mouse = [mouse_condition[:, self.vae_time_compression_ratio*(i - self.windows_size) + pad_t:i * self.vae_time_compression_ratio + pad_t,:] for i in range(num_frame_per_block)] + else: + group_mouse = [mouse_condition[:, self.vae_time_compression_ratio*(i - self.windows_size) + pad_t:i * self.vae_time_compression_ratio + pad_t,:] for i in range(N_feats)] + + group_mouse = torch.stack(group_mouse, dim = 1) + + S = th * tw + group_mouse = group_mouse.unsqueeze(-1).expand(B, num_frame_per_block, pad_t, C, S) + group_mouse = group_mouse.permute(0, 4, 1, 2, 3).reshape(B * S, num_frame_per_block, pad_t * C) + + group_mouse = torch.cat([hidden_states, group_mouse], dim = -1) + group_mouse = self.mouse_mlp(group_mouse) + # qkv + mouse_qkv, _ = self.t_qkv(group_mouse) + q, k, v = rearrange(mouse_qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num) # BHW F H C + q = self.img_attn_q_norm(q).to(v) + k = self.img_attn_k_norm(k).to(v) + # rope embd + + + # freqs_cis = (self.freqs_cos, self.freqs_sin) + + + q, k = _apply_rotary_emb_qk(q, k, freqs_cis[0], freqs_cis[1], start_offset=start_frame) + ## TODO: adding cache here + if is_causal: + if kv_cache_mouse is None: + assert q.shape[0] == k.shape[0] and q.shape[0] % 880 == 0 # == 880, f"{q.shape[0]},{k.shape[0]}" + padded_length = math.ceil(q.shape[1] / 32) * 32 - q.shape[1] + padded_q = torch.cat( + [q, + torch.zeros([q.shape[0], padded_length, q.shape[2], q.shape[3]], + device=q.device, dtype=v.dtype)], + dim=1 + ) + padded_k = torch.cat( + [k, torch.zeros([k.shape[0], padded_length, k.shape[2], k.shape[3]], + device=k.device, dtype=v.dtype)], + dim=1 + ) + padded_v = torch.cat( + [v, torch.zeros([v.shape[0], padded_length, v.shape[2], v.shape[3]], + device=v.device, dtype=v.dtype)], + dim=1 + ) + attn = flex_attention( + query=padded_q.transpose(2, 1), # after: B, HW, F, C + key=padded_k.transpose(2, 1), + value=padded_v.transpose(2, 1), + block_mask=block_mask_mouse + )[:, :, :-padded_length].transpose(2, 1) + else: + current_start = start_frame + current_end = current_start + q.shape[1] + + assert q.shape[1] == num_frame_per_block + sink_size = 0 + max_attention_size = self.local_attn_size + sink_tokens = sink_size * 1 + kv_cache_size = kv_cache_mouse["k"].shape[1] + num_new_tokens = q.shape[1] + + + if (current_end > kv_cache_mouse["global_end_index"].item()) and ( + num_new_tokens + kv_cache_mouse["local_end_index"].item() > kv_cache_size): + num_evicted_tokens = num_new_tokens + kv_cache_mouse["local_end_index"].item() - kv_cache_size + num_rolled_tokens = kv_cache_mouse["local_end_index"].item() - num_evicted_tokens - sink_tokens + kv_cache_mouse["k"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \ + kv_cache_mouse["k"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone() + kv_cache_mouse["v"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \ + kv_cache_mouse["v"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone() + # Insert the new keys/values at the end + local_end_index = kv_cache_mouse["local_end_index"].item() + current_end - \ + kv_cache_mouse["global_end_index"].item() - num_evicted_tokens + local_start_index = local_end_index - num_new_tokens + else: + local_end_index = kv_cache_mouse["local_end_index"].item() + current_end - kv_cache_mouse["global_end_index"].item() + local_start_index = local_end_index - num_new_tokens + kv_cache_mouse["k"][:, local_start_index:local_end_index] = k + kv_cache_mouse["v"][:, local_start_index:local_end_index] = v + + attn = self.mouse_attn_layer( + q, + kv_cache_mouse["k"][:, max(0, local_end_index - max_attention_size):local_end_index], + kv_cache_mouse["v"][:, max(0, local_end_index - max_attention_size):local_end_index], + ) + kv_cache_mouse["global_end_index"].fill_(current_end) + kv_cache_mouse["local_end_index"].fill_(local_end_index) + else: + attn = self.mouse_attn_layer(q, k, v) + # Compute cu_squlens and max_seqlen for flash attention + # qk norm + attn = rearrange(attn, '(b S) T h d -> b (T S) (h d)',b=B) + + hidden_states = rearrange(x, "(B S) T C -> B (T S) C", B=B) + attn, _ = self.proj_mouse(attn) + + hidden_states = hidden_states + attn + + if self.enable_keyboard and keyboard_condition is not None: + pad = keyboard_condition[:, 0:1, :].expand(-1, pad_t, -1) + keyboard_condition = torch.cat([pad, keyboard_condition], dim=1) + if is_causal and kv_cache_keyboard is not None: + keyboard_condition = keyboard_condition[:, self.vae_time_compression_ratio*(N_feats - num_frame_per_block - self.windows_size) + pad_t:, :] # keyboard_condition[:, self.vae_time_compression_ratio*(start_frame - self.windows_size) + pad_t:start_frame * self.vae_time_compression_ratio + pad_t,:] + keyboard_condition = self.keyboard_embed(keyboard_condition) + group_keyboard = [keyboard_condition[:, self.vae_time_compression_ratio*(i - self.windows_size) + pad_t:i * self.vae_time_compression_ratio + pad_t,:] for i in range(num_frame_per_block)] + else: + keyboard_condition = self.keyboard_embed(keyboard_condition) + group_keyboard = [keyboard_condition[:, self.vae_time_compression_ratio*(i - self.windows_size) + pad_t:i * self.vae_time_compression_ratio + pad_t,:] for i in range(N_feats)] + group_keyboard = torch.stack(group_keyboard, dim = 1) # B F RW C + group_keyboard = group_keyboard.reshape(shape=(group_keyboard.shape[0],group_keyboard.shape[1],-1)) + # apply cross attn + mouse_q, _ = self.mouse_attn_q(hidden_states) + keyboard_kv, _ = self.keyboard_attn_kv(group_keyboard) + + B, L, HD = mouse_q.shape + D = HD // self.heads_num + q = mouse_q.view(B, L, self.heads_num, D) + + B, L, KHD = keyboard_kv.shape + k, v = keyboard_kv.view(B, L, 2, self.heads_num, D).permute(2, 0, 1, 3, 4) + + # Compute cu_squlens and max_seqlen for flash attention + # qk norm + + q = self.key_attn_q_norm(q).to(v) + k = self.key_attn_k_norm(k).to(v) + S = th * tw + assert S == 880 + # position embed + if use_rope_keyboard: + B, TS, H, D = q.shape + T_ = TS // S + q = q.view(B, T_, S, H, D).transpose(1, 2).reshape(B * S, T_, H, D) + q, k = _apply_rotary_emb_qk(q, k, freqs_cis[0], freqs_cis[1], start_offset=start_frame) + + k1, k2, k3, k4 = k.shape + k = k.expand(S, k2, k3, k4) + v = v.expand(S, k2, k3, k4) + + + if is_causal: + if kv_cache_keyboard is None: + assert q.shape[0] == k.shape[0] and q.shape[0] % 880 == 0 + + padded_length = math.ceil(q.shape[1] / 32) * 32 - q.shape[1] + padded_q = torch.cat( + [q, + torch.zeros([q.shape[0], padded_length, q.shape[2], q.shape[3]], + device=q.device, dtype=v.dtype)], + dim=1 + ) + padded_k = torch.cat( + [k, torch.zeros([k.shape[0], padded_length, k.shape[2], k.shape[3]], + device=k.device, dtype=v.dtype)], + dim=1 + ) + padded_v = torch.cat( + [v, torch.zeros([v.shape[0], padded_length, v.shape[2], v.shape[3]], + device=v.device, dtype=v.dtype)], + dim=1 + ) + attn = flex_attention( + query=padded_q.transpose(2, 1), # after: B, HW, F, C + key=padded_k.transpose(2, 1), + value=padded_v.transpose(2, 1), + block_mask=block_mask_keyboard + )[:, :, :-padded_length].transpose(2, 1) + else: + current_start = start_frame + current_end = current_start + k.shape[1] + assert k.shape[1] == num_frame_per_block + sink_size = 0 + max_attention_size = self.local_attn_size + sink_tokens = sink_size * 1 + kv_cache_size = kv_cache_keyboard["k"].shape[1] + num_new_tokens = k.shape[1] + + if (current_end > kv_cache_keyboard["global_end_index"].item()) and ( + num_new_tokens + kv_cache_keyboard["local_end_index"].item() > kv_cache_size): + num_evicted_tokens = num_new_tokens + kv_cache_keyboard["local_end_index"].item() - kv_cache_size + num_rolled_tokens = kv_cache_keyboard["local_end_index"].item() - num_evicted_tokens - sink_tokens + kv_cache_keyboard["k"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \ + kv_cache_keyboard["k"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone() + kv_cache_keyboard["v"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \ + kv_cache_keyboard["v"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone() + # Insert the new keys/values at the end + local_end_index = kv_cache_keyboard["local_end_index"].item() + current_end - \ + kv_cache_keyboard["global_end_index"].item() - num_evicted_tokens + local_start_index = local_end_index - num_new_tokens + else: + local_end_index = kv_cache_keyboard["local_end_index"].item() + current_end - kv_cache_keyboard["global_end_index"].item() + local_start_index = local_end_index - num_new_tokens + assert k.shape[0] == 880 # BS == 1 or the cache should not be saved/ load method should be modified + kv_cache_keyboard["k"][:, local_start_index:local_end_index] = k[:1] + kv_cache_keyboard["v"][:, local_start_index:local_end_index] = v[:1] + + attn = self.keyboard_attn_layer( + q, + kv_cache_keyboard["k"][:, max(0, local_end_index - max_attention_size):local_end_index].repeat(S, 1, 1, 1), + kv_cache_keyboard["v"][:, max(0, local_end_index - max_attention_size):local_end_index].repeat(S, 1, 1, 1), + ) + + kv_cache_keyboard["global_end_index"].fill_(current_end) + kv_cache_keyboard["local_end_index"].fill_(local_end_index) + else: + attn = self.keyboard_attn_layer(q, k, v) + attn = rearrange(attn, '(B S) T H D -> B (T S) (H D)', S=S) + else: + if is_causal: + if kv_cache_keyboard is None: + + padded_length = math.ceil(q.shape[1] / 32) * 32 - q.shape[1] + padded_q = torch.cat( + [q, + torch.zeros([q.shape[0], padded_length, q.shape[2], q.shape[3]], + device=q.device, dtype=v.dtype)], + dim=1 + ) + padded_k = torch.cat( + [k, torch.zeros([k.shape[0], padded_length, k.shape[2], k.shape[3]], + device=k.device, dtype=v.dtype)], + dim=1 + ) + padded_v = torch.cat( + [v, torch.zeros([v.shape[0], padded_length, v.shape[2], v.shape[3]], + device=v.device, dtype=v.dtype)], + dim=1 + ) + attn = flex_attention( + query=padded_q.transpose(2, 1), # after: B, HW, F, C + key=padded_k.transpose(2, 1), + value=padded_v.transpose(2, 1), + block_mask=block_mask_keyboard + )[:, :, :-padded_length].transpose(2, 1) + else: + current_start = start_frame + current_end = current_start + k.shape[1] + assert k.shape[1] == num_frame_per_block + sink_size = 0 + max_attention_size = self.local_attn_size + sink_tokens = sink_size * 1 + kv_cache_size = kv_cache_keyboard["k"].shape[1] + num_new_tokens = k.shape[1] + + + if (current_end > kv_cache_keyboard["global_end_index"].item()) and ( + num_new_tokens + kv_cache_keyboard["local_end_index"].item() > kv_cache_size): + num_evicted_tokens = num_new_tokens + kv_cache_keyboard["local_end_index"].item() - kv_cache_size + num_rolled_tokens = kv_cache_keyboard["local_end_index"].item() - num_evicted_tokens - sink_tokens + kv_cache_keyboard["k"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \ + kv_cache_keyboard["k"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone() + kv_cache_keyboard["v"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \ + kv_cache_keyboard["v"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone() + # Insert the new keys/values at the end + local_end_index = kv_cache_keyboard["local_end_index"].item() + current_end - \ + kv_cache_keyboard["global_end_index"].item() - num_evicted_tokens + local_start_index = local_end_index - num_new_tokens + + + else: + local_end_index = kv_cache_keyboard["local_end_index"].item() + current_end - kv_cache_keyboard["global_end_index"].item() + local_start_index = local_end_index - num_new_tokens + kv_cache_keyboard["k"][:, local_start_index:local_end_index] = k + kv_cache_keyboard["v"][:, local_start_index:local_end_index] = v + attn = self.keyboard_attn_layer( + q, + kv_cache_keyboard["k"][:, max(0, local_end_index - max_attention_size):local_end_index], + kv_cache_keyboard["v"][:, max(0, local_end_index - max_attention_size):local_end_index], + ) + kv_cache_keyboard["global_end_index"].fill_(current_end) + kv_cache_keyboard["local_end_index"].fill_(local_end_index) + else: + attn = self.keyboard_attn_layer(q, k, v) + attn = rearrange(attn, 'B L H D -> B L (H D)') + attn, _ = self.proj_keyboard(attn) + hidden_states = hidden_states + attn + return hidden_states \ No newline at end of file diff --git a/fastvideo/models/dits/matrix_game/causal_model.py b/fastvideo/models/dits/matrix_game/causal_model.py new file mode 100644 index 000000000..fc859e5a8 --- /dev/null +++ b/fastvideo/models/dits/matrix_game/causal_model.py @@ -0,0 +1,1099 @@ +# SPDX-License-Identifier: Apache-2.0 + +import math +from typing import Any + +PYTEST_DONT_REWRITE = True + +import torch +import torch.nn as nn +import torch.distributed as dist + +from torch.nn.attention.flex_attention import create_block_mask, flex_attention +from torch.nn.attention.flex_attention import BlockMask +from fastvideo.forward_context import get_forward_context + +flex_attention = torch.compile( + flex_attention, dynamic=False, mode="max-autotune-no-cudagraphs") + +from fastvideo.attention import LocalAttention +from fastvideo.configs.models.dits.matrixgame import MatrixGameWanVideoConfig +from fastvideo.distributed.parallel_state import get_sp_world_size +from fastvideo.layers.layernorm import (FP32LayerNorm, LayerNormScaleShift, + RMSNorm, ScaleResidual, + ScaleResidualLayerNormScaleShift) +from fastvideo.layers.linear import ReplicatedLinear +from fastvideo.layers.mlp import MLP +from fastvideo.layers.rotary_embedding import (get_rotary_pos_embed) +from fastvideo.layers.visual_embedding import PatchEmbed, TimestepEmbedder, ModulateProjection +from fastvideo.logger import init_logger +from fastvideo.models.dits.base import BaseDiT +from fastvideo.models.dits.wanvideo import (WanT2VCrossAttention, + WanImageEmbedding) +from fastvideo.platforms import AttentionBackendEnum, current_platform + +from .action_module import ActionModule +from .model import MatrixGameCrossAttention + +logger = init_logger(__name__) + + +def causal_rope_apply(x, grid_sizes, freqs, start_frame=0): + n, c = x.size(2), x.size(3) // 2 + + # split freqs + freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1) + + # loop over samples + output = [] + f, h, w = grid_sizes.tolist() + + for i in range(len(x)): + seq_len = f * h * w + + # precompute multipliers + x_i = torch.view_as_complex(x[i, :seq_len].to(torch.float64).reshape( + seq_len, n, -1, 2)) + freqs_i = torch.cat([ + freqs[0][start_frame:start_frame + f].view(f, 1, 1, -1).expand(f, h, w, -1), + freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1), + freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1) + ], + dim=-1).reshape(seq_len, 1, -1) + + # apply rotary embedding + x_i = torch.view_as_real(x_i * freqs_i).flatten(2) + x_i = torch.cat([x_i, x[i, seq_len:]]) + + # append to collection + output.append(x_i) + return torch.stack(output).type_as(x) + + +def rope_params(max_seq_len, dim, theta=10000): + assert dim % 2 == 0 + freqs = torch.outer( + torch.arange(max_seq_len), + 1.0 / torch.pow(theta, + torch.arange(0, dim, 2).to(torch.float64).div(dim))) + freqs = torch.polar(torch.ones_like(freqs), freqs) + return freqs + + +class CausalMatrixGameTimeImageEmbedding(nn.Module): + def __init__( + self, + dim: int, + time_freq_dim: int, + image_embed_dim: int | None = None, + ): + super().__init__() + + self.time_embedder = TimestepEmbedder( + dim, frequency_embedding_size=time_freq_dim, act_layer="silu") + self.time_modulation = ModulateProjection(dim, + factor=6, + act_layer="silu") + + self.image_embedder = None + if image_embed_dim is not None: + self.image_embedder = WanImageEmbedding(image_embed_dim, dim) + + def forward( + self, + timestep: torch.Tensor, + encoder_hidden_states: torch.Tensor, + encoder_hidden_states_image: torch.Tensor | None = None, + timestep_seq_len: int | None = None, + ): + temb = self.time_embedder(timestep, timestep_seq_len) + timestep_proj = self.time_modulation(temb) + + if encoder_hidden_states_image is not None: + assert self.image_embedder is not None + encoder_hidden_states_image = self.image_embedder( + encoder_hidden_states_image) + + return temb, timestep_proj, None, encoder_hidden_states_image + + +class CausalMatrixGameSelfAttention(nn.Module): + + def __init__(self, + dim: int, + num_heads: int, + local_attn_size: int = -1, + sink_size: int = 0, + qk_norm: bool = True, + eps: float = 1e-6) -> None: + assert dim % num_heads == 0 + super().__init__() + self.dim = dim + self.num_heads = num_heads + self.head_dim = dim // num_heads + self.local_attn_size = local_attn_size + self.sink_size = sink_size + self.qk_norm = qk_norm + self.eps = eps + self._freqs_cache = None + + self.attn = LocalAttention( + num_heads=num_heads, + head_size=self.head_dim, + dropout_rate=0, + softmax_scale=None, + causal=False, + supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, + AttentionBackendEnum.TORCH_SDPA)) + + def forward( + self, + q: torch.Tensor, + k: torch.Tensor, + v: torch.Tensor, + freqs_cis: tuple[torch.Tensor, torch.Tensor], + block_mask: BlockMask, + kv_cache: dict | None = None, + current_start: int = 0, + cache_start: int | None = None, + grid_sizes: torch.Tensor | None = None + ): + if cache_start is None: + cache_start = current_start + + # Calculate start_frame for causal mode + if kv_cache is not None and grid_sizes is not None: + frame_seqlen = int(grid_sizes[1] * grid_sizes[2]) + start_frame = current_start // frame_seqlen + else: + start_frame = 0 + + if self._freqs_cache is None or self._freqs_cache.device != q.device: + d = self.head_dim + self._freqs_cache = torch.cat([ + rope_params(1024, d - 4 * (d // 6)), + rope_params(1024, 2 * (d // 6)), + rope_params(1024, 2 * (d // 6)) + ], dim=1).to(q.device) + + freqs = self._freqs_cache + + roped_query = causal_rope_apply(q, grid_sizes, freqs, start_frame=start_frame).type_as(v) + roped_key = causal_rope_apply(k, grid_sizes, freqs, start_frame=start_frame).type_as(v) + + if kv_cache is None: + padded_length = math.ceil(q.shape[1] / 128) * 128 - q.shape[1] + padded_roped_query = torch.cat( + [roped_query, + torch.zeros([q.shape[0], padded_length, q.shape[2], q.shape[3]], + device=q.device, dtype=v.dtype)], + dim=1 + ) + + padded_roped_key = torch.cat( + [roped_key, torch.zeros([k.shape[0], padded_length, k.shape[2], k.shape[3]], + device=k.device, dtype=v.dtype)], + dim=1 + ) + + padded_v = torch.cat( + [v, torch.zeros([v.shape[0], padded_length, v.shape[2], v.shape[3]], + device=v.device, dtype=v.dtype)], + dim=1 + ) + + x = flex_attention( + query=padded_roped_query.transpose(2, 1), + key=padded_roped_key.transpose(2, 1), + value=padded_v.transpose(2, 1), + block_mask=block_mask + )[:, :, :-padded_length].transpose(2, 1) + else: + # Calculate frame_seqlen correctly from grid_sizes (single frame token count) + if grid_sizes is not None: + frame_seqlen = int(grid_sizes[1] * grid_sizes[2]) + else: + # Fallback: assume q.shape[1] is single frame (shouldn't happen in causal mode) + frame_seqlen = q.shape[1] + logger.warning("grid_sizes not provided, using q.shape[1] as frame_seqlen") + + current_end = current_start + roped_query.shape[1] + sink_tokens = self.sink_size * frame_seqlen + + # Compute max_attention_size dynamically based on actual frame_seqlen + max_attention_size = (15 * frame_seqlen if self.local_attn_size == -1 + else self.local_attn_size * frame_seqlen) + + kv_cache_size = kv_cache["k"].shape[1] + num_new_tokens = roped_query.shape[1] + + if (current_end > kv_cache["global_end_index"].item()) and ( + num_new_tokens + kv_cache["local_end_index"].item() > kv_cache_size): + + num_evicted_tokens = num_new_tokens + kv_cache["local_end_index"].item() - kv_cache_size + num_rolled_tokens = kv_cache["local_end_index"].item() - num_evicted_tokens - sink_tokens + kv_cache["k"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \ + kv_cache["k"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone() + kv_cache["v"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \ + kv_cache["v"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone() + local_end_index = kv_cache["local_end_index"].item() + current_end - \ + kv_cache["global_end_index"].item() - num_evicted_tokens + local_start_index = local_end_index - num_new_tokens + kv_cache["k"][:, local_start_index:local_end_index] = roped_key + kv_cache["v"][:, local_start_index:local_end_index] = v + else: + local_end_index = kv_cache["local_end_index"].item() + current_end - kv_cache["global_end_index"].item() + local_start_index = local_end_index - num_new_tokens + + kv_cache["k"][:, local_start_index:local_end_index] = roped_key + kv_cache["v"][:, local_start_index:local_end_index] = v + + kv_start = max(0, local_end_index - max_attention_size) + k_for_attn = kv_cache["k"][:, kv_start:local_end_index] + v_for_attn = kv_cache["v"][:, kv_start:local_end_index] + + x = torch.nn.functional.scaled_dot_product_attention( + roped_query.transpose(1, 2), + k_for_attn.transpose(1, 2), + v_for_attn.transpose(1, 2), + ).transpose(1, 2) + + kv_cache["global_end_index"].fill_(current_end) + kv_cache["local_end_index"].fill_(local_end_index) + + return x + + +class CausalMatrixGameTransformerBlock(nn.Module): + + def __init__(self, + dim: int, + ffn_dim: int, + num_heads: int, + local_attn_size: int = -1, + sink_size: int = 0, + qk_norm: str = "rms_norm_across_heads", + cross_attn_norm: bool = False, + eps: float = 1e-6, + added_kv_proj_dim: int | None = None, + supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, + prefix: str = "", + action_config: dict | None = None, + block_idx: int = 0): + super().__init__() + action_config = action_config or {} + + self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False) + self.to_q = ReplicatedLinear(dim, dim, bias=True) + self.to_k = ReplicatedLinear(dim, dim, bias=True) + self.to_v = ReplicatedLinear(dim, dim, bias=True) + + self.to_out = ReplicatedLinear(dim, dim, bias=True) + self.attn1 = CausalMatrixGameSelfAttention( + dim, + num_heads, + local_attn_size=local_attn_size, + sink_size=sink_size, + qk_norm=qk_norm, + eps=eps) + self.hidden_dim = dim + self.num_attention_heads = num_heads + self.local_attn_size = local_attn_size + dim_head = dim // num_heads + if qk_norm == "rms_norm": + self.norm_q = RMSNorm(dim_head, eps=eps) + self.norm_k = RMSNorm(dim_head, eps=eps) + elif qk_norm == "rms_norm_across_heads": + self.norm_q = RMSNorm(dim, eps=eps) + self.norm_k = RMSNorm(dim, eps=eps) + else: + print("QK Norm type not supported") + raise Exception + assert cross_attn_norm is True + self.self_attn_residual_norm = ScaleResidualLayerNormScaleShift( + dim, + norm_type="layer", + eps=eps, + elementwise_affine=True, + dtype=torch.float32) + + if added_kv_proj_dim is not None: + self.attn2 = MatrixGameCrossAttention(dim, + num_heads, + qk_norm=qk_norm, + eps=eps) + else: + self.attn2 = WanT2VCrossAttention(dim, + num_heads, + qk_norm=qk_norm, + eps=eps) + self.cross_attn_residual_norm = ScaleResidualLayerNormScaleShift( + dim, + norm_type="layer", + eps=eps, + elementwise_affine=False, + dtype=torch.float32) + + + if len(action_config) > 0 and block_idx in action_config.get("blocks", []): + self.action_model = ActionModule( + heads_num=action_config["heads_num"], + hidden_size=action_config["hidden_size"], + img_hidden_size=action_config["img_hidden_size"], + mouse_hidden_dim=action_config.get("mouse_hidden_dim", 0), + keyboard_hidden_dim=action_config["keyboard_hidden_dim"], + mouse_dim_in=action_config.get("mouse_dim_in", 0), + keyboard_dim_in=action_config["keyboard_dim_in"], + enable_mouse=action_config["enable_mouse"], + enable_keyboard=action_config["enable_keyboard"], + windows_size=action_config["windows_size"], + rope_theta=action_config["rope_theta"], + rope_dim_list=action_config["rope_dim_list"], + mouse_qk_dim_list=action_config.get("mouse_qk_dim_list", [8, 28, 28]), + patch_size=action_config["patch_size"], + local_attn_size=local_attn_size, + qk_norm=action_config["qk_norm"], + qkv_bias=action_config["qkv_bias"], + vae_time_compression_ratio=action_config["vae_time_compression_ratio"], + ) + else: + self.action_model = None + + self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh") + self.mlp_residual = ScaleResidual() + + self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + temb: torch.Tensor, + freqs_cis: tuple[torch.Tensor, torch.Tensor], + block_mask: BlockMask, + grid_sizes: torch.Tensor | None = None, + mouse_cond: torch.Tensor | None = None, + keyboard_cond: torch.Tensor | None = None, + block_mask_mouse: BlockMask | None = None, + block_mask_keyboard: BlockMask | None = None, + num_frame_per_block: int = 1, + use_rope_keyboard: bool = True, + kv_cache: dict | None = None, + kv_cache_mouse: dict | None = None, + kv_cache_keyboard: dict | None = None, + crossattn_cache: dict | None = None, + current_start: int = 0, + cache_start: int | None = None, + ) -> torch.Tensor: + if hidden_states.dim() == 4: + hidden_states = hidden_states.squeeze(1) + + assert temb.ndim == 4 + num_frames = temb.shape[1] + frame_seqlen = hidden_states.shape[1] // num_frames + bs, seq_length, _ = hidden_states.shape + orig_dtype = hidden_states.dtype + # hidden_states = hidden_states.float() + + e = self.scale_shift_table + temb + assert e.shape == (bs, num_frames, 6, self.hidden_dim) + shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(6, dim=2) + + norm_hidden_states = (self.norm1(hidden_states).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * + (1 + scale_msa) + shift_msa).flatten(1, 2) + query, _ = self.to_q(norm_hidden_states) + key, _ = self.to_k(norm_hidden_states) + value, _ = self.to_v(norm_hidden_states) + + if self.norm_q is not None: + query = self.norm_q.forward_native(query) + if self.norm_k is not None: + key = self.norm_k.forward_native(key) + + query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1)) + key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1)) + value = value.squeeze(1).unflatten(2, (self.num_attention_heads, -1)) + + attn_output = self.attn1(query, key, value, freqs_cis, block_mask, kv_cache, current_start, cache_start, grid_sizes) + attn_output = attn_output.flatten(2) + attn_output, _ = self.to_out(attn_output) + attn_output = attn_output.squeeze(1) + + null_shift = null_scale = torch.tensor([0], device=hidden_states.device) + norm_hidden_states, hidden_states = self.self_attn_residual_norm( + hidden_states, attn_output, gate_msa, null_shift, null_scale) + + # norm3 weights are loaded into self_attn_residual_norm.norm + attn_output = self.attn2(norm_hidden_states, + context=encoder_hidden_states, + context_lens=None, + crossattn_cache=crossattn_cache) + # residual connection + hidden_states = hidden_states + attn_output + + if self.action_model is not None: + if mouse_cond is not None or keyboard_cond is not None: + start_frame = current_start // frame_seqlen if grid_sizes is not None else 0 + + hidden_states = self.action_model( + hidden_states, + grid_sizes[0], grid_sizes[1], grid_sizes[2], + mouse_cond, keyboard_cond, + block_mask_mouse, block_mask_keyboard, + is_causal=True, + kv_cache_mouse=kv_cache_mouse, + kv_cache_keyboard=kv_cache_keyboard, + start_frame=start_frame, + use_rope_keyboard=use_rope_keyboard, + num_frame_per_block=num_frame_per_block + ) + + + ff_output = self.ffn( + (self.norm1(hidden_states).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * + + (1 + c_scale_msa) + c_shift_msa).flatten(1, 2) + ) + hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa) + + # Convert back to original dtype before return + return hidden_states.to(orig_dtype) + + +_DEFAULT_MATRIXGAME_CONFIG = MatrixGameWanVideoConfig() + + +class CausalMatrixGameWanModel(BaseDiT): + supports_action_input = True + + _fsdp_shard_conditions = _DEFAULT_MATRIXGAME_CONFIG._fsdp_shard_conditions + _compile_conditions = _DEFAULT_MATRIXGAME_CONFIG._compile_conditions + _supported_attention_backends = _DEFAULT_MATRIXGAME_CONFIG._supported_attention_backends + param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.param_names_mapping + reverse_param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.reverse_param_names_mapping + lora_param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.lora_param_names_mapping + + def __init__(self, + config: MatrixGameWanVideoConfig, + hf_config: dict[str, Any], + **kwargs) -> None: + super().__init__(config=config, hf_config=hf_config) + + inner_dim = config.num_attention_heads * config.attention_head_dim + self.hidden_size = config.hidden_size + self.num_attention_heads = config.num_attention_heads + self.attention_head_dim = config.attention_head_dim + self.in_channels = config.in_channels + self.out_channels = config.out_channels + self.num_channels_latents = config.num_channels_latents + self.patch_size = config.patch_size + + arch_cfg = getattr(config, "arch_config", None) + self.local_attn_size = getattr(arch_cfg, 'local_attn_size', getattr(config, 'local_attn_size', -1)) if arch_cfg else getattr(config, 'local_attn_size', -1) + self.sink_size = getattr(arch_cfg, 'sink_size', getattr(config, 'sink_size', 0)) if arch_cfg else getattr(config, 'sink_size', 0) + + # 1. Patch & position embedding + self.patch_embedding = PatchEmbed(in_chans=config.in_channels, + embed_dim=inner_dim, + patch_size=config.patch_size, + flatten=False) + + # 2. Condition embeddings + self.condition_embedder = CausalMatrixGameTimeImageEmbedding( + dim=inner_dim, + time_freq_dim=config.freq_dim, + image_embed_dim=config.image_dim, + ) + + # 2.1. Get action config + arch_cfg = getattr(config, "arch_config", None) + self.action_config = getattr(arch_cfg, "action_config", {}) if arch_cfg is not None else {} + + # 3. Transformer blocks + self.blocks = nn.ModuleList([ + CausalMatrixGameTransformerBlock( + inner_dim, + config.ffn_dim, + config.num_attention_heads, + self.local_attn_size, + self.sink_size, + config.qk_norm, + config.cross_attn_norm, + config.eps, + config.added_kv_proj_dim, + self._supported_attention_backends, + prefix=f"{getattr(config, 'prefix', 'Wan')}.blocks.{i}", + action_config=self.action_config, + block_idx=i) + for i in range(config.num_layers) + ]) + + # 4. Output norm & projection + self.norm_out = LayerNormScaleShift(inner_dim, + norm_type="layer", + eps=config.eps, + elementwise_affine=False, + dtype=torch.float32) + self.proj_out = nn.Linear( + inner_dim, config.out_channels * math.prod(config.patch_size)) + self.scale_shift_table = nn.Parameter( + torch.randn(1, 2, inner_dim) / inner_dim**0.5) + + self.gradient_checkpointing = False + + self.block_mask = None + self.block_mask_keyboard = None + self.block_mask_mouse = None + self.use_rope_keyboard = True + + self.num_frame_per_block = getattr(arch_cfg, 'num_frames_per_block', getattr(config, 'num_frame_per_block', 1)) if arch_cfg else getattr(config, 'num_frame_per_block', 1) + + self.__post_init__() + + @staticmethod + def _prepare_blockwise_causal_attn_mask( + device: torch.device | str, num_frames: int = 9, + frame_seqlen: int = 880, num_frame_per_block: int = 1, local_attn_size: int = -1 + ) -> BlockMask: + total_length = num_frames * frame_seqlen + padded_length = math.ceil(total_length / 128) * 128 - total_length + + ends = torch.zeros(total_length + padded_length, + device=device, dtype=torch.long) + + frame_indices = torch.arange( + start=0, + end=total_length, + step=frame_seqlen * num_frame_per_block, + device=device + ) + + for tmp in frame_indices: + ends[tmp:tmp + frame_seqlen * num_frame_per_block] = tmp + \ + frame_seqlen * num_frame_per_block + + def attention_mask(b, h, q_idx, kv_idx): + if local_attn_size == -1: + return (kv_idx < ends[q_idx]) | (q_idx == kv_idx) + else: + return ((kv_idx < ends[q_idx]) & (kv_idx >= (ends[q_idx] - local_attn_size * frame_seqlen))) | (q_idx == kv_idx) + + block_mask = create_block_mask(attention_mask, B=None, H=None, Q_LEN=total_length + padded_length, + KV_LEN=total_length + padded_length, _compile=False, device=device) + + if not dist.is_initialized() or dist.get_rank() == 0: + print(f" cache a block wise causal mask with block size of {num_frame_per_block} frames") + + return block_mask + + @staticmethod + def _prepare_blockwise_causal_attn_mask_keyboard( + device: torch.device | str, num_frames: int = 9, + frame_seqlen: int = 880, num_frame_per_block: int = 1, local_attn_size: int = -1 + ) -> BlockMask: + total_length2 = num_frames * frame_seqlen + padded_length2 = math.ceil(total_length2 / 32) * 32 - total_length2 + padded_length_kv2 = math.ceil(num_frames / 32) * 32 - num_frames + ends2 = torch.zeros(total_length2 + padded_length2, + device=device, dtype=torch.long) + + frame_indices2 = torch.arange( + start=0, + end=total_length2, + step=frame_seqlen * num_frame_per_block, + device=device + ) + cnt = num_frame_per_block + for tmp in frame_indices2: + ends2[tmp:tmp + frame_seqlen * num_frame_per_block] = cnt + cnt += num_frame_per_block + + def attention_mask2(b, h, q_idx, kv_idx): + if local_attn_size == -1: + return (kv_idx < ends2[q_idx]) | (q_idx == kv_idx) + else: + return ((kv_idx < ends2[q_idx]) & (kv_idx >= (ends2[q_idx] - local_attn_size))) | (q_idx == kv_idx) + + block_mask2 = create_block_mask(attention_mask2, B=None, H=None, Q_LEN=total_length2 + padded_length2, + KV_LEN=num_frames + padded_length_kv2, _compile=False, device=device) + + if not dist.is_initialized() or dist.get_rank() == 0: + print(f" cache a block wise causal mask for keyboard with block size of {num_frame_per_block} frames") + + return block_mask2 + + @staticmethod + def _prepare_blockwise_causal_attn_mask_action( + device: torch.device | str, num_frames: int = 9, + frame_seqlen: int = 1, num_frame_per_block: int = 1, local_attn_size: int = -1 + ) -> BlockMask: + total_length2 = num_frames * frame_seqlen + padded_length2 = math.ceil(total_length2 / 32) * 32 - total_length2 + padded_length_kv2 = math.ceil(num_frames / 32) * 32 - num_frames + ends2 = torch.zeros(total_length2 + padded_length2, + device=device, dtype=torch.long) + + frame_indices2 = torch.arange( + start=0, + end=total_length2, + step=frame_seqlen * num_frame_per_block, + device=device + ) + cnt = num_frame_per_block + for tmp in frame_indices2: + ends2[tmp:tmp + frame_seqlen * num_frame_per_block] = cnt + cnt += num_frame_per_block + + def attention_mask2(b, h, q_idx, kv_idx): + if local_attn_size == -1: + return (kv_idx < ends2[q_idx]) | (q_idx == kv_idx) + else: + return ((kv_idx < ends2[q_idx]) & (kv_idx >= (ends2[q_idx] - local_attn_size))) | (q_idx == kv_idx) + + block_mask2 = create_block_mask(attention_mask2, B=None, H=None, Q_LEN=total_length2 + padded_length2, + KV_LEN=num_frames + padded_length_kv2, _compile=False, device=device) + + if not dist.is_initialized() or dist.get_rank() == 0: + print(f" cache a block wise causal mask for action with block size of {num_frame_per_block} frames") + + return block_mask2 + + def _forward_inference( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor | list[torch.Tensor], + timestep: torch.LongTensor, + encoder_hidden_states_image: torch.Tensor | list[torch.Tensor] | None = None, + mouse_cond: torch.Tensor | None = None, + keyboard_cond: torch.Tensor | None = None, + kv_cache: dict = None, + kv_cache_mouse: dict = None, + kv_cache_keyboard: dict = None, + crossattn_cache: dict = None, + current_start: int = 0, + cache_start: int = 0, + start_frame: int = 0, + **kwargs + ) -> torch.Tensor: + # Extract num_frame_per_block from kwargs + effective_num_frame_per_block = kwargs.pop('num_frame_per_block', self.num_frame_per_block) + + if mouse_cond is not None or keyboard_cond is not None: + assert self.action_config is not None and len(self.action_config) > 0 + + if encoder_hidden_states is not None and not isinstance(encoder_hidden_states, torch.Tensor): + encoder_hidden_states = encoder_hidden_states[0] + if isinstance(encoder_hidden_states_image, list) and len(encoder_hidden_states_image) > 0: + encoder_hidden_states_image = encoder_hidden_states_image[0] + else: + encoder_hidden_states_image = None + + ctx = get_forward_context() + batch = getattr(ctx, "forward_batch", None) + + if batch is not None and getattr(batch, "image_latent", None) is not None: + image_latent = batch.image_latent + if isinstance(image_latent, torch.Tensor) and image_latent.ndim == 5 and hidden_states.ndim == 5: + _, _, num_frames, _, _ = hidden_states.shape + start = start_frame + end = start + num_frames + cond = image_latent + if cond.shape[2] >= end: + cond = cond[:, :, start:end] + elif cond.shape[2] > start: + cond = cond[:, :, start:] + pad_frames = num_frames - cond.shape[2] + if pad_frames > 0: + pad = torch.zeros( + cond.shape[0], + cond.shape[1], + pad_frames, + cond.shape[3], + cond.shape[4], + device=cond.device, + dtype=cond.dtype, + ) + cond = torch.cat([cond, pad], dim=2) + else: + pad = torch.zeros( + cond.shape[0], + cond.shape[1], + num_frames, + cond.shape[3], + cond.shape[4], + device=cond.device, + dtype=cond.dtype, + ) + cond = pad + hidden_states = torch.cat( + [hidden_states, cond.to(dtype=hidden_states.dtype)], dim=1 + ) + + batch_size, num_channels, num_frames, height, width = hidden_states.shape + p_t, p_h, p_w = self.patch_size + post_patch_num_frames = num_frames // p_t + post_patch_height = height // p_h + post_patch_width = width // p_w + + d = self.hidden_size // self.num_attention_heads + rope_dim_list = [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)] + freqs_cos, freqs_sin = get_rotary_pos_embed( + (post_patch_num_frames * get_sp_world_size(), post_patch_height, post_patch_width), + self.hidden_size, + self.num_attention_heads, + rope_dim_list, + dtype=torch.float32 if current_platform.is_mps() else torch.float64, + rope_theta=10000, + start_frame=start_frame + ) + freqs_cos = freqs_cos.to(hidden_states.device) + freqs_sin = freqs_sin.to(hidden_states.device) + freqs_cis = (freqs_cos, freqs_sin) if freqs_cos is not None else None + + hidden_states = self.patch_embedding(hidden_states) + grid_sizes = torch.tensor([post_patch_num_frames, post_patch_height, post_patch_width], + device=hidden_states.device) + hidden_states = hidden_states.flatten(2).transpose(1, 2) + + if timestep.dim() == 2: + timestep = timestep.flatten() + + temb, timestep_proj, encoder_hidden_states_text, encoder_hidden_states_image = self.condition_embedder( + timestep, encoder_hidden_states, encoder_hidden_states_image) + timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=(batch_size, post_patch_num_frames)) + + if encoder_hidden_states_text is not None: + if isinstance(encoder_hidden_states_text, list): + encoder_hidden_states_text = encoder_hidden_states_text[0] + elif encoder_hidden_states_text.ndim == 2: + encoder_hidden_states_text = encoder_hidden_states_text.unsqueeze(0) + + if encoder_hidden_states_image is not None: + if encoder_hidden_states_text is not None: + encoder_hidden_states = torch.concat( + [encoder_hidden_states_image, encoder_hidden_states_text], dim=1) + else: + encoder_hidden_states = encoder_hidden_states_image + + if self.block_mask is None: + self.block_mask = self._prepare_blockwise_causal_attn_mask( + device=hidden_states.device, + num_frames=num_frames, + frame_seqlen=post_patch_height * post_patch_width, + num_frame_per_block=self.num_frame_per_block, + local_attn_size=self.local_attn_size + ) + if self.block_mask_keyboard is None: + if not self.use_rope_keyboard: + self.block_mask_keyboard = self._prepare_blockwise_causal_attn_mask_keyboard( + device=hidden_states.device, + num_frames=num_frames, + frame_seqlen=post_patch_height * post_patch_width, + num_frame_per_block=self.num_frame_per_block, + local_attn_size=self.local_attn_size + ) + else: + self.block_mask_keyboard = self._prepare_blockwise_causal_attn_mask_action( + device=hidden_states.device, + num_frames=num_frames, + frame_seqlen=1, + num_frame_per_block=self.num_frame_per_block, + local_attn_size=self.local_attn_size + ) + if self.block_mask_mouse is None: + self.block_mask_mouse = self._prepare_blockwise_causal_attn_mask_action( + device=hidden_states.device, + num_frames=num_frames, + frame_seqlen=1, + num_frame_per_block=self.num_frame_per_block, + local_attn_size=self.local_attn_size + ) + if kv_cache is None: + kv_cache = [None] * len(self.blocks) + if kv_cache_mouse is None: + kv_cache_mouse = [None] * len(self.blocks) + if kv_cache_keyboard is None: + kv_cache_keyboard = [None] * len(self.blocks) + if crossattn_cache is None: + crossattn_cache = [None] * len(self.blocks) + + for block_index, block in enumerate(self.blocks): + if torch.is_grad_enabled() and self.gradient_checkpointing: + kwargs.update({ + "kv_cache": kv_cache[block_index], + "kv_cache_mouse": kv_cache_mouse[block_index] if kv_cache_mouse else None, + "kv_cache_keyboard": kv_cache_keyboard[block_index] if kv_cache_keyboard else None, + "crossattn_cache": crossattn_cache[block_index] if crossattn_cache else None, + "current_start": current_start, + "cache_start": cache_start, + "num_frame_per_block": effective_num_frame_per_block, + }) + hidden_states = self._gradient_checkpointing_func( + block, hidden_states, **kwargs) + else: + kwargs.update({ + "kv_cache": kv_cache[block_index], + "kv_cache_mouse": kv_cache_mouse[block_index] if kv_cache_mouse else None, + "kv_cache_keyboard": kv_cache_keyboard[block_index] if kv_cache_keyboard else None, + "crossattn_cache": crossattn_cache[block_index] if crossattn_cache else None, + "current_start": current_start, + "cache_start": cache_start, + }) + hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, freqs_cis, + block_mask=self.block_mask, + grid_sizes=grid_sizes, + mouse_cond=mouse_cond, + keyboard_cond=keyboard_cond, + block_mask_mouse=self.block_mask_mouse, + block_mask_keyboard=self.block_mask_keyboard, + num_frame_per_block=effective_num_frame_per_block, + use_rope_keyboard=self.use_rope_keyboard, + **kwargs) + + temb = temb.unflatten(dim=0, sizes=(batch_size, post_patch_num_frames)).unsqueeze(2) + shift, scale = (self.scale_shift_table.unsqueeze(1) + temb).chunk(2, dim=2) + hidden_states = self.norm_out(hidden_states, shift, scale) + hidden_states = self.proj_out(hidden_states) + + hidden_states = self.unpatchify(hidden_states, grid_sizes) + + return hidden_states + + + def _forward_train( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor | list[torch.Tensor], + timestep: torch.LongTensor, + encoder_hidden_states_image: torch.Tensor | list[torch.Tensor] | None = None, + mouse_cond: torch.Tensor | None = None, + keyboard_cond: torch.Tensor | None = None, + **kwargs + ) -> torch.Tensor: + if mouse_cond is not None or keyboard_cond is not None: + assert self.action_config is not None and len(self.action_config) > 0 + + if encoder_hidden_states is not None and not isinstance(encoder_hidden_states, torch.Tensor): + encoder_hidden_states = encoder_hidden_states[0] + if isinstance(encoder_hidden_states_image, list) and len(encoder_hidden_states_image) > 0: + encoder_hidden_states_image = encoder_hidden_states_image[0] + else: + encoder_hidden_states_image = None + + ctx = get_forward_context() + batch = getattr(ctx, "forward_batch", None) + if batch is not None and getattr(batch, "image_latent", None) is not None: + image_latent = batch.image_latent + if isinstance(image_latent, torch.Tensor) and image_latent.ndim == 5 and hidden_states.ndim == 5: + _, _, num_frames, _, _ = hidden_states.shape + cond = image_latent + if cond.shape[2] >= num_frames: + cond = cond[:, :, :num_frames] + else: + pad_frames = num_frames - cond.shape[2] + pad = torch.zeros( + cond.shape[0], + cond.shape[1], + pad_frames, + cond.shape[3], + cond.shape[4], + device=cond.device, + dtype=cond.dtype, + ) + cond = torch.cat([cond, pad], dim=2) + hidden_states = torch.cat( + [hidden_states, cond.to(dtype=hidden_states.dtype)], dim=1 + ) + + batch_size, num_channels, num_frames, height, width = hidden_states.shape + p_t, p_h, p_w = self.patch_size + post_patch_num_frames = num_frames // p_t + post_patch_height = height // p_h + post_patch_width = width // p_w + + d = self.hidden_size // self.num_attention_heads + rope_dim_list = [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)] + freqs_cos, freqs_sin = get_rotary_pos_embed( + (post_patch_num_frames * get_sp_world_size(), post_patch_height, post_patch_width), + self.hidden_size, + self.num_attention_heads, + rope_dim_list, + dtype=torch.float32 if current_platform.is_mps() else torch.float64, + rope_theta=10000 + ) + freqs_cos = freqs_cos.to(hidden_states.device) + freqs_sin = freqs_sin.to(hidden_states.device) + freqs_cis = (freqs_cos, freqs_sin) if freqs_cos is not None else None + + if self.block_mask is None: + self.block_mask = self._prepare_blockwise_causal_attn_mask( + device=hidden_states.device, + num_frames=num_frames, + frame_seqlen=post_patch_height * post_patch_width, + num_frame_per_block=self.num_frame_per_block, + local_attn_size=self.local_attn_size + ) + if self.block_mask_keyboard is None: + if not self.use_rope_keyboard: + self.block_mask_keyboard = self._prepare_blockwise_causal_attn_mask_keyboard( + device=hidden_states.device, + num_frames=num_frames, + frame_seqlen=post_patch_height * post_patch_width, + num_frame_per_block=self.num_frame_per_block, + local_attn_size=self.local_attn_size + ) + else: + self.block_mask_keyboard = self._prepare_blockwise_causal_attn_mask_action( + device=hidden_states.device, + num_frames=num_frames, + frame_seqlen=1, + num_frame_per_block=self.num_frame_per_block, + local_attn_size=self.local_attn_size + ) + if self.block_mask_mouse is None: + self.block_mask_mouse = self._prepare_blockwise_causal_attn_mask_action( + device=hidden_states.device, + num_frames=num_frames, + frame_seqlen=1, + num_frame_per_block=self.num_frame_per_block, + local_attn_size=self.local_attn_size + ) + + hidden_states = self.patch_embedding(hidden_states) + grid_sizes = torch.tensor([post_patch_num_frames, post_patch_height, post_patch_width], + device=hidden_states.device) + hidden_states = hidden_states.flatten(2).transpose(1, 2) + + if timestep.dim() == 2: + timestep = timestep.flatten() + + temb, timestep_proj, encoder_hidden_states_text, encoder_hidden_states_image = self.condition_embedder( + timestep, encoder_hidden_states, encoder_hidden_states_image) + timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=(batch_size, post_patch_num_frames)) + + if encoder_hidden_states_text is not None: + if isinstance(encoder_hidden_states_text, list): + encoder_hidden_states_text = encoder_hidden_states_text[0] + elif encoder_hidden_states_text.ndim == 2: + encoder_hidden_states_text = encoder_hidden_states_text.unsqueeze(0) + + if encoder_hidden_states_image is not None: + if encoder_hidden_states_text is not None: + encoder_hidden_states = torch.concat( + [encoder_hidden_states_image, encoder_hidden_states_text], dim=1) + else: + encoder_hidden_states = encoder_hidden_states_image + else: + encoder_hidden_states = encoder_hidden_states_text + + kwargs = dict( + e=timestep_proj, + grid_sizes=grid_sizes, + freqs_cis=freqs_cis, + encoder_hidden_states=encoder_hidden_states, + mouse_cond=mouse_cond, + keyboard_cond=keyboard_cond, + block_mask=self.block_mask, + block_mask_mouse=self.block_mask_mouse, + block_mask_keyboard=self.block_mask_keyboard, + use_rope_keyboard=self.use_rope_keyboard, + num_frame_per_block=self.num_frame_per_block + ) + + # Initialize kv_cache, kv_cache_mouse, kv_cache_keyboard, crossattn_cache if not provided + kv_cache = kwargs.pop("kv_cache", [None] * len(self.blocks)) + kv_cache_mouse = kwargs.pop("kv_cache_mouse", [None] * len(self.blocks)) + kv_cache_keyboard = kwargs.pop("kv_cache_keyboard", [None] * len(self.blocks)) + crossattn_cache = kwargs.pop("crossattn_cache", [None] * len(self.blocks)) + + for i, block in enumerate(self.blocks): + block_crossattn_cache = None + if crossattn_cache is not None: + if crossattn_cache[i] is None: + crossattn_cache[i] = {"is_init": False} + block_crossattn_cache = crossattn_cache[i] + + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = self._gradient_checkpointing_func( + block, hidden_states, encoder_hidden_states, timestep_proj, freqs_cis, + kv_cache=kv_cache[i], + kv_cache_mouse=kv_cache_mouse[i], + kv_cache_keyboard=kv_cache_keyboard[i], + crossattn_cache=block_crossattn_cache, + block_mask=self.block_mask, + grid_sizes=grid_sizes, + mouse_cond=mouse_cond, + keyboard_cond=keyboard_cond, + block_mask_mouse=self.block_mask_mouse, + block_mask_keyboard=self.block_mask_keyboard, + num_frame_per_block=self.num_frame_per_block, + use_rope_keyboard=self.use_rope_keyboard) + else: + hidden_states = block( + hidden_states, + encoder_hidden_states, + timestep_proj, + freqs_cis, + kv_cache=kv_cache[i], + kv_cache_mouse=kv_cache_mouse[i], + kv_cache_keyboard=kv_cache_keyboard[i], + crossattn_cache=block_crossattn_cache, + block_mask=self.block_mask, + grid_sizes=grid_sizes, + mouse_cond=mouse_cond, + keyboard_cond=keyboard_cond, + block_mask_mouse=self.block_mask_mouse, + block_mask_keyboard=self.block_mask_keyboard, + num_frame_per_block=self.num_frame_per_block, + use_rope_keyboard=self.use_rope_keyboard) + + temb = temb.unflatten(dim=0, sizes=(batch_size, post_patch_num_frames)).unsqueeze(2) + shift, scale = (self.scale_shift_table.unsqueeze(1) + temb).chunk(2, dim=2) + hidden_states = self.norm_out(hidden_states, shift, scale) + hidden_states = self.proj_out(hidden_states) + + hidden_states = self.unpatchify(hidden_states, grid_sizes) + + return hidden_states + + def forward(self, *args, **kwargs): + if kwargs.get('kv_cache') is not None: + return self._forward_inference(*args, **kwargs) + else: + return self._forward_train(*args, **kwargs) + + def unpatchify(self, x, grid_sizes): + c = self.proj_out.out_features // math.prod(self.patch_size) + p_t, p_h, p_w = self.patch_size + f, h, w = grid_sizes.tolist() + + x = x[:, :f * h * w].view(-1, f, h, w, p_t, p_h, p_w, c) + x = x.permute(0, 7, 1, 4, 2, 5, 3, 6) + x = x.reshape(-1, c, f * p_t, h * p_h, w * p_w) + return x + + def init_weights(self): + for m in self.modules(): + if isinstance(m, nn.Linear): + nn.init.xavier_uniform_(m.weight) + if m.bias is not None: + nn.init.zeros_(m.bias) + + nn.init.xavier_uniform_(self.patch_embedding.proj.weight.flatten(1)) + for m in self.condition_embedder.time_embedder.mlp.modules(): + if isinstance(m, nn.Linear): + nn.init.normal_(m.weight, std=.02) + + nn.init.zeros_(self.proj_out.weight) + + for block in self.blocks: + if block.action_model is not None: + try: + nn.init.zeros_(block.action_model.proj_mouse.weight) + if block.action_model.proj_mouse.bias is not None: + nn.init.zeros_(block.action_model.proj_mouse.bias) + nn.init.zeros_(block.action_model.proj_keyboard.weight) + if block.action_model.proj_keyboard.bias is not None: + nn.init.zeros_(block.action_model.proj_keyboard.bias) + except AttributeError: + pass diff --git a/fastvideo/models/dits/matrix_game/model.py b/fastvideo/models/dits/matrix_game/model.py new file mode 100644 index 000000000..30891164d --- /dev/null +++ b/fastvideo/models/dits/matrix_game/model.py @@ -0,0 +1,469 @@ +# SPDX-License-Identifier: Apache-2.0 + +import math +from typing import Any + +import torch +import torch.nn as nn + +from fastvideo.attention import DistributedAttention +from fastvideo.configs.models.dits.matrixgame import MatrixGameWanVideoConfig +from fastvideo.distributed.parallel_state import get_sp_world_size +from fastvideo.layers.layernorm import (FP32LayerNorm, LayerNormScaleShift, + RMSNorm, ScaleResidual, + ScaleResidualLayerNormScaleShift) +from fastvideo.layers.linear import ReplicatedLinear +from fastvideo.layers.mlp import MLP +from fastvideo.layers.rotary_embedding import (_apply_rotary_emb, + get_rotary_pos_embed) +from fastvideo.layers.visual_embedding import PatchEmbed, TimestepEmbedder, ModulateProjection +from fastvideo.logger import init_logger +from fastvideo.models.dits.base import BaseDiT +from fastvideo.models.dits.wanvideo import (WanSelfAttention, + WanI2VCrossAttention, + WanT2VCrossAttention, + WanImageEmbedding) +from fastvideo.platforms import AttentionBackendEnum, current_platform + +# Import ActionModule +from .action_module import ActionModule + +logger = init_logger(__name__) + + +class MatrixGameTimeImageEmbedding(nn.Module): + def __init__( + self, + dim: int, + time_freq_dim: int, + image_embed_dim: int | None = None, + ): + super().__init__() + + self.time_embedder = TimestepEmbedder( + dim, frequency_embedding_size=time_freq_dim, act_layer="silu") + self.time_modulation = ModulateProjection(dim, + factor=6, + act_layer="silu") + + self.image_embedder = None + if image_embed_dim is not None: + self.image_embedder = WanImageEmbedding(image_embed_dim, dim) + + def forward( + self, + timestep: torch.Tensor, + encoder_hidden_states: torch.Tensor, # Kept for interface compatibility + encoder_hidden_states_image: torch.Tensor | None = None, + timestep_seq_len: int | None = None, + ): + temb = self.time_embedder(timestep, timestep_seq_len) + timestep_proj = self.time_modulation(temb) + + # MatrixGame does not use text embeddings, so we ignore encoder_hidden_states + # and return None for the text embedding part + + if encoder_hidden_states_image is not None: + assert self.image_embedder is not None + encoder_hidden_states_image = self.image_embedder( + encoder_hidden_states_image) + + return temb, timestep_proj, None, encoder_hidden_states_image + + +class MatrixGameCrossAttention(WanSelfAttention): + def forward(self, x, context, context_lens=None, crossattn_cache=None): + r""" + Args: + x(Tensor): Shape [B, L1, C] + context(Tensor): Shape [B, L2, C] - typically 257 image tokens + context_lens(Tensor): Shape [B] + crossattn_cache(dict): Optional cache for k/v during inference + """ + b, n, d = x.size(0), self.num_heads, self.head_dim + + # compute query, key, value + q = self.norm_q(self.to_q(x)[0]).view(b, -1, n, d) + + if crossattn_cache is not None: + if not crossattn_cache["is_init"]: + crossattn_cache["is_init"] = True + k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d) + v = self.to_v(context)[0].view(b, -1, n, d) + crossattn_cache["k"] = k + crossattn_cache["v"] = v + else: + k = crossattn_cache["k"] + v = crossattn_cache["v"] + else: + k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d) + v = self.to_v(context)[0].view(b, -1, n, d) + + # compute attention + x = self.attn(q, k, v) + + # output + x = x.flatten(2) + x, _ = self.to_out(x) + return x + + +class MatrixGameTransformerBlock(nn.Module): + + def __init__(self, + dim: int, + ffn_dim: int, + num_heads: int, + qk_norm: str = "rms_norm_across_heads", + cross_attn_norm: bool = False, + eps: float = 1e-6, + added_kv_proj_dim: int | None = None, + supported_attention_backends: tuple[AttentionBackendEnum, ...] + | None = None, + prefix: str = "", + action_config: dict | None = None): + super().__init__() + action_config = action_config or {} + + # 1. Self-attention + self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False) + self.to_q = ReplicatedLinear(dim, dim, bias=True) + self.to_k = ReplicatedLinear(dim, dim, bias=True) + self.to_v = ReplicatedLinear(dim, dim, bias=True) + + self.to_out = ReplicatedLinear(dim, dim, bias=True) + self.attn1 = DistributedAttention( + num_heads=num_heads, + head_size=dim // num_heads, + causal=False, + supported_attention_backends=supported_attention_backends, + prefix=f"{prefix}.attn1") + self.hidden_dim = dim + self.num_attention_heads = num_heads + dim_head = dim // num_heads + if qk_norm == "rms_norm": + self.norm_q = RMSNorm(dim_head, eps=eps) + self.norm_k = RMSNorm(dim_head, eps=eps) + elif qk_norm == "rms_norm_across_heads": + # LTX applies qk norm across all heads + self.norm_q = RMSNorm(dim, eps=eps) + self.norm_k = RMSNorm(dim, eps=eps) + else: + print("QK Norm type not supported") + raise Exception + assert cross_attn_norm is True + self.self_attn_residual_norm = ScaleResidualLayerNormScaleShift( + dim, + norm_type="layer", + eps=eps, + elementwise_affine=True, + dtype=torch.float32, + compute_dtype=torch.float32) + + # 2. Cross-attention + if added_kv_proj_dim is not None: + # I2V + self.attn2 = WanI2VCrossAttention(dim, + num_heads, + qk_norm=qk_norm, + eps=eps) + else: + # T2V + self.attn2 = WanT2VCrossAttention(dim, + num_heads, + qk_norm=qk_norm, + eps=eps) + self.cross_attn_residual_norm = ScaleResidualLayerNormScaleShift( + dim, + norm_type="layer", + eps=eps, + elementwise_affine=False, + dtype=torch.float32, + compute_dtype=torch.float32) + + # 2.1. Action Module Integration + self.use_action_module = len(action_config) > 0 + if self.use_action_module: + self.action_model = ActionModule(**action_config) + else: + self.action_model = None + + # 3. Feed-forward + self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh") + self.mlp_residual = ScaleResidual() + + self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + temb: torch.Tensor, + freqs_cis: tuple[torch.Tensor, torch.Tensor], + # Action Module specific args + grid_sizes: torch.Tensor | None = None, + mouse_cond: torch.Tensor | None = None, + keyboard_cond: torch.Tensor | None = None, + ) -> torch.Tensor: + if hidden_states.dim() == 4: + hidden_states = hidden_states.squeeze(1) + orig_dtype = hidden_states.dtype + + if temb.dim() == 4: + # temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v) + shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = ( + self.scale_shift_table.unsqueeze(0) + temb.float() + ).chunk(6, dim=2) + # batch_size, seq_len, 1, inner_dim + shift_msa = shift_msa.squeeze(2) + scale_msa = scale_msa.squeeze(2) + gate_msa = gate_msa.squeeze(2) + c_shift_msa = c_shift_msa.squeeze(2) + c_scale_msa = c_scale_msa.squeeze(2) + c_gate_msa = c_gate_msa.squeeze(2) + else: + # temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B) + e = self.scale_shift_table + temb.float() + shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk( + 6, dim=1) + assert shift_msa.dtype == torch.float32 + + # 1. Self-attention + norm_hidden_states = (self.norm1(hidden_states.float()) * + (1 + scale_msa) + shift_msa).to(orig_dtype) + query, _ = self.to_q(norm_hidden_states) + key, _ = self.to_k(norm_hidden_states) + value, _ = self.to_v(norm_hidden_states) + + if self.norm_q is not None: + query = self.norm_q(query) + if self.norm_k is not None: + key = self.norm_k(key) + + query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1)) + key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1)) + value = value.squeeze(1).unflatten(2, (self.num_attention_heads, -1)) + + # Apply rotary embeddings + cos, sin = freqs_cis + query, key = _apply_rotary_emb(query, cos, sin, + is_neox_style=False), _apply_rotary_emb( + key, cos, sin, is_neox_style=False) + + attn_output, _ = self.attn1(query, key, value) + attn_output = attn_output.flatten(2) + attn_output, _ = self.to_out(attn_output) + attn_output = attn_output.squeeze(1) + + null_shift = null_scale = torch.tensor([0], device=hidden_states.device) + norm_hidden_states, hidden_states = self.self_attn_residual_norm( + hidden_states, attn_output, gate_msa, null_shift, null_scale) + norm_hidden_states, hidden_states = norm_hidden_states.to( + orig_dtype), hidden_states.to(orig_dtype) + + # 2. Cross-attention + attn_output = self.attn2(norm_hidden_states, + context=encoder_hidden_states, + context_lens=None) + norm_hidden_states, hidden_states = self.cross_attn_residual_norm( + hidden_states, attn_output, 1, c_shift_msa, c_scale_msa) + norm_hidden_states, hidden_states = norm_hidden_states.to( + orig_dtype), hidden_states.to(orig_dtype) + + # ================= Action Module ================= + if self.action_model is not None: + if mouse_cond is not None or keyboard_cond is not None: + # grid_sizes is expected to be [F, H, W] + # ActionModule implementation takes hidden_states directly + hidden_states = self.action_model( + hidden_states, + grid_sizes[0], grid_sizes[1], grid_sizes[2], + mouse_cond, keyboard_cond, + num_frame_per_block=grid_sizes[0], + ) + # ================================================= + + # 3. Feed-forward + ff_output = self.ffn(norm_hidden_states) + hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa) + hidden_states = hidden_states.to(orig_dtype) + + return hidden_states + +_DEFAULT_MATRIXGAME_CONFIG = MatrixGameWanVideoConfig() + + +class MatrixGameWanModel(BaseDiT): + # Marker for action input support (Matrix-Game) + supports_action_input = True + + _fsdp_shard_conditions = _DEFAULT_MATRIXGAME_CONFIG._fsdp_shard_conditions + _compile_conditions = _DEFAULT_MATRIXGAME_CONFIG._compile_conditions + _supported_attention_backends = _DEFAULT_MATRIXGAME_CONFIG._supported_attention_backends + param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.param_names_mapping + reverse_param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.reverse_param_names_mapping + lora_param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.lora_param_names_mapping + + def __init__(self, + config: MatrixGameWanVideoConfig, + hf_config: dict[str, Any], + **kwargs) -> None: + super().__init__(config=config, hf_config=hf_config) + + inner_dim = config.num_attention_heads * config.attention_head_dim + self.hidden_size = config.hidden_size + self.num_attention_heads = config.num_attention_heads + self.patch_size = config.patch_size + + # 1. Patch & position embedding + self.patch_embedding = PatchEmbed(in_chans=config.in_channels, + embed_dim=inner_dim, + patch_size=config.patch_size, + flatten=False) + + # 2. Condition embeddings + self.condition_embedder = MatrixGameTimeImageEmbedding( + dim=inner_dim, + time_freq_dim=config.freq_dim, + image_embed_dim=config.image_dim, + ) + + # 2.1. Get action config + self.action_config = getattr(config, 'action_config', {}) + + # 3. Transformer blocks + self.blocks = nn.ModuleList([ + MatrixGameTransformerBlock( + inner_dim, + config.ffn_dim, + config.num_attention_heads, + config.qk_norm, + config.cross_attn_norm, + config.eps, + config.added_kv_proj_dim, + self._supported_attention_backends, + prefix=f"{getattr(config, 'prefix', 'Wan')}.blocks.{i}", + action_config=self.action_config) + for i in range(config.num_layers) + ]) + + # 4. Output norm & projection + self.norm_out = LayerNormScaleShift(inner_dim, + norm_type="layer", + eps=config.eps, + elementwise_affine=False, + dtype=torch.float32, + compute_dtype=torch.float32) + self.proj_out = nn.Linear( + inner_dim, config.out_channels * math.prod(config.patch_size)) + self.scale_shift_table = nn.Parameter( + torch.randn(1, 2, inner_dim) / inner_dim**0.5) + + self.gradient_checkpointing = False + + def forward(self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor | list[torch.Tensor], + timestep: torch.LongTensor, + encoder_hidden_states_image: torch.Tensor + | list[torch.Tensor] | None = None, + # Action inputs + mouse_cond: torch.Tensor | None = None, + keyboard_cond: torch.Tensor | None = None, + **kwargs) -> torch.Tensor: + if encoder_hidden_states is not None and not isinstance(encoder_hidden_states, torch.Tensor): + encoder_hidden_states = encoder_hidden_states[0] + if isinstance(encoder_hidden_states_image, + list) and len(encoder_hidden_states_image) > 0: + encoder_hidden_states_image = encoder_hidden_states_image[0] + else: + encoder_hidden_states_image = None + + batch_size, num_channels, num_frames, height, width = hidden_states.shape + p_t, p_h, p_w = self.patch_size + post_patch_num_frames = num_frames // p_t + post_patch_height = height // p_h + post_patch_width = width // p_w + + # Get rotary embeddings + d = self.hidden_size // self.num_attention_heads + rope_dim_list = [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)] + freqs_cos, freqs_sin = get_rotary_pos_embed( + (post_patch_num_frames * get_sp_world_size(), post_patch_height, + post_patch_width), + self.hidden_size, + self.num_attention_heads, + rope_dim_list, + dtype=torch.float32 if current_platform.is_mps() else torch.float64, + rope_theta=10000) + freqs_cos = freqs_cos.to(hidden_states.device) + freqs_sin = freqs_sin.to(hidden_states.device) + freqs_cis = (freqs_cos.float(), + freqs_sin.float()) if freqs_cos is not None else None + + hidden_states = self.patch_embedding(hidden_states) + hidden_states = hidden_states.flatten(2).transpose(1, 2) + + if timestep.dim() == 2: + timestep = timestep.flatten() + + temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder( + timestep, encoder_hidden_states, encoder_hidden_states_image) + timestep_proj = timestep_proj.unflatten(1, (6, -1)) + + if encoder_hidden_states is not None: + if isinstance(encoder_hidden_states, list): + encoder_hidden_states = encoder_hidden_states[0] + elif encoder_hidden_states.ndim == 2: + encoder_hidden_states = encoder_hidden_states.unsqueeze(0) + else: + # encoder_hidden_states is None (e.g. no text encoder) + # MatrixGame uses image-action cross-attn. + pass + + if encoder_hidden_states_image is not None: + if encoder_hidden_states is not None: + encoder_hidden_states = torch.concat( + [encoder_hidden_states_image, encoder_hidden_states], dim=1) + else: + encoder_hidden_states = encoder_hidden_states_image + + # This is [F, H, W] for the ActionModule + grid_sizes = torch.tensor([ + post_patch_num_frames, post_patch_height, post_patch_width + ], + device=hidden_states.device) + + # Blocks + for block in self.blocks: + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = self._gradient_checkpointing_func( + block, hidden_states, encoder_hidden_states, timestep_proj, + freqs_cis, + grid_sizes=grid_sizes, + mouse_cond=mouse_cond, + keyboard_cond=keyboard_cond) + else: + hidden_states = block( + hidden_states, + encoder_hidden_states, + timestep_proj, + freqs_cis, + grid_sizes=grid_sizes, + mouse_cond=mouse_cond, + keyboard_cond=keyboard_cond) + + # Output + shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, + dim=1) + hidden_states = self.norm_out(hidden_states, shift, scale) + hidden_states = self.proj_out(hidden_states) + + hidden_states = hidden_states.reshape(batch_size, post_patch_num_frames, + post_patch_height, + post_patch_width, p_t, p_h, p_w, + -1) + hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6) + output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3) + + return output \ No newline at end of file diff --git a/fastvideo/models/dits/matrix_game/utils.py b/fastvideo/models/dits/matrix_game/utils.py new file mode 100644 index 000000000..7f365eed7 --- /dev/null +++ b/fastvideo/models/dits/matrix_game/utils.py @@ -0,0 +1,414 @@ +from __future__ import annotations + +import asyncio +import os +import random + +import torch +from PIL import Image + +from fastvideo.distributed.parallel_state import get_local_torch_device +from fastvideo.utils import logger + + +CAM_VALUE = 0.1 +CAMERA_MAP = { + "i": [CAM_VALUE, 0], "k": [-CAM_VALUE, 0], + "j": [0, -CAM_VALUE], "l": [0, CAM_VALUE], "u": [0, 0] +} + +KEYBOARD_MAP_4 = { # base_distilled_model (universal): W/S/A/D + "w": [1, 0, 0, 0], "s": [0, 1, 0, 0], + "a": [0, 0, 1, 0], "d": [0, 0, 0, 1], "q": [0, 0, 0, 0] +} +KEYBOARD_MAP_2 = { # gta_distilled_model: W/S only (steering via mouse) + "w": [1, 0], "s": [0, 1], "q": [0, 0] +} +KEYBOARD_MAP_7 = { # templerun_distilled_model: still/w/s/left/right/a/d + "q": [1, 0, 0, 0, 0, 0, 0], # still + "w": [0, 1, 0, 0, 0, 0, 0], # forward + "s": [0, 0, 1, 0, 0, 0, 0], # back + "j": [0, 0, 0, 1, 0, 0, 0], # left (swipe) + "l": [0, 0, 0, 0, 1, 0, 0], # right (swipe) + "a": [0, 0, 0, 0, 0, 1, 0], # a + "d": [0, 0, 0, 0, 0, 0, 1], # d +} +KEYBOARD_MAP = KEYBOARD_MAP_4 # Default for backward compatibility + +def expand_action_to_frames(action: dict, num_frames: int) -> tuple[torch.Tensor, torch.Tensor]: + result = {} + for key, tensor in action.items(): + if tensor is not None: + # Expand to [num_frames, D] then unsqueeze to [1, num_frames, D] + result[key] = tensor.unsqueeze(0).repeat(num_frames, 1).unsqueeze(0) + else: + result[key] = None + + if "mouse" not in result or result["mouse"] is None: + # keyboard device if available, otherwise default + device = result.get("keyboard", torch.tensor([])).device if result.get("keyboard") is not None else get_local_torch_device() + result["mouse"] = torch.zeros(1, num_frames, 2, device=device) + + return result["keyboard"], result["mouse"] + +def get_current_action(mode="universal"): + + CAM_VALUE = 0.1 + if mode == 'universal': + logger.info("") + logger.info('-'*30) + logger.info("PRESS [I, K, J, L, U] FOR CAMERA TRANSFORM\n (I: up, K: down, J: left, L: right, U: no move)") + logger.info("PRESS [W, S, A, D, Q] FOR MOVEMENT\n (W: forward, S: back, A: left, D: right, Q: no move)") + logger.info('-'*30) + CAMERA_VALUE_MAP = { + "i": [CAM_VALUE, 0], + "k": [-CAM_VALUE, 0], + "j": [0, -CAM_VALUE], + "l": [0, CAM_VALUE], + "u": [0, 0] + } + KEYBOARD_IDX = { + "w": [1, 0, 0, 0], "s": [0, 1, 0, 0], "a": [0, 0, 1, 0], "d": [0, 0, 0, 1], + "q": [0, 0, 0, 0] + } + flag = 0 + while flag != 1: + try: + idx_mouse = input('Please input the mouse action (e.g. `U`):\n').strip().lower() + idx_keyboard = input('Please input the keyboard action (e.g. `W`):\n').strip().lower() + if idx_mouse in CAMERA_VALUE_MAP and idx_keyboard in KEYBOARD_IDX: + flag = 1 + except Exception: + pass + mouse_cond = torch.tensor(CAMERA_VALUE_MAP[idx_mouse]).cuda() + keyboard_cond = torch.tensor(KEYBOARD_IDX[idx_keyboard]).cuda() + elif mode == 'gta_drive': + logger.info("") + logger.info('-'*30) + logger.info("PRESS [W, S, A, D, Q] FOR MOVEMENT\n (W: forward, S: back, A: left, D: right, Q: no move)") + logger.info('-'*30) + CAMERA_VALUE_MAP = { + "a": [0, -CAM_VALUE], + "d": [0, CAM_VALUE], + "q": [0, 0] + } + KEYBOARD_IDX = { + "w": [1, 0], "s": [0, 1], + "q": [0, 0] + } + flag = 0 + while flag != 1: + try: + indexes = input('Please input the actions (split with ` `):\n(e.g. `W` for forward, `W A` for forward and left)\n').strip().lower().split(' ') + idx_mouse = [] + idx_keyboard = [] + for i in indexes: + if i in CAMERA_VALUE_MAP.keys(): + idx_mouse += [i] + elif i in KEYBOARD_IDX.keys(): + idx_keyboard += [i] + if len(idx_mouse) == 0: + idx_mouse += ['q'] + if len(idx_keyboard) == 0: + idx_keyboard += ['q'] + assert idx_mouse in [['a'], ['d'], ['q']] and idx_keyboard in [['q'], ['w'], ['s']] + flag = 1 + except Exception: + pass + mouse_cond = torch.tensor(CAMERA_VALUE_MAP[idx_mouse[0]]).cuda() + keyboard_cond = torch.tensor(KEYBOARD_IDX[idx_keyboard[0]]).cuda() + elif mode == 'templerun': + logger.info("") + logger.info('-'*30) + logger.info("PRESS [W, S, A, D, Z, C, Q] FOR ACTIONS\n (W: jump, S: slide, A: left side, D: right side, Z: turn left, C: turn right, Q: no move)") + logger.info('-'*30) + KEYBOARD_IDX = { + "w": [0, 1, 0, 0, 0, 0, 0], "s": [0, 0, 1, 0, 0, 0, 0], + "a": [0, 0, 0, 0, 0, 1, 0], "d": [0, 0, 0, 0, 0, 0, 1], + "z": [0, 0, 0, 1, 0, 0, 0], "c": [0, 0, 0, 0, 1, 0, 0], + "q": [1, 0, 0, 0, 0, 0, 0] + } + flag = 0 + while flag != 1: + try: + idx_keyboard = input('Please input the action: \n(e.g. `W` for forward, `Z` for turning left)\n').strip().lower() + if idx_keyboard in KEYBOARD_IDX.keys(): + flag = 1 + except Exception: + pass + keyboard_cond = torch.tensor(KEYBOARD_IDX[idx_keyboard]).cuda() + + if mode != 'templerun': + return { + "mouse": mouse_cond, + "keyboard": keyboard_cond + } + return { + "keyboard": keyboard_cond + } + +async def get_current_action_async(mode="universal"): + return await asyncio.to_thread(get_current_action, mode) + +def load_initial_image(image_path: str = None) -> Image.Image: + if image_path and os.path.exists(image_path): + return Image.open(image_path).convert("RGB") + logger.warning("No image provided, creating placeholder...") + return Image.new("RGB", (640, 352), (128, 128, 128)) + +def create_action_presets(num_frames: int, keyboard_dim: int = 4, seed: int = None): + if keyboard_dim not in (2, 4, 7): + raise ValueError(f"keyboard_dim must be 2, 4, or 7, got {keyboard_dim}") + if num_frames % 4 != 1: + raise ValueError("Matrix-Game conditioning expects num_frames to be 4k+1.") + + # Set seed for reproducibility if provided + if seed is not None: + random.seed(seed) + + num_samples_per_action = 4 + + # Define actions based on keyboard_dim + if keyboard_dim == 4: + # Universal model: W, S, A, D + actions_single_action = ["forward", "left", "right"] + actions_double_action = ["forward_left", "forward_right"] + actions_single_camera = ["camera_l", "camera_r"] + keyboard_idx = {"forward": 0, "back": 1, "left": 2, "right": 3} + elif keyboard_dim == 2: + # GTA model: W, S only (steering via mouse) + actions_single_action = ["forward", "back"] + actions_double_action = [] + actions_single_camera = ["camera_l", "camera_r"] + keyboard_idx = {"forward": 0, "back": 1} + else: # keyboard_dim == 7 + # Temple Run model: still, w, s, left, right, a, d (no mouse) + actions_single_action = ["forward", "back", "left", "right"] + actions_double_action = [] + actions_single_camera = [] # No mouse for Temple Run + keyboard_idx = {"still": 0, "forward": 1, "back": 2, "left": 3, "right": 4, "a": 5, "d": 6} + + actions_to_test = ( + actions_double_action * 5 + actions_single_camera * 5 + actions_single_action * 5 + ) + for action in (actions_single_action + actions_double_action): + for camera in actions_single_camera: + actions_to_test.append(f"{action}_{camera}") + + # Ensure we have at least some actions + if not actions_to_test: + actions_to_test = actions_single_action * 5 + + base_action = actions_single_action + actions_single_camera + cam_value = 0.1 + camera_value_map = { + "camera_up": [cam_value, 0], + "camera_down": [-cam_value, 0], + "camera_l": [0, -cam_value], + "camera_r": [0, cam_value], + "camera_ur": [cam_value, cam_value], + "camera_ul": [cam_value, -cam_value], + "camera_dr": [-cam_value, cam_value], + "camera_dl": [-cam_value, -cam_value], + } + + data = [] + for action_name in actions_to_test: + keyboard_condition = torch.zeros((num_samples_per_action, keyboard_dim)) + mouse_condition = torch.zeros((num_samples_per_action, 2)) + + for sub_act in base_action: + if sub_act not in action_name: + continue + if sub_act in camera_value_map: + mouse_condition = torch.tensor( + [camera_value_map[sub_act] for _ in range(num_samples_per_action)], + dtype=mouse_condition.dtype, + ) + elif sub_act in keyboard_idx: + keyboard_condition[:, keyboard_idx[sub_act]] = 1 + + data.append({ + "keyboard_condition": keyboard_condition, + "mouse_condition": mouse_condition, + }) + + keyboard_condition = torch.zeros((num_frames, keyboard_dim)) + mouse_condition = torch.zeros((num_frames, 2)) + current_frame = 0 + selections = [12] + + while current_frame < num_frames: + rd_frame = selections[random.randint(0, len(selections) - 1)] + entry = data[random.randint(0, len(data) - 1)] + key_seq = entry["keyboard_condition"] + mouse_seq = entry["mouse_condition"] + + if current_frame == 0: + keyboard_condition[:1] = key_seq[:1] + mouse_condition[:1] = mouse_seq[:1] + current_frame = 1 + else: + rd_frame = min(rd_frame, num_frames - current_frame) + repeat_time = rd_frame // 4 + keyboard_condition[current_frame:current_frame + rd_frame] = key_seq.repeat(repeat_time, 1) + mouse_condition[current_frame:current_frame + rd_frame] = mouse_seq.repeat(repeat_time, 1) + current_frame += rd_frame + + return {"keyboard": keyboard_condition, "mouse": mouse_condition} + +def parse_config(config, mode="universal"): + assert mode in ['universal', 'gta_drive', 'templerun'] + key_data = {} + mouse_data = {} + if mode != 'templerun': + key, mouse = config + else: + key = config + + for i in range(len(key)): + if mode == 'templerun': + still, w, s, left, right, a, d = key[i] + elif mode == 'universal': + w, s, a, d = key[i] + else: + w, s, a, d = key[i][0], key[i][1], mouse[i][1] < 0, mouse[i][1] > 0 + if mode == 'universal': + mouse_y, mouse_x = mouse[i] + mouse_y = -1 * mouse_y + + key_data[i] = {"W": bool(w), "A": bool(a), "S": bool(s), "D": bool(d)} + if mode == 'templerun': + key_data[i].update({"left": bool(left), "right": bool(right)}) + + if mode == 'universal': + if i == 0: + mouse_data[i] = (320, 352 // 2) + else: + global_scale_factor = 0.1 + mouse_scale_x = 15 * global_scale_factor + mouse_scale_y = 15 * 4 * global_scale_factor + mouse_data[i] = ( + mouse_data[i - 1][0] + mouse_x * mouse_scale_x, + mouse_data[i - 1][1] + mouse_y * mouse_scale_y, + ) + return key_data, mouse_data + +# NOTE: drawing functions are commented out to avoid cv2/libGL dependency. +# +# def draw_rounded_rectangle(image, top_left, bottom_right, color, radius=10, alpha=0.5): +# overlay = image.copy() +# x1, y1 = top_left +# x2, y2 = bottom_right +# +# cv2.rectangle(overlay, (x1 + radius, y1), (x2 - radius, y2), color, -1) +# cv2.rectangle(overlay, (x1, y1 + radius), (x2, y2 - radius), color, -1) +# cv2.ellipse(overlay, (x1 + radius, y1 + radius), (radius, radius), 180, 0, 90, color, -1) +# cv2.ellipse(overlay, (x2 - radius, y1 + radius), (radius, radius), 270, 0, 90, color, -1) +# cv2.ellipse(overlay, (x1 + radius, y2 - radius), (radius, radius), 90, 0, 90, color, -1) +# cv2.ellipse(overlay, (x2 - radius, y2 - radius), (radius, radius), 0, 0, 90, color, -1) +# cv2.addWeighted(overlay, alpha, image, 1 - alpha, 0, image) +# +# def draw_keys_on_frame(frame, keys, key_size=(80, 50), spacing=20, bottom_margin=30, mode='universal'): +# h, w, _ = frame.shape +# horison_shift = 90 +# vertical_shift = -20 +# horizon_shift_all = 50 +# key_positions = { +# "W": (w // 2 - key_size[0] // 2 - horison_shift - horizon_shift_all, +# h - bottom_margin - key_size[1] * 2 + vertical_shift - 20), +# "A": (w // 2 - key_size[0] * 2 + 5 - horison_shift - horizon_shift_all, +# h - bottom_margin - key_size[1] + vertical_shift), +# "S": (w // 2 - key_size[0] // 2 - horison_shift - horizon_shift_all, +# h - bottom_margin - key_size[1] + vertical_shift), +# "D": (w // 2 + key_size[0] - 5 - horison_shift - horizon_shift_all, +# h - bottom_margin - key_size[1] + vertical_shift), +# } +# key_icon = {"W": "W", "A": "A", "S": "S", "D": "D", "left": "left", "right": "right"} +# if mode == 'templerun': +# key_positions.update({ +# "left": (w // 2 + key_size[0] * 2 + spacing * 2 - horison_shift - horizon_shift_all, +# h - bottom_margin - key_size[1] + vertical_shift), +# "right": (w // 2 + key_size[0] * 3 + spacing * 7 - horison_shift - horizon_shift_all, +# h - bottom_margin - key_size[1] + vertical_shift) +# }) +# +# for key, (x, y) in key_positions.items(): +# is_pressed = keys.get(key, False) +# top_left = (x, y) +# if key in ["left", "right"]: +# bottom_right = (x + key_size[0] + 40, y + key_size[1]) +# else: +# bottom_right = (x + key_size[0], y + key_size[1]) +# +# color = (0, 255, 0) if is_pressed else (200, 200, 200) +# alpha = 0.8 if is_pressed else 0.5 +# draw_rounded_rectangle(frame, top_left, bottom_right, color, radius=10, alpha=alpha) +# +# text_size = cv2.getTextSize(key, cv2.FONT_HERSHEY_SIMPLEX, 0.8, 2)[0] +# if key in ["left", "right"]: +# text_x = x + (key_size[0] + 40 - text_size[0]) // 2 +# else: +# text_x = x + (key_size[0] - text_size[0]) // 2 +# text_y = y + (key_size[1] + text_size[1]) // 2 +# cv2.putText(frame, key_icon[key], (text_x, text_y), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 0), 2) +# +# def overlay_icon(frame, icon, position, scale=1.0, rotation=0): +# x, y = position +# h, w, _ = icon.shape +# +# scaled_width = int(w * scale) +# scaled_height = int(h * scale) +# icon_resized = cv2.resize(icon, (scaled_width, scaled_height), interpolation=cv2.INTER_AREA) +# +# center = (scaled_width // 2, scaled_height // 2) +# rotation_matrix = cv2.getRotationMatrix2D(center, rotation, 1.0) +# icon_rotated = cv2.warpAffine( +# icon_resized, rotation_matrix, (scaled_width, scaled_height), +# flags=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT, borderValue=(0, 0, 0, 0) +# ) +# +# h, w, _ = icon_rotated.shape +# frame_h, frame_w, _ = frame.shape +# +# top_left_x = max(0, int(x - w // 2)) +# top_left_y = max(0, int(y - h // 2)) +# bottom_right_x = min(frame_w, int(x + w // 2)) +# bottom_right_y = min(frame_h, int(y + h // 2)) +# +# icon_x_start = max(0, int(-x + w // 2)) +# icon_y_start = max(0, int(-y + h // 2)) +# icon_x_end = icon_x_start + (bottom_right_x - top_left_x) +# icon_y_end = icon_y_start + (bottom_right_y - top_left_y) +# +# icon_region = icon_rotated[icon_y_start:icon_y_end, icon_x_start:icon_x_end] +# alpha = icon_region[:, :, 3] / 255.0 +# icon_rgb = icon_region[:, :, :3] +# +# frame_region = frame[top_left_y:bottom_right_y, top_left_x:bottom_right_x] +# for c in range(3): +# frame_region[:, :, c] = (1 - alpha) * frame_region[:, :, c] + alpha * icon_rgb[:, :, c] +# frame[top_left_y:bottom_right_y, top_left_x:bottom_right_x] = frame_region +# +# def process_video(input_video, output_video, config, mouse_icon_path, +# mouse_scale=1.0, mouse_rotation=0, process_icon=True, mode='universal'): +# key_data, mouse_data = parse_config(config, mode=mode) +# fps = 12 +# +# mouse_icon = cv2.imread(mouse_icon_path, cv2.IMREAD_UNCHANGED) +# +# out_video = [] +# for frame_idx, frame in enumerate(input_video): +# frame = np.ascontiguousarray(frame) +# if process_icon: +# keys = key_data.get(frame_idx, {"W": False, "A": False, "S": False, "D": False, "left": False, "right": False}) +# draw_keys_on_frame(frame, keys, key_size=(50, 50), spacing=10, bottom_margin=20, mode=mode) +# if mode == 'universal': +# frame_width = frame.shape[1] +# frame_height = frame.shape[0] +# mouse_position = mouse_data.get(frame_idx, (frame_width // 2, frame_height // 2)) +# overlay_icon(frame, mouse_icon, mouse_position, scale=mouse_scale, rotation=mouse_rotation) +# out_video.append(frame / 255) +# +# export_to_video(out_video, output_video, fps=fps) +# logger.info(f"Video saved to {output_video}") diff --git a/fastvideo/models/dits/wanvideo.py b/fastvideo/models/dits/wanvideo.py index 359db5742..480a36cc8 100644 --- a/fastvideo/models/dits/wanvideo.py +++ b/fastvideo/models/dits/wanvideo.py @@ -1,6 +1,7 @@ # SPDX-License-Identifier: Apache-2.0 import math +from contextlib import nullcontext from typing import Any import torch @@ -26,6 +27,7 @@ from fastvideo.layers.visual_embedding import (ModulateProjection, PatchEmbed, from fastvideo.logger import init_logger from fastvideo.models.dits.base import BaseDiT from fastvideo.platforms import AttentionBackendEnum, current_platform +from fastvideo.layers.quantization import QuantizationConfig from fastvideo.distributed.parallel_state import get_sp_world_size @@ -106,7 +108,9 @@ class WanSelfAttention(nn.Module): window_size=(-1, -1), qk_norm=True, eps=1e-6, - parallel_attention=False) -> None: + parallel_attention=False, + quant_config: QuantizationConfig | None = None, + prefix: str = "") -> None: assert dim % num_heads == 0 super().__init__() self.dim = dim @@ -118,10 +122,10 @@ class WanSelfAttention(nn.Module): self.parallel_attention = parallel_attention # layers - self.to_q = ReplicatedLinear(dim, dim) - self.to_k = ReplicatedLinear(dim, dim) - self.to_v = ReplicatedLinear(dim, dim) - self.to_out = ReplicatedLinear(dim, dim) + self.to_q = ReplicatedLinear(dim, dim, quant_config=quant_config, prefix=f"{prefix}.to_q") + self.to_k = ReplicatedLinear(dim, dim, quant_config=quant_config, prefix=f"{prefix}.to_k") + self.to_v = ReplicatedLinear(dim, dim, quant_config=quant_config, prefix=f"{prefix}.to_v") + self.to_out = ReplicatedLinear(dim, dim, quant_config=quant_config, prefix=f"{prefix}.to_out") self.norm_q = RMSNorm(dim, eps=eps) if qk_norm else nn.Identity() self.norm_k = RMSNorm(dim, eps=eps) if qk_norm else nn.Identity() @@ -194,13 +198,15 @@ class WanI2VCrossAttention(WanSelfAttention): qk_norm=True, eps=1e-6, supported_attention_backends: tuple[AttentionBackendEnum, ...] - | None = None + | None = None, + quant_config: QuantizationConfig | None = None, + prefix: str = "", ) -> None: super().__init__(dim, num_heads, window_size, qk_norm, eps, - supported_attention_backends) + supported_attention_backends, quant_config=quant_config, prefix=prefix) - self.add_k_proj = ReplicatedLinear(dim, dim) - self.add_v_proj = ReplicatedLinear(dim, dim) + self.add_k_proj = ReplicatedLinear(dim, dim, quant_config=quant_config, prefix=f"{prefix}.add_k_proj") + self.add_v_proj = ReplicatedLinear(dim, dim, quant_config=quant_config, prefix=f"{prefix}.add_v_proj") self.norm_added_k = RMSNorm(dim, eps=eps) if qk_norm else nn.Identity() self.norm_added_q = RMSNorm(dim, eps=eps) if qk_norm else nn.Identity() @@ -246,16 +252,17 @@ class WanTransformerBlock(nn.Module): added_kv_proj_dim: int | None = None, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, + quant_config: QuantizationConfig | None = None, prefix: str = ""): super().__init__() # 1. Self-attention self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False) - self.to_q = ReplicatedLinear(dim, dim, bias=True) - self.to_k = ReplicatedLinear(dim, dim, bias=True) - self.to_v = ReplicatedLinear(dim, dim, bias=True) + self.to_q = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_q") + self.to_k = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_k") + self.to_v = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_v") - self.to_out = ReplicatedLinear(dim, dim, bias=True) + self.to_out = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_out") self.attn1 = DistributedAttention( num_heads=num_heads, head_size=dim // num_heads, @@ -290,13 +297,17 @@ class WanTransformerBlock(nn.Module): self.attn2 = WanI2VCrossAttention(dim, num_heads, qk_norm=qk_norm, - eps=eps) + eps=eps, + quant_config=quant_config, + prefix=f"{prefix}.attn2") else: # T2V self.attn2 = WanT2VCrossAttention(dim, num_heads, qk_norm=qk_norm, - eps=eps) + eps=eps, + quant_config=quant_config, + prefix=f"{prefix}.attn2") self.cross_attn_residual_norm = ScaleResidualLayerNormScaleShift( dim, norm_type="layer", @@ -306,7 +317,7 @@ class WanTransformerBlock(nn.Module): compute_dtype=torch.float32) # 3. Feed-forward - self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh") + self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh", quant_config=quant_config, prefix=f"{prefix}.ffn") self.mlp_residual = ScaleResidual() self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5) @@ -406,17 +417,17 @@ class WanTransformerBlock_VSA(nn.Module): added_kv_proj_dim: int | None = None, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, + quant_config: QuantizationConfig | None = None, prefix: str = ""): super().__init__() # 1. Self-attention self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False) - self.to_q = ReplicatedLinear(dim, dim, bias=True) - self.to_k = ReplicatedLinear(dim, dim, bias=True) - self.to_v = ReplicatedLinear(dim, dim, bias=True) - self.to_gate_compress = ReplicatedLinear(dim, dim, bias=True) - - self.to_out = ReplicatedLinear(dim, dim, bias=True) + self.to_q = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_q") + self.to_k = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_k") + self.to_v = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_v") + self.to_gate_compress = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_gate_compress") + self.to_out = ReplicatedLinear(dim, dim, bias=True, quant_config=quant_config, prefix=f"{prefix}.to_out") self.attn1 = DistributedAttention_VSA( num_heads=num_heads, head_size=dim // num_heads, @@ -451,13 +462,17 @@ class WanTransformerBlock_VSA(nn.Module): self.attn2 = WanI2VCrossAttention(dim, num_heads, qk_norm=qk_norm, - eps=eps) + eps=eps, + quant_config=quant_config, + prefix=f"{prefix}.attn2") else: # T2V self.attn2 = WanT2VCrossAttention(dim, num_heads, qk_norm=qk_norm, - eps=eps) + eps=eps, + quant_config=quant_config, + prefix=f"{prefix}.attn2") self.cross_attn_residual_norm = ScaleResidualLayerNormScaleShift( dim, norm_type="layer", @@ -467,7 +482,7 @@ class WanTransformerBlock_VSA(nn.Module): compute_dtype=torch.float32) # 3. Feed-forward - self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh") + self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh", quant_config=quant_config, prefix=f"{prefix}.ffn") self.mlp_residual = ScaleResidual() self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5) @@ -556,6 +571,7 @@ class WanTransformer3DModel(BaseDiT): def __init__(self, config: WanVideoConfig, hf_config: dict[str, Any]) -> None: super().__init__(config=config, hf_config=hf_config) + self.quant_config = config.quant_config inner_dim = config.num_attention_heads * config.attention_head_dim self.hidden_size = config.hidden_size @@ -594,6 +610,7 @@ class WanTransformer3DModel(BaseDiT): config.eps, config.added_kv_proj_dim, self._supported_attention_backends, + quant_config=config.quant_config, prefix=f"{config.prefix}.blocks.{i}") for i in range(config.num_layers) ]) diff --git a/fastvideo/models/loader/component_loader.py b/fastvideo/models/loader/component_loader.py index 9dfde7cb2..837faea44 100644 --- a/fastvideo/models/loader/component_loader.py +++ b/fastvideo/models/loader/component_loader.py @@ -561,6 +561,7 @@ class TokenizerLoader(ComponentLoader): # If parsing fails, fall through to AutoTokenizer below. pass + tokenizer = AutoTokenizer.from_pretrained( resolved_model_path, # "/tokenizer" # in v0, this was same string as encoder_name "ClipTextModel" diff --git a/fastvideo/models/loader/fsdp_load.py b/fastvideo/models/loader/fsdp_load.py index f0e449414..59730196b 100644 --- a/fastvideo/models/loader/fsdp_load.py +++ b/fastvideo/models/loader/fsdp_load.py @@ -56,7 +56,9 @@ def set_default_dtype(dtype: torch.dtype) -> Generator[None, None, None]: torch.set_default_dtype(old_dtype) -# Supports optional torch.compile for FSDP-wrapped models during training +USE_FP8 = False +USE_FP4 = False # TODO: move to FastVideoArgs +# TODO(PY): add compile option def maybe_load_fsdp_model( model_cls: type[nn.Module], init_params: dict[str, Any], @@ -157,6 +159,8 @@ def maybe_load_fsdp_model( # Avoid unintended computation graph accumulation during inference if isinstance(p, torch.nn.Parameter): p.requires_grad = False + if USE_FP4: + convert_model_to_fp4(model) compile_in_loader = enable_torch_compile and training_mode if compile_in_loader: @@ -165,6 +169,8 @@ def maybe_load_fsdp_model( compile_kwargs) model = torch.compile(model, **compile_kwargs) logger.info("torch.compile enabled for %s", type(model).__name__) + # elif USE_FP8: + # convert_model_to_fp8(model) return model diff --git a/fastvideo/models/schedulers/scheduling_rcm.py b/fastvideo/models/schedulers/scheduling_rcm.py index 71cc714b0..e69de29bb 100644 --- a/fastvideo/models/schedulers/scheduling_rcm.py +++ b/fastvideo/models/schedulers/scheduling_rcm.py @@ -1,326 +0,0 @@ -# SPDX-License-Identifier: Apache-2.0 -# rCM (recurrent Consistency Model) Scheduler for TurboDiffusion support -# -# This scheduler implements the rCM sampling method from TurboDiffusion, -# enabling 1-4 step video generation with distilled checkpoints. -# -# Reference: -# TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times -# https://arxiv.org/pdf/2512.16093 - -import math -from dataclasses import dataclass -from typing import Any - -import torch -from diffusers.configuration_utils import ConfigMixin, register_to_config -from diffusers.schedulers.scheduling_utils import SchedulerMixin -from diffusers.utils import BaseOutput - -from fastvideo.logger import init_logger -from fastvideo.models.schedulers.base import BaseScheduler - -logger = init_logger(__name__) - - -@dataclass -class RCMSchedulerOutput(BaseOutput): - """ - Output class for the RCM scheduler's `step` function output. - - Args: - prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, ...)`): - Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used - as next model input in the denoising loop. - """ - - prev_sample: torch.FloatTensor - - -class RCMScheduler(SchedulerMixin, ConfigMixin, BaseScheduler): - """ - rCM (recurrent Consistency Model) scheduler for TurboDiffusion. - - This scheduler implements the rCM sampling method which enables 1-4 step - video generation using distilled checkpoints. It uses: - 1. TrigFlow → RectifiedFlow timestep conversion - 2. SDE sampling formula: x = (1 - t_next) * (x - t_cur * v_pred) + t_next * noise - - Args: - num_train_timesteps (`int`, defaults to 1000): - The number of diffusion steps used to train the model. - sigma_max (`float`, defaults to 80.0): - The initial sigma value for rCM sampling. Controls the noise level - at the start of sampling. - mid_timesteps (`list[float]`, *optional*): - Custom intermediate timesteps. If None, uses optimized defaults - [1.5, 1.4, 1.0] for better visual quality. - """ - - _compatibles: list[Any] = [] - order = 1 - - @register_to_config - def __init__( - self, - num_train_timesteps: int = 1000, - sigma_max: float = 80.0, - mid_timesteps: list[float] | None = None, - ): - # Default mid timesteps optimized for visual quality - if mid_timesteps is None: - mid_timesteps = [1.5, 1.4, 1.0] - self._mid_timesteps = mid_timesteps - - self.num_train_timesteps = num_train_timesteps - self.sigma_max = sigma_max - - # Initialize with default timesteps (will be set properly via set_timesteps) - self.timesteps = torch.tensor([1.0, 0.0], dtype=torch.float64) - self.sigmas = self.timesteps.clone() - - self._step_index: int | None = None - self._begin_index: int | None = None - - BaseScheduler.__init__(self) - - @property - def step_index(self) -> int | None: - """ - The index counter for current timestep. Increases by 1 after each scheduler step. - """ - return self._step_index - - @property - def begin_index(self) -> int | None: - """ - The index for the first timestep. - """ - return self._begin_index - - @property - def init_noise_sigma(self) -> float: - """ - Initial noise sigma for scaling latents. - - In rCM, initial noise is scaled by the first sigma value: - x_0 = noise * sigmas[0] - - This property is used by LatentPreparationStage to scale initial latents. - """ - return float(self.sigmas[0]) - - def set_begin_index(self, begin_index: int = 0) -> None: - """ - Sets the begin index for the scheduler. - - Args: - begin_index (`int`): - The begin index for the scheduler. - """ - self._begin_index = begin_index - - def set_shift(self, shift: float) -> None: - """ - rCM doesn't use shift parameter, but required by BaseScheduler. - """ - pass - - def set_timesteps( - self, - num_inference_steps: int, - device: str | torch.device | None = None, - sigma_max: float | None = None, - ) -> None: - """ - Sets the discrete timesteps used for the rCM sampling process. - - The timesteps are computed using TrigFlow → RectifiedFlow conversion: - 1. Start with atan(sigma_max) and intermediate values - 2. Convert via: t = sin(t) / (cos(t) + sin(t)) - - Args: - num_inference_steps (`int`): - The number of diffusion steps (1-4 for rCM). - device (`str` or `torch.device`, *optional*): - The device to move timesteps to. - sigma_max (`float`, *optional*): - Override the initial sigma value. - """ - if num_inference_steps < 1 or num_inference_steps > 4: - logger.warning( - "rCM is optimized for 1-4 steps, got %d steps. " - "Performance may be suboptimal.", num_inference_steps - ) - - self.num_inference_steps = num_inference_steps - - if sigma_max is not None: - self.sigma_max = sigma_max - - # Build timestep schedule - mid_t = self._mid_timesteps[:num_inference_steps - 1] - - # TrigFlow timesteps: [atan(sigma_max), mid_t..., 0] - t_steps = torch.tensor( - [math.atan(self.sigma_max), *mid_t, 0], - dtype=torch.float64, - device=device, - ) - - # Convert TrigFlow → RectifiedFlow: t = sin(t) / (cos(t) + sin(t)) - t_steps = torch.sin(t_steps) / (torch.cos(t_steps) + torch.sin(t_steps)) - - # Store raw sigmas for use in step() formula - self.sigmas = t_steps.clone() - - # Scale timesteps by 1000 for model input (as per TurboDiffusion) - self.timesteps = t_steps * 1000 - - self._step_index = None - self._begin_index = None - - logger.debug("rCM timesteps (scaled): %s", self.timesteps.tolist()) - logger.debug("rCM sigmas (raw): %s", self.sigmas.tolist()) - - def _init_step_index(self, timestep: torch.FloatTensor | None = None) -> None: - """Initialize step index at the beginning of sampling.""" - if self._begin_index is None: - self._step_index = 0 - else: - self._step_index = self._begin_index - - def scale_model_input( - self, - sample: torch.Tensor, - timestep: int | None = None, - ) -> torch.Tensor: - """ - rCM doesn't scale model input, returns sample as-is. - """ - return sample - - def scale_noise( - self, - sample: torch.FloatTensor, - timestep: torch.FloatTensor | None = None, - noise: torch.FloatTensor | None = None, - ) -> torch.FloatTensor: - """ - Scale initial noise for rCM sampling. - - In rCM, initial noise is scaled by the first timestep (raw sigma): - x_0 = noise * t_steps[0] - - Args: - sample: Not used (for API compatibility) - timestep: Not used (for API compatibility) - noise: The noise tensor to scale - - Returns: - Scaled noise tensor ready for sampling - """ - if noise is None: - raise ValueError("noise must be provided for rCM scale_noise") - - # Use raw sigma (not scaled timestep) for initial noise scaling - t_initial = self.sigmas[0] - return noise.to(torch.float64) * t_initial - - def step( - self, - model_output: torch.FloatTensor, - timestep: int | torch.Tensor, - sample: torch.FloatTensor, - generator: torch.Generator | None = None, - return_dict: bool = True, - ) -> RCMSchedulerOutput | tuple[torch.FloatTensor, ...]: - """ - Predict the sample from the previous timestep using rCM update rule. - - The rCM update formula is: - x_{t+1} = (1 - t_next) * (x_t - t_cur * v_pred) + t_next * noise - - Args: - model_output (`torch.FloatTensor`): - The velocity prediction from the model (v_pred). - timestep (`int` or `torch.Tensor`): - The current timestep index (not the actual timestep value). - For rCM, this should be the index into self.timesteps. - sample (`torch.FloatTensor`): - Current sample x_t. - generator (`torch.Generator`, *optional*): - Random number generator for noise. - return_dict (`bool`): - Whether to return RCMSchedulerOutput or tuple. - - Returns: - `RCMSchedulerOutput` or `tuple`: - The denoised sample for the next step. - """ - if self._step_index is None: - self._init_step_index() - - assert self._step_index is not None - - # Get current and next sigma values (raw, unscaled) for rCM formula - # Note: self.timesteps is scaled by 1000 for model input, - # but we need raw values for the step formula - t_cur = self.sigmas[self._step_index] - # On the final step, t_next should be 0 (fully denoised) - if self._step_index + 1 < len(self.sigmas): - t_next = self.sigmas[self._step_index + 1] - else: - t_next = torch.tensor(0.0, device=sample.device, dtype=torch.float64) - - # Ensure we're working in float64 for precision - sample = sample.to(torch.float64) - model_output = model_output.to(torch.float64) - - # rCM update: x = (1 - t_next) * (x - t_cur * v_pred) + t_next * noise - x_denoised = sample - t_cur * model_output - - # Generate noise for SDE sampling - if isinstance(generator, list): - generator = generator[0] - noise = torch.randn( - sample.shape, - dtype=torch.float32, - device="cpu", - generator=generator, - ).to(sample.device).to(torch.float64) - - prev_sample = (1 - t_next) * x_denoised + t_next * noise - - # Increment step counter - self._step_index += 1 - - # Cast back to model output dtype - prev_sample = prev_sample.to(model_output.dtype) - - if not return_dict: - return (prev_sample,) - - return RCMSchedulerOutput(prev_sample=prev_sample) - - def add_noise( - self, - original_samples: torch.Tensor, - noise: torch.Tensor, - timesteps: torch.IntTensor, - ) -> torch.Tensor: - """ - Add noise to samples (forward diffusion process). - - Not typically used for rCM inference, but provided for API compatibility. - """ - raise NotImplementedError( - "add_noise is not implemented for RCMScheduler. " - "Use scale_noise for initializing the sampling process." - ) - - def __len__(self) -> int: - return self.config.num_train_timesteps - -# Entry point for model registry -EntryClass = RCMScheduler diff --git a/fastvideo/pipelines/basic/turbodiffusion/turbodiffusion_pipeline.py b/fastvideo/pipelines/basic/turbodiffusion/turbodiffusion_pipeline.py index f49f0c91c..8ec834c48 100644 --- a/fastvideo/pipelines/basic/turbodiffusion/turbodiffusion_pipeline.py +++ b/fastvideo/pipelines/basic/turbodiffusion/turbodiffusion_pipeline.py @@ -11,8 +11,16 @@ from fastvideo.fastvideo_args import FastVideoArgs from fastvideo.logger import init_logger from fastvideo.models.schedulers.scheduling_rcm import RCMScheduler from fastvideo.pipelines import ComposedPipelineBase, LoRAPipeline +<<<<<<< HEAD from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage, DenoisingStage, InputValidationStage, LatentPreparationStage, TextEncodingStage, TimestepPreparationStage) +======= +from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage, + DenoisingStage, InputValidationStage, + LatentPreparationStage, + TextEncodingStage, + TimestepPreparationStage) +>>>>>>> matthew/clean logger = init_logger(__name__) @@ -24,7 +32,13 @@ class TurboDiffusionPipeline(LoRAPipeline, ComposedPipelineBase): Uses RCM scheduler and SLA attention for fast, high-quality video generation. """ +<<<<<<< HEAD _required_config_modules = ["text_encoder", "tokenizer", "vae", "transformer", "scheduler"] +======= + _required_config_modules = [ + "text_encoder", "tokenizer", "vae", "transformer", "scheduler" + ] +>>>>>>> matthew/clean def initialize_pipeline(self, fastvideo_args: FastVideoArgs): # Use RCM scheduler for TurboDiffusion @@ -34,7 +48,12 @@ class TurboDiffusionPipeline(LoRAPipeline, ComposedPipelineBase): def create_pipeline_stages(self, fastvideo_args: FastVideoArgs) -> None: """Set up pipeline stages with proper dependency injection.""" +<<<<<<< HEAD self.add_stage(stage_name="input_validation_stage", stage=InputValidationStage()) +======= + self.add_stage(stage_name="input_validation_stage", + stage=InputValidationStage()) +>>>>>>> matthew/clean self.add_stage(stage_name="prompt_encoding_stage", stage=TextEncodingStage( @@ -42,6 +61,7 @@ class TurboDiffusionPipeline(LoRAPipeline, ComposedPipelineBase): tokenizers=[self.get_module("tokenizer")], )) +<<<<<<< HEAD self.add_stage(stage_name="conditioning_stage", stage=ConditioningStage()) self.add_stage(stage_name="timestep_preparation_stage", @@ -59,6 +79,31 @@ class TurboDiffusionPipeline(LoRAPipeline, ComposedPipelineBase): pipeline=self)) self.add_stage(stage_name="decoding_stage", stage=DecodingStage(vae=self.get_module("vae"), pipeline=self)) +======= + self.add_stage(stage_name="conditioning_stage", + stage=ConditioningStage()) + + self.add_stage(stage_name="timestep_preparation_stage", + stage=TimestepPreparationStage( + scheduler=self.get_module("scheduler"))) + + self.add_stage(stage_name="latent_preparation_stage", + stage=LatentPreparationStage( + scheduler=self.get_module("scheduler"), + transformer=self.get_module("transformer", None))) + + self.add_stage(stage_name="denoising_stage", + stage=DenoisingStage( + transformer=self.get_module("transformer"), + transformer_2=self.get_module("transformer_2", None), + scheduler=self.get_module("scheduler"), + vae=self.get_module("vae"), + pipeline=self)) + + self.add_stage(stage_name="decoding_stage", + stage=DecodingStage(vae=self.get_module("vae"), + pipeline=self)) +>>>>>>> matthew/clean EntryClass = TurboDiffusionPipeline diff --git a/fastvideo/pipelines/stages/conditioning.py b/fastvideo/pipelines/stages/conditioning.py index 80fb3b6cc..bbc639ec8 100644 --- a/fastvideo/pipelines/stages/conditioning.py +++ b/fastvideo/pipelines/stages/conditioning.py @@ -86,9 +86,5 @@ class ConditioningStage(PipelineStage): def verify_output(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> VerificationResult: """Verify conditioning stage outputs.""" result = VerificationResult() - if batch.prompt_embeds is None or not batch.prompt_embeds: - batch.do_classifier_free_guidance = False - return result - if batch.do_classifier_free_guidance or batch.prompt_embeds: - result.add_check("prompt_embeds", batch.prompt_embeds, V.list_not_empty) + result.add_check("prompt_embeds", batch.prompt_embeds, V.list_not_empty) return result diff --git a/fastvideo/pipelines/stages/decoding.py b/fastvideo/pipelines/stages/decoding.py index cc88fbd7e..a6c9e2448 100644 --- a/fastvideo/pipelines/stages/decoding.py +++ b/fastvideo/pipelines/stages/decoding.py @@ -244,4 +244,4 @@ class DecodingStage(PipelineStage): del pipeline.modules["vae"] fastvideo_args.model_loaded["vae"] = False - return batch + return batch \ No newline at end of file diff --git a/fastvideo/pipelines/stages/denoising.py b/fastvideo/pipelines/stages/denoising.py index 1efeec734..64bdfdffa 100644 --- a/fastvideo/pipelines/stages/denoising.py +++ b/fastvideo/pipelines/stages/denoising.py @@ -142,6 +142,14 @@ class DenoisingStage(PipelineStage): }, ) + action_kwargs = self.prepare_extra_func_kwargs( + self.transformer.forward, + { + "mouse_cond": batch.mouse_cond, + "keyboard_cond": batch.keyboard_cond, + }, + ) + action_kwargs = self.prepare_extra_func_kwargs( self.transformer.forward, { @@ -158,6 +166,7 @@ class DenoisingStage(PipelineStage): }, ) + # Get latents and embeddings latents = batch.latents prompt_embeds = batch.prompt_embeds diff --git a/fastvideo/pipelines/stages/matrixgame_denoising.py b/fastvideo/pipelines/stages/matrixgame_denoising.py index 7e1496857..e69de29bb 100644 --- a/fastvideo/pipelines/stages/matrixgame_denoising.py +++ b/fastvideo/pipelines/stages/matrixgame_denoising.py @@ -1,739 +0,0 @@ -from __future__ import annotations -from collections.abc import Callable -from dataclasses import dataclass -from typing import Any - -import torch # type: ignore - -from fastvideo.distributed import get_local_torch_device -from fastvideo.fastvideo_args import FastVideoArgs -from fastvideo.forward_context import set_forward_context -from fastvideo.logger import init_logger -from fastvideo.models.utils import pred_noise_to_pred_video, pred_noise_to_x_bound -from fastvideo.pipelines.pipeline_batch_info import ForwardBatch -from fastvideo.pipelines.stages.denoising import DenoisingStage -from fastvideo.pipelines.stages.validators import StageValidators as V -from fastvideo.pipelines.stages.validators import VerificationResult - -try: - from fastvideo.attention.backends.video_sparse_attn import (VideoSparseAttentionBackend) - vsa_available = True -except ImportError: - vsa_available = False - VideoSparseAttentionBackend = None # type: ignore - -logger = init_logger(__name__) - - -@dataclass -class BlockProcessingContext: - """Dataclass contains for block processing.""" - - batch: ForwardBatch - - block_idx: int - start_index: int - - kv_cache1: list[dict[Any, Any]] - kv_cache2: list[dict[Any, Any]] | None - kv_cache_mouse: list[dict[Any, Any]] | None - kv_cache_keyboard: list[dict[Any, Any]] | None - crossattn_cache: list[dict[Any, Any]] - - timesteps: torch.Tensor - block_sizes: list[int] - noise_pool: list[torch.Tensor] | None - - fastvideo_args: FastVideoArgs - target_dtype: torch.dtype - autocast_enabled: bool - boundary_timestep: float | None - high_noise_timesteps: torch.Tensor | None - context_noise: float - - image_kwargs: dict[str, Any] - pos_cond_kwargs: dict[str, Any] - - def get_kv_cache(self, timestep_val: float) -> list[dict[Any, Any]]: - if self.boundary_timestep is not None: - if timestep_val >= self.boundary_timestep: - return self.kv_cache1 - else: - assert self.kv_cache2 is not None, "kv_cache2 is not initialized" - return self.kv_cache2 - return self.kv_cache1 - - -class MatrixGameCausalDenoisingStage(DenoisingStage): - - def __init__(self, transformer, scheduler, pipeline=None, transformer_2=None, vae=None) -> None: - super().__init__(transformer, scheduler, pipeline, transformer_2, vae) - self.transformer = transformer - self.transformer_2 = transformer_2 - self.vae = vae - self.num_transformer_blocks = len(self.transformer.blocks) - - if hasattr(self.transformer, 'config') and hasattr(self.transformer.config, 'arch_config'): - self.num_frame_per_block = getattr(self.transformer.config.arch_config, 'num_frames_per_block', - getattr(self.transformer, 'num_frame_per_block', 1)) - self.sliding_window_num_frames = getattr(self.transformer.config.arch_config, 'sliding_window_num_frames', - 15) - else: - self.num_frame_per_block = getattr(self.transformer, 'num_frame_per_block', 1) - self.sliding_window_num_frames = 15 - - try: - self.local_attn_size = getattr(self.transformer, "local_attn_size", -1) - except Exception: - self.local_attn_size = -1 - - assert self.local_attn_size != -1, (f"local_attn_size must be set for Matrix-Game causal inference, " - f"got {self.local_attn_size}. Check MatrixGameWanVideoArchConfig.") - assert self.num_frame_per_block > 0, (f"num_frame_per_block must be positive, got {self.num_frame_per_block}") - - logger.info("MatrixGame causal inference initialized: " - "local_attn_size=%s, num_frame_per_block=%s", self.local_attn_size, self.num_frame_per_block) - - self.action_config = getattr(self.transformer, 'action_config', {}) - self.use_action_module = len(self.action_config) > 0 - - self._streaming_initialized: bool = False - self._streaming_ctx: BlockProcessingContext | None = None - - def forward( - self, - batch: ForwardBatch, - fastvideo_args: FastVideoArgs, - ) -> ForwardBatch: - target_dtype = torch.bfloat16 - autocast_enabled = (target_dtype != torch.float32) and not fastvideo_args.disable_autocast - - latent_seq_length = batch.latents.shape[-1] * batch.latents.shape[-2] - patch_size = self.transformer.patch_size - patch_ratio = patch_size[-1] * patch_size[-2] - self.frame_seq_length = latent_seq_length // patch_ratio - - timesteps = torch.tensor(fastvideo_args.pipeline_config.dmd_denoising_steps, dtype=torch.long).cpu() - if fastvideo_args.pipeline_config.warp_denoising_step: - scheduler_timesteps = torch.cat((self.scheduler.timesteps.cpu(), torch.tensor([0], dtype=torch.float32))) - timesteps = scheduler_timesteps[1000 - timesteps] - timesteps = timesteps.to(get_local_torch_device()) - - boundary_ratio = getattr(fastvideo_args.pipeline_config.dit_config, 'boundary_ratio', None) - if boundary_ratio is not None: - boundary_timestep = boundary_ratio * self.scheduler.num_train_timesteps - high_noise_timesteps = timesteps[timesteps >= boundary_timestep] - else: - boundary_timestep = None - high_noise_timesteps = None - - image_embeds = batch.image_embeds - if len(image_embeds) > 0: - assert torch.isnan(image_embeds[0]).sum() == 0 - image_embeds = [image_embed.to(target_dtype) for image_embed in image_embeds] - - # directly set the kwarg. - image_kwargs = {"encoder_hidden_states_image": image_embeds} - pos_cond_kwargs: dict[str, Any] = {} - - assert batch.latents is not None, "latents must be provided" - latents = batch.latents - b, c, t, h, w = latents.shape - prompt_embeds = batch.prompt_embeds - assert torch.isnan(prompt_embeds[0]).sum() == 0 - - kv_cache1 = self._initialize_kv_cache(batch_size=latents.shape[0], dtype=target_dtype, device=latents.device) - kv_cache2 = None - if boundary_timestep is not None: - kv_cache2 = self._initialize_kv_cache(batch_size=latents.shape[0], - dtype=target_dtype, - device=latents.device) - - kv_cache_mouse = None - kv_cache_keyboard = None - if self.use_action_module: - kv_cache_mouse, kv_cache_keyboard = self._initialize_action_kv_cache(batch_size=latents.shape[0], - dtype=target_dtype, - device=latents.device) - - crossattn_cache = self._initialize_crossattn_cache( - batch_size=latents.shape[0], - max_text_len=257, # 1 CLS + 256 patch tokens - dtype=target_dtype, - device=latents.device) - - if t % self.num_frame_per_block != 0: - raise ValueError("num_frames must be divisible by num_frame_per_block for causal denoising") - num_blocks = t // self.num_frame_per_block - block_sizes = [self.num_frame_per_block] * num_blocks - start_index = 0 - - if boundary_timestep is not None: - block_sizes[0] = 1 - - # NOTE: MatrixGame does NOT process the first frame separately. - # The first frame information is already encoded in batch.image_latent (cond_concat) - # and will be used by the model via channel concatenation: torch.cat([x, cond_concat], dim=1) - - ctx = BlockProcessingContext( - batch=batch, - block_idx=0, - start_index=0, - kv_cache1=kv_cache1, - kv_cache2=kv_cache2, - kv_cache_mouse=kv_cache_mouse, - kv_cache_keyboard=kv_cache_keyboard, - crossattn_cache=crossattn_cache, - timesteps=timesteps, - block_sizes=block_sizes, - noise_pool=None, - fastvideo_args=fastvideo_args, - target_dtype=target_dtype, - autocast_enabled=autocast_enabled, - boundary_timestep=boundary_timestep, - high_noise_timesteps=high_noise_timesteps, - context_noise=getattr(fastvideo_args.pipeline_config, "context_noise", 0), - image_kwargs=image_kwargs, - pos_cond_kwargs=pos_cond_kwargs, - ) - - context_noise = getattr(fastvideo_args.pipeline_config, "context_noise", 0) - - with self.progress_bar(total=len(block_sizes) * len(timesteps)) as progress_bar: - for block_idx, current_num_frames in enumerate(block_sizes): - ctx.block_idx = block_idx - ctx.start_index = start_index - current_latents = latents[:, :, start_index:start_index + current_num_frames, :, :] - - action_kwargs = self._prepare_action_kwargs(batch, start_index, current_num_frames) - - current_latents = self._process_single_block( - current_latents=current_latents, - batch=batch, - start_index=start_index, - current_num_frames=current_num_frames, - timesteps=timesteps, - ctx=ctx, - action_kwargs=action_kwargs, - progress_bar=progress_bar, - ) - - latents[:, :, start_index:start_index + current_num_frames, :, :] = current_latents - - # Update KV caches with clean context - self._update_context_cache( - current_latents=current_latents, - batch=batch, - start_index=start_index, - current_num_frames=current_num_frames, - ctx=ctx, - action_kwargs=action_kwargs, - context_noise=context_noise, - ) - - start_index += current_num_frames - - if boundary_timestep is not None: - num_frames_to_remove = self.num_frame_per_block - 1 - if num_frames_to_remove > 0: - latents = latents[:, :, :-num_frames_to_remove, :, :] - - batch.latents = latents - return batch - - def _prepare_action_kwargs(self, batch: ForwardBatch, start_index: int, num_frames: int) -> dict[str, Any]: - action_kwargs: dict[str, Any] = {} - if not self.use_action_module: - return action_kwargs - - vae_time_compression_ratio = 4 - end_frame_idx = 1 + vae_time_compression_ratio * (start_index + num_frames - 1) - if hasattr(batch, 'mouse_cond') and batch.mouse_cond is not None: - action_kwargs['mouse_cond'] = batch.mouse_cond[:, :end_frame_idx] - if hasattr(batch, 'keyboard_cond') and batch.keyboard_cond is not None: - action_kwargs['keyboard_cond'] = batch.keyboard_cond[:, :end_frame_idx] - - # CRITICAL: Pass num_frame_per_block to model - this should be num_frames (current block size) - action_kwargs['num_frame_per_block'] = num_frames - return action_kwargs - - def _initialize_kv_cache(self, batch_size: int, dtype: torch.dtype, device: torch.device) -> list[dict]: - kv_cache = [] - num_attention_heads = self.transformer.num_attention_heads - attention_head_dim = getattr(self.transformer, 'attention_head_dim', - self.transformer.hidden_size // num_attention_heads) - if self.local_attn_size != -1: - kv_cache_size = self.local_attn_size * self.frame_seq_length - else: - kv_cache_size = self.frame_seq_length * self.sliding_window_num_frames - - for _ in range(self.num_transformer_blocks): - kv_cache.append({ - "k": - torch.zeros([batch_size, kv_cache_size, num_attention_heads, attention_head_dim], - dtype=dtype, - device=device), - "v": - torch.zeros([batch_size, kv_cache_size, num_attention_heads, attention_head_dim], - dtype=dtype, - device=device), - "global_end_index": - torch.tensor([0], dtype=torch.long, device=device), - "local_end_index": - torch.tensor([0], dtype=torch.long, device=device), - }) - - return kv_cache - - def _initialize_action_kv_cache(self, batch_size: int, dtype: torch.dtype, device: torch.device): - kv_cache_mouse = [] - kv_cache_keyboard = [] - - action_heads = self.action_config.get('heads_num', 16) - mouse_head_dim = self.action_config.get('mouse_hidden_dim', 1024) // action_heads - keyboard_head_dim = self.action_config.get('keyboard_hidden_dim', 1024) // action_heads - - kv_cache_size = self.local_attn_size if self.local_attn_size != -1 else 15 - - for _ in range(self.num_transformer_blocks): - kv_cache_keyboard.append({ - "k": - torch.zeros([batch_size, kv_cache_size, action_heads, keyboard_head_dim], dtype=dtype, device=device), - "v": - torch.zeros([batch_size, kv_cache_size, action_heads, keyboard_head_dim], dtype=dtype, device=device), - "global_end_index": - torch.tensor([0], dtype=torch.long, device=device), - "local_end_index": - torch.tensor([0], dtype=torch.long, device=device), - }) - kv_cache_mouse.append({ - "k": - torch.zeros([batch_size * self.frame_seq_length, kv_cache_size, action_heads, mouse_head_dim], - dtype=dtype, - device=device), - "v": - torch.zeros([batch_size * self.frame_seq_length, kv_cache_size, action_heads, mouse_head_dim], - dtype=dtype, - device=device), - "global_end_index": - torch.tensor([0], dtype=torch.long, device=device), - "local_end_index": - torch.tensor([0], dtype=torch.long, device=device), - }) - - return kv_cache_mouse, kv_cache_keyboard - - def _initialize_crossattn_cache(self, batch_size: int, max_text_len: int, dtype: torch.dtype, - device: torch.device) -> list[dict]: - crossattn_cache = [] - num_attention_heads = self.transformer.num_attention_heads - attention_head_dim = getattr(self.transformer, 'attention_head_dim', - self.transformer.hidden_size // num_attention_heads) - for _ in range(self.num_transformer_blocks): - crossattn_cache.append({ - "k": - torch.zeros([batch_size, max_text_len, num_attention_heads, attention_head_dim], - dtype=dtype, - device=device), - "v": - torch.zeros([batch_size, max_text_len, num_attention_heads, attention_head_dim], - dtype=dtype, - device=device), - "is_init": - False, - }) - return crossattn_cache - - def _process_single_block( - self, - current_latents: torch.Tensor, - batch: ForwardBatch, - start_index: int, - current_num_frames: int, - timesteps: torch.Tensor, - ctx: BlockProcessingContext, - action_kwargs: dict[str, Any], - noise_generator: Callable[[tuple, torch.dtype, int], torch.Tensor] - | None = None, - progress_bar: Any | None = None, - ) -> torch.Tensor: - prompt_embeds = batch.prompt_embeds - noise_latents_btchw = current_latents.permute(0, 2, 1, 3, 4) - - for i, t_cur in enumerate(timesteps): - if ctx.boundary_timestep is not None and t_cur < ctx.boundary_timestep: - current_model = self.transformer_2 if self.transformer_2 is not None else self.transformer - else: - current_model = self.transformer - - noise_latents = noise_latents_btchw.clone() - latent_model_input = current_latents.to(ctx.target_dtype) - - independent_first_frame = getattr(self.transformer, 'independent_first_frame', False) - if batch.image_latent is not None and independent_first_frame and start_index == 0: - latent_model_input = torch.cat([latent_model_input, batch.image_latent.to(ctx.target_dtype)], dim=2) - - # t_expand needs to be [batch * frames] to match flattened pred_noise/noise_latents - t_expand = t_cur.repeat(latent_model_input.shape[0] * current_num_frames) - - # Build attention metadata if VSA is available - if vsa_available and self.attn_backend == VideoSparseAttentionBackend: - self.attn_metadata_builder_cls = self.attn_backend.get_builder_cls() - if self.attn_metadata_builder_cls is not None: - self.attn_metadata_builder = self.attn_metadata_builder_cls() - h, w = current_latents.shape[-2:] - attn_metadata = self.attn_metadata_builder.build( - current_timestep=i, - raw_latent_shape=(current_num_frames, h, w), - patch_size=ctx.fastvideo_args.pipeline_config.dit_config.patch_size, - VSA_sparsity=ctx.fastvideo_args.VSA_sparsity, - device=get_local_torch_device(), - ) - assert attn_metadata is not None, "attn_metadata cannot be None" - else: - attn_metadata = None - else: - attn_metadata = None - - with torch.autocast(device_type="cuda", - dtype=ctx.target_dtype, - enabled=ctx.autocast_enabled), \ - set_forward_context(current_timestep=i, - attn_metadata=attn_metadata, - forward_batch=batch): - # Expand timestep to per-frame format [batch, num_frames] for causal model - t_expanded_noise = t_cur * torch.ones((latent_model_input.shape[0], current_num_frames), - device=latent_model_input.device, - dtype=torch.long) - - model_kwargs = { - "kv_cache": ctx.get_kv_cache(t_cur), - "crossattn_cache": ctx.crossattn_cache, - "current_start": start_index * self.frame_seq_length, - "start_frame": start_index, - } - - if self.use_action_module and current_model == self.transformer: - model_kwargs.update({ - "kv_cache_mouse": ctx.kv_cache_mouse, - "kv_cache_keyboard": ctx.kv_cache_keyboard, - }) - model_kwargs.update(action_kwargs) - - pred_noise_btchw = current_model( - latent_model_input, - prompt_embeds, - t_expanded_noise, - **ctx.image_kwargs, - **ctx.pos_cond_kwargs, - **model_kwargs, - ).permute(0, 2, 1, 3, 4) - - if ctx.boundary_timestep is not None and t_cur >= ctx.boundary_timestep: - pred_video_btchw = pred_noise_to_x_bound( - pred_noise=pred_noise_btchw.flatten(0, 1), - noise_input_latent=noise_latents.flatten(0, 1), - timestep=t_expand, - boundary_timestep=torch.ones_like(t_expand) * ctx.boundary_timestep, - scheduler=self.scheduler).unflatten(0, pred_noise_btchw.shape[:2]) - else: - pred_video_btchw = pred_noise_to_pred_video(pred_noise=pred_noise_btchw.flatten(0, 1), - noise_input_latent=noise_latents.flatten(0, 1), - timestep=t_expand, - scheduler=self.scheduler).unflatten( - 0, pred_noise_btchw.shape[:2]) - - if i < len(timesteps) - 1: - next_timestep = timesteps[i + 1] * torch.ones([1], dtype=torch.long, device=pred_video_btchw.device) - - # Use custom noise generator if provided (for streaming), else generate - if noise_generator is not None: - noise = noise_generator(pred_video_btchw.shape, pred_video_btchw.dtype, i) - else: - noise = torch.randn( - pred_video_btchw.shape, - dtype=pred_video_btchw.dtype, - generator=(batch.generator[0] if isinstance(batch.generator, list) else batch.generator)).to( - pred_video_btchw.device) - - noise_btchw = noise - if ctx.boundary_timestep is not None and ctx.high_noise_timesteps is not None and i < len( - ctx.high_noise_timesteps) - 1: - noise_latents_btchw = self.scheduler.add_noise_high( - pred_video_btchw.flatten(0, 1), noise_btchw.flatten(0, 1), next_timestep, - torch.ones_like(next_timestep) * ctx.boundary_timestep).unflatten( - 0, pred_video_btchw.shape[:2]) - elif ctx.boundary_timestep is not None and ctx.high_noise_timesteps is not None and i == len( - ctx.high_noise_timesteps) - 1: - noise_latents_btchw = pred_video_btchw - else: - noise_latents_btchw = self.scheduler.add_noise(pred_video_btchw.flatten(0, 1), - noise_btchw.flatten(0, 1), next_timestep).unflatten( - 0, pred_video_btchw.shape[:2]) - current_latents = noise_latents_btchw.permute(0, 2, 1, 3, 4) - else: - current_latents = pred_video_btchw.permute(0, 2, 1, 3, 4) - - if progress_bar is not None: - progress_bar.update() - - return current_latents - - def _update_context_cache( - self, - current_latents: torch.Tensor, - batch: ForwardBatch, - start_index: int, - current_num_frames: int, - ctx: BlockProcessingContext, - action_kwargs: dict[str, Any], - context_noise: float, - ) -> None: - prompt_embeds = batch.prompt_embeds - latents_device = current_latents.device - - # Expand context timestep to per-frame format [batch, num_frames] for causal model - t_context = torch.ones([current_latents.shape[0], current_num_frames], device=latents_device, - dtype=torch.long) * int(context_noise) - context_bcthw = current_latents.to(ctx.target_dtype) - - with torch.autocast(device_type="cuda", - dtype=ctx.target_dtype, - enabled=ctx.autocast_enabled), \ - set_forward_context(current_timestep=0, - attn_metadata=None, - forward_batch=batch): - - context_model_kwargs = { - "kv_cache": ctx.kv_cache1, - "crossattn_cache": ctx.crossattn_cache, - "current_start": start_index * self.frame_seq_length, - "start_frame": start_index, - } - - if self.use_action_module: - context_model_kwargs.update({ - "kv_cache_mouse": ctx.kv_cache_mouse, - "kv_cache_keyboard": ctx.kv_cache_keyboard, - }) - context_model_kwargs.update(action_kwargs) - - if ctx.boundary_timestep is not None and self.transformer_2 is not None: - self.transformer_2( - context_bcthw, - prompt_embeds, - t_context, - kv_cache=ctx.kv_cache2, - crossattn_cache=ctx.crossattn_cache, - current_start=start_index * self.frame_seq_length, - start_frame=start_index, - **ctx.image_kwargs, - **ctx.pos_cond_kwargs, - ) - - self.transformer( - context_bcthw, - prompt_embeds, - t_context, - **ctx.image_kwargs, - **ctx.pos_cond_kwargs, - **context_model_kwargs, - ) - - def streaming_reset(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch: - target_dtype = torch.bfloat16 - autocast_enabled = (target_dtype != torch.float32) and not fastvideo_args.disable_autocast - - latent_seq_length = batch.latents.shape[-1] * batch.latents.shape[-2] - patch_size = self.transformer.patch_size - patch_ratio = patch_size[-1] * patch_size[-2] - self.frame_seq_length = latent_seq_length // patch_ratio - - timesteps = torch.tensor(fastvideo_args.pipeline_config.dmd_denoising_steps, dtype=torch.long).cpu() - if fastvideo_args.pipeline_config.warp_denoising_step: - scheduler_timesteps = torch.cat((self.scheduler.timesteps.cpu(), torch.tensor([0], dtype=torch.float32))) - timesteps = scheduler_timesteps[1000 - timesteps] - timesteps = timesteps.to(get_local_torch_device()) - - boundary_ratio = getattr(fastvideo_args.pipeline_config.dit_config, 'boundary_ratio', None) - if boundary_ratio is not None: - boundary_timestep = boundary_ratio * self.scheduler.num_train_timesteps - high_noise_timesteps = timesteps[timesteps >= boundary_timestep] - else: - boundary_timestep = None - high_noise_timesteps = None - - image_embeds = batch.image_embeds - if len(image_embeds) > 0: - assert torch.isnan(image_embeds[0]).sum() == 0 - image_embeds = [image_embed.to(target_dtype) for image_embed in image_embeds] - - # directly set the kwarg. - image_kwargs = {"encoder_hidden_states_image": image_embeds} - pos_cond_kwargs: dict[str, Any] = {} - - assert batch.latents is not None, "latents must be provided" - latents = batch.latents - b, c, t, h, w = latents.shape - prompt_embeds = batch.prompt_embeds - assert torch.isnan(prompt_embeds[0]).sum() == 0 - - # Initialize caches - kv_cache1 = self._initialize_kv_cache(batch_size=latents.shape[0], dtype=target_dtype, device=latents.device) - kv_cache2 = None - if boundary_timestep is not None: - kv_cache2 = self._initialize_kv_cache(batch_size=latents.shape[0], - dtype=target_dtype, - device=latents.device) - - kv_cache_mouse = None - kv_cache_keyboard = None - if self.use_action_module: - kv_cache_mouse, kv_cache_keyboard = self._initialize_action_kv_cache(batch_size=latents.shape[0], - dtype=target_dtype, - device=latents.device) - - crossattn_cache = self._initialize_crossattn_cache( - batch_size=latents.shape[0], - max_text_len=257, # 1 CLS + 256 patch tokens - dtype=target_dtype, - device=latents.device) - - # Calculate block sizes - if t % self.num_frame_per_block != 0: - raise ValueError("num_frames must be divisible by num_frame_per_block for causal denoising") - num_blocks = t // self.num_frame_per_block - block_sizes = [self.num_frame_per_block] * num_blocks - if boundary_timestep is not None: - block_sizes[0] = 1 - - # Pre-allocate noise pool - num_denoising_steps = len(timesteps) - noise_shape = (b, self.num_frame_per_block, c, h, w) - noise_pool = [ - torch.randn( - noise_shape, - dtype=target_dtype, - device=latents.device, - ) for _ in range(num_denoising_steps - 1) - ] - - # Create and store context - self._streaming_ctx = BlockProcessingContext( - batch=batch, - block_idx=0, - start_index=0, - kv_cache1=kv_cache1, - kv_cache2=kv_cache2, - kv_cache_mouse=kv_cache_mouse, - kv_cache_keyboard=kv_cache_keyboard, - crossattn_cache=crossattn_cache, - timesteps=timesteps, - block_sizes=block_sizes, - noise_pool=noise_pool, - fastvideo_args=fastvideo_args, - target_dtype=target_dtype, - autocast_enabled=autocast_enabled, - boundary_timestep=boundary_timestep, - high_noise_timesteps=high_noise_timesteps, - context_noise=getattr(fastvideo_args.pipeline_config, "context_noise", 0), - image_kwargs=image_kwargs, - pos_cond_kwargs=pos_cond_kwargs, - ) - - self._streaming_initialized = True - return batch - - def streaming_step(self, - keyboard_action: torch.Tensor | None = None, - mouse_action: torch.Tensor | None = None) -> ForwardBatch: - if not self._streaming_initialized or self._streaming_ctx is None: - raise RuntimeError("Streaming not initialized! Call streaming_reset first.") - - ctx = self._streaming_ctx - if ctx.block_idx >= len(ctx.block_sizes): - return ctx.batch - - batch = ctx.batch - latents = batch.latents - assert latents is not None, "latents must be set in batch" - - current_num_frames = ctx.block_sizes[ctx.block_idx] - start_index = ctx.start_index - - current_latents = latents[:, :, start_index:start_index + current_num_frames, :, :] - - # Update batch with new actions for this block - if keyboard_action is not None or mouse_action is not None: - vae_ratio = 4 - start_frame = 0 if start_index == 0 else 1 + vae_ratio * (start_index - 1) - - if keyboard_action is not None: - n = keyboard_action.shape[1] - batch.keyboard_cond[:, start_frame:start_frame + n] = keyboard_action.to(batch.keyboard_cond.device) - if mouse_action is not None: - n = mouse_action.shape[1] - batch.mouse_cond[:, start_frame:start_frame + n] = mouse_action.to(batch.mouse_cond.device) - - action_kwargs = self._prepare_action_kwargs(batch, start_index, current_num_frames) - - # Create noise generator that uses pre-allocated noise pool - def streaming_noise_generator(shape: tuple, dtype: torch.dtype, step_idx: int) -> torch.Tensor: - if ctx.noise_pool is not None and step_idx < len(ctx.noise_pool): - return ctx.noise_pool[step_idx][:, :shape[1], :, :, :].to(latents.device) - else: - # Fallback to dynamic allocation if pool not available - return torch.randn( - shape, - dtype=dtype, - generator=(batch.generator[0] if isinstance(batch.generator, list) else batch.generator)).to( - latents.device) - - current_latents = self._process_single_block( - current_latents=current_latents, - batch=batch, - start_index=start_index, - current_num_frames=current_num_frames, - timesteps=ctx.timesteps, - ctx=ctx, - action_kwargs=action_kwargs, - noise_generator=streaming_noise_generator, - ) - - latents[:, :, start_index:start_index + current_num_frames, :, :] = current_latents - - # Update KV caches with clean context - self._update_context_cache( - current_latents=current_latents, - batch=batch, - start_index=start_index, - current_num_frames=current_num_frames, - ctx=ctx, - action_kwargs=action_kwargs, - context_noise=ctx.context_noise, - ) - - # Advance streaming state - ctx.start_index += current_num_frames - ctx.block_idx += 1 - - return batch - - def streaming_clear(self) -> None: - self._streaming_initialized = False - self._streaming_ctx = None - - def verify_input(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> VerificationResult: - result = VerificationResult() - result.add_check("latents", batch.latents, [V.is_tensor, V.with_dims(5)]) - result.add_check("prompt_embeds", batch.prompt_embeds, V.list_not_empty) - result.add_check("image_embeds", batch.image_embeds, V.is_list) - result.add_check("image_latent", batch.image_latent, V.none_or_tensor_with_dims(5)) - result.add_check("num_inference_steps", batch.num_inference_steps, V.positive_int) - result.add_check("guidance_scale", batch.guidance_scale, V.positive_float) - result.add_check("eta", batch.eta, V.non_negative_float) - result.add_check("generator", batch.generator, V.generator_or_list_generators) - result.add_check("do_classifier_free_guidance", batch.do_classifier_free_guidance, V.bool_value) - result.add_check("negative_prompt_embeds", batch.negative_prompt_embeds, - lambda x: not batch.do_classifier_free_guidance or V.list_not_empty(x)) - return result diff --git a/fastvideo/pipelines/stages/text_encoding.py b/fastvideo/pipelines/stages/text_encoding.py index b62d98bbc..b8a246567 100644 --- a/fastvideo/pipelines/stages/text_encoding.py +++ b/fastvideo/pipelines/stages/text_encoding.py @@ -85,9 +85,6 @@ class TextEncodingStage(PipelineStage): encoder_index=all_indices, return_attention_mask=True, ) - if self._last_audio_embeds is not None: - batch.extra["ltx2_audio_negative_embeds"] = self._last_audio_embeds - assert batch.negative_prompt_embeds is not None for ne in neg_embeds_list: batch.negative_prompt_embeds.append(ne) @@ -104,7 +101,8 @@ class TextEncodingStage(PipelineStage): # result.add_check( # "negative_prompt", batch.negative_prompt, lambda x: not batch. # do_classifier_free_guidance or V.string_not_empty(x)) - result.add_check("do_classifier_free_guidance", batch.do_classifier_free_guidance, V.bool_value) + result.add_check("do_classifier_free_guidance", + batch.do_classifier_free_guidance, V.bool_value) result.add_check("prompt_embeds", batch.prompt_embeds, V.is_list) result.add_check("negative_prompt_embeds", batch.negative_prompt_embeds, V.none_or_list) return result @@ -185,8 +183,8 @@ class TextEncodingStage(PipelineStage): preprocess_funcs = fastvideo_args.pipeline_config.preprocess_text_funcs postprocess_funcs = fastvideo_args.pipeline_config.postprocess_text_funcs encoder_cfgs = fastvideo_args.pipeline_config.text_encoder_configs - is_ltx2 = getattr(fastvideo_args.pipeline_config.dit_config, "prefix", "") == "ltx2" - + is_ltx2 = getattr(fastvideo_args.pipeline_config.dit_config, "prefix", + "") == "ltx2" if return_type not in ("list", "dict", "stack"): raise ValueError(f"Invalid return_type '{return_type}'. Expected one of: 'list', 'dict', 'stack'") @@ -361,3 +359,69 @@ class Cosmos25TextEncodingStage(PipelineStage): result.add_check("negative_prompt_embeds", batch.negative_prompt_embeds, lambda x: not batch.do_classifier_free_guidance or V.list_not_empty(x)) return result + + +class Cosmos25TextEncodingStage(PipelineStage): + """Cosmos 2.5 text encoding stage. + + Cosmos 2.5 uses Reason1 (Qwen2.5-VL) and relies on the encoder's + `compute_text_embeddings_online()`. + """ + + def __init__(self, text_encoder) -> None: + super().__init__() + self.text_encoder = text_encoder + + @torch.no_grad() + def forward(self, batch: ForwardBatch, + fastvideo_args: FastVideoArgs) -> ForwardBatch: + assert batch.prompt is not None + prompts = [batch.prompt] if isinstance(batch.prompt, + str) else batch.prompt + + encoder = self.text_encoder + if not hasattr(encoder, "compute_text_embeddings_online"): + raise RuntimeError( + "Cosmos25TextEncodingStage requires text_encoder.compute_text_embeddings_online()" + ) + + with set_forward_context(current_timestep=0, attn_metadata=None): + prompt_embeds = encoder.compute_text_embeddings_online( + {"text": prompts}, "text") + + batch.prompt_embeds = [prompt_embeds] + + if batch.do_classifier_free_guidance: + neg = batch.negative_prompt + neg_prompts = ([neg] * + len(prompts)) if isinstance(neg, str) else neg + with set_forward_context(current_timestep=0, attn_metadata=None): + neg_embeds = encoder.compute_text_embeddings_online( + {"text": neg_prompts}, "text") + batch.negative_prompt_embeds = [neg_embeds] + else: + batch.negative_prompt_embeds = [] + + return batch + + def verify_input(self, batch: ForwardBatch, + fastvideo_args: FastVideoArgs) -> VerificationResult: + result = VerificationResult() + result.add_check("prompt", batch.prompt, V.string_or_list_strings) + result.add_check( + "negative_prompt", + batch.negative_prompt, + lambda x: + (not batch.do_classifier_free_guidance) or isinstance(x, str), + ) + return result + + def verify_output(self, batch: ForwardBatch, + fastvideo_args: FastVideoArgs) -> VerificationResult: + result = VerificationResult() + result.add_check("prompt_embeds", batch.prompt_embeds, + V.list_of_tensors_min_dims(2)) + result.add_check( + "negative_prompt_embeds", batch.negative_prompt_embeds, lambda x: + not batch.do_classifier_free_guidance or V.list_not_empty(x)) + return result diff --git a/fastvideo/tests/nightly/test_fp4_linear.py b/fastvideo/tests/nightly/test_fp4_linear.py new file mode 100644 index 000000000..d75731339 --- /dev/null +++ b/fastvideo/tests/nightly/test_fp4_linear.py @@ -0,0 +1,156 @@ +#!/usr/bin/env python3 +import math +import sys +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fastvideo.layers.fp4linear import LinearFWD4BWD16 + +def fail(msg: str): + print(f"[FAIL] {msg}") + sys.exit(1) + +def metric_report(ref, test, name, cos_min, rel_l2_max): + diff = (test - ref).float() + ref32 = ref.float() + test32 = test.float() + + mean_abs_diff = diff.abs().mean().item() + max_abs_diff = diff.abs().max().item() + rel_l2 = (diff.norm() / (ref32.norm() + 1e-12)).item() + cos = F.cosine_similarity(ref32.reshape(-1), test32.reshape(-1), dim=0).item() + + print(f" [{name}] cos: {cos:.6f}, rel_l2: {rel_l2:.6e}, " + f"mean_abs: {mean_abs_diff:.6e}, max_abs: {max_abs_diff:.6e}") + if cos_min is not None and cos < cos_min: + fail(f"{name}: cosine similarity too low: {cos:.6f} < {cos_min}") + if rel_l2_max is not None and rel_l2 > rel_l2_max: + fail(f"{name}: relative L2 too high: {rel_l2:.6e} > {rel_l2_max}") + +def run_case(case, fwd_cos_min=0.98, fwd_rel_l2_max=0.15, + bwd_cos_min=0.999, bwd_rel_l2_max=5e-3): + device = torch.device("cuda") + torch.manual_seed(0) + + B = case["batch_size"] + S = case["seq_len"] + IN = case["in_features"] + OUT = case["out_features"] + NAME = case["name"] + M = B * S + + print(f"\n== Case: {NAME} | M={M}, K={IN}, N={OUT} ==") + + # Reference BF16 Linear + ref = nn.Linear(IN, OUT, bias=True).to(device=device, dtype=torch.bfloat16) + + # FP4-forward layer (stores fp32, casts to activation dtype internally) + try: + LinearFWD4BWD16 + except NameError: + print("LinearFWD4BWD16 is not defined/imported. Please import it before running.") + sys.exit(1) + + test_layer = LinearFWD4BWD16( + IN, OUT, bias=True, backend="cutlass", block_size=16, use_128x4_sf_layout=True + ).to(device) + + # Initialize identically from the same fp32 master params + with torch.no_grad(): + w32 = torch.empty(OUT, IN, device=device, dtype=torch.float32) + nn.init.kaiming_uniform_(w32, a=math.sqrt(5.0)) + b32 = torch.empty(OUT, device=device, dtype=torch.float32) + nn.init.uniform_(b32, -1 / math.sqrt(IN), 1 / math.sqrt(IN)) + + test_layer.weight.copy_(w32) + if test_layer.bias is not None: + test_layer.bias.copy_(b32) + + ref.weight.copy_(w32.to(dtype=torch.bfloat16)) + if ref.bias is not None: + ref.bias.copy_(b32.to(dtype=torch.bfloat16)) + + # Inputs + x = torch.randn(M, IN, device=device, dtype=torch.bfloat16, requires_grad=True) + x_ref = x.clone().detach().requires_grad_(True) + + # -------- Forward comparison -------- + with torch.inference_mode(): + y_ref = ref(x_ref) + y_fp4 = test_layer(x) + + print("Forward metrics (FP4-forward vs BF16):") + metric_report(y_ref, y_fp4, "forward", fwd_cos_min, fwd_rel_l2_max) + + # -------- Backward comparison -------- + # Inject the SAME external grad_out so gradients are comparable. + grad_out = torch.randn_like(y_ref) + + # Clear grads + for p in ref.parameters(): + if p.grad is not None: + p.grad.zero_() + for p in test_layer.parameters(): + if p.grad is not None: + p.grad.zero_() + if x.grad is not None: + x.grad.zero_() + if x_ref.grad is not None: + x_ref.grad.zero_() + + # Recompute forward with grad graph (no inference_mode) + y_ref = ref(x_ref) + y_fp4 = test_layer(x) + + # Backward with same grad_out + y_ref.backward(grad_out) + y_fp4.backward(grad_out) + + # Collect grads (cast to fp32 for fair comparison) + gx_ref = x_ref.grad.float() + gw_ref = ref.weight.grad.float() + gb_ref = ref.bias.grad.float() if ref.bias is not None else None + + gx_fp4 = x.grad.float() + # test_layer stores weight in fp32 → grad is fp32 already + gw_fp4 = test_layer.weight.grad.float() + gb_fp4 = test_layer.bias.grad.float() if test_layer.bias is not None else None + + print("Backward metrics (grads parity with injected grad_out):") + metric_report(gx_ref, gx_fp4, "grad_x", bwd_cos_min, bwd_rel_l2_max) + metric_report(gw_ref, gw_fp4, "grad_w", bwd_cos_min, bwd_rel_l2_max) + if gb_ref is not None and gb_fp4 is not None: + metric_report(gb_ref, gb_fp4, "grad_b", bwd_cos_min, bwd_rel_l2_max) + +def main(): + if not torch.cuda.is_available(): + print("CUDA is required.") + sys.exit(1) + + # Your exact config list + configs = [ + {'batch_size': 1, 'seq_len': 1, 'in_features': 256, 'out_features': 5120, 'name': 'fc_in_256x5120'}, + {'batch_size': 1, 'seq_len': 1, 'in_features': 5120, 'out_features': 5120, 'name': 'fc_out_5120x5120'}, + {'batch_size': 1, 'seq_len': 1, 'in_features': 5120, 'out_features': 30720,'name': 'fc_5120x30720'}, + {'batch_size': 1, 'seq_len': 512,'in_features': 4096, 'out_features': 5120, 'name': 'fc_in_512x4096x5120'}, + {'batch_size': 1, 'seq_len': 512,'in_features': 5120, 'out_features': 5120, 'name': 'fc_out_512x5120x5120'}, + {'batch_size': 1, 'seq_len': 32768,'in_features': 5120,'out_features': 5120, 'name': 'to_q_32768x5120x5120'}, + {'batch_size': 1, 'seq_len': 32768,'in_features': 5120,'out_features': 13824,'name': 'ffn_fc_in_32768x5120x13824'}, + {'batch_size': 1, 'seq_len': 32768,'in_features': 13824,'out_features': 5120, 'name': 'ffn_fc_out_32768x13824x5120'}, + ] + + # Tolerances: forward is approximate (FP4), backward should match closely + fwd_cos_min = 0.98 + fwd_rel_l2_max = 0.15 + bwd_cos_min = 0.999 + bwd_rel_l2_max = 5e-3 + + for case in configs: + run_case(case, fwd_cos_min, fwd_rel_l2_max, bwd_cos_min, bwd_rel_l2_max) + torch.cuda.synchronize() + + print("\n[OK] All forward & backward checks passed for all cases.") + +if __name__ == "__main__": + main() diff --git a/fastvideo/tests/nightly/test_sage3_with_16bit_bwd.py b/fastvideo/tests/nightly/test_sage3_with_16bit_bwd.py new file mode 100644 index 000000000..6ab2080d3 --- /dev/null +++ b/fastvideo/tests/nightly/test_sage3_with_16bit_bwd.py @@ -0,0 +1,137 @@ +#!/usr/bin/env python3 +# SPDX-License-Identifier: Apache-2.0 + +import sys +import math +import torch +import torch.nn.functional as F + +# ---- Import function-under-test ---- +# Make sure this import path matches your project layout. +# If it's in the same file, remove this import and keep the definition above. +from fastvideo.attention.backends.sage_attn3 import attn_forward_4bit_fwd_16bit_bwd +from fastvideo.attention.backends.sage_attn3 import sageattn_blackwell_with_16bit_bwd + +# ---- Import FlashAttention forward (public) ---- +# We DO NOT use _wrapped_* internals. Prefer the top-level alias; fall back if needed. +try: + from flash_attn import flash_attn_func as flash_attn +except Exception: + try: + from flash_attn.flash_attn_interface import flash_attn_func as flash_attn + except Exception as e: + print("Could not import flash_attn function:", e) + sys.exit(1) + + +def fail(msg: str): + print(f"[FAIL] {msg}") + sys.exit(1) + + +def metric_report(ref, test, name, cos_min=None, rel_l2_max=None): + ref32 = ref.float() + test32 = test.float() + diff = test32 - ref32 + + mean_abs = diff.abs().mean().item() + max_abs = diff.abs().max().item() + rel_l2 = (diff.norm() / (ref32.norm() + 1e-12)).item() + cos = F.cosine_similarity(ref32.reshape(-1), test32.reshape(-1), dim=0).item() + + print(f" [{name}] cos={cos:.6f} rel_l2={rel_l2:.6e} mean_abs={mean_abs:.6e} max_abs={max_abs:.6e}") + if cos_min is not None and cos < cos_min: + fail(f"{name}: cosine too low: {cos:.6f} < {cos_min}") + if rel_l2_max is not None and rel_l2 > rel_l2_max: + fail(f"{name}: relative L2 too high: {rel_l2:.6e} > {rel_l2_max}") + + +def run_case(case, fwd_cos_min=0.95, fwd_rel_l2_max=0.25, bwd_cos_min=0.95, bwd_rel_l2_max=0.25): + if not torch.cuda.is_available(): + print("CUDA is required.") + sys.exit(1) + + device = torch.device("cuda") + torch.manual_seed(case.get("seed", 0)) + + B = case["B"] + L = case["L"] + H = case["H"] + D = case["D"] + dtype = case["dtype"] + causal = case["causal"] + name = case["name"] + + print(f"\n== Case: {name} | B={B}, L={L}, H={H}, D={D}, dtype={str(dtype).split('.')[-1]}, causal={causal} ==") + + # Inputs in BLHD (flash_attn expects BLHD) + q = torch.randn(B, L, H, D, device=device, dtype=dtype, requires_grad=True) + k = torch.randn(B, L, H, D, device=device, dtype=dtype, requires_grad=True) + v = torch.randn(B, L, H, D, device=device, dtype=dtype, requires_grad=True) + + softmax_scale = 1.0 / math.sqrt(D) + + # ----- Forward ----- + # Our kernel (forward uses sage, backward will use FA inside the autograd fn) + y_sage = sageattn_blackwell_with_16bit_bwd(q, k, v, is_causal=causal) + + # Sanity checks + if y_sage.shape != (B, L, H, D): + fail(f"Output shape mismatch: got {tuple(y_sage.shape)}, expected {(B, L, H, D)}") + if not torch.isfinite(y_sage).all(): + fail("Non-finite values detected in forward output") + + # FlashAttention reference forward (public function) + with torch.inference_mode(): + y_fa = attn_forward_4bit_fwd_16bit_bwd(q, k, v, is_causal=causal) + + print("Forward numeric check vs FlashAttention:") + metric_report(y_fa, y_sage, "forward", cos_min=fwd_cos_min, rel_l2_max=fwd_rel_l2_max) + + # ----- Backward ----- + # Inject the SAME grad_out for both models so gradients are comparable + grad_out = torch.randn_like(y_sage) + + # Keep separate copies with grad for the FA path + q_ref = q.detach().clone().requires_grad_(True) + k_ref = k.detach().clone().requires_grad_(True) + v_ref = v.detach().clone().requires_grad_(True) + + # Backprop through our function + # (It recomputes FA fwd+bwd internally in 16-bit per your implementation) + for t in (q, k, v): + if t.grad is not None: + t.grad.zero_() + y_sage.backward(grad_out) + dq_sage, dk_sage, dv_sage = q.grad.detach(), k.grad.detach(), v.grad.detach() + + # Backprop through FA public forward + for t in (q_ref, k_ref, v_ref): + if t.grad is not None: + t.grad.zero_() + y_ref = attn_forward_4bit_fwd_16bit_bwd(q_ref, k_ref, v_ref, is_causal=causal) + y_ref.backward(grad_out) + + dq_ref, dk_ref, dv_ref = q_ref.grad.detach(), k_ref.grad.detach(), v_ref.grad.detach() + + print("Backward gradient checks vs FlashAttention:") + metric_report(dq_ref, dq_sage, "grad_q", cos_min=bwd_cos_min, rel_l2_max=bwd_rel_l2_max) + metric_report(dk_ref, dk_sage, "grad_k", cos_min=bwd_cos_min, rel_l2_max=bwd_rel_l2_max) + metric_report(dv_ref, dv_sage, "grad_v", cos_min=bwd_cos_min, rel_l2_max=bwd_rel_l2_max) + +def main(): + # Keep sizes moderate so it runs quickly; adjust as you like. + cases = [ + {"name": "bf16_noncausal_L16384", "B": 1, "L": 1024, "H": 8, "D": 128, "dtype": torch.bfloat16, "causal": False}, + {"name": "bf16_noncausal_L2048", "B": 1, "L": 9360, "H": 8, "D": 128, "dtype": torch.bfloat16, "causal": False}, + ] + + for case in cases: + run_case(case) + torch.cuda.synchronize() + + print("\n[OK] All tests passed.") + + +if __name__ == "__main__": + main() diff --git a/fastvideo/tests/nightly/test_torch_fp4_qk.py b/fastvideo/tests/nightly/test_torch_fp4_qk.py new file mode 100644 index 000000000..3493c3dc0 --- /dev/null +++ b/fastvideo/tests/nightly/test_torch_fp4_qk.py @@ -0,0 +1,85 @@ +#!/usr/bin/env python3 +import math +import sys +import torch +import torch.nn.functional as F +from fastvideo.attention.backends.sage_attn3 import fp_16_qk, sage_attn_qk_torch + + +def fail(msg: str): + print(f"[FAIL] {msg}") + sys.exit(1) + + +def metric_report(ref, test, name, cos_min=None, rel_l2_max=None): + diff = (test - ref).float() + ref32 = ref.float() + test32 = test.float() + + mean_abs_diff = diff.abs().mean().item() + max_abs_diff = diff.abs().max().item() + rel_l2 = (diff.norm() / (ref32.norm() + 1e-12)).item() + cos = F.cosine_similarity(ref32.reshape(-1), test32.reshape(-1), dim=0).item() + + print(f" [{name}] cos: {cos:.6f}, rel_l2: {rel_l2:.6e}, " + f"mean_abs: {mean_abs_diff:.6e}, max_abs: {max_abs_diff:.6e}") + + if cos_min is not None and cos < cos_min: + fail(f"{name}: cosine similarity too low: {cos:.6f} < {cos_min}") + if rel_l2_max is not None and rel_l2 > rel_l2_max: + fail(f"{name}: relative L2 too high: {rel_l2:.6e} > {rel_l2_max}") + + +def run_case(case, fwd_cos_min=0.98, fwd_rel_l2_max=2.0e-1): + device = torch.device("cuda") + torch.manual_seed(0) + + B, H, L, D = case["B"], case["H"], case["L"], case["D"] + NAME = case["name"] + assert B == 1, "sage_attn_qk_torch requires B == 1" + + print(f"\n== Case: {NAME} | B={B}, H={H}, L={L}, D={D} ==") + + dtype = torch.bfloat16 + q = torch.randn(B, H, L, D, device=device, dtype=dtype) + k = torch.randn(B, H, L, D, device=device, dtype=dtype) + + with torch.inference_mode(): + # Reference computes q @ k^T / sqrt(D) + p_ref = fp_16_qk(q, k.transpose(-1, -2)) # [B,H,L,L] + p_sage = sage_attn_qk_torch(q, k, False) # [B,H,L,L] + print("Forward metrics (sage_attn FP4 vs fp16/bf16 reference):") + metric_report(p_ref, p_sage, "forward", fwd_cos_min, fwd_rel_l2_max) + + +def main(): + try: + _ = fp_16_qk + _ = sage_attn_qk_torch + except NameError: + print("Please import/define fp_16_qk and sage_attn_qk_torch before running.") + sys.exit(1) + + if not torch.cuda.is_available(): + print("CUDA is required.") + sys.exit(1) + + configs = [ + {"B": 1, "H": 4, "L": 256, "D": 128, "name": "H4_L256_D128"}, + {"B": 1, "H": 8, "L": 512, "D": 128, "name": "H8_L512_D128"}, + {"B": 1, "H": 8, "L": 256, "D": 256, "name": "H8_L256_D256"}, + {"B": 1, "H": 16, "L": 384, "D": 128, "name": "H16_L384_D128"}, + ] + + fwd_cos_min = 0.98 + fwd_rel_l2_max = 2.0e-1 # relax for FP4 probabilities + + for case in configs: + run_case(case, fwd_cos_min=fwd_cos_min, fwd_rel_l2_max=fwd_rel_l2_max) + torch.cuda.synchronize() + + print("\n[OK] All forward diff checks passed for all cases.") + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-I2V/FLASH_ATTN/A woman sits at a wooden table by the window in a cozy café. She reaches out with her right hand, pi.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-I2V/FLASH_ATTN/A woman sits at a wooden table by the window in a cozy café. She reaches out with her right hand, pi.mp4 new file mode 100644 index 000000000..f17abf832 Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-I2V/FLASH_ATTN/A woman sits at a wooden table by the window in a cozy café. She reaches out with her right hand, pi.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-I2V/TORCH_SDPA/A woman sits at a wooden table by the window in a cozy café. She reaches out with her right hand, pi.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-I2V/TORCH_SDPA/A woman sits at a wooden table by the window in a cozy café. She reaches out with her right hand, pi.mp4 new file mode 100644 index 000000000..74bb31dfb Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-I2V/TORCH_SDPA/A woman sits at a wooden table by the window in a cozy café. She reaches out with her right hand, pi.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-T2V/FLASH_ATTN/In a realistic photography style, a white boy around seven or eight years old sits on a park bench,.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-T2V/FLASH_ATTN/In a realistic photography style, a white boy around seven or eight years old sits on a park bench,.mp4 new file mode 100644 index 000000000..e6a2ec0df Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-T2V/FLASH_ATTN/In a realistic photography style, a white boy around seven or eight years old sits on a park bench,.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-T2V/TORCH_SDPA/In a realistic photography style, a white boy around seven or eight years old sits on a park bench,.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-T2V/TORCH_SDPA/In a realistic photography style, a white boy around seven or eight years old sits on a park bench,.mp4 new file mode 100644 index 000000000..d4403e860 Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-T2V/TORCH_SDPA/In a realistic photography style, a white boy around seven or eight years old sits on a park bench,.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-VC/FLASH_ATTN/A person rides a motorcycle along a long, straight road that stretches between a body of water and a.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-VC/FLASH_ATTN/A person rides a motorcycle along a long, straight road that stretches between a body of water and a.mp4 new file mode 100644 index 000000000..eac144587 Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-VC/FLASH_ATTN/A person rides a motorcycle along a long, straight road that stretches between a body of water and a.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-VC/TORCH_SDPA/A person rides a motorcycle along a long, straight road that stretches between a body of water and a.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-VC/TORCH_SDPA/A person rides a motorcycle along a long, straight road that stretches between a body of water and a.mp4 new file mode 100644 index 000000000..9386e7050 Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/LongCat-Video-VC/TORCH_SDPA/A person rides a motorcycle along a long, straight road that stretches between a body of water and a.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/Matrix-Game-2.0-Diffusers-Base/FLASH_ATTN/video.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/Matrix-Game-2.0-Diffusers-Base/FLASH_ATTN/video.mp4 new file mode 100644 index 000000000..deca1faad Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/Matrix-Game-2.0-Diffusers-Base/FLASH_ATTN/video.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 new file mode 100644 index 000000000..83eede69b Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of_1.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of_1.mp4 new file mode 100644 index 000000000..83eede69b Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of_1.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.2-I2V-A14B-Diffusers/SLA_ATTN/An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space reali.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.2-I2V-A14B-Diffusers/SLA_ATTN/An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space reali.mp4 new file mode 100644 index 000000000..b12da4d45 Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.2-I2V-A14B-Diffusers/SLA_ATTN/An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space reali.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.2-I2V-A14B-Diffusers/SLA_ATTN/steps4_An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space reali_ssim.json b/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.2-I2V-A14B-Diffusers/SLA_ATTN/steps4_An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space reali_ssim.json new file mode 100644 index 000000000..db690f38d --- /dev/null +++ b/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.2-I2V-A14B-Diffusers/SLA_ATTN/steps4_An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space reali_ssim.json @@ -0,0 +1,11 @@ +{ + "mean_ssim": 0.9805062346988254, + "min_ssim": 0.9636241793632507, + "max_ssim": 0.9951534271240234, + "reference_video": "/mnt/fast-disks/hao_lab/loay/FastVideo/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.2-I2V-A14B-Diffusers/SLA_ATTN/An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space reali.mp4", + "generated_video": "/mnt/fast-disks/hao_lab/loay/FastVideo/fastvideo/tests/ssim/generated_videos/TurboWan2.2-I2V-A14B-Diffusers/SLA_ATTN/An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space reali.mp4", + "parameters": { + "num_inference_steps": 4, + "prompt": "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot." + } +} \ No newline at end of file diff --git a/fastvideo/tests/ssim/L40S_reference_videos/ltx2_diffusers/FLASH_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/ltx2_diffusers/FLASH_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 new file mode 100644 index 000000000..7acf878af Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/ltx2_diffusers/FLASH_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 differ diff --git a/fastvideo/tests/ssim/L40S_reference_videos/ltx2_diffusers/TORCH_SDPA/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 b/fastvideo/tests/ssim/L40S_reference_videos/ltx2_diffusers/TORCH_SDPA/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 new file mode 100644 index 000000000..545fb2d69 Binary files /dev/null and b/fastvideo/tests/ssim/L40S_reference_videos/ltx2_diffusers/TORCH_SDPA/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4 differ diff --git a/fastvideo/training/distillation_pipeline.py b/fastvideo/training/distillation_pipeline.py index 319f89e4f..2c6742369 100644 --- a/fastvideo/training/distillation_pipeline.py +++ b/fastvideo/training/distillation_pipeline.py @@ -1498,4 +1498,4 @@ class DistillationPipeline(TrainingPipeline): logger.info("Profiler stopped.") if get_sp_group(): - cleanup_dist_env_and_memory() + cleanup_dist_env_and_memory() \ No newline at end of file diff --git a/fastvideo/training/training_pipeline.py b/fastvideo/training/training_pipeline.py index cd8f7e0b0..7f0279e41 100644 --- a/fastvideo/training/training_pipeline.py +++ b/fastvideo/training/training_pipeline.py @@ -37,6 +37,7 @@ from fastvideo.forward_context import set_forward_context from fastvideo.logger import init_logger from fastvideo.pipelines import (ComposedPipelineBase, ForwardBatch, LoRAPipeline, TrainingBatch) from fastvideo.platforms import current_platform +from training_utils import traverse_swap_module from fastvideo.training.activation_checkpoint import (apply_activation_checkpointing) from fastvideo.training.trackers import (DummyTracker, TrackerType, initialize_trackers, Trackers) from fastvideo.training.training_utils import (clip_grad_norm_while_handling_failing_dtensor_cases, @@ -54,6 +55,116 @@ except Exception: logger = init_logger(__name__) +def swap_sage_attn3(obj: Any, obj_path: str) -> int: + """ + If `obj` has an attribute named `attn_impl`, replace it with SageAttention3Impl, + carrying over `.causal` and `.softmax_scale`. Everything else set to None. + Returns number of swaps performed (0 or 1). + """ + # Quick reject: no attribute + if not hasattr(obj, "attn_impl"): + return 0 + + # Access may raise if it's a property; guard it. + try: + old_impl = getattr(obj, "attn_impl") + except Exception: + return 0 + + # Already None or already of the desired type => skip + if old_impl is None or isinstance(old_impl, SageAttention3Impl): + return 0 + + # Extract fields with sensible defaults + causal = getattr(old_impl, "causal", False) + softmax_scale = getattr(old_impl, "softmax_scale", 1.0) + + + new_impl = SageAttention3Impl( + num_heads=None, + head_size=None, + causal=causal, + softmax_scale=softmax_scale, + num_kv_heads=None, + prefix="", + ) + setattr(obj, "attn_impl", new_impl) + return 1 + + +def swap_fp4_linear(obj: Any, obj_path: str) -> int: + """ + Replace torch.nn.Linear and ReplicatedLinear layers with LinearFWD4BWD16, + preserving the input/output dimensions and bias setting. + Returns number of swaps performed. + """ + from fastvideo.layers.fp4linear import fp4_linear_forward + from fastvideo.layers.linear import ReplicatedLinear + + swaps_performed = 0 + + # Find all Linear-like attributes in the object + for attr_name in dir(obj): + if attr_name.startswith('_'): + continue + + try: + attr_value = getattr(obj, attr_name) + except Exception: + continue + + # Check if it's a linear layer we want to replace + should_replace = False + in_features = None + out_features = None + bias = None + + # if isinstance(attr_value, torch.nn.Linear): + # should_replace = True + # in_features = attr_value.in_features + # out_features = attr_value.out_features + # bias = attr_value.bias is not None + + if isinstance(attr_value, ReplicatedLinear): + should_replace = True + in_features = attr_value.input_size + out_features = attr_value.output_size + # For ReplicatedLinear, check if bias exists + bias = hasattr(attr_value, 'bias') and attr_value.bias is not None + + + if should_replace and in_features is not None and out_features is not None: + # Create new LinearFWD4BWD16 layer with same dimensions + # new_layer = LinearFWD4BWD16( + # in_features=in_features, + # out_features=out_features, + # bias=bias, + # backend="cutlass", # Default backend + # block_size=16, # Default block size + # use_128x4_sf_layout=True # Default layout + # ) + + # Copy weights if possible + # if hasattr(attr_value, 'weight') and attr_value.weight is not None: + # with torch.no_grad(): + # new_layer.weight.copy_(attr_value.weight.data.float()) + + # if bias and hasattr(attr_value, 'bias') and attr_value.bias is not None: + # with torch.no_grad(): + # new_layer.bias.copy_(attr_value.bias.data.float()) + layer = getattr(obj, attr_name) # Replace the attribute + layer.forward = types.MethodType(fp4_linear_forward, layer) + + # from fpdb import ForkedPdb; ForkedPdb().set_trace() + # setattr(obj, attr_name, new_layer) + # setattr(new_layer, "weight", old_layer.weight) + # setattr(new_layer, "bias", old_layer.bias) + swaps_performed += 1 + + + + + return swaps_performed class TrainingPipeline(LoRAPipeline, ABC): """ A pipeline for training a model. All training pipelines should inherit from this class. @@ -112,8 +223,17 @@ class TrainingPipeline(LoRAPipeline, ABC): self.transformer, checkpointing_type=training_args.enable_gradient_checkpointing_type) if self.transformer_2 is not None: self.transformer_2 = apply_activation_checkpointing( - self.transformer_2, checkpointing_type=training_args.enable_gradient_checkpointing_type) - + self.transformer_2, + checkpointing_type=training_args. + enable_gradient_checkpointing_type) + + + if training_args.generator_4bit_attn: + num_swaps = traverse_swap_module(self.transformer, swap_fn=swap_sage_attn3) + logger.info(f"Swapped {num_swaps} attn_impl to SageAttention3Impl instances in self.transformer") + if training_args.generator_4bit_linear: + num_swaps = traverse_swap_module(self.transformer, swap_fn=swap_fp4_linear) + logger.info(f"Swapped {num_swaps} linear layers to LinearFWD4BWD16 instances in self.transformer") noise_scheduler = self.modules["scheduler"] self.set_trainable() params_to_optimize = self.transformer.parameters() diff --git a/fastvideo/training/training_utils.py b/fastvideo/training/training_utils.py index c539f2608..30e07c826 100644 --- a/fastvideo/training/training_utils.py +++ b/fastvideo/training/training_utils.py @@ -1702,3 +1702,105 @@ class EMA_FSDP: def apply_to_model(self, module: torch.nn.Module) -> _ApplyEMACtx: return EMA_FSDP._ApplyEMACtx(self, module) + + +import types +from typing import Callable, Any, Optional +import torch +import torch.nn as nn + +def traverse_swap_module( + root: Any, + swap_fn: Callable[[Any, str], int], + *, + verbose: bool = False, + max_depth: Optional[int] = None, +) -> int: + """ + Traverse `root` (e.g., your transformer) and apply `swap_fn` to every reachable object. + - Cycle-safe via `visited` set. + - Handles nn.Module submodules, regular attributes, and common containers (list/tuple/set/dict). + + Args: + root: object graph root (e.g., transformer) + swap_fn: function(obj, obj_path) -> int (number of swaps it performed on `obj`) + verbose: print what gets swapped + max_depth: optional limit to traversal depth (0 = only root) + + Returns: + Total number of swaps performed across the graph. + """ + visited = set() + + def is_traversable(x: Any) -> bool: + # Skip obvious non-containers / primitives / callables / modules + return not isinstance( + x, + ( + str, bytes, bytearray, memoryview, + int, float, bool, complex, + types.BuiltinFunctionType, types.FunctionType, types.MethodType, + types.ModuleType, + ), + ) + + def walk(obj: Any, obj_path: str, depth: int) -> int: + if max_depth is not None and depth > max_depth: + return 0 + + oid = id(obj) + if oid in visited: + return 0 + visited.add(oid) + + swaps = 0 + # First, try to swap on the current object + swaps += swap_fn(obj, obj_path) + + + # Then, traverse children + # 1) PyTorch modules: prefer named_children() for proper module graph traversal + if nn is not None and isinstance(obj, nn.Module): + + for name, child in obj.named_children(): + if is_traversable(child): + swaps += walk(child, f"{obj_path}.{name}", depth + 1) + + + # Also traverse non-module attributes living on the module + d = getattr(obj, "__dict__", None) + if isinstance(d, dict): + for k, v in d.items(): + if is_traversable(v): + # Avoid double-visiting known submodules (already covered by named_children) + if not (nn is not None and isinstance(v, nn.Module)): + swaps += walk(v, f"{obj_path}.{k}", depth + 1) + + return swaps + + # 2) Generic Python objects: __dict__ attrs + d = getattr(obj, "__dict__", None) + if isinstance(d, dict): + for k, v in d.items(): + if is_traversable(v): + swaps += walk(v, f"{obj_path}.{k}", depth + 1) + + # 3) Common containers + if isinstance(obj, (list, tuple, set, frozenset)): + for idx, v in enumerate(obj): + if is_traversable(v): + swaps += walk(v, f"{obj_path}[{idx}]", depth + 1) + elif isinstance(obj, dict): + for k, v in obj.items(): + if is_traversable(v): + k_show = repr(k) + if len(k_show) > 40: + k_show = k_show[:37] + "..." + swaps += walk(v, f"{obj_path}[{k_show}]", depth + 1) + + return swaps + + return walk(root, "root", 0) + + + diff --git a/fused_attention.py b/fused_attention.py new file mode 100644 index 000000000..e3d9f1cef --- /dev/null +++ b/fused_attention.py @@ -0,0 +1,763 @@ +""" +Fused Attention +=============== + +This is a Triton implementation of the Flash Attention v2 algorithm from Tri Dao (https://tridao.me/publications/flash2/flash2.pdf) + +Credits: OpenAI kernel team + +Extra Credits: + +* Original flash attention paper (https://arxiv.org/abs/2205.14135) +* Rabe and Staats (https://arxiv.org/pdf/2112.05682v2.pdf) + +""" + +import pytest +import torch +import os + +import triton +import triton.language as tl +from triton.tools.tensor_descriptor import TensorDescriptor + +DEVICE = triton.runtime.driver.active.get_active_torch_device() + + +def is_hip(): + return triton.runtime.driver.active.get_current_target().backend == "hip" + + +def is_cuda(): + return triton.runtime.driver.active.get_current_target().backend == "cuda" + + +def supports_host_descriptor(): + return is_cuda() and torch.cuda.get_device_capability()[0] >= 9 + + +def is_blackwell(): + return is_cuda() and torch.cuda.get_device_capability()[0] == 10 + + +def is_hopper(): + return is_cuda() and torch.cuda.get_device_capability()[0] == 9 + + +@triton.jit +def _attn_fwd_inner(acc, l_i, m_i, q, # + desc_k, desc_v, # + offset_y, dtype: tl.constexpr, start_m, qk_scale, # + BLOCK_M: tl.constexpr, HEAD_DIM: tl.constexpr, BLOCK_N: tl.constexpr, # + STAGE: tl.constexpr, offs_m: tl.constexpr, offs_n: tl.constexpr, # + N_CTX: tl.constexpr, warp_specialize: tl.constexpr, IS_HOPPER: tl.constexpr): + # range of values handled by this stage + if STAGE == 1: + lo, hi = 0, start_m * BLOCK_M + elif STAGE == 2: + lo, hi = start_m * BLOCK_M, (start_m + 1) * BLOCK_M + lo = tl.multiple_of(lo, BLOCK_M) + # causal = False + else: + lo, hi = 0, N_CTX + offsetk_y = offset_y + lo + if dtype == tl.float8e5: + offsetv_y = offset_y * HEAD_DIM + lo + else: + offsetv_y = offset_y + lo + # loop over k, v and update accumulator + for start_n in tl.range(lo, hi, BLOCK_N, warp_specialize=warp_specialize): + start_n = tl.multiple_of(start_n, BLOCK_N) + # -- compute qk ---- + k = desc_k.load([offsetk_y, 0]).T + qk = tl.dot(q, k) + if STAGE == 2: + mask = offs_m[:, None] >= (start_n + offs_n[None, :]) + qk = qk * qk_scale + tl.where(mask, 0, -1.0e6) + m_ij = tl.maximum(m_i, tl.max(qk, 1)) + qk -= m_ij[:, None] + else: + m_ij = tl.maximum(m_i, tl.max(qk, 1) * qk_scale) + qk = qk * qk_scale - m_ij[:, None] + p = tl.math.exp2(qk) + # -- compute correction factor + alpha = tl.math.exp2(m_i - m_ij) + l_ij = tl.sum(p, 1) + # -- update output accumulator -- + if not IS_HOPPER and warp_specialize and BLOCK_M == 128 and HEAD_DIM == 128: + BM: tl.constexpr = acc.shape[0] + BN: tl.constexpr = acc.shape[1] + acc0, acc1 = acc.reshape([BM, 2, BN // 2]).permute(0, 2, 1).split() + acc0 = acc0 * alpha[:, None] + acc1 = acc1 * alpha[:, None] + acc = tl.join(acc0, acc1).permute(0, 2, 1).reshape([BM, BN]) + else: + acc = acc * alpha[:, None] + # prepare p and v for the dot + if dtype == tl.float8e5: + v = desc_v.load([0, offsetv_y]).T + else: + v = desc_v.load([offsetv_y, 0]) + p = p.to(dtype) + # note that this non transposed v for FP8 is only supported on Blackwell + acc = tl.dot(p, v.to(dtype), acc) + # update m_i and l_i + # place this at the end of the loop to reduce register pressure + l_i = l_i * alpha + l_ij + m_i = m_ij + offsetk_y += BLOCK_N + offsetv_y += BLOCK_N + return acc, l_i, m_i + + +def _host_descriptor_pre_hook(nargs): + BLOCK_M = nargs["BLOCK_M"] + BLOCK_N = nargs["BLOCK_N"] + HEAD_DIM = nargs["HEAD_DIM"] + if not isinstance(nargs["desc_q"], TensorDescriptor): + return + nargs["desc_q"].block_shape = [BLOCK_M, HEAD_DIM] + if nargs["FP8_OUTPUT"]: + nargs["desc_v"].block_shape = [HEAD_DIM, BLOCK_N] + else: + nargs["desc_v"].block_shape = [BLOCK_N, HEAD_DIM] + nargs["desc_k"].block_shape = [BLOCK_N, HEAD_DIM] + nargs["desc_o"].block_shape = [BLOCK_M, HEAD_DIM] + + +if is_hip(): + NUM_STAGES_OPTIONS = [1] +elif supports_host_descriptor(): + NUM_STAGES_OPTIONS = [2, 3, 4] +else: + NUM_STAGES_OPTIONS = [2, 3, 4] + +configs = [ + triton.Config({'BLOCK_M': BM, 'BLOCK_N': BN}, num_stages=s, num_warps=w, pre_hook=_host_descriptor_pre_hook) \ + for BM in [64, 128]\ + for BN in [32, 64, 128]\ + for s in NUM_STAGES_OPTIONS \ + for w in [4, 8]\ +] +if "PYTEST_VERSION" in os.environ: + # Use a single config in testing for reproducibility + configs = [ + triton.Config(dict(BLOCK_M=128, BLOCK_N=64), num_stages=2, num_warps=4, pre_hook=_host_descriptor_pre_hook), + ] + + +def keep(conf): + BLOCK_M = conf.kwargs["BLOCK_M"] + BLOCK_N = conf.kwargs["BLOCK_N"] + return not (is_cuda() and torch.cuda.get_device_capability()[0] == 9 and BLOCK_M * BLOCK_N < 128 * 128 + and conf.num_warps == 8) + + +def prune_invalid_configs(configs, named_args, **kwargs): + N_CTX = kwargs["N_CTX"] + + # Filter out configs where BLOCK_M > N_CTX + return [conf for conf in configs if conf.kwargs.get("BLOCK_M", 0) <= N_CTX] + + +@triton.jit +def _maybe_make_tensor_desc(desc_or_ptr, shape, strides, block_shape): + if isinstance(desc_or_ptr, tl.tensor_descriptor): + return desc_or_ptr + else: + return tl.make_tensor_descriptor(desc_or_ptr, shape, strides, block_shape) + + +@triton.autotune(configs=list(filter(keep, configs)), key=["N_CTX", "HEAD_DIM", "FP8_OUTPUT", "warp_specialize"], + prune_configs_by={'early_config_prune': prune_invalid_configs}) +@triton.jit +def _attn_fwd(sm_scale, M, # + Z, H, desc_q, desc_k, desc_v, desc_o, N_CTX, # + HEAD_DIM: tl.constexpr, # + BLOCK_M: tl.constexpr, # + BLOCK_N: tl.constexpr, # + FP8_OUTPUT: tl.constexpr, # + STAGE: tl.constexpr, # + warp_specialize: tl.constexpr, # + IS_HOPPER: tl.constexpr, # + ): + dtype = tl.float8e5 if FP8_OUTPUT else tl.float16 + tl.static_assert(BLOCK_N <= HEAD_DIM) + start_m = tl.program_id(0) + off_hz = tl.program_id(1) + off_z = off_hz // H + off_h = off_hz % H + + y_dim = Z * H * N_CTX + desc_q = _maybe_make_tensor_desc(desc_q, shape=[y_dim, HEAD_DIM], strides=[HEAD_DIM, 1], + block_shape=[BLOCK_M, HEAD_DIM]) + if FP8_OUTPUT: + desc_v = _maybe_make_tensor_desc(desc_v, shape=[HEAD_DIM, y_dim], strides=[N_CTX, 1], + block_shape=[HEAD_DIM, BLOCK_N]) + else: + desc_v = _maybe_make_tensor_desc(desc_v, shape=[y_dim, HEAD_DIM], strides=[HEAD_DIM, 1], + block_shape=[BLOCK_N, HEAD_DIM]) + desc_k = _maybe_make_tensor_desc(desc_k, shape=[y_dim, HEAD_DIM], strides=[HEAD_DIM, 1], + block_shape=[BLOCK_N, HEAD_DIM]) + desc_o = _maybe_make_tensor_desc(desc_o, shape=[y_dim, HEAD_DIM], strides=[HEAD_DIM, 1], + block_shape=[BLOCK_M, HEAD_DIM]) + + offset_y = off_z * (N_CTX * H) + off_h * N_CTX + qo_offset_y = offset_y + start_m * BLOCK_M + # initialize offsets + offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M) + offs_n = tl.arange(0, BLOCK_N) + # initialize pointer to m and l + m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf") + l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0 + acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32) + # load scales + qk_scale = sm_scale + qk_scale *= 1.44269504 # 1/log(2) + # load q: it will stay in SRAM throughout + q = desc_q.load([qo_offset_y, 0]) + # stage 1: off-band + # For causal = True, STAGE = 3 and _attn_fwd_inner gets 1 as its STAGE + # For causal = False, STAGE = 1, and _attn_fwd_inner gets 3 as its STAGE + if STAGE & 1: + acc, l_i, m_i = _attn_fwd_inner(acc, l_i, m_i, q, # + desc_k, desc_v, # + offset_y, dtype, start_m, qk_scale, # + BLOCK_M, HEAD_DIM, BLOCK_N, # + 4 - STAGE, offs_m, offs_n, N_CTX, # + warp_specialize, IS_HOPPER) + # stage 2: on-band + if STAGE & 2: + acc, l_i, m_i = _attn_fwd_inner(acc, l_i, m_i, q, # + desc_k, desc_v, # + offset_y, dtype, start_m, qk_scale, # + BLOCK_M, HEAD_DIM, BLOCK_N, # + 2, offs_m, offs_n, N_CTX, # + warp_specialize, IS_HOPPER) + # epilogue + m_i += tl.math.log2(l_i) + acc = acc / l_i[:, None] + m_ptrs = M + off_hz * N_CTX + offs_m + tl.store(m_ptrs, m_i) + desc_o.store([qo_offset_y, 0], acc.to(dtype)) + + +@triton.jit +def _attn_bwd_preprocess(O, DO, # + Delta, # + Z, H, N_CTX, # + BLOCK_M: tl.constexpr, HEAD_DIM: tl.constexpr # + ): + off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M) + off_hz = tl.program_id(1) + off_n = tl.arange(0, HEAD_DIM) + # load + o = tl.load(O + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :]) + do = tl.load(DO + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :]).to(tl.float32) + delta = tl.sum(o * do, axis=1) + # write-back + tl.store(Delta + off_hz * N_CTX + off_m, delta) + + +# The main inner-loop logic for computing dK and dV. +@triton.jit +def _attn_bwd_dkdv(dk, dv, # + Q, k, v, sm_scale, # + DO, # + M, D, # + # shared by Q/K/V/DO. + stride_tok, stride_d, # + H, N_CTX, BLOCK_M1: tl.constexpr, # + BLOCK_N1: tl.constexpr, # + HEAD_DIM: tl.constexpr, # + # Filled in by the wrapper. + start_n, start_m, num_steps, # + MASK: tl.constexpr): + offs_m = start_m + tl.arange(0, BLOCK_M1) + offs_n = start_n + tl.arange(0, BLOCK_N1) + offs_k = tl.arange(0, HEAD_DIM) + qT_ptrs = Q + offs_m[None, :] * stride_tok + offs_k[:, None] * stride_d + do_ptrs = DO + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d + # BLOCK_N1 must be a multiple of BLOCK_M1, otherwise the code wouldn't work. + tl.static_assert(BLOCK_N1 % BLOCK_M1 == 0) + curr_m = start_m + step_m = BLOCK_M1 + for blk_idx in range(num_steps): + qT = tl.load(qT_ptrs) + # Load m before computing qk to reduce pipeline stall. + offs_m = curr_m + tl.arange(0, BLOCK_M1) + m = tl.load(M + offs_m) + qkT = tl.dot(k, qT) + pT = tl.math.exp2(qkT - m[None, :]) + # Autoregressive masking. + if MASK: + mask = (offs_m[None, :] >= offs_n[:, None]) + pT = tl.where(mask, pT, 0.0) + do = tl.load(do_ptrs) + # Compute dV. + ppT = pT + ppT = ppT.to(tl.float16) + dv += tl.dot(ppT, do.to(tl.float16)) + # D (= delta) is pre-divided by ds_scale. + Di = tl.load(D + offs_m) + # Compute dP and dS. + dpT = tl.dot(v, tl.trans(do)).to(tl.float32) + dsT = pT * (dpT - Di[None, :]) + dsT = dsT.to(tl.float16) + dk += tl.dot(dsT, tl.trans(qT.to(tl.float16))) + # Increment pointers. + curr_m += step_m + qT_ptrs += step_m * stride_tok + do_ptrs += step_m * stride_tok + return dk, dv + + +# the main inner-loop logic for computing dQ +@triton.jit +def _attn_bwd_dq(dq, q, K, V, # + do, m, D, + # shared by Q/K/V/DO. + stride_tok, stride_d, # + H, N_CTX, # + BLOCK_M2: tl.constexpr, # + BLOCK_N2: tl.constexpr, # + HEAD_DIM: tl.constexpr, + # Filled in by the wrapper. + start_m, start_n, num_steps, # + MASK: tl.constexpr): + offs_m = start_m + tl.arange(0, BLOCK_M2) + offs_n = start_n + tl.arange(0, BLOCK_N2) + offs_k = tl.arange(0, HEAD_DIM) + kT_ptrs = K + offs_n[None, :] * stride_tok + offs_k[:, None] * stride_d + vT_ptrs = V + offs_n[None, :] * stride_tok + offs_k[:, None] * stride_d + # D (= delta) is pre-divided by ds_scale. + Di = tl.load(D + offs_m) + # BLOCK_M2 must be a multiple of BLOCK_N2, otherwise the code wouldn't work. + tl.static_assert(BLOCK_M2 % BLOCK_N2 == 0) + curr_n = start_n + step_n = BLOCK_N2 + for blk_idx in range(num_steps): + kT = tl.load(kT_ptrs) + vT = tl.load(vT_ptrs) + qk = tl.dot(q, kT) + p = tl.math.exp2(qk - m) + # Autoregressive masking. + if MASK: + offs_n = curr_n + tl.arange(0, BLOCK_N2) + mask = (offs_m[:, None] >= offs_n[None, :]) + p = tl.where(mask, p, 0.0) + # Compute dP and dS. + dp = tl.dot(do, vT).to(tl.float32) + ds = p * (dp - Di[:, None]) + ds = ds.to(tl.float16) + # Compute dQ. + # NOTE: We need to de-scale dq in the end, because kT was pre-scaled. + dq += tl.dot(ds, tl.trans(kT.to(tl.float16))) + # Increment pointers. + curr_n += step_n + kT_ptrs += step_n * stride_tok + vT_ptrs += step_n * stride_tok + return dq + + +@triton.jit +def _attn_bwd(Q, K, V, sm_scale, # + DO, # + DQ, DK, DV, # + M, D, + # shared by Q/K/V/DO. + stride_z, stride_h, stride_tok, stride_d, # + H, N_CTX, # + BLOCK_M1: tl.constexpr, # + BLOCK_N1: tl.constexpr, # + BLOCK_M2: tl.constexpr, # + BLOCK_N2: tl.constexpr, # + BLK_SLICE_FACTOR: tl.constexpr, # + HEAD_DIM: tl.constexpr): + LN2: tl.constexpr = 0.6931471824645996 # = ln(2) + + bhid = tl.program_id(2) + off_chz = (bhid * N_CTX).to(tl.int64) + adj = (stride_h * (bhid % H) + stride_z * (bhid // H)).to(tl.int64) + pid = tl.program_id(0) + + # offset pointers for batch/head + Q += adj + K += adj + V += adj + DO += adj + DQ += adj + DK += adj + DV += adj + M += off_chz + D += off_chz + + # load scales + offs_k = tl.arange(0, HEAD_DIM) + + start_n = pid * BLOCK_N1 + start_m = start_n + + MASK_BLOCK_M1: tl.constexpr = BLOCK_M1 // BLK_SLICE_FACTOR + offs_n = start_n + tl.arange(0, BLOCK_N1) + + dv = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32) + dk = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32) + + # load K and V: they stay in SRAM throughout the inner loop. + k = tl.load(K + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d) + v = tl.load(V + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d) + + num_steps = BLOCK_N1 // MASK_BLOCK_M1 + + dk, dv = _attn_bwd_dkdv(dk, dv, # + Q, k, v, sm_scale, # + DO, # + M, D, # + stride_tok, stride_d, # + H, N_CTX, # + MASK_BLOCK_M1, BLOCK_N1, HEAD_DIM, # + start_n, start_m, num_steps, # + MASK=True # + ) + + start_m += num_steps * MASK_BLOCK_M1 + num_steps = (N_CTX - start_m) // BLOCK_M1 + + # Compute dK and dV for non-masked blocks. + dk, dv = _attn_bwd_dkdv( # + dk, dv, # + Q, k, v, sm_scale, # + DO, # + M, D, # + stride_tok, stride_d, # + H, N_CTX, # + BLOCK_M1, BLOCK_N1, HEAD_DIM, # + start_n, start_m, num_steps, # + MASK=False # + ) + + dv_ptrs = DV + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d + tl.store(dv_ptrs, dv) + + # Write back dK. + dk *= sm_scale + dk_ptrs = DK + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d + tl.store(dk_ptrs, dk) + + # THIS BLOCK DOES DQ: + start_m = pid * BLOCK_M2 + end_n = start_m + BLOCK_M2 + + MASK_BLOCK_N2: tl.constexpr = BLOCK_N2 // BLK_SLICE_FACTOR + offs_m = start_m + tl.arange(0, BLOCK_M2) + + q = tl.load(Q + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d) + dq = tl.zeros([BLOCK_M2, HEAD_DIM], dtype=tl.float32) + do = tl.load(DO + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d) + + m = tl.load(M + offs_m) + m = m[:, None] + + # Compute dQ for masked (diagonal) blocks. + # NOTE: This code scans each row of QK^T backward (from right to left, + # but inside each call to _attn_bwd_dq, from left to right), but that's + # not due to anything important. I just wanted to reuse the loop + # structure for dK & dV above as much as possible. + num_steps = BLOCK_M2 // MASK_BLOCK_N2 + dq = _attn_bwd_dq(dq, q, K, V, # + do, m, D, # + stride_tok, stride_d, # + H, N_CTX, # + BLOCK_M2, MASK_BLOCK_N2, HEAD_DIM, # + start_m, end_n - num_steps * MASK_BLOCK_N2, num_steps, # + MASK=True # + ) + end_n -= num_steps * MASK_BLOCK_N2 + # stage 2 + num_steps = end_n // BLOCK_N2 + dq = _attn_bwd_dq(dq, q, K, V, # + do, m, D, # + stride_tok, stride_d, # + H, N_CTX, # + BLOCK_M2, BLOCK_N2, HEAD_DIM, # + start_m, end_n - num_steps * BLOCK_N2, num_steps, # + MASK=False # + ) + # Write back dQ. + dq_ptrs = DQ + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d + dq *= LN2 + tl.store(dq_ptrs, dq) + + +class _attention(torch.autograd.Function): + + @staticmethod + def forward(ctx, q, k, v, causal, sm_scale, warp_specialize=True): + # shape constraints + HEAD_DIM_Q, HEAD_DIM_K = q.shape[-1], k.shape[-1] + # when v is in float8_e5m2 it is transposed. + HEAD_DIM_V = v.shape[-1] + assert HEAD_DIM_Q == HEAD_DIM_K and HEAD_DIM_K == HEAD_DIM_V + assert HEAD_DIM_K in {16, 32, 64, 128, 256} + o = torch.empty_like(q) + stage = 3 if causal else 1 + extra_kern_args = {} + # Tuning for AMD target + if is_hip(): + waves_per_eu = 3 if HEAD_DIM_K <= 64 else 2 + extra_kern_args = {"waves_per_eu": waves_per_eu, "allow_flush_denorm": True} + + M = torch.empty((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32) + # Use device_descriptor for Hopper + warpspec. + if supports_host_descriptor() and not (is_hopper() and warp_specialize): + # Note that on Hopper we cannot perform a FP8 dot with a non-transposed second tensor + y_dim = q.shape[0] * q.shape[1] * q.shape[2] + + dummy_block = [1, 1] + desc_q = TensorDescriptor(q, shape=[y_dim, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], block_shape=dummy_block) + if q.dtype == torch.float8_e5m2: + desc_v = TensorDescriptor(v, shape=[HEAD_DIM_K, y_dim], strides=[q.shape[2], 1], + block_shape=dummy_block) + else: + desc_v = TensorDescriptor(v, shape=[y_dim, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], + block_shape=dummy_block) + desc_k = TensorDescriptor(k, shape=[y_dim, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], block_shape=dummy_block) + desc_o = TensorDescriptor(o, shape=[y_dim, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], block_shape=dummy_block) + else: + desc_q = q + desc_v = v + desc_k = k + desc_o = o + + def alloc_fn(size: int, align: int, _): + return torch.empty(size, dtype=torch.int8, device="cuda") + + triton.set_allocator(alloc_fn) + + def grid(META): + return (triton.cdiv(q.shape[2], META["BLOCK_M"]), q.shape[0] * q.shape[1], 1) + + ctx.grid = grid + if is_blackwell() and warp_specialize: + if HEAD_DIM_K == 128 and q.dtype == torch.float16: + extra_kern_args["maxnreg"] = 168 + else: + extra_kern_args["maxnreg"] = 80 + _attn_fwd[grid]( + sm_scale, M, # + q.shape[0], q.shape[1], # + desc_q, desc_k, desc_v, desc_o, # + N_CTX=q.shape[2], # + HEAD_DIM=HEAD_DIM_K, # + FP8_OUTPUT=q.dtype == torch.float8_e5m2, # + STAGE=stage, # + warp_specialize=warp_specialize, # + IS_HOPPER=is_hopper(), # + **extra_kern_args) + + ctx.save_for_backward(q, k, v, o, M) + ctx.sm_scale = sm_scale + ctx.HEAD_DIM = HEAD_DIM_K + ctx.causal = causal + return o + + @staticmethod + def backward(ctx, do): + q, k, v, o, M = ctx.saved_tensors + do = do.contiguous() + assert do.is_contiguous() + assert q.stride() == k.stride() == v.stride() == o.stride() == do.stride() + dq = torch.empty_like(q) + dk = torch.empty_like(k) + dv = torch.empty_like(v) + BATCH, N_HEAD, N_CTX = q.shape[:3] + PRE_BLOCK = 128 + NUM_WARPS, NUM_STAGES = 4, 5 + BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 32, 128, 128, 32 + BLK_SLICE_FACTOR = 2 + RCP_LN2 = 1.4426950408889634 # = 1.0 / ln(2) + arg_k = k + arg_k = arg_k * (ctx.sm_scale * RCP_LN2) + PRE_BLOCK = 128 + assert N_CTX % PRE_BLOCK == 0 + pre_grid = (N_CTX // PRE_BLOCK, BATCH * N_HEAD) + delta = torch.empty_like(M) + _attn_bwd_preprocess[pre_grid]( + o, do, # + delta, # + BATCH, N_HEAD, N_CTX, # + BLOCK_M=PRE_BLOCK, HEAD_DIM=ctx.HEAD_DIM # + ) + grid = (N_CTX // BLOCK_N1, 1, BATCH * N_HEAD) + _attn_bwd[grid]( + q, arg_k, v, ctx.sm_scale, do, dq, dk, dv, # + M, delta, # + q.stride(0), q.stride(1), q.stride(2), q.stride(3), # + N_HEAD, N_CTX, # + BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1, # + BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2, # + BLK_SLICE_FACTOR=BLK_SLICE_FACTOR, # + HEAD_DIM=ctx.HEAD_DIM, # + num_warps=NUM_WARPS, # + num_stages=NUM_STAGES # + ) + + return dq, dk, dv, None, None, None, None + + +attention = _attention.apply + +TORCH_HAS_FP8 = hasattr(torch, 'float8_e5m2') + + +@pytest.mark.parametrize("Z", [1, 4]) +@pytest.mark.parametrize("H", [2, 48]) +@pytest.mark.parametrize("N_CTX", [128, 1024, (2 if is_hip() else 4) * 1024]) +@pytest.mark.parametrize("HEAD_DIM", [64, 128]) +@pytest.mark.parametrize("causal", [True]) # FIXME: Non-causal tests do not pass at the moment. +@pytest.mark.parametrize("warp_specialize", [False, True] if is_blackwell() else [False]) +@pytest.mark.parametrize("mode", ["fwd", "bwd"]) +@pytest.mark.parametrize("provider", ["triton-fp16"] + (["triton-fp8"] if TORCH_HAS_FP8 else [])) +def test_op(Z, H, N_CTX, HEAD_DIM, causal, warp_specialize, mode, provider, dtype=torch.float16): + if mode == "fwd" and "fp16" in provider: + pytest.skip("Avoid running the forward computation twice.") + if mode == "bwd" and "fp8" in provider: + pytest.skip("Backward pass with FP8 is not supported.") + torch.manual_seed(20) + q = (torch.empty((Z, H, N_CTX, HEAD_DIM), dtype=dtype, device=DEVICE).normal_(mean=0.0, std=0.5).requires_grad_()) + k = (torch.empty((Z, H, N_CTX, HEAD_DIM), dtype=dtype, device=DEVICE).normal_(mean=0.0, std=0.5).requires_grad_()) + v = (torch.empty((Z, H, N_CTX, HEAD_DIM), dtype=dtype, device=DEVICE).normal_(mean=0.0, std=0.5).requires_grad_()) + sm_scale = 0.5 + # reference implementation + ref_dtype = dtype + if mode == "fwd" and "fp8" in provider: + ref_dtype = torch.float32 + q = q.to(ref_dtype) + k = k.to(ref_dtype) + v = v.to(ref_dtype) + M = torch.tril(torch.ones((N_CTX, N_CTX), device=DEVICE)) + p = torch.matmul(q, k.transpose(2, 3)) * sm_scale + if causal: + p[:, :, M == 0] = float("-inf") + p = torch.softmax(p.float(), dim=-1) + p = p.to(ref_dtype) + # p = torch.exp(p) + ref_out = torch.matmul(p, v).half() + if mode == "bwd": + dout = torch.randn_like(q) + ref_out.backward(dout) + ref_dv, v.grad = v.grad.clone(), None + ref_dk, k.grad = k.grad.clone(), None + ref_dq, q.grad = q.grad.clone(), None + # triton implementation + if mode == "fwd" and "fp8" in provider: + q = q.to(torch.float8_e5m2) + k = k.to(torch.float8_e5m2) + v = v.permute(0, 1, 3, 2).contiguous() + v = v.permute(0, 1, 3, 2) + v = v.to(torch.float8_e5m2) + tri_out = attention(q, k, v, causal, sm_scale, warp_specialize).half() + if mode == "fwd": + atol = 3 if "fp8" in provider else 1e-2 + torch.testing.assert_close(tri_out, ref_out, atol=atol, rtol=0) + return + tri_out.backward(dout) + tri_dv, v.grad = v.grad.clone(), None + tri_dk, k.grad = k.grad.clone(), None + tri_dq, q.grad = q.grad.clone(), None + # compare + torch.testing.assert_close(tri_out, ref_out, atol=1e-2, rtol=0) + rtol = 0.0 + # Relative tolerance workaround for known hardware limitation of CDNA2 GPU. + # For details see https://pytorch.org/docs/stable/notes/numerical_accuracy.html#reduced-precision-fp16-and-bf16-gemms-and-convolutions-on-amd-instinct-mi200-devices + if torch.version.hip is not None and triton.runtime.driver.active.get_current_target().arch == "gfx90a": + rtol = 1e-2 + torch.testing.assert_close(tri_dv, ref_dv, atol=1e-2, rtol=rtol) + torch.testing.assert_close(tri_dk, ref_dk, atol=1e-2, rtol=rtol) + torch.testing.assert_close(tri_dq, ref_dq, atol=1e-2, rtol=rtol) + + +try: + from flash_attn.flash_attn_interface import \ + flash_attn_qkvpacked_func as flash_attn_func + HAS_FLASH = True +except BaseException: + HAS_FLASH = False + +TORCH_HAS_FP8 = hasattr(torch, 'float8_e5m2') +BATCH, N_HEADS = 4, 32 +# vary seq length for fixed head and batch=4 +configs = [] +for HEAD_DIM in [64, 128]: + for mode in ["fwd", "bwd"]: + for causal in [True, False]: + # Enable warpspec for causal fwd on Hopper + enable_ws = mode == "fwd" and (is_blackwell() or (is_hopper() and not causal)) + for warp_specialize in [False, True] if enable_ws else [False]: + configs.append( + triton.testing.Benchmark( + x_names=["N_CTX"], + x_vals=[2**i for i in range(10, 15)], + line_arg="provider", + line_vals=["triton-fp16"] + (["triton-fp8"] if TORCH_HAS_FP8 else []) + + (["flash"] if HAS_FLASH else []), + line_names=["Triton [FP16]"] + (["Triton [FP8]"] if TORCH_HAS_FP8 else []) + + (["Flash-2"] if HAS_FLASH else []), + styles=[("red", "-"), ("blue", "-"), ("green", "-")], + ylabel="TFLOPS", + plot_name= + f"fused-attention-batch{BATCH}-head{N_HEADS}-d{HEAD_DIM}-{mode}-causal={causal}-warp_specialize={warp_specialize}", + args={ + "H": N_HEADS, + "BATCH": BATCH, + "HEAD_DIM": HEAD_DIM, + "mode": mode, + "causal": causal, + "warp_specialize": warp_specialize, + }, + )) + + +@triton.testing.perf_report(configs) +def bench_flash_attention(BATCH, H, N_CTX, HEAD_DIM, causal, warp_specialize, mode, provider, device=DEVICE): + assert mode in ["fwd", "bwd"] + dtype = torch.float16 + if "triton" in provider: + q = torch.randn((BATCH, H, N_CTX, HEAD_DIM), dtype=dtype, device=device, requires_grad=True) + k = torch.randn((BATCH, H, N_CTX, HEAD_DIM), dtype=dtype, device=device, requires_grad=True) + v = torch.randn((BATCH, H, N_CTX, HEAD_DIM), dtype=dtype, device=device, requires_grad=True) + if mode == "fwd" and "fp8" in provider: + q = q.to(torch.float8_e5m2) + k = k.to(torch.float8_e5m2) + v = v.permute(0, 1, 3, 2).contiguous() + v = v.permute(0, 1, 3, 2) + v = v.to(torch.float8_e5m2) + sm_scale = 1.3 + fn = lambda: attention(q, k, v, causal, sm_scale, warp_specialize) + if mode == "bwd": + o = fn() + do = torch.randn_like(o) + fn = lambda: o.backward(do, retain_graph=True) + ms = triton.testing.do_bench(fn) + + if provider == "flash": + qkv = torch.randn((BATCH, N_CTX, 3, H, HEAD_DIM), dtype=dtype, device=device, requires_grad=True) + fn = lambda: flash_attn_func(qkv, causal=causal) + if mode == "bwd": + o = fn() + do = torch.randn_like(o) + fn = lambda: o.backward(do, retain_graph=True) + ms = triton.testing.do_bench(fn) + flops_per_matmul = 2.0 * BATCH * H * N_CTX * N_CTX * HEAD_DIM + total_flops = 2 * flops_per_matmul + if causal: + total_flops *= 0.5 + if mode == "bwd": + total_flops *= 2.5 # 2.0(bwd) + 0.5(recompute) + return total_flops * 1e-12 / (ms * 1e-3) + + +if __name__ == "__main__": + # only works on post-Ampere GPUs right now + bench_flash_attention.run(save_path=".", print_data=True) diff --git a/logs/wan_t2v_14B_finetune_sage3_1940.err b/logs/wan_t2v_14B_finetune_sage3_1940.err new file mode 100644 index 000000000..8bf3a3f49 --- /dev/null +++ b/logs/wan_t2v_14B_finetune_sage3_1940.err @@ -0,0 +1,4 @@ +[2025-12-18T14:34:37.184] error: execve(): torchrun: No such file or directory +srun: error: hpc-rack-1-8: task 1: Exited with exit code 2 +[2025-12-18T14:34:37.189] error: execve(): torchrun: No such file or directory +srun: error: hpc-rack-1-3: task 0: Exited with exit code 2 diff --git a/logs/wan_t2v_14B_finetune_sage3_1941.err b/logs/wan_t2v_14B_finetune_sage3_1941.err new file mode 100644 index 000000000..14489b5b7 --- /dev/null +++ b/logs/wan_t2v_14B_finetune_sage3_1941.err @@ -0,0 +1,908 @@ +W1218 20:00:23.836000 2210460 torch/distributed/run.py:803] +W1218 20:00:23.836000 2210460 torch/distributed/run.py:803] ***************************************** +W1218 20:00:23.836000 2210460 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W1218 20:00:23.836000 2210460 torch/distributed/run.py:803] ***************************************** +W1218 20:00:23.836000 353305 torch/distributed/run.py:803] +W1218 20:00:23.836000 353305 torch/distributed/run.py:803] ***************************************** +W1218 20:00:23.836000 353305 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W1218 20:00:23.836000 353305 torch/distributed/run.py:803] ***************************************** + Fetching 41 files: 0%| | 0/41 [00:00 +[rank3]: main(args) +[rank3]: ~~~~^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank3]: pipeline.train() +[rank3]: ~~~~~~~~~~~~~~^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train +[rank3]: self._log_validation(self.transformer, self.training_args, +[rank3]: ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: self.init_steps) +[rank3]: ^^^^^^^^^^^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank3]: return func(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation +[rank3]: output_batch = self.validation_pipeline.forward( +[rank3]: batch, training_args) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank3]: return func(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward +[rank3]: batch = stage(batch, fastvideo_args) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ +[rank3]: result = self.forward(batch, fastvideo_args) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward +[rank3]: noise_pred = current_model( +[rank3]: latent_model_input, +[rank3]: ...<4 lines>... +[rank3]: **pos_cond_kwargs, +[rank3]: ) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank3]: return self._call_impl(*args, **kwargs) +[rank3]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank3]: return inner() +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank3]: result = forward_call(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward +[rank3]: hidden_states = block(hidden_states, encoder_hidden_states, +[rank3]: timestep_proj, freqs_cis) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank3]: return self._call_impl(*args, **kwargs) +[rank3]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank3]: return forward_call(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward +[rank3]: return self.checkpoint_fn( # type: ignore[misc] +[rank3]: ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: self._checkpoint_wrapped_module, *args, **kwargs +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: ) +[rank3]: ^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner +[rank3]: return disable_fn(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank3]: return fn(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint +[rank3]: ret = function(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank3]: return self._call_impl(*args, **kwargs) +[rank3]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank3]: return inner() +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank3]: result = forward_call(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward +[rank3]: attn_output, _ = self.attn1(query, key, value) +[rank3]: ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank3]: return self._call_impl(*args, **kwargs) +[rank3]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank3]: return forward_call(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank3]: return fn(*args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward +[rank3]: output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward +[rank3]: output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd +[rank3]: p = sage_attn_qk_torch(q, k).to(q.dtype) +[rank3]: ~~~~~~~~~~~~~~~~~~^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch +[rank3]: qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) +[rank3]: ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +[rank3]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 3 has a total capacity of 184.31 GiB of which 67.28 GiB is free. Including non-PyTorch memory, this process has 117.02 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank1]: Traceback (most recent call last): +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank1]: main(args) +[rank1]: ~~~~^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank1]: pipeline.train() +[rank1]: ~~~~~~~~~~~~~~^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train +[rank1]: self._log_validation(self.transformer, self.training_args, +[rank1]: ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: self.init_steps) +[rank1]: ^^^^^^^^^^^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank1]: return func(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation +[rank1]: output_batch = self.validation_pipeline.forward( +[rank1]: batch, training_args) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank1]: return func(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward +[rank1]: batch = stage(batch, fastvideo_args) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ +[rank1]: result = self.forward(batch, fastvideo_args) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward +[rank1]: noise_pred = current_model( +[rank1]: latent_model_input, +[rank1]: ...<4 lines>... +[rank1]: **pos_cond_kwargs, +[rank1]: ) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank1]: return self._call_impl(*args, **kwargs) +[rank1]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank1]: return inner() +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank1]: result = forward_call(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward +[rank1]: hidden_states = block(hidden_states, encoder_hidden_states, +[rank1]: timestep_proj, freqs_cis) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank1]: return self._call_impl(*args, **kwargs) +[rank1]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank1]: return forward_call(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward +[rank1]: return self.checkpoint_fn( # type: ignore[misc] +[rank1]: ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: self._checkpoint_wrapped_module, *args, **kwargs +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: ) +[rank1]: ^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner +[rank1]: return disable_fn(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank1]: return fn(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint +[rank1]: ret = function(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank1]: return self._call_impl(*args, **kwargs) +[rank1]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank1]: return inner() +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank1]: result = forward_call(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward +[rank1]: attn_output, _ = self.attn1(query, key, value) +[rank1]: ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank1]: return self._call_impl(*args, **kwargs) +[rank1]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank1]: return forward_call(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank1]: return fn(*args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward +[rank1]: output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward +[rank1]: output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd +[rank1]: p = sage_attn_qk_torch(q, k).to(q.dtype) +[rank1]: ~~~~~~~~~~~~~~~~~~^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch +[rank1]: qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) +[rank1]: ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +[rank1]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 1 has a total capacity of 184.31 GiB of which 67.28 GiB is free. Including non-PyTorch memory, this process has 117.02 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) + 0%| | 0/50 [00:00 +[rank4]: main(args) +[rank4]: ~~~~^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank4]: pipeline.train() +[rank4]: ~~~~~~~~~~~~~~^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train +[rank4]: self._log_validation(self.transformer, self.training_args, +[rank4]: ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: self.init_steps) +[rank4]: ^^^^^^^^^^^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank4]: return func(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation +[rank4]: output_batch = self.validation_pipeline.forward( +[rank4]: batch, training_args) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank4]: return func(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward +[rank4]: batch = stage(batch, fastvideo_args) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ +[rank4]: result = self.forward(batch, fastvideo_args) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward +[rank4]: noise_pred = current_model( +[rank4]: latent_model_input, +[rank4]: ...<4 lines>... +[rank4]: **pos_cond_kwargs, +[rank4]: ) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank4]: return self._call_impl(*args, **kwargs) +[rank4]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank4]: return inner() +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank4]: result = forward_call(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward +[rank4]: hidden_states = block(hidden_states, encoder_hidden_states, +[rank4]: timestep_proj, freqs_cis) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank4]: return self._call_impl(*args, **kwargs) +[rank4]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank4]: return forward_call(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward +[rank4]: return self.checkpoint_fn( # type: ignore[misc] +[rank4]: ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: self._checkpoint_wrapped_module, *args, **kwargs +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: ) +[rank4]: ^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner +[rank4]: return disable_fn(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank4]: return fn(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint +[rank4]: ret = function(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank4]: return self._call_impl(*args, **kwargs) +[rank4]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank4]: return inner() +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank4]: result = forward_call(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward +[rank4]: attn_output, _ = self.attn1(query, key, value) +[rank4]: ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank4]: return self._call_impl(*args, **kwargs) +[rank4]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank4]: return forward_call(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank4]: return fn(*args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward +[rank4]: output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward +[rank4]: output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd +[rank4]: p = sage_attn_qk_torch(q, k).to(q.dtype) +[rank4]: ~~~~~~~~~~~~~~~~~~^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch +[rank4]: qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) +[rank4]: ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +[rank4]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 0 has a total capacity of 184.31 GiB of which 67.28 GiB is free. Including non-PyTorch memory, this process has 117.02 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank5]: Traceback (most recent call last): +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank5]: main(args) +[rank5]: ~~~~^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank5]: pipeline.train() +[rank5]: ~~~~~~~~~~~~~~^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train +[rank5]: self._log_validation(self.transformer, self.training_args, +[rank5]: ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: self.init_steps) +[rank5]: ^^^^^^^^^^^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank5]: return func(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation +[rank5]: output_batch = self.validation_pipeline.forward( +[rank5]: batch, training_args) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank5]: return func(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward +[rank5]: batch = stage(batch, fastvideo_args) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ +[rank5]: result = self.forward(batch, fastvideo_args) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward +[rank5]: noise_pred = current_model( +[rank5]: latent_model_input, +[rank5]: ...<4 lines>... +[rank5]: **pos_cond_kwargs, +[rank5]: ) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank5]: return self._call_impl(*args, **kwargs) +[rank5]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank5]: return inner() +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank5]: result = forward_call(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward +[rank5]: hidden_states = block(hidden_states, encoder_hidden_states, +[rank5]: timestep_proj, freqs_cis) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank5]: return self._call_impl(*args, **kwargs) +[rank5]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank5]: return forward_call(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward +[rank5]: return self.checkpoint_fn( # type: ignore[misc] +[rank5]: ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: self._checkpoint_wrapped_module, *args, **kwargs +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: ) +[rank5]: ^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner +[rank5]: return disable_fn(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank5]: return fn(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint +[rank5]: ret = function(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank5]: return self._call_impl(*args, **kwargs) +[rank5]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank5]: return inner() +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank5]: result = forward_call(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward +[rank5]: attn_output, _ = self.attn1(query, key, value) +[rank5]: ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank5]: return self._call_impl(*args, **kwargs) +[rank5]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank5]: return forward_call(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank5]: return fn(*args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward +[rank5]: output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward +[rank5]: output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd +[rank5]: p = sage_attn_qk_torch(q, k).to(q.dtype) +[rank5]: ~~~~~~~~~~~~~~~~~~^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch +[rank5]: qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) +[rank5]: ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +[rank5]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 1 has a total capacity of 184.31 GiB of which 67.28 GiB is free. Including non-PyTorch memory, this process has 117.02 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in + main(args) + ~~~~^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main + pipeline.train() + ~~~~~~~~~~~~~~^^ + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train + self._log_validation(self.transformer, self.training_args, + ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + self.init_steps) + ^^^^^^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context + return func(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation + output_batch = self.validation_pipeline.forward( + batch, training_args) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context + return func(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward + batch = stage(batch, fastvideo_args) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ + result = self.forward(batch, fastvideo_args) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward + noise_pred = current_model( + latent_model_input, + ...<4 lines>... + **pos_cond_kwargs, + ) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl + return self._call_impl(*args, **kwargs) + ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl + return inner() + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner + result = forward_call(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward + hidden_states = block(hidden_states, encoder_hidden_states, + timestep_proj, freqs_cis) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl + return self._call_impl(*args, **kwargs) + ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl + return forward_call(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward + return self.checkpoint_fn( # type: ignore[misc] + ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ + self._checkpoint_wrapped_module, *args, **kwargs + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + ) + ^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner + return disable_fn(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn + return fn(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint + ret = function(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl + return self._call_impl(*args, **kwargs) + ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl + return inner() + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner + result = forward_call(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward + attn_output, _ = self.attn1(query, key, value) + ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl + return self._call_impl(*args, **kwargs) + ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl + return forward_call(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn + return fn(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward + output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward + output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd + p = sage_attn_qk_torch(q, k).to(q.dtype) + ~~~~~~~~~~~~~~~~~~^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch + qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) + ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 0 has a total capacity of 184.31 GiB of which 67.15 GiB is free. Including non-PyTorch memory, this process has 117.14 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank0]: Traceback (most recent call last): +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank0]: main(args) +[rank0]: ~~~~^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank0]: pipeline.train() +[rank0]: ~~~~~~~~~~~~~~^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train +[rank0]: self._log_validation(self.transformer, self.training_args, +[rank0]: ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: self.init_steps) +[rank0]: ^^^^^^^^^^^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank0]: return func(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation +[rank0]: output_batch = self.validation_pipeline.forward( +[rank0]: batch, training_args) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank0]: return func(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward +[rank0]: batch = stage(batch, fastvideo_args) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ +[rank0]: result = self.forward(batch, fastvideo_args) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward +[rank0]: noise_pred = current_model( +[rank0]: latent_model_input, +[rank0]: ...<4 lines>... +[rank0]: **pos_cond_kwargs, +[rank0]: ) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank0]: return self._call_impl(*args, **kwargs) +[rank0]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank0]: return inner() +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank0]: result = forward_call(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward +[rank0]: hidden_states = block(hidden_states, encoder_hidden_states, +[rank0]: timestep_proj, freqs_cis) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank0]: return self._call_impl(*args, **kwargs) +[rank0]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank0]: return forward_call(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward +[rank0]: return self.checkpoint_fn( # type: ignore[misc] +[rank0]: ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: self._checkpoint_wrapped_module, *args, **kwargs +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: ) +[rank0]: ^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner +[rank0]: return disable_fn(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank0]: return fn(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint +[rank0]: ret = function(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank0]: return self._call_impl(*args, **kwargs) +[rank0]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank0]: return inner() +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank0]: result = forward_call(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward +[rank0]: attn_output, _ = self.attn1(query, key, value) +[rank0]: ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank0]: return self._call_impl(*args, **kwargs) +[rank0]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank0]: return forward_call(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank0]: return fn(*args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward +[rank0]: output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward +[rank0]: output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd +[rank0]: p = sage_attn_qk_torch(q, k).to(q.dtype) +[rank0]: ~~~~~~~~~~~~~~~~~~^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch +[rank0]: qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) +[rank0]: ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 0 has a total capacity of 184.31 GiB of which 67.15 GiB is free. Including non-PyTorch memory, this process has 117.14 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank6]: Traceback (most recent call last): +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank6]: main(args) +[rank6]: ~~~~^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank6]: pipeline.train() +[rank6]: ~~~~~~~~~~~~~~^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train +[rank6]: self._log_validation(self.transformer, self.training_args, +[rank6]: ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: self.init_steps) +[rank6]: ^^^^^^^^^^^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank6]: return func(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation +[rank6]: output_batch = self.validation_pipeline.forward( +[rank6]: batch, training_args) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank6]: return func(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward +[rank6]: batch = stage(batch, fastvideo_args) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ +[rank6]: result = self.forward(batch, fastvideo_args) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward +[rank6]: noise_pred = current_model( +[rank6]: latent_model_input, +[rank6]: ...<4 lines>... +[rank6]: **pos_cond_kwargs, +[rank6]: ) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank6]: return self._call_impl(*args, **kwargs) +[rank6]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank6]: return inner() +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank6]: result = forward_call(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward +[rank6]: hidden_states = block(hidden_states, encoder_hidden_states, +[rank6]: timestep_proj, freqs_cis) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank6]: return self._call_impl(*args, **kwargs) +[rank6]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank6]: return forward_call(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward +[rank6]: return self.checkpoint_fn( # type: ignore[misc] +[rank6]: ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: self._checkpoint_wrapped_module, *args, **kwargs +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: ) +[rank6]: ^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner +[rank6]: return disable_fn(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank6]: return fn(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint +[rank6]: ret = function(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank6]: return self._call_impl(*args, **kwargs) +[rank6]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank6]: return inner() +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank6]: result = forward_call(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward +[rank6]: attn_output, _ = self.attn1(query, key, value) +[rank6]: ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank6]: return self._call_impl(*args, **kwargs) +[rank6]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank6]: return forward_call(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank6]: return fn(*args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward +[rank6]: output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward +[rank6]: output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd +[rank6]: p = sage_attn_qk_torch(q, k).to(q.dtype) +[rank6]: ~~~~~~~~~~~~~~~~~~^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch +[rank6]: qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) +[rank6]: ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +[rank6]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 2 has a total capacity of 184.31 GiB of which 67.28 GiB is free. Including non-PyTorch memory, this process has 117.02 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank7]: Traceback (most recent call last): +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank7]: main(args) +[rank7]: ~~~~^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank7]: pipeline.train() +[rank7]: ~~~~~~~~~~~~~~^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train +[rank7]: self._log_validation(self.transformer, self.training_args, +[rank7]: ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: self.init_steps) +[rank7]: ^^^^^^^^^^^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank7]: return func(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation +[rank7]: output_batch = self.validation_pipeline.forward( +[rank7]: batch, training_args) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank7]: return func(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward +[rank7]: batch = stage(batch, fastvideo_args) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ +[rank7]: result = self.forward(batch, fastvideo_args) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward +[rank7]: noise_pred = current_model( +[rank7]: latent_model_input, +[rank7]: ...<4 lines>... +[rank7]: **pos_cond_kwargs, +[rank7]: ) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank7]: return self._call_impl(*args, **kwargs) +[rank7]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank7]: return inner() +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank7]: result = forward_call(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward +[rank7]: hidden_states = block(hidden_states, encoder_hidden_states, +[rank7]: timestep_proj, freqs_cis) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank7]: return self._call_impl(*args, **kwargs) +[rank7]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank7]: return forward_call(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward +[rank7]: return self.checkpoint_fn( # type: ignore[misc] +[rank7]: ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: self._checkpoint_wrapped_module, *args, **kwargs +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: ) +[rank7]: ^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner +[rank7]: return disable_fn(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank7]: return fn(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint +[rank7]: ret = function(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank7]: return self._call_impl(*args, **kwargs) +[rank7]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank7]: return inner() +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank7]: result = forward_call(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward +[rank7]: attn_output, _ = self.attn1(query, key, value) +[rank7]: ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank7]: return self._call_impl(*args, **kwargs) +[rank7]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank7]: return forward_call(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank7]: return fn(*args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward +[rank7]: output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward +[rank7]: output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd +[rank7]: p = sage_attn_qk_torch(q, k).to(q.dtype) +[rank7]: ~~~~~~~~~~~~~~~~~~^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch +[rank7]: qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) +[rank7]: ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +[rank7]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 3 has a total capacity of 184.31 GiB of which 67.28 GiB is free. Including non-PyTorch memory, this process has 117.02 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank2]: Traceback (most recent call last): +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank2]: main(args) +[rank2]: ~~~~^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank2]: pipeline.train() +[rank2]: ~~~~~~~~~~~~~~^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 581, in train +[rank2]: self._log_validation(self.transformer, self.training_args, +[rank2]: ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: self.init_steps) +[rank2]: ^^^^^^^^^^^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank2]: return func(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 791, in _log_validation +[rank2]: output_batch = self.validation_pipeline.forward( +[rank2]: batch, training_args) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context +[rank2]: return func(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/composed_pipeline_base.py", line 358, in forward +[rank2]: batch = stage(batch, fastvideo_args) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/base.py", line 169, in __call__ +[rank2]: result = self.forward(batch, fastvideo_args) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/pipelines/stages/denoising.py", line 391, in forward +[rank2]: noise_pred = current_model( +[rank2]: latent_model_input, +[rank2]: ...<4 lines>... +[rank2]: **pos_cond_kwargs, +[rank2]: ) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank2]: return self._call_impl(*args, **kwargs) +[rank2]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank2]: return inner() +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank2]: result = forward_call(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 729, in forward +[rank2]: hidden_states = block(hidden_states, encoder_hidden_states, +[rank2]: timestep_proj, freqs_cis) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank2]: return self._call_impl(*args, **kwargs) +[rank2]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank2]: return forward_call(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward +[rank2]: return self.checkpoint_fn( # type: ignore[misc] +[rank2]: ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: self._checkpoint_wrapped_module, *args, **kwargs +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: ) +[rank2]: ^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_compile.py", line 53, in inner +[rank2]: return disable_fn(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank2]: return fn(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/utils/checkpoint.py", line 503, in checkpoint +[rank2]: ret = function(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank2]: return self._call_impl(*args, **kwargs) +[rank2]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1881, in _call_impl +[rank2]: return inner() +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1829, in inner +[rank2]: result = forward_call(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/models/dits/wanvideo.py", line 375, in forward +[rank2]: attn_output, _ = self.attn1(query, key, value) +[rank2]: ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl +[rank2]: return self._call_impl(*args, **kwargs) +[rank2]: ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl +[rank2]: return forward_call(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn +[rank2]: return fn(*args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/layer.py", line 117, in forward +[rank2]: output = self.attn_impl.forward(q, k, v, ctx_attn_metadata) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 379, in forward +[rank2]: output = attn_forward_4bit_fwd_16bit_bwd(query, key, value) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 108, in attn_forward_4bit_fwd_16bit_bwd +[rank2]: p = sage_attn_qk_torch(q, k).to(q.dtype) +[rank2]: ~~~~~~~~~~~~~~~~~~^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/attention/backends/sage_attn3.py", line 165, in sage_attn_qk_torch +[rank2]: qk = qk / torch.sqrt(torch.tensor(D, device=q.device, dtype=q.dtype)) +[rank2]: ~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 96.56 GiB. GPU 2 has a total capacity of 184.31 GiB of which 67.28 GiB is free. Including non-PyTorch memory, this process has 117.02 GiB memory in use. Of the allocated memory 107.23 GiB is allocated by PyTorch, and 2.23 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank3]:[W1218 20:07:44.535428492 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W1218 20:07:44.666017699 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W1218 20:07:44.191777715 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W1218 20:07:45.217607691 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W1218 20:07:45.241774202 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W1218 20:07:45.275828693 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W1218 20:07:45.277266674 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank0]:[W1218 20:07:46.978083362 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W1218 20:07:48.393000 353305 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 353380 closing signal SIGTERM +W1218 20:07:48.394000 353305 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 353381 closing signal SIGTERM +W1218 20:07:48.395000 353305 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 353382 closing signal SIGTERM +W1218 20:07:49.110000 2210460 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 2210533 closing signal SIGTERM +W1218 20:07:49.111000 2210460 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 2210534 closing signal SIGTERM +W1218 20:07:49.111000 2210460 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 2210535 closing signal SIGTERM +E1218 20:07:49.311000 353305 torch/distributed/elastic/multiprocessing/api.py:882] failed (exitcode: 1) local_rank: 3 (pid: 353383) of binary: /home/hal-matthewn/FastVideo-Quantization/.venv/bin/python3 +Traceback (most recent call last): + File "/home/hal-matthewn/FastVideo-Quantization/.venv/bin/torchrun", line 10, in + sys.exit(main()) + ~~~~^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper + return f(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/run.py", line 936, in main + run(args) + ~~~^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/run.py", line 927, in run + elastic_launch( + ~~~~~~~~~~~~~~~ + config=config, + ~~~~~~~~~~~~~~ + entrypoint=cmd, + ~~~~~~~~~~~~~~~ + )(*cmd_args) + ~^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/launcher/api.py", line 156, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/launcher/api.py", line 293, in launch_agent + raise ChildFailedError( + ...<2 lines>... + ) +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +fastvideo/training/wan_training_pipeline.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2025-12-18_20:07:48 + host : hpc-rack-1-16.slurm-compute.slurm.svc.cluster.local + rank : 3 (local_rank: 3) + exitcode : 1 (pid: 353383) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ +srun: error: hpc-rack-1-16: task 1: Exited with exit code 1 +E1218 20:07:49.943000 2210460 torch/distributed/elastic/multiprocessing/api.py:882] failed (exitcode: 1) local_rank: 3 (pid: 2210536) of binary: /home/hal-matthewn/FastVideo-Quantization/.venv/bin/python3 +Traceback (most recent call last): + File "/home/hal-matthewn/FastVideo-Quantization/.venv/bin/torchrun", line 10, in + sys.exit(main()) + ~~~~^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper + return f(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/run.py", line 936, in main + run(args) + ~~~^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/run.py", line 927, in run + elastic_launch( + ~~~~~~~~~~~~~~~ + config=config, + ~~~~~~~~~~~~~~ + entrypoint=cmd, + ~~~~~~~~~~~~~~~ + )(*cmd_args) + ~^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/launcher/api.py", line 156, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/launcher/api.py", line 293, in launch_agent + raise ChildFailedError( + ...<2 lines>... + ) +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +fastvideo/training/wan_training_pipeline.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2025-12-18_20:07:49 + host : hpc-rack-1-3.slurm-compute.slurm.svc.cluster.local + rank : 7 (local_rank: 3) + exitcode : 1 (pid: 2210536) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ +srun: error: hpc-rack-1-3: task 0: Exited with exit code 1 diff --git a/logs/wan_t2v_14B_finetune_sage3_2074.err b/logs/wan_t2v_14B_finetune_sage3_2074.err new file mode 100644 index 000000000..4b7c9f718 --- /dev/null +++ b/logs/wan_t2v_14B_finetune_sage3_2074.err @@ -0,0 +1,627 @@ +W0105 20:45:36.871000 2647951 torch/distributed/run.py:803] +W0105 20:45:36.871000 2647951 torch/distributed/run.py:803] ***************************************** +W0105 20:45:36.871000 2647951 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0105 20:45:36.871000 2647951 torch/distributed/run.py:803] ***************************************** +W0105 20:45:36.904000 456319 torch/distributed/run.py:803] +W0105 20:45:36.904000 456319 torch/distributed/run.py:803] ***************************************** +W0105 20:45:36.904000 456319 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0105 20:45:36.904000 456319 torch/distributed/run.py:803] ***************************************** + Fetching 41 files: 0%| | 0/41 [00:00 +[rank3]: main(args) +[rank3]: ~~~~^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank3]: pipeline.train() +[rank3]: ~~~~~~~~~~~~~~^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank3]: return fn(self, *args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train +[rank3]: training_batch = self.train_one_step(training_batch) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank3]: return fn(self, *args, **kwargs) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step +[rank3]: training_batch = self._transformer_forward_and_compute_loss( +[rank3]: training_batch) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss +[rank3]: sharded_pred = shard_latents_across_sp(model_pred) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp +[rank3]: latents = rearrange(latents, +[rank3]: ~~~~~~~~~^^^^^^^^^ +[rank3]: "b c (n s) h w -> b c n s h w", +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: n=sp_world_size).contiguous() +[rank3]: ^^^^^^^^^^^^^^^^ +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange +[rank3]: return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) +[rank3]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce +[rank3]: raise EinopsError(message + "\n {}".format(e)) +[rank3]: einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". +[rank3]: Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. +[rank3]: Shape mismatch, can't divide axis of length 20 in chunks of 8 +[rank2]: Traceback (most recent call last): +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 532, in reduce +[rank2]: return _apply_recipe( +[rank2]: backend, recipe, cast(Tensor, tensor), reduction_type=reduction, axes_lengths=hashable_axes_lengths +[rank2]: ) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 235, in _apply_recipe +[rank2]: init_shapes, axes_reordering, reduced_axes, added_axes, final_shapes, n_axes_w_added = _reconstruct_from_shape( +[rank2]: ~~~~~~~~~~~~~~~~~~~~~~~^ +[rank2]: recipe, backend.shape(tensor), axes_lengths +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: ) +[rank2]: ^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 188, in _reconstruct_from_shape_uncached +[rank2]: raise EinopsError(f"Shape mismatch, can't divide axis of length {length} in chunks of {known_product}") +[rank2]: einops.EinopsError: Shape mismatch, can't divide axis of length 20 in chunks of 8 + +[rank2]: During handling of the above exception, another exception occurred: + +[rank2]: Traceback (most recent call last): +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank2]: main(args) +[rank2]: ~~~~^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank2]: pipeline.train() +[rank2]: ~~~~~~~~~~~~~~^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank2]: return fn(self, *args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train +[rank2]: training_batch = self.train_one_step(training_batch) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank2]: return fn(self, *args, **kwargs) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step +[rank2]: training_batch = self._transformer_forward_and_compute_loss( +[rank2]: training_batch) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss +[rank2]: sharded_pred = shard_latents_across_sp(model_pred) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp +[rank2]: latents = rearrange(latents, +[rank2]: ~~~~~~~~~^^^^^^^^^ +[rank2]: "b c (n s) h w -> b c n s h w", +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: n=sp_world_size).contiguous() +[rank2]: ^^^^^^^^^^^^^^^^ +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange +[rank2]: return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) +[rank2]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce +[rank2]: raise EinopsError(message + "\n {}".format(e)) +[rank2]: einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". +[rank2]: Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. +[rank2]: Shape mismatch, can't divide axis of length 20 in chunks of 8 +[rank1]: Traceback (most recent call last): +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 532, in reduce +[rank1]: return _apply_recipe( +[rank1]: backend, recipe, cast(Tensor, tensor), reduction_type=reduction, axes_lengths=hashable_axes_lengths +[rank1]: ) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 235, in _apply_recipe +[rank1]: init_shapes, axes_reordering, reduced_axes, added_axes, final_shapes, n_axes_w_added = _reconstruct_from_shape( +[rank1]: ~~~~~~~~~~~~~~~~~~~~~~~^ +[rank1]: recipe, backend.shape(tensor), axes_lengths +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: ) +[rank1]: ^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 188, in _reconstruct_from_shape_uncached +[rank1]: raise EinopsError(f"Shape mismatch, can't divide axis of length {length} in chunks of {known_product}") +[rank1]: einops.EinopsError: Shape mismatch, can't divide axis of length 20 in chunks of 8 + +[rank1]: During handling of the above exception, another exception occurred: + +[rank1]: Traceback (most recent call last): +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank1]: main(args) +[rank1]: ~~~~^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank1]: pipeline.train() +[rank1]: ~~~~~~~~~~~~~~^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank1]: return fn(self, *args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train +[rank1]: training_batch = self.train_one_step(training_batch) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank1]: return fn(self, *args, **kwargs) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step +[rank1]: training_batch = self._transformer_forward_and_compute_loss( +[rank1]: training_batch) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss +[rank1]: sharded_pred = shard_latents_across_sp(model_pred) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp +[rank1]: latents = rearrange(latents, +[rank1]: ~~~~~~~~~^^^^^^^^^ +[rank1]: "b c (n s) h w -> b c n s h w", +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: n=sp_world_size).contiguous() +[rank1]: ^^^^^^^^^^^^^^^^ +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange +[rank1]: return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) +[rank1]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce +[rank1]: raise EinopsError(message + "\n {}".format(e)) +[rank1]: einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". +[rank1]: Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. +[rank1]: Shape mismatch, can't divide axis of length 20 in chunks of 8 + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 235, in _apply_recipe + init_shapes, axes_reordering, reduced_axes, added_axes, final_shapes, n_axes_w_added = _reconstruct_from_shape( + ~~~~~~~~~~~~~~~~~~~~~~~^ + recipe, backend.shape(tensor), axes_lengths + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + ) + ^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 188, in _reconstruct_from_shape_uncached + raise EinopsError(f"Shape mismatch, can't divide axis of length {length} in chunks of {known_product}") +einops.EinopsError: Shape mismatch, can't divide axis of length 20 in chunks of 8 + +During handling of the above exception, another exception occurred: + +Traceback (most recent call last): + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in + main(args) + ~~~~^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main + pipeline.train() + ~~~~~~~~~~~~~~^^ + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped + return fn(self, *args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train + training_batch = self.train_one_step(training_batch) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped + return fn(self, *args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step + training_batch = self._transformer_forward_and_compute_loss( + training_batch) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss + sharded_pred = shard_latents_across_sp(model_pred) + File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp + latents = rearrange(latents, + ~~~~~~~~~^^^^^^^^^ + "b c (n s) h w -> b c n s h w", + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + n=sp_world_size).contiguous() + ^^^^^^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange + return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce + raise EinopsError(message + "\n {}".format(e)) +einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". + Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. + Shape mismatch, can't divide axis of length 20 in chunks of 8 +[rank0]: Traceback (most recent call last): +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 532, in reduce +[rank0]: return _apply_recipe( +[rank0]: backend, recipe, cast(Tensor, tensor), reduction_type=reduction, axes_lengths=hashable_axes_lengths +[rank0]: ) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 235, in _apply_recipe +[rank0]: init_shapes, axes_reordering, reduced_axes, added_axes, final_shapes, n_axes_w_added = _reconstruct_from_shape( +[rank0]: ~~~~~~~~~~~~~~~~~~~~~~~^ +[rank0]: recipe, backend.shape(tensor), axes_lengths +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: ) +[rank0]: ^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 188, in _reconstruct_from_shape_uncached +[rank0]: raise EinopsError(f"Shape mismatch, can't divide axis of length {length} in chunks of {known_product}") +[rank0]: einops.EinopsError: Shape mismatch, can't divide axis of length 20 in chunks of 8 + +[rank0]: During handling of the above exception, another exception occurred: + +[rank0]: Traceback (most recent call last): +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank0]: main(args) +[rank0]: ~~~~^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank0]: pipeline.train() +[rank0]: ~~~~~~~~~~~~~~^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank0]: return fn(self, *args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train +[rank0]: training_batch = self.train_one_step(training_batch) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank0]: return fn(self, *args, **kwargs) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step +[rank0]: training_batch = self._transformer_forward_and_compute_loss( +[rank0]: training_batch) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss +[rank0]: sharded_pred = shard_latents_across_sp(model_pred) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp +[rank0]: latents = rearrange(latents, +[rank0]: ~~~~~~~~~^^^^^^^^^ +[rank0]: "b c (n s) h w -> b c n s h w", +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: n=sp_world_size).contiguous() +[rank0]: ^^^^^^^^^^^^^^^^ +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange +[rank0]: return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) +[rank0]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce +[rank0]: raise EinopsError(message + "\n {}".format(e)) +[rank0]: einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". +[rank0]: Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. +[rank0]: Shape mismatch, can't divide axis of length 20 in chunks of 8 + Steps: 0%| | 0/4000 [00:00 +[rank7]: main(args) +[rank7]: ~~~~^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank7]: pipeline.train() +[rank7]: ~~~~~~~~~~~~~~^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank7]: return fn(self, *args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train +[rank7]: training_batch = self.train_one_step(training_batch) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank7]: return fn(self, *args, **kwargs) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step +[rank7]: training_batch = self._transformer_forward_and_compute_loss( +[rank7]: training_batch) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss +[rank7]: sharded_pred = shard_latents_across_sp(model_pred) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp +[rank7]: latents = rearrange(latents, +[rank7]: ~~~~~~~~~^^^^^^^^^ +[rank7]: "b c (n s) h w -> b c n s h w", +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: n=sp_world_size).contiguous() +[rank7]: ^^^^^^^^^^^^^^^^ +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange +[rank7]: return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) +[rank7]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce +[rank7]: raise EinopsError(message + "\n {}".format(e)) +[rank7]: einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". +[rank7]: Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. +[rank7]: Shape mismatch, can't divide axis of length 20 in chunks of 8 +[rank6]: Traceback (most recent call last): +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 532, in reduce +[rank6]: return _apply_recipe( +[rank6]: backend, recipe, cast(Tensor, tensor), reduction_type=reduction, axes_lengths=hashable_axes_lengths +[rank6]: ) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 235, in _apply_recipe +[rank6]: init_shapes, axes_reordering, reduced_axes, added_axes, final_shapes, n_axes_w_added = _reconstruct_from_shape( +[rank6]: ~~~~~~~~~~~~~~~~~~~~~~~^ +[rank6]: recipe, backend.shape(tensor), axes_lengths +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: ) +[rank6]: ^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 188, in _reconstruct_from_shape_uncached +[rank6]: raise EinopsError(f"Shape mismatch, can't divide axis of length {length} in chunks of {known_product}") +[rank6]: einops.EinopsError: Shape mismatch, can't divide axis of length 20 in chunks of 8 + +[rank6]: During handling of the above exception, another exception occurred: + +[rank6]: Traceback (most recent call last): +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank6]: main(args) +[rank6]: ~~~~^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank6]: pipeline.train() +[rank6]: ~~~~~~~~~~~~~~^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank6]: return fn(self, *args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train +[rank6]: training_batch = self.train_one_step(training_batch) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank6]: return fn(self, *args, **kwargs) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step +[rank6]: training_batch = self._transformer_forward_and_compute_loss( +[rank6]: training_batch) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss +[rank6]: sharded_pred = shard_latents_across_sp(model_pred) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp +[rank6]: latents = rearrange(latents, +[rank6]: ~~~~~~~~~^^^^^^^^^ +[rank6]: "b c (n s) h w -> b c n s h w", +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: n=sp_world_size).contiguous() +[rank6]: ^^^^^^^^^^^^^^^^ +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange +[rank6]: return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) +[rank6]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce +[rank6]: raise EinopsError(message + "\n {}".format(e)) +[rank6]: einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". +[rank6]: Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. +[rank6]: Shape mismatch, can't divide axis of length 20 in chunks of 8 +[rank4]: Traceback (most recent call last): +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 532, in reduce +[rank4]: return _apply_recipe( +[rank4]: backend, recipe, cast(Tensor, tensor), reduction_type=reduction, axes_lengths=hashable_axes_lengths +[rank4]: ) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 235, in _apply_recipe +[rank4]: init_shapes, axes_reordering, reduced_axes, added_axes, final_shapes, n_axes_w_added = _reconstruct_from_shape( +[rank4]: ~~~~~~~~~~~~~~~~~~~~~~~^ +[rank4]: recipe, backend.shape(tensor), axes_lengths +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: ) +[rank4]: ^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 188, in _reconstruct_from_shape_uncached +[rank4]: raise EinopsError(f"Shape mismatch, can't divide axis of length {length} in chunks of {known_product}") +[rank4]: einops.EinopsError: Shape mismatch, can't divide axis of length 20 in chunks of 8 + +[rank4]: During handling of the above exception, another exception occurred: + +[rank4]: Traceback (most recent call last): +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank4]: main(args) +[rank4]: ~~~~^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank4]: pipeline.train() +[rank4]: ~~~~~~~~~~~~~~^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank4]: return fn(self, *args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train +[rank4]: training_batch = self.train_one_step(training_batch) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank4]: return fn(self, *args, **kwargs) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step +[rank4]: training_batch = self._transformer_forward_and_compute_loss( +[rank4]: training_batch) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss +[rank4]: sharded_pred = shard_latents_across_sp(model_pred) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp +[rank4]: latents = rearrange(latents, +[rank4]: ~~~~~~~~~^^^^^^^^^ +[rank4]: "b c (n s) h w -> b c n s h w", +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: n=sp_world_size).contiguous() +[rank4]: ^^^^^^^^^^^^^^^^ +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange +[rank4]: return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) +[rank4]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce +[rank4]: raise EinopsError(message + "\n {}".format(e)) +[rank4]: einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". +[rank4]: Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. +[rank4]: Shape mismatch, can't divide axis of length 20 in chunks of 8 +[rank5]: Traceback (most recent call last): +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 532, in reduce +[rank5]: return _apply_recipe( +[rank5]: backend, recipe, cast(Tensor, tensor), reduction_type=reduction, axes_lengths=hashable_axes_lengths +[rank5]: ) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 235, in _apply_recipe +[rank5]: init_shapes, axes_reordering, reduced_axes, added_axes, final_shapes, n_axes_w_added = _reconstruct_from_shape( +[rank5]: ~~~~~~~~~~~~~~~~~~~~~~~^ +[rank5]: recipe, backend.shape(tensor), axes_lengths +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: ) +[rank5]: ^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 188, in _reconstruct_from_shape_uncached +[rank5]: raise EinopsError(f"Shape mismatch, can't divide axis of length {length} in chunks of {known_product}") +[rank5]: einops.EinopsError: Shape mismatch, can't divide axis of length 20 in chunks of 8 + +[rank5]: During handling of the above exception, another exception occurred: + +[rank5]: Traceback (most recent call last): +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 74, in +[rank5]: main(args) +[rank5]: ~~~~^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/wan_training_pipeline.py", line 61, in main +[rank5]: pipeline.train() +[rank5]: ~~~~~~~~~~~~~~^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank5]: return fn(self, *args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 741, in train +[rank5]: training_batch = self.train_one_step(training_batch) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/profiler.py", line 450, in wrapped +[rank5]: return fn(self, *args, **kwargs) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 638, in train_one_step +[rank5]: training_batch = self._transformer_forward_and_compute_loss( +[rank5]: training_batch) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_pipeline.py", line 560, in _transformer_forward_and_compute_loss +[rank5]: sharded_pred = shard_latents_across_sp(model_pred) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/fastvideo/training/training_utils.py", line 870, in shard_latents_across_sp +[rank5]: latents = rearrange(latents, +[rank5]: ~~~~~~~~~^^^^^^^^^ +[rank5]: "b c (n s) h w -> b c n s h w", +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: n=sp_world_size).contiguous() +[rank5]: ^^^^^^^^^^^^^^^^ +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 600, in rearrange +[rank5]: return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) +[rank5]: File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/einops/einops.py", line 542, in reduce +[rank5]: raise EinopsError(message + "\n {}".format(e)) +[rank5]: einops.EinopsError: Error while processing rearrange-reduction pattern "b c (n s) h w -> b c n s h w". +[rank5]: Input tensor shape: torch.Size([1, 16, 20, 56, 104]). Additional info: {'n': 8}. +[rank5]: Shape mismatch, can't divide axis of length 20 in chunks of 8 + Steps: 0%| | 0/4000 [00:29 + sys.exit(main()) + ~~~~^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper + return f(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/run.py", line 936, in main + run(args) + ~~~^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/run.py", line 927, in run + elastic_launch( + ~~~~~~~~~~~~~~~ + config=config, + ~~~~~~~~~~~~~~ + entrypoint=cmd, + ~~~~~~~~~~~~~~~ + )(*cmd_args) + ~^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/launcher/api.py", line 156, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/launcher/api.py", line 293, in launch_agent + raise ChildFailedError( + ...<2 lines>... + ) +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +fastvideo/training/wan_training_pipeline.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2026-01-05_22:30:21 + host : hpc-rack-1-8.slurm-compute.slurm.svc.cluster.local + rank : 5 (local_rank: 1) + exitcode : 1 (pid: 456393) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ +E0105 22:30:22.008000 2647951 torch/distributed/elastic/multiprocessing/api.py:882] failed (exitcode: 1) local_rank: 1 (pid: 2648023) of binary: /home/hal-matthewn/FastVideo-Quantization/.venv/bin/python3 +Traceback (most recent call last): + File "/home/hal-matthewn/FastVideo-Quantization/.venv/bin/torchrun", line 10, in + sys.exit(main()) + ~~~~^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper + return f(*args, **kwargs) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/run.py", line 936, in main + run(args) + ~~~^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/run.py", line 927, in run + elastic_launch( + ~~~~~~~~~~~~~~~ + config=config, + ~~~~~~~~~~~~~~ + entrypoint=cmd, + ~~~~~~~~~~~~~~~ + )(*cmd_args) + ~^^^^^^^^^^^ + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/launcher/api.py", line 156, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + File "/home/hal-matthewn/FastVideo-Quantization/.venv/lib/python3.13/site-packages/torch/distributed/launcher/api.py", line 293, in launch_agent + raise ChildFailedError( + ...<2 lines>... + ) +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +fastvideo/training/wan_training_pipeline.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2026-01-05_22:30:20 + host : hpc-rack-1-13.slurm-compute.slurm.svc.cluster.local + rank : 1 (local_rank: 1) + exitcode : 1 (pid: 2648023) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ +srun: error: hpc-rack-1-13: task 1: Exited with exit code 1 +srun: error: hpc-rack-1-8: task 0: Exited with exit code 1 diff --git a/nvfp4_utils.py b/nvfp4_utils.py new file mode 100644 index 000000000..afefe3742 --- /dev/null +++ b/nvfp4_utils.py @@ -0,0 +1,221 @@ +# SPDX-License-Identifier: Apache-2.0 +# Adapted from https://github.com/triton-lang/triton/blob/main/python/triton_kernels/triton_kernels/numerics_details/mxfp_details/_upcast_from_mxfp.py +# and https://github.com/triton-lang/triton/blob/main/python/triton_kernels/triton_kernels/numerics_details/mxfp_details/_downcast_to_mxfp.py + +import triton +import triton.language as tl +from triton.language.target_info import cuda_capability_geq + +MXFP_BLOCK_SIZE = tl.constexpr(16) + +@triton.jit +def _compute_quant_and_scale(src_tensor, valid_src_mask, use_global_sf=True, two_level_quant_P=False): + BLOCK_SIZE_OUT_DIM: tl.constexpr = src_tensor.shape[0] + BLOCK_SIZE_QUANT_DIM: tl.constexpr = src_tensor.shape[1] + BLOCK_SIZE_QUANT_MX_SCALE: tl.constexpr = src_tensor.shape[1] // MXFP_BLOCK_SIZE + + # Explicit cast to fp32 since most ops are not supported on bfloat16. We avoid needless conversions to and from bf16 + f32_tensor = src_tensor.to(tl.float32) + abs_tensor = tl.abs(f32_tensor) + abs_tensor = tl.where(valid_src_mask, abs_tensor, -1.0) # Don't consider padding tensors in scale computation + + if two_level_quant_P: + # row max from SageAttn3 paper + global_max_val = tl.max(f32_tensor, axis=1, keep_dims=True) # (BLOCK_SIZE_OUT_DIM, 1) + global_max_val = tl.maximum(global_max_val, 1e-8) + s_enc = ((6 * 448) / global_max_val).reshape([BLOCK_SIZE_OUT_DIM, 1, 1]) + s_dec = (1 / s_enc) + + abs_tensor = tl.reshape(abs_tensor, [BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, MXFP_BLOCK_SIZE]) + + if use_global_sf and not two_level_quant_P: + global_max_val = tl.max(abs_tensor) + # Avoid division by zero: if all values are padding (max is 0), use a default scale + global_max_val = tl.maximum(global_max_val, 1e-8) + s_enc = (6 * 448) / global_max_val + s_dec = (1 / s_enc) + elif not two_level_quant_P and not use_global_sf: + s_dec = 1.0 + s_enc = 1.0 + + max_val = tl.max(abs_tensor, axis=2, keep_dims=True) # (BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, 1) # per block maxima + s_dec_b = max_val / 6 # (BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, 1) + s_dec_b_e4m3 = (s_dec_b * s_enc).to(tl.float8e4nv) # (BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, 1) + s_enc_b = 1 / (s_dec_b_e4m3.to(tl.float32) * s_dec) # (BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, 1) + + f32_tensor = tl.reshape(f32_tensor, [BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, MXFP_BLOCK_SIZE]) + quant_tensor = f32_tensor * s_enc_b + + # Reshape the tensors after scaling + quant_tensor = quant_tensor.reshape([BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_DIM]) + # Set the invalid portions of the tensor to 0. This will ensure that any padding tensors are 0 in the mx format. + quant_tensor = tl.where(valid_src_mask, quant_tensor, 0.0) + dequant_scale = s_dec_b_e4m3.reshape([BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE]) + + if cuda_capability_geq(10, 0): + # Convert scaled values to two f32 lanes and use PTX cvt to e2m1x2 with two f32 operands. + pairs = tl.reshape(quant_tensor, [BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_DIM // 2, 2]) + lo_f, hi_f = tl.split(pairs) + lo_f32 = lo_f.to(tl.float32) + hi_f32 = hi_f.to(tl.float32) + + # Inline PTX: cvt.rn.satfinite.e2m1x2.f32 takes two f32 sources and produces one .b8 packed e2m1x2. + out_tensor = tl.inline_asm_elementwise( + """ + { + .reg .b8 r; + cvt.rn.satfinite.e2m1x2.f32 r, $1, $2; + mov.b32 $0, {r, r, r, r}; + } + """, + constraints="=r,f,f", + args=[hi_f32, lo_f32], + dtype=tl.uint8, + is_pure=True, + pack=1, + ) + else: + quant_tensor = quant_tensor.to(tl.uint32, bitcast=True) + signs = quant_tensor & 0x80000000 + exponents = (quant_tensor >> 23) & 0xFF + mantissas_orig = (quant_tensor & 0x7FFFFF) + + # For RTNE: 0.25 < x < 0.75 maps to 0.5 (denormal); exactly 0.25 maps to 0.0 + E8_BIAS = 127 + E2_BIAS = 1 + # Move implicit bit 1 at the beginning to mantissa for denormals + is_subnormal = exponents < E8_BIAS + adjusted_exponents = tl.core.sub(E8_BIAS, exponents + 1, sanitize_overflow=False) + mantissas_pre = (0x400000 | (mantissas_orig >> 1)) + mantissas = tl.where(is_subnormal, mantissas_pre >> adjusted_exponents, mantissas_orig) + + # For normal numbers, we change the bias from 127 to 1, and for subnormals, we keep exponent as 0. + exponents = tl.maximum(exponents, E8_BIAS - E2_BIAS) - (E8_BIAS - E2_BIAS) + + # Combine sign, exponent, and mantissa, while saturating + # Round to nearest, ties to even (RTNE): use guard/sticky and LSB to decide increment + m2bits = mantissas >> 21 + lsb_keep = (m2bits >> 1) & 0x1 + guard = m2bits & 0x1 + IS_SRC_FP32: tl.constexpr = src_tensor.dtype == tl.float32 + if IS_SRC_FP32: + bit0_dropped = (mantissas_orig & 0x1) != 0 + mask = (1 << tl.minimum(adjusted_exponents, 31)) - 1 + dropped_post = (mantissas_pre & mask) != 0 + sticky = is_subnormal & (bit0_dropped | dropped_post) + sticky |= ((mantissas & 0x1FFFFF) != 0).to(tl.uint32) + else: + sticky = ((mantissas & 0x1FFFFF) != 0).to(tl.uint32) + round_inc = guard & (sticky | lsb_keep) + e2m1_tmp = tl.minimum((((exponents << 2) | m2bits) + round_inc) >> 1, 0x7) + e2m1_value = ((signs >> 28) | e2m1_tmp).to(tl.uint8) + + e2m1_value = tl.reshape(e2m1_value, [BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_DIM // 2, 2]) + evens, odds = tl.split(e2m1_value) + out_tensor = evens | (odds << 4) + + return out_tensor, dequant_scale, s_dec + +@triton.jit +def _compute_dequant( + mx_tensor, + scale, + s_dec, + BLOCK_SIZE_OUT_DIM: tl.constexpr, + BLOCK_SIZE_QUANT_DIM: tl.constexpr, + dst_dtype: tl.constexpr, +): + tl.static_assert(BLOCK_SIZE_QUANT_DIM % MXFP_BLOCK_SIZE == 0, f"Block size along quantization block must be a multiple of {MXFP_BLOCK_SIZE=}") + # uint8 signifies two fp4 e2m1 values packed into a single byte + mx_tensor_dtype: tl.constexpr = mx_tensor.dtype + tl.static_assert(dst_dtype == tl.float16 or dst_dtype == tl.bfloat16 or dst_dtype == tl.float32) + tl.static_assert( + mx_tensor_dtype == tl.uint8 + or ((mx_tensor_dtype == tl.float8e4nv or mx_tensor_dtype == tl.float8e5) or mx_tensor_dtype == dst_dtype), + "mx_tensor_ptr must be uint8 or float8 or dst_dtype") + tl.static_assert(scale.dtype == tl.float8e4nv, "scale must be float8e4nv") + + # Determine if we are dealing with fp8 types. + is_fp4: tl.constexpr = mx_tensor_dtype == tl.uint8 + BLOCK_SIZE_QUANT_MX_SCALE: tl.constexpr = BLOCK_SIZE_QUANT_DIM // MXFP_BLOCK_SIZE + + # Upcast the scale to the destination type. + if dst_dtype == tl.bfloat16: + dst_scale = scale.to(tl.bfloat16) + else: + dst_scale = scale.to(tl.float32) + if dst_dtype == tl.float16: + dst_scale = dst_scale.to(tl.float16) + + # Now upcast the tensor. + intermediate_dtype: tl.constexpr = tl.bfloat16 if dst_dtype == tl.float32 else dst_dtype + if cuda_capability_geq(10, 0): + assert is_fp4 + packed_u32 = tl.inline_asm_elementwise( + asm=""" + { + .reg .b8 in_8; + .reg .f16x2 out; + cvt.u8.u32 in_8, $1; + cvt.rn.f16x2.e2m1x2 out, in_8; + mov.b32 $0, out; + } + """, + constraints="=r,r", + args=[mx_tensor], # tl.uint8 passed in as a 32-bit reg with value in low 8 bits + dtype=tl.uint32, + is_pure=True, + pack=1, + ) + lo_u16 = (packed_u32 & 0xFFFF).to(tl.uint16) + hi_u16 = (packed_u32 >> 16).to(tl.uint16) + lo_f16 = lo_u16.to(tl.float16, bitcast=True) + hi_f16 = hi_u16.to(tl.float16, bitcast=True) + + if intermediate_dtype == tl.float16: + x0, x1 = lo_f16, hi_f16 + else: + x0 = lo_f16.to(intermediate_dtype) + x1 = hi_f16.to(intermediate_dtype) + + dst_tensor = tl.interleave(x0, x1) + + else: + assert is_fp4 + dst_bias: tl.constexpr = 127 if intermediate_dtype == tl.bfloat16 else 15 # exponent bias + dst_0p5: tl.constexpr = 16128 if intermediate_dtype == tl.bfloat16 else 0x3800 + dst_m_bits: tl.constexpr = 7 if intermediate_dtype == tl.bfloat16 else 10 # mantissa bits + # e2m1 + em0 = mx_tensor & 0x07 + em1 = mx_tensor & 0x70 + x0 = (em0.to(tl.uint16) << (dst_m_bits - 1)) | ((mx_tensor & 0x08).to(tl.uint16) << 12) + x1 = (em1.to(tl.uint16) << (dst_m_bits - 5)) | ((mx_tensor & 0x80).to(tl.uint16) << 8) + # Three cases: + # 1) x is normal and non-zero: Correct bias + x0 = tl.where((em0 & 0x06) != 0, x0 + ((dst_bias - 1) << dst_m_bits), x0) + x1 = tl.where((em1 & 0x60) != 0, x1 + ((dst_bias - 1) << dst_m_bits), x1) + # 2) x is subnormal (x == 0bs001 where s is the sign): Map to +-0.5 in the dst type + x0 = tl.where(em0 == 0x01, dst_0p5 | (x0 & 0x8000), x0) + x1 = tl.where(em1 == 0x10, dst_0p5 | (x1 & 0x8000), x1) + # 3) x is zero, do nothing + dst_tensor = tl.interleave(x0, x1).to(intermediate_dtype, bitcast=True) + + dst_tensor = dst_tensor.to(dst_dtype) + + # Reshape for proper broadcasting: the scale was stored with a 16‐sized “inner” grouping. + dst_tensor = dst_tensor.reshape([BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, MXFP_BLOCK_SIZE]) + dst_scale = dst_scale.reshape([BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, 1]) + scale = scale.reshape(dst_scale.shape) + + out_tensor = dst_tensor * dst_scale * s_dec # NVFP4 has the additional global scale factor + if dst_dtype == tl.float32: + max_fin = 3.4028234663852886e+38 + elif dst_dtype == tl.bfloat16: + max_fin = 3.3895313892515355e+38 + else: + tl.static_assert(dst_dtype == tl.float16) + max_fin = 65504 + out_tensor = tl.clamp(out_tensor, min=-max_fin, max=max_fin) + out_tensor = out_tensor.reshape([BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_DIM]) + out_tensor = out_tensor.to(dst_dtype) + return out_tensor diff --git a/pyproject.toml b/pyproject.toml index 5479aff4c..33813350d 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -23,7 +23,7 @@ dependencies = [ # Machine Learning & Transformers "transformers==4.57.3", "tokenizers>=0.20.1", - "sentencepiece==0.2.0", + "sentencepiece==0.2.1", "timm==1.0.11", "peft>=0.15.0", "diffusers>=0.33.1", @@ -98,6 +98,7 @@ torchvision = [ { index = "pytorch-cpu", marker = "sys_platform != 'linux'" }, { index = "pytorch-cu128", marker = "sys_platform == 'linux'" }, ] +fastvideo-kernel = { path = "fastvideo-kernel", editable = true } [[tool.uv.index]] name = "pytorch-cpu" diff --git a/qat_attn.py b/qat_attn.py new file mode 100644 index 000000000..0f792e502 --- /dev/null +++ b/qat_attn.py @@ -0,0 +1,1119 @@ +# SPDX-License-Identifier: Apache-2.0 +# Adapted from https://github.com/triton-lang/triton/blob/main/python/tutorials/06-fused-attention.py + +import os + +import torch +import triton +import triton.language as tl +from triton.tools.tensor_descriptor import TensorDescriptor + +from quant_utils import fake_quantize_q, fake_quantize_kv, fake_quantize +# from fastvideo.attention.backends.sageattn.api import triton_group_mean + + +def is_cuda(): + return triton.runtime.driver.active.get_current_target().backend == "cuda" + + +def supports_host_descriptor(): + return is_cuda() and torch.cuda.get_device_capability()[0] >= 9 + + +def is_blackwell(): + return is_cuda() and torch.cuda.get_device_capability()[0] == 10 + + +def is_hopper(): + return is_cuda() and torch.cuda.get_device_capability()[0] == 9 + + +@triton.jit +def _mul_alpha(acc, alpha, BM: tl.constexpr, BN: tl.constexpr): + acc0, acc1 = acc.reshape([BM, 2, BN // 2]).permute(0, 2, 1).split() + acc0 = acc0 * alpha[:, None] + acc1 = acc1 * alpha[:, None] + acc = tl.join(acc0, acc1).permute(0, 2, 1).reshape([BM, BN]) + return acc + + +@triton.jit +def _attn_fwd_inner(acc, high_prec_acc, l_i, m_i, q, q_valid, + desc_k, desc_v, + offset_y, dtype: tl.constexpr, start_m, qk_scale, + BLOCK_M: tl.constexpr, HEAD_DIM: tl.constexpr, BLOCK_N: tl.constexpr, + STAGE: tl.constexpr, offs_m: tl.constexpr, offs_n: tl.constexpr, + N_CTX: tl.constexpr, warp_specialize: tl.constexpr, IS_HOPPER: tl.constexpr, + IS_QAT: tl.constexpr, + fake_quant_P: tl.constexpr = True, + two_level_quant_P: tl.constexpr = False, + use_global_sf_P: tl.constexpr = True): + # range of values handled by this stage (kv blocks) + if STAGE == 1: + lo, hi = 0, start_m * BLOCK_M + elif STAGE == 2: + lo, hi = start_m * BLOCK_M, (start_m + 1) * BLOCK_M + lo = tl.multiple_of(lo, BLOCK_M) + # causal = False + else: + lo, hi = 0, N_CTX + offsetk_y = offset_y + lo # offset from the start of the current batch-head combination + if dtype == tl.float8e5: + offsetv_y = offset_y * HEAD_DIM + lo + else: + offsetv_y = offset_y + lo + # loop over k, v and update accumulator + for start_n in tl.range(lo, hi, BLOCK_N, warp_specialize=warp_specialize): + start_n = tl.multiple_of(start_n, BLOCK_N) + kv_valid = (start_n + offs_n) < N_CTX + # -- compute qk ---- + k = desc_k.load([offsetk_y, 0]) + k = tl.where(kv_valid[:, None], k, 0.0) + qk = tl.dot(q, tl.trans(k)) + if STAGE == 2: + mask = offs_m[:, None] >= (start_n + offs_n[None, :]) + qk = qk * qk_scale + tl.where(mask & kv_valid[None, :], 0, -1.0e6) + m_ij = tl.maximum(m_i, tl.max(qk, 1)) + qk -= m_ij[:, None] + else: + qk = tl.where(kv_valid[None, :], qk, -1.0e6) + m_ij = tl.maximum(m_i, tl.max(qk, 1) * qk_scale) + qk = qk * qk_scale - m_ij[:, None] + p = tl.math.exp2(qk) + if IS_QAT: + p, high_prec_p = fake_quantize( + src_tensor=p, + valid_src_mask=tl.full(shape=p.shape, value=1.0, dtype=p.dtype) == 1.0, + BLOCK_SIZE_OUT_DIM=BLOCK_M, + BLOCK_SIZE_QUANT_DIM=BLOCK_N, + dst_dtype=dtype, + two_level_quant_P=two_level_quant_P, + use_global_sf=use_global_sf_P + ) + l_ij = tl.sum(high_prec_p, 1) + else: + l_ij = tl.sum(p, 1) + # -- compute correction factor + alpha = tl.math.exp2(m_i - m_ij) + # -- update output accumulator -- + if not IS_HOPPER and warp_specialize and BLOCK_M == 128 and HEAD_DIM == 128: + BM: tl.constexpr = acc.shape[0] + BN: tl.constexpr = acc.shape[1] + acc = _mul_alpha(acc, alpha, BM, BN) + if IS_QAT: + high_prec_acc = _mul_alpha(high_prec_acc, alpha, BM, BN) + else: + acc = acc * alpha[:, None] + if IS_QAT: + high_prec_acc = high_prec_acc * alpha[:, None] + # prepare p and v for the dot + if dtype == tl.float8e5: + v = desc_v.load([0, offsetv_y]).T + else: + v = desc_v.load([offsetv_y, 0]) + v = tl.where(kv_valid[:, None], v, 0.0) + p = p.to(dtype) + # note that this non transposed v for FP8 is only supported on Blackwell + acc = tl.dot(p, v.to(dtype), acc) + if IS_QAT: + high_prec_acc = tl.dot(high_prec_p, v, high_prec_acc) + # update m_i and l_i + # place this at the end of the loop to reduce register pressure + l_i = l_i * alpha + l_ij + m_i = m_ij + offsetk_y += BLOCK_N + offsetv_y += BLOCK_N + if IS_QAT: + return acc, high_prec_acc, l_i, m_i + else: + return acc, acc, l_i, m_i + + +def _host_descriptor_pre_hook(nargs): + BLOCK_M = nargs["BLOCK_M"] + BLOCK_N = nargs["BLOCK_N"] + HEAD_DIM = nargs["HEAD_DIM"] + if not isinstance(nargs["desc_q"], TensorDescriptor): + return + nargs["desc_q"].block_shape = [BLOCK_M, HEAD_DIM] + if nargs["FP8_OUTPUT"]: + nargs["desc_v"].block_shape = [HEAD_DIM, BLOCK_N] + else: + nargs["desc_v"].block_shape = [BLOCK_N, HEAD_DIM] + nargs["desc_k"].block_shape = [BLOCK_N, HEAD_DIM] + nargs["desc_o"].block_shape = [1, BLOCK_M, HEAD_DIM] + if "desc_high_prec_o" in nargs and isinstance(nargs["desc_high_prec_o"], TensorDescriptor): + nargs["desc_high_prec_o"].block_shape = [1, BLOCK_M, HEAD_DIM] + + +NUM_STAGES_OPTIONS = [2, 3, 4] + +configs = [ + triton.Config({'BLOCK_M': BM, 'BLOCK_N': BN}, num_stages=s, num_warps=w, pre_hook=_host_descriptor_pre_hook) + for BM in [64, 128] + for BN in [32, 64, 128] + for s in NUM_STAGES_OPTIONS + for w in [4, 8] +] +if "PYTEST_VERSION" in os.environ: + # Use a single config in testing for reproducibility + configs = [ + triton.Config(dict(BLOCK_M=128, BLOCK_N=64), num_stages=2, num_warps=4, pre_hook=_host_descriptor_pre_hook), + ] + + +def keep(conf): + BLOCK_M = conf.kwargs["BLOCK_M"] + BLOCK_N = conf.kwargs["BLOCK_N"] + return not (is_cuda() and torch.cuda.get_device_capability()[0] == 9 and BLOCK_M * BLOCK_N < 128 * 128 + and conf.num_warps == 8) + + +def prune_invalid_configs(configs, named_args, **kwargs): + N_CTX_Q = kwargs["N_CTX_Q"] + N_CTX_KV = kwargs["N_CTX_KV"] + return [ + conf for conf in configs + if conf.kwargs.get("BLOCK_M", 0) <= N_CTX_Q + and conf.kwargs.get("BLOCK_N", 0) <= N_CTX_KV + and conf.kwargs.get("BLOCK_N", 0) % conf.kwargs.get("BLOCK_M", 0) == 0 + ] + + +@triton.jit +def _maybe_make_tensor_desc(desc_or_ptr, shape, strides, block_shape): + if isinstance(desc_or_ptr, tl.tensor_descriptor): + return desc_or_ptr + else: + return tl.make_tensor_descriptor(desc_or_ptr, shape, strides, block_shape) + + +# @triton.autotune(configs=list(filter(keep, configs)), key=["N_CTX_Q", "N_CTX_KV", "HEAD_DIM", "FP8_OUTPUT", "warp_specialize"], +# prune_configs_by={'early_config_prune': prune_invalid_configs}, cache_results=True) +@triton.jit +def _attn_fwd(sm_scale, M, + Z, H, desc_q, desc_k, desc_v, desc_o, desc_high_prec_o, N_CTX_Q, N_CTX_KV, + HEAD_DIM: tl.constexpr, + BLOCK_M: tl.constexpr, + BLOCK_N: tl.constexpr, + FP8_OUTPUT: tl.constexpr, + STAGE: tl.constexpr, + warp_specialize: tl.constexpr, + IS_HOPPER: tl.constexpr, + IS_QAT: tl.constexpr, + fake_quant_P: tl.constexpr = True, + two_level_quant_P: tl.constexpr = False, + use_global_sf_P: tl.constexpr = True, + ): + dtype = tl.float8e5 if FP8_OUTPUT else tl.bfloat16 + tl.static_assert(BLOCK_N <= HEAD_DIM) + start_m = tl.program_id(0) + off_hz = tl.program_id(1) + off_z = off_hz // H # since it's Z then H in the shape + off_h = off_hz % H + + y_dim_q = Z * H * N_CTX_Q + y_dim_kv = Z * H * N_CTX_KV + desc_q = _maybe_make_tensor_desc(desc_q, shape=[y_dim_q, HEAD_DIM], strides=[HEAD_DIM, 1], + block_shape=[BLOCK_M, HEAD_DIM]) + if FP8_OUTPUT: + desc_v = _maybe_make_tensor_desc(desc_v, shape=[HEAD_DIM, y_dim_kv], strides=[N_CTX_KV, 1], + block_shape=[HEAD_DIM, BLOCK_N]) + else: + desc_v = _maybe_make_tensor_desc(desc_v, shape=[y_dim_kv, HEAD_DIM], strides=[HEAD_DIM, 1], + block_shape=[BLOCK_N, HEAD_DIM]) + desc_k = _maybe_make_tensor_desc(desc_k, shape=[y_dim_kv, HEAD_DIM], strides=[HEAD_DIM, 1], + block_shape=[BLOCK_N, HEAD_DIM]) + desc_o = _maybe_make_tensor_desc( + desc_o, shape=[Z * H, N_CTX_Q, HEAD_DIM], + strides=[N_CTX_Q * HEAD_DIM, HEAD_DIM, 1], + block_shape=[1, BLOCK_M, HEAD_DIM] + ) + if IS_QAT: + desc_high_prec_o = _maybe_make_tensor_desc( + desc_high_prec_o, shape=[Z * H, N_CTX_Q, HEAD_DIM], + strides=[N_CTX_Q * HEAD_DIM, HEAD_DIM, 1], + block_shape=[1, BLOCK_M, HEAD_DIM] + ) + + offset_y_q = off_z * (N_CTX_Q * H) + off_h * N_CTX_Q # offset for query tensor + offset_y_kv = off_z * (N_CTX_KV * H) + off_h * N_CTX_KV # offset for key/value tensors + qo_offset_y = offset_y_q + start_m * BLOCK_M + # initialize offsets + offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M) + offs_n = tl.arange(0, BLOCK_N) # for kv blocks + q_valid = offs_m < N_CTX_Q + # initialize pointer to m and l + m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf") + l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0 + acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32) + if IS_QAT: + high_prec_acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32) + else: + # dummy value + high_prec_acc = acc + # load scales + qk_scale = sm_scale + qk_scale *= 1.44269504 # 1/log(2) + # load q: it will stay in SRAM throughout + q = desc_q.load([qo_offset_y, 0]) # load from start of q block and start of the entire head dimension + q = tl.where(q_valid[:, None], q, 0.0) + # stage 1: off-band + # For causal = True, STAGE = 3 and _attn_fwd_inner gets 1 as its STAGE + # For causal = False, STAGE = 1, and _attn_fwd_inner gets 3 as its STAGE + if STAGE & 1: + acc, high_prec_acc, l_i, m_i = _attn_fwd_inner( + acc, high_prec_acc, l_i, m_i, q, q_valid, + desc_k, desc_v, + offset_y_kv, dtype, start_m, qk_scale, + BLOCK_M, HEAD_DIM, BLOCK_N, + 4 - STAGE, offs_m, offs_n, N_CTX_KV, + warp_specialize, IS_HOPPER, IS_QAT, fake_quant_P, two_level_quant_P, use_global_sf_P + ) + # stage 2: on-band + if STAGE & 2: + acc, high_prec_acc, l_i, m_i = _attn_fwd_inner( + acc, high_prec_acc, l_i, m_i, q, q_valid, + desc_k, desc_v, + offset_y_kv, dtype, start_m, qk_scale, + BLOCK_M, HEAD_DIM, BLOCK_N, + 2, offs_m, offs_n, N_CTX_KV, + warp_specialize, IS_HOPPER, IS_QAT, fake_quant_P, two_level_quant_P, use_global_sf_P + ) + # epilogue + m_i += tl.math.log2(l_i) + acc = acc / l_i[:, None] + if IS_QAT: + high_prec_acc = high_prec_acc / l_i[:, None] + m_ptrs = M + off_hz * N_CTX_Q + offs_m + tl.store(m_ptrs, m_i, mask=q_valid) + desc_o.store([off_hz, start_m * BLOCK_M, 0], acc[None, :, :]) + if IS_QAT: + desc_high_prec_o.store([off_hz, start_m * BLOCK_M, 0], high_prec_acc[None, :, :]) + + +@triton.jit +def _attn_bwd_preprocess(O, DO, + Delta, + Z, H, N_CTX, + BLOCK_M: tl.constexpr, HEAD_DIM: tl.constexpr + ): + off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M) + off_hz = tl.program_id(1) + off_n = tl.arange(0, HEAD_DIM) + valid = off_m < N_CTX + # load + o = tl.load(O + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :], mask=valid[:, None]) + do = tl.load(DO + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :], mask=valid[:, None]).to(tl.float32) + delta = tl.sum(o * do, axis=1) + # write-back + tl.store(Delta + off_hz * N_CTX + off_m, delta, mask=valid) + + +# The main inner-loop logic for computing dK and dV. +@triton.jit +def _attn_bwd_dkdv(dk, dv, + Q, k, v, qk_scale, + DO, + M, D, Q_MEAN, + # shared by Q/K/V/DO. + stride_tok, stride_d, + H, N_CTX, BLOCK_M1: tl.constexpr, + BLOCK_N1: tl.constexpr, + HEAD_DIM: tl.constexpr, + # Filled in by the wrapper. + start_n, start_m, num_steps, + MASK: tl.constexpr, + IS_QAT: tl.constexpr, + two_level_quant_P: tl.constexpr = False, + fake_quant_P: tl.constexpr = True, + SMOOTH_Q: tl.constexpr = False, + use_global_sf_P: tl.constexpr = True, + warp_specialize: tl.constexpr = False): + offs_m = start_m + tl.arange(0, BLOCK_M1) + offs_n = start_n + tl.arange(0, BLOCK_N1) + offs_k = tl.arange(0, HEAD_DIM) + q_ptrs = Q + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d + do_ptrs = DO + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d + if SMOOTH_Q: + q_m_ptrs = Q_MEAN + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d + # BLOCK_N1 must be a multiple of BLOCK_M1, otherwise the code wouldn't work. + tl.static_assert(BLOCK_N1 % BLOCK_M1 == 0) + curr_m = start_m + step_m = BLOCK_M1 + for blk_idx in range(num_steps): + offs_m = curr_m + tl.arange(0, BLOCK_M1) + q_valid = offs_m < N_CTX + q = tl.load(q_ptrs, mask=q_valid[:, None]) + if SMOOTH_Q: + q_m = tl.load(q_m_ptrs, mask=q_valid[:, None]) + # Load m before computing qk to reduce pipeline stall. + m = tl.load(M + offs_m, mask=q_valid) + qk = tl.dot(q, tl.trans(k)) + qk = qk * qk_scale + # Autoregressive masking - apply BEFORE exp2 to match forward pass behavior + if MASK: + mask = (offs_m[:, None] >= offs_n[None, :]) + qk = tl.where(mask, qk, -1.0e6) + p = tl.math.exp2(qk - m[:, None]) + do = tl.load(do_ptrs, mask=q_valid[:, None]) + # Compute dV. + p_quant = p + if IS_QAT and fake_quant_P: + p_quant, _ = fake_quantize( + src_tensor=p, + valid_src_mask=tl.full(shape=p.shape, value=1.0, dtype=p.dtype) == 1.0, + BLOCK_SIZE_OUT_DIM=BLOCK_M1, + BLOCK_SIZE_QUANT_DIM=BLOCK_N1, + dst_dtype=p.dtype, + two_level_quant_P=two_level_quant_P, + use_global_sf=use_global_sf_P + ) + dv += tl.dot(tl.trans(p_quant.to(tl.bfloat16)), do) + # D (= delta) is pre-divided by ds_scale. + Di = tl.load(D + offs_m, mask=q_valid) + # Compute dP and dS. + dp = tl.dot(do, tl.trans(v)) + ds = p * (dp - Di[:, None]) + ds = ds.to(tl.bfloat16) + dk += tl.dot(tl.trans(ds), q) + if SMOOTH_Q: + dk += tl.sum(ds, axis=1, keep_dims=True) * q_m + # Increment pointers. + curr_m += step_m + q_ptrs += step_m * stride_tok + do_ptrs += step_m * stride_tok + if SMOOTH_Q: + q_m_ptrs += step_m * stride_tok + return dk, dv + + +# The main inner-loop logic for computing dQ +@triton.jit +def _attn_bwd_dq(dq, q, K, V, + do, m, D, qk_scale, + # shared by Q/K/V/DO. + stride_tok, stride_d, + H, N_CTX, + K_MEAN, + BLOCK_M2: tl.constexpr, + BLOCK_N2: tl.constexpr, + HEAD_DIM: tl.constexpr, + # Filled in by the wrapper. + start_m, start_n, num_steps, + MASK: tl.constexpr, + SMOOTH_K: tl.constexpr, + warp_specialize: tl.constexpr = False): + offs_m = start_m + tl.arange(0, BLOCK_M2) + offs_n = start_n + tl.arange(0, BLOCK_N2) + offs_k = tl.arange(0, HEAD_DIM) + k_ptrs = K + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d + v_ptrs = V + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d + q_valid = offs_m < N_CTX + # D (= delta) is pre-divided by ds_scale. + Di = tl.load(D + offs_m, mask=q_valid) + # BLOCK_M2 must be a multiple of BLOCK_N2, otherwise the code wouldn't work. + tl.static_assert(BLOCK_M2 % BLOCK_N2 == 0) + curr_n = start_n + step_n = BLOCK_N2 + + if SMOOTH_K: + k_m = tl.load(K_MEAN + offs_k) + + for blk_idx in range(num_steps): + # bounds checking for kv block (dynamic) + offs_n = curr_n + tl.arange(0, BLOCK_N2) + kv_valid = offs_n < N_CTX + k = tl.load(k_ptrs, mask=kv_valid[:, None]) + v = tl.load(v_ptrs, mask=kv_valid[:, None]) + + qk = tl.dot(q, tl.trans(k)) + qk = qk * qk_scale + # Autoregressive masking - apply BEFORE exp2 to match forward pass behavior + if MASK: + mask = (offs_m[:, None] >= offs_n[None, :]) + qk = tl.where(mask, qk, -1.0e6) + p = tl.math.exp2(qk - m) + # Compute dP and dS. + dp = tl.dot(do, tl.trans(v)) + ds = p * (dp - Di[:, None]) + # Compute dQ. + # NOTE: We need to de-scale dq in the end, because k was pre-scaled. + dq += tl.dot(ds.to(tl.bfloat16), k) + + if SMOOTH_K: + dq += tl.sum(ds, axis=1, keep_dims=True) * k_m[None, :] + # Increment pointers. + curr_n += step_n + k_ptrs += step_n * stride_tok + v_ptrs += step_n * stride_tok + return dq + + +@triton.jit +def _compute_cross_attn_pointer_offsets(bhid, H, N_CTX_Q, stride_z_q, stride_z_kv, stride_h_q, stride_h_kv): + """Helper function to compute pointer offsets for cross-attention backward kernels.""" + off_chz = (bhid * N_CTX_Q).to(tl.int64) + adj_q = (stride_h_q * (bhid % H) + stride_z_q * (bhid // H)) + adj_kv = (stride_h_kv * (bhid % H) + stride_z_kv * (bhid // H)) + return off_chz, adj_q, adj_kv + + +@triton.jit +def _attn_bwd_dq_cross(Q, K, V, sm_scale, + DO, DQ, + M, D, + stride_z_q, stride_z_kv, stride_h_q, + stride_h_kv, stride_tok_q, stride_tok_kv, stride_d_q, stride_d_kv, + H, N_CTX_Q, N_CTX_KV, + K_MEAN, + BLOCK_M2: tl.constexpr, + BLOCK_N2: tl.constexpr, + HEAD_DIM: tl.constexpr, + SMOOTH_K: tl.constexpr, + warp_specialize: tl.constexpr = False): + # Apply scale AFTER dot product for better precision + RCP_LN2: tl.constexpr = 1.4426950408889634 # = 1.0 / ln(2) + qk_scale = sm_scale * RCP_LN2 + + bhid = tl.program_id(2) + off_chz, adj_q, adj_kv = _compute_cross_attn_pointer_offsets(bhid, H, N_CTX_Q, stride_z_q, stride_z_kv, stride_h_q, stride_h_kv) + + Q += adj_q + K += adj_kv + V += adj_kv + DO += adj_q + DQ += adj_q + M += off_chz + D += off_chz + + pid = tl.program_id(0) + start_m = pid * BLOCK_M2 + offs_m = start_m + tl.arange(0, BLOCK_M2) + offs_k = tl.arange(0, HEAD_DIM) + + q_valid = offs_m < N_CTX_Q + q = tl.load(Q + offs_m[:, None] * stride_tok_q + offs_k[None, :] * stride_d_q, mask=q_valid[:, None]) + dq = tl.zeros([BLOCK_M2, HEAD_DIM], dtype=tl.float32) + do = tl.load(DO + offs_m[:, None] * stride_tok_q + offs_k[None, :] * stride_d_q, mask=q_valid[:, None]) + m = tl.load(M + offs_m, mask=q_valid)[:, None] + + Di = tl.load(D + offs_m, mask=q_valid) + num_steps = (N_CTX_KV + BLOCK_N2 - 1) // BLOCK_N2 + if SMOOTH_K: + k_m = tl.load(K_MEAN + offs_k) + for step in range(num_steps): + start_n = step * BLOCK_N2 + offs_n = start_n + tl.arange(0, BLOCK_N2) + kv_valid = offs_n < N_CTX_KV + + k = tl.load(K + offs_n[:, None] * stride_tok_kv + offs_k[None, :] * stride_d_kv, mask=kv_valid[:, None]) + v = tl.load(V + offs_n[:, None] * stride_tok_kv + offs_k[None, :] * stride_d_kv, mask=kv_valid[:, None]) + + qk = tl.dot(q, tl.trans(k)) + qk = qk * qk_scale + p = tl.math.exp2(qk - m) + + dp = tl.dot(do, tl.trans(v)) + ds = p * (dp - Di[:, None]) + ds = ds.to(tl.bfloat16) + dq += tl.dot(ds, k) + + if SMOOTH_K: + dq += tl.sum(ds, axis=1, keep_dims=True) * k_m[None, :] + + # NOTE: dq is scaled by sm_scale since K is not pre-scaled + dq *= sm_scale + dq_ptrs = DQ + offs_m[:, None] * stride_tok_q + offs_k[None, :] * stride_d_q + tl.store(dq_ptrs, dq, mask=q_valid[:, None]) + + +@triton.jit +def _attn_bwd_dkdv_cross(Q, K, V, sm_scale, + DO, DK, DV, + M, D, Q_MEAN, + stride_z_q, stride_z_kv, stride_h_q, stride_h_kv, + stride_tok_q, stride_tok_kv, stride_d_q, stride_d_kv, + H, N_CTX_Q, N_CTX_KV, + BLOCK_M1: tl.constexpr, + BLOCK_N1: tl.constexpr, + HEAD_DIM: tl.constexpr, + IS_QAT: tl.constexpr, + two_level_quant_P: tl.constexpr = False, + fake_quant_P: tl.constexpr = True, + SMOOTH_Q: tl.constexpr = False, + use_global_sf_P: tl.constexpr = True, + warp_specialize: tl.constexpr = False + ): + # Apply scale AFTER dot product for better precision + RCP_LN2: tl.constexpr = 1.4426950408889634 # = 1.0 / ln(2) + qk_scale = sm_scale * RCP_LN2 + + bhid = tl.program_id(2) + off_chz, adj_q, adj_kv = _compute_cross_attn_pointer_offsets(bhid, H, N_CTX_Q, stride_z_q, stride_z_kv, stride_h_q, stride_h_kv) + + Q += adj_q + K += adj_kv + V += adj_kv + DO += adj_q + DK += adj_kv + DV += adj_kv + M += off_chz + D += off_chz + + if SMOOTH_Q: + Q_MEAN += adj_q + + pid = tl.program_id(0) + start_n = pid * BLOCK_N1 + offs_n = start_n + tl.arange(0, BLOCK_N1) + offs_k = tl.arange(0, HEAD_DIM) + + dk = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32) + dv = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32) + + kv_valid = offs_n < N_CTX_KV + k_block = tl.load(K + offs_n[:, None] * stride_tok_kv + offs_k[None, :] * stride_d_kv, mask=kv_valid[:, None]) + v_block = tl.load(V + offs_n[:, None] * stride_tok_kv + offs_k[None, :] * stride_d_kv, mask=kv_valid[:, None]) + + num_q_steps = (N_CTX_Q + BLOCK_M1 - 1) // BLOCK_M1 + for step in range(num_q_steps): + start_m = step * BLOCK_M1 + offs_m = start_m + tl.arange(0, BLOCK_M1) + q_valid = offs_m < N_CTX_Q + + q = tl.load(Q + offs_m[:, None] * stride_tok_q + offs_k[None, :] * stride_d_q, mask=q_valid[:, None]) + do = tl.load(DO + offs_m[:, None] * stride_tok_q + offs_k[None, :] * stride_d_q, mask=q_valid[:, None]) + m = tl.load(M + offs_m, mask=q_valid) + + qk = tl.dot(q, tl.trans(k_block)) + # Apply scale AFTER dot product (matches forward pass, better precision) + qk = qk * qk_scale + p = tl.math.exp2(qk - m[:, None]) + p_quant = p + if IS_QAT and fake_quant_P: + p_quant, _ = fake_quantize( + src_tensor=p, + valid_src_mask=tl.full(shape=p.shape, value=1.0, dtype=p.dtype) == 1.0, + BLOCK_SIZE_OUT_DIM=BLOCK_M1, + BLOCK_SIZE_QUANT_DIM=BLOCK_N1, + dst_dtype=p.dtype, + two_level_quant_P=two_level_quant_P, + use_global_sf=use_global_sf_P + ) + dv += tl.dot(tl.trans(p_quant.to(tl.bfloat16)), do) + + dp = tl.dot(do, tl.trans(v_block)) + Di = tl.load(D + offs_m, mask=q_valid) + ds = p * (dp - Di[:, None]) + ds = ds.to(tl.bfloat16) + dk += tl.dot(tl.trans(ds), q) + + if SMOOTH_Q: + q_m = tl.load(Q_MEAN + offs_m[:, None] * stride_tok_q + offs_k[None, :] * stride_d_q, mask=q_valid[:, None]) + dk += tl.sum(ds, axis=1, keep_dims=True) * q_m + + dv_ptrs = DV + offs_n[:, None] * stride_tok_kv + offs_k[None, :] * stride_d_kv + tl.store(dv_ptrs, dv, mask=kv_valid[:, None]) + + dk *= sm_scale + dk_ptrs = DK + offs_n[:, None] * stride_tok_kv + offs_k[None, :] * stride_d_kv + tl.store(dk_ptrs, dk, mask=kv_valid[:, None]) + + +@triton.jit +def _attn_bwd(Q, K, V, sm_scale, + DO, + DQ, DK, DV, + M, D, Q_MEAN, + # shared by Q/K/V/DO. + stride_z, stride_h, stride_tok, stride_d, + H, N_CTX, + K_MEAN, + BLOCK_M1: tl.constexpr, + BLOCK_N1: tl.constexpr, + BLOCK_M2: tl.constexpr, + BLOCK_N2: tl.constexpr, + BLK_SLICE_FACTOR: tl.constexpr, + HEAD_DIM: tl.constexpr, + CAUSAL: tl.constexpr, + IS_QAT: tl.constexpr, + SMOOTH_K: tl.constexpr, + two_level_quant_P: tl.constexpr = False, + fake_quant_P: tl.constexpr = True, + SMOOTH_Q: tl.constexpr = False, + use_global_sf_P: tl.constexpr = True, + warp_specialize: tl.constexpr = False): + + bhid = tl.program_id(2) + off_chz = (bhid * N_CTX).to(tl.int64) + adj = (stride_h * (bhid % H) + stride_z * (bhid // H)).to(tl.int64) + pid = tl.program_id(0) + + # offset pointers for batch/head + Q += adj + K += adj + V += adj + DO += adj + DQ += adj + DK += adj + DV += adj + M += off_chz + D += off_chz + + if SMOOTH_Q: + Q_MEAN += adj + + # load scales + offs_k = tl.arange(0, HEAD_DIM) + + start_n = pid * BLOCK_N1 + start_m = start_n + + MASK_BLOCK_M1: tl.constexpr = BLOCK_M1 // BLK_SLICE_FACTOR + offs_n = start_n + tl.arange(0, BLOCK_N1) + kv_valid = offs_n < N_CTX + + dv = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32) + dk = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32) + + # load K and V: they stay in SRAM throughout the inner loop. + k = tl.load(K + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d, mask=kv_valid[:, None]) + v = tl.load(V + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d, mask=kv_valid[:, None]) + + # qk_scale is expressed in base-2 domain because we use exp2/log2. + # Two equivalent ways to form qk_scale: + # - **Post-scale (default)**: qk = dot(q, k^T) * (sm_scale / ln2) + # - **Pre-scale (Flash-aligned)**: k is pre-multiplied by (sm_scale / ln2) and qk_scale = 1 + RCP_LN2: tl.constexpr = 1.4426950408889634 # = 1.0 / ln(2) + qk_scale = sm_scale * RCP_LN2 + + # For causal attention, process diagonal block with masking, then rest without + # For non-causal attention, process all blocks without masking + if CAUSAL: + num_steps = BLOCK_N1 // MASK_BLOCK_M1 + + dk, dv = _attn_bwd_dkdv( + dk, dv, + Q, k, v, qk_scale, + DO, + M, D, Q_MEAN, + stride_tok, stride_d, + H, N_CTX, + MASK_BLOCK_M1, BLOCK_N1, HEAD_DIM, + start_n, start_m, num_steps, + MASK=True, + IS_QAT=IS_QAT, + two_level_quant_P=two_level_quant_P, + fake_quant_P=fake_quant_P, + SMOOTH_Q=SMOOTH_Q, + use_global_sf_P=use_global_sf_P, + warp_specialize=warp_specialize + ) + + start_m += num_steps * MASK_BLOCK_M1 + num_steps = (N_CTX - start_m + BLOCK_M1 - 1) // BLOCK_M1 + else: + # For non-causal, start from 0 and process all Q blocks + start_m = 0 + num_steps = (N_CTX + BLOCK_M1 - 1) // BLOCK_M1 + + # Compute dK and dV for non-masked blocks. + dk, dv = _attn_bwd_dkdv( + dk, dv, + Q, k, v, qk_scale, + DO, + M, D, Q_MEAN, + stride_tok, stride_d, + H, N_CTX, + BLOCK_M1, BLOCK_N1, HEAD_DIM, + start_n, start_m, num_steps, + MASK=False, + IS_QAT=IS_QAT, + two_level_quant_P=two_level_quant_P, + fake_quant_P=fake_quant_P, + SMOOTH_Q=SMOOTH_Q, + use_global_sf_P=use_global_sf_P, + warp_specialize=warp_specialize + ) + + dv_ptrs = DV + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d + tl.store(dv_ptrs, dv, mask=kv_valid[:, None]) + + # Write back dK. + dk *= sm_scale + dk_ptrs = DK + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d + tl.store(dk_ptrs, dk, mask=kv_valid[:, None]) + + # THIS BLOCK DOES DQ: + start_m = pid * BLOCK_M2 + end_n = start_m + BLOCK_M2 + + MASK_BLOCK_N2: tl.constexpr = BLOCK_N2 // BLK_SLICE_FACTOR + offs_m = start_m + tl.arange(0, BLOCK_M2) + q_valid = offs_m < N_CTX + + q = tl.load(Q + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d, mask=q_valid[:, None]) + + dq = tl.zeros([BLOCK_M2, HEAD_DIM], dtype=tl.float32) + do = tl.load(DO + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d, mask=q_valid[:, None]) + + m = tl.load(M + offs_m, mask=q_valid)[:, None] + + # Compute dQ for masked (diagonal) blocks. + # NOTE: This code scans each row of QK^T backward (from right to left, + # but inside each call to _attn_bwd_dq, from left to right), but that's + # not due to anything important. I just wanted to reuse the loop + # structure for dK & dV above as much as possible. + if CAUSAL: + num_steps = BLOCK_M2 // MASK_BLOCK_N2 + dq = _attn_bwd_dq( + dq, q, K, V, + do, m, D, qk_scale, + stride_tok, stride_d, + H, N_CTX, + K_MEAN, + BLOCK_M2, MASK_BLOCK_N2, HEAD_DIM, + start_m, end_n - num_steps * MASK_BLOCK_N2, num_steps, + MASK=True, + SMOOTH_K=SMOOTH_K, + warp_specialize=warp_specialize, + ) + end_n -= num_steps * MASK_BLOCK_N2 + # stage 2: process KV blocks from 0 to end_n (before diagonal), ceiling division to include remainder + num_steps = (end_n + BLOCK_N2 - 1) // BLOCK_N2 + start_n = 0 + else: + # For non-causal, process all KV blocks from 0 to N_CTX (ceiling division to include remainder) + num_steps = (N_CTX + BLOCK_N2 - 1) // BLOCK_N2 + start_n = 0 + # stage 2 + dq = _attn_bwd_dq( + dq, q, K, V, + do, m, D, qk_scale, + stride_tok, stride_d, + H, N_CTX, + K_MEAN, + BLOCK_M2, BLOCK_N2, HEAD_DIM, + start_m, start_n, num_steps, + MASK=False, + SMOOTH_K=SMOOTH_K, + warp_specialize=warp_specialize, + ) + # Write back dQ. + # NOTE: dq is scaled by sm_scale since K is not pre-scaled (unlike original which used LN2) + dq_ptrs = DQ + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d + dq *= sm_scale + tl.store(dq_ptrs, dq, mask=q_valid[:, None]) + + +class _attention(torch.autograd.Function): + + @staticmethod + def forward( + ctx, + q, + k, + v, + causal, + sm_scale, + use_qat_qkv_backward=True, + smooth_k=True, + warp_specialize=True, + IS_QAT=True, + two_level_quant_P=True, + fake_quant_P=True, + use_high_prec_o=False, + smooth_q=False, + use_global_sf_P=True, + use_global_sf_QKV=True, + ): + # shape constraints + HEAD_DIM_Q, HEAD_DIM_K = q.shape[-1], k.shape[-1] + # when v is in float8_e5m2 it is transposed. + HEAD_DIM_V = v.shape[-1] + assert HEAD_DIM_Q == HEAD_DIM_K and HEAD_DIM_K == HEAD_DIM_V + assert HEAD_DIM_K in {16, 32, 64, 128, 256} + + # Support different sequence lengths for q and k/v (needed for cross attention) + N_CTX_Q = q.shape[2] # Query sequence length + N_CTX_KV = k.shape[2] # Key/Value sequence length (may differ from query) + assert k.shape[2] == v.shape[2], "k and v must have the same sequence length" + + # smoothing k from SageAttn + ctx.k_mean = None + if smooth_k: + k_mean = k.mean(dim=(0, 1, 2), keepdim=True) + k = k - k_mean + ctx.k_mean = k_mean.view(-1) + + q_orig, k_orig, v_orig = None, None, None + if not use_qat_qkv_backward: + q_orig, k_orig, v_orig = q, k, v + ctx.q_orig, ctx.k_orig, ctx.v_orig = q_orig, k_orig, v_orig + + o = torch.empty_like(q) + if IS_QAT: + high_prec_o = torch.empty_like(q) + else: + # Initialize to a dummy value + high_prec_o = o + stage = 3 if causal else 1 + extra_kern_args = {} + + M = torch.empty((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32) # (Z, H, N_CTX_Q) + # Use device_descriptor for Hopper + warpspec. + if supports_host_descriptor() and not (is_hopper() and warp_specialize): + # Note that on Hopper we cannot perform a FP8 dot with a non-transposed second tensor + y_dim_q = q.shape[0] * q.shape[1] * q.shape[2] + y_dim_kv = k.shape[0] * k.shape[1] * k.shape[2] + + dummy_block = [1, 1] + # (Z, H, N_CTX_Q, HEAD_DIM_K) -> (Z*H*N_CTX_Q, HEAD_DIM_K) + desc_q = TensorDescriptor(q, shape=[y_dim_q, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], block_shape=dummy_block) + if q.dtype == torch.float8_e5m2: + desc_v = TensorDescriptor( + v, shape=[HEAD_DIM_K, y_dim_kv], strides=[k.shape[2], 1], + block_shape=dummy_block + ) + else: + desc_v = TensorDescriptor( + v, shape=[y_dim_kv, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], + block_shape=dummy_block + ) + desc_k = TensorDescriptor(k, shape=[y_dim_kv, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], block_shape=dummy_block) + # Use 3D descriptor for output to handle N_CTX_Q boundaries correctly + dummy_block3d = [1, 1, 1] + ZH = q.shape[0] * q.shape[1] + desc_o = TensorDescriptor( + o, shape=[ZH, N_CTX_Q, HEAD_DIM_K], + strides=[N_CTX_Q * HEAD_DIM_K, HEAD_DIM_K, 1], + block_shape=dummy_block3d + ) + if IS_QAT: + desc_high_prec_o = TensorDescriptor( + high_prec_o, shape=[ZH, N_CTX_Q, HEAD_DIM_K], + strides=[N_CTX_Q * HEAD_DIM_K, HEAD_DIM_K, 1], + block_shape=dummy_block3d + ) + else: + desc_high_prec_o = desc_o # Use regular output descriptor when not in QAT mode + else: + desc_q = q + desc_v = v + desc_k = k + desc_o = o + if IS_QAT: + desc_high_prec_o = high_prec_o + else: + desc_high_prec_o = o # Use regular output when not in QAT mode + + def alloc_fn(size: int, align: int, _): + return torch.empty(size, dtype=torch.int8, device="cuda") + + triton.set_allocator(alloc_fn) + + def grid(META): + # (ceil(N_CTX / BLOCK_M), Z * H, 1) + return (triton.cdiv(q.shape[2], META["BLOCK_M"]), q.shape[0] * q.shape[1], 1) + + ctx.grid = grid + if is_blackwell() and warp_specialize: + if HEAD_DIM_K == 128 and q.dtype == torch.float16: + extra_kern_args["maxnreg"] = 168 + else: + extra_kern_args["maxnreg"] = 80 + + BLOCK_M, BLOCK_N = 32, 32 + if IS_QAT: + fake_q = torch.empty_like(q) + fake_k = torch.empty_like(k) + fake_v = torch.empty_like(v) + + # override desc_q, desc_k, desc_v with fake_q, fake_k, fake_v + if supports_host_descriptor() and not (is_hopper() and warp_specialize): + # (Z, H, N_CTX_Q, HEAD_DIM_K) -> (Z*H*N_CTX_Q, HEAD_DIM_K) + desc_q = TensorDescriptor(fake_q, shape=[y_dim_q, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], block_shape=dummy_block) + desc_k = TensorDescriptor(fake_k, shape=[y_dim_kv, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], block_shape=dummy_block) + desc_v = TensorDescriptor(fake_v, shape=[y_dim_kv, HEAD_DIM_K], strides=[HEAD_DIM_K, 1], block_shape=dummy_block) + else: + desc_q = fake_q + desc_k = fake_k + desc_v = fake_v + + H = q.shape[1] + grid_1 = (triton.cdiv(q.shape[2], BLOCK_M), q.shape[0] * q.shape[1], 1) + grid_2 = (triton.cdiv(k.shape[2], BLOCK_N), q.shape[0] * q.shape[1], 1) + + fake_quantize_q[grid_1]( + q, fake_q, + q.stride(0), q.stride(1), + q.stride(2), q.stride(3), + fake_q.stride(0), fake_q.stride(1), + fake_q.stride(2), fake_q.stride(3), + H, N_CTX_Q, + BLOCK_M=BLOCK_M, HEAD_DIM=HEAD_DIM_K, + use_global_sf=use_global_sf_QKV, + ) + fake_quantize_kv[grid_2]( + k, v, fake_k, fake_v, + k.stride(0), k.stride(1), + k.stride(2), k.stride(3), + fake_k.stride(0), fake_k.stride(1), + fake_k.stride(2), fake_k.stride(3), + H, N_CTX_KV, + BLOCK_N=BLOCK_N, HEAD_DIM=HEAD_DIM_K, + use_global_sf=use_global_sf_QKV, + ) + + # Apply pre-hook to set block shapes on tensor descriptors + _host_descriptor_pre_hook({ + "BLOCK_M": BLOCK_M, + "BLOCK_N": BLOCK_N, + "HEAD_DIM": HEAD_DIM_K, + "desc_q": desc_q, + "desc_k": desc_k, + "desc_v": desc_v, + "desc_o": desc_o, + "desc_high_prec_o": desc_high_prec_o, + "FP8_OUTPUT": q.dtype == torch.float8_e5m2, + }) + + _attn_fwd[grid]( + sm_scale, M, + q.shape[0], q.shape[1], + desc_q, desc_k, desc_v, desc_o, desc_high_prec_o, + N_CTX_Q=N_CTX_Q, + N_CTX_KV=N_CTX_KV, + HEAD_DIM=HEAD_DIM_K, + BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N, + FP8_OUTPUT=q.dtype == torch.float8_e5m2, + STAGE=stage, + warp_specialize=warp_specialize, + IS_HOPPER=is_hopper(), + IS_QAT=IS_QAT, + fake_quant_P=fake_quant_P, + two_level_quant_P=two_level_quant_P, + use_global_sf_P=use_global_sf_P, + num_warps=4, + num_stages=2, + **extra_kern_args + ) + o_for_bwd = high_prec_o if IS_QAT and use_high_prec_o else o + + if IS_QAT: + q = fake_q + k = fake_k + v = fake_v + + ctx.save_for_backward(q, k, v, o_for_bwd, M) + ctx.sm_scale = sm_scale + ctx.HEAD_DIM = HEAD_DIM_K + ctx.causal = causal + ctx.IS_QAT = IS_QAT + ctx.use_qat_qkv_backward = use_qat_qkv_backward + ctx.smooth_k = smooth_k + ctx.two_level_quant_P = two_level_quant_P + ctx.fake_quant_P = fake_quant_P + ctx.smooth_q = smooth_q + ctx.use_global_sf_P = use_global_sf_P + ctx.warp_specialize = warp_specialize + return o + + @staticmethod + def backward(ctx, do): + q, k, v, o_for_bwd, M = ctx.saved_tensors + do = do.contiguous() + assert do.is_contiguous() + dq = torch.empty_like(q) + dk = torch.empty_like(k) + dv = torch.empty_like(v) + BATCH, N_HEAD, N_CTX_Q = q.shape[:3] + N_CTX_KV = k.shape[2] + assert k.shape[2] == v.shape[2], "k and v must have the same sequence length" + PRE_BLOCK = 128 + NUM_STAGES = 3 + NUM_WARPS = 4 + BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 32, 32, 32, 32 + if not ctx.use_qat_qkv_backward: + q = ctx.q_orig + k = ctx.k_orig + v = ctx.v_orig + BLK_SLICE_FACTOR = 1 if ctx.IS_QAT else 2 # must be 1 for QAT + # NOTE: K is NOT pre-scaled here - scaling is applied AFTER qk dot product in kernels + # This improves precision by avoiding rounding errors in K before the dot product + arg_k = k + pre_grid = ((N_CTX_Q + PRE_BLOCK - 1) // PRE_BLOCK, BATCH * N_HEAD) + delta = torch.empty_like(M) + _attn_bwd_preprocess[pre_grid]( + o_for_bwd, do, + delta, + BATCH, N_HEAD, N_CTX_Q, + BLOCK_M=PRE_BLOCK, HEAD_DIM=ctx.HEAD_DIM + ) + + q_m = None + if ctx.smooth_q: + # _, q_m = triton_group_mean(q) + q_m = q_m.repeat_interleave(q.shape[2] // q_m.shape[2], dim=2) # B,H,L,D + + if N_CTX_Q == N_CTX_KV: + # Use existing kernel for self-attention (same sequence lengths) + grid = ((N_CTX_KV + BLOCK_N1 - 1) // BLOCK_N1, 1, BATCH * N_HEAD) + _attn_bwd[grid]( + q, arg_k, v, ctx.sm_scale, do, dq, dk, dv, + M, delta, q_m, + q.stride(0), q.stride(1), q.stride(2), q.stride(3), + N_HEAD, N_CTX_KV, + ctx.k_mean, + BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1, + BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2, + BLK_SLICE_FACTOR=BLK_SLICE_FACTOR, + HEAD_DIM=ctx.HEAD_DIM, + CAUSAL=ctx.causal, + IS_QAT=ctx.IS_QAT, + SMOOTH_K=ctx.smooth_k, + two_level_quant_P=ctx.two_level_quant_P, + fake_quant_P=ctx.fake_quant_P, + SMOOTH_Q=ctx.smooth_q, + use_global_sf_P=ctx.use_global_sf_P, + warp_specialize=ctx.warp_specialize, + num_warps=NUM_WARPS, + num_stages=NUM_STAGES + ) + else: + # Use separate kernels for cross-attention (different sequence lengths) + grid_dq = ((N_CTX_Q + BLOCK_M2 - 1) // BLOCK_M2, 1, BATCH * N_HEAD) + _attn_bwd_dq_cross[grid_dq]( + q, arg_k, v, ctx.sm_scale, do, dq, M, delta, + q.stride(0), k.stride(0), q.stride(1), k.stride(1), q.stride(2), k.stride(2), q.stride(3), k.stride(3), + N_HEAD, N_CTX_Q, N_CTX_KV, + ctx.k_mean, + BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2, + HEAD_DIM=ctx.HEAD_DIM, + SMOOTH_K=ctx.smooth_k, + warp_specialize=ctx.warp_specialize, + num_warps=NUM_WARPS, + num_stages=NUM_STAGES, + ) + grid_dkdv = ((N_CTX_KV + BLOCK_N1 - 1) // BLOCK_N1, 1, BATCH * N_HEAD) + _attn_bwd_dkdv_cross[grid_dkdv]( + q, arg_k, v, ctx.sm_scale, do, dk, dv, M, delta, q_m, + q.stride(0), k.stride(0), q.stride(1), k.stride(1), q.stride(2), k.stride(2), q.stride(3), k.stride(3), + N_HEAD, N_CTX_Q, N_CTX_KV, + BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1, + HEAD_DIM=ctx.HEAD_DIM, + IS_QAT=ctx.IS_QAT, + two_level_quant_P=ctx.two_level_quant_P, + fake_quant_P=ctx.fake_quant_P, + SMOOTH_Q=ctx.smooth_q, + use_global_sf_P=ctx.use_global_sf_P, + warp_specialize=ctx.warp_specialize, + num_warps=NUM_WARPS, + num_stages=NUM_STAGES, + ) + + return dq, dk, dv, None, None, None, None, None, None, None, None, None, None, None, None + + +attention = _attention.apply \ No newline at end of file diff --git a/quant_utils.py b/quant_utils.py new file mode 100644 index 000000000..8b824f4b3 --- /dev/null +++ b/quant_utils.py @@ -0,0 +1,77 @@ +import triton +import triton.language as tl +from nvfp4_utils import _compute_quant_and_scale, _compute_dequant + +@triton.jit +def fake_quantize(src_tensor, valid_src_mask, BLOCK_SIZE_OUT_DIM: tl.constexpr, + BLOCK_SIZE_QUANT_DIM: tl.constexpr, + dst_dtype: tl.constexpr, + use_global_sf: tl.constexpr = True, + two_level_quant_P: tl.constexpr = False): + high_prec_src_tensor = src_tensor + src_tensor, src_scale, src_s_dec = _compute_quant_and_scale(src_tensor=src_tensor, + valid_src_mask=valid_src_mask, + use_global_sf=use_global_sf, + two_level_quant_P=two_level_quant_P) + src_tensor = _compute_dequant(mx_tensor=src_tensor, + scale=src_scale, + s_dec=src_s_dec, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=dst_dtype) + return src_tensor, high_prec_src_tensor.to(src_tensor.dtype) + +@triton.jit +def fake_quantize_q(Q, fake_Q, stride_z_q, stride_h_q, + stride_tok_q, stride_d_q, + fake_stride_z_q, fake_stride_h_q, + fake_stride_tok_q, fake_stride_d_q, + H, N_CTX_Q, + BLOCK_M: tl.constexpr, + HEAD_DIM: tl.constexpr, + use_global_sf: tl.constexpr = True): + bhid = tl.program_id(1) + adj_q = (stride_h_q * (bhid % H) + stride_z_q * (bhid // H)) + fake_adj_q = (fake_stride_h_q * (bhid % H) + fake_stride_z_q * (bhid // H)) + Q += adj_q + fake_Q += fake_adj_q + + pid = tl.program_id(0) + start_m = pid * BLOCK_M + offs_m = start_m + tl.arange(0, BLOCK_M) + offs_k = tl.arange(0, HEAD_DIM) + + q_valid = offs_m < N_CTX_Q + q = tl.load(Q + offs_m[:, None] * stride_tok_q + offs_k[None, :] * stride_d_q, mask=q_valid[:, None], other=0.0) + q, _ = fake_quantize(src_tensor=q, valid_src_mask=q_valid[:, None], BLOCK_SIZE_OUT_DIM=BLOCK_M, BLOCK_SIZE_QUANT_DIM=HEAD_DIM, dst_dtype=q.dtype, use_global_sf=use_global_sf) + tl.store(fake_Q + offs_m[:, None] * fake_stride_tok_q + offs_k[None, :] * fake_stride_d_q, q, mask=q_valid[:, None]) + +@triton.jit +def fake_quantize_kv(K, V, fake_K, fake_V, stride_z_kv, stride_h_kv, + stride_tok_kv, stride_d_kv, + fake_stride_z_kv, fake_stride_h_kv, + fake_stride_tok_kv, fake_stride_d_kv, + H, N_CTX_KV, + BLOCK_N: tl.constexpr, + HEAD_DIM: tl.constexpr, + use_global_sf: tl.constexpr = True): + bhid = tl.program_id(1) + adj_kv = (stride_h_kv * (bhid % H) + stride_z_kv * (bhid // H)) + fake_adj_kv = (fake_stride_h_kv * (bhid % H) + fake_stride_z_kv * (bhid // H)) + K += adj_kv + V += adj_kv + fake_K += fake_adj_kv + fake_V += fake_adj_kv + + pid = tl.program_id(0) + start_n = pid * BLOCK_N + offs_n = start_n + tl.arange(0, BLOCK_N) + offs_k = tl.arange(0, HEAD_DIM) + + kv_valid = offs_n < N_CTX_KV + k_block = tl.load(K + offs_n[:, None] * stride_tok_kv + offs_k[None, :] * stride_d_kv, mask=kv_valid[:, None], other=0.0) + v_block = tl.load(V + offs_n[:, None] * stride_tok_kv + offs_k[None, :] * stride_d_kv, mask=kv_valid[:, None], other=0.0) + k, _ = fake_quantize(src_tensor=k_block, valid_src_mask=kv_valid[:, None], BLOCK_SIZE_OUT_DIM=BLOCK_N, BLOCK_SIZE_QUANT_DIM=HEAD_DIM, dst_dtype=k_block.dtype, use_global_sf=use_global_sf) + v, _ = fake_quantize(src_tensor=v_block, valid_src_mask=kv_valid[:, None], BLOCK_SIZE_OUT_DIM=BLOCK_N, BLOCK_SIZE_QUANT_DIM=HEAD_DIM, dst_dtype=v_block.dtype, use_global_sf=use_global_sf) + tl.store(fake_K + offs_n[:, None] * fake_stride_tok_kv + offs_k[None, :] * fake_stride_d_kv, k, mask=kv_valid[:, None]) + tl.store(fake_V + offs_n[:, None] * fake_stride_tok_kv + offs_k[None, :] * fake_stride_d_kv, v, mask=kv_valid[:, None]) diff --git a/test_fake_quant.py b/test_fake_quant.py new file mode 100644 index 000000000..f544ee08d --- /dev/null +++ b/test_fake_quant.py @@ -0,0 +1,524 @@ +#!/usr/bin/env python3 +""" +Precision test to compare numerical differences for fake_quantize triton function: +1. fake_quantize (triton implementation) vs reference implementations +2. Tests various shapes, dtypes, and value ranges +3. Evaluates cosine similarity, max diff, and mean diff between input and output +""" + +import torch +import triton +import triton.language as tl +from flashinfer import SfLayout, nvfp4_quantize, e2m1_and_ufp8sf_scale_to_float +from nvfp4_utils import _compute_quant_and_scale, _compute_dequant +from typing import Optional + +# MXFP_BLOCK_SIZE is 16 - use Python int for runtime checks +MXFP_BLOCK_SIZE = 16 + +DEVICE = torch.device("cuda") + + +def cosine_similarity(tensor1, tensor2): + """ + Compute cosine similarity between two tensors. + + Args: + tensor1: First tensor + tensor2: Second tensor (same shape as tensor1) + + Returns: + Cosine similarity value (scalar) + - Returns 1.0 if both tensors are zero (identical zero vectors) + - Returns 0.0 if only one tensor is zero (orthogonal to non-zero vector) + """ + # Flatten tensors for computation + t1_flat = tensor1.flatten().float() + t2_flat = tensor2.flatten().float() + + # Compute cosine similarity: (A · B) / (||A|| * ||B||) + dot_product = torch.dot(t1_flat, t2_flat) + norm1 = torch.norm(t1_flat) + norm2 = torch.norm(t2_flat) + + # Handle zero vectors + if norm1 == 0 and norm2 == 0: + # Both are zero vectors - they are identical, so similarity is 1.0 + return 1.0 + elif norm1 == 0 or norm2 == 0: + # One is zero, one is not - they are orthogonal, so similarity is 0.0 + return 0.0 + + cos_sim = dot_product / (norm1 * norm2) + return cos_sim.item() + + +@triton.jit +def fake_quantize(src_tensor, valid_src_mask, BLOCK_SIZE_OUT_DIM: tl.constexpr, + BLOCK_SIZE_QUANT_DIM: tl.constexpr, + dst_dtype: tl.constexpr, + mx_tensor_dtype: tl.constexpr = tl.uint8): + """ + Fake quantize function - matches API from qat_attn.py. + """ + high_prec_src_tensor = src_tensor + src_tensor, src_scale, src_s_dec = _compute_quant_and_scale( + src_tensor=src_tensor, + valid_src_mask=valid_src_mask, + mx_tensor_dtype=mx_tensor_dtype + ) + src_tensor = _compute_dequant( + mx_tensor=src_tensor, + scale=src_scale, + s_dec=src_s_dec, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=dst_dtype + ) + return src_tensor, high_prec_src_tensor + + +def get_fake_quant_reference(x: torch.Tensor): + """ + Reference implementation using FlashInfer for comparison. + """ + orig_shape = x.shape + orig_dtype = x.dtype + device = x.device + x = x.view(-1, x.shape[-1]) + x_global_sf = (448 * 6) / x.float().abs().nan_to_num().max() + x_fp4, x_scale = nvfp4_quantize(x, x_global_sf, sfLayout=SfLayout.layout_128x4, do_shuffle=False) + x_dequant = e2m1_and_ufp8sf_scale_to_float(x_fp4, x_scale, 1 / x_global_sf) + return x_dequant.view(orig_shape).to(orig_dtype).to(device) + + +@triton.jit +def fake_quantize_kernel( + src_ptr, + dst_ptr, + BLOCK_SIZE_OUT_DIM: tl.constexpr, + BLOCK_SIZE_QUANT_DIM: tl.constexpr, + dst_dtype: tl.constexpr, + mx_tensor_dtype: tl.constexpr, + stride_src_outer, + stride_src_quant, + stride_dst_outer, + stride_dst_quant, + outer_dim, + quant_dim, +): + """ + Kernel wrapper to call fake_quantize on a block of data. + """ + outer_idx = tl.program_id(0) + quant_idx = tl.program_id(1) + + # Compute offsets + start_outer = outer_idx * BLOCK_SIZE_OUT_DIM + start_quant = quant_idx * BLOCK_SIZE_QUANT_DIM + + # Create offset arrays + offs_outer = tl.arange(0, BLOCK_SIZE_OUT_DIM)[:, None] + offs_quant = tl.arange(0, BLOCK_SIZE_QUANT_DIM)[None, :] + + # Create masks for valid elements + mask_outer = (start_outer + offs_outer) < outer_dim + mask_quant = (start_quant + offs_quant) < quant_dim + full_mask = mask_outer & mask_quant + + # Load source tensor + src_offsets = (start_outer + offs_outer) * stride_src_outer + (start_quant + offs_quant) * stride_src_quant + src_tensor = tl.load(src_ptr + src_offsets, mask=full_mask, other=0.0) + + # Call fake_quantize with valid_src_mask parameter + quantized_tensor, high_prec_tensor = fake_quantize( + src_tensor=src_tensor, + valid_src_mask=full_mask, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=dst_dtype, + mx_tensor_dtype=mx_tensor_dtype + ) + + # Store result + dst_offsets = (start_outer + offs_outer) * stride_dst_outer + (start_quant + offs_quant) * stride_dst_quant + tl.store(dst_ptr + dst_offsets, quantized_tensor, mask=full_mask) + + +def triton_fake_quantize( + x: torch.Tensor, + BLOCK_SIZE_OUT_DIM: int = 128, + BLOCK_SIZE_QUANT_DIM: int = 128, + use_fp4: bool = True, # True for fp4 (uint8), False for fp8 (float8e4nv) + dst_dtype: Optional[torch.dtype] = None +) -> torch.Tensor: + """ + Call fake_quantize triton function on a tensor. + + Args: + x: Input tensor (2D or can be reshaped to 2D) + BLOCK_SIZE_OUT_DIM: Block size for outer dimension + BLOCK_SIZE_QUANT_DIM: Block size for quantization dimension (must be multiple of 16) + use_fp4: If True, use fp4 (uint8), else use fp8 (float8e4nv) + dst_dtype: Output dtype (defaults to input dtype) + + Returns: + Fake quantized tensor + """ + assert x.is_cuda, "Input must be on CUDA" + assert BLOCK_SIZE_QUANT_DIM % 16 == 0, f"BLOCK_SIZE_QUANT_DIM must be multiple of 16" + + orig_shape = x.shape + orig_dtype = x.dtype + + # Reshape to 2D + x_2d = x.view(-1, x.shape[-1]) + outer_dim, quant_dim = x_2d.shape + + if dst_dtype is None: + dst_dtype = orig_dtype + + # Map torch dtype to triton dtype + dtype_map = { + torch.float32: tl.float32, + torch.float16: tl.float16, + torch.bfloat16: tl.bfloat16, + } + triton_dst_dtype = dtype_map.get(dst_dtype, tl.float16) + + # Allocate output + output = torch.empty_like(x_2d, dtype=dst_dtype) + + # Launch kernel with appropriate quantization dtype + grid = ( + triton.cdiv(outer_dim, BLOCK_SIZE_OUT_DIM), + triton.cdiv(quant_dim, BLOCK_SIZE_QUANT_DIM), + ) + + if use_fp4: + # Use fp4 (uint8) + fake_quantize_kernel[grid]( + src_ptr=x_2d, + dst_ptr=output, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=triton_dst_dtype, + mx_tensor_dtype=tl.uint8, # fp4 uses uint8 + stride_src_outer=x_2d.stride(0), + stride_src_quant=x_2d.stride(1), + stride_dst_outer=output.stride(0), + stride_dst_quant=output.stride(1), + outer_dim=outer_dim, + quant_dim=quant_dim, + ) + else: + # Use fp8 (float8e4nv) + fake_quantize_kernel[grid]( + src_ptr=x_2d, + dst_ptr=output, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=triton_dst_dtype, + mx_tensor_dtype=tl.float8e4nv, # fp8 uses float8e4nv + stride_src_outer=x_2d.stride(0), + stride_src_quant=x_2d.stride(1), + stride_dst_outer=output.stride(0), + stride_dst_quant=output.stride(1), + outer_dim=outer_dim, + quant_dim=quant_dim, + ) + + return output.view(orig_shape).to(dst_dtype) + + +def test_fake_quantize_basic(): + """Test basic functionality of fake_quantize.""" + torch.manual_seed(42) + + # Test parameters + shape = (128, 128) + dtype = torch.bfloat16 + + # Create input tensor + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Test triton fake_quantize + x_fq = triton_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + # Check that outputs have correct shape + assert x_fq.shape == x.shape + assert x_fq.dtype == x.dtype + + # Check that outputs are finite + assert torch.isfinite(x_fq).all() + + # Compare input and output + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" Input vs Output - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ Basic test passed.") + + +def test_fake_quantize_different_shapes(): + """Test fake_quantize with different input shapes.""" + torch.manual_seed(42) + + test_configs = [ + (128, 64), # Small + (256, 128), # Medium + (512, 256), # Large + (128, 256), # Rectangular + (256, 128), # Rectangular (reversed) + ] + + dtype = torch.bfloat16 + + for outer_dim, quant_dim in test_configs: + # Create input tensor + x = torch.randn((outer_dim, quant_dim), dtype=dtype, device=DEVICE) + + # Test triton fake_quantize + x_fq = triton_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + # Compare input and output + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" Shape ({outer_dim}, {quant_dim}) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print(f"✓ Shape test passed for ({outer_dim}, {quant_dim})") + + +def test_fake_quantize_different_dtypes(): + """Test fake_quantize with different input dtypes.""" + torch.manual_seed(42) + + shape = (128, 128) + dtypes = [torch.float32, torch.float16, torch.bfloat16] + + for dtype in dtypes: + # Create input tensor + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Test triton fake_quantize + x_fq = triton_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + assert x_fq.shape == x.shape + assert x_fq.dtype == x.dtype + assert torch.isfinite(x_fq).all() + + # Compare input and output + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" Dtype {dtype} - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print(f"✓ Dtype test passed for {dtype}") + + +def test_fake_quantize_3d_4d_tensors(): + """Test fake_quantize with 3D and 4D tensors (reshaped to 2D internally).""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + + # Test 3D tensor (B, H, D) + print("\nTesting 3D tensor (B, H, D)") + x_3d = torch.randn((2, 8, 128), dtype=dtype, device=DEVICE) + x_fq_3d = triton_fake_quantize(x_3d, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + assert x_fq_3d.shape == x_3d.shape + assert torch.isfinite(x_fq_3d).all() + + max_diff = (x_fq_3d - x_3d).abs().max() + mean_diff = (x_fq_3d - x_3d).abs().mean() + cos_sim = cosine_similarity(x_fq_3d, x_3d) + print(f" 3D (2, 8, 128) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Test 4D tensor (B, H, L, D) + print("\nTesting 4D tensor (B, H, L, D)") + x_4d = torch.randn((1, 8, 256, 128), dtype=dtype, device=DEVICE) + x_fq_4d = triton_fake_quantize(x_4d, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + assert x_fq_4d.shape == x_4d.shape + assert torch.isfinite(x_fq_4d).all() + + max_diff = (x_fq_4d - x_4d).abs().max() + mean_diff = (x_fq_4d - x_4d).abs().mean() + cos_sim = cosine_similarity(x_fq_4d, x_4d) + print(f" 4D (1, 8, 256, 128) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ 3D/4D tensor test passed.") + + +def test_fake_quantize_edge_cases(): + """Test edge cases like zeros, ones, extreme values.""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + shape = (128, 128) + + test_cases = [ + ("zeros", torch.zeros(shape, dtype=dtype, device=DEVICE)), + ("ones", torch.ones(shape, dtype=dtype, device=DEVICE)), + ("negative_ones", -torch.ones(shape, dtype=dtype, device=DEVICE)), + ("very_small", torch.randn(shape, dtype=dtype, device=DEVICE) * 1e-6), + ("very_large", torch.randn(shape, dtype=dtype, device=DEVICE) * 1e6), + ] + + for name, x in test_cases: + x_fq = triton_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" {name} - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ Edge cases test passed.") + + +def test_fake_quantize_vs_reference(): + """Test triton fake_quantize vs reference implementation.""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + shape = (128, 128) + + # Create input tensor + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Test triton fake_quantize + x_fq_triton = triton_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + # Test reference implementation + x_fq_ref = get_fake_quant_reference(x) + + # Compare triton vs reference + max_diff = (x_fq_triton - x_fq_ref).abs().max() + mean_diff = (x_fq_triton - x_fq_ref).abs().mean() + cos_sim = cosine_similarity(x_fq_triton, x_fq_ref) + print(f" Triton vs Reference - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Also compare each vs input + max_diff_triton = (x_fq_triton - x).abs().max() + mean_diff_triton = (x_fq_triton - x).abs().mean() + cos_sim_triton = cosine_similarity(x_fq_triton, x) + print(f" Triton vs Input - Max diff: {max_diff_triton.item():.6f}, Mean diff: {mean_diff_triton.item():.6f}, Cosine sim: {cos_sim_triton:.6f}") + + max_diff_ref = (x_fq_ref - x).abs().max() + mean_diff_ref = (x_fq_ref - x).abs().mean() + cos_sim_ref = cosine_similarity(x_fq_ref, x) + print(f" Reference vs Input - Max diff: {max_diff_ref.item():.6f}, Mean diff: {mean_diff_ref.item():.6f}, Cosine sim: {cos_sim_ref:.6f}") + + print("✓ Reference comparison test passed.") + + +def test_fake_quantize_fp8(): + """Test fake_quantize with FP8 quantization.""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + shape = (128, 128) + + # Create input tensor + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Test triton fake_quantize with FP8 + x_fq = triton_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=False, dst_dtype=dtype) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" FP8 - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ FP8 test passed.") + + +def test_fake_quantize_attention_shapes(): + """Test fake_quantize with attention-like shapes (similar to test_qat_attn.py).""" + torch.manual_seed(42) + + test_configs = [ + (2, 4, 128, 64), # Z, H, N_CTX, HEAD_DIM - Medium + (1, 8, 256, 128), # Large head dim + (1, 40, 9360, 128), # WAN shape + ] + + dtype = torch.bfloat16 + + for Z, H, N_CTX, HEAD_DIM in test_configs: + # Create input tensor in BLHD format (B, L, H, D) = (Z, N_CTX, H, HEAD_DIM) + x = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + # Test triton fake_quantize + x_fq = triton_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + # Compare input and output + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print(f"✓ Attention shape test passed for (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM})") + + +def test_fake_quantize_non_divisible_blocks(): + """Test fake_quantize with shapes that are not divisible by block sizes.""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + + # Test with non-divisible dimensions + test_shapes = [ + (100, 100), # Not divisible by 128 + (150, 200), # Not divisible by 128 + (256, 100), # One dimension divisible, one not + ] + + for shape in test_shapes: + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Test triton fake_quantize + x_fq = triton_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=128, use_fp4=True, dst_dtype=dtype) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" Shape {shape} - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print(f"✓ Non-divisible block test passed for {shape}") + + +if __name__ == "__main__": + print("Running fake_quantize tests...") + print(f"Device: {DEVICE}") + print() + + test_fake_quantize_basic() + test_fake_quantize_different_shapes() + test_fake_quantize_different_dtypes() + test_fake_quantize_3d_4d_tensors() + test_fake_quantize_edge_cases() + test_fake_quantize_vs_reference() + test_fake_quantize_attention_shapes() + test_fake_quantize_non_divisible_blocks() + + print() + print("All tests passed! ✓") diff --git a/test_fake_quant_attn.py b/test_fake_quant_attn.py new file mode 100644 index 000000000..f7cc0931e --- /dev/null +++ b/test_fake_quant_attn.py @@ -0,0 +1,416 @@ +#!/usr/bin/env python3 +""" +Test script for fake quantization of attention computation using get_fake_quant method. +Tests attention computation with fake-quantized Q, K, V tensors and compares with reference. +""" + +import torch +import torch.nn.functional as F +from math import sqrt +from test_quantization_precision import get_fake_quant +import sys +import os +from contextlib import contextmanager + + +# Global flag to control verbosity +VERBOSE = False + + +@contextmanager +def suppress_stdout(): + """Context manager to suppress stdout.""" + if not VERBOSE: + with open(os.devnull, 'w') as devnull: + old_stdout = sys.stdout + sys.stdout = devnull + try: + yield + finally: + sys.stdout = old_stdout + else: + yield + + +def scaled_dot_product_attention_reference(q, k, v, causal=False, scale=None): + """ + Reference implementation of scaled dot-product attention. + + Args: + q: Query tensor [..., seq_len_q, head_dim] + k: Key tensor [..., seq_len_kv, head_dim] + v: Value tensor [..., seq_len_kv, head_dim] + causal: Whether to apply causal masking + scale: Scale factor (default: 1/sqrt(head_dim)) + + Returns: + Output tensor [..., seq_len_q, head_dim] + """ + if scale is None: + scale = 1.0 / sqrt(q.shape[-1]) + + # Compute attention scores: Q @ K^T + scores = torch.matmul(q, k.transpose(-2, -1)) * scale + + # Apply causal mask if needed + if causal: + seq_len = scores.shape[-1] + causal_mask = torch.triu(torch.ones(seq_len, seq_len, device=scores.device, dtype=scores.dtype), diagonal=1) + causal_mask = causal_mask.masked_fill(causal_mask == 1, float('-inf')) + scores = scores + causal_mask + + # Apply softmax + attn_weights = F.softmax(scores, dim=-1) + + # Apply attention weights to values + output = torch.matmul(attn_weights, v) + + return output + + +def attention_with_fake_quant(q, k, v, causal=False, scale=None, quantize_qkv=True): + """ + Compute attention with fake quantization applied to Q, K, V tensors. + + Args: + q: Query tensor [..., seq_len_q, head_dim] + k: Key tensor [..., seq_len_kv, head_dim] + v: Value tensor [..., seq_len_kv, head_dim] + causal: Whether to apply causal masking + scale: Scale factor (default: 1/sqrt(head_dim)) + quantize_qkv: Whether to apply fake quantization to Q, K, V + + Returns: + Output tensor [..., seq_len_q, head_dim] + """ + if quantize_qkv: + # Apply fake quantization to Q, K, V + if VERBOSE: + print(" Applying fake quantization to Q, K, V...") + with suppress_stdout(): + q = get_fake_quant(q) + k = get_fake_quant(k) + v = get_fake_quant(v) + + # Use PyTorch's scaled_dot_product_attention + return F.scaled_dot_product_attention( + q, k, v, + is_causal=causal, + scale=scale + ) + + +def test_attention_fake_quant_basic(): + """Test basic attention computation with fake quantization.""" + print("\n" + "="*60) + print("Test 1: Basic Attention with Fake Quantization") + print("="*60) + + if not torch.cuda.is_available(): + print("CUDA not available, skipping test") + return + + device = "cuda" + dtype = torch.bfloat16 + + # Test parameters + batch_size = 2 + num_heads = 4 + seq_len = 128 + head_dim = 64 + + # Create input tensors + torch.manual_seed(42) + q = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + k = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + v = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + + print(f"Input shapes: Q={q.shape}, K={k.shape}, V={v.shape}") + print(f"Input dtype: {dtype}") + print(f"Q stats: min={q.min().item():.4f}, max={q.max().item():.4f}, mean={q.mean().item():.4f}") + + # Reference attention (no quantization) + print("\nComputing reference attention (no quantization)...") + with torch.no_grad(): + ref_output = F.scaled_dot_product_attention(q, k, v, is_causal=False) + + # Attention with fake quantization + print("\nComputing attention with fake quantization...") + with torch.no_grad(): + quant_output = attention_with_fake_quant(q, k, v, causal=False, quantize_qkv=True) + + # Compare results + diff = torch.abs(ref_output - quant_output) + max_diff = diff.max().item() + mean_diff = diff.mean().item() + relative_diff = (diff / (torch.abs(ref_output) + 1e-8)).max().item() + + print(f"\nResults:") + print(f" Reference output stats: min={ref_output.min().item():.4f}, max={ref_output.max().item():.4f}, mean={ref_output.mean().item():.4f}") + print(f" Quantized output stats: min={quant_output.min().item():.4f}, max={quant_output.max().item():.4f}, mean={quant_output.mean().item():.4f}") + print(f" Max absolute difference: {max_diff:.6f}") + print(f" Mean absolute difference: {mean_diff:.6f}") + print(f" Max relative difference: {relative_diff:.6f}") + + # Check that outputs are finite + assert torch.isfinite(ref_output).all(), "Reference output contains NaN/Inf" + assert torch.isfinite(quant_output).all(), "Quantized output contains NaN/Inf" + + print("✓ Basic test passed!") + return max_diff, mean_diff, relative_diff + + +def test_attention_fake_quant_causal(): + """Test causal attention with fake quantization.""" + print("\n" + "="*60) + print("Test 2: Causal Attention with Fake Quantization") + print("="*60) + + if not torch.cuda.is_available(): + print("CUDA not available, skipping test") + return + + device = "cuda" + dtype = torch.bfloat16 + + batch_size = 1 + num_heads = 2 + seq_len = 64 + head_dim = 32 + + torch.manual_seed(42) + q = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + k = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + v = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + + print(f"Input shapes: Q={q.shape}, K={k.shape}, V={v.shape}") + + # Reference causal attention + print("\nComputing reference causal attention...") + with torch.no_grad(): + ref_output = F.scaled_dot_product_attention(q, k, v, is_causal=True) + + # Causal attention with fake quantization + print("\nComputing causal attention with fake quantization...") + with torch.no_grad(): + quant_output = attention_with_fake_quant(q, k, v, causal=True, quantize_qkv=True) + + # Compare results + diff = torch.abs(ref_output - quant_output) + max_diff = diff.max().item() + mean_diff = diff.mean().item() + + print(f"\nResults:") + print(f" Max absolute difference: {max_diff:.6f}") + print(f" Mean absolute difference: {mean_diff:.6f}") + + assert torch.isfinite(ref_output).all() + assert torch.isfinite(quant_output).all() + + print("✓ Causal attention test passed!") + return max_diff, mean_diff + + +def test_attention_fake_quant_different_shapes(): + """Test attention with different tensor shapes.""" + print("\n" + "="*60) + print("Test 3: Different Shapes") + print("="*60) + + if not torch.cuda.is_available(): + print("CUDA not available, skipping test") + return + + device = "cuda" + dtype = torch.bfloat16 + + test_configs = [ + (1, 1, 32, 16, "small"), + (2, 4, 128, 64, "medium"), + (1, 8, 256, 128, "large_seq"), + (2, 4, 64, 256, "large_head_dim"), + ] + + results = [] + + for batch_size, num_heads, seq_len, head_dim, name in test_configs: + print(f"\nTesting {name} configuration: B={batch_size}, H={num_heads}, L={seq_len}, D={head_dim}") + + torch.manual_seed(42) + q = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + k = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + v = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + + with torch.no_grad(): + ref_output = F.scaled_dot_product_attention(q, k, v, is_causal=False) + quant_output = attention_with_fake_quant(q, k, v, causal=False, quantize_qkv=True) + + diff = torch.abs(ref_output - quant_output) + max_diff = diff.max().item() + mean_diff = diff.mean().item() + + print(f" Max diff: {max_diff:.6f}, Mean diff: {mean_diff:.6f}") + + assert torch.isfinite(ref_output).all() + assert torch.isfinite(quant_output).all() + assert ref_output.shape == quant_output.shape + + results.append((name, max_diff, mean_diff)) + + print("\n✓ All shape tests passed!") + return results + + +def test_attention_fake_quant_gradient(): + """Test that gradients can flow through fake-quantized attention.""" + print("\n" + "="*60) + print("Test 4: Gradient Flow") + print("="*60) + + if not torch.cuda.is_available(): + print("CUDA not available, skipping test") + return + + device = "cuda" + dtype = torch.bfloat16 + + batch_size = 1 + num_heads = 2 + seq_len = 32 + head_dim = 32 + + torch.manual_seed(42) + q = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype, requires_grad=True) + k = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype, requires_grad=True) + v = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype, requires_grad=True) + + print(f"Input shapes: Q={q.shape}, K={k.shape}, V={v.shape}") + print("Note: get_fake_quant uses no_grad internally, so gradients won't flow through quantization") + + # Forward pass with fake quantization + # Note: get_fake_quant uses no_grad, so we'll test that it doesn't crash + with torch.no_grad(): + q_quant = get_fake_quant(q) + k_quant = get_fake_quant(k) + v_quant = get_fake_quant(v) + + # Now compute attention with quantized tensors (requires_grad=False) + q_quant.requires_grad_(True) + k_quant.requires_grad_(True) + v_quant.requires_grad_(True) + + output = F.scaled_dot_product_attention(q_quant, k_quant, v_quant, is_causal=False) + + # Backward pass + loss = output.sum() + loss.backward() + + # Check that gradients exist + assert q_quant.grad is not None + assert k_quant.grad is not None + assert v_quant.grad is not None + + print("✓ Gradient flow test passed!") + print(f" Gradient shapes: Q={q_quant.grad.shape}, K={k_quant.grad.shape}, V={v_quant.grad.shape}") + + +def test_attention_fake_quant_precision(): + """Test precision comparison between quantized and non-quantized attention.""" + print("\n" + "="*60) + print("Test 5: Precision Analysis") + print("="*60) + + if not torch.cuda.is_available(): + print("CUDA not available, skipping test") + return + + device = "cuda" + dtype = torch.bfloat16 + + batch_size = 2 + num_heads = 4 + seq_len = 128 + head_dim = 64 + + torch.manual_seed(42) + q = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + k = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + v = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype) + + print(f"Input shapes: Q={q.shape}, K={k.shape}, V={v.shape}") + + # Compute reference + with torch.no_grad(): + ref_output = F.scaled_dot_product_attention(q, k, v, is_causal=False) + + # Compute with quantization + with torch.no_grad(): + quant_output = attention_with_fake_quant(q, k, v, causal=False, quantize_qkv=True) + + # Detailed analysis + diff = torch.abs(ref_output - quant_output) + + print("\nPrecision Analysis:") + print(f" Output shape: {ref_output.shape}") + print(f" Total elements: {ref_output.numel()}") + print(f" Max absolute error: {diff.max().item():.6f}") + print(f" Mean absolute error: {diff.mean().item():.6f}") + print(f" Median absolute error: {diff.median().item():.6f}") + print(f" Std of errors: {diff.std().item():.6f}") + + # Relative error + relative_diff = diff / (torch.abs(ref_output) + 1e-8) + print(f" Max relative error: {relative_diff.max().item():.6f}") + print(f" Mean relative error: {relative_diff.mean().item():.6f}") + + # Error distribution (convert to float32 for quantile computation) + error_percentiles = torch.quantile(diff.float(), torch.tensor([0.5, 0.9, 0.95, 0.99], device=device, dtype=torch.float32)) + print(f" Error percentiles (50th, 90th, 95th, 99th): {error_percentiles.cpu().tolist()}") + + print("\n✓ Precision analysis completed!") + + +def main(): + """Run all tests.""" + global VERBOSE + + # Check for verbose flag + VERBOSE = "--verbose" in sys.argv or "-v" in sys.argv + + print("="*60) + print("Fake Quantization Attention Test Suite") + print("="*60) + + if not torch.cuda.is_available(): + print("ERROR: CUDA is not available. These tests require CUDA.") + sys.exit(1) + + print(f"CUDA Device: {torch.cuda.get_device_name()}") + print(f"PyTorch version: {torch.__version__}") + if VERBOSE: + print("Verbose mode: ON (will show quantization details)") + print() + + try: + # Run tests + test_attention_fake_quant_basic() + test_attention_fake_quant_causal() + test_attention_fake_quant_different_shapes() + test_attention_fake_quant_gradient() + test_attention_fake_quant_precision() + + print("\n" + "="*60) + print("All tests passed! ✓") + print("="*60) + + except Exception as e: + print(f"\n❌ Test failed with error: {e}") + import traceback + traceback.print_exc() + sys.exit(1) + + +if __name__ == "__main__": + main() + + diff --git a/test_fake_quant_basic.py b/test_fake_quant_basic.py new file mode 100644 index 000000000..c29a45b3b --- /dev/null +++ b/test_fake_quant_basic.py @@ -0,0 +1,883 @@ +#!/usr/bin/env python3 +""" +Basic tests for the fake_quantize function. +Tests core functionality with simple inputs and edge cases. +""" + +import torch +import triton +import triton.language as tl +from typing import Tuple +from qat_attn import fake_quantize +from nvfp4_utils import _downcast_to_mxfp, _upcast_from_mxfp, MXFP_BLOCK_SIZE, _compute_quant_and_scale, _compute_dequant + +# Import get_fake_quant from test_quantization_precision +try: + from test_quantization_precision import get_fake_quant +except ImportError: + # Fallback: define get_fake_quant here if import fails + from flashinfer import SfLayout, nvfp4_quantize, e2m1_and_ufp8sf_scale_to_float + def get_fake_quant(x: torch.Tensor): + orig_shape = x.shape + orig_dtype = x.dtype + device = x.device + x = x.view(-1, x.shape[-1]) + x_global_sf = (448 * 6) / x.float().abs().nan_to_num().max() + x_fp4, x_scale = nvfp4_quantize(x, x_global_sf, sfLayout=SfLayout.layout_128x4, do_shuffle=False) + x_dequant = e2m1_and_ufp8sf_scale_to_float(x_fp4, x_scale, 1 / x_global_sf) + val = x_dequant.view(orig_shape).to(orig_dtype).to(device) + return val + +DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") + + +def cosine_similarity(tensor1, tensor2): + """ + Compute cosine similarity between two tensors. + + Args: + tensor1: First tensor + tensor2: Second tensor (same shape as tensor1) + + Returns: + Cosine similarity value (scalar) + """ + # Flatten tensors for computation + t1_flat = tensor1.flatten().float() + t2_flat = tensor2.flatten().float() + + # Compute cosine similarity: (A · B) / (||A|| * ||B||) + dot_product = torch.dot(t1_flat, t2_flat) + norm1 = torch.norm(t1_flat) + norm2 = torch.norm(t2_flat) + + # Avoid division by zero + if norm1 == 0 or norm2 == 0: + return 0.0 + + cos_sim = dot_product / (norm1 * norm2) + return cos_sim.item() + + +@triton.jit +def fake_quantize_kernel( + src_ptr, + dst_ptr, + high_prec_ptr, + BLOCK_SIZE_OUT_DIM: tl.constexpr, + BLOCK_SIZE_QUANT_DIM: tl.constexpr, + dst_dtype: tl.constexpr, + mx_tensor_dtype: tl.constexpr, + stride_src_outer, + stride_src_quant, + stride_dst_outer, + stride_dst_quant, + stride_high_prec_outer, + stride_high_prec_quant, + outer_dim, + quant_dim, +): + """ + Kernel wrapper to test fake_quantize function. + """ + outer_idx = tl.program_id(0) + quant_idx = tl.program_id(1) + + # Compute offsets + start_outer = outer_idx * BLOCK_SIZE_OUT_DIM + start_quant = quant_idx * BLOCK_SIZE_QUANT_DIM + + # Create offset arrays + offs_outer = tl.arange(0, BLOCK_SIZE_OUT_DIM)[:, None] + offs_quant = tl.arange(0, BLOCK_SIZE_QUANT_DIM)[None, :] + + # Create masks for valid elements + mask_outer = (start_outer + offs_outer) < outer_dim + mask_quant = (start_quant + offs_quant) < quant_dim + full_mask = mask_outer & mask_quant + + # Load source tensor + src_offsets = (start_outer + offs_outer) * stride_src_outer + (start_quant + offs_quant) * stride_src_quant + src_tensor = tl.load(src_ptr + src_offsets, mask=full_mask, other=0.0) + + # Call fake_quantize + quantized_tensor, high_prec_tensor = fake_quantize( + src_tensor=src_tensor, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=dst_dtype, + mx_tensor_dtype=mx_tensor_dtype + ) + + # Store results + dst_offsets = (start_outer + offs_outer) * stride_dst_outer + (start_quant + offs_quant) * stride_dst_quant + high_prec_offsets = (start_outer + offs_outer) * stride_high_prec_outer + (start_quant + offs_quant) * stride_high_prec_quant + + tl.store(dst_ptr + dst_offsets, quantized_tensor, mask=full_mask) + tl.store(high_prec_ptr + high_prec_offsets, high_prec_tensor, mask=full_mask) + + +def call_fake_quantize( + x: torch.Tensor, + BLOCK_SIZE_OUT_DIM: int = 128, + BLOCK_SIZE_QUANT_DIM: int = 128, + dst_dtype: torch.dtype = None, + mx_tensor_dtype = tl.float8e4nv +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Call fake_quantize on a tensor. + + Args: + x: Input tensor (2D or can be reshaped to 2D) + BLOCK_SIZE_OUT_DIM: Block size for outer dimension + BLOCK_SIZE_QUANT_DIM: Block size for quantization dimension (must be multiple of 16) + dst_dtype: Output dtype (defaults to input dtype) + mx_tensor_dtype: Quantization dtype (tl.float8e4nv or tl.uint8 for fp4) + + Returns: + Tuple of (quantized_tensor, high_precision_tensor) + """ + assert x.is_cuda, "Input must be on CUDA" + assert BLOCK_SIZE_QUANT_DIM % 16 == 0, f"BLOCK_SIZE_QUANT_DIM must be multiple of 16, got {BLOCK_SIZE_QUANT_DIM}" + + orig_shape = x.shape + orig_dtype = x.dtype + + # Reshape to 2D + x_2d = x.view(-1, x.shape[-1]) + outer_dim, quant_dim = x_2d.shape + + if dst_dtype is None: + dst_dtype = orig_dtype + + # Map torch dtype to triton dtype + dtype_map = { + torch.float32: tl.float32, + torch.float16: tl.float16, + torch.bfloat16: tl.bfloat16, + } + triton_dst_dtype = dtype_map.get(dst_dtype, tl.float16) + + # Allocate outputs + quantized_output = torch.empty_like(x_2d, dtype=dst_dtype) + high_prec_output = torch.empty_like(x_2d, dtype=torch.float32) + + # Launch kernel + grid = ( + triton.cdiv(outer_dim, BLOCK_SIZE_OUT_DIM), + triton.cdiv(quant_dim, BLOCK_SIZE_QUANT_DIM), + ) + + fake_quantize_kernel[grid]( + src_ptr=x_2d, + dst_ptr=quantized_output, + high_prec_ptr=high_prec_output, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=triton_dst_dtype, + mx_tensor_dtype=mx_tensor_dtype, + stride_src_outer=x_2d.stride(0), + stride_src_quant=x_2d.stride(1), + stride_dst_outer=quantized_output.stride(0), + stride_dst_quant=quantized_output.stride(1), + stride_high_prec_outer=high_prec_output.stride(0), + stride_high_prec_quant=high_prec_output.stride(1), + outer_dim=outer_dim, + quant_dim=quant_dim, + ) + + return quantized_output.view(orig_shape).to(dst_dtype), high_prec_output.view(orig_shape) + + +def test_basic_functionality(): + """Test basic functionality of fake_quantize.""" + print("Test 1: Basic functionality") + + torch.manual_seed(42) + x = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) + + quantized, high_prec = call_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + + # Check shapes + assert quantized.shape == x.shape, f"Shape mismatch: {quantized.shape} vs {x.shape}" + assert high_prec.shape == x.shape, f"Shape mismatch: {high_prec.shape} vs {x.shape}" + + # Check that quantization happened (values should be different) + diff = (quantized.float() - x.float()).abs() + max_diff = diff.max().item() + mean_diff = diff.mean().item() + assert max_diff > 0, "Quantization should change values" + + # Compute cosine similarity between quantized and original + cos_sim = cosine_similarity(quantized, x) + + # Check that high_prec matches original + high_prec_diff = (high_prec - x.float()).abs().max().item() + assert high_prec_diff < 1e-5, f"High precision tensor should match original, got diff: {high_prec_diff}" + + print(f" ✓ Shape check passed") + print(f" ✓ Quantization occurred (max diff: {max_diff:.6f}, mean diff: {mean_diff:.6f}, cosine sim: {cos_sim:.6f})") + print(f" ✓ High precision tensor matches original (diff: {high_prec_diff:.6e})") + + +def test_different_shapes(): + """Test fake_quantize with different tensor shapes.""" + print("\nTest 2: Different shapes") + + torch.manual_seed(42) + test_shapes = [ + (64, 64), + (128, 128), + (256, 128), + (128, 256), + (32, 64), + ] + + for shape in test_shapes: + x = torch.randn(shape, device=DEVICE, dtype=torch.float16) + + # Use block sizes that divide evenly + block_out = min(128, shape[0]) + block_quant = min(128, shape[1]) + + # Ensure block_quant is multiple of 16 + block_quant = (block_quant // 16) * 16 + if block_quant == 0: + block_quant = 16 + + quantized, high_prec = call_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=block_out, + BLOCK_SIZE_QUANT_DIM=block_quant + ) + + assert quantized.shape == x.shape, f"Shape mismatch for {shape}" + assert high_prec.shape == x.shape, f"Shape mismatch for {shape}" + + print(f" ✓ Shape {shape} passed") + + +def test_different_dtypes(): + """Test fake_quantize with different input dtypes.""" + print("\nTest 3: Different dtypes") + + torch.manual_seed(42) + shape = (128, 64) + dtypes = [torch.float16, torch.bfloat16, torch.float32] + + for dtype in dtypes: + x = torch.randn(shape, device=DEVICE, dtype=dtype) + + quantized, high_prec = call_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=128, + BLOCK_SIZE_QUANT_DIM=64, + dst_dtype=dtype + ) + + assert quantized.dtype == dtype, f"Dtype mismatch: {quantized.dtype} vs {dtype}" + assert quantized.shape == x.shape, f"Shape mismatch for dtype {dtype}" + + print(f" ✓ Dtype {dtype} passed") + + +def test_edge_cases(): + """Test fake_quantize with edge cases.""" + print("\nTest 4: Edge cases") + + # Test with zeros + print(" Testing zeros...") + x_zeros = torch.zeros(128, 64, device=DEVICE, dtype=torch.float16) + quantized, high_prec = call_fake_quantize(x_zeros, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + assert quantized.shape == x_zeros.shape + # Quantized zeros should still be close to zero + assert quantized.abs().max().item() < 1e-3, "Quantized zeros should be close to zero" + print(" ✓ Zeros passed") + + # Test with ones (FAIL) + print(" Testing ones...") + x_ones = torch.ones(128, 64, device=DEVICE, dtype=torch.float16) + print(f"x_ones: {x_ones}") + quantized, high_prec = call_fake_quantize(x_ones, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + print(f"quantized: {quantized}") + print(f"high_prec: {high_prec}") + assert quantized.shape == x_ones.shape + # Quantized ones should be close to 1.0 + diff = (quantized - x_ones).abs().max().item() + assert diff < 0.1, f"Quantized ones should be close to 1.0, got max diff: {diff}" + print(f" ✓ Ones passed (max diff: {diff:.6f})") + + # Test with small values + print(" Testing small values...") + x_small = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) * 0.001 + quantized, high_prec = call_fake_quantize(x_small, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + assert quantized.shape == x_small.shape + print(" ✓ Small values passed") + + # Test with large values + print(" Testing large values...") + x_large = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) * 100.0 + quantized, high_prec = call_fake_quantize(x_large, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + assert quantized.shape == x_large.shape + print(" ✓ Large values passed") + + +def test_fp4_quantization(): + """Test fake_quantize with FP4 quantization (uint8).""" + print("\nTest 5: FP4 quantization (uint8)") + + torch.manual_seed(42) + x = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) + + quantized, high_prec = call_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=128, + BLOCK_SIZE_QUANT_DIM=64, + mx_tensor_dtype=tl.uint8 # FP4 uses uint8 + ) + + assert quantized.shape == x.shape + assert high_prec.shape == x.shape + + # Check that quantization happened + diff = (quantized.float() - x.float()).abs() + max_diff = diff.max().item() + mean_diff = diff.mean().item() + assert max_diff > 0, "Quantization should change values" + + # Compute cosine similarity between quantized and original + cos_sim = cosine_similarity(quantized, x) + + print(f" ✓ FP4 quantization passed (max diff: {max_diff:.6f}, mean diff: {mean_diff:.6f}, cosine sim: {cos_sim:.6f})") + + + +def test_finite_values(): + """Test that fake_quantize produces finite values.""" + print("\nTest 7: Finite values check") + + torch.manual_seed(42) + x = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) + + quantized, high_prec = call_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + + # Check for NaN and Inf + assert not torch.isnan(quantized).any(), "Quantized tensor contains NaN" + assert not torch.isinf(quantized).any(), "Quantized tensor contains Inf" + assert not torch.isnan(high_prec).any(), "High precision tensor contains NaN" + assert not torch.isinf(high_prec).any(), "High precision tensor contains Inf" + + print(" ✓ All values are finite") + + +def test_different_block_sizes(): + """Test fake_quantize with different block sizes.""" + print("\nTest 8: Different block sizes") + + torch.manual_seed(42) + x = torch.randn(256, 128, device=DEVICE, dtype=torch.float16) + + block_configs = [ + (64, 64), + (128, 64), + (64, 128), + (128, 128), + ] + + for block_out, block_quant in block_configs: + quantized, high_prec = call_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=block_out, + BLOCK_SIZE_QUANT_DIM=block_quant + ) + + assert quantized.shape == x.shape + assert high_prec.shape == x.shape + + print(f" ✓ Block size ({block_out}, {block_quant}) passed") + + +# ============================================================================ +# MXFP Kernel Tests +# ============================================================================ + +@triton.jit +def mxfp_fake_quantize_kernel( + src_ptr, + dst_ptr, + high_prec_ptr, + BLOCK_SIZE_OUT_DIM: tl.constexpr, + BLOCK_SIZE_QUANT_DIM: tl.constexpr, + dst_dtype: tl.constexpr, + mx_tensor_dtype: tl.constexpr, + stride_src_outer, + stride_src_quant, + stride_dst_outer, + stride_dst_quant, + stride_high_prec_outer, + stride_high_prec_quant, + outer_dim, + quant_dim, +): + """ + Kernel wrapper to test MXFP fake_quantize using _downcast_to_mxfp and _upcast_from_mxfp. + """ + outer_idx = tl.program_id(0) + quant_idx = tl.program_id(1) + + # Determine if we are dealing with fp4 types + is_fp4: tl.constexpr = mx_tensor_dtype == tl.uint8 + K_DIVISOR: tl.constexpr = 2 if is_fp4 else 1 + BLOCK_SIZE_QUANT_MX_SCALE: tl.constexpr = BLOCK_SIZE_QUANT_DIM // MXFP_BLOCK_SIZE + BLOCK_SIZE_QUANT_MX_TENSOR: tl.constexpr = BLOCK_SIZE_QUANT_DIM // K_DIVISOR + + # Allocate temporary storage for quantized tensor and scale + # These need to be stored in shared memory or we need to allocate them + # For simplicity, we'll compute quantize and dequantize in the same kernel + + # Compute offsets + start_outer = outer_idx * BLOCK_SIZE_OUT_DIM + start_quant = quant_idx * BLOCK_SIZE_QUANT_DIM + start_mx_scale_quant = quant_idx * BLOCK_SIZE_QUANT_MX_SCALE + start_mx_quant = quant_idx * BLOCK_SIZE_QUANT_MX_TENSOR + + # Create offset arrays + offs_outer = tl.arange(0, BLOCK_SIZE_OUT_DIM)[:, None] + offs_quant = tl.arange(0, BLOCK_SIZE_QUANT_DIM)[None, :] + offs_mx_scale_quant = tl.arange(0, BLOCK_SIZE_QUANT_MX_SCALE)[None, :] + offs_mx_quant = tl.arange(0, BLOCK_SIZE_QUANT_MX_TENSOR)[None, :] + + # Create masks for valid elements + mask_outer = (start_outer + offs_outer) < outer_dim + mask_quant = (start_quant + offs_quant) < quant_dim + full_mask = mask_outer & mask_quant + + # Load source tensor + src_offsets = (start_outer + offs_outer) * stride_src_outer + (start_quant + offs_quant) * stride_src_quant + src_tensor = tl.load(src_ptr + src_offsets, mask=full_mask, other=0.0) + + # Store high precision tensor (original) + high_prec_offsets = (start_outer + offs_outer) * stride_high_prec_outer + (start_quant + offs_quant) * stride_high_prec_quant + tl.store(high_prec_ptr + high_prec_offsets, src_tensor.to(tl.float32), mask=full_mask) + + # Quantize using _compute_quant_and_scale (same as fake_quantize uses internally) + valid_mask = tl.full(src_tensor.shape, 1, tl.int1) + quantized_tensor, scale_tensor, s_dec = _compute_quant_and_scale( + src_tensor=src_tensor, + valid_src_mask=valid_mask, + mx_tensor_dtype=mx_tensor_dtype + ) + + # Dequantize using _compute_dequant (same as fake_quantize uses internally) + dequantized_tensor = _compute_dequant( + mx_tensor=quantized_tensor, + scale=scale_tensor, + s_dec=s_dec, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=dst_dtype + ) + + # Store result + dst_offsets = (start_outer + offs_outer) * stride_dst_outer + (start_quant + offs_quant) * stride_dst_quant + tl.store(dst_ptr + dst_offsets, dequantized_tensor, mask=full_mask) + + +def call_mxfp_fake_quantize( + x: torch.Tensor, + BLOCK_SIZE_OUT_DIM: int = 128, + BLOCK_SIZE_QUANT_DIM: int = 128, + dst_dtype: torch.dtype = None, + mx_tensor_dtype = tl.uint8 # Default to FP4 (uint8) +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Call MXFP fake_quantize on a tensor using _downcast_to_mxfp and _upcast_from_mxfp logic. + + Args: + x: Input tensor (2D or can be reshaped to 2D) + BLOCK_SIZE_OUT_DIM: Block size for outer dimension + BLOCK_SIZE_QUANT_DIM: Block size for quantization dimension (must be multiple of 16) + dst_dtype: Output dtype (defaults to input dtype) + mx_tensor_dtype: Quantization dtype (tl.uint8 for fp4 or tl.float8e4nv for fp8) + + Returns: + Tuple of (quantized_tensor, high_precision_tensor) + """ + assert x.is_cuda, "Input must be on CUDA" + assert BLOCK_SIZE_QUANT_DIM % 16 == 0, f"BLOCK_SIZE_QUANT_DIM must be multiple of 16, got {BLOCK_SIZE_QUANT_DIM}" + + orig_shape = x.shape + orig_dtype = x.dtype + + # Reshape to 2D + x_2d = x.view(-1, x.shape[-1]) + outer_dim, quant_dim = x_2d.shape + + if dst_dtype is None: + dst_dtype = orig_dtype + + # Map torch dtype to triton dtype + dtype_map = { + torch.float32: tl.float32, + torch.float16: tl.float16, + torch.bfloat16: tl.bfloat16, + } + triton_dst_dtype = dtype_map.get(dst_dtype, tl.float16) + + # Allocate outputs + quantized_output = torch.empty_like(x_2d, dtype=dst_dtype) + high_prec_output = torch.empty_like(x_2d, dtype=torch.float32) + + # Launch kernel + grid = ( + triton.cdiv(outer_dim, BLOCK_SIZE_OUT_DIM), + triton.cdiv(quant_dim, BLOCK_SIZE_QUANT_DIM), + ) + + mxfp_fake_quantize_kernel[grid]( + src_ptr=x_2d, + dst_ptr=quantized_output, + high_prec_ptr=high_prec_output, + BLOCK_SIZE_OUT_DIM=BLOCK_SIZE_OUT_DIM, + BLOCK_SIZE_QUANT_DIM=BLOCK_SIZE_QUANT_DIM, + dst_dtype=triton_dst_dtype, + mx_tensor_dtype=mx_tensor_dtype, + stride_src_outer=x_2d.stride(0), + stride_src_quant=x_2d.stride(1), + stride_dst_outer=quantized_output.stride(0), + stride_dst_quant=quantized_output.stride(1), + stride_high_prec_outer=high_prec_output.stride(0), + stride_high_prec_quant=high_prec_output.stride(1), + outer_dim=outer_dim, + quant_dim=quant_dim, + ) + + return quantized_output.view(orig_shape).to(dst_dtype), high_prec_output.view(orig_shape) + + +def test_mxfp_basic_functionality(): + """Test basic functionality of MXFP fake_quantize.""" + print("\n" + "=" * 60) + print("MXFP Kernel Tests") + print("=" * 60) + print("\nTest 1: Basic functionality (MXFP)") + + torch.manual_seed(42) + x = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) + + quantized, high_prec = call_mxfp_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + + # Check shapes + assert quantized.shape == x.shape, f"Shape mismatch: {quantized.shape} vs {x.shape}" + assert high_prec.shape == x.shape, f"Shape mismatch: {high_prec.shape} vs {x.shape}" + + # Check that quantization happened (values should be different) + diff = (quantized.float() - x.float()).abs() + max_diff = diff.max().item() + mean_diff = diff.mean().item() + assert max_diff > 0, "Quantization should change values" + + # Compute cosine similarity between quantized and original + cos_sim = cosine_similarity(quantized, x) + + # Check that high_prec matches original + high_prec_diff = (high_prec - x.float()).abs().max().item() + assert high_prec_diff < 1e-5, f"High precision tensor should match original, got diff: {high_prec_diff}" + + print(f" ✓ Shape check passed") + print(f" ✓ Quantization occurred (max diff: {max_diff:.6f}, mean diff: {mean_diff:.6f}, cosine sim: {cos_sim:.6f})") + print(f" ✓ High precision tensor matches original (diff: {high_prec_diff:.6e})") + + +def test_mxfp_different_shapes(): + """Test MXFP fake_quantize with different tensor shapes.""" + print("\nTest 2: Different shapes (MXFP)") + + torch.manual_seed(42) + test_shapes = [ + (64, 64), + (128, 128), + (256, 128), + (128, 256), + (32, 64), + ] + + for shape in test_shapes: + x = torch.randn(shape, device=DEVICE, dtype=torch.float16) + + # Use block sizes that divide evenly + block_out = min(128, shape[0]) + block_quant = min(128, shape[1]) + + # Ensure block_quant is multiple of 16 + block_quant = (block_quant // 16) * 16 + if block_quant == 0: + block_quant = 16 + + quantized, high_prec = call_mxfp_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=block_out, + BLOCK_SIZE_QUANT_DIM=block_quant + ) + + assert quantized.shape == x.shape, f"Shape mismatch for {shape}" + assert high_prec.shape == x.shape, f"Shape mismatch for {shape}" + + print(f" ✓ Shape {shape} passed") + + +def test_mxfp_different_dtypes(): + """Test MXFP fake_quantize with different input dtypes.""" + print("\nTest 3: Different dtypes (MXFP)") + + torch.manual_seed(42) + shape = (128, 64) + dtypes = [torch.float16, torch.bfloat16, torch.float32] + + for dtype in dtypes: + x = torch.randn(shape, device=DEVICE, dtype=dtype) + + quantized, high_prec = call_mxfp_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=128, + BLOCK_SIZE_QUANT_DIM=64, + dst_dtype=dtype + ) + + assert quantized.dtype == dtype, f"Dtype mismatch: {quantized.dtype} vs {dtype}" + assert quantized.shape == x.shape, f"Shape mismatch for dtype {dtype}" + + print(f" ✓ Dtype {dtype} passed") + + +def test_mxfp_edge_cases(): + """Test MXFP fake_quantize with edge cases.""" + print("\nTest 4: Edge cases (MXFP)") + + # Test with zeros + print(" Testing zeros...") + x_zeros = torch.zeros(128, 64, device=DEVICE, dtype=torch.float16) + quantized, high_prec = call_mxfp_fake_quantize(x_zeros, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + assert quantized.shape == x_zeros.shape + # Quantized zeros should still be close to zero + assert quantized.abs().max().item() < 1e-3, "Quantized zeros should be close to zero" + print(" ✓ Zeros passed") + + # Test with ones + print(" Testing ones...") + x_ones = torch.ones(128, 64, device=DEVICE, dtype=torch.float16) + quantized, high_prec = call_mxfp_fake_quantize(x_ones, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + assert quantized.shape == x_ones.shape + # Quantized ones should be close to 1.0 + diff = (quantized - x_ones).abs().max().item() + assert diff < 0.1, f"Quantized ones should be close to 1.0, got max diff: {diff}" + print(f" ✓ Ones passed (max diff: {diff:.6f})") + + # Test with small values + print(" Testing small values...") + x_small = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) * 0.001 + quantized, high_prec = call_mxfp_fake_quantize(x_small, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + assert quantized.shape == x_small.shape + print(" ✓ Small values passed") + + # Test with large values + print(" Testing large values...") + x_large = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) * 100.0 + quantized, high_prec = call_mxfp_fake_quantize(x_large, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + assert quantized.shape == x_large.shape + print(" ✓ Large values passed") + + +def test_mxfp_fp4_quantization(): + """Test MXFP fake_quantize with FP4 quantization (uint8).""" + print("\nTest 5: FP4 quantization (MXFP, uint8)") + + torch.manual_seed(42) + x = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) + + quantized, high_prec = call_mxfp_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=128, + BLOCK_SIZE_QUANT_DIM=64, + mx_tensor_dtype=tl.uint8 # FP4 uses uint8 + ) + + assert quantized.shape == x.shape + assert high_prec.shape == x.shape + + # Check that quantization happened + diff = (quantized.float() - x.float()).abs() + max_diff = diff.max().item() + mean_diff = diff.mean().item() + assert max_diff > 0, "Quantization should change values" + + # Compute cosine similarity between quantized and original + cos_sim = cosine_similarity(quantized, x) + + print(f" ✓ FP4 quantization passed (max diff: {max_diff:.6f}, mean diff: {mean_diff:.6f}, cosine sim: {cos_sim:.6f})") + + +def test_mxfp_fp8_quantization(): + """Test MXFP fake_quantize with FP8 quantization (float8e4nv).""" + print("\nTest 6: FP8 quantization (MXFP, float8e4nv)") + + torch.manual_seed(42) + x = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) + + quantized, high_prec = call_mxfp_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=128, + BLOCK_SIZE_QUANT_DIM=64, + mx_tensor_dtype=tl.float8e4nv # FP8 uses float8e4nv + ) + + assert quantized.shape == x.shape + assert high_prec.shape == x.shape + + # Check that quantization happened + diff = (quantized.float() - x.float()).abs() + max_diff = diff.max().item() + mean_diff = diff.mean().item() + assert max_diff > 0, "Quantization should change values" + + # Compute cosine similarity between quantized and original + cos_sim = cosine_similarity(quantized, x) + + print(f" ✓ FP8 quantization passed (max diff: {max_diff:.6f}, mean diff: {mean_diff:.6f}, cosine sim: {cos_sim:.6f})") + + +def test_mxfp_finite_values(): + """Test that MXFP fake_quantize produces finite values.""" + print("\nTest 7: Finite values check (MXFP)") + + torch.manual_seed(42) + x = torch.randn(128, 64, device=DEVICE, dtype=torch.float16) + + quantized, high_prec = call_mxfp_fake_quantize(x, BLOCK_SIZE_OUT_DIM=128, BLOCK_SIZE_QUANT_DIM=64) + + # Check for NaN and Inf + assert not torch.isnan(quantized).any(), "Quantized tensor contains NaN" + assert not torch.isinf(quantized).any(), "Quantized tensor contains Inf" + assert not torch.isnan(high_prec).any(), "High precision tensor contains NaN" + assert not torch.isinf(high_prec).any(), "High precision tensor contains Inf" + + print(" ✓ All values are finite") + + +def test_mxfp_different_block_sizes(): + """Test MXFP fake_quantize with different block sizes.""" + print("\nTest 8: Different block sizes (MXFP)") + + torch.manual_seed(42) + x = torch.randn(256, 128, device=DEVICE, dtype=torch.float16) + + block_configs = [ + (64, 64), + (128, 64), + (64, 128), + (128, 128), + ] + + for block_out, block_quant in block_configs: + quantized, high_prec = call_mxfp_fake_quantize( + x, + BLOCK_SIZE_OUT_DIM=block_out, + BLOCK_SIZE_QUANT_DIM=block_quant + ) + + assert quantized.shape == x.shape + assert high_prec.shape == x.shape + + print(f" ✓ Block size ({block_out}, {block_quant}) passed") + + +def test_get_fake_quant_metrics(): + """Test get_fake_quant function with cosine similarity, mean_diff, and max_diff metrics.""" + print("\n" + "=" * 60) + print("Testing get_fake_quant with cosine similarity, mean_diff, and max_diff") + print("=" * 60) + + if not torch.cuda.is_available(): + print("CUDA not available, skipping test") + return + + device = torch.device("cuda") + torch.manual_seed(42) + + # Test with different shapes and dtypes + test_cases = [ + ((128, 64), torch.bfloat16, "2D bfloat16"), + ((256, 128), torch.bfloat16, "2D bfloat16 large"), + ((2, 8, 128), torch.bfloat16, "3D bfloat16"), + ((128, 64), torch.float16, "2D float16"), + ((1, 8, 256, 128), torch.bfloat16, "4D bfloat16"), + ] + + for shape, dtype, name in test_cases: + print(f"\nTest: {name} - Shape: {shape}") + x = torch.randn(*shape, device=device, dtype=dtype) + x_fq = get_fake_quant(x) + + # Check shapes match + assert x_fq.shape == x.shape, f"Shape mismatch: {x_fq.shape} vs {x.shape}" + assert x_fq.dtype == x.dtype, f"Dtype mismatch: {x_fq.dtype} vs {x.dtype}" + + # Compute metrics + diff = (x_fq.float() - x.float()).abs() + max_diff = diff.max().item() + mean_diff = diff.mean().item() + cos_sim = cosine_similarity(x_fq, x) + + print(f" Max diff: {max_diff:.6f}") + print(f" Mean diff: {mean_diff:.6f}") + print(f" Cosine similarity: {cos_sim:.6f}") + + # Basic sanity checks + assert max_diff < 10.0, f"Max difference too large: {max_diff}" + assert cos_sim > 0.5, f"Cosine similarity too low: {cos_sim}" + + print("\n✓ All metric tests passed!") + + +if __name__ == "__main__": + if not torch.cuda.is_available(): + print("CUDA not available. These tests require CUDA.") + exit(1) + + print("=" * 60) + print("Running basic tests for fake_quantize function") + print("=" * 60) + print(f"Device: {DEVICE}") + print() + + try: + test_basic_functionality() + test_different_shapes() + test_different_dtypes() + test_edge_cases() + test_fp4_quantization() + test_finite_values() + + # Test get_fake_quant with metrics + test_get_fake_quant_metrics() + + # Run MXFP kernel tests + # test_mxfp_basic_functionality() + # test_mxfp_different_shapes() + # test_mxfp_different_dtypes() + # test_mxfp_edge_cases() + # test_mxfp_fp4_quantization() + # test_mxfp_fp8_quantization() + # test_mxfp_finite_values() + # test_mxfp_different_block_sizes() + + print("\n" + "=" * 60) + print("All tests passed! ✓") + print("=" * 60) + except Exception as e: + print(f"\n❌ Test failed with error: {e}") + import traceback + traceback.print_exc() + exit(1) + + diff --git a/test_fake_quantize_kernels.py b/test_fake_quantize_kernels.py new file mode 100644 index 000000000..01f991e14 --- /dev/null +++ b/test_fake_quantize_kernels.py @@ -0,0 +1,452 @@ +#!/usr/bin/env python3 +""" +Precision test for fake_quantize_q and fake_quantize_kv kernels. + +Tests that the kernels correctly: +1. Quantize and dequantize Q, K, V tensors +2. Preserve approximate values within quantization error +3. Handle edge cases (zeros, small/large values, padding) +4. Work with different shapes and dtypes +""" + +import torch +import triton +import triton.language as tl +import numpy as np +from typing import Tuple, Dict, Any +import unittest + +from quant_utils import fake_quantize_q, fake_quantize_kv + + +def cosine_similarity(tensor1: torch.Tensor, tensor2: torch.Tensor) -> float: + """Compute cosine similarity between two tensors.""" + t1_flat = tensor1.flatten().float() + t2_flat = tensor2.flatten().float() + + dot_product = torch.dot(t1_flat, t2_flat) + norm1 = torch.norm(t1_flat) + norm2 = torch.norm(t2_flat) + + if norm1 == 0 or norm2 == 0: + return 0.0 + + return (dot_product / (norm1 * norm2)).item() + + +def compute_metrics(original: torch.Tensor, quantized: torch.Tensor, mask: torch.Tensor = None) -> Dict[str, float]: + """Compute precision metrics between original and quantized tensors.""" + if mask is not None: + # Only compute metrics on valid (non-padding) elements + original = original[mask] + quantized = quantized[mask] + + diff = (original.float() - quantized.float()).abs() + + metrics = { + 'max_diff': diff.max().item(), + 'mean_diff': diff.mean().item(), + 'std_diff': diff.std().item(), + 'cosine_similarity': cosine_similarity(original, quantized), + } + + # Relative error (avoid division by zero) + abs_original = original.float().abs() + rel_error = diff / (abs_original + 1e-8) + metrics['max_relative_error'] = rel_error.max().item() + metrics['mean_relative_error'] = rel_error.mean().item() + + return metrics + + +def test_fake_quantize_q( + q: torch.Tensor, + BLOCK_M: int = 128, + HEAD_DIM: int = 128, + verbose: bool = False +) -> Tuple[torch.Tensor, Dict[str, Any]]: + """ + Test fake_quantize_q kernel. + + Args: + q: Input Q tensor with shape (Z, H, N_CTX_Q, HEAD_DIM) + BLOCK_M: Block size for sequence dimension + HEAD_DIM: Head dimension (must be multiple of 16) + verbose: Whether to print detailed information + + Returns: + Tuple of (fake_q, metrics_dict) + """ + assert q.is_cuda, "Input must be on CUDA" + assert q.dim() == 4, f"Q must be 4D, got shape {q.shape}" + assert HEAD_DIM % 16 == 0, f"HEAD_DIM must be multiple of 16, got {HEAD_DIM}" + + Z, H, N_CTX_Q, HEAD_DIM_actual = q.shape + assert HEAD_DIM_actual == HEAD_DIM, f"HEAD_DIM mismatch: {HEAD_DIM_actual} != {HEAD_DIM}" + + # Create output tensor + fake_q = torch.empty_like(q) + + # Launch kernel + grid = (triton.cdiv(N_CTX_Q, BLOCK_M), Z * H, 1) + + fake_quantize_q[grid]( + q, fake_q, + q.stride(0), q.stride(1), # stride_z_q, stride_h_q + q.stride(2), q.stride(3), # stride_tok_q, stride_d_q + fake_q.stride(0), fake_q.stride(1), # fake_stride_z_q, fake_stride_h_q + fake_q.stride(2), fake_q.stride(3), # fake_stride_tok_q, fake_stride_d_q + H, N_CTX_Q, + BLOCK_M=BLOCK_M, HEAD_DIM=HEAD_DIM + ) + + # Compute metrics + metrics = compute_metrics(q, fake_q) + metrics['shape'] = list(q.shape) + metrics['dtype'] = str(q.dtype) + + if verbose: + print(f"fake_quantize_q test:") + print(f" Shape: {q.shape}") + print(f" Dtype: {q.dtype}") + print(f" Max diff: {metrics['max_diff']:.6f}") + print(f" Mean diff: {metrics['mean_diff']:.6f}") + print(f" Cosine similarity: {metrics['cosine_similarity']:.6f}") + print(f" Max relative error: {metrics['max_relative_error']:.6f}") + + return fake_q, metrics + + +def test_fake_quantize_kv( + k: torch.Tensor, + v: torch.Tensor, + BLOCK_N: int = 128, + HEAD_DIM: int = 128, + verbose: bool = False +) -> Tuple[torch.Tensor, torch.Tensor, Dict[str, Any]]: + """ + Test fake_quantize_kv kernel. + + Args: + k: Input K tensor with shape (Z, H, N_CTX_KV, HEAD_DIM) + v: Input V tensor with shape (Z, H, N_CTX_KV, HEAD_DIM) + BLOCK_N: Block size for sequence dimension + HEAD_DIM: Head dimension (must be multiple of 16) + verbose: Whether to print detailed information + + Returns: + Tuple of (fake_k, fake_v, metrics_dict) + """ + assert k.is_cuda and v.is_cuda, "Inputs must be on CUDA" + assert k.shape == v.shape, f"K and V must have same shape: {k.shape} != {v.shape}" + assert k.dim() == 4, f"K must be 4D, got shape {k.shape}" + assert HEAD_DIM % 16 == 0, f"HEAD_DIM must be multiple of 16, got {HEAD_DIM}" + + Z, H, N_CTX_KV, HEAD_DIM_actual = k.shape + assert HEAD_DIM_actual == HEAD_DIM, f"HEAD_DIM mismatch: {HEAD_DIM_actual} != {HEAD_DIM}" + + # Create output tensors + fake_k = torch.empty_like(k) + fake_v = torch.empty_like(v) + + # Launch kernel + grid = (triton.cdiv(N_CTX_KV, BLOCK_N), Z * H, 1) + + fake_quantize_kv[grid]( + k, v, fake_k, fake_v, + k.stride(0), k.stride(1), # stride_z_kv, stride_h_kv + k.stride(2), k.stride(3), # stride_tok_kv, stride_d_kv + fake_k.stride(0), fake_k.stride(1), # fake_stride_z_kv, fake_stride_h_kv + fake_k.stride(2), fake_k.stride(3), # fake_stride_tok_kv, fake_stride_d_kv + H, N_CTX_KV, + BLOCK_N=BLOCK_N, HEAD_DIM=HEAD_DIM + ) + + # Compute metrics for both K and V + metrics_k = compute_metrics(k, fake_k) + metrics_v = compute_metrics(v, fake_v) + + metrics = { + 'k': metrics_k, + 'v': metrics_v, + 'shape': list(k.shape), + 'dtype': str(k.dtype), + } + + if verbose: + print(f"fake_quantize_kv test:") + print(f" Shape: {k.shape}") + print(f" Dtype: {k.dtype}") + print(f" K - Max diff: {metrics_k['max_diff']:.6f}, Cosine sim: {metrics_k['cosine_similarity']:.6f}") + print(f" V - Max diff: {metrics_v['max_diff']:.6f}, Cosine sim: {metrics_v['cosine_similarity']:.6f}") + + return fake_k, fake_v, metrics + + +class FakeQuantizeKernelsTest: + """Test suite for fake_quantize_q and fake_quantize_kv kernels.""" + + def __init__(self, device: str = "cuda", dtype: torch.dtype = torch.bfloat16): + self.device = device + self.dtype = dtype + self.HEAD_DIM = 128 + self.BLOCK_M = 128 + self.BLOCK_N = 128 + + def test_basic_q(self, verbose: bool = False) -> Dict[str, Any]: + """Test fake_quantize_q with basic random input.""" + print("\n" + "="*60) + print("Test: Basic Q quantization") + print("="*60) + + Z, H, N_CTX_Q = 2, 8, 256 + q = torch.randn(Z, H, N_CTX_Q, self.HEAD_DIM, device=self.device, dtype=self.dtype) + + fake_q, metrics = test_fake_quantize_q(q, BLOCK_M=self.BLOCK_M, HEAD_DIM=self.HEAD_DIM, verbose=verbose) + + # Assertions + assert fake_q.shape == q.shape, f"Shape mismatch: {fake_q.shape} != {q.shape}" + assert fake_q.dtype == q.dtype, f"Dtype mismatch: {fake_q.dtype} != {q.dtype}" + assert metrics['cosine_similarity'] > 0.9, f"Cosine similarity too low: {metrics['cosine_similarity']}" + + print("✓ Basic Q test passed") + return metrics + + def test_basic_kv(self, verbose: bool = False) -> Dict[str, Any]: + """Test fake_quantize_kv with basic random inputs.""" + print("\n" + "="*60) + print("Test: Basic KV quantization") + print("="*60) + + Z, H, N_CTX_KV = 2, 8, 512 + k = torch.randn(Z, H, N_CTX_KV, self.HEAD_DIM, device=self.device, dtype=self.dtype) + v = torch.randn(Z, H, N_CTX_KV, self.HEAD_DIM, device=self.device, dtype=self.dtype) + + fake_k, fake_v, metrics = test_fake_quantize_kv( + k, v, BLOCK_N=self.BLOCK_N, HEAD_DIM=self.HEAD_DIM, verbose=verbose + ) + + # Assertions + assert fake_k.shape == k.shape, f"K shape mismatch: {fake_k.shape} != {k.shape}" + assert fake_v.shape == v.shape, f"V shape mismatch: {fake_v.shape} != {v.shape}" + assert metrics['k']['cosine_similarity'] > 0.9, f"K cosine similarity too low: {metrics['k']['cosine_similarity']}" + assert metrics['v']['cosine_similarity'] > 0.9, f"V cosine similarity too low: {metrics['v']['cosine_similarity']}" + + print("✓ Basic KV test passed") + return metrics + + def test_small_values(self, verbose: bool = False) -> Dict[str, Any]: + """Test with small input values.""" + print("\n" + "="*60) + print("Test: Small values") + print("="*60) + + Z, H, N_CTX_Q = 1, 4, 128 + q = torch.randn(Z, H, N_CTX_Q, self.HEAD_DIM, device=self.device, dtype=self.dtype) * 0.01 + + fake_q, metrics = test_fake_quantize_q(q, BLOCK_M=self.BLOCK_M, HEAD_DIM=self.HEAD_DIM, verbose=verbose) + + print("✓ Small values test passed") + return metrics + + def test_large_values(self, verbose: bool = False) -> Dict[str, Any]: + """Test with large input values.""" + print("\n" + "="*60) + print("Test: Large values") + print("="*60) + + Z, H, N_CTX_Q = 1, 4, 128 + q = torch.randn(Z, H, N_CTX_Q, self.HEAD_DIM, device=self.device, dtype=self.dtype) * 10.0 + + fake_q, metrics = test_fake_quantize_q(q, BLOCK_M=self.BLOCK_M, HEAD_DIM=self.HEAD_DIM, verbose=verbose) + + print("✓ Large values test passed") + return metrics + + def test_zeros(self, verbose: bool = False) -> Dict[str, Any]: + """Test with zero input.""" + print("\n" + "="*60) + print("Test: Zeros") + print("="*60) + + Z, H, N_CTX_Q = 1, 4, 128 + q = torch.zeros(Z, H, N_CTX_Q, self.HEAD_DIM, device=self.device, dtype=self.dtype) + + fake_q, metrics = test_fake_quantize_q(q, BLOCK_M=self.BLOCK_M, HEAD_DIM=self.HEAD_DIM, verbose=verbose) + + # Zeros should remain approximately zero after quantization + max_val = fake_q.abs().max().item() + assert max_val < 1e-3, f"Zeros should remain near zero, got max: {max_val}" + + print("✓ Zeros test passed") + return metrics + + def test_different_shapes(self, verbose: bool = False) -> Dict[str, Any]: + """Test with different tensor shapes.""" + print("\n" + "="*60) + print("Test: Different shapes") + print("="*60) + + test_shapes = [ + (1, 1, 64, self.HEAD_DIM), # Single batch, single head, small seq + (1, 8, 128, self.HEAD_DIM), # Single batch, multiple heads + (4, 16, 256, self.HEAD_DIM), # Multiple batches, many heads + (2, 8, 512, self.HEAD_DIM), # Medium size + ] + + all_metrics = [] + for shape in test_shapes: + Z, H, N_CTX_Q, HEAD_DIM = shape + q = torch.randn(*shape, device=self.device, dtype=self.dtype) + + fake_q, metrics = test_fake_quantize_q( + q, BLOCK_M=self.BLOCK_M, HEAD_DIM=self.HEAD_DIM, verbose=verbose + ) + + assert fake_q.shape == q.shape + assert metrics['cosine_similarity'] > 0.85 + all_metrics.append(metrics) + + if verbose: + print(f" Shape {shape}: cosine_sim={metrics['cosine_similarity']:.4f}") + + print("✓ Different shapes test passed") + return all_metrics + + def test_different_dtypes(self, verbose: bool = False) -> Dict[str, Any]: + """Test with different dtypes.""" + print("\n" + "="*60) + print("Test: Different dtypes") + print("="*60) + + Z, H, N_CTX_Q = 1, 4, 128 + dtypes = [torch.bfloat16, torch.float16, torch.float32] + + all_metrics = [] + for dtype in dtypes: + q = torch.randn(Z, H, N_CTX_Q, self.HEAD_DIM, device=self.device, dtype=dtype) + + fake_q, metrics = test_fake_quantize_q( + q, BLOCK_M=self.BLOCK_M, HEAD_DIM=self.HEAD_DIM, verbose=verbose + ) + + assert fake_q.dtype == q.dtype + assert metrics['cosine_similarity'] > 0.85 + all_metrics.append(metrics) + + if verbose: + print(f" {dtype}: cosine_sim={metrics['cosine_similarity']:.4f}") + + print("✓ Different dtypes test passed") + return all_metrics + + def test_non_divisible_sequence_length(self, verbose: bool = False) -> Dict[str, Any]: + """Test with sequence lengths that don't divide evenly by block size.""" + print("\n" + "="*60) + print("Test: Non-divisible sequence length") + print("="*60) + + Z, H = 1, 4 + # Use sequence length that doesn't divide evenly by BLOCK_M + N_CTX_Q = 150 # Not divisible by 128 + + q = torch.randn(Z, H, N_CTX_Q, self.HEAD_DIM, device=self.device, dtype=self.dtype) + + fake_q, metrics = test_fake_quantize_q(q, BLOCK_M=self.BLOCK_M, HEAD_DIM=self.HEAD_DIM, verbose=verbose) + + # Check that padding is handled correctly (last block should have zeros for padding) + assert fake_q.shape == q.shape + assert metrics['cosine_similarity'] > 0.85 + + print("✓ Non-divisible sequence length test passed") + return metrics + + def test_kv_different_shapes(self, verbose: bool = False) -> Dict[str, Any]: + """Test fake_quantize_kv with different shapes.""" + print("\n" + "="*60) + print("Test: KV different shapes") + print("="*60) + + test_shapes = [ + (1, 1, 64, self.HEAD_DIM), + (1, 8, 128, self.HEAD_DIM), + (2, 16, 256, self.HEAD_DIM), + ] + + all_metrics = [] + for shape in test_shapes: + Z, H, N_CTX_KV, HEAD_DIM = shape + k = torch.randn(*shape, device=self.device, dtype=self.dtype) + v = torch.randn(*shape, device=self.device, dtype=self.dtype) + + fake_k, fake_v, metrics = test_fake_quantize_kv( + k, v, BLOCK_N=self.BLOCK_N, HEAD_DIM=self.HEAD_DIM, verbose=verbose + ) + + assert fake_k.shape == k.shape + assert fake_v.shape == v.shape + assert metrics['k']['cosine_similarity'] > 0.85 + assert metrics['v']['cosine_similarity'] > 0.85 + all_metrics.append(metrics) + + print("✓ KV different shapes test passed") + return all_metrics + + def run_all_tests(self, verbose: bool = False) -> Dict[str, Any]: + """Run all tests.""" + print("\n" + "="*60) + print("Running Fake Quantize Kernels Precision Tests") + print("="*60) + + if not torch.cuda.is_available(): + print("CUDA not available, skipping tests") + return {} + + results = {} + + try: + results['basic_q'] = self.test_basic_q(verbose=verbose) + results['basic_kv'] = self.test_basic_kv(verbose=verbose) + results['small_values'] = self.test_small_values(verbose=verbose) + results['large_values'] = self.test_large_values(verbose=verbose) + results['zeros'] = self.test_zeros(verbose=verbose) + results['different_shapes'] = self.test_different_shapes(verbose=verbose) + results['different_dtypes'] = self.test_different_dtypes(verbose=verbose) + results['non_divisible_seq'] = self.test_non_divisible_sequence_length(verbose=verbose) + results['kv_different_shapes'] = self.test_kv_different_shapes(verbose=verbose) + + print("\n" + "="*60) + print("✓ All tests passed!") + print("="*60) + + except Exception as e: + print(f"\n✗ Test failed with error: {e}") + import traceback + traceback.print_exc() + raise + + return results + + +def main(): + """Main function to run the precision tests.""" + tester = FakeQuantizeKernelsTest(device="cuda", dtype=torch.bfloat16) + results = tester.run_all_tests(verbose=True) + + # Print summary + print("\n" + "="*60) + print("Test Summary") + print("="*60) + + if 'basic_q' in results: + print(f"Basic Q - Cosine similarity: {results['basic_q']['cosine_similarity']:.6f}") + print(f"Basic Q - Max diff: {results['basic_q']['max_diff']:.6f}") + + if 'basic_kv' in results: + print(f"Basic K - Cosine similarity: {results['basic_kv']['k']['cosine_similarity']:.6f}") + print(f"Basic V - Cosine similarity: {results['basic_kv']['v']['cosine_similarity']:.6f}") + + +if __name__ == "__main__": + main() + diff --git a/test_get_fake_quant.py b/test_get_fake_quant.py new file mode 100644 index 000000000..a8d4fad6f --- /dev/null +++ b/test_get_fake_quant.py @@ -0,0 +1,336 @@ +#!/usr/bin/env python3 +""" +Precision test to compare numerical differences for get_fake_quant Python function: +1. get_fake_quant (Python implementation using FlashInfer) +2. Tests various shapes, dtypes, and value ranges +3. Evaluates cosine similarity, max diff, and mean diff between input and output +""" + +import torch +from test_quantization_precision import get_fake_quant + +DEVICE = torch.device("cuda") + + +def cosine_similarity(tensor1, tensor2): + """ + Compute cosine similarity between two tensors. + + Args: + tensor1: First tensor + tensor2: Second tensor (same shape as tensor1) + + Returns: + Cosine similarity value (scalar) + - Returns 1.0 if both tensors are zero (identical zero vectors) + - Returns 0.0 if only one tensor is zero (orthogonal to non-zero vector) + """ + # Flatten tensors for computation + t1_flat = tensor1.flatten().float() + t2_flat = tensor2.flatten().float() + + # Compute cosine similarity: (A · B) / (||A|| * ||B||) + dot_product = torch.dot(t1_flat, t2_flat) + norm1 = torch.norm(t1_flat) + norm2 = torch.norm(t2_flat) + + # Handle zero vectors + if norm1 == 0 and norm2 == 0: + # Both are zero vectors - they are identical, so similarity is 1.0 + return 1.0 + elif norm1 == 0 or norm2 == 0: + # One is zero, one is not - they are orthogonal, so similarity is 0.0 + return 0.0 + + cos_sim = dot_product / (norm1 * norm2) + return cos_sim.item() + + +def test_get_fake_quant_basic(): + """Test basic functionality of get_fake_quant.""" + torch.manual_seed(42) + + # Test parameters + shape = (128, 128) + dtype = torch.bfloat16 + + # Create input tensor + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Test get_fake_quant + x_fq = get_fake_quant(x) + + # Check that outputs have correct shape + assert x_fq.shape == x.shape + assert x_fq.dtype == x.dtype + + # Check that outputs are finite + assert torch.isfinite(x_fq).all() + + # Compare input and output + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" Input vs Output - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ Basic test passed.") + + +def test_get_fake_quant_different_shapes(): + """Test get_fake_quant with different input shapes.""" + torch.manual_seed(42) + + test_configs = [ + (128, 64), # Small + (256, 128), # Medium + (512, 256), # Large + (128, 256), # Rectangular + (256, 128), # Rectangular (reversed) + ] + + dtype = torch.bfloat16 + + for outer_dim, quant_dim in test_configs: + # Create input tensor + x = torch.randn((outer_dim, quant_dim), dtype=dtype, device=DEVICE) + + # Test get_fake_quant + x_fq = get_fake_quant(x) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + # Compare input and output + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" Shape ({outer_dim}, {quant_dim}) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print(f"✓ Shape test passed for ({outer_dim}, {quant_dim})") + + +def test_get_fake_quant_different_dtypes(): + """Test get_fake_quant with different input dtypes.""" + torch.manual_seed(42) + + shape = (128, 128) + dtypes = [torch.float32, torch.float16, torch.bfloat16] + + for dtype in dtypes: + # Create input tensor + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Test get_fake_quant + x_fq = get_fake_quant(x) + + assert x_fq.shape == x.shape + assert x_fq.dtype == x.dtype + assert torch.isfinite(x_fq).all() + + # Compare input and output + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" Dtype {dtype} - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print(f"✓ Dtype test passed for {dtype}") + + +def test_get_fake_quant_3d_4d_tensors(): + """Test get_fake_quant with 3D and 4D tensors (reshaped to 2D internally).""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + + # Test 3D tensor (B, H, D) + print("\nTesting 3D tensor (B, H, D)") + x_3d = torch.randn((2, 8, 128), dtype=dtype, device=DEVICE) + x_fq_3d = get_fake_quant(x_3d) + + assert x_fq_3d.shape == x_3d.shape + assert torch.isfinite(x_fq_3d).all() + + max_diff = (x_fq_3d - x_3d).abs().max() + mean_diff = (x_fq_3d - x_3d).abs().mean() + cos_sim = cosine_similarity(x_fq_3d, x_3d) + print(f" 3D (2, 8, 128) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Test 4D tensor (B, H, L, D) + print("\nTesting 4D tensor (B, H, L, D)") + x_4d = torch.randn((1, 8, 256, 128), dtype=dtype, device=DEVICE) + x_fq_4d = get_fake_quant(x_4d) + + assert x_fq_4d.shape == x_4d.shape + assert torch.isfinite(x_fq_4d).all() + + max_diff = (x_fq_4d - x_4d).abs().max() + mean_diff = (x_fq_4d - x_4d).abs().mean() + cos_sim = cosine_similarity(x_fq_4d, x_4d) + print(f" 4D (1, 8, 256, 128) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ 3D/4D tensor test passed.") + + +def test_get_fake_quant_edge_cases(): + """Test edge cases like zeros, ones, extreme values.""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + shape = (128, 128) + + test_cases = [ + ("zeros", torch.zeros(shape, dtype=dtype, device=DEVICE)), + ("ones", torch.ones(shape, dtype=dtype, device=DEVICE)), + ("negative_ones", -torch.ones(shape, dtype=dtype, device=DEVICE)), + ("very_small", torch.randn(shape, dtype=dtype, device=DEVICE) * 1e-6), + ("very_large", torch.randn(shape, dtype=dtype, device=DEVICE) * 1e6), + ] + + for name, x in test_cases: + x_fq = get_fake_quant(x) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" {name} - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ Edge cases test passed.") + + +def test_get_fake_quant_attention_shapes(): + """Test get_fake_quant with attention-like shapes (similar to test_qat_attn.py).""" + torch.manual_seed(42) + + test_configs = [ + (2, 4, 128, 64), # Z, H, N_CTX, HEAD_DIM - Medium + (1, 8, 256, 128), # Large head dim + (1, 40, 9360, 128), # WAN shape + ] + + dtype = torch.bfloat16 + + for Z, H, N_CTX, HEAD_DIM in test_configs: + # Create input tensor in BLHD format (B, L, H, D) = (Z, N_CTX, H, HEAD_DIM) + x = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + # Test get_fake_quant + x_fq = get_fake_quant(x) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + # Compare input and output + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print(f"✓ Attention shape test passed for (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM})") + + +def test_get_fake_quant_non_divisible_blocks(): + """Test get_fake_quant with shapes that are not divisible by block sizes.""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + + # Test with non-divisible dimensions + test_shapes = [ + (100, 100), # Not divisible by 128 + (150, 200), # Not divisible by 128 + (256, 100), # One dimension divisible, one not + ] + + for shape in test_shapes: + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Test get_fake_quant + x_fq = get_fake_quant(x) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" Shape {shape} - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print(f"✓ Non-divisible block test passed for {shape}") + + +def test_get_fake_quant_idempotency(): + """Test that quantizing twice produces similar results (idempotency check).""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + shape = (128, 128) + + # Create input tensor + x = torch.randn(shape, dtype=dtype, device=DEVICE) + + # Quantize once + x_fq1 = get_fake_quant(x) + + # Quantize again + x_fq2 = get_fake_quant(x_fq1) + + # Compare the two quantizations + max_diff = (x_fq1 - x_fq2).abs().max() + mean_diff = (x_fq1 - x_fq2).abs().mean() + cos_sim = cosine_similarity(x_fq1, x_fq2) + print(f" First quantization vs Second quantization - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ Idempotency test passed.") + + +def test_get_fake_quant_different_value_ranges(): + """Test get_fake_quant with different value ranges.""" + torch.manual_seed(42) + + dtype = torch.bfloat16 + shape = (128, 128) + + test_cases = [ + ("normal_range", lambda: torch.randn(shape, dtype=dtype, device=DEVICE)), + ("small_range", lambda: torch.randn(shape, dtype=dtype, device=DEVICE) * 0.01), + ("medium_range", lambda: torch.randn(shape, dtype=dtype, device=DEVICE) * 1.0), + ("large_range", lambda: torch.randn(shape, dtype=dtype, device=DEVICE) * 10.0), + ("very_large_range", lambda: torch.randn(shape, dtype=dtype, device=DEVICE) * 100.0), + ] + + for name, gen_func in test_cases: + x = gen_func() + x_fq = get_fake_quant(x) + + assert x_fq.shape == x.shape + assert torch.isfinite(x_fq).all() + + max_diff = (x_fq - x).abs().max() + mean_diff = (x_fq - x).abs().mean() + cos_sim = cosine_similarity(x_fq, x) + print(f" {name} - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + print("✓ Different value ranges test passed.") + + +if __name__ == "__main__": + print("Running get_fake_quant precision tests...") + print(f"Device: {DEVICE}") + print() + + test_get_fake_quant_basic() + test_get_fake_quant_different_shapes() + test_get_fake_quant_different_dtypes() + test_get_fake_quant_3d_4d_tensors() + test_get_fake_quant_edge_cases() + test_get_fake_quant_attention_shapes() + test_get_fake_quant_non_divisible_blocks() + test_get_fake_quant_idempotency() + test_get_fake_quant_different_value_ranges() + + print() + print("All tests passed! ✓") + diff --git a/test_qat_attn.py b/test_qat_attn.py new file mode 100644 index 000000000..d36ad3a85 --- /dev/null +++ b/test_qat_attn.py @@ -0,0 +1,1503 @@ +#!/usr/bin/env python3 +""" +Simple test for QAT attention implementation. +Tests forward and backward passes with and without QAT enabled. +""" + +import torch +from qat_attn import _attention +from fused_attention import attention as fused_attention +from math import sqrt + +attention = _attention.apply +DEVICE = torch.device("cuda") + + +def qat_attn_wrapper(q_BLHD, k_BLHD, v_BLHD, is_causal=False, sm_scale=None): + """ + Wrapper function that mimics qat_attn from sage_attn3.py. + Converts from BLHD format to BHLD format, calls attention, then converts back. + + Args: + q_BLHD: Query tensor in (B, L, H, D) format + k_BLHD: Key tensor in (B, L, H, D) format + v_BLHD: Value tensor in (B, L, H, D) format + is_causal: Whether to apply causal masking + sm_scale: Scale factor for attention scores (if None, uses 1.0 / sqrt(D)) + + Returns: + Output tensor in (B, L, H, D) format + """ + if sm_scale is None: + sm_scale = 1.0 / sqrt(q_BLHD.shape[-1]) + + # Convert from BLHD to BHLD format + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + + # Call attention with BHLD format + o_BHLD = attention(q_BHLD, k_BHLD, v_BHLD, is_causal, sm_scale) + + # Convert back from BHLD to BLHD format + return o_BHLD.permute(0, 2, 1, 3).contiguous() + + +def fused_attn_wrapper(q_BLHD, k_BLHD, v_BLHD, is_causal=False, sm_scale=None, warp_specialize=True): + """ + Wrapper function for fused_attention from fused_attention.py. + Converts from BLHD format to BHLD format, calls attention, then converts back. + + Args: + q_BLHD: Query tensor in (B, L, H, D) format + k_BLHD: Key tensor in (B, L, H, D) format + v_BLHD: Value tensor in (B, L, H, D) format + is_causal: Whether to apply causal masking + sm_scale: Scale factor for attention scores (if None, uses 1.0 / sqrt(D)) + warp_specialize: Whether to use warp specialization (default: True) + + Returns: + Output tensor in (B, L, H, D) format + """ + if sm_scale is None: + sm_scale = 1.0 / sqrt(q_BLHD.shape[-1]) + + # Convert from BLHD to BHLD format + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + + # Call fused attention with BHLD format + # Note: fused_attention expects inputs in BHLD format and only supports same sequence lengths + o_BHLD = fused_attention(q_BHLD, k_BHLD, v_BHLD, is_causal, sm_scale, warp_specialize) + + # Convert back from BHLD to BLHD format + return o_BHLD.permute(0, 2, 1, 3).contiguous() + + +def cosine_similarity(tensor1, tensor2): + """ + Compute cosine similarity between two tensors. + + Args: + tensor1: First tensor + tensor2: Second tensor (same shape as tensor1) + + Returns: + Cosine similarity value (scalar) + """ + # Flatten tensors for computation + t1_flat = tensor1.flatten().float() + t2_flat = tensor2.flatten().float() + + # Compute cosine similarity: (A · B) / (||A|| * ||B||) + dot_product = torch.dot(t1_flat, t2_flat) + norm1 = torch.norm(t1_flat) + norm2 = torch.norm(t2_flat) + + # Avoid division by zero + if norm1 == 0 or norm2 == 0: + return 0.0 + + cos_sim = dot_product / (norm1 * norm2) + return cos_sim.item() + + +def naive_attention(q, k, v, causal, sm_scale): + """ + Naive PyTorch implementation of attention for comparison. + + Args: + q: Query tensor of shape (Z, H, N_CTX_Q, HEAD_DIM) + k: Key tensor of shape (Z, H, N_CTX_KV, HEAD_DIM) + v: Value tensor of shape (Z, H, N_CTX_KV, HEAD_DIM) + causal: Whether to apply causal masking (only meaningful when N_CTX_Q == N_CTX_KV) + sm_scale: Scale factor for attention scores + + Returns: + Output tensor of shape (Z, H, N_CTX_Q, HEAD_DIM) + """ + # Compute attention scores: QK^T + scores = torch.matmul(q, k.transpose(-2, -1)) * sm_scale + + # Apply causal mask if needed + # Note: Causal masking is only meaningful when Q and KV have the same sequence length + if causal: + N_CTX_Q = q.shape[-2] + N_CTX_KV = k.shape[-2] + if N_CTX_Q == N_CTX_KV: + # Standard causal masking: query i can only attend to keys 0 to i + mask = torch.tril(torch.ones(N_CTX_Q, N_CTX_KV, device=scores.device, dtype=scores.dtype)) + scores = scores.masked_fill(mask == 0, float("-inf")) + else: + # For different sequence lengths, causal masking doesn't apply in the same way + # In cross-attention, typically no causal masking is applied + pass + + # Apply softmax + attn_weights = torch.softmax(scores, dim=-1) + + # Apply attention weights to values + out = torch.matmul(attn_weights, v) + + return out + +def test_qat_attention_forward(): + """Test forward pass with QAT enabled vs disabled.""" + torch.manual_seed(42) + + # Test parameters + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False + + # Create input tensors in BLHD format (B, L, H, D) = (Z, N_CTX, H, HEAD_DIM) + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + + # Test with QAT (using wrapper that does permute/contiguous) + out_qat_BLHD = qat_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention comparison + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + # Check that outputs have correct shape + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + + # Check that outputs are finite + assert torch.isfinite(out_qat_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare QAT and naive outputs + max_diff = (out_qat_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_qat_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_qat_BLHD, out_naive_BLHD) + print(f" QAT vs Naive Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM} - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # QAT output should be reasonably close to naive (within tolerance due to quantization) + # Using a reasonable tolerance for float16 and quantization effects + # assert max_diff < 1.0, f"QAT and naive outputs should be reasonably close, got max_diff={max_diff.item():.6f}" + + print("✓ Forward pass test passed.") + + +def test_qat_attention_backward(): + """Test backward pass with QAT enabled.""" + torch.manual_seed(42) + + # Test parameters + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format for QAT + q_qat_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + k_qat_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + v_qat_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + + # Create input tensors for naive (same values, convert to BHLD) + q_naive_BHLD = q_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + k_naive_BHLD = k_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + v_naive_BHLD = v_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + + # Forward pass with QAT (using wrapper) + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + + # Forward pass with naive + out_naive_BHLD = naive_attention(q_naive_BHLD, k_naive_BHLD, v_naive_BHLD, causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + # Create dummy gradient (same for both, in BLHD format) + # Ensure gradient is contiguous as required by the backward function + dout = torch.randn_like(out_qat_BLHD).contiguous() + + # Backward pass for QAT + out_qat_BLHD.backward(dout) + + # Backward pass for naive + out_naive_BLHD.backward(dout) + + # Check that gradients are computed + assert q_qat_BLHD.grad is not None + assert k_qat_BLHD.grad is not None + assert v_qat_BLHD.grad is not None + assert q_naive_BHLD.grad is not None + assert k_naive_BHLD.grad is not None + assert v_naive_BHLD.grad is not None + + # Check that gradients have correct shape + assert q_qat_BLHD.grad.shape == q_qat_BLHD.shape + assert k_qat_BLHD.grad.shape == k_qat_BLHD.shape + assert v_qat_BLHD.grad.shape == v_qat_BLHD.shape + + # Check that gradients are finite + assert torch.isfinite(q_qat_BLHD.grad).all() + assert torch.isfinite(k_qat_BLHD.grad).all() + assert torch.isfinite(v_qat_BLHD.grad).all() + + # Convert naive gradients to BLHD format for comparison + q_naive_grad_BLHD = q_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + k_naive_grad_BLHD = k_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + v_naive_grad_BLHD = v_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + + # Compare gradients with naive + dq_diff = (q_qat_BLHD.grad - q_naive_grad_BLHD).abs() + dk_diff = (k_qat_BLHD.grad - k_naive_grad_BLHD).abs() + dv_diff = (v_qat_BLHD.grad - v_naive_grad_BLHD).abs() + + dq_cos_sim = cosine_similarity(q_qat_BLHD.grad, q_naive_grad_BLHD) + dk_cos_sim = cosine_similarity(k_qat_BLHD.grad, k_naive_grad_BLHD) + dv_cos_sim = cosine_similarity(v_qat_BLHD.grad, v_naive_grad_BLHD) + + print(f" dQ - Max diff: {dq_diff.max().item():.6f}, Mean diff: {dq_diff.mean().item():.6f}, Cosine sim: {dq_cos_sim:.6f}") + print(f" dK - Max diff: {dk_diff.max().item():.6f}, Mean diff: {dk_diff.mean().item():.6f}, Cosine sim: {dk_cos_sim:.6f}") + print(f" dV - Max diff: {dv_diff.max().item():.6f}, Mean diff: {dv_diff.mean().item():.6f}, Cosine sim: {dv_cos_sim:.6f}") + + # Gradients should be reasonably close (within tolerance due to quantization) + # assert dq_diff.max() < 2.0, f"dQ gradients should be reasonably close, got max_diff={dq_diff.max().item():.6f}" + # assert dk_diff.max() < 2.0, f"dK gradients should be reasonably close, got max_diff={dk_diff.max().item():.6f}" + # assert dv_diff.max() < 2.0, f"dV gradients should be reasonably close, got max_diff={dv_diff.max().item():.6f}" + + print("✓ Backward pass test passed.") + + +def test_qat_attention_different_shapes(): + """Test QAT attention with different input shapes.""" + torch.manual_seed(42) + + test_configs = [ + (2, 4, 128, 64), # Medium + (1, 8, 256, 128), # Large head dim + ] + + dtype = torch.bfloat16 + causal = True + + for Z, H, N_CTX, HEAD_DIM in test_configs: + # Create input tensors in BLHD format + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + sm_scale = 1.0 / sqrt(HEAD_DIM) + out_qat_BLHD = qat_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_qat_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare outputs + max_diff = (out_qat_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_qat_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_qat_BLHD, out_naive_BLHD) + print(f" (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs should be reasonably close + # assert max_diff < 1.0, f"QAT and naive outputs should be reasonably close for shape (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}), got max_diff={max_diff.item():.6f}" + + print(f"✓ Shape test passed for (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM})") + + +def test_qat_attention_non_causal(): + """Test QAT attention with non-causal masking.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False + + # Create input tensors in BLHD format + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + out_qat_BLHD = qat_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_qat_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare outputs + max_diff = (out_qat_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_qat_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_qat_BLHD, out_naive_BLHD) + print(f" QAT vs Naive (non-causal) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs should be reasonably close + # assert max_diff < 1.0, f"QAT and naive outputs should be reasonably close for non-causal, got max_diff={max_diff.item():.6f}" + + print("✓ Non-causal test passed.") + + +def test_qat_attention_causal(): + """Test QAT attention with causal masking.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + out_qat_BLHD = qat_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_qat_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare outputs + max_diff = (out_qat_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_qat_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_qat_BLHD, out_naive_BLHD) + print(f" QAT vs Naive (causal) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs should be reasonably close + # assert max_diff < 1.0, f"QAT and naive outputs should be reasonably close for causal, got max_diff={max_diff.item():.6f}" + + print("✓ Causal test passed.") + + +def test_qat_attention_causal_backward(): + """Test backward pass of QAT attention with causal masking vs naive.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format for QAT + q_qat_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + k_qat_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + v_qat_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + + # Create input tensors for naive (same values, convert to BHLD) + q_naive_BHLD = q_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + k_naive_BHLD = k_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + v_naive_BHLD = v_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + + # Forward pass with QAT + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + + # Forward pass with naive + out_naive_BHLD = naive_attention(q_naive_BHLD, k_naive_BHLD, v_naive_BHLD, causal, sm_scale) + out_naive_BHLD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + # Create dummy gradient + dout = torch.randn_like(out_qat_BLHD).contiguous() + + # Backward pass for QAT + out_qat_BLHD.backward(dout) + + # Backward pass for naive + out_naive_BHLD.backward(dout) + + # Check that gradients are computed + assert q_qat_BLHD.grad is not None + assert k_qat_BLHD.grad is not None + assert v_qat_BLHD.grad is not None + + # Convert naive gradients to BLHD format for comparison + q_naive_grad_BLHD = q_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + k_naive_grad_BLHD = k_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + v_naive_grad_BLHD = v_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + + # Compare gradients + dq_diff = (q_qat_BLHD.grad - q_naive_grad_BLHD).abs() + dk_diff = (k_qat_BLHD.grad - k_naive_grad_BLHD).abs() + dv_diff = (v_qat_BLHD.grad - v_naive_grad_BLHD).abs() + + dq_cos_sim = cosine_similarity(q_qat_BLHD.grad, q_naive_grad_BLHD) + dk_cos_sim = cosine_similarity(k_qat_BLHD.grad, k_naive_grad_BLHD) + dv_cos_sim = cosine_similarity(v_qat_BLHD.grad, v_naive_grad_BLHD) + + print(f" dQ - Max diff: {dq_diff.max().item():.6f}, Mean diff: {dq_diff.mean().item():.6f}, Cosine sim: {dq_cos_sim:.6f}") + print(f" dK - Max diff: {dk_diff.max().item():.6f}, Mean diff: {dk_diff.mean().item():.6f}, Cosine sim: {dk_cos_sim:.6f}") + print(f" dV - Max diff: {dv_diff.max().item():.6f}, Mean diff: {dv_diff.mean().item():.6f}, Cosine sim: {dv_cos_sim:.6f}") + + print("✓ Causal backward test passed.") + + +def test_qat_attention_different_seq_lengths(): + """Test QAT attention with different sequence lengths for Q and KV (cross-attention).""" + torch.manual_seed(42) + + test_configs = [ + (2, 4, 64, 128, 64, False), # Q shorter than KV, non-causal + (2, 4, 128, 64, 64, False), # Q longer than KV, non-causal + (1, 8, 256, 128, 64, False), # Q longer than KV, larger dimensions + ] + + dtype = torch.bfloat16 + + for Z, H, N_CTX_Q, N_CTX_KV, HEAD_DIM, causal in test_configs: + sm_scale = 1.0 / sqrt(HEAD_DIM) + + # Create input tensors in BLHD format with different sequence lengths + q_BLHD = torch.randn((Z, N_CTX_Q, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX_KV, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX_KV, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + # Test with QAT (using wrapper that does permute/contiguous) + out_qat_BLHD = qat_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention comparison + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + # Check that outputs have correct shape (should match Q sequence length) + assert out_qat_BLHD.shape == (Z, N_CTX_Q, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX_Q, H, HEAD_DIM) + + # Check that outputs are finite + assert torch.isfinite(out_qat_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare QAT and naive outputs + max_diff = (out_qat_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_qat_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_qat_BLHD, out_naive_BLHD) + print(f" (N_CTX_Q={N_CTX_Q}, N_CTX_KV={N_CTX_KV}, causal={causal}) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs should be reasonably close + # assert max_diff < 1.0, f"QAT and naive outputs should be reasonably close for (N_CTX_Q={N_CTX_Q}, N_CTX_KV={N_CTX_KV}), got max_diff={max_diff.item():.6f}" + + print(f"✓ Different sequence lengths test passed for (N_CTX_Q={N_CTX_Q}, N_CTX_KV={N_CTX_KV})") + + +def test_qat_attention_different_seq_lengths_backward(): + """Test backward pass of QAT attention with different sequence lengths for Q and KV (cross-attention).""" + torch.manual_seed(42) + + test_configs = [ + (2, 4, 64, 128, 64, False), # Q shorter than KV, non-causal + (2, 4, 128, 64, 64, False), # Q longer than KV, non-causal + ] + + dtype = torch.bfloat16 + + for Z, H, N_CTX_Q, N_CTX_KV, HEAD_DIM, causal in test_configs: + sm_scale = 1.0 / sqrt(HEAD_DIM) + + # Create input tensors in BLHD format with different sequence lengths + q_qat_BLHD = torch.randn((Z, N_CTX_Q, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + k_qat_BLHD = torch.randn((Z, N_CTX_KV, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + v_qat_BLHD = torch.randn((Z, N_CTX_KV, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + + # Create input tensors for naive (same values, convert to BHLD) + q_naive_BHLD = q_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + k_naive_BHLD = k_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + v_naive_BHLD = v_qat_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + + # Forward pass with QAT + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + + # Forward pass with naive + out_naive_BHLD = naive_attention(q_naive_BHLD, k_naive_BHLD, v_naive_BHLD, causal, sm_scale) + out_naive_BHLD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + # Create dummy gradient + dout = torch.randn_like(out_qat_BLHD).contiguous() + + # Backward pass for QAT + out_qat_BLHD.backward(dout) + + # Backward pass for naive + out_naive_BHLD.backward(dout) + + # Check that gradients are computed + assert q_qat_BLHD.grad is not None + assert k_qat_BLHD.grad is not None + assert v_qat_BLHD.grad is not None + assert q_naive_BHLD.grad is not None + assert k_naive_BHLD.grad is not None + assert v_naive_BHLD.grad is not None + + # Check that gradients have correct shape + assert q_qat_BLHD.grad.shape == q_qat_BLHD.shape + assert k_qat_BLHD.grad.shape == k_qat_BLHD.shape + assert v_qat_BLHD.grad.shape == v_qat_BLHD.shape + + # Check that gradients are finite + assert torch.isfinite(q_qat_BLHD.grad).all() + assert torch.isfinite(k_qat_BLHD.grad).all() + assert torch.isfinite(v_qat_BLHD.grad).all() + + # Convert naive gradients to BLHD format for comparison + q_naive_grad_BLHD = q_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + k_naive_grad_BLHD = k_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + v_naive_grad_BLHD = v_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + + # Compare gradients + dq_diff = (q_qat_BLHD.grad - q_naive_grad_BLHD).abs() + dk_diff = (k_qat_BLHD.grad - k_naive_grad_BLHD).abs() + dv_diff = (v_qat_BLHD.grad - v_naive_grad_BLHD).abs() + + dq_cos_sim = cosine_similarity(q_qat_BLHD.grad, q_naive_grad_BLHD) + dk_cos_sim = cosine_similarity(k_qat_BLHD.grad, k_naive_grad_BLHD) + dv_cos_sim = cosine_similarity(v_qat_BLHD.grad, v_naive_grad_BLHD) + + print(f" (N_CTX_Q={N_CTX_Q}, N_CTX_KV={N_CTX_KV}) - dQ: Max diff: {dq_diff.max().item():.6f}, Mean diff: {dq_diff.mean().item():.6f}, Cosine sim: {dq_cos_sim:.6f}") + print(f" (N_CTX_Q={N_CTX_Q}, N_CTX_KV={N_CTX_KV}) - dK: Max diff: {dk_diff.max().item():.6f}, Mean diff: {dk_diff.mean().item():.6f}, Cosine sim: {dk_cos_sim:.6f}") + print(f" (N_CTX_Q={N_CTX_Q}, N_CTX_KV={N_CTX_KV}) - dV: Max diff: {dv_diff.max().item():.6f}, Mean diff: {dv_diff.mean().item():.6f}, Cosine sim: {dv_cos_sim:.6f}") + + # Gradients should be reasonably close (within tolerance due to quantization) + # assert dq_diff.max() < 2.0, f"dQ gradients should be reasonably close, got max_diff={dq_diff.max().item():.6f}" + # assert dk_diff.max() < 2.0, f"dK gradients should be reasonably close, got max_diff={dk_diff.max().item():.6f}" + # assert dv_diff.max() < 2.0, f"dV gradients should be reasonably close, got max_diff={dv_diff.max().item():.6f}" + + print(f"✓ Cross-attention backward test passed for (N_CTX_Q={N_CTX_Q}, N_CTX_KV={N_CTX_KV})") + + +def test_fused_attention_forward(): + """Test forward pass of fused attention vs naive PyTorch implementation.""" + torch.manual_seed(42) + + # Test parameters + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 # fused_attention uses float16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False + + # Create input tensors in BLHD format (B, L, H, D) = (Z, N_CTX, H, HEAD_DIM) + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + # Test with fused attention (using wrapper that does permute/contiguous) + out_fused_BLHD = fused_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention comparison + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + # Check that outputs have correct shape + assert out_fused_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + + # Check that outputs are finite + assert torch.isfinite(out_fused_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare fused and naive outputs + max_diff = (out_fused_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_fused_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_fused_BLHD, out_naive_BLHD) + print(f" Fused vs Naive - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Fused attention should be reasonably close to naive (within tolerance for float16) + # assert max_diff < 1e-1, f"Fused and naive outputs should be reasonably close, got max_diff={max_diff.item():.6f}" + + print("✓ Fused attention forward pass test passed.") + + +def test_fused_attention_backward(): + """Test backward pass of fused attention vs naive PyTorch implementation.""" + torch.manual_seed(42) + + # Test parameters + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format for fused attention + q_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + k_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + v_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + + # Create input tensors for naive (same values, convert to BHLD) + q_naive_BHLD = q_fused_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + k_naive_BHLD = k_fused_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + v_naive_BHLD = v_fused_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + + # Forward pass with fused attention (using wrapper) + out_fused_BLHD = fused_attn_wrapper(q_fused_BLHD, k_fused_BLHD, v_fused_BLHD, causal, sm_scale) + + # Forward pass with naive + out_naive_BHLD = naive_attention(q_naive_BHLD, k_naive_BHLD, v_naive_BHLD, causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + # Create dummy gradient (same for both, in BLHD format) + dout = torch.randn_like(out_fused_BLHD).contiguous() + + # Backward pass for fused attention + out_fused_BLHD.backward(dout) + + # Backward pass for naive + out_naive_BLHD.backward(dout) + + # Check that gradients are computed + assert q_fused_BLHD.grad is not None + assert k_fused_BLHD.grad is not None + assert v_fused_BLHD.grad is not None + assert q_naive_BHLD.grad is not None + assert k_naive_BHLD.grad is not None + assert v_naive_BHLD.grad is not None + + # Check that gradients have correct shape + assert q_fused_BLHD.grad.shape == q_fused_BLHD.shape + assert k_fused_BLHD.grad.shape == k_fused_BLHD.shape + assert v_fused_BLHD.grad.shape == v_fused_BLHD.shape + + # Check that gradients are finite + assert torch.isfinite(q_fused_BLHD.grad).all() + assert torch.isfinite(k_fused_BLHD.grad).all() + assert torch.isfinite(v_fused_BLHD.grad).all() + + # Convert naive gradients to BLHD format for comparison + q_naive_grad_BLHD = q_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + k_naive_grad_BLHD = k_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + v_naive_grad_BLHD = v_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + + # Compare gradients with naive + dq_diff = (q_fused_BLHD.grad - q_naive_grad_BLHD).abs() + dk_diff = (k_fused_BLHD.grad - k_naive_grad_BLHD).abs() + dv_diff = (v_fused_BLHD.grad - v_naive_grad_BLHD).abs() + + dq_cos_sim = cosine_similarity(q_fused_BLHD.grad, q_naive_grad_BLHD) + dk_cos_sim = cosine_similarity(k_fused_BLHD.grad, k_naive_grad_BLHD) + dv_cos_sim = cosine_similarity(v_fused_BLHD.grad, v_naive_grad_BLHD) + + print(f" dQ - Max diff: {dq_diff.max().item():.6f}, Mean diff: {dq_diff.mean().item():.6f}, Cosine sim: {dq_cos_sim:.6f}") + print(f" dK - Max diff: {dk_diff.max().item():.6f}, Mean diff: {dk_diff.mean().item():.6f}, Cosine sim: {dk_cos_sim:.6f}") + print(f" dV - Max diff: {dv_diff.max().item():.6f}, Mean diff: {dv_diff.mean().item():.6f}, Cosine sim: {dv_cos_sim:.6f}") + + # Gradients should be reasonably close (within tolerance for float16) + # assert dq_diff.max() < 1e-1, f"dQ gradients should be reasonably close, got max_diff={dq_diff.max().item():.6f}" + # assert dk_diff.max() < 1e-1, f"dK gradients should be reasonably close, got max_diff={dk_diff.max().item():.6f}" + # assert dv_diff.max() < 1e-1, f"dV gradients should be reasonably close, got max_diff={dv_diff.max().item():.6f}" + + print("✓ Fused attention backward pass test passed.") + + +def test_fused_attention_different_shapes(): + """Test fused attention with different input shapes.""" + torch.manual_seed(42) + + test_configs = [ + (1, 2, 64, 32), # Small + (2, 4, 128, 64), # Medium + (1, 8, 256, 128), # Large head dim + ] + + dtype = torch.bfloat16 + causal = True + + for Z, H, N_CTX, HEAD_DIM in test_configs: + # Create input tensors in BLHD format + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + sm_scale = 1.0 / sqrt(HEAD_DIM) + out_fused_BLHD = fused_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + assert out_fused_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_fused_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare outputs + max_diff = (out_fused_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_fused_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_fused_BLHD, out_naive_BLHD) + print(f" (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs should be reasonably close + # assert max_diff < 1e-1, f"Fused and naive outputs should be reasonably close for shape (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}), got max_diff={max_diff.item():.6f}" + + print(f"✓ Fused attention shape test passed for (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM})") + + +def test_fused_attention_non_causal(): + """Test fused attention with non-causal masking.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False + + # Create input tensors in BLHD format + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + out_fused_BLHD = fused_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + assert out_fused_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_fused_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare outputs + max_diff = (out_fused_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_fused_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_fused_BLHD, out_naive_BLHD) + print(f" Fused vs Naive (non-causal) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs should be reasonably close + # assert max_diff < 1e-1, f"Fused and naive outputs should be reasonably close for non-causal, got max_diff={max_diff.item():.6f}" + + print("✓ Fused attention non-causal test passed.") + + +def test_fused_attention_causal(): + """Test fused attention with causal masking.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + out_fused_BLHD = fused_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + assert out_fused_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_fused_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare outputs + max_diff = (out_fused_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_fused_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_fused_BLHD, out_naive_BLHD) + print(f" Fused vs Naive (causal) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs should be reasonably close + # assert max_diff < 1e-1, f"Fused and naive outputs should be reasonably close for causal, got max_diff={max_diff.item():.6f}" + + print("✓ Fused attention causal test passed.") + + +def test_fused_attention_causal_backward(): + """Test backward pass of fused attention with causal masking vs naive.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format for fused attention + q_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + k_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + v_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + + # Create input tensors for naive (same values, convert to BHLD) + q_naive_BHLD = q_fused_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + k_naive_BHLD = k_fused_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + v_naive_BHLD = v_fused_BLHD.clone().permute(0, 2, 1, 3).contiguous().detach().requires_grad_(True) + + # Forward pass with fused attention + out_fused_BLHD = fused_attn_wrapper(q_fused_BLHD, k_fused_BLHD, v_fused_BLHD, causal, sm_scale) + + # Forward pass with naive + out_naive_BHLD = naive_attention(q_naive_BHLD, k_naive_BHLD, v_naive_BHLD, causal, sm_scale) + out_naive_BHLD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + + # Create dummy gradient + dout = torch.randn_like(out_fused_BLHD).contiguous() + + # Backward pass for fused attention + out_fused_BLHD.backward(dout) + + # Backward pass for naive + out_naive_BHLD.backward(dout) + + # Check that gradients are computed + assert q_fused_BLHD.grad is not None + assert k_fused_BLHD.grad is not None + assert v_fused_BLHD.grad is not None + + # Convert naive gradients to BLHD format for comparison + q_naive_grad_BLHD = q_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + k_naive_grad_BLHD = k_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + v_naive_grad_BLHD = v_naive_BHLD.grad.permute(0, 2, 1, 3).contiguous() + + # Compare gradients + dq_diff = (q_fused_BLHD.grad - q_naive_grad_BLHD).abs() + dk_diff = (k_fused_BLHD.grad - k_naive_grad_BLHD).abs() + dv_diff = (v_fused_BLHD.grad - v_naive_grad_BLHD).abs() + + dq_cos_sim = cosine_similarity(q_fused_BLHD.grad, q_naive_grad_BLHD) + dk_cos_sim = cosine_similarity(k_fused_BLHD.grad, k_naive_grad_BLHD) + dv_cos_sim = cosine_similarity(v_fused_BLHD.grad, v_naive_grad_BLHD) + + print(f" dQ - Max diff: {dq_diff.max().item():.6f}, Mean diff: {dq_diff.mean().item():.6f}, Cosine sim: {dq_cos_sim:.6f}") + print(f" dK - Max diff: {dk_diff.max().item():.6f}, Mean diff: {dk_diff.mean().item():.6f}, Cosine sim: {dk_cos_sim:.6f}") + print(f" dV - Max diff: {dv_diff.max().item():.6f}, Mean diff: {dv_diff.mean().item():.6f}, Cosine sim: {dv_cos_sim:.6f}") + + print("✓ Fused attention causal backward test passed.") + + +def test_fused_vs_qat_attention_forward(): + """Test forward pass comparing fused attention vs QAT attention.""" + torch.manual_seed(42) + + # Test parameters + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False + + # Create input tensors in BLHD format + # Use float16 for both (QAT can handle float16, though it typically uses float16) + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + + # Test with fused attention + out_fused_BLHD = fused_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Test with QAT attention (convert to float16 for QAT) + q_qat_BLHD = q_BLHD.clone().to(torch.bfloat16) + k_qat_BLHD = k_BLHD.clone().to(torch.bfloat16) + v_qat_BLHD = v_BLHD.clone().to(torch.bfloat16) + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + + # Convert QAT output back to float16 for comparison + out_qat_BLHD = out_qat_BLHD.to(torch.bfloat16) + + # Check that outputs have correct shape + assert out_fused_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + + # Check that outputs are finite + assert torch.isfinite(out_fused_BLHD).all() + assert torch.isfinite(out_qat_BLHD).all() + + # Compare fused and QAT outputs + max_diff = (out_fused_BLHD - out_qat_BLHD).abs().max() + mean_diff = (out_fused_BLHD - out_qat_BLHD).abs().mean() + cos_sim = cosine_similarity(out_fused_BLHD, out_qat_BLHD) + print(f" Fused vs QAT - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # They may differ due to quantization in QAT and different implementations + # assert max_diff < 1.0, f"Fused and QAT outputs should be reasonably close, got max_diff={max_diff.item():.6f}" + + print("✓ Fused vs QAT forward pass test passed.") + + +def test_fused_vs_qat_attention_backward(): + """Test backward pass comparing fused attention vs QAT attention.""" + torch.manual_seed(42) + + # Test parameters + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format for fused attention (float16) + q_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE, requires_grad=True) + k_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE, requires_grad=True) + v_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE, requires_grad=True) + + # Create input tensors for QAT (float16, same values) + q_qat_BLHD = q_fused_BLHD.clone().to(torch.bfloat16).detach().requires_grad_(True) + k_qat_BLHD = k_fused_BLHD.clone().to(torch.bfloat16).detach().requires_grad_(True) + v_qat_BLHD = v_fused_BLHD.clone().to(torch.bfloat16).detach().requires_grad_(True) + + # Forward pass with fused attention + out_fused_BLHD = fused_attn_wrapper(q_fused_BLHD, k_fused_BLHD, v_fused_BLHD, causal, sm_scale) + + # Forward pass with QAT + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + + # Create dummy gradient (same for both, in BLHD format) + dout = torch.randn_like(out_fused_BLHD).contiguous() + dout_qat = dout.to(torch.bfloat16) + + # Backward pass for fused attention + out_fused_BLHD.backward(dout) + + # Backward pass for QAT + out_qat_BLHD.backward(dout_qat) + + # Check that gradients are computed + assert q_fused_BLHD.grad is not None + assert k_fused_BLHD.grad is not None + assert v_fused_BLHD.grad is not None + assert q_qat_BLHD.grad is not None + assert k_qat_BLHD.grad is not None + assert v_qat_BLHD.grad is not None + + # Check that gradients have correct shape + assert q_fused_BLHD.grad.shape == q_fused_BLHD.shape + assert k_fused_BLHD.grad.shape == k_fused_BLHD.shape + assert v_fused_BLHD.grad.shape == v_fused_BLHD.shape + + # Check that gradients are finite + assert torch.isfinite(q_fused_BLHD.grad).all() + assert torch.isfinite(k_fused_BLHD.grad).all() + assert torch.isfinite(v_fused_BLHD.grad).all() + + # Convert QAT gradients to float16 for comparison + q_qat_grad_f16 = q_qat_BLHD.grad.to(torch.bfloat16) + k_qat_grad_f16 = k_qat_BLHD.grad.to(torch.bfloat16) + v_qat_grad_f16 = v_qat_BLHD.grad.to(torch.bfloat16) + + # Compare gradients + dq_diff = (q_fused_BLHD.grad - q_qat_grad_f16).abs() + dk_diff = (k_fused_BLHD.grad - k_qat_grad_f16).abs() + dv_diff = (v_fused_BLHD.grad - v_qat_grad_f16).abs() + + dq_cos_sim = cosine_similarity(q_fused_BLHD.grad, q_qat_grad_f16) + dk_cos_sim = cosine_similarity(k_fused_BLHD.grad, k_qat_grad_f16) + dv_cos_sim = cosine_similarity(v_fused_BLHD.grad, v_qat_grad_f16) + + print(f" dQ - Max diff: {dq_diff.max().item():.6f}, Mean diff: {dq_diff.mean().item():.6f}, Cosine sim: {dq_cos_sim:.6f}") + print(f" dK - Max diff: {dk_diff.max().item():.6f}, Mean diff: {dk_diff.mean().item():.6f}, Cosine sim: {dk_cos_sim:.6f}") + print(f" dV - Max diff: {dv_diff.max().item():.6f}, Mean diff: {dv_diff.mean().item():.6f}, Cosine sim: {dv_cos_sim:.6f}") + + # Gradients may differ due to quantization in QAT and different implementations + # assert dq_diff.max() < 2.0, f"dQ gradients should be reasonably close, got max_diff={dq_diff.max().item():.6f}" + # assert dk_diff.max() < 2.0, f"dK gradients should be reasonably close, got max_diff={dk_diff.max().item():.6f}" + # assert dv_diff.max() < 2.0, f"dV gradients should be reasonably close, got max_diff={dv_diff.max().item():.6f}" + + print("✓ Fused vs QAT backward pass test passed.") + + +def test_fused_vs_qat_attention_different_shapes(): + """Test fused vs QAT attention with different input shapes.""" + torch.manual_seed(42) + + test_configs = [ + (2, 4, 128, 64), # Medium + (1, 8, 256, 128), # Large head dim + ] + + causal = True + + for Z, H, N_CTX, HEAD_DIM in test_configs: + sm_scale = 1.0 / sqrt(HEAD_DIM) + + # Create input tensors in BLHD format (float16) + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + + # Test with fused attention + out_fused_BLHD = fused_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Test with QAT attention (convert to float16) + q_qat_BLHD = q_BLHD.clone().to(torch.bfloat16) + k_qat_BLHD = k_BLHD.clone().to(torch.bfloat16) + v_qat_BLHD = v_BLHD.clone().to(torch.bfloat16) + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + out_qat_BLHD = out_qat_BLHD.to(torch.bfloat16) + + assert out_fused_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_fused_BLHD).all() + assert torch.isfinite(out_qat_BLHD).all() + + # Compare outputs + max_diff = (out_fused_BLHD - out_qat_BLHD).abs().max() + mean_diff = (out_fused_BLHD - out_qat_BLHD).abs().mean() + cos_sim = cosine_similarity(out_fused_BLHD, out_qat_BLHD) + print(f" (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs may differ due to quantization in QAT + # assert max_diff < 1.0, f"Fused and QAT outputs should be reasonably close for shape (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}), got max_diff={max_diff.item():.6f}" + + print(f"✓ Fused vs QAT shape test passed for (Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM})") + + +def test_fused_vs_qat_attention_non_causal(): + """Test fused vs QAT attention with non-causal masking.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False + + # Create input tensors in BLHD format (float16) + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + + # Test with fused attention + out_fused_BLHD = fused_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Test with QAT attention (convert to float16) + q_qat_BLHD = q_BLHD.clone().to(torch.bfloat16) + k_qat_BLHD = k_BLHD.clone().to(torch.bfloat16) + v_qat_BLHD = v_BLHD.clone().to(torch.bfloat16) + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + out_qat_BLHD = out_qat_BLHD.to(torch.bfloat16) + + assert out_fused_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_fused_BLHD).all() + assert torch.isfinite(out_qat_BLHD).all() + + # Compare outputs + max_diff = (out_fused_BLHD - out_qat_BLHD).abs().max() + mean_diff = (out_fused_BLHD - out_qat_BLHD).abs().mean() + cos_sim = cosine_similarity(out_fused_BLHD, out_qat_BLHD) + print(f" Fused vs QAT (non-causal) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs may differ due to quantization in QAT + # assert max_diff < 1.0, f"Fused and QAT outputs should be reasonably close for non-causal, got max_diff={max_diff.item():.6f}" + + print("✓ Fused vs QAT non-causal test passed.") + + +def test_fused_vs_qat_attention_causal(): + """Test fused vs QAT attention with causal masking.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format (float16) + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE) + + # Test with fused attention + out_fused_BLHD = fused_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Test with QAT attention (convert to float16) + q_qat_BLHD = q_BLHD.clone().to(torch.bfloat16) + k_qat_BLHD = k_BLHD.clone().to(torch.bfloat16) + v_qat_BLHD = v_BLHD.clone().to(torch.bfloat16) + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + out_qat_BLHD = out_qat_BLHD.to(torch.bfloat16) + + assert out_fused_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert torch.isfinite(out_fused_BLHD).all() + assert torch.isfinite(out_qat_BLHD).all() + + # Compare outputs + max_diff = (out_fused_BLHD - out_qat_BLHD).abs().max() + mean_diff = (out_fused_BLHD - out_qat_BLHD).abs().mean() + cos_sim = cosine_similarity(out_fused_BLHD, out_qat_BLHD) + print(f" Fused vs QAT (causal) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # Outputs may differ due to quantization in QAT + # assert max_diff < 1.0, f"Fused and QAT outputs should be reasonably close for causal, got max_diff={max_diff.item():.6f}" + + print("✓ Fused vs QAT causal test passed.") + + +def test_fused_vs_qat_attention_causal_backward(): + """Test backward pass of fused vs QAT attention with causal masking.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 2, 4, 128, 64 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = True + + # Create input tensors in BLHD format for fused attention (float16) + q_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE, requires_grad=True) + k_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE, requires_grad=True) + v_fused_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=torch.bfloat16, device=DEVICE, requires_grad=True) + + # Create input tensors for QAT (float16, same values) + q_qat_BLHD = q_fused_BLHD.clone().to(torch.bfloat16).detach().requires_grad_(True) + k_qat_BLHD = k_fused_BLHD.clone().to(torch.bfloat16).detach().requires_grad_(True) + v_qat_BLHD = v_fused_BLHD.clone().to(torch.bfloat16).detach().requires_grad_(True) + + # Forward pass with fused attention + out_fused_BLHD = fused_attn_wrapper(q_fused_BLHD, k_fused_BLHD, v_fused_BLHD, causal, sm_scale) + + # Forward pass with QAT + out_qat_BLHD = qat_attn_wrapper(q_qat_BLHD, k_qat_BLHD, v_qat_BLHD, causal, sm_scale) + + # Create dummy gradient + dout = torch.randn_like(out_fused_BLHD).contiguous() + dout_qat = dout.to(torch.bfloat16) + + # Backward pass for fused attention + out_fused_BLHD.backward(dout) + + # Backward pass for QAT + out_qat_BLHD.backward(dout_qat) + + # Check that gradients are computed + assert q_fused_BLHD.grad is not None + assert k_fused_BLHD.grad is not None + assert v_fused_BLHD.grad is not None + assert q_qat_BLHD.grad is not None + assert k_qat_BLHD.grad is not None + assert v_qat_BLHD.grad is not None + + # Convert QAT gradients to float16 for comparison + q_qat_grad_f16 = q_qat_BLHD.grad.to(torch.bfloat16) + k_qat_grad_f16 = k_qat_BLHD.grad.to(torch.bfloat16) + v_qat_grad_f16 = v_qat_BLHD.grad.to(torch.bfloat16) + + # Compare gradients + dq_diff = (q_fused_BLHD.grad - q_qat_grad_f16).abs() + dk_diff = (k_fused_BLHD.grad - k_qat_grad_f16).abs() + dv_diff = (v_fused_BLHD.grad - v_qat_grad_f16).abs() + + dq_cos_sim = cosine_similarity(q_fused_BLHD.grad, q_qat_grad_f16) + dk_cos_sim = cosine_similarity(k_fused_BLHD.grad, k_qat_grad_f16) + dv_cos_sim = cosine_similarity(v_fused_BLHD.grad, v_qat_grad_f16) + + print(f" dQ - Max diff: {dq_diff.max().item():.6f}, Mean diff: {dq_diff.mean().item():.6f}, Cosine sim: {dq_cos_sim:.6f}") + print(f" dK - Max diff: {dk_diff.max().item():.6f}, Mean diff: {dk_diff.mean().item():.6f}, Cosine sim: {dk_cos_sim:.6f}") + print(f" dV - Max diff: {dv_diff.max().item():.6f}, Mean diff: {dv_diff.mean().item():.6f}, Cosine sim: {dv_cos_sim:.6f}") + + print("✓ Fused vs QAT causal backward test passed.") + + +def test_qat_attention_wan_shape_forward(): + """Test QAT attention forward pass with WAN shape [1, 40, 9360, 128].""" + torch.manual_seed(42) + + # WAN shape: [1, 40, 9360, 128] = (Z, H, N_CTX, HEAD_DIM) + Z, H, N_CTX, HEAD_DIM = 1, 40, 9360, 128 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False # WAN uses non-causal attention + + print(f" Testing WAN shape: Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}") + + # Create input tensors in BLHD format (B, L, H, D) = (Z, N_CTX, H, HEAD_DIM) + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE) + + # Test with QAT (using wrapper that does permute/contiguous) + out_qat_BLHD = qat_attn_wrapper(q_BLHD.clone(), k_BLHD.clone(), v_BLHD.clone(), causal, sm_scale) + + # Convert to BHLD for naive attention comparison + q_BHLD = q_BLHD.permute(0, 2, 1, 3).contiguous() + k_BHLD = k_BLHD.permute(0, 2, 1, 3).contiguous() + v_BHLD = v_BLHD.permute(0, 2, 1, 3).contiguous() + out_naive_BHLD = naive_attention(q_BHLD.clone(), k_BHLD.clone(), v_BHLD.clone(), causal, sm_scale) + print(f" out_naive_BHLD shape (before permute): {out_naive_BHLD.shape}") + out_naive_BLHD = out_naive_BHLD.permute(0, 2, 1, 3).contiguous() + print(f" out_naive_BLHD shape (after permute): {out_naive_BLHD.shape}") + print(f" Expected shape: {(Z, N_CTX, H, HEAD_DIM)}") + print(f" out_qat_BLHD shape: {out_qat_BLHD.shape}") + + # Check that outputs have correct shape + assert out_qat_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + assert out_naive_BLHD.shape == (Z, N_CTX, H, HEAD_DIM) + + # Check that outputs are finite + assert torch.isfinite(out_qat_BLHD).all() + assert torch.isfinite(out_naive_BLHD).all() + + # Compare QAT and naive outputs + max_diff = (out_qat_BLHD - out_naive_BLHD).abs().max() + mean_diff = (out_qat_BLHD - out_naive_BLHD).abs().mean() + cos_sim = cosine_similarity(out_qat_BLHD, out_naive_BLHD) + print(f" QAT vs Naive (WAN shape) - Max diff: {max_diff.item():.6f}, Mean diff: {mean_diff.item():.6f}, Cosine sim: {cos_sim:.6f}") + + # QAT output should be reasonably close to naive (within tolerance due to quantization) + print("✓ WAN shape forward pass test passed.") + + +def test_qat_attention_wan_shape_backward(): + """Test QAT attention backward pass with WAN shape [1, 40, 9360, 128].""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 1, 40, 9360, 128 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False # WAN is non-causal + + print(f" Testing WAN shape: Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}") + + # ----------------------------- + # CREATE BLHD INPUTS FOR QAT + # ----------------------------- + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + + # ----------------------------- + # CREATE BHLD INPUTS FOR NAIVE + # ----------------------------- + q_BHLD = q_BLHD.detach().permute(0,2,1,3).contiguous().requires_grad_(True) + k_BHLD = k_BLHD.detach().permute(0,2,1,3).contiguous().requires_grad_(True) + v_BHLD = v_BLHD.detach().permute(0,2,1,3).contiguous().requires_grad_(True) + + # ----------------------------- + # FORWARD — QAT + # ----------------------------- + out_qat_BLHD = qat_attn_wrapper(q_BLHD, k_BLHD, v_BLHD, causal, sm_scale) + + # ----------------------------- + # FORWARD — NAIVE + # ----------------------------- + out_naive_BHLD = naive_attention(q_BHLD, k_BHLD, v_BHLD, causal, sm_scale) + + # Convert naive output back to BLHD for comparing outputs + out_naive_BLHD = out_naive_BHLD.permute(0,2,1,3).contiguous() + + # ----------------------------- + # BACKWARD + # ----------------------------- + dout_BLHD = torch.randn_like(out_qat_BLHD).contiguous() + + # QAT backward (BLHD dout) + out_qat_BLHD.backward(dout_BLHD) + + # ❗ FIXED: convert BLHD dout → BHLD for naive backward + dout_BHLD = dout_BLHD.permute(0,2,1,3).contiguous() + + out_naive_BHLD.backward(dout_BHLD) + + # ----------------------------- + # GRADIENT FORMAT FIX + # Convert naive BHLD grads → BLHD + # ----------------------------- + q_naive_grad_BLHD = q_BHLD.grad.permute(0,2,1,3).contiguous() + k_naive_grad_BLHD = k_BHLD.grad.permute(0,2,1,3).contiguous() + v_naive_grad_BLHD = v_BHLD.grad.permute(0,2,1,3).contiguous() + + # ----------------------------- + # PRINT GRADIENTS + # ----------------------------- + print("\n === QAT Gradients ===") + print(f" dQ_QAT - Shape: {q_BLHD.grad.shape}, Min: {q_BLHD.grad.min().item():.6f}, Max: {q_BLHD.grad.max().item():.6f}, Mean: {q_BLHD.grad.mean().item():.6f}, Std: {q_BLHD.grad.std().item():.6f}") + print(f" dK_QAT - Shape: {k_BLHD.grad.shape}, Min: {k_BLHD.grad.min().item():.6f}, Max: {k_BLHD.grad.max().item():.6f}, Mean: {k_BLHD.grad.mean().item():.6f}, Std: {k_BLHD.grad.std().item():.6f}") + print(f" dV_QAT - Shape: {v_BLHD.grad.shape}, Min: {v_BLHD.grad.min().item():.6f}, Max: {v_BLHD.grad.max().item():.6f}, Mean: {v_BLHD.grad.mean().item():.6f}, Std: {v_BLHD.grad.std().item():.6f}") + + print("\n === Naive Gradients ===") + print(f" dQ_Naive - Shape: {q_naive_grad_BLHD.shape}, Min: {q_naive_grad_BLHD.min().item():.6f}, Max: {q_naive_grad_BLHD.max().item():.6f}, Mean: {q_naive_grad_BLHD.mean().item():.6f}, Std: {q_naive_grad_BLHD.std().item():.6f}") + print(f" dK_Naive - Shape: {k_naive_grad_BLHD.shape}, Min: {k_naive_grad_BLHD.min().item():.6f}, Max: {k_naive_grad_BLHD.max().item():.6f}, Mean: {k_naive_grad_BLHD.mean().item():.6f}, Std: {k_naive_grad_BLHD.std().item():.6f}") + print(f" dV_Naive - Shape: {v_naive_grad_BLHD.shape}, Min: {v_naive_grad_BLHD.min().item():.6f}, Max: {v_naive_grad_BLHD.max().item():.6f}, Mean: {v_naive_grad_BLHD.mean().item():.6f}, Std: {v_naive_grad_BLHD.std().item():.6f}") + + # Print sample values (first few elements) + # print("\n === Sample Gradient Values (first 10 elements) ===") + # print(f" dQ_QAT[0,0,0,:10]: {q_BLHD.grad[0,0,0,:10].cpu().tolist()}") + # print(f" dQ_Naive[0,0,0,:10]: {q_naive_grad_BLHD[0,0,0,:10].cpu().tolist()}") + # print(f" dK_QAT[0,0,0,:10]: {k_BLHD.grad[0,0,0,:10].cpu().tolist()}") + # print(f" dK_Naive[0,0,0,:10]: {k_naive_grad_BLHD[0,0,0,:10].cpu().tolist()}") + # print(f" dV_QAT[0,0,0,:10]: {v_BLHD.grad[0,0,0,:10].cpu().tolist()}") + # print(f" dV_Naive[0,0,0,:10]: {v_naive_grad_BLHD[0,0,0,:10].cpu().tolist()}") + + # ----------------------------- + # COMPARE + # ----------------------------- + dq_diff = (q_BLHD.grad - q_naive_grad_BLHD).abs() + dk_diff = (k_BLHD.grad - k_naive_grad_BLHD).abs() + dv_diff = (v_BLHD.grad - v_naive_grad_BLHD).abs() + + dq_cos = cosine_similarity(q_BLHD.grad, q_naive_grad_BLHD) + dk_cos = cosine_similarity(k_BLHD.grad, k_naive_grad_BLHD) + dv_cos = cosine_similarity(v_BLHD.grad, v_naive_grad_BLHD) + + print("\n === Gradient Comparison ===") + print(f" dQ - Max diff: {dq_diff.max().item():.6f}, Mean diff: {dq_diff.mean().item():.6f}, Cos: {dq_cos:.6f}") + print(f" dK - Max diff: {dk_diff.max().item():.6f}, Mean diff: {dk_diff.mean().item():.6f}, Cos: {dk_cos:.6f}") + print(f" dV - Max diff: {dv_diff.max().item():.6f}, Mean diff: {dv_diff.mean().item():.6f}, Cos: {dv_cos:.6f}") + + print("\n✓ WAN shape backward pass test passed.") + + + +def test_qat_attention_wan_shape_non_divisible_64(): + """Test QAT attention where N_CTX is not divisible by 64.""" + torch.manual_seed(42) + + Z, H, N_CTX, HEAD_DIM = 1, 40, 9367, 128 + dtype = torch.bfloat16 + sm_scale = 1.0 / sqrt(HEAD_DIM) + causal = False + + print(f" Testing WAN non-divisible: Z={Z}, H={H}, N_CTX={N_CTX}, HEAD_DIM={HEAD_DIM}") + print(f" N_CTX % 64 = {N_CTX % 64}") + + # ----------------------------- + # CREATE BLHD INPUTS + # ----------------------------- + q_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + k_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + v_BLHD = torch.randn((Z, N_CTX, H, HEAD_DIM), dtype=dtype, device=DEVICE, requires_grad=True) + + # ----------------------------- + # BHLD for naive + # ----------------------------- + q_BHLD = q_BLHD.detach().permute(0,2,1,3).contiguous().requires_grad_(True) + k_BHLD = k_BLHD.detach().permute(0,2,1,3).contiguous().requires_grad_(True) + v_BHLD = v_BLHD.detach().permute(0,2,1,3).contiguous().requires_grad_(True) + + # ----------------------------- + # FORWARD + # ----------------------------- + out_qat_BLHD = qat_attn_wrapper(q_BLHD, k_BLHD, v_BLHD, causal, sm_scale) + out_naive_BHLD = naive_attention(q_BHLD, k_BHLD, v_BHLD, causal, sm_scale) + out_naive_BLHD = out_naive_BHLD.permute(0,2,1,3).contiguous() + + # ----------------------------- + # BACKWARD + # ----------------------------- + dout_BLHD = torch.randn_like(out_qat_BLHD).contiguous() + + # QAT backward + out_qat_BLHD.backward(dout_BLHD) + + # FIXED: Convert dout to BHLD for naive + dout_BHLD = dout_BLHD.permute(0,2,1,3).contiguous() + + out_naive_BHLD.backward(dout_BHLD) + + # ----------------------------- + # GRADIENT FORMAT FIX + # ----------------------------- + q_naive_grad_BLHD = q_BHLD.grad.permute(0,2,1,3).contiguous() + k_naive_grad_BLHD = k_BHLD.grad.permute(0,2,1,3).contiguous() + v_naive_grad_BLHD = v_BHLD.grad.permute(0,2,1,3).contiguous() + + # ----------------------------- + # PRINT GRADIENTS + # ----------------------------- + print("\n === QAT Gradients ===") + print(f" dQ_QAT - Shape: {q_BLHD.grad.shape}, Min: {q_BLHD.grad.min().item():.6f}, Max: {q_BLHD.grad.max().item():.6f}, Mean: {q_BLHD.grad.mean().item():.6f}, Std: {q_BLHD.grad.std().item():.6f}") + print(f" dK_QAT - Shape: {k_BLHD.grad.shape}, Min: {k_BLHD.grad.min().item():.6f}, Max: {k_BLHD.grad.max().item():.6f}, Mean: {k_BLHD.grad.mean().item():.6f}, Std: {k_BLHD.grad.std().item():.6f}") + print(f" dV_QAT - Shape: {v_BLHD.grad.shape}, Min: {v_BLHD.grad.min().item():.6f}, Max: {v_BLHD.grad.max().item():.6f}, Mean: {v_BLHD.grad.mean().item():.6f}, Std: {v_BLHD.grad.std().item():.6f}") + + print("\n === Naive Gradients ===") + print(f" dQ_Naive - Shape: {q_naive_grad_BLHD.shape}, Min: {q_naive_grad_BLHD.min().item():.6f}, Max: {q_naive_grad_BLHD.max().item():.6f}, Mean: {q_naive_grad_BLHD.mean().item():.6f}, Std: {q_naive_grad_BLHD.std().item():.6f}") + print(f" dK_Naive - Shape: {k_naive_grad_BLHD.shape}, Min: {k_naive_grad_BLHD.min().item():.6f}, Max: {k_naive_grad_BLHD.max().item():.6f}, Mean: {k_naive_grad_BLHD.mean().item():.6f}, Std: {k_naive_grad_BLHD.std().item():.6f}") + print(f" dV_Naive - Shape: {v_naive_grad_BLHD.shape}, Min: {v_naive_grad_BLHD.min().item():.6f}, Max: {v_naive_grad_BLHD.max().item():.6f}, Mean: {v_naive_grad_BLHD.mean().item():.6f}, Std: {v_naive_grad_BLHD.std().item():.6f}") + + # Print sample values (first few elements) + # print("\n === Sample Gradient Values (first 10 elements) ===") + # print(f" dQ_QAT[0,0,0,:10]: {q_BLHD.grad[0,0,0,:10].cpu().tolist()}") + # print(f" dQ_Naive[0,0,0,:10]: {q_naive_grad_BLHD[0,0,0,:10].cpu().tolist()}") + # print(f" dK_QAT[0,0,0,:10]: {k_BLHD.grad[0,0,0,:10].cpu().tolist()}") + # print(f" dK_Naive[0,0,0,:10]: {k_naive_grad_BLHD[0,0,0,:10].cpu().tolist()}") + # print(f" dV_QAT[0,0,0,:10]: {v_BLHD.grad[0,0,0,:10].cpu().tolist()}") + # print(f" dV_Naive[0,0,0,:10]: {v_naive_grad_BLHD[0,0,0,:10].cpu().tolist()}") + + # ----------------------------- + # COMPARE + # ----------------------------- + dq_diff = (q_BLHD.grad - q_naive_grad_BLHD).abs() + dk_diff = (k_BLHD.grad - k_naive_grad_BLHD).abs() + dv_diff = (v_BLHD.grad - v_naive_grad_BLHD).abs() + + dq_cos = cosine_similarity(q_BLHD.grad, q_naive_grad_BLHD) + dk_cos = cosine_similarity(k_BLHD.grad, k_naive_grad_BLHD) + dv_cos = cosine_similarity(v_BLHD.grad, v_naive_grad_BLHD) + + print("\n === Gradient Comparison ===") + print(f" dQ - Max diff: {dq_diff.max().item():.6f}, Mean: {dq_diff.mean().item():.6f}, Cos: {dq_cos:.6f}") + print(f" dK - Max diff: {dk_diff.max().item():.6f}, Mean: {dk_diff.mean().item():.6f}, Cos: {dk_cos:.6f}") + print(f" dV - Max diff: {dv_diff.max().item():.6f}, Mean: {dv_diff.mean().item():.6f}, Cos: {dv_cos:.6f}") + + print("\n✓ WAN shape non-divisible-by-64 forward/backward test passed.") + + +if __name__ == "__main__": + print("Running QAT attention tests...") + print(f"Device: {DEVICE}") + print() + + test_qat_attention_forward() + test_qat_attention_backward() + test_qat_attention_different_shapes() + test_qat_attention_non_causal() + test_qat_attention_causal() + test_qat_attention_causal_backward() + test_qat_attention_different_seq_lengths() + test_qat_attention_different_seq_lengths_backward() + test_qat_attention_wan_shape_forward() + test_qat_attention_wan_shape_backward() + test_qat_attention_wan_shape_non_divisible_64() + + print() + print("Running Fused attention tests...") + print() + + # test_fused_attention_forward() + # test_fused_attention_backward() + # test_fused_attention_different_shapes() + # test_fused_attention_non_causal() + # test_fused_attention_causal() + # test_fused_attention_causal_backward() + + # print() + # print("Running Fused vs QAT attention comparison tests...") + # print() + + # test_fused_vs_qat_attention_forward() + # test_fused_vs_qat_attention_backward() + # test_fused_vs_qat_attention_different_shapes() + # test_fused_vs_qat_attention_non_causal() + # test_fused_vs_qat_attention_causal() + # test_fused_vs_qat_attention_causal_backward() + + print() + print("All tests passed! ✓") \ No newline at end of file diff --git a/test_quantization_precision.py b/test_quantization_precision.py new file mode 100644 index 000000000..c72a01ab3 --- /dev/null +++ b/test_quantization_precision.py @@ -0,0 +1,461 @@ +#!/usr/bin/env python3 +""" +Precision test to compare numerical differences between two quantization methods: +1. dequant(quant(X) @ quant(W)) +2. dequant(quant(X)) @ dequant(quant(W)) + +Uses FlashInfer API for 4-bit quantization producing bf16 matrices. +""" + +import torch +import numpy as np +from flashinfer import SfLayout, mm_fp4, nvfp4_quantize, e2m1_and_ufp8sf_scale_to_float +import unittest +from typing import Tuple, Dict, Any + +import torch +import triton +import triton.language as tl +import triton.profiler as proton +from triton.tools.tensor_descriptor import TensorDescriptor +from triton.tools.mxfp import MXFP4Tensor, MXScaleTensor + +def get_fake_quant(x: torch.Tensor): + # TODO: dtype needs to be float32? + orig_shape = x.shape + orig_dtype = x.dtype + device = x.device + x = x.view(-1, x.shape[-1]) + x_global_sf = (448 * 6) / x.float().abs().nan_to_num().max() + x_fp4, x_scale = nvfp4_quantize(x, x_global_sf, sfLayout=SfLayout.layout_128x4, do_shuffle=False) + x_dequant = e2m1_and_ufp8sf_scale_to_float(x_fp4, x_scale, 1 / x_global_sf) # Note: this function does dequantization on the cpu + return x_dequant.view(orig_shape).to(orig_dtype).to(device) + + +def generate_fake_quant(a_f32: torch.Tensor, + b_f32: torch.Tensor, + vec_size: int = 16, + max_mxf4: float = 6.0, + eps: float = 1e-8): + """ + Fake-quantize fp32 A (M,K) and B (K,N) with nvfp4 (E2M1) using FP8 E4M3 (finite-only) block scales. + Returns dequantized fp32 tensors (A_qdq, B_qdq). Assumes: + - M % 128 == 0, N % 128 == 0, K % (vec_size * 4) == 0 + - A is (M, K), B is (K, N) + - Scales are uniform per 128x64 (A) and 64x128 (B) tiles (matching your kernel’s layout) + """ + assert a_f32.is_cuda and b_f32.is_cuda + M, K = a_f32.shape + Kb, N = b_f32.shape + assert Kb == K, "A and B inner dims must match (A:(M,K), B:(K,N))" + assert M % 128 == 0 and N % 128 == 0, "M,N must be multiples of 128" + assert K % (vec_size * 4) == 0, f"K must be multiple of {vec_size*4}" + + device = a_f32.device + + # ---- tile-scale helpers (produce FP8 E4M3 scales with shape [*, *, 32, 16]) ---- + def _tile_scales_A(a_fp32): + m_chunks = M // 128 + k_chunks = K // (vec_size * 4) # 64-wide tiles along K + x = a_fp32.view(m_chunks, 128, k_chunks, 4, vec_size) + tile_max = x.abs().amax(dim=(1, 3, 4)) # [m_chunks, k_chunks] + scale = torch.clamp(tile_max / max_mxf4, min=eps) # fp32 + s = scale[:, :, None, None].expand(m_chunks, k_chunks, 32, 16) + return s.to(torch.float8_e4m3fn) # [m_chunks, k_chunks, 32, 16] + + def _tile_scales_B(b_fp32): + k_chunks = K // (vec_size * 4) + n_chunks = N // 128 + x = b_fp32.view(k_chunks, 4, vec_size, n_chunks, 128) + tile_max = x.abs().amax(dim=(1, 2, 4)) # [k_chunks, n_chunks] + scale = torch.clamp(tile_max / max_mxf4, min=eps).t() # [n_chunks, k_chunks] + s = scale[:, :, None, None].expand(n_chunks, k_chunks, 32, 16) + return s.to(torch.float8_e4m3fn) # [n_chunks, k_chunks, 32, 16] + + # ---- expand tile scales to per-element fp32 ---- + def _expand_A_scale_full(a_scale_fp8): + m_chunks, k_chunks, _, _ = a_scale_fp8.shape + base = a_scale_fp8[..., 0, 0].to(a_f32.dtype) # [m_chunks, k_chunks] + tile = base[:, :, None, None].expand(m_chunks, k_chunks, 128, 64) + return tile.permute(0, 2, 1, 3).reshape(M, k_chunks * 64).contiguous() # [M,K] + + def _expand_B_scale_full(b_scale_fp8): + n_chunks, k_chunks, _, _ = b_scale_fp8.shape + base = b_scale_fp8[..., 0, 0].to(b_f32.dtype) # [n_chunks, k_chunks] + tile = base[:, :, None, None].expand(n_chunks, k_chunks, 64, 128) + return tile.permute(1, 2, 0, 3).reshape(k_chunks * 64, n_chunks * 128).contiguous() # [K,N] + + # ---- 1) compute FP8 E4M3 tile scales from fp32 ---- + a_scale_fp8 = _tile_scales_A(a_f32) # [M//128, K//64, 32, 16] + b_scale_fp8 = _tile_scales_B(b_f32) # [N//128, K//64, 32, 16] + + # ---- 2) expand scales to full fp32 for quant/dequant math ---- + a_scale_full = _expand_A_scale_full(a_scale_fp8) # [M, K] fp32 + b_scale_full = _expand_B_scale_full(b_scale_fp8) # [K, N] fp32 + + # ---- 3) quantize to nvfp4 codes (E2M1) ---- + a_codes = MXFP4Tensor(data=(a_f32 / a_scale_full)).data # [M, K] uint8 (low nibble) + b_codes = MXFP4Tensor(data=(b_f32 / b_scale_full)).data # [K, N] uint8 + + # ---- 4) dequantize back to fp32 (fake-quant output) ---- + a_qdq = MXFP4Tensor(size=(M, K), device=device); a_qdq.data = a_codes + b_qdq = MXFP4Tensor(size=(K, N), device=device); b_qdq.data = b_codes + + a_qdq = a_qdq.to(torch.float32).to(a_f32.dtype) * a_scale_full + b_qdq = b_qdq.to(torch.float32).to(b_f32.dtype) * b_scale_full + + return a_qdq, b_qdq + + +class QuantizationPrecisionTest: + """Test class for comparing quantization precision methods.""" + + def __init__(self, device: str = "cuda", dtype: torch.dtype = torch.bfloat16): + self.device = device + self.dtype = dtype + self.sf_layout = SfLayout.layout_128x4 + + def _compute_scale_factor(self, tensor: torch.Tensor) -> torch.Tensor: + """Compute scale factor for quantization using the same method as FastVideo.""" + maxabs = tensor.float().abs().nan_to_num().max() + maxabs = torch.maximum(maxabs, torch.tensor(1e-12, device=tensor.device, dtype=maxabs.dtype)) + return (448.0 * 6.0) / maxabs + + def quantize_matrix(self, matrix: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Quantize a matrix to 4-bit using FlashInfer. + + Args: + matrix: Input matrix to quantize + + Returns: + Tuple of (quantized_matrix, inverse_scale_factor) + """ + # scale_factor = self._compute_scale_factor(matrix) + scale_factor = torch.tensor(1.0, device=matrix.device, dtype=torch.float32) + quantized, inv_scale = nvfp4_quantize( + matrix, + scale_factor, + sfLayout=self.sf_layout, + do_shuffle=False + ) + return quantized, inv_scale, scale_factor + + def dequantize_matrix(self, shape: Tuple[int, int], quantized_matrix: torch.Tensor, inv_scale: torch.Tensor, global_scale_factor: torch.Tensor) -> torch.Tensor: + """ + Dequantize a 4-bit matrix back to bf16. + + Note: FlashInfer doesn't provide direct dequantization, so we use a workaround + by multiplying with an identity matrix to extract the dequantized values. + + Args: + quantized_matrix: 4-bit quantized matrix + inv_scale: Inverse scale factor from quantization + + Returns: + Dequantized matrix in bf16 + """ + # Create identity matrix for dequantization + if len(quantized_matrix.shape) == 2: + # 2D case + identity = torch.eye(shape[1], device=self.device, dtype=self.dtype) + identity_fp4, identity_inv_scale = nvfp4_quantize( + identity, + torch.tensor(1.0, device=self.device, dtype=torch.float32), + sfLayout=self.sf_layout, + do_shuffle=False + ) + dequantized = torch.empty(shape, device=self.device, dtype=self.dtype) + mm_fp4( + quantized_matrix, identity_fp4.T, inv_scale, identity_inv_scale.T, + 1 / global_scale_factor, + self.dtype, dequantized, + block_size=16, + use_8x4_sf_layout=False, + backend="cutlass" + ) + else: + # 3D case (batch dimension) + identity = torch.eye(quantized_matrix.shape[-1], device=self.device, dtype=self.dtype) + dequantized = torch.empty_like(quantized_matrix) + for i in range(quantized_matrix.shape[0]): + mm_fp4( + quantized_matrix[i], identity, inv_scale, inv_scale, + torch.tensor(1.0, device=self.device, dtype=torch.float32), + self.dtype, dequantized[i], + block_size=16, + use_8x4_sf_layout=False, + backend="cutlass" + ) + + return dequantized + + def real_quant_matmul(self, X: torch.Tensor, W: torch.Tensor) -> torch.Tensor: + """ + Method 1: dequant(quant(X) @ quant(W)) + Perform matrix multiplication in quantized space, then dequantize. + """ + # Quantize both matrices + X_quant, X_inv_scale, X_global_scale_factor = self.quantize_matrix(X) + W_quant, W_inv_scale, W_global_scale_factor = self.quantize_matrix(W) + + # Compute alpha scaling factor + alpha = 1.0 / (X_global_scale_factor * W_global_scale_factor) + + result = torch.empty((X.shape[0], W.shape[0]), device=self.device, dtype=self.dtype) + mm_fp4( + X_quant, W_quant.T, X_inv_scale, W_inv_scale.T, alpha, + self.dtype, result, + block_size=16, + use_8x4_sf_layout=False, + backend="cutlass" + ) + + + return result + + def fake_quant_matmul(self, X: torch.Tensor, W: torch.Tensor) -> torch.Tensor: + """ + Method 2: dequant(quant(X)) @ dequant(quant(W)) + Dequantize both matrices first, then perform matrix multiplication. + """ + # Quantize both matrices + + X_quant, X_inv_scale, X_global_scale_factor = self.quantize_matrix(X) + W_quant, W_inv_scale, W_global_scale_factor = self.quantize_matrix(W) + # Dequantize both matrices + X_dequant = self.dequantize_matrix(X.shape, X_quant, X_inv_scale, X_global_scale_factor) + W_dequant = self.dequantize_matrix(W.shape, W_quant, W_inv_scale, W_global_scale_factor) + + result = torch.matmul(X_dequant, W_dequant.T) + + + # X_fake_quant, W_fake_quant_T = generate_fake_quant(X, W.T.contiguous()) + # import pdb; pdb.set_trace() + # result = torch.matmul(X_fake_quant, W_fake_quant_T) + + return result + + def compare_methods(self, X: torch.Tensor, W: torch.Tensor) -> Dict[str, Any]: + """ + Compare the two quantization methods and return detailed analysis. + + Args: + X: Input matrix X + W: Weight matrix W + + Returns: + Dictionary containing comparison results + """ + # Ensure matrices are on correct device and dtype + X = X.to(device=self.device, dtype=self.dtype) + W = W.to(device=self.device, dtype=self.dtype) + + # Compute results using both methods + result1 = self.real_quant_matmul(X, W) + result2 = self.fake_quant_matmul(X, W) + import pdb; pdb.set_trace() + # Compute reference result (no quantization) + reference = torch.matmul(X, W.T) + # Calculate differences + diff_methods = torch.abs(result1 - result2) + diff1_ref = torch.abs(result1 - reference) + diff2_ref = torch.abs(result2 - reference) + + # Statistical analysis + analysis = { + 'method1_result': result1, + 'method2_result': result2, + 'reference_result': reference, + 'method_difference': { + 'max': torch.max(diff_methods).item(), + 'mean': torch.mean(diff_methods).item(), + 'std': torch.std(diff_methods).item(), + 'relative_max': torch.max(diff_methods / (torch.abs(reference) + 1e-8)).item(), + 'relative_mean': torch.mean(diff_methods / (torch.abs(reference) + 1e-8)).item(), + }, + 'method1_vs_reference': { + 'max': torch.max(diff1_ref).item(), + 'mean': torch.mean(diff1_ref).item(), + 'relative_max': torch.max(diff1_ref / (torch.abs(reference) + 1e-8)).item(), + }, + 'method2_vs_reference': { + 'max': torch.max(diff2_ref).item(), + 'mean': torch.mean(diff2_ref).item(), + 'relative_max': torch.max(diff2_ref / (torch.abs(reference) + 1e-8)).item(), + }, + 'input_shapes': { + 'X': list(X.shape), + 'W': list(W.shape), + } + } + + return analysis + + def run_comprehensive_test(self) -> Dict[str, Any]: + """Run comprehensive tests with different matrix sizes and distributions.""" + test_results = {} + + # Test cases with different sizes and distributions + test_cases = [ + # (X_shape, W_shape, distribution_name) + ((128, 128), (256, 128), "small_matrices"), + ((256, 512), (1024, 512), "medium_matrices"), + ((1024, 2048), (4096, 2048), "large_matrices"), + ] + + distributions = [ + ("normal", lambda shape: torch.randn(shape, device=self.device)), + ("uniform", lambda shape: torch.rand(shape, device=self.device) * 2 - 1), + ] + + for x_shape, w_shape, size_name in test_cases: + test_results[size_name] = {} + + for dist_name, dist_func in distributions: + # Generate test matrices + X = dist_func(x_shape).to(dtype=self.dtype) + W = dist_func(w_shape).to(dtype=self.dtype) + + # Run comparison + analysis = self.compare_methods(X, W) + test_results[size_name][dist_name] = analysis + + print(f"Test: {size_name} - {dist_name}") + print(f" Max difference between methods: {analysis['method_difference']['max']:.6f}") + print(f" Mean difference between methods: {analysis['method_difference']['mean']:.6f}") + + print(f" Max difference between method1 (real quant) and reference: {analysis['method1_vs_reference']['max']:.6f}") + print(f" Mean difference between method1 (real quant) and reference: {analysis['method1_vs_reference']['mean']:.6f}") + print(f" Max difference between method2 (fake quant) and reference: {analysis['method2_vs_reference']['max']:.6f}") + print(f" Mean difference between method2 (fake quant) and reference: {analysis['method2_vs_reference']['mean']:.6f}") + print() + + return test_results + + + +def main(): + """Main function to run the precision test.""" + print("Starting Quantization Precision Test") + print("=" * 50) + + # Check if CUDA is available + if not torch.cuda.is_available(): + print("CUDA not available, using CPU (may be slower)") + device = "cpu" + else: + device = "cuda" + print(f"Using CUDA device: {torch.cuda.get_device_name()}") + + # Create tester + tester = QuantizationPrecisionTest(device=device) + + # Run comprehensive test + print("\nRunning comprehensive tests...") + results = tester.run_comprehensive_test() + + +def test_get_fake_quant(): + """Test the get_fake_quant function with various inputs.""" + print("Testing get_fake_quant function...") + + if not torch.cuda.is_available(): + print("CUDA not available, skipping test") + return + + device = "cuda" + dtype = torch.bfloat16 + + # Test 1: Basic functionality - 2D tensor + print("\nTest 1: Basic 2D tensor") + x = torch.randn(128, 64, device=device, dtype=dtype) + x_fq = get_fake_quant(x) + assert x_fq.shape == x.shape, f"Shape mismatch: {x_fq.shape} vs {x.shape}" + assert x_fq.dtype == x.dtype, f"Dtype mismatch: {x_fq.dtype} vs {x.dtype}" + assert x_fq.device == x.device, f"Device mismatch: {x_fq.device} vs {x.device}" + max_diff = (x - x_fq).abs().max().item() + print(f" Max difference: {max_diff:.6f}") + assert max_diff < 1.0, f"Difference too large: {max_diff}" + + # Test 2: 3D tensor (B, H, D) + print("\nTest 2: 3D tensor [B, H, D]") + x = torch.randn(2, 8, 128, device=device, dtype=dtype) + x_fq = get_fake_quant(x) + assert x_fq.shape == x.shape, f"Shape mismatch: {x_fq.shape} vs {x.shape}" + assert x_fq.dtype == x.dtype, f"Dtype mismatch: {x_fq.dtype} vs {x.dtype}" + max_diff = (x - x_fq).abs().max().item() + print(f" Max difference: {max_diff:.6f}") + assert max_diff < 1.0, f"Difference too large: {max_diff}" + + # Test 3: 4D tensor (B, H, L, D) + print("\nTest 3: 4D tensor [B, H, L, D]") + x = torch.randn(1, 8, 256, 128, device=device, dtype=dtype) + x_fq = get_fake_quant(x) + assert x_fq.shape == x.shape, f"Shape mismatch: {x_fq.shape} vs {x.shape}" + assert x_fq.dtype == x.dtype, f"Dtype mismatch: {x_fq.dtype} vs {x.dtype}" + max_diff = (x - x_fq).abs().max().item() + print(f" Max difference: {max_diff:.6f}") + assert max_diff < 1.0, f"Difference too large: {max_diff}" + + # Test 4: Small values + print("\nTest 4: Small values") + x = torch.randn(128, 64, device=device, dtype=dtype) * 0.01 + x_fq = get_fake_quant(x) + assert x_fq.shape == x.shape + max_diff = (x - x_fq).abs().max().item() + print(f" Max difference: {max_diff:.6f}") + + # Test 5: Large values + print("\nTest 5: Large values") + x = torch.randn(128, 64, device=device, dtype=dtype) * 10.0 + x_fq = get_fake_quant(x) + assert x_fq.shape == x.shape + max_diff = (x - x_fq).abs().max().item() + print(f" Max difference: {max_diff:.6f}") + + # Test 6: Zeros + print("\nTest 6: Zeros") + x = torch.zeros(128, 64, device=device, dtype=dtype) + x_fq = get_fake_quant(x) + assert x_fq.shape == x.shape + max_diff = (x - x_fq).abs().max().item() + print(f" Max difference: {max_diff:.6f}") + assert max_diff < 1e-5, f"Zeros should remain zeros: {max_diff}" + + # Test 7: Different dtypes + print("\nTest 7: Float16 dtype") + x = torch.randn(128, 64, device=device, dtype=torch.float16) + x_fq = get_fake_quant(x) + assert x_fq.shape == x.shape + assert x_fq.dtype == x.dtype + max_diff = (x - x_fq).abs().max().item() + print(f" Max difference: {max_diff:.6f}") + + # Test 8: Idempotency check (quantizing twice should be similar) + print("\nTest 8: Idempotency") + x = torch.randn(128, 64, device=device, dtype=dtype) + x_fq1 = get_fake_quant(x) + x_fq2 = get_fake_quant(x_fq1) + # Second quantization should be close to first + max_diff = (x_fq1 - x_fq2).abs().max().item() + print(f" Max difference between two quantizations: {max_diff:.6f}") + + # Test 9: Requires grad + print("\nTest 9: Gradient flow") + x = torch.randn(128, 64, device=device, dtype=dtype, requires_grad=True) + x_fq = get_fake_quant(x) + # Note: get_fake_quant uses no_grad internally, so x_fq won't have grad + # But we can check it doesn't crash + assert x_fq.shape == x.shape + print(" Gradient test passed (no crash)") + + print("\n✓ All tests passed!") + + +if __name__ == "__main__": + import sys + if len(sys.argv) > 1 and sys.argv[1] == "test_fake_quant": + test_get_fake_quant() + else: + main() diff --git a/test_scale_debug.py b/test_scale_debug.py new file mode 100644 index 000000000..4785b7313 --- /dev/null +++ b/test_scale_debug.py @@ -0,0 +1,92 @@ +#!/usr/bin/env python3 +""" +Debug script to understand the scale computation issue. +""" + +import torch +import triton +import triton.language as tl +from nvfp4_utils import _compute_quant_and_scale, _compute_dequant + +DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") + +@triton.jit +def debug_scale_kernel( + src_ptr, + scale_ptr, + s_dec_ptr, + BLOCK_SIZE_OUT_DIM: tl.constexpr, + BLOCK_SIZE_QUANT_DIM: tl.constexpr, + stride_src_outer, + stride_src_quant, + outer_dim, + quant_dim, +): + """Debug kernel to print scale values.""" + outer_idx = tl.program_id(0) + quant_idx = tl.program_id(1) + + start_outer = outer_idx * BLOCK_SIZE_OUT_DIM + start_quant = quant_idx * BLOCK_SIZE_QUANT_DIM + + offs_outer = tl.arange(0, BLOCK_SIZE_OUT_DIM)[:, None] + offs_quant = tl.arange(0, BLOCK_SIZE_QUANT_DIM)[None, :] + + mask_outer = (start_outer + offs_outer) < outer_dim + mask_quant = (start_quant + offs_quant) < quant_dim + full_mask = mask_outer & mask_quant + + src_offsets = (start_outer + offs_outer) * stride_src_outer + (start_quant + offs_quant) * stride_src_quant + src_tensor = tl.load(src_ptr + src_offsets, mask=full_mask, other=0.0) + + valid_mask = tl.full(src_tensor.shape, 1, tl.int1) + quantized_tensor, scale_tensor, s_dec = _compute_quant_and_scale( + src_tensor=src_tensor, + valid_src_mask=valid_mask, + mx_tensor_dtype=tl.float8e4nv + ) + + # Store scale and s_dec for inspection + # scale_tensor has shape [BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE] where BLOCK_SIZE_QUANT_MX_SCALE = BLOCK_SIZE_QUANT_DIM // 16 + BLOCK_SIZE_QUANT_MX_SCALE: tl.constexpr = BLOCK_SIZE_QUANT_DIM // 16 + scale_offsets = (start_outer + offs_outer) * stride_src_outer + (start_quant // 16) * stride_src_quant + tl.store(scale_ptr + scale_offsets, scale_tensor.reshape([BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE])[:, 0].to(tl.float32), mask=mask_outer) + tl.store(s_dec_ptr + outer_idx, s_dec) + + +if __name__ == "__main__": + if not torch.cuda.is_available(): + print("CUDA not available") + exit(1) + + # Test with all ones + x = torch.ones(128, 64, device=DEVICE, dtype=torch.float16) + x_2d = x.view(-1, x.shape[-1]) + outer_dim, quant_dim = x_2d.shape + + scale_output = torch.zeros(outer_dim, quant_dim // 16, device=DEVICE, dtype=torch.float32) + s_dec_output = torch.zeros(outer_dim // 128, device=DEVICE, dtype=torch.float32) + + grid = ( + triton.cdiv(outer_dim, 128), + triton.cdiv(quant_dim, 64), + ) + + debug_scale_kernel[grid]( + src_ptr=x_2d, + scale_ptr=scale_output, + s_dec_ptr=s_dec_output, + BLOCK_SIZE_OUT_DIM=128, + BLOCK_SIZE_QUANT_DIM=64, + stride_src_outer=x_2d.stride(0), + stride_src_quant=x_2d.stride(1), + outer_dim=outer_dim, + quant_dim=quant_dim, + ) + + print(f"Scale values (first block): {scale_output[0, :4]}") + print(f"s_dec value: {s_dec_output[0]}") + print(f"Expected s_dec: {1.0 / (6 * 448 / 1.0)} = {1.0 / 2688}") + print(f"Expected scale: {(1.0 / 6) * 2688} = {2688 / 6} = {448}") + + diff --git a/test_triton_fp4_vs_cutlass.py b/test_triton_fp4_vs_cutlass.py new file mode 100644 index 000000000..c028425fb --- /dev/null +++ b/test_triton_fp4_vs_cutlass.py @@ -0,0 +1,276 @@ +import torch +from triton.tools.tensor_descriptor import TensorDescriptor +from triton.tools.mxfp import MXFP4Tensor +from triton_fp4 import block_scaled_matmul +from flashinfer import SfLayout, mm_fp4, nvfp4_quantize + + + +def triton_nvfp4(A: torch.Tensor, + B: torch.Tensor, + out_dtype: torch.dtype = torch.float16): + """ + Block-scaled nvfp4 GEMM: (A @ B) with A:(M,K), B:(K,N) given in fp32/bf16. + Returns the GEMM output in `out_dtype`. + + Requirements (match your kernel): + - M % 128 == 0, N % 128 == 0, K % 64 == 0 (since vec_size=16 and 4 groups → 64) + - Device: CUDA, Blackwell (SM=10) + """ + assert A.is_cuda and B.is_cuda, "Inputs must be on CUDA" + assert A.dtype in (torch.float32, torch.bfloat16), "A must be fp32/bf16" + assert B.dtype in (torch.float32, torch.bfloat16), "B must be fp32/bf16" + M, K = A.shape + Kb, N = B.shape + assert Kb == K, "Inner dims must match" + assert M % 128 == 0 and N % 128 == 0 and K % 64 == 0, "Shape must respect tiling" + + # ---- kernel constants / config ---- + BLOCK_M = 128 + BLOCK_N = 256 + BLOCK_K = 256 + VEC_SIZE = 16 + ELEM_PER_BYTE_A = 2 + ELEM_PER_BYTE_B = 2 + MAX_MXF4 = 6.0 # nvfp4 (E2M1) max magnitude + EPS = 1e-8 + + device = A.device + + # ---- tile-scale helpers (FP8 E4M3, finite-only) ---- + # A tiles are 128 x 64 along (M,K) + def _tile_scales_A(a_fp32): + m_chunks = M // 128 + k_chunks = K // (VEC_SIZE * 4) # 64 wide + x = a_fp32.view(m_chunks, 128, k_chunks, 4, VEC_SIZE) + tile_max = x.abs().amax(dim=(1, 3, 4)) # [m_chunks, k_chunks] + scale = torch.clamp(tile_max / MAX_MXF4, min=EPS) # fp32 + s = scale[:, :, None, None].expand(m_chunks, k_chunks, 32, 16) + return s.to(torch.float8_e4m3fn) # [m_chunks, k_chunks, 32, 16] + + # B tiles are 64 x 128 over (K,N) + def _tile_scales_B(b_fp32): + k_chunks = K // (VEC_SIZE * 4) + n_chunks = N // 128 + x = b_fp32.view(k_chunks, 4, VEC_SIZE, n_chunks, 128) + tile_max = x.abs().amax(dim=(1, 2, 4)) # [k_chunks, n_chunks] + scale = torch.clamp(tile_max / MAX_MXF4, min=EPS).t() # [n_chunks, k_chunks] + s = scale[:, :, None, None].expand(n_chunks, k_chunks, 32, 16) + return s.to(torch.float8_e4m3fn) # [n_chunks, k_chunks, 32, 16] + + def _expand_A_scale_full(a_scale_fp8): + m_chunks, k_chunks, _, _ = a_scale_fp8.shape + base = a_scale_fp8[..., 0, 0].to(A.dtype) # [m_chunks, k_chunks] + tile = base[:, :, None, None].expand(m_chunks, k_chunks, 128, 64) + return tile.permute(0, 2, 1, 3).reshape(M, k_chunks * 64).contiguous() # [M,K] + + def _expand_B_scale_full(b_scale_fp8): + n_chunks, k_chunks, _, _ = b_scale_fp8.shape + base = b_scale_fp8[..., 0, 0].to(B.dtype) # [n_chunks, k_chunks] + tile = base[:, :, None, None].expand(n_chunks, k_chunks, 64, 128) + return tile.permute(1, 2, 0, 3).reshape(k_chunks * 64, n_chunks * 128).contiguous() # [K,N] + + # ---- 1) scales from fp32 ---- + a_scale_fp8 = _tile_scales_A(A) # [M//128, K//64, 32, 16] + b_scale_fp8 = _tile_scales_B(B) # [N//128, K//64, 32, 16] + a_scale_full = _expand_A_scale_full(a_scale_fp8) # [M,K] fp32 + b_scale_full = _expand_B_scale_full(b_scale_fp8) # [K,N] fp32 + + # ---- 2) quantize to nvfp4 codes (E2M1) ---- + a_codes = MXFP4Tensor(data=(A / a_scale_full)).data # [M,K] uint8 (low nibble used) + b_codes_KN = MXFP4Tensor(data=(B / b_scale_full)).data # [K,N] uint8 + + # Kernel expects B laid out as (N, K) packed along K (like the original example) + b_codes_NK = b_codes_KN.transpose(0, 1).contiguous() # [N,K] + + # ---- 3) pack to 2-per-byte along K and build descriptors ---- + a_fp4 = MXFP4Tensor(size=(M, K), device=device); a_fp4.data = a_codes + b_fp4_desc = MXFP4Tensor(size=(N, K), device=device); b_fp4_desc.data = b_codes_NK + + a_packed = a_fp4.to_packed_tensor(dim=1) # (M, K/2) + b_packed = b_fp4_desc.to_packed_tensor(dim=1) # (N, K/2) + + a_desc = TensorDescriptor.from_tensor(a_packed, [BLOCK_M, BLOCK_K // ELEM_PER_BYTE_A]) + b_desc = TensorDescriptor.from_tensor(b_packed, [BLOCK_N, BLOCK_K // ELEM_PER_BYTE_B]) + + # reps for scale swizzle + rep_m = BLOCK_M // 128 + rep_n = BLOCK_N // 128 + rep_k = BLOCK_K // VEC_SIZE // 4 + + # reshape FP8 scales into 5D TMA blocks [1, rep_*, rep_k, 2, 256] + a_scale_5d = a_scale_fp8.reshape(1, a_scale_fp8.shape[0], a_scale_fp8.shape[1], 2, 256) + b_scale_5d = b_scale_fp8.reshape(1, b_scale_fp8.shape[0], b_scale_fp8.shape[1], 2, 256) + a_scale_desc = TensorDescriptor.from_tensor(a_scale_5d, [1, rep_m, rep_k, 2, 256]) + b_scale_desc = TensorDescriptor.from_tensor(b_scale_5d, [1, rep_n, rep_k, 2, 256]) + + a_packed_cutlass, a_scale_fp8_cutlass = nvfp4_quantize( + A, torch.tensor(1.0, device=device, dtype=torch.float32), sfLayout=SfLayout.layout_128x4, do_shuffle=False + ) + import pdb; pdb.set_trace() + # ---- 4) launch GEMM ---- + configs = { + "BLOCK_SIZE_M": BLOCK_M, + "BLOCK_SIZE_N": BLOCK_N, + "BLOCK_SIZE_K": BLOCK_K, + "num_stages": 4, + "ELEM_PER_BYTE_A": ELEM_PER_BYTE_A, + "ELEM_PER_BYTE_B": ELEM_PER_BYTE_B, + "VEC_SIZE": VEC_SIZE, + } + # Uses your already-defined block_scaled_matmul(...) + C = block_scaled_matmul( + a_desc, a_scale_desc, + b_desc, b_scale_desc, + out_dtype, M, N, K, + rep_m, rep_n, rep_k, configs + ) + return C + + + +def cutlass_nvfp4(A: torch.Tensor, + B: torch.Tensor, + out_dtype: torch.dtype = torch.bfloat16, + sf_layout: SfLayout = SfLayout.layout_128x4, + block_size: int = 16, + do_shuffle: bool = False, + scale_factor_A: torch.Tensor | float | None = None, + scale_factor_B: torch.Tensor | float | None = None) -> torch.Tensor: + """ + FlashInfer/CUTLASS nvfp4 GEMM: returns (A @ B) using 4-bit E2M1 operands with per-tile scaling. + + Args: + A: (M, K) tensor, fp32 or bf16, CUDA. + B: (K, N) tensor, fp32 or bf16, CUDA. + out_dtype: torch.float32, torch.float16, or torch.bfloat16 (recommended). + sf_layout: FlashInfer scale-factor layout (default SfLayout.layout_128x4). + block_size: internal tiling, default 16. + do_shuffle: pass-through to nvfp4_quantize. + scale_factor_A, scale_factor_B: optional global scale factors (float or 0-dim tensor). + If None, defaults to 1.0 (FlashInfer still applies its internal per-tile scales). + + Returns: + C: (M, N) tensor in out_dtype + """ + assert A.is_cuda and B.is_cuda, "Inputs must be CUDA tensors" + assert A.dim() == 2 and B.dim() == 2, "A and B must be 2D" + M, K = A.shape + Kb, N = B.shape + assert K == Kb, f"Incompatible shapes: A is (M={M},K={K}), B is (K={Kb},N={N})" + + # Cast inputs to supported accumulation input types for FlashInfer quantizer. + # (fp32/bf16 are fine; we keep user dtype if already bf16 to save bandwidth.) + A_in = A if A.dtype in (torch.bfloat16, torch.float32) else A.float() + B_in = B if B.dtype in (torch.bfloat16, torch.float32) else B.float() + B_in = B_in.T.contiguous() + + # Choose global scaling (FlashInfer also uses per-tile scales internally). + if scale_factor_A is None: + scale_factor_A = torch.tensor(1.0, device=A.device, dtype=torch.float32) + elif not torch.is_tensor(scale_factor_A): + scale_factor_A = torch.tensor(float(scale_factor_A), device=A.device, dtype=torch.float32) + + if scale_factor_B is None: + scale_factor_B = torch.tensor(1.0, device=B.device, dtype=torch.float32) + elif not torch.is_tensor(scale_factor_B): + scale_factor_B = torch.tensor(float(scale_factor_B), device=B.device, dtype=torch.float32) + + # Quantize A and B to nvfp4; inv scales are FlashInfer’s per-tile inverse scales. + # Returned quantized matrices are FP4-packed formats understood by mm_fp4. + A_q, A_inv_sf = nvfp4_quantize( + A_in, scale_factor_A, sfLayout=sf_layout, do_shuffle=do_shuffle + ) + B_q, B_inv_sf = nvfp4_quantize( + B_in, scale_factor_B, sfLayout=sf_layout, do_shuffle=do_shuffle + ) + + # Alpha rescales the product of the two global scale factors. + alpha = 1.0 / (scale_factor_A * scale_factor_B) + + # Pick FlashInfer flag for the scale-factor layout + use_8x4_sf_layout = (sf_layout == SfLayout.layout_8x4) + + # Allocate output and call CUTLASS-backed FP4 GEMM. + C = torch.empty((M, N), device=A.device, dtype=out_dtype) + mm_fp4( + A_q, B_q.T, # note the transpose on B_q (expecting KxN → NxK in kernel) + A_inv_sf, B_inv_sf.T, # matching transposition of per-tile inverse scales + alpha, # global alpha + out_dtype, C, # output dtype & buffer + block_size=block_size, + use_8x4_sf_layout=use_8x4_sf_layout, + backend="cutlass" + ) + return C + + +def main(): + import argparse, math + torch.manual_seed(0) + + parser = argparse.ArgumentParser() + parser.add_argument("--M", type=int, default=256, help="Rows of A / C (multiple of 128)") + parser.add_argument("--N", type=int, default=256, help="Cols of B / C (multiple of 128)") + parser.add_argument("--K", type=int, default=512, help="Inner dim (multiple of 64)") + parser.add_argument("--dtype", type=str, choices=["fp32", "bf16"], default="bf16") + parser.add_argument("--trials", type=int, default=1) + parser.add_argument("--verbose", action="store_true") + args = parser.parse_args() + + if not torch.cuda.is_available(): + raise RuntimeError("CUDA required") + + M, N, K = args.M, args.N, args.K + assert M % 128 == 0 and N % 128 == 0 and K % 64 == 0, "Shape must satisfy: M%128==0, N%128==0, K%64==0." + + dtype = torch.bfloat16 if args.dtype == "bf16" else torch.float32 + device = "cuda" + + def stats(name, X): + X = X.float() + return f"{name}: max={X.abs().max().item():.6e} mean={X.abs().mean().item():.6e} std={X.std().item():.6e}" + + for t in range(args.trials): + # Inputs + A = torch.randn((M, K), device=device, dtype=dtype) + B = torch.randn((K, N), device=device, dtype=dtype) + + # GEMMs + C_triton = triton_nvfp4(A, B, out_dtype=torch.bfloat16 if dtype is torch.bfloat16 else torch.float32) + C_cutlass = cutlass_nvfp4(A, B, out_dtype=torch.bfloat16 if dtype is torch.bfloat16 else torch.float32) + # Bring to fp32 for error analysis + Ct = C_triton.float() + Cc = C_cutlass.float() + Cref = (A.float() @ B.float()) + # Triton vs Cutlass + diff_tc = (Ct - Cc) + denom_tc = Cc.abs().clamp_min(1e-8) + rel_tc = (diff_tc.abs() / denom_tc) + + # Each vs reference (optional sanity) + diff_t_ref = (Ct - Cref) + diff_c_ref = (Cc - Cref) + rel_t_ref = diff_t_ref.abs() / Cref.abs().clamp_min(1e-8) + rel_c_ref = diff_c_ref.abs() / Cref.abs().clamp_min(1e-8) + + + print("\nTriton vs Cutlass:") + print(f" abs diff -> max={diff_tc.abs().max().item():.6e} mean={diff_tc.abs().mean().item():.6e}") + print(f" rel diff -> max={rel_tc.max().item():.6e} mean={rel_tc.mean().item():.6e}") + + print("\nTriton vs Reference:") + print(f" abs diff -> max={diff_t_ref.abs().max().item():.6e} mean={diff_t_ref.abs().mean().item():.6e}") + print(f" rel diff -> max={rel_t_ref.max().item():.6e} mean={rel_t_ref.mean().item():.6e}") + + print("\nCutlass vs Reference:") + print(f" abs diff -> max={diff_c_ref.abs().max().item():.6e} mean={diff_c_ref.abs().mean().item():.6e}") + print(f" rel diff -> max={rel_c_ref.max().item():.6e} mean={rel_c_ref.mean().item():.6e}") + + if args.verbose: + # a tiny sample print to eyeball values + i, j = 0, 0 + print(f"\nSample C[0,0]: triton={Ct[i,j].item():.6e}, cutlass={Cc[i,j].item():.6e}, ref={Cref[i,j].item():.6e}") + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/triton_fp4.py b/triton_fp4.py new file mode 100644 index 000000000..782838132 --- /dev/null +++ b/triton_fp4.py @@ -0,0 +1,320 @@ +import argparse + +import torch +import triton +import triton.language as tl +import triton.profiler as proton +from triton.tools.tensor_descriptor import TensorDescriptor +from triton.tools.mxfp import MXFP4Tensor, MXScaleTensor +from test_quantization_precision import generate_fake_quant + +VEC_SIZE = 16 +MAX_MXF4 = 6.0 # max magnitude for E2M1 (nvfp4) with M∈{0,1}, E∈{0..3} +EPS = 1e-8 + + + +def is_cuda(): + return triton.runtime.driver.active.get_current_target().backend == "cuda" + + +def supports_block_scaling(): + return is_cuda() and torch.cuda.get_device_capability()[0] == 10 + + +def _matmul_launch_metadata(grid, kernel, args): + ret = {} + M, N, K = args["M"], args["N"], args["K"] + kernel_name = kernel.name + if "ELEM_PER_BYTE_A" and "ELEM_PER_BYTE_B" and "VEC_SIZE" in args: + kernel_name += "_nvfp4" + ret["name"] = f"{kernel_name} [M={M}, N={N}, K={K}]" + ret["flops"] = 2.0 * M * N * K + return ret + + +@triton.jit(launch_metadata=_matmul_launch_metadata) +def block_scaled_matmul_kernel( # + a_desc, # + a_scale_desc, # + b_desc, # + b_scale_desc, # + c_desc, # + M: tl.constexpr, # + N: tl.constexpr, # + K: tl.constexpr, # + output_type: tl.constexpr, # + ELEM_PER_BYTE_A: tl.constexpr, # + ELEM_PER_BYTE_B: tl.constexpr, # + VEC_SIZE: tl.constexpr, # + BLOCK_M: tl.constexpr, # + BLOCK_N: tl.constexpr, # + BLOCK_K: tl.constexpr, # + rep_m: tl.constexpr, # + rep_n: tl.constexpr, # + rep_k: tl.constexpr, # + NUM_STAGES: tl.constexpr, # +): # + if output_type == 0: + output_dtype = tl.float32 + elif output_type == 1: + output_dtype = tl.float16 + elif output_type == 2: + output_dtype = tl.bfloat16 + elif output_type == 3: + output_dtype = tl.float8e4nv + + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_M) + pid_m = pid % num_pid_m + pid_n = pid // num_pid_m + offs_am = pid_m * BLOCK_M + offs_bn = pid_n * BLOCK_N + offs_k_a = 0 + offs_k_b = 0 + offs_scale_m = pid_m * rep_m + offs_scale_n = pid_n * rep_n + offs_scale_k = 0 + + MIXED_PREC: tl.constexpr = ELEM_PER_BYTE_A == 1 and ELEM_PER_BYTE_B == 2 + + accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32) + for k in tl.range(0, tl.cdiv(K, BLOCK_K), num_stages=NUM_STAGES): + a = a_desc.load([offs_am, offs_k_a]) + b = b_desc.load([offs_bn, offs_k_b]) + scale_a = a_scale_desc.load([0, offs_scale_m, offs_scale_k, 0, 0]) + scale_b = b_scale_desc.load([0, offs_scale_n, offs_scale_k, 0, 0]) + + scale_a = scale_a.reshape(rep_m, rep_k, 32, 4, 4).trans(0, 3, 2, 1, 4).reshape(BLOCK_M, BLOCK_K // VEC_SIZE) + scale_b = scale_b.reshape(rep_n, rep_k, 32, 4, 4).trans(0, 3, 2, 1, 4).reshape(BLOCK_N, BLOCK_K // VEC_SIZE) + + if MIXED_PREC: + accumulator = tl.dot_scaled(a, scale_a, "e4m3", b.T, scale_b, "e2m1", accumulator) + elif ELEM_PER_BYTE_A == 2 and ELEM_PER_BYTE_B == 2: + accumulator = tl.dot_scaled(a, scale_a, "e2m1", b.T, scale_b, "e2m1", accumulator) + else: + accumulator = tl.dot_scaled(a, scale_a, "e4m3", b.T, scale_b, "e4m3", accumulator) + + offs_k_a += BLOCK_K // ELEM_PER_BYTE_A + offs_k_b += BLOCK_K // ELEM_PER_BYTE_B + offs_scale_k += rep_k + + c_desc.store([offs_am, offs_bn], accumulator.to(output_dtype)) + + +def block_scaled_matmul(a_desc, a_scale_desc, b_desc, b_scale_desc, dtype_dst, M, N, K, rep_m, rep_n, rep_k, configs): + output = torch.empty((M, N), dtype=dtype_dst, device="cuda") + if dtype_dst == torch.float32: + dtype_dst = 0 + elif dtype_dst == torch.float16: + dtype_dst = 1 + elif dtype_dst == torch.bfloat16: + dtype_dst = 2 + elif dtype_dst == torch.float8_e4m3fn: + dtype_dst = 3 + else: + raise ValueError(f"Unsupported dtype: {dtype_dst}") + + BLOCK_M = configs["BLOCK_SIZE_M"] + BLOCK_N = configs["BLOCK_SIZE_N"] + c_desc = TensorDescriptor.from_tensor(output, [BLOCK_M, BLOCK_N]) + + grid = (triton.cdiv(M, BLOCK_M) * triton.cdiv(N, BLOCK_N), 1) + block_scaled_matmul_kernel[grid]( + a_desc, + a_scale_desc, + b_desc, + b_scale_desc, + c_desc, + M, + N, + K, + dtype_dst, + configs["ELEM_PER_BYTE_A"], + configs["ELEM_PER_BYTE_B"], + configs["VEC_SIZE"], + configs["BLOCK_SIZE_M"], + configs["BLOCK_SIZE_N"], + configs["BLOCK_SIZE_K"], + rep_m, + rep_n, + rep_k, + configs["num_stages"], + ) + return output +def initialize_block_scaled(M, N, K, block_scale_type="nvfp4", compute_reference=False, input_output_dtype=torch.bfloat16): + assert block_scale_type == "nvfp4", f"Only nvfp4 supported here (got {block_scale_type})" + + # --- constants --- + BLOCK_M = 128 + BLOCK_N = 256 + BLOCK_K = 256 + VEC_SIZE = 16 + ELEM_PER_BYTE_A = 2 + ELEM_PER_BYTE_B = 2 + device = "cuda" + MAX_MXF4 = 6.0 + EPS = 1e-8 + + assert M % 128 == 0 + assert N % 128 == 0 + assert K % (VEC_SIZE * 4) == 0 + + # --- helpers: tile scales (FP8 E4M3) --- + def _tile_scales_A(a_fp32): + m_chunks = M // 128 + k_chunks = K // (VEC_SIZE * 4) # 64-wide tiles along K + x = a_fp32.view(m_chunks, 128, k_chunks, 4, VEC_SIZE) + tile_max = x.abs().amax(dim=(1, 3, 4)) # [m_chunks, k_chunks] + scale = torch.clamp(tile_max / MAX_MXF4, min=EPS) # fp32 + s = scale[:, :, None, None].expand(m_chunks, k_chunks, 32, 16) + return s.to(torch.float8_e4m3fn) # [m_chunks, k_chunks, 32, 16] + + def _tile_scales_B(b_fp32): # b_fp32 is (K, N) + k_chunks = K // (VEC_SIZE * 4) + n_chunks = N // 128 + x = b_fp32.view(k_chunks, 4, VEC_SIZE, n_chunks, 128) + tile_max = x.abs().amax(dim=(1, 2, 4)) # [k_chunks, n_chunks] + scale = torch.clamp(tile_max / MAX_MXF4, min=EPS).t() # [n_chunks, k_chunks] + s = scale[:, :, None, None].expand(n_chunks, k_chunks, 32, 16) + return s.to(torch.float8_e4m3fn) # [n_chunks, k_chunks, 32, 16] + + # --- expand tile scales to per-element fp32 (for quant/dequant math) --- + def _expand_A_scale_full(a_scale_fp8): + m_chunks, k_chunks, _, _ = a_scale_fp8.shape + base = a_scale_fp8[..., 0, 0].to(input_output_dtype) # [m_chunks, k_chunks] + tile = base[:, :, None, None].expand(m_chunks, k_chunks, 128, 64) + return tile.permute(0, 2, 1, 3).reshape(M, k_chunks * 64).contiguous() # [M, K] + + def _expand_B_scale_full(b_scale_fp8): + n_chunks, k_chunks, _, _ = b_scale_fp8.shape + base = b_scale_fp8[..., 0, 0].to(input_output_dtype) # [n_chunks, k_chunks] + tile = base[:, :, None, None].expand(n_chunks, k_chunks, 64, 128) + return tile.permute(1, 2, 0, 3).reshape(k_chunks * 64, n_chunks * 128).contiguous() # [K, N] + + # --- 1) fp32 sources (pre-quant) --- + torch.manual_seed(0) + a_f32_pre = torch.randn((M, K), device=device, dtype=input_output_dtype) + b_f32_pre = torch.randn((K, N), device=device, dtype=input_output_dtype) + + # --- 2) tile scales (FP8 E4M3, finite-only) --- + a_scale_fp8 = _tile_scales_A(a_f32_pre) # [M//128, K//64, 32, 16] + b_scale_fp8 = _tile_scales_B(b_f32_pre) # [N//128, K//64, 32, 16] + + # --- 3) expand scales to full for quantization --- + a_scale_full = _expand_A_scale_full(a_scale_fp8) # [M, K] fp32 + b_scale_full = _expand_B_scale_full(b_scale_fp8) # [K, N] fp32 + + # --- 4) quantize to nvfp4 codes (E2M1) --- + a_codes = MXFP4Tensor(data=(a_f32_pre / a_scale_full)).data # [M, K] uint8 (nibble used) + b_codes_KN = MXFP4Tensor(data=(b_f32_pre / b_scale_full)).data # [K, N] uint8 + + # IMPORTANT: the kernel expects B laid out as (N, K_packed) (like original) + # so create a view for descriptors with shape (N, K) and pack along dim=1. + b_codes_NK = b_codes_KN.transpose(0, 1).contiguous() # [N, K] uint8 + + # --- 5) pack to 2-per-byte along K and build descriptors --- + a_fp4 = MXFP4Tensor(size=(M, K), device=device); a_fp4.data = a_codes + b_fp4_for_desc = MXFP4Tensor(size=(N, K), device=device); b_fp4_for_desc.data = b_codes_NK + + a_packed = a_fp4.to_packed_tensor(dim=1) # (M, K/2) + b_packed = b_fp4_for_desc.to_packed_tensor(dim=1) # (N, K/2) <-- matches kernel indexing + + a_desc = TensorDescriptor.from_tensor(a_packed, [BLOCK_M, BLOCK_K // ELEM_PER_BYTE_A]) + b_desc = TensorDescriptor.from_tensor(b_packed, [BLOCK_N, BLOCK_K // ELEM_PER_BYTE_B]) + + rep_m = BLOCK_M // 128 + rep_n = BLOCK_N // 128 + rep_k = BLOCK_K // VEC_SIZE // 4 + + # reshape FP8 scales to 5D TMA blocks [1, rep_*, rep_k, 2, 256] + a_scale_5d = a_scale_fp8.reshape(1, a_scale_fp8.shape[0], a_scale_fp8.shape[1], 2, 256) + b_scale_5d = b_scale_fp8.reshape(1, b_scale_fp8.shape[0], b_scale_fp8.shape[1], 2, 256) + a_scale_desc = TensorDescriptor.from_tensor(a_scale_5d, [1, rep_m, rep_k, 2, 256]) + b_scale_desc = TensorDescriptor.from_tensor(b_scale_5d, [1, rep_n, rep_k, 2, 256]) + + # --- 6) dequantize back to fp32 for the reference (post-quant tensors) --- + a_ref_fp4 = MXFP4Tensor(size=(M, K), device=device); a_ref_fp4.data = a_codes + b_ref_fp4 = MXFP4Tensor(size=(K, N), device=device); b_ref_fp4.data = b_codes_KN + + a_ref_f32, b_ref_f32 = generate_fake_quant(a_f32_pre, b_f32_pre) + + reference = a_ref_f32 @ b_ref_f32 if compute_reference else None + + configs = { + "BLOCK_SIZE_M": BLOCK_M, + "BLOCK_SIZE_N": BLOCK_N, + "BLOCK_SIZE_K": BLOCK_K, + "num_stages": 4, + "ELEM_PER_BYTE_A": ELEM_PER_BYTE_A, + "ELEM_PER_BYTE_B": ELEM_PER_BYTE_B, + "VEC_SIZE": VEC_SIZE, + } + # RETURN DESCRIPTORS (not raw tensors!) + return a_desc, a_scale_desc, b_desc, b_scale_desc, rep_m, rep_n, rep_k, configs, reference + + +def validate_block_scaled(M, N, K, block_scale_type="nvfp4", input_output_dtype=torch.bfloat16): + a_desc, a_scale, b_desc, b_scale, rep_m, rep_n, rep_k, configs, reference = initialize_block_scaled( + M, N, K, block_scale_type, compute_reference=True) + output = block_scaled_matmul(a_desc, a_scale, b_desc, b_scale, input_output_dtype, M, N, K, rep_m, rep_n, rep_k, configs) + torch.testing.assert_close(reference, output.to(input_output_dtype), atol=1e-3, rtol=1e-3) + print(f"✅ (pass {block_scale_type})") + + +def bench_block_scaled(K, block_scale_type="nvfp4", reps=10): + assert K % 128 == 0 + M = 8192 + N = 8192 + print(f"Problem Shape = {M}x{N}x{K}") + + a_desc, a_scale, b_desc, b_scale, rep_m, rep_n, rep_k, configs, _ = initialize_block_scaled( + M, N, K, block_scale_type, compute_reference=False) + _ = block_scaled_matmul(a_desc, a_scale, b_desc, b_scale, torch.float16, M, N, K, rep_m, rep_n, rep_k, configs) + + proton.activate(0) + for _ in range(reps): + _ = block_scaled_matmul(a_desc, a_scale, b_desc, b_scale, torch.float16, M, N, K, rep_m, rep_n, rep_k, configs) + proton.deactivate(0) + print("Done benchmarking") + + +def show_profile(profile_name): + import triton.profiler.viewer as proton_viewer + + metric_names = ["time/ms"] + metric_names = ["tflop/s"] + metric_names + file_name = f"{profile_name}.hatchet" + tree, metrics = proton_viewer.parse(metric_names, file_name) + proton_viewer.print_tree(tree, metrics) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("-K", type=int, required=False, default=512) + parser.add_argument("--K_range", type=int, nargs=2) + parser.add_argument("--K_step", type=int, default=512) + parser.add_argument("--bench", action="store_true", default=True) + parser.add_argument("--format", type=str, choices=["nvfp4"], default="nvfp4") + parser.add_argument("--input_output_dtype", type=str, choices=["float32", "float16", "bfloat16"], default="bfloat16") + args = parser.parse_args() + + if not supports_block_scaling(): + print("⛔ This example requires GPU support for block scaled matmul") + else: + if args.K and args.K_range is None: + args.K_range = [args.K, args.K] + args.K_step = 1 # doesn't matter as long as it's not 0 + + torch.manual_seed(42) + dtype =torch.float32 if args.input_output_dtype == "float32" else torch.float16 if args.input_output_dtype == "float16" else torch.bfloat16 + validate_block_scaled(8192, 8192, 8192, block_scale_type=args.format, input_output_dtype=dtype) + + + if args.bench: + proton.start("block_scaled_matmul", hook="triton") + proton.deactivate(0) # Skip argument creation + for K in range(args.K_range[0], args.K_range[1] + 1, args.K_step): + bench_block_scaled(K, reps=10000, block_scale_type=args.format) + proton.finalize() + show_profile("block_scaled_matmul") \ No newline at end of file diff --git a/triton_original.py b/triton_original.py new file mode 100644 index 000000000..b52103e6e --- /dev/null +++ b/triton_original.py @@ -0,0 +1,302 @@ +import argparse + +import torch +import triton +import triton.language as tl +import triton.profiler as proton +from triton.tools.tensor_descriptor import TensorDescriptor +from triton.tools.mxfp import MXFP4Tensor, MXScaleTensor + + +def is_cuda(): + return triton.runtime.driver.active.get_current_target().backend == "cuda" + + +def supports_block_scaling(): + return is_cuda() and torch.cuda.get_device_capability()[0] == 10 + + +def _matmul_launch_metadata(grid, kernel, args): + ret = {} + M, N, K = args["M"], args["N"], args["K"] + kernel_name = kernel.name + if "ELEM_PER_BYTE_A" and "ELEM_PER_BYTE_B" and "VEC_SIZE" in args: + if args["ELEM_PER_BYTE_A"] == 1 and args["ELEM_PER_BYTE_B"] == 1: + kernel_name += "_mxfp8" + elif args["ELEM_PER_BYTE_A"] == 1 and args["ELEM_PER_BYTE_B"] == 2: + kernel_name += "_mixed" + elif args["ELEM_PER_BYTE_A"] == 2 and args["ELEM_PER_BYTE_B"] == 2: + if args["VEC_SIZE"] == 16: + kernel_name += "_nvfp4" + elif args["VEC_SIZE"] == 32: + kernel_name += "_mxfp4" + ret["name"] = f"{kernel_name} [M={M}, N={N}, K={K}]" + ret["flops"] = 2.0 * M * N * K + return ret + + +@triton.jit(launch_metadata=_matmul_launch_metadata) +def block_scaled_matmul_kernel( # + a_desc, # + a_scale_desc, # + b_desc, # + b_scale_desc, # + c_desc, # + M: tl.constexpr, # + N: tl.constexpr, # + K: tl.constexpr, # + output_type: tl.constexpr, # + ELEM_PER_BYTE_A: tl.constexpr, # + ELEM_PER_BYTE_B: tl.constexpr, # + VEC_SIZE: tl.constexpr, # + BLOCK_M: tl.constexpr, # + BLOCK_N: tl.constexpr, # + BLOCK_K: tl.constexpr, # + rep_m: tl.constexpr, # + rep_n: tl.constexpr, # + rep_k: tl.constexpr, # + NUM_STAGES: tl.constexpr, # +): # + if output_type == 0: + output_dtype = tl.float32 + elif output_type == 1: + output_dtype = tl.float16 + elif output_type == 2: + output_dtype = tl.float8e4nv + + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_M) + pid_m = pid % num_pid_m + pid_n = pid // num_pid_m + offs_am = pid_m * BLOCK_M + offs_bn = pid_n * BLOCK_N + offs_k_a = 0 + offs_k_b = 0 + offs_scale_m = pid_m * rep_m + offs_scale_n = pid_n * rep_n + offs_scale_k = 0 + + MIXED_PREC: tl.constexpr = ELEM_PER_BYTE_A == 1 and ELEM_PER_BYTE_B == 2 + + accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32) + for k in tl.range(0, tl.cdiv(K, BLOCK_K), num_stages=NUM_STAGES): + a = a_desc.load([offs_am, offs_k_a]) + b = b_desc.load([offs_bn, offs_k_b]) + scale_a = a_scale_desc.load([0, offs_scale_m, offs_scale_k, 0, 0]) + scale_b = b_scale_desc.load([0, offs_scale_n, offs_scale_k, 0, 0]) + + scale_a = scale_a.reshape(rep_m, rep_k, 32, 4, 4).trans(0, 3, 2, 1, 4).reshape(BLOCK_M, BLOCK_K // VEC_SIZE) + scale_b = scale_b.reshape(rep_n, rep_k, 32, 4, 4).trans(0, 3, 2, 1, 4).reshape(BLOCK_N, BLOCK_K // VEC_SIZE) + + if MIXED_PREC: + accumulator = tl.dot_scaled(a, scale_a, "e4m3", b.T, scale_b, "e2m1", accumulator) + elif ELEM_PER_BYTE_A == 2 and ELEM_PER_BYTE_B == 2: + accumulator = tl.dot_scaled(a, scale_a, "e2m1", b.T, scale_b, "e2m1", accumulator) + else: + accumulator = tl.dot_scaled(a, scale_a, "e4m3", b.T, scale_b, "e4m3", accumulator) + + offs_k_a += BLOCK_K // ELEM_PER_BYTE_A + offs_k_b += BLOCK_K // ELEM_PER_BYTE_B + offs_scale_k += rep_k + + c_desc.store([offs_am, offs_bn], accumulator.to(output_dtype)) + + +def block_scaled_matmul(a_desc, a_scale_desc, b_desc, b_scale_desc, dtype_dst, M, N, K, rep_m, rep_n, rep_k, configs): + output = torch.empty((M, N), dtype=dtype_dst, device="cuda") + if dtype_dst == torch.float32: + dtype_dst = 0 + elif dtype_dst == torch.float16: + dtype_dst = 1 + elif dtype_dst == torch.float8_e4m3fn: + dtype_dst = 2 + else: + raise ValueError(f"Unsupported dtype: {dtype_dst}") + + BLOCK_M = configs["BLOCK_SIZE_M"] + BLOCK_N = configs["BLOCK_SIZE_N"] + c_desc = TensorDescriptor.from_tensor(output, [BLOCK_M, BLOCK_N]) + + grid = (triton.cdiv(M, BLOCK_M) * triton.cdiv(N, BLOCK_N), 1) + block_scaled_matmul_kernel[grid]( + a_desc, + a_scale_desc, + b_desc, + b_scale_desc, + c_desc, + M, + N, + K, + dtype_dst, + configs["ELEM_PER_BYTE_A"], + configs["ELEM_PER_BYTE_B"], + configs["VEC_SIZE"], + configs["BLOCK_SIZE_M"], + configs["BLOCK_SIZE_N"], + configs["BLOCK_SIZE_K"], + rep_m, + rep_n, + rep_k, + configs["num_stages"], + ) + return output + + +def initialize_block_scaled(M, N, K, block_scale_type="nvfp4", compute_reference=False): + BLOCK_M = 128 + BLOCK_N = 256 + BLOCK_K = 256 if "fp4" in block_scale_type else 128 + VEC_SIZE = 16 if block_scale_type == "nvfp4" else 32 + assert block_scale_type in ["nvfp4", "mxfp4", "mxfp8", "mixed"], f"Invalid block scale type: {block_scale_type}" + ELEM_PER_BYTE_A = 2 if "fp4" in block_scale_type else 1 + ELEM_PER_BYTE_B = 1 if block_scale_type == "mxfp8" else 2 + + device = "cuda" + a_ref = MXFP4Tensor(size=(M, K), device=device).random() + # Similar to Hopper's wgmma symmetric fp8 instruction, the RHS is expected + # to be in col-major layout for Blackwell's tcgen05.mma when using fp4 operands. + # To conform to the expected semantics of tl.dot_scaled, (M, K) x (K, N), + # the data is generated in col-major layout, packed along K for fp4, and then + # logically transposed. Note that if one operand is of fp8 precision, unlike Hopper, + # Blackwell supports both row-major and col-major layouts for the RHS matrix. + # For the mixed-precision case, the fp4 RHS can be either in row or col-major layout. + # But for performance reason, it is recommended to use col-major layout. If TMA is used + # for the fp4 RHS operand load in mixed-precision dot, as in this tutorial, it must be + # in col-major layout. + b_ref = MXFP4Tensor(size=(N, K), device=device).random() + if block_scale_type in ["mxfp8", "mixed"]: + a_ref = a_ref.to(torch.float32) + a = a_ref.to(torch.float8_e4m3fn) + else: + # Pack two fp4 elements per byte along K + a = a_ref.to_packed_tensor(dim=1) + + if block_scale_type == "mxfp8": + b_ref = b_ref.to(torch.float32) + b = b_ref.to(torch.float8_e4m3fn) + else: + b = b_ref.to_packed_tensor(dim=1) + + b_ref = b_ref.to(torch.float32).T + + a_desc = TensorDescriptor.from_tensor(a, [BLOCK_M, BLOCK_K // ELEM_PER_BYTE_A]) + b_desc = TensorDescriptor.from_tensor(b, [BLOCK_N, BLOCK_K // ELEM_PER_BYTE_B]) + + a_scale_shape = [M // 128, K // VEC_SIZE // 4, 32, 16] + b_scale_shape = [N // 128, K // VEC_SIZE // 4, 32, 16] + epsilon = 1e-8 + a_scale = torch.rand(a_scale_shape, device=device) + epsilon + b_scale = torch.rand(b_scale_shape, device=device) + epsilon + if block_scale_type == "nvfp4": + a_scale = a_scale.to(torch.float8_e4m3fn) + b_scale = b_scale.to(torch.float8_e4m3fn) + a_scale_ref = a_scale + b_scale_ref = b_scale + elif block_scale_type in ["mxfp4", "mxfp8", "mixed"]: + a_scale_ref = MXScaleTensor(a_scale) + b_scale_ref = MXScaleTensor(b_scale) + a_scale = a_scale_ref.data + b_scale = b_scale_ref.data + + rep_m = BLOCK_M // 128 + rep_n = BLOCK_N // 128 + rep_k = BLOCK_K // VEC_SIZE // 4 + + # Use 5D TMA descriptor [1, rep_m, rep_k, 2, 256] with uint8 elements. + # With 256 elements we better utilize the L2 and don't require the TMA + # engine to emit many small messages (16B) messages as with 32x16xu8. + a_scale_block_shape = [1, rep_m, rep_k, 2, 256] + b_scale_block_shape = [1, rep_n, rep_k, 2, 256] + a_scale = a_scale.reshape(1, a_scale_shape[0], a_scale.shape[1], 2, 256) + b_scale = b_scale.reshape(1, b_scale_shape[0], b_scale.shape[1], 2, 256) + a_scale_desc = TensorDescriptor.from_tensor(a_scale, block_shape=a_scale_block_shape) + b_scale_desc = TensorDescriptor.from_tensor(b_scale, block_shape=b_scale_block_shape) + + reference = None + if compute_reference: + a_scale_ref = a_scale_ref.to(torch.float32) + b_scale_ref = b_scale_ref.to(torch.float32) + + def unpack_scale(packed): + packed = packed.reshape(*packed.shape[:-2], 32, 4, 4) + num_chunk_m, num_chunk_k, _, _, _ = packed.shape + return packed.permute(0, 3, 2, 1, 4).reshape(num_chunk_m * 128, num_chunk_k * 4).contiguous() + + a_scale_ref = unpack_scale(a_scale_ref).repeat_interleave(VEC_SIZE, dim=1)[:M, :K] + b_scale_ref = unpack_scale(b_scale_ref).repeat_interleave(VEC_SIZE, dim=1).T.contiguous()[:K, :N] + reference = torch.matmul(a_ref.to(torch.float32) * a_scale_ref, b_ref * b_scale_ref) + + configs = { + "BLOCK_SIZE_M": BLOCK_M, + "BLOCK_SIZE_N": BLOCK_N, + "BLOCK_SIZE_K": BLOCK_K, + "num_stages": 4, + "ELEM_PER_BYTE_A": ELEM_PER_BYTE_A, + "ELEM_PER_BYTE_B": ELEM_PER_BYTE_B, + "VEC_SIZE": VEC_SIZE, + } + return a_desc, a_scale_desc, b_desc, b_scale_desc, rep_m, rep_n, rep_k, configs, reference + + +def validate_block_scaled(M, N, K, block_scale_type="nvfp4"): + a_desc, a_scale, b_desc, b_scale, rep_m, rep_n, rep_k, configs, reference = initialize_block_scaled( + M, N, K, block_scale_type, compute_reference=True) + output = block_scaled_matmul(a_desc, a_scale, b_desc, b_scale, torch.float16, M, N, K, rep_m, rep_n, rep_k, configs) + torch.testing.assert_close(reference, output.to(torch.float32), atol=1e-3, rtol=1e-3) + print(f"✅ (pass {block_scale_type})") + + +def bench_block_scaled(K, block_scale_type="nvfp4", reps=10): + assert K % 128 == 0 + M = 8192 + N = 8192 + print(f"Problem Shape = {M}x{N}x{K}") + + a_desc, a_scale, b_desc, b_scale, rep_m, rep_n, rep_k, configs, _ = initialize_block_scaled( + M, N, K, block_scale_type, compute_reference=False) + _ = block_scaled_matmul(a_desc, a_scale, b_desc, b_scale, torch.float16, M, N, K, rep_m, rep_n, rep_k, configs) + + proton.activate(0) + for _ in range(reps): + _ = block_scaled_matmul(a_desc, a_scale, b_desc, b_scale, torch.float16, M, N, K, rep_m, rep_n, rep_k, configs) + proton.deactivate(0) + print("Done benchmarking") + + +def show_profile(profile_name): + import triton.profiler.viewer as proton_viewer + + metric_names = ["time/ms"] + metric_names = ["tflop/s"] + metric_names + file_name = f"{profile_name}.hatchet" + tree, metrics = proton_viewer.parse(metric_names, file_name) + proton_viewer.print_tree(tree, metrics) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("-K", type=int, required=False, default=8192) + parser.add_argument("--K_range", type=int, nargs=2) + parser.add_argument("--K_step", type=int, default=512) + parser.add_argument("--bench", action="store_true", default=True) + parser.add_argument("--format", type=str, choices=["mxfp4", "nvfp4", "mxfp8", "mixed"], default="nvfp4") + args = parser.parse_args() + + if not supports_block_scaling(): + print("⛔ This example requires GPU support for block scaled matmul") + else: + if args.K and args.K_range is None: + args.K_range = [args.K, args.K] + args.K_step = 1 # doesn't matter as long as it's not 0 + + torch.manual_seed(42) + + validate_block_scaled(8192, 8192, 8192, block_scale_type=args.format) + + if args.bench: + proton.start("block_scaled_matmul", hook="triton") + proton.deactivate(0) # Skip argument creation + for K in range(args.K_range[0], args.K_range[1] + 1, args.K_step): + bench_block_scaled(K, reps=10000, block_scale_type=args.format) + proton.finalize() + show_profile("block_scaled_matmul") \ No newline at end of file diff --git a/uv.lock b/uv.lock new file mode 100644 index 000000000..45c905407 --- /dev/null +++ b/uv.lock @@ -0,0 +1,4405 @@ +version = 1 +revision = 3 +requires-python = ">=3.10" +resolution-markers = [ + "python_full_version >= '3.13' and sys_platform == 'darwin'", + "python_full_version >= '3.13' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version >= '3.13' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version >= '3.13' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.12.*' and sys_platform == 'darwin'", + "python_full_version == '3.12.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.11.*' and sys_platform == 'darwin'", + "python_full_version == '3.11.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version < '3.11' and sys_platform == 'darwin'", + "python_full_version < '3.11' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux'", +] + +[options] +prerelease-mode = "allow" + +[[package]] +name = "absl-py" +version = "2.3.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/10/2a/c93173ffa1b39c1d0395b7e842bbdc62e556ca9d8d3b5572926f3e4ca752/absl_py-2.3.1.tar.gz", hash = "sha256:a97820526f7fbfd2ec1bce83f3f25e3a14840dac0d8e02a0b71cd75db3f77fc9", size = 116588, upload-time = "2025-07-03T09:31:44.05Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/8f/aa/ba0014cc4659328dc818a28827be78e6d97312ab0cb98105a770924dc11e/absl_py-2.3.1-py3-none-any.whl", hash = "sha256:eeecf07f0c2a93ace0772c92e596ace6d3d3996c042b2128459aaae2a76de11d", size = 135811, upload-time = "2025-07-03T09:31:42.253Z" }, +] + +[[package]] +name = "accelerate" +version = "1.0.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "huggingface-hub" }, + { name = "numpy" }, + { name = "packaging" }, + { name = "psutil" }, + { name = "pyyaml" }, + { name = "safetensors" }, + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/a0/04/7f1b79a02f66a7e87e1716849b3bd295445cc46ecb63cb7521e6f7a72282/accelerate-1.0.1.tar.gz", hash = "sha256:e8f95fc2db14915dc0a9182edfcf3068e5ddb2fa310b583717ad44e5c442399c", size = 333238, upload-time = "2024-10-12T02:59:21.245Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/2c/92/48aec3736ca778ffe5fa68e19e3c18917cba4de43fa46fe6176cccafe267/accelerate-1.0.1-py3-none-any.whl", hash = "sha256:c6aa0c7b8a797cb150471e90e3ca36ac41f5d4b40512cdd6f058b8bf25589467", size = 330926, upload-time = "2024-10-12T02:59:19.195Z" }, +] + +[[package]] +name = "aiofiles" +version = "24.1.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/0b/03/a88171e277e8caa88a4c77808c20ebb04ba74cc4681bf1e9416c862de237/aiofiles-24.1.0.tar.gz", hash = "sha256:22a075c9e5a3810f0c2e48f3008c94d68c65d763b9b03857924c99e57355166c", size = 30247, upload-time = "2024-06-24T11:02:03.584Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/a5/45/30bb92d442636f570cb5651bc661f52b610e2eec3f891a5dc3a4c3667db0/aiofiles-24.1.0-py3-none-any.whl", hash = "sha256:b4ec55f4195e3eb5d7abd1bf7e061763e864dd4954231fb8539a0ef8bb8260e5", size = 15896, upload-time = "2024-06-24T11:02:01.529Z" }, +] + +[[package]] +name = "aiohappyeyeballs" +version = "2.6.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/26/30/f84a107a9c4331c14b2b586036f40965c128aa4fee4dda5d3d51cb14ad54/aiohappyeyeballs-2.6.1.tar.gz", hash = "sha256:c3f9d0113123803ccadfdf3f0faa505bc78e6a72d1cc4806cbd719826e943558", size = 22760, upload-time = "2025-03-12T01:42:48.764Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/0f/15/5bf3b99495fb160b63f95972b81750f18f7f4e02ad051373b669d17d44f2/aiohappyeyeballs-2.6.1-py3-none-any.whl", hash = "sha256:f349ba8f4b75cb25c99c5c2d84e997e485204d2902a9597802b0371f09331fb8", size = 15265, upload-time = "2025-03-12T01:42:47.083Z" }, +] + +[[package]] +name = "aiohttp" +version = "3.13.2" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "aiohappyeyeballs" }, + { name = "aiosignal" }, + { name = "async-timeout", marker = "python_full_version < '3.11'" }, + { name = "attrs" }, + { name = "frozenlist" }, + { name = "multidict" }, + { name = "propcache" }, + { name = "yarl" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/1c/ce/3b83ebba6b3207a7135e5fcaba49706f8a4b6008153b4e30540c982fae26/aiohttp-3.13.2.tar.gz", hash = "sha256:40176a52c186aefef6eb3cad2cdd30cd06e3afbe88fe8ab2af9c0b90f228daca", size = 7837994, upload-time = "2025-10-28T20:59:39.937Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/6d/34/939730e66b716b76046dedfe0842995842fa906ccc4964bba414ff69e429/aiohttp-3.13.2-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:2372b15a5f62ed37789a6b383ff7344fc5b9f243999b0cd9b629d8bc5f5b4155", size = 736471, upload-time = "2025-10-28T20:55:27.924Z" }, + { url = "https://files.pythonhosted.org/packages/fd/cf/dcbdf2df7f6ca72b0bb4c0b4509701f2d8942cf54e29ca197389c214c07f/aiohttp-3.13.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:e7f8659a48995edee7229522984bd1009c1213929c769c2daa80b40fe49a180c", size = 493985, upload-time = "2025-10-28T20:55:29.456Z" }, + { url = "https://files.pythonhosted.org/packages/9d/87/71c8867e0a1d0882dcbc94af767784c3cb381c1c4db0943ab4aae4fed65e/aiohttp-3.13.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:939ced4a7add92296b0ad38892ce62b98c619288a081170695c6babe4f50e636", size = 489274, upload-time = "2025-10-28T20:55:31.134Z" }, + { url = "https://files.pythonhosted.org/packages/38/0f/46c24e8dae237295eaadd113edd56dee96ef6462adf19b88592d44891dc5/aiohttp-3.13.2-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6315fb6977f1d0dd41a107c527fee2ed5ab0550b7d885bc15fee20ccb17891da", size = 1668171, upload-time = "2025-10-28T20:55:36.065Z" }, + { url = "https://files.pythonhosted.org/packages/eb/c6/4cdfb4440d0e28483681a48f69841fa5e39366347d66ef808cbdadddb20e/aiohttp-3.13.2-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:6e7352512f763f760baaed2637055c49134fd1d35b37c2dedfac35bfe5cf8725", size = 1636036, upload-time = "2025-10-28T20:55:37.576Z" }, + { url = "https://files.pythonhosted.org/packages/84/37/8708cf678628216fb678ab327a4e1711c576d6673998f4f43e86e9ae90dd/aiohttp-3.13.2-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e09a0a06348a2dd73e7213353c90d709502d9786219f69b731f6caa0efeb46f5", size = 1727975, upload-time = "2025-10-28T20:55:39.457Z" }, + { url = "https://files.pythonhosted.org/packages/e6/2e/3ebfe12fdcb9b5f66e8a0a42dffcd7636844c8a018f261efb2419f68220b/aiohttp-3.13.2-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a09a6d073fb5789456545bdee2474d14395792faa0527887f2f4ec1a486a59d3", size = 1815823, upload-time = "2025-10-28T20:55:40.958Z" }, + { url = "https://files.pythonhosted.org/packages/a1/4f/ca2ef819488cbb41844c6cf92ca6dd15b9441e6207c58e5ae0e0fc8d70ad/aiohttp-3.13.2-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b59d13c443f8e049d9e94099c7e412e34610f1f49be0f230ec656a10692a5802", size = 1669374, upload-time = "2025-10-28T20:55:42.745Z" }, + { url = "https://files.pythonhosted.org/packages/f8/fe/1fe2e1179a0d91ce09c99069684aab619bf2ccde9b20bd6ca44f8837203e/aiohttp-3.13.2-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:20db2d67985d71ca033443a1ba2001c4b5693fe09b0e29f6d9358a99d4d62a8a", size = 1555315, upload-time = "2025-10-28T20:55:44.264Z" }, + { url = "https://files.pythonhosted.org/packages/5a/2b/f3781899b81c45d7cbc7140cddb8a3481c195e7cbff8e36374759d2ab5a5/aiohttp-3.13.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:960c2fc686ba27b535f9fd2b52d87ecd7e4fd1cf877f6a5cba8afb5b4a8bd204", size = 1639140, upload-time = "2025-10-28T20:55:46.626Z" }, + { url = "https://files.pythonhosted.org/packages/72/27/c37e85cd3ece6f6c772e549bd5a253d0c122557b25855fb274224811e4f2/aiohttp-3.13.2-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:6c00dbcf5f0d88796151e264a8eab23de2997c9303dd7c0bf622e23b24d3ce22", size = 1645496, upload-time = "2025-10-28T20:55:48.933Z" }, + { url = "https://files.pythonhosted.org/packages/66/20/3af1ab663151bd3780b123e907761cdb86ec2c4e44b2d9b195ebc91fbe37/aiohttp-3.13.2-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:fed38a5edb7945f4d1bcabe2fcd05db4f6ec7e0e82560088b754f7e08d93772d", size = 1697625, upload-time = "2025-10-28T20:55:50.377Z" }, + { url = "https://files.pythonhosted.org/packages/95/eb/ae5cab15efa365e13d56b31b0d085a62600298bf398a7986f8388f73b598/aiohttp-3.13.2-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:b395bbca716c38bef3c764f187860e88c724b342c26275bc03e906142fc5964f", size = 1542025, upload-time = "2025-10-28T20:55:51.861Z" }, + { url = "https://files.pythonhosted.org/packages/e9/2d/1683e8d67ec72d911397fe4e575688d2a9b8f6a6e03c8fdc9f3fd3d4c03f/aiohttp-3.13.2-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:204ffff2426c25dfda401ba08da85f9c59525cdc42bda26660463dd1cbcfec6f", size = 1714918, upload-time = "2025-10-28T20:55:53.515Z" }, + { url = "https://files.pythonhosted.org/packages/99/a2/ffe8e0e1c57c5e542d47ffa1fcf95ef2b3ea573bf7c4d2ee877252431efc/aiohttp-3.13.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:05c4dd3c48fb5f15db31f57eb35374cb0c09afdde532e7fb70a75aede0ed30f6", size = 1656113, upload-time = "2025-10-28T20:55:55.438Z" }, + { url = "https://files.pythonhosted.org/packages/0d/42/d511aff5c3a2b06c09d7d214f508a4ad8ac7799817f7c3d23e7336b5e896/aiohttp-3.13.2-cp310-cp310-win32.whl", hash = "sha256:e574a7d61cf10351d734bcddabbe15ede0eaa8a02070d85446875dc11189a251", size = 432290, upload-time = "2025-10-28T20:55:56.96Z" }, + { url = "https://files.pythonhosted.org/packages/8b/ea/1c2eb7098b5bad4532994f2b7a8228d27674035c9b3234fe02c37469ef14/aiohttp-3.13.2-cp310-cp310-win_amd64.whl", hash = "sha256:364f55663085d658b8462a1c3f17b2b84a5c2e1ba858e1b79bff7b2e24ad1514", size = 455075, upload-time = "2025-10-28T20:55:58.373Z" }, + { url = "https://files.pythonhosted.org/packages/35/74/b321e7d7ca762638cdf8cdeceb39755d9c745aff7a64c8789be96ddf6e96/aiohttp-3.13.2-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:4647d02df098f6434bafd7f32ad14942f05a9caa06c7016fdcc816f343997dd0", size = 743409, upload-time = "2025-10-28T20:56:00.354Z" }, + { url = "https://files.pythonhosted.org/packages/99/3d/91524b905ec473beaf35158d17f82ef5a38033e5809fe8742e3657cdbb97/aiohttp-3.13.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:e3403f24bcb9c3b29113611c3c16a2a447c3953ecf86b79775e7be06f7ae7ccb", size = 497006, upload-time = "2025-10-28T20:56:01.85Z" }, + { url = "https://files.pythonhosted.org/packages/eb/d3/7f68bc02a67716fe80f063e19adbd80a642e30682ce74071269e17d2dba1/aiohttp-3.13.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:43dff14e35aba17e3d6d5ba628858fb8cb51e30f44724a2d2f0c75be492c55e9", size = 493195, upload-time = "2025-10-28T20:56:03.314Z" }, + { url = "https://files.pythonhosted.org/packages/98/31/913f774a4708775433b7375c4f867d58ba58ead833af96c8af3621a0d243/aiohttp-3.13.2-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e2a9ea08e8c58bb17655630198833109227dea914cd20be660f52215f6de5613", size = 1747759, upload-time = "2025-10-28T20:56:04.904Z" }, + { url = "https://files.pythonhosted.org/packages/e8/63/04efe156f4326f31c7c4a97144f82132c3bb21859b7bb84748d452ccc17c/aiohttp-3.13.2-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:53b07472f235eb80e826ad038c9d106c2f653584753f3ddab907c83f49eedead", size = 1704456, upload-time = "2025-10-28T20:56:06.986Z" }, + { url = "https://files.pythonhosted.org/packages/8e/02/4e16154d8e0a9cf4ae76f692941fd52543bbb148f02f098ca73cab9b1c1b/aiohttp-3.13.2-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e736c93e9c274fce6419af4aac199984d866e55f8a4cec9114671d0ea9688780", size = 1807572, upload-time = "2025-10-28T20:56:08.558Z" }, + { url = "https://files.pythonhosted.org/packages/34/58/b0583defb38689e7f06798f0285b1ffb3a6fb371f38363ce5fd772112724/aiohttp-3.13.2-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ff5e771f5dcbc81c64898c597a434f7682f2259e0cd666932a913d53d1341d1a", size = 1895954, upload-time = "2025-10-28T20:56:10.545Z" }, + { url = "https://files.pythonhosted.org/packages/6b/f3/083907ee3437425b4e376aa58b2c915eb1a33703ec0dc30040f7ae3368c6/aiohttp-3.13.2-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a3b6fb0c207cc661fa0bf8c66d8d9b657331ccc814f4719468af61034b478592", size = 1747092, upload-time = "2025-10-28T20:56:12.118Z" }, + { url = "https://files.pythonhosted.org/packages/ac/61/98a47319b4e425cc134e05e5f3fc512bf9a04bf65aafd9fdcda5d57ec693/aiohttp-3.13.2-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:97a0895a8e840ab3520e2288db7cace3a1981300d48babeb50e7425609e2e0ab", size = 1606815, upload-time = "2025-10-28T20:56:14.191Z" }, + { url = "https://files.pythonhosted.org/packages/97/4b/e78b854d82f66bb974189135d31fce265dee0f5344f64dd0d345158a5973/aiohttp-3.13.2-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:9e8f8afb552297aca127c90cb840e9a1d4bfd6a10d7d8f2d9176e1acc69bad30", size = 1723789, upload-time = "2025-10-28T20:56:16.101Z" }, + { url = "https://files.pythonhosted.org/packages/ed/fc/9d2ccc794fc9b9acd1379d625c3a8c64a45508b5091c546dea273a41929e/aiohttp-3.13.2-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:ed2f9c7216e53c3df02264f25d824b079cc5914f9e2deba94155190ef648ee40", size = 1718104, upload-time = "2025-10-28T20:56:17.655Z" }, + { url = "https://files.pythonhosted.org/packages/66/65/34564b8765ea5c7d79d23c9113135d1dd3609173da13084830f1507d56cf/aiohttp-3.13.2-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:99c5280a329d5fa18ef30fd10c793a190d996567667908bef8a7f81f8202b948", size = 1785584, upload-time = "2025-10-28T20:56:19.238Z" }, + { url = "https://files.pythonhosted.org/packages/30/be/f6a7a426e02fc82781afd62016417b3948e2207426d90a0e478790d1c8a4/aiohttp-3.13.2-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:2ca6ffef405fc9c09a746cb5d019c1672cd7f402542e379afc66b370833170cf", size = 1595126, upload-time = "2025-10-28T20:56:20.836Z" }, + { url = "https://files.pythonhosted.org/packages/e5/c7/8e22d5d28f94f67d2af496f14a83b3c155d915d1fe53d94b66d425ec5b42/aiohttp-3.13.2-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:47f438b1a28e926c37632bff3c44df7d27c9b57aaf4e34b1def3c07111fdb782", size = 1800665, upload-time = "2025-10-28T20:56:22.922Z" }, + { url = "https://files.pythonhosted.org/packages/d1/11/91133c8b68b1da9fc16555706aa7276fdf781ae2bb0876c838dd86b8116e/aiohttp-3.13.2-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:9acda8604a57bb60544e4646a4615c1866ee6c04a8edef9b8ee6fd1d8fa2ddc8", size = 1739532, upload-time = "2025-10-28T20:56:25.924Z" }, + { url = "https://files.pythonhosted.org/packages/17/6b/3747644d26a998774b21a616016620293ddefa4d63af6286f389aedac844/aiohttp-3.13.2-cp311-cp311-win32.whl", hash = "sha256:868e195e39b24aaa930b063c08bb0c17924899c16c672a28a65afded9c46c6ec", size = 431876, upload-time = "2025-10-28T20:56:27.524Z" }, + { url = "https://files.pythonhosted.org/packages/c3/63/688462108c1a00eb9f05765331c107f95ae86f6b197b865d29e930b7e462/aiohttp-3.13.2-cp311-cp311-win_amd64.whl", hash = "sha256:7fd19df530c292542636c2a9a85854fab93474396a52f1695e799186bbd7f24c", size = 456205, upload-time = "2025-10-28T20:56:29.062Z" }, + { url = "https://files.pythonhosted.org/packages/29/9b/01f00e9856d0a73260e86dd8ed0c2234a466c5c1712ce1c281548df39777/aiohttp-3.13.2-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:b1e56bab2e12b2b9ed300218c351ee2a3d8c8fdab5b1ec6193e11a817767e47b", size = 737623, upload-time = "2025-10-28T20:56:30.797Z" }, + { url = "https://files.pythonhosted.org/packages/5a/1b/4be39c445e2b2bd0aab4ba736deb649fabf14f6757f405f0c9685019b9e9/aiohttp-3.13.2-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:364e25edaabd3d37b1db1f0cbcee8c73c9a3727bfa262b83e5e4cf3489a2a9dc", size = 492664, upload-time = "2025-10-28T20:56:32.708Z" }, + { url = "https://files.pythonhosted.org/packages/28/66/d35dcfea8050e131cdd731dff36434390479b4045a8d0b9d7111b0a968f1/aiohttp-3.13.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:c5c94825f744694c4b8db20b71dba9a257cd2ba8e010a803042123f3a25d50d7", size = 491808, upload-time = "2025-10-28T20:56:34.57Z" }, + { url = "https://files.pythonhosted.org/packages/00/29/8e4609b93e10a853b65f8291e64985de66d4f5848c5637cddc70e98f01f8/aiohttp-3.13.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ba2715d842ffa787be87cbfce150d5e88c87a98e0b62e0f5aa489169a393dbbb", size = 1738863, upload-time = "2025-10-28T20:56:36.377Z" }, + { url = "https://files.pythonhosted.org/packages/9d/fa/4ebdf4adcc0def75ced1a0d2d227577cd7b1b85beb7edad85fcc87693c75/aiohttp-3.13.2-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:585542825c4bc662221fb257889e011a5aa00f1ae4d75d1d246a5225289183e3", size = 1700586, upload-time = "2025-10-28T20:56:38.034Z" }, + { url = "https://files.pythonhosted.org/packages/da/04/73f5f02ff348a3558763ff6abe99c223381b0bace05cd4530a0258e52597/aiohttp-3.13.2-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:39d02cb6025fe1aabca329c5632f48c9532a3dabccd859e7e2f110668972331f", size = 1768625, upload-time = "2025-10-28T20:56:39.75Z" }, + { url = "https://files.pythonhosted.org/packages/f8/49/a825b79ffec124317265ca7d2344a86bcffeb960743487cb11988ffb3494/aiohttp-3.13.2-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:e67446b19e014d37342f7195f592a2a948141d15a312fe0e700c2fd2f03124f6", size = 1867281, upload-time = "2025-10-28T20:56:41.471Z" }, + { url = "https://files.pythonhosted.org/packages/b9/48/adf56e05f81eac31edcfae45c90928f4ad50ef2e3ea72cb8376162a368f8/aiohttp-3.13.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4356474ad6333e41ccefd39eae869ba15a6c5299c9c01dfdcfdd5c107be4363e", size = 1752431, upload-time = "2025-10-28T20:56:43.162Z" }, + { url = "https://files.pythonhosted.org/packages/30/ab/593855356eead019a74e862f21523db09c27f12fd24af72dbc3555b9bfd9/aiohttp-3.13.2-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:eeacf451c99b4525f700f078becff32c32ec327b10dcf31306a8a52d78166de7", size = 1562846, upload-time = "2025-10-28T20:56:44.85Z" }, + { url = "https://files.pythonhosted.org/packages/39/0f/9f3d32271aa8dc35036e9668e31870a9d3b9542dd6b3e2c8a30931cb27ae/aiohttp-3.13.2-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:d8a9b889aeabd7a4e9af0b7f4ab5ad94d42e7ff679aaec6d0db21e3b639ad58d", size = 1699606, upload-time = "2025-10-28T20:56:46.519Z" }, + { url = "https://files.pythonhosted.org/packages/2c/3c/52d2658c5699b6ef7692a3f7128b2d2d4d9775f2a68093f74bca06cf01e1/aiohttp-3.13.2-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:fa89cb11bc71a63b69568d5b8a25c3ca25b6d54c15f907ca1c130d72f320b76b", size = 1720663, upload-time = "2025-10-28T20:56:48.528Z" }, + { url = "https://files.pythonhosted.org/packages/9b/d4/8f8f3ff1fb7fb9e3f04fcad4e89d8a1cd8fc7d05de67e3de5b15b33008ff/aiohttp-3.13.2-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:8aa7c807df234f693fed0ecd507192fc97692e61fee5702cdc11155d2e5cadc8", size = 1737939, upload-time = "2025-10-28T20:56:50.77Z" }, + { url = "https://files.pythonhosted.org/packages/03/d3/ddd348f8a27a634daae39a1b8e291ff19c77867af438af844bf8b7e3231b/aiohttp-3.13.2-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:9eb3e33fdbe43f88c3c75fa608c25e7c47bbd80f48d012763cb67c47f39a7e16", size = 1555132, upload-time = "2025-10-28T20:56:52.568Z" }, + { url = "https://files.pythonhosted.org/packages/39/b8/46790692dc46218406f94374903ba47552f2f9f90dad554eed61bfb7b64c/aiohttp-3.13.2-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:9434bc0d80076138ea986833156c5a48c9c7a8abb0c96039ddbb4afc93184169", size = 1764802, upload-time = "2025-10-28T20:56:54.292Z" }, + { url = "https://files.pythonhosted.org/packages/ba/e4/19ce547b58ab2a385e5f0b8aa3db38674785085abcf79b6e0edd1632b12f/aiohttp-3.13.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:ff15c147b2ad66da1f2cbb0622313f2242d8e6e8f9b79b5206c84523a4473248", size = 1719512, upload-time = "2025-10-28T20:56:56.428Z" }, + { url = "https://files.pythonhosted.org/packages/70/30/6355a737fed29dcb6dfdd48682d5790cb5eab050f7b4e01f49b121d3acad/aiohttp-3.13.2-cp312-cp312-win32.whl", hash = "sha256:27e569eb9d9e95dbd55c0fc3ec3a9335defbf1d8bc1d20171a49f3c4c607b93e", size = 426690, upload-time = "2025-10-28T20:56:58.736Z" }, + { url = "https://files.pythonhosted.org/packages/0a/0d/b10ac09069973d112de6ef980c1f6bb31cb7dcd0bc363acbdad58f927873/aiohttp-3.13.2-cp312-cp312-win_amd64.whl", hash = "sha256:8709a0f05d59a71f33fd05c17fc11fcb8c30140506e13c2f5e8ee1b8964e1b45", size = 453465, upload-time = "2025-10-28T20:57:00.795Z" }, + { url = "https://files.pythonhosted.org/packages/bf/78/7e90ca79e5aa39f9694dcfd74f4720782d3c6828113bb1f3197f7e7c4a56/aiohttp-3.13.2-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:7519bdc7dfc1940d201651b52bf5e03f5503bda45ad6eacf64dda98be5b2b6be", size = 732139, upload-time = "2025-10-28T20:57:02.455Z" }, + { url = "https://files.pythonhosted.org/packages/db/ed/1f59215ab6853fbaa5c8495fa6cbc39edfc93553426152b75d82a5f32b76/aiohttp-3.13.2-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:088912a78b4d4f547a1f19c099d5a506df17eacec3c6f4375e2831ec1d995742", size = 490082, upload-time = "2025-10-28T20:57:04.784Z" }, + { url = "https://files.pythonhosted.org/packages/68/7b/fe0fe0f5e05e13629d893c760465173a15ad0039c0a5b0d0040995c8075e/aiohttp-3.13.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:5276807b9de9092af38ed23ce120539ab0ac955547b38563a9ba4f5b07b95293", size = 489035, upload-time = "2025-10-28T20:57:06.894Z" }, + { url = "https://files.pythonhosted.org/packages/d2/04/db5279e38471b7ac801d7d36a57d1230feeee130bbe2a74f72731b23c2b1/aiohttp-3.13.2-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1237c1375eaef0db4dcd7c2559f42e8af7b87ea7d295b118c60c36a6e61cb811", size = 1720387, upload-time = "2025-10-28T20:57:08.685Z" }, + { url = "https://files.pythonhosted.org/packages/31/07/8ea4326bd7dae2bd59828f69d7fdc6e04523caa55e4a70f4a8725a7e4ed2/aiohttp-3.13.2-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:96581619c57419c3d7d78703d5b78c1e5e5fc0172d60f555bdebaced82ded19a", size = 1688314, upload-time = "2025-10-28T20:57:10.693Z" }, + { url = "https://files.pythonhosted.org/packages/48/ab/3d98007b5b87ffd519d065225438cc3b668b2f245572a8cb53da5dd2b1bc/aiohttp-3.13.2-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a2713a95b47374169409d18103366de1050fe0ea73db358fc7a7acb2880422d4", size = 1756317, upload-time = "2025-10-28T20:57:12.563Z" }, + { url = "https://files.pythonhosted.org/packages/97/3d/801ca172b3d857fafb7b50c7c03f91b72b867a13abca982ed6b3081774ef/aiohttp-3.13.2-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:228a1cd556b3caca590e9511a89444925da87d35219a49ab5da0c36d2d943a6a", size = 1858539, upload-time = "2025-10-28T20:57:14.623Z" }, + { url = "https://files.pythonhosted.org/packages/f7/0d/4764669bdf47bd472899b3d3db91fffbe925c8e3038ec591a2fd2ad6a14d/aiohttp-3.13.2-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ac6cde5fba8d7d8c6ac963dbb0256a9854e9fafff52fbcc58fdf819357892c3e", size = 1739597, upload-time = "2025-10-28T20:57:16.399Z" }, + { url = "https://files.pythonhosted.org/packages/c4/52/7bd3c6693da58ba16e657eb904a5b6decfc48ecd06e9ac098591653b1566/aiohttp-3.13.2-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:f2bef8237544f4e42878c61cef4e2839fee6346dc60f5739f876a9c50be7fcdb", size = 1555006, upload-time = "2025-10-28T20:57:18.288Z" }, + { url = "https://files.pythonhosted.org/packages/48/30/9586667acec5993b6f41d2ebcf96e97a1255a85f62f3c653110a5de4d346/aiohttp-3.13.2-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:16f15a4eac3bc2d76c45f7ebdd48a65d41b242eb6c31c2245463b40b34584ded", size = 1683220, upload-time = "2025-10-28T20:57:20.241Z" }, + { url = "https://files.pythonhosted.org/packages/71/01/3afe4c96854cfd7b30d78333852e8e851dceaec1c40fd00fec90c6402dd2/aiohttp-3.13.2-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:bb7fb776645af5cc58ab804c58d7eba545a97e047254a52ce89c157b5af6cd0b", size = 1712570, upload-time = "2025-10-28T20:57:22.253Z" }, + { url = "https://files.pythonhosted.org/packages/11/2c/22799d8e720f4697a9e66fd9c02479e40a49de3de2f0bbe7f9f78a987808/aiohttp-3.13.2-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:e1b4951125ec10c70802f2cb09736c895861cd39fd9dcb35107b4dc8ae6220b8", size = 1733407, upload-time = "2025-10-28T20:57:24.37Z" }, + { url = "https://files.pythonhosted.org/packages/34/cb/90f15dd029f07cebbd91f8238a8b363978b530cd128488085b5703683594/aiohttp-3.13.2-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:550bf765101ae721ee1d37d8095f47b1f220650f85fe1af37a90ce75bab89d04", size = 1550093, upload-time = "2025-10-28T20:57:26.257Z" }, + { url = "https://files.pythonhosted.org/packages/69/46/12dce9be9d3303ecbf4d30ad45a7683dc63d90733c2d9fe512be6716cd40/aiohttp-3.13.2-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:fe91b87fc295973096251e2d25a811388e7d8adf3bd2b97ef6ae78bc4ac6c476", size = 1758084, upload-time = "2025-10-28T20:57:28.349Z" }, + { url = "https://files.pythonhosted.org/packages/f9/c8/0932b558da0c302ffd639fc6362a313b98fdf235dc417bc2493da8394df7/aiohttp-3.13.2-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e0c8e31cfcc4592cb200160344b2fb6ae0f9e4effe06c644b5a125d4ae5ebe23", size = 1716987, upload-time = "2025-10-28T20:57:30.233Z" }, + { url = "https://files.pythonhosted.org/packages/5d/8b/f5bd1a75003daed099baec373aed678f2e9b34f2ad40d85baa1368556396/aiohttp-3.13.2-cp313-cp313-win32.whl", hash = "sha256:0740f31a60848d6edb296a0df827473eede90c689b8f9f2a4cdde74889eb2254", size = 425859, upload-time = "2025-10-28T20:57:32.105Z" }, + { url = "https://files.pythonhosted.org/packages/5d/28/a8a9fc6957b2cee8902414e41816b5ab5536ecf43c3b1843c10e82c559b2/aiohttp-3.13.2-cp313-cp313-win_amd64.whl", hash = "sha256:a88d13e7ca367394908f8a276b89d04a3652044612b9a408a0bb22a5ed976a1a", size = 452192, upload-time = "2025-10-28T20:57:34.166Z" }, + { url = "https://files.pythonhosted.org/packages/9b/36/e2abae1bd815f01c957cbf7be817b3043304e1c87bad526292a0410fdcf9/aiohttp-3.13.2-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:2475391c29230e063ef53a66669b7b691c9bfc3f1426a0f7bcdf1216bdbac38b", size = 735234, upload-time = "2025-10-28T20:57:36.415Z" }, + { url = "https://files.pythonhosted.org/packages/ca/e3/1ee62dde9b335e4ed41db6bba02613295a0d5b41f74a783c142745a12763/aiohttp-3.13.2-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:f33c8748abef4d8717bb20e8fb1b3e07c6adacb7fd6beaae971a764cf5f30d61", size = 490733, upload-time = "2025-10-28T20:57:38.205Z" }, + { url = "https://files.pythonhosted.org/packages/1a/aa/7a451b1d6a04e8d15a362af3e9b897de71d86feac3babf8894545d08d537/aiohttp-3.13.2-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:ae32f24bbfb7dbb485a24b30b1149e2f200be94777232aeadba3eecece4d0aa4", size = 491303, upload-time = "2025-10-28T20:57:40.122Z" }, + { url = "https://files.pythonhosted.org/packages/57/1e/209958dbb9b01174870f6a7538cd1f3f28274fdbc88a750c238e2c456295/aiohttp-3.13.2-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5d7f02042c1f009ffb70067326ef183a047425bb2ff3bc434ead4dd4a4a66a2b", size = 1717965, upload-time = "2025-10-28T20:57:42.28Z" }, + { url = "https://files.pythonhosted.org/packages/08/aa/6a01848d6432f241416bc4866cae8dc03f05a5a884d2311280f6a09c73d6/aiohttp-3.13.2-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:93655083005d71cd6c072cdab54c886e6570ad2c4592139c3fb967bfc19e4694", size = 1667221, upload-time = "2025-10-28T20:57:44.869Z" }, + { url = "https://files.pythonhosted.org/packages/87/4f/36c1992432d31bbc789fa0b93c768d2e9047ec8c7177e5cd84ea85155f36/aiohttp-3.13.2-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:0db1e24b852f5f664cd728db140cf11ea0e82450471232a394b3d1a540b0f906", size = 1757178, upload-time = "2025-10-28T20:57:47.216Z" }, + { url = "https://files.pythonhosted.org/packages/ac/b4/8e940dfb03b7e0f68a82b88fd182b9be0a65cb3f35612fe38c038c3112cf/aiohttp-3.13.2-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b009194665bcd128e23eaddef362e745601afa4641930848af4c8559e88f18f9", size = 1838001, upload-time = "2025-10-28T20:57:49.337Z" }, + { url = "https://files.pythonhosted.org/packages/d7/ef/39f3448795499c440ab66084a9db7d20ca7662e94305f175a80f5b7e0072/aiohttp-3.13.2-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c038a8fdc8103cd51dbd986ecdce141473ffd9775a7a8057a6ed9c3653478011", size = 1716325, upload-time = "2025-10-28T20:57:51.327Z" }, + { url = "https://files.pythonhosted.org/packages/d7/51/b311500ffc860b181c05d91c59a1313bdd05c82960fdd4035a15740d431e/aiohttp-3.13.2-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:66bac29b95a00db411cd758fea0e4b9bdba6d549dfe333f9a945430f5f2cc5a6", size = 1547978, upload-time = "2025-10-28T20:57:53.554Z" }, + { url = "https://files.pythonhosted.org/packages/31/64/b9d733296ef79815226dab8c586ff9e3df41c6aff2e16c06697b2d2e6775/aiohttp-3.13.2-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:4ebf9cfc9ba24a74cf0718f04aac2a3bbe745902cc7c5ebc55c0f3b5777ef213", size = 1682042, upload-time = "2025-10-28T20:57:55.617Z" }, + { url = "https://files.pythonhosted.org/packages/3f/30/43d3e0f9d6473a6db7d472104c4eff4417b1e9df01774cb930338806d36b/aiohttp-3.13.2-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:a4b88ebe35ce54205c7074f7302bd08a4cb83256a3e0870c72d6f68a3aaf8e49", size = 1680085, upload-time = "2025-10-28T20:57:57.59Z" }, + { url = "https://files.pythonhosted.org/packages/16/51/c709f352c911b1864cfd1087577760ced64b3e5bee2aa88b8c0c8e2e4972/aiohttp-3.13.2-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:98c4fb90bb82b70a4ed79ca35f656f4281885be076f3f970ce315402b53099ae", size = 1728238, upload-time = "2025-10-28T20:57:59.525Z" }, + { url = "https://files.pythonhosted.org/packages/19/e2/19bd4c547092b773caeb48ff5ae4b1ae86756a0ee76c16727fcfd281404b/aiohttp-3.13.2-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:ec7534e63ae0f3759df3a1ed4fa6bc8f75082a924b590619c0dd2f76d7043caa", size = 1544395, upload-time = "2025-10-28T20:58:01.914Z" }, + { url = "https://files.pythonhosted.org/packages/cf/87/860f2803b27dfc5ed7be532832a3498e4919da61299b4a1f8eb89b8ff44d/aiohttp-3.13.2-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:5b927cf9b935a13e33644cbed6c8c4b2d0f25b713d838743f8fe7191b33829c4", size = 1742965, upload-time = "2025-10-28T20:58:03.972Z" }, + { url = "https://files.pythonhosted.org/packages/67/7f/db2fc7618925e8c7a601094d5cbe539f732df4fb570740be88ed9e40e99a/aiohttp-3.13.2-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:88d6c017966a78c5265d996c19cdb79235be5e6412268d7e2ce7dee339471b7a", size = 1697585, upload-time = "2025-10-28T20:58:06.189Z" }, + { url = "https://files.pythonhosted.org/packages/0c/07/9127916cb09bb38284db5036036042b7b2c514c8ebaeee79da550c43a6d6/aiohttp-3.13.2-cp314-cp314-win32.whl", hash = "sha256:f7c183e786e299b5d6c49fb43a769f8eb8e04a2726a2bd5887b98b5cc2d67940", size = 431621, upload-time = "2025-10-28T20:58:08.636Z" }, + { url = "https://files.pythonhosted.org/packages/fb/41/554a8a380df6d3a2bba8a7726429a23f4ac62aaf38de43bb6d6cde7b4d4d/aiohttp-3.13.2-cp314-cp314-win_amd64.whl", hash = "sha256:fe242cd381e0fb65758faf5ad96c2e460df6ee5b2de1072fe97e4127927e00b4", size = 457627, upload-time = "2025-10-28T20:58:11Z" }, + { url = "https://files.pythonhosted.org/packages/c7/8e/3824ef98c039d3951cb65b9205a96dd2b20f22241ee17d89c5701557c826/aiohttp-3.13.2-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:f10d9c0b0188fe85398c61147bbd2a657d616c876863bfeff43376e0e3134673", size = 767360, upload-time = "2025-10-28T20:58:13.358Z" }, + { url = "https://files.pythonhosted.org/packages/a4/0f/6a03e3fc7595421274fa34122c973bde2d89344f8a881b728fa8c774e4f1/aiohttp-3.13.2-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:e7c952aefdf2460f4ae55c5e9c3e80aa72f706a6317e06020f80e96253b1accd", size = 504616, upload-time = "2025-10-28T20:58:15.339Z" }, + { url = "https://files.pythonhosted.org/packages/c6/aa/ed341b670f1bc8a6f2c6a718353d13b9546e2cef3544f573c6a1ff0da711/aiohttp-3.13.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:c20423ce14771d98353d2e25e83591fa75dfa90a3c1848f3d7c68243b4fbded3", size = 509131, upload-time = "2025-10-28T20:58:17.693Z" }, + { url = "https://files.pythonhosted.org/packages/7f/f0/c68dac234189dae5c4bbccc0f96ce0cc16b76632cfc3a08fff180045cfa4/aiohttp-3.13.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e96eb1a34396e9430c19d8338d2ec33015e4a87ef2b4449db94c22412e25ccdf", size = 1864168, upload-time = "2025-10-28T20:58:20.113Z" }, + { url = "https://files.pythonhosted.org/packages/8f/65/75a9a76db8364b5d0e52a0c20eabc5d52297385d9af9c35335b924fafdee/aiohttp-3.13.2-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:23fb0783bc1a33640036465019d3bba069942616a6a2353c6907d7fe1ccdaf4e", size = 1719200, upload-time = "2025-10-28T20:58:22.583Z" }, + { url = "https://files.pythonhosted.org/packages/f5/55/8df2ed78d7f41d232f6bd3ff866b6f617026551aa1d07e2f03458f964575/aiohttp-3.13.2-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:2e1a9bea6244a1d05a4e57c295d69e159a5c50d8ef16aa390948ee873478d9a5", size = 1843497, upload-time = "2025-10-28T20:58:24.672Z" }, + { url = "https://files.pythonhosted.org/packages/e9/e0/94d7215e405c5a02ccb6a35c7a3a6cfff242f457a00196496935f700cde5/aiohttp-3.13.2-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0a3d54e822688b56e9f6b5816fb3de3a3a64660efac64e4c2dc435230ad23bad", size = 1935703, upload-time = "2025-10-28T20:58:26.758Z" }, + { url = "https://files.pythonhosted.org/packages/0b/78/1eeb63c3f9b2d1015a4c02788fb543141aad0a03ae3f7a7b669b2483f8d4/aiohttp-3.13.2-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7a653d872afe9f33497215745da7a943d1dc15b728a9c8da1c3ac423af35178e", size = 1792738, upload-time = "2025-10-28T20:58:29.787Z" }, + { url = "https://files.pythonhosted.org/packages/41/75/aaf1eea4c188e51538c04cc568040e3082db263a57086ea74a7d38c39e42/aiohttp-3.13.2-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:56d36e80d2003fa3fc0207fac644216d8532e9504a785ef9a8fd013f84a42c61", size = 1624061, upload-time = "2025-10-28T20:58:32.529Z" }, + { url = "https://files.pythonhosted.org/packages/9b/c2/3b6034de81fbcc43de8aeb209073a2286dfb50b86e927b4efd81cf848197/aiohttp-3.13.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:78cd586d8331fb8e241c2dd6b2f4061778cc69e150514b39a9e28dd050475661", size = 1789201, upload-time = "2025-10-28T20:58:34.618Z" }, + { url = "https://files.pythonhosted.org/packages/c9/38/c15dcf6d4d890217dae79d7213988f4e5fe6183d43893a9cf2fe9e84ca8d/aiohttp-3.13.2-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:20b10bbfbff766294fe99987f7bb3b74fdd2f1a2905f2562132641ad434dcf98", size = 1776868, upload-time = "2025-10-28T20:58:38.835Z" }, + { url = "https://files.pythonhosted.org/packages/04/75/f74fd178ac81adf4f283a74847807ade5150e48feda6aef024403716c30c/aiohttp-3.13.2-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:9ec49dff7e2b3c85cdeaa412e9d438f0ecd71676fde61ec57027dd392f00c693", size = 1790660, upload-time = "2025-10-28T20:58:41.507Z" }, + { url = "https://files.pythonhosted.org/packages/e7/80/7368bd0d06b16b3aba358c16b919e9c46cf11587dc572091031b0e9e3ef0/aiohttp-3.13.2-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:94f05348c4406450f9d73d38efb41d669ad6cd90c7ee194810d0eefbfa875a7a", size = 1617548, upload-time = "2025-10-28T20:58:43.674Z" }, + { url = "https://files.pythonhosted.org/packages/7d/4b/a6212790c50483cb3212e507378fbe26b5086d73941e1ec4b56a30439688/aiohttp-3.13.2-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:fa4dcb605c6f82a80c7f95713c2b11c3b8e9893b3ebd2bc9bde93165ed6107be", size = 1817240, upload-time = "2025-10-28T20:58:45.787Z" }, + { url = "https://files.pythonhosted.org/packages/ff/f7/ba5f0ba4ea8d8f3c32850912944532b933acbf0f3a75546b89269b9b7dde/aiohttp-3.13.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:cf00e5db968c3f67eccd2778574cf64d8b27d95b237770aa32400bd7a1ca4f6c", size = 1762334, upload-time = "2025-10-28T20:58:47.936Z" }, + { url = "https://files.pythonhosted.org/packages/7e/83/1a5a1856574588b1cad63609ea9ad75b32a8353ac995d830bf5da9357364/aiohttp-3.13.2-cp314-cp314t-win32.whl", hash = "sha256:d23b5fe492b0805a50d3371e8a728a9134d8de5447dce4c885f5587294750734", size = 464685, upload-time = "2025-10-28T20:58:50.642Z" }, + { url = "https://files.pythonhosted.org/packages/9f/4d/d22668674122c08f4d56972297c51a624e64b3ed1efaa40187607a7cb66e/aiohttp-3.13.2-cp314-cp314t-win_amd64.whl", hash = "sha256:ff0a7b0a82a7ab905cbda74006318d1b12e37c797eb1b0d4eb3e316cf47f658f", size = 498093, upload-time = "2025-10-28T20:58:52.782Z" }, +] + +[[package]] +name = "aiosignal" +version = "1.4.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "frozenlist" }, + { name = "typing-extensions", marker = "python_full_version < '3.13'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/61/62/06741b579156360248d1ec624842ad0edf697050bbaf7c3e46394e106ad1/aiosignal-1.4.0.tar.gz", hash = "sha256:f47eecd9468083c2029cc99945502cb7708b082c232f9aca65da147157b251c7", size = 25007, upload-time = "2025-07-03T22:54:43.528Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/fb/76/641ae371508676492379f16e2fa48f4e2c11741bd63c48be4b12a6b09cba/aiosignal-1.4.0-py3-none-any.whl", hash = "sha256:053243f8b92b990551949e63930a839ff0cf0b0ebbe0597b0f3fb19e1a0fe82e", size = 7490, upload-time = "2025-07-03T22:54:42.156Z" }, +] + +[[package]] +name = "amdsmi" +version = "7.0.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/45/c1/330da195623ec7d9f699be2dbec98df364b1def9b48aa169f1abe369804f/amdsmi-7.0.2.tar.gz", hash = "sha256:3e622e48c630b889045a6f57387455cdf082066348718172dd8af6d275baf8f2", size = 61577, upload-time = "2025-10-11T05:17:44.898Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/2b/cf/bacc741d9926d662e76fd694f5bc63dd4c2471e32cedfb1c3cca7e47aa3c/amdsmi-7.0.2-py3-none-any.whl", hash = "sha256:db5aa757f8ed82dfd799c4d39e2828542678dc4e485b0ab7fabe5f398fec5652", size = 64366, upload-time = "2025-10-11T05:17:44.047Z" }, +] + +[[package]] +name = "aniso8601" +version = "10.0.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/8b/8d/52179c4e3f1978d3d9a285f98c706642522750ef343e9738286130423730/aniso8601-10.0.1.tar.gz", hash = "sha256:25488f8663dd1528ae1f54f94ac1ea51ae25b4d531539b8bc707fed184d16845", size = 47190, upload-time = "2025-04-18T17:29:42.995Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/59/75/e0e10dc7ed1408c28e03a6cb2d7a407f99320eb953f229d008a7a6d05546/aniso8601-10.0.1-py2.py3-none-any.whl", hash = "sha256:eb19717fd4e0db6de1aab06f12450ab92144246b257423fe020af5748c0cb89e", size = 52848, upload-time = "2025-04-18T17:29:41.492Z" }, +] + +[[package]] +name = "annotated-doc" +version = "0.0.4" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/57/ba/046ceea27344560984e26a590f90bc7f4a75b06701f653222458922b558c/annotated_doc-0.0.4.tar.gz", hash = "sha256:fbcda96e87e9c92ad167c2e53839e57503ecfda18804ea28102353485033faa4", size = 7288, upload-time = "2025-11-10T22:07:42.062Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/1e/d3/26bf1008eb3d2daa8ef4cacc7f3bfdc11818d111f7e2d0201bc6e3b49d45/annotated_doc-0.0.4-py3-none-any.whl", hash = "sha256:571ac1dc6991c450b25a9c2d84a3705e2ae7a53467b5d111c24fa8baabbed320", size = 5303, upload-time = "2025-11-10T22:07:40.673Z" }, +] + +[[package]] +name = "annotated-types" +version = "0.7.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/ee/67/531ea369ba64dcff5ec9c3402f9f51bf748cec26dde048a2f973a4eea7f5/annotated_types-0.7.0.tar.gz", hash = "sha256:aff07c09a53a08bc8cfccb9c85b05f1aa9a2a6f23728d790723543408344ce89", size = 16081, upload-time = "2024-05-20T21:33:25.928Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/78/b6/6307fbef88d9b5ee7421e68d78a9f162e0da4900bc5f5793f6d3d0e34fb8/annotated_types-0.7.0-py3-none-any.whl", hash = "sha256:1f02e8b43a8fbbc3f3e0d4f0f4bfc8131bcb4eebe8849b8e5c773f3a1c582a53", size = 13643, upload-time = "2024-05-20T21:33:24.1Z" }, +] + +[[package]] +name = "ansicon" +version = "1.89.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/b6/e2/1c866404ddbd280efedff4a9f15abfe943cb83cde6e895022370f3a61f85/ansicon-1.89.0.tar.gz", hash = "sha256:e4d039def5768a47e4afec8e89e83ec3ae5a26bf00ad851f914d1240b444d2b1", size = 67312, upload-time = "2019-04-29T20:23:57.314Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/75/f9/f1c10e223c7b56a38109a3f2eb4e7fe9a757ea3ed3a166754fb30f65e466/ansicon-1.89.0-py2.py3-none-any.whl", hash = "sha256:f1def52d17f65c2c9682cf8370c03f541f410c1752d6a14029f97318e4b9dfec", size = 63675, upload-time = "2019-04-29T20:23:53.83Z" }, +] + +[[package]] +name = "anyio" +version = "4.11.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "exceptiongroup", marker = "python_full_version < '3.11'" }, + { name = "idna" }, + { name = "sniffio" }, + { name = "typing-extensions", marker = "python_full_version < '3.13'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/c6/78/7d432127c41b50bccba979505f272c16cbcadcc33645d5fa3a738110ae75/anyio-4.11.0.tar.gz", hash = "sha256:82a8d0b81e318cc5ce71a5f1f8b5c4e63619620b63141ef8c995fa0db95a57c4", size = 219094, upload-time = "2025-09-23T09:19:12.58Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/15/b3/9b1a8074496371342ec1e796a96f99c82c945a339cd81a8e73de28b4cf9e/anyio-4.11.0-py3-none-any.whl", hash = "sha256:0287e96f4d26d4149305414d4e3bc32f0dcd0862365a4bddea19d7a1ec38c4fc", size = 109097, upload-time = "2025-09-23T09:19:10.601Z" }, +] + +[[package]] +name = "async-timeout" +version = "5.0.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/a5/ae/136395dfbfe00dfc94da3f3e136d0b13f394cba8f4841120e34226265780/async_timeout-5.0.1.tar.gz", hash = "sha256:d9321a7a3d5a6a5e187e824d2fa0793ce379a202935782d555d6e9d2735677d3", size = 9274, upload-time = "2024-11-06T16:41:39.6Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/fe/ba/e2081de779ca30d473f21f5b30e0e737c438205440784c7dfc81efc2b029/async_timeout-5.0.1-py3-none-any.whl", hash = "sha256:39e3809566ff85354557ec2398b55e096c8364bacac9405a7a1fa429e77fe76c", size = 6233, upload-time = "2024-11-06T16:41:37.9Z" }, +] + +[[package]] +name = "attrs" +version = "25.4.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/6b/5c/685e6633917e101e5dcb62b9dd76946cbb57c26e133bae9e0cd36033c0a9/attrs-25.4.0.tar.gz", hash = "sha256:16d5969b87f0859ef33a48b35d55ac1be6e42ae49d5e853b597db70c35c57e11", size = 934251, upload-time = "2025-10-06T13:54:44.725Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/3a/2a/7cc015f5b9f5db42b7d48157e23356022889fc354a2813c15934b7cb5c0e/attrs-25.4.0-py3-none-any.whl", hash = "sha256:adcf7e2a1fb3b36ac48d97835bb6d8ade15b8dcce26aba8bf1d14847b57a3373", size = 67615, upload-time = "2025-10-06T13:54:43.17Z" }, +] + +[[package]] +name = "audioop-lts" +version = "0.2.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/38/53/946db57842a50b2da2e0c1e34bd37f36f5aadba1a929a3971c5d7841dbca/audioop_lts-0.2.2.tar.gz", hash = "sha256:64d0c62d88e67b98a1a5e71987b7aa7b5bcffc7dcee65b635823dbdd0a8dbbd0", size = 30686, upload-time = "2025-08-05T16:43:17.409Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/de/d4/94d277ca941de5a507b07f0b592f199c22454eeaec8f008a286b3fbbacd6/audioop_lts-0.2.2-cp313-abi3-macosx_10_13_universal2.whl", hash = "sha256:fd3d4602dc64914d462924a08c1a9816435a2155d74f325853c1f1ac3b2d9800", size = 46523, upload-time = "2025-08-05T16:42:20.836Z" }, + { url = "https://files.pythonhosted.org/packages/f8/5a/656d1c2da4b555920ce4177167bfeb8623d98765594af59702c8873f60ec/audioop_lts-0.2.2-cp313-abi3-macosx_10_13_x86_64.whl", hash = "sha256:550c114a8df0aafe9a05442a1162dfc8fec37e9af1d625ae6060fed6e756f303", size = 27455, upload-time = "2025-08-05T16:42:22.283Z" }, + { url = "https://files.pythonhosted.org/packages/1b/83/ea581e364ce7b0d41456fb79d6ee0ad482beda61faf0cab20cbd4c63a541/audioop_lts-0.2.2-cp313-abi3-macosx_11_0_arm64.whl", hash = "sha256:9a13dc409f2564de15dd68be65b462ba0dde01b19663720c68c1140c782d1d75", size = 26997, upload-time = "2025-08-05T16:42:23.849Z" }, + { url = "https://files.pythonhosted.org/packages/b8/3b/e8964210b5e216e5041593b7d33e97ee65967f17c282e8510d19c666dab4/audioop_lts-0.2.2-cp313-abi3-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:51c916108c56aa6e426ce611946f901badac950ee2ddaf302b7ed35d9958970d", size = 85844, upload-time = "2025-08-05T16:42:25.208Z" }, + { url = "https://files.pythonhosted.org/packages/c7/2e/0a1c52faf10d51def20531a59ce4c706cb7952323b11709e10de324d6493/audioop_lts-0.2.2-cp313-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:47eba38322370347b1c47024defbd36374a211e8dd5b0dcbce7b34fdb6f8847b", size = 85056, upload-time = "2025-08-05T16:42:26.559Z" }, + { url = "https://files.pythonhosted.org/packages/75/e8/cd95eef479656cb75ab05dfece8c1f8c395d17a7c651d88f8e6e291a63ab/audioop_lts-0.2.2-cp313-abi3-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ba7c3a7e5f23e215cb271516197030c32aef2e754252c4c70a50aaff7031a2c8", size = 93892, upload-time = "2025-08-05T16:42:27.902Z" }, + { url = "https://files.pythonhosted.org/packages/5c/1e/a0c42570b74f83efa5cca34905b3eef03f7ab09fe5637015df538a7f3345/audioop_lts-0.2.2-cp313-abi3-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:def246fe9e180626731b26e89816e79aae2276f825420a07b4a647abaa84becc", size = 96660, upload-time = "2025-08-05T16:42:28.9Z" }, + { url = "https://files.pythonhosted.org/packages/50/d5/8a0ae607ca07dbb34027bac8db805498ee7bfecc05fd2c148cc1ed7646e7/audioop_lts-0.2.2-cp313-abi3-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:e160bf9df356d841bb6c180eeeea1834085464626dc1b68fa4e1d59070affdc3", size = 79143, upload-time = "2025-08-05T16:42:29.929Z" }, + { url = "https://files.pythonhosted.org/packages/12/17/0d28c46179e7910bfb0bb62760ccb33edb5de973052cb2230b662c14ca2e/audioop_lts-0.2.2-cp313-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:4b4cd51a57b698b2d06cb9993b7ac8dfe89a3b2878e96bc7948e9f19ff51dba6", size = 84313, upload-time = "2025-08-05T16:42:30.949Z" }, + { url = "https://files.pythonhosted.org/packages/84/ba/bd5d3806641564f2024e97ca98ea8f8811d4e01d9b9f9831474bc9e14f9e/audioop_lts-0.2.2-cp313-abi3-musllinux_1_2_ppc64le.whl", hash = "sha256:4a53aa7c16a60a6857e6b0b165261436396ef7293f8b5c9c828a3a203147ed4a", size = 93044, upload-time = "2025-08-05T16:42:31.959Z" }, + { url = "https://files.pythonhosted.org/packages/f9/5e/435ce8d5642f1f7679540d1e73c1c42d933331c0976eb397d1717d7f01a3/audioop_lts-0.2.2-cp313-abi3-musllinux_1_2_riscv64.whl", hash = "sha256:3fc38008969796f0f689f1453722a0f463da1b8a6fbee11987830bfbb664f623", size = 78766, upload-time = "2025-08-05T16:42:33.302Z" }, + { url = "https://files.pythonhosted.org/packages/ae/3b/b909e76b606cbfd53875693ec8c156e93e15a1366a012f0b7e4fb52d3c34/audioop_lts-0.2.2-cp313-abi3-musllinux_1_2_s390x.whl", hash = "sha256:15ab25dd3e620790f40e9ead897f91e79c0d3ce65fe193c8ed6c26cffdd24be7", size = 87640, upload-time = "2025-08-05T16:42:34.854Z" }, + { url = "https://files.pythonhosted.org/packages/30/e7/8f1603b4572d79b775f2140d7952f200f5e6c62904585d08a01f0a70393a/audioop_lts-0.2.2-cp313-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:03f061a1915538fd96272bac9551841859dbb2e3bf73ebe4a23ef043766f5449", size = 86052, upload-time = "2025-08-05T16:42:35.839Z" }, + { url = "https://files.pythonhosted.org/packages/b5/96/c37846df657ccdda62ba1ae2b6534fa90e2e1b1742ca8dcf8ebd38c53801/audioop_lts-0.2.2-cp313-abi3-win32.whl", hash = "sha256:3bcddaaf6cc5935a300a8387c99f7a7fbbe212a11568ec6cf6e4bc458c048636", size = 26185, upload-time = "2025-08-05T16:42:37.04Z" }, + { url = "https://files.pythonhosted.org/packages/34/a5/9d78fdb5b844a83da8a71226c7bdae7cc638861085fff7a1d707cb4823fa/audioop_lts-0.2.2-cp313-abi3-win_amd64.whl", hash = "sha256:a2c2a947fae7d1062ef08c4e369e0ba2086049a5e598fda41122535557012e9e", size = 30503, upload-time = "2025-08-05T16:42:38.427Z" }, + { url = "https://files.pythonhosted.org/packages/34/25/20d8fde083123e90c61b51afb547bb0ea7e77bab50d98c0ab243d02a0e43/audioop_lts-0.2.2-cp313-abi3-win_arm64.whl", hash = "sha256:5f93a5db13927a37d2d09637ccca4b2b6b48c19cd9eda7b17a2e9f77edee6a6f", size = 24173, upload-time = "2025-08-05T16:42:39.704Z" }, + { url = "https://files.pythonhosted.org/packages/58/a7/0a764f77b5c4ac58dc13c01a580f5d32ae8c74c92020b961556a43e26d02/audioop_lts-0.2.2-cp313-cp313t-macosx_10_13_universal2.whl", hash = "sha256:73f80bf4cd5d2ca7814da30a120de1f9408ee0619cc75da87d0641273d202a09", size = 47096, upload-time = "2025-08-05T16:42:40.684Z" }, + { url = "https://files.pythonhosted.org/packages/aa/ed/ebebedde1a18848b085ad0fa54b66ceb95f1f94a3fc04f1cd1b5ccb0ed42/audioop_lts-0.2.2-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:106753a83a25ee4d6f473f2be6b0966fc1c9af7e0017192f5531a3e7463dce58", size = 27748, upload-time = "2025-08-05T16:42:41.992Z" }, + { url = "https://files.pythonhosted.org/packages/cb/6e/11ca8c21af79f15dbb1c7f8017952ee8c810c438ce4e2b25638dfef2b02c/audioop_lts-0.2.2-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:fbdd522624141e40948ab3e8cdae6e04c748d78710e9f0f8d4dae2750831de19", size = 27329, upload-time = "2025-08-05T16:42:42.987Z" }, + { url = "https://files.pythonhosted.org/packages/84/52/0022f93d56d85eec5da6b9da6a958a1ef09e80c39f2cc0a590c6af81dcbb/audioop_lts-0.2.2-cp313-cp313t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:143fad0311e8209ece30a8dbddab3b65ab419cbe8c0dde6e8828da25999be911", size = 92407, upload-time = "2025-08-05T16:42:44.336Z" }, + { url = "https://files.pythonhosted.org/packages/87/1d/48a889855e67be8718adbc7a01f3c01d5743c325453a5e81cf3717664aad/audioop_lts-0.2.2-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:dfbbc74ec68a0fd08cfec1f4b5e8cca3d3cd7de5501b01c4b5d209995033cde9", size = 91811, upload-time = "2025-08-05T16:42:45.325Z" }, + { url = "https://files.pythonhosted.org/packages/98/a6/94b7213190e8077547ffae75e13ed05edc488653c85aa5c41472c297d295/audioop_lts-0.2.2-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:cfcac6aa6f42397471e4943e0feb2244549db5c5d01efcd02725b96af417f3fe", size = 100470, upload-time = "2025-08-05T16:42:46.468Z" }, + { url = "https://files.pythonhosted.org/packages/e9/e9/78450d7cb921ede0cfc33426d3a8023a3bda755883c95c868ee36db8d48d/audioop_lts-0.2.2-cp313-cp313t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:752d76472d9804ac60f0078c79cdae8b956f293177acd2316cd1e15149aee132", size = 103878, upload-time = "2025-08-05T16:42:47.576Z" }, + { url = "https://files.pythonhosted.org/packages/4f/e2/cd5439aad4f3e34ae1ee852025dc6aa8f67a82b97641e390bf7bd9891d3e/audioop_lts-0.2.2-cp313-cp313t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:83c381767e2cc10e93e40281a04852facc4cd9334550e0f392f72d1c0a9c5753", size = 84867, upload-time = "2025-08-05T16:42:49.003Z" }, + { url = "https://files.pythonhosted.org/packages/68/4b/9d853e9076c43ebba0d411e8d2aa19061083349ac695a7d082540bad64d0/audioop_lts-0.2.2-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:c0022283e9556e0f3643b7c3c03f05063ca72b3063291834cca43234f20c60bb", size = 90001, upload-time = "2025-08-05T16:42:50.038Z" }, + { url = "https://files.pythonhosted.org/packages/58/26/4bae7f9d2f116ed5593989d0e521d679b0d583973d203384679323d8fa85/audioop_lts-0.2.2-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:a2d4f1513d63c795e82948e1305f31a6d530626e5f9f2605408b300ae6095093", size = 99046, upload-time = "2025-08-05T16:42:51.111Z" }, + { url = "https://files.pythonhosted.org/packages/b2/67/a9f4fb3e250dda9e9046f8866e9fa7d52664f8985e445c6b4ad6dfb55641/audioop_lts-0.2.2-cp313-cp313t-musllinux_1_2_riscv64.whl", hash = "sha256:c9c8e68d8b4a56fda8c025e538e639f8c5953f5073886b596c93ec9b620055e7", size = 84788, upload-time = "2025-08-05T16:42:52.198Z" }, + { url = "https://files.pythonhosted.org/packages/70/f7/3de86562db0121956148bcb0fe5b506615e3bcf6e63c4357a612b910765a/audioop_lts-0.2.2-cp313-cp313t-musllinux_1_2_s390x.whl", hash = "sha256:96f19de485a2925314f5020e85911fb447ff5fbef56e8c7c6927851b95533a1c", size = 94472, upload-time = "2025-08-05T16:42:53.59Z" }, + { url = "https://files.pythonhosted.org/packages/f1/32/fd772bf9078ae1001207d2df1eef3da05bea611a87dd0e8217989b2848fa/audioop_lts-0.2.2-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:e541c3ef484852ef36545f66209444c48b28661e864ccadb29daddb6a4b8e5f5", size = 92279, upload-time = "2025-08-05T16:42:54.632Z" }, + { url = "https://files.pythonhosted.org/packages/4f/41/affea7181592ab0ab560044632571a38edaf9130b84928177823fbf3176a/audioop_lts-0.2.2-cp313-cp313t-win32.whl", hash = "sha256:d5e73fa573e273e4f2e5ff96f9043858a5e9311e94ffefd88a3186a910c70917", size = 26568, upload-time = "2025-08-05T16:42:55.627Z" }, + { url = "https://files.pythonhosted.org/packages/28/2b/0372842877016641db8fc54d5c88596b542eec2f8f6c20a36fb6612bf9ee/audioop_lts-0.2.2-cp313-cp313t-win_amd64.whl", hash = "sha256:9191d68659eda01e448188f60364c7763a7ca6653ed3f87ebb165822153a8547", size = 30942, upload-time = "2025-08-05T16:42:56.674Z" }, + { url = "https://files.pythonhosted.org/packages/ee/ca/baf2b9cc7e96c179bb4a54f30fcd83e6ecb340031bde68f486403f943768/audioop_lts-0.2.2-cp313-cp313t-win_arm64.whl", hash = "sha256:c174e322bb5783c099aaf87faeb240c8d210686b04bd61dfd05a8e5a83d88969", size = 24603, upload-time = "2025-08-05T16:42:57.571Z" }, + { url = "https://files.pythonhosted.org/packages/5c/73/413b5a2804091e2c7d5def1d618e4837f1cb82464e230f827226278556b7/audioop_lts-0.2.2-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:f9ee9b52f5f857fbaf9d605a360884f034c92c1c23021fb90b2e39b8e64bede6", size = 47104, upload-time = "2025-08-05T16:42:58.518Z" }, + { url = "https://files.pythonhosted.org/packages/ae/8c/daa3308dc6593944410c2c68306a5e217f5c05b70a12e70228e7dd42dc5c/audioop_lts-0.2.2-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:49ee1a41738a23e98d98b937a0638357a2477bc99e61b0f768a8f654f45d9b7a", size = 27754, upload-time = "2025-08-05T16:43:00.132Z" }, + { url = "https://files.pythonhosted.org/packages/4e/86/c2e0f627168fcf61781a8f72cab06b228fe1da4b9fa4ab39cfb791b5836b/audioop_lts-0.2.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:5b00be98ccd0fc123dcfad31d50030d25fcf31488cde9e61692029cd7394733b", size = 27332, upload-time = "2025-08-05T16:43:01.666Z" }, + { url = "https://files.pythonhosted.org/packages/c7/bd/35dce665255434f54e5307de39e31912a6f902d4572da7c37582809de14f/audioop_lts-0.2.2-cp314-cp314t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:a6d2e0f9f7a69403e388894d4ca5ada5c47230716a03f2847cfc7bd1ecb589d6", size = 92396, upload-time = "2025-08-05T16:43:02.991Z" }, + { url = "https://files.pythonhosted.org/packages/2d/d2/deeb9f51def1437b3afa35aeb729d577c04bcd89394cb56f9239a9f50b6f/audioop_lts-0.2.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f9b0b8a03ef474f56d1a842af1a2e01398b8f7654009823c6d9e0ecff4d5cfbf", size = 91811, upload-time = "2025-08-05T16:43:04.096Z" }, + { url = "https://files.pythonhosted.org/packages/76/3b/09f8b35b227cee28cc8231e296a82759ed80c1a08e349811d69773c48426/audioop_lts-0.2.2-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:2b267b70747d82125f1a021506565bdc5609a2b24bcb4773c16d79d2bb260bbd", size = 100483, upload-time = "2025-08-05T16:43:05.085Z" }, + { url = "https://files.pythonhosted.org/packages/0b/15/05b48a935cf3b130c248bfdbdea71ce6437f5394ee8533e0edd7cfd93d5e/audioop_lts-0.2.2-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0337d658f9b81f4cd0fdb1f47635070cc084871a3d4646d9de74fdf4e7c3d24a", size = 103885, upload-time = "2025-08-05T16:43:06.197Z" }, + { url = "https://files.pythonhosted.org/packages/83/80/186b7fce6d35b68d3d739f228dc31d60b3412105854edb975aa155a58339/audioop_lts-0.2.2-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:167d3b62586faef8b6b2275c3218796b12621a60e43f7e9d5845d627b9c9b80e", size = 84899, upload-time = "2025-08-05T16:43:07.291Z" }, + { url = "https://files.pythonhosted.org/packages/49/89/c78cc5ac6cb5828f17514fb12966e299c850bc885e80f8ad94e38d450886/audioop_lts-0.2.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:0d9385e96f9f6da847f4d571ce3cb15b5091140edf3db97276872647ce37efd7", size = 89998, upload-time = "2025-08-05T16:43:08.335Z" }, + { url = "https://files.pythonhosted.org/packages/4c/4b/6401888d0c010e586c2ca50fce4c903d70a6bb55928b16cfbdfd957a13da/audioop_lts-0.2.2-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:48159d96962674eccdca9a3df280e864e8ac75e40a577cc97c5c42667ffabfc5", size = 99046, upload-time = "2025-08-05T16:43:09.367Z" }, + { url = "https://files.pythonhosted.org/packages/de/f8/c874ca9bb447dae0e2ef2e231f6c4c2b0c39e31ae684d2420b0f9e97ee68/audioop_lts-0.2.2-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:8fefe5868cd082db1186f2837d64cfbfa78b548ea0d0543e9b28935ccce81ce9", size = 84843, upload-time = "2025-08-05T16:43:10.749Z" }, + { url = "https://files.pythonhosted.org/packages/3e/c0/0323e66f3daebc13fd46b36b30c3be47e3fc4257eae44f1e77eb828c703f/audioop_lts-0.2.2-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:58cf54380c3884fb49fdd37dfb7a772632b6701d28edd3e2904743c5e1773602", size = 94490, upload-time = "2025-08-05T16:43:12.131Z" }, + { url = "https://files.pythonhosted.org/packages/98/6b/acc7734ac02d95ab791c10c3f17ffa3584ccb9ac5c18fd771c638ed6d1f5/audioop_lts-0.2.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:088327f00488cdeed296edd9215ca159f3a5a5034741465789cad403fcf4bec0", size = 92297, upload-time = "2025-08-05T16:43:13.139Z" }, + { url = "https://files.pythonhosted.org/packages/13/c3/c3dc3f564ce6877ecd2a05f8d751b9b27a8c320c2533a98b0c86349778d0/audioop_lts-0.2.2-cp314-cp314t-win32.whl", hash = "sha256:068aa17a38b4e0e7de771c62c60bbca2455924b67a8814f3b0dee92b5820c0b3", size = 27331, upload-time = "2025-08-05T16:43:14.19Z" }, + { url = "https://files.pythonhosted.org/packages/72/bb/b4608537e9ffcb86449091939d52d24a055216a36a8bf66b936af8c3e7ac/audioop_lts-0.2.2-cp314-cp314t-win_amd64.whl", hash = "sha256:a5bf613e96f49712073de86f20dbdd4014ca18efd4d34ed18c75bd808337851b", size = 31697, upload-time = "2025-08-05T16:43:15.193Z" }, + { url = "https://files.pythonhosted.org/packages/f6/22/91616fe707a5c5510de2cac9b046a30defe7007ba8a0c04f9c08f27df312/audioop_lts-0.2.2-cp314-cp314t-win_arm64.whl", hash = "sha256:b492c3b040153e68b9fdaff5913305aaaba5bb433d8a7f73d5cf6a64ed3cc1dd", size = 25206, upload-time = "2025-08-05T16:43:16.444Z" }, +] + +[[package]] +name = "av" +version = "16.1.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/78/cd/3a83ffbc3cc25b39721d174487fb0d51a76582f4a1703f98e46170ce83d4/av-16.1.0.tar.gz", hash = "sha256:a094b4fd87a3721dacf02794d3d2c82b8d712c85b9534437e82a8a978c175ffd", size = 4285203, upload-time = "2026-01-11T07:31:33.772Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/97/51/2217a9249409d2e88e16e3f16f7c0def9fd3e7ffc4238b2ec211f9935bdb/av-16.1.0-cp310-cp310-macosx_11_0_x86_64.whl", hash = "sha256:2395748b0c34fe3a150a1721e4f3d4487b939520991b13e7b36f8926b3b12295", size = 26942590, upload-time = "2026-01-09T20:17:58.588Z" }, + { url = "https://files.pythonhosted.org/packages/bf/cd/a7070f4febc76a327c38808e01e2ff6b94531fe0b321af54ea3915165338/av-16.1.0-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:72d7ac832710a158eeb7a93242370aa024a7646516291c562ee7f14a7ea881fd", size = 21507910, upload-time = "2026-01-09T20:18:02.309Z" }, + { url = "https://files.pythonhosted.org/packages/ae/30/ec812418cd9b297f0238fe20eb0747d8a8b68d82c5f73c56fe519a274143/av-16.1.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:6cbac833092e66b6b0ac4d81ab077970b8ca874951e9c3974d41d922aaa653ed", size = 38738309, upload-time = "2026-01-09T20:18:04.701Z" }, + { url = "https://files.pythonhosted.org/packages/3a/b8/6c5795bf1f05f45c5261f8bce6154e0e5e86b158a6676650ddd77c28805e/av-16.1.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:eb990672d97c18f99c02f31c8d5750236f770ffe354b5a52c5f4d16c5e65f619", size = 40293006, upload-time = "2026-01-09T20:18:07.238Z" }, + { url = "https://files.pythonhosted.org/packages/a7/44/5e183bcb9333fc3372ee6e683be8b0c9b515a506894b2d32ff465430c074/av-16.1.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:05ad70933ac3b8ef896a820ea64b33b6cca91a5fac5259cb9ba7fa010435be15", size = 40123516, upload-time = "2026-01-09T20:18:09.955Z" }, + { url = "https://files.pythonhosted.org/packages/12/1d/b5346d582a3c3d958b4d26a2cc63ce607233582d956121eb20d2bbe55c2e/av-16.1.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:d831a1062a3c47520bf99de6ec682bd1d64a40dfa958e5457bb613c5270e7ce3", size = 41463289, upload-time = "2026-01-09T20:18:12.459Z" }, + { url = "https://files.pythonhosted.org/packages/fa/31/acc946c0545f72b8d0d74584cb2a0ade9b7dfe2190af3ef9aa52a2e3c0b1/av-16.1.0-cp310-cp310-win_amd64.whl", hash = "sha256:358ab910fef3c5a806c55176f2b27e5663b33c4d0a692dafeb049c6ed71f8aff", size = 31754959, upload-time = "2026-01-09T20:18:14.718Z" }, + { url = "https://files.pythonhosted.org/packages/48/d0/b71b65d1b36520dcb8291a2307d98b7fc12329a45614a303ff92ada4d723/av-16.1.0-cp311-cp311-macosx_11_0_x86_64.whl", hash = "sha256:e88ad64ee9d2b9c4c5d891f16c22ae78e725188b8926eb88187538d9dd0b232f", size = 26927747, upload-time = "2026-01-09T20:18:16.976Z" }, + { url = "https://files.pythonhosted.org/packages/2f/79/720a5a6ccdee06eafa211b945b0a450e3a0b8fc3d12922f0f3c454d870d2/av-16.1.0-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:cb296073fa6935724de72593800ba86ae49ed48af03960a4aee34f8a611f442b", size = 21492232, upload-time = "2026-01-09T20:18:19.266Z" }, + { url = "https://files.pythonhosted.org/packages/8e/4f/a1ba8d922f2f6d1a3d52419463ef26dd6c4d43ee364164a71b424b5ae204/av-16.1.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:720edd4d25aa73723c1532bb0597806d7b9af5ee34fc02358782c358cfe2f879", size = 39291737, upload-time = "2026-01-09T20:18:21.513Z" }, + { url = "https://files.pythonhosted.org/packages/1a/31/fc62b9fe8738d2693e18d99f040b219e26e8df894c10d065f27c6b4f07e3/av-16.1.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:c7f2bc703d0df260a1fdf4de4253c7f5500ca9fc57772ea241b0cb241bcf972e", size = 40846822, upload-time = "2026-01-09T20:18:24.275Z" }, + { url = "https://files.pythonhosted.org/packages/53/10/ab446583dbce730000e8e6beec6ec3c2753e628c7f78f334a35cad0317f4/av-16.1.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:d69c393809babada7d54964d56099e4b30a3e1f8b5736ca5e27bd7be0e0f3c83", size = 40675604, upload-time = "2026-01-09T20:18:26.866Z" }, + { url = "https://files.pythonhosted.org/packages/31/d7/1003be685277005f6d63fd9e64904ee222fe1f7a0ea70af313468bb597db/av-16.1.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:441892be28582356d53f282873c5a951592daaf71642c7f20165e3ddcb0b4c63", size = 42015955, upload-time = "2026-01-09T20:18:29.461Z" }, + { url = "https://files.pythonhosted.org/packages/2f/4a/fa2a38ee9306bf4579f556f94ecbc757520652eb91294d2a99c7cf7623b9/av-16.1.0-cp311-cp311-win_amd64.whl", hash = "sha256:273a3e32de64819e4a1cd96341824299fe06f70c46f2288b5dc4173944f0fd62", size = 31750339, upload-time = "2026-01-09T20:18:32.249Z" }, + { url = "https://files.pythonhosted.org/packages/9c/84/2535f55edcd426cebec02eb37b811b1b0c163f26b8d3f53b059e2ec32665/av-16.1.0-cp312-cp312-macosx_11_0_x86_64.whl", hash = "sha256:640f57b93f927fba8689f6966c956737ee95388a91bd0b8c8b5e0481f73513d6", size = 26945785, upload-time = "2026-01-09T20:18:34.486Z" }, + { url = "https://files.pythonhosted.org/packages/b6/17/ffb940c9e490bf42e86db4db1ff426ee1559cd355a69609ec1efe4d3a9eb/av-16.1.0-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:ae3fb658eec00852ebd7412fdc141f17f3ddce8afee2d2e1cf366263ad2a3b35", size = 21481147, upload-time = "2026-01-09T20:18:36.716Z" }, + { url = "https://files.pythonhosted.org/packages/15/c1/e0d58003d2d83c3921887d5c8c9b8f5f7de9b58dc2194356a2656a45cfdc/av-16.1.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:27ee558d9c02a142eebcbe55578a6d817fedfde42ff5676275504e16d07a7f86", size = 39517197, upload-time = "2026-01-11T09:57:31.937Z" }, + { url = "https://files.pythonhosted.org/packages/32/77/787797b43475d1b90626af76f80bfb0c12cfec5e11eafcfc4151b8c80218/av-16.1.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:7ae547f6d5fa31763f73900d43901e8c5fa6367bb9a9840978d57b5a7ae14ed2", size = 41174337, upload-time = "2026-01-11T09:57:35.792Z" }, + { url = "https://files.pythonhosted.org/packages/8e/ac/d90df7f1e3b97fc5554cf45076df5045f1e0a6adf13899e10121229b826c/av-16.1.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:8cf065f9d438e1921dc31fc7aa045790b58aee71736897866420d80b5450f62a", size = 40817720, upload-time = "2026-01-11T09:57:39.039Z" }, + { url = "https://files.pythonhosted.org/packages/80/6f/13c3a35f9dbcebafd03fe0c4cbd075d71ac8968ec849a3cfce406c35a9d2/av-16.1.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:a345877a9d3cc0f08e2bc4ec163ee83176864b92587afb9d08dff50f37a9a829", size = 42267396, upload-time = "2026-01-11T09:57:42.115Z" }, + { url = "https://files.pythonhosted.org/packages/c8/b9/275df9607f7fb44317ccb1d4be74827185c0d410f52b6e2cd770fe209118/av-16.1.0-cp312-cp312-win_amd64.whl", hash = "sha256:f49243b1d27c91cd8c66fdba90a674e344eb8eb917264f36117bf2b6879118fd", size = 31752045, upload-time = "2026-01-11T09:57:45.106Z" }, + { url = "https://files.pythonhosted.org/packages/75/2a/63797a4dde34283dd8054219fcb29294ba1c25d68ba8c8c8a6ae53c62c45/av-16.1.0-cp313-cp313-macosx_11_0_x86_64.whl", hash = "sha256:ce2a1b3d8bf619f6c47a9f28cfa7518ff75ddd516c234a4ee351037b05e6a587", size = 26916715, upload-time = "2026-01-11T09:57:47.682Z" }, + { url = "https://files.pythonhosted.org/packages/d2/c4/0b49cf730d0ae8cda925402f18ae814aef351f5772d14da72dd87ff66448/av-16.1.0-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:408dbe6a2573ca58a855eb8cd854112b33ea598651902c36709f5f84c991ed8e", size = 21452167, upload-time = "2026-01-11T09:57:50.606Z" }, + { url = "https://files.pythonhosted.org/packages/51/23/408806503e8d5d840975aad5699b153aaa21eb6de41ade75248a79b7a37f/av-16.1.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:57f657f86652a160a8a01887aaab82282f9e629abf94c780bbdbb01595d6f0f7", size = 39215659, upload-time = "2026-01-11T09:57:53.757Z" }, + { url = "https://files.pythonhosted.org/packages/c4/19/a8528d5bba592b3903f44c28dab9cc653c95fcf7393f382d2751a1d1523e/av-16.1.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:adbad2b355c2ee4552cac59762809d791bda90586d134a33c6f13727fb86cb3a", size = 40874970, upload-time = "2026-01-11T09:57:56.802Z" }, + { url = "https://files.pythonhosted.org/packages/e8/24/2dbcdf0e929ad56b7df078e514e7bd4ca0d45cba798aff3c8caac097d2f7/av-16.1.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:f42e1a68ec2aebd21f7eb6895be69efa6aa27eec1670536876399725bbda4b99", size = 40530345, upload-time = "2026-01-11T09:58:00.421Z" }, + { url = "https://files.pythonhosted.org/packages/54/27/ae91b41207f34e99602d1c72ab6ffd9c51d7c67e3fbcd4e3a6c0e54f882c/av-16.1.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:58fe47aeaef0f100c40ec8a5de9abbd37f118d3ca03829a1009cf288e9aef67c", size = 41972163, upload-time = "2026-01-11T09:58:03.756Z" }, + { url = "https://files.pythonhosted.org/packages/fc/7a/22158fb923b2a9a00dfab0e96ef2e8a1763a94dd89e666a5858412383d46/av-16.1.0-cp313-cp313-win_amd64.whl", hash = "sha256:565093ebc93b2f4b76782589564869dadfa83af5b852edebedd8fee746457d06", size = 31729230, upload-time = "2026-01-11T09:58:07.254Z" }, + { url = "https://files.pythonhosted.org/packages/7f/f1/878f8687d801d6c4565d57ebec08449c46f75126ebca8e0fed6986599627/av-16.1.0-cp313-cp313t-macosx_11_0_x86_64.whl", hash = "sha256:574081a24edb98343fd9f473e21ae155bf61443d4ec9d7708987fa597d6b04b2", size = 27008769, upload-time = "2026-01-11T09:58:10.266Z" }, + { url = "https://files.pythonhosted.org/packages/30/f1/bd4ce8c8b5cbf1d43e27048e436cbc9de628d48ede088a1d0a993768eb86/av-16.1.0-cp313-cp313t-macosx_14_0_arm64.whl", hash = "sha256:9ab00ea29c25ebf2ea1d1e928d7babb3532d562481c5d96c0829212b70756ad0", size = 21590588, upload-time = "2026-01-11T09:58:12.629Z" }, + { url = "https://files.pythonhosted.org/packages/1d/dd/c81f6f9209201ff0b5d5bed6da6c6e641eef52d8fbc930d738c3f4f6f75d/av-16.1.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:a84a91188c1071f238a9523fd42dbe567fb2e2607b22b779851b2ce0eac1b560", size = 40638029, upload-time = "2026-01-11T09:58:15.399Z" }, + { url = "https://files.pythonhosted.org/packages/15/4d/07edff82b78d0459a6e807e01cd280d3180ce832efc1543de80d77676722/av-16.1.0-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:c2cd0de4dd022a7225ff224fde8e7971496d700be41c50adaaa26c07bb50bf97", size = 41970776, upload-time = "2026-01-11T09:58:19.075Z" }, + { url = "https://files.pythonhosted.org/packages/da/9d/1f48b354b82fa135d388477cd1b11b81bdd4384bd6a42a60808e2ec2d66b/av-16.1.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:0816143530624a5a93bc5494f8c6eeaf77549b9366709c2ac8566c1e9bff6df5", size = 41764751, upload-time = "2026-01-11T09:58:22.788Z" }, + { url = "https://files.pythonhosted.org/packages/2f/c7/a509801e98db35ec552dd79da7bdbcff7104044bfeb4c7d196c1ce121593/av-16.1.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:e3a28053af29644696d0c007e897d19b1197585834660a54773e12a40b16974c", size = 43034355, upload-time = "2026-01-11T09:58:26.125Z" }, + { url = "https://files.pythonhosted.org/packages/36/8b/e5f530d9e8f640da5f5c5f681a424c65f9dd171c871cd255d8a861785a6e/av-16.1.0-cp313-cp313t-win_amd64.whl", hash = "sha256:2e3e67144a202b95ed299d165232533989390a9ea3119d37eccec697dc6dbb0c", size = 31947047, upload-time = "2026-01-11T09:58:31.867Z" }, + { url = "https://files.pythonhosted.org/packages/df/18/8812221108c27d19f7e5f486a82c827923061edf55f906824ee0fcaadf50/av-16.1.0-cp314-cp314-macosx_11_0_x86_64.whl", hash = "sha256:39a634d8e5a87e78ea80772774bfd20c0721f0d633837ff185f36c9d14ffede4", size = 26916179, upload-time = "2026-01-11T09:58:36.506Z" }, + { url = "https://files.pythonhosted.org/packages/38/ef/49d128a9ddce42a2766fe2b6595bd9c49e067ad8937a560f7838a541464e/av-16.1.0-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:0ba32fb9e9300948a7fa9f8a3fc686e6f7f77599a665c71eb2118fdfd2c743f9", size = 21460168, upload-time = "2026-01-11T09:58:39.231Z" }, + { url = "https://files.pythonhosted.org/packages/e6/a9/b310d390844656fa74eeb8c2750e98030877c75b97551a23a77d3f982741/av-16.1.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:ca04d17815182d34ce3edc53cbda78a4f36e956c0fd73e3bab249872a831c4d7", size = 39210194, upload-time = "2026-01-11T09:58:42.138Z" }, + { url = "https://files.pythonhosted.org/packages/0c/7b/e65aae179929d0f173af6e474ad1489b5b5ad4c968a62c42758d619e54cf/av-16.1.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:ee0e8de2e124a9ef53c955fe2add6ee7c56cc8fd83318265549e44057db77142", size = 40811675, upload-time = "2026-01-11T09:58:45.871Z" }, + { url = "https://files.pythonhosted.org/packages/54/3f/5d7edefd26b6a5187d6fac0f5065ee286109934f3dea607ef05e53f05b31/av-16.1.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:22bf77a2f658827043a1e184b479c3bf25c4c43ab32353677df2d119f080e28f", size = 40543942, upload-time = "2026-01-11T09:58:49.759Z" }, + { url = "https://files.pythonhosted.org/packages/1b/24/f8b17897b67be0900a211142f5646a99d896168f54d57c81f3e018853796/av-16.1.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:2dd419d262e6a71cab206d80bbf28e0a10d0f227b671cdf5e854c028faa2d043", size = 41924336, upload-time = "2026-01-11T09:58:53.344Z" }, + { url = "https://files.pythonhosted.org/packages/1c/cf/d32bc6bbbcf60b65f6510c54690ed3ae1c4ca5d9fafbce835b6056858686/av-16.1.0-cp314-cp314-win_amd64.whl", hash = "sha256:53585986fd431cd436f290fba662cfb44d9494fbc2949a183de00acc5b33fa88", size = 31735077, upload-time = "2026-01-11T09:58:56.684Z" }, + { url = "https://files.pythonhosted.org/packages/53/f4/9b63dc70af8636399bd933e9df4f3025a0294609510239782c1b746fc796/av-16.1.0-cp314-cp314t-macosx_11_0_x86_64.whl", hash = "sha256:76f5ed8495cf41e1209a5775d3699dc63fdc1740b94a095e2485f13586593205", size = 27014423, upload-time = "2026-01-11T09:58:59.703Z" }, + { url = "https://files.pythonhosted.org/packages/d1/da/787a07a0d6ed35a0888d7e5cfb8c2ffa202f38b7ad2c657299fac08eb046/av-16.1.0-cp314-cp314t-macosx_14_0_arm64.whl", hash = "sha256:8d55397190f12a1a3ae7538be58c356cceb2bf50df1b33523817587748ce89e5", size = 21595536, upload-time = "2026-01-11T09:59:02.508Z" }, + { url = "https://files.pythonhosted.org/packages/d8/f4/9a7d8651a611be6e7e3ab7b30bb43779899c8cac5f7293b9fb634c44a3f3/av-16.1.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:9d51d9037437218261b4bbf9df78a95e216f83d7774fbfe8d289230b5b2e28e2", size = 40642490, upload-time = "2026-01-11T09:59:05.842Z" }, + { url = "https://files.pythonhosted.org/packages/6b/e4/eb79bc538a94b4ff93cd4237d00939cba797579f3272490dd0144c165a21/av-16.1.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:0ce07a89c15644407f49d942111ca046e323bbab0a9078ff43ee57c9b4a50dad", size = 41976905, upload-time = "2026-01-11T09:59:09.169Z" }, + { url = "https://files.pythonhosted.org/packages/5e/f5/f6db0dd86b70167a4d55ee0d9d9640983c570d25504f2bde42599f38241e/av-16.1.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:cac0c074892ea97113b53556ff41c99562db7b9f09f098adac1f08318c2acad5", size = 41770481, upload-time = "2026-01-11T09:59:12.74Z" }, + { url = "https://files.pythonhosted.org/packages/9e/8b/33651d658e45e16ab7671ea5fcf3d20980ea7983234f4d8d0c63c65581a5/av-16.1.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:7dec3dcbc35a187ce450f65a2e0dda820d5a9e6553eea8344a1459af11c98649", size = 43036824, upload-time = "2026-01-11T09:59:16.507Z" }, + { url = "https://files.pythonhosted.org/packages/83/41/7f13361db54d7e02f11552575c0384dadaf0918138f4eaa82ea03a9f9580/av-16.1.0-cp314-cp314t-win_amd64.whl", hash = "sha256:6f90dc082ff2068ddbe77618400b44d698d25d9c4edac57459e250c16b33d700", size = 31948164, upload-time = "2026-01-11T09:59:19.501Z" }, +] + +[[package]] +name = "blessed" +version = "1.23.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "jinxed", marker = "sys_platform == 'win32'" }, + { name = "wcwidth" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/c6/70/057c35a79a2015a6e35e45101d710b7a84b92af0c1104399bb83d33cd0d4/blessed-1.23.0.tar.gz", hash = "sha256:56591a32966f704f6131f1400af4151d9e8f5f4144133a5ca034019763dee77b", size = 6745236, upload-time = "2025-11-03T02:50:12.633Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/bc/f4/668909d1273be078ce5fb9a6d75bcfd3bf0832f238b4b09e0eb020d0eec9/blessed-1.23.0-py3-none-any.whl", hash = "sha256:4c432dcde0d45112372d1d096b2c4c0a6a5db1b94d546124872d0c3e64b5ea26", size = 95330, upload-time = "2025-11-03T02:50:10.64Z" }, +] + +[[package]] +name = "blinker" +version = "1.9.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/21/28/9b3f50ce0e048515135495f198351908d99540d69bfdc8c1d15b73dc55ce/blinker-1.9.0.tar.gz", hash = "sha256:b4ce2265a7abece45e7cc896e98dbebe6cead56bcf805a3d23136d145f5445bf", size = 22460, upload-time = "2024-11-08T17:25:47.436Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/10/cb/f2ad4230dc2eb1a74edf38f1a38b9b52277f75bef262d8908e60d957e13c/blinker-1.9.0-py3-none-any.whl", hash = "sha256:ba0efaa9080b619ff2f3459d1d500c57bddea4a6b424b60a91141db6fd2f08bc", size = 8458, upload-time = "2024-11-08T17:25:46.184Z" }, +] + +[[package]] +name = "certifi" +version = "2025.11.12" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/a2/8c/58f469717fa48465e4a50c014a0400602d3c437d7c0c468e17ada824da3a/certifi-2025.11.12.tar.gz", hash = "sha256:d8ab5478f2ecd78af242878415affce761ca6bc54a22a27e026d7c25357c3316", size = 160538, upload-time = "2025-11-12T02:54:51.517Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/70/7d/9bc192684cea499815ff478dfcdc13835ddf401365057044fb721ec6bddb/certifi-2025.11.12-py3-none-any.whl", hash = "sha256:97de8790030bbd5c2d96b7ec782fc2f7820ef8dba6db909ccf95449f2d062d4b", size = 159438, upload-time = "2025-11-12T02:54:49.735Z" }, +] + +[[package]] +name = "cfgv" +version = "3.4.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/11/74/539e56497d9bd1d484fd863dd69cbbfa653cd2aa27abfe35653494d85e94/cfgv-3.4.0.tar.gz", hash = "sha256:e52591d4c5f5dead8e0f673fb16db7949d2cfb3f7da4582893288f0ded8fe560", size = 7114, upload-time = "2023-08-12T20:38:17.776Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/c5/55/51844dd50c4fc7a33b653bfaba4c2456f06955289ca770a5dbd5fd267374/cfgv-3.4.0-py2.py3-none-any.whl", hash = "sha256:b7265b1f29fd3316bfcd2b330d63d024f2bfd8bcb8b0272f8e19a504856c48f9", size = 7249, upload-time = "2023-08-12T20:38:16.269Z" }, +] + +[[package]] +name = "charset-normalizer" +version = "3.4.4" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/13/69/33ddede1939fdd074bce5434295f38fae7136463422fe4fd3e0e89b98062/charset_normalizer-3.4.4.tar.gz", hash = "sha256:94537985111c35f28720e43603b8e7b43a6ecfb2ce1d3058bbe955b73404e21a", size = 129418, upload-time = "2025-10-14T04:42:32.879Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/1f/b8/6d51fc1d52cbd52cd4ccedd5b5b2f0f6a11bbf6765c782298b0f3e808541/charset_normalizer-3.4.4-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:e824f1492727fa856dd6eda4f7cee25f8518a12f3c4a56a74e8095695089cf6d", size = 209709, upload-time = "2025-10-14T04:40:11.385Z" }, + { url = "https://files.pythonhosted.org/packages/5c/af/1f9d7f7faafe2ddfb6f72a2e07a548a629c61ad510fe60f9630309908fef/charset_normalizer-3.4.4-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4bd5d4137d500351a30687c2d3971758aac9a19208fc110ccb9d7188fbe709e8", size = 148814, upload-time = "2025-10-14T04:40:13.135Z" }, + { url = "https://files.pythonhosted.org/packages/79/3d/f2e3ac2bbc056ca0c204298ea4e3d9db9b4afe437812638759db2c976b5f/charset_normalizer-3.4.4-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:027f6de494925c0ab2a55eab46ae5129951638a49a34d87f4c3eda90f696b4ad", size = 144467, upload-time = "2025-10-14T04:40:14.728Z" }, + { url = "https://files.pythonhosted.org/packages/ec/85/1bf997003815e60d57de7bd972c57dc6950446a3e4ccac43bc3070721856/charset_normalizer-3.4.4-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:f820802628d2694cb7e56db99213f930856014862f3fd943d290ea8438d07ca8", size = 162280, upload-time = "2025-10-14T04:40:16.14Z" }, + { url = "https://files.pythonhosted.org/packages/3e/8e/6aa1952f56b192f54921c436b87f2aaf7c7a7c3d0d1a765547d64fd83c13/charset_normalizer-3.4.4-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:798d75d81754988d2565bff1b97ba5a44411867c0cf32b77a7e8f8d84796b10d", size = 159454, upload-time = "2025-10-14T04:40:17.567Z" }, + { url = "https://files.pythonhosted.org/packages/36/3b/60cbd1f8e93aa25d1c669c649b7a655b0b5fb4c571858910ea9332678558/charset_normalizer-3.4.4-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9d1bb833febdff5c8927f922386db610b49db6e0d4f4ee29601d71e7c2694313", size = 153609, upload-time = "2025-10-14T04:40:19.08Z" }, + { url = "https://files.pythonhosted.org/packages/64/91/6a13396948b8fd3c4b4fd5bc74d045f5637d78c9675585e8e9fbe5636554/charset_normalizer-3.4.4-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:9cd98cdc06614a2f768d2b7286d66805f94c48cde050acdbbb7db2600ab3197e", size = 151849, upload-time = "2025-10-14T04:40:20.607Z" }, + { url = "https://files.pythonhosted.org/packages/b7/7a/59482e28b9981d105691e968c544cc0df3b7d6133152fb3dcdc8f135da7a/charset_normalizer-3.4.4-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:077fbb858e903c73f6c9db43374fd213b0b6a778106bc7032446a8e8b5b38b93", size = 151586, upload-time = "2025-10-14T04:40:21.719Z" }, + { url = "https://files.pythonhosted.org/packages/92/59/f64ef6a1c4bdd2baf892b04cd78792ed8684fbc48d4c2afe467d96b4df57/charset_normalizer-3.4.4-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:244bfb999c71b35de57821b8ea746b24e863398194a4014e4c76adc2bbdfeff0", size = 145290, upload-time = "2025-10-14T04:40:23.069Z" }, + { url = "https://files.pythonhosted.org/packages/6b/63/3bf9f279ddfa641ffa1962b0db6a57a9c294361cc2f5fcac997049a00e9c/charset_normalizer-3.4.4-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:64b55f9dce520635f018f907ff1b0df1fdc31f2795a922fb49dd14fbcdf48c84", size = 163663, upload-time = "2025-10-14T04:40:24.17Z" }, + { url = "https://files.pythonhosted.org/packages/ed/09/c9e38fc8fa9e0849b172b581fd9803bdf6e694041127933934184e19f8c3/charset_normalizer-3.4.4-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:faa3a41b2b66b6e50f84ae4a68c64fcd0c44355741c6374813a800cd6695db9e", size = 151964, upload-time = "2025-10-14T04:40:25.368Z" }, + { url = "https://files.pythonhosted.org/packages/d2/d1/d28b747e512d0da79d8b6a1ac18b7ab2ecfd81b2944c4c710e166d8dd09c/charset_normalizer-3.4.4-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:6515f3182dbe4ea06ced2d9e8666d97b46ef4c75e326b79bb624110f122551db", size = 161064, upload-time = "2025-10-14T04:40:26.806Z" }, + { url = "https://files.pythonhosted.org/packages/bb/9a/31d62b611d901c3b9e5500c36aab0ff5eb442043fb3a1c254200d3d397d9/charset_normalizer-3.4.4-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:cc00f04ed596e9dc0da42ed17ac5e596c6ccba999ba6bd92b0e0aef2f170f2d6", size = 155015, upload-time = "2025-10-14T04:40:28.284Z" }, + { url = "https://files.pythonhosted.org/packages/1f/f3/107e008fa2bff0c8b9319584174418e5e5285fef32f79d8ee6a430d0039c/charset_normalizer-3.4.4-cp310-cp310-win32.whl", hash = "sha256:f34be2938726fc13801220747472850852fe6b1ea75869a048d6f896838c896f", size = 99792, upload-time = "2025-10-14T04:40:29.613Z" }, + { url = "https://files.pythonhosted.org/packages/eb/66/e396e8a408843337d7315bab30dbf106c38966f1819f123257f5520f8a96/charset_normalizer-3.4.4-cp310-cp310-win_amd64.whl", hash = "sha256:a61900df84c667873b292c3de315a786dd8dac506704dea57bc957bd31e22c7d", size = 107198, upload-time = "2025-10-14T04:40:30.644Z" }, + { url = "https://files.pythonhosted.org/packages/b5/58/01b4f815bf0312704c267f2ccb6e5d42bcc7752340cd487bc9f8c3710597/charset_normalizer-3.4.4-cp310-cp310-win_arm64.whl", hash = "sha256:cead0978fc57397645f12578bfd2d5ea9138ea0fac82b2f63f7f7c6877986a69", size = 100262, upload-time = "2025-10-14T04:40:32.108Z" }, + { url = "https://files.pythonhosted.org/packages/ed/27/c6491ff4954e58a10f69ad90aca8a1b6fe9c5d3c6f380907af3c37435b59/charset_normalizer-3.4.4-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:6e1fcf0720908f200cd21aa4e6750a48ff6ce4afe7ff5a79a90d5ed8a08296f8", size = 206988, upload-time = "2025-10-14T04:40:33.79Z" }, + { url = "https://files.pythonhosted.org/packages/94/59/2e87300fe67ab820b5428580a53cad894272dbb97f38a7a814a2a1ac1011/charset_normalizer-3.4.4-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5f819d5fe9234f9f82d75bdfa9aef3a3d72c4d24a6e57aeaebba32a704553aa0", size = 147324, upload-time = "2025-10-14T04:40:34.961Z" }, + { url = "https://files.pythonhosted.org/packages/07/fb/0cf61dc84b2b088391830f6274cb57c82e4da8bbc2efeac8c025edb88772/charset_normalizer-3.4.4-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:a59cb51917aa591b1c4e6a43c132f0cdc3c76dbad6155df4e28ee626cc77a0a3", size = 142742, upload-time = "2025-10-14T04:40:36.105Z" }, + { url = "https://files.pythonhosted.org/packages/62/8b/171935adf2312cd745d290ed93cf16cf0dfe320863ab7cbeeae1dcd6535f/charset_normalizer-3.4.4-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8ef3c867360f88ac904fd3f5e1f902f13307af9052646963ee08ff4f131adafc", size = 160863, upload-time = "2025-10-14T04:40:37.188Z" }, + { url = "https://files.pythonhosted.org/packages/09/73/ad875b192bda14f2173bfc1bc9a55e009808484a4b256748d931b6948442/charset_normalizer-3.4.4-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d9e45d7faa48ee908174d8fe84854479ef838fc6a705c9315372eacbc2f02897", size = 157837, upload-time = "2025-10-14T04:40:38.435Z" }, + { url = "https://files.pythonhosted.org/packages/6d/fc/de9cce525b2c5b94b47c70a4b4fb19f871b24995c728e957ee68ab1671ea/charset_normalizer-3.4.4-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:840c25fb618a231545cbab0564a799f101b63b9901f2569faecd6b222ac72381", size = 151550, upload-time = "2025-10-14T04:40:40.053Z" }, + { url = "https://files.pythonhosted.org/packages/55/c2/43edd615fdfba8c6f2dfbd459b25a6b3b551f24ea21981e23fb768503ce1/charset_normalizer-3.4.4-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:ca5862d5b3928c4940729dacc329aa9102900382fea192fc5e52eb69d6093815", size = 149162, upload-time = "2025-10-14T04:40:41.163Z" }, + { url = "https://files.pythonhosted.org/packages/03/86/bde4ad8b4d0e9429a4e82c1e8f5c659993a9a863ad62c7df05cf7b678d75/charset_normalizer-3.4.4-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:d9c7f57c3d666a53421049053eaacdd14bbd0a528e2186fcb2e672effd053bb0", size = 150019, upload-time = "2025-10-14T04:40:42.276Z" }, + { url = "https://files.pythonhosted.org/packages/1f/86/a151eb2af293a7e7bac3a739b81072585ce36ccfb4493039f49f1d3cae8c/charset_normalizer-3.4.4-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:277e970e750505ed74c832b4bf75dac7476262ee2a013f5574dd49075879e161", size = 143310, upload-time = "2025-10-14T04:40:43.439Z" }, + { url = "https://files.pythonhosted.org/packages/b5/fe/43dae6144a7e07b87478fdfc4dbe9efd5defb0e7ec29f5f58a55aeef7bf7/charset_normalizer-3.4.4-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:31fd66405eaf47bb62e8cd575dc621c56c668f27d46a61d975a249930dd5e2a4", size = 162022, upload-time = "2025-10-14T04:40:44.547Z" }, + { url = "https://files.pythonhosted.org/packages/80/e6/7aab83774f5d2bca81f42ac58d04caf44f0cc2b65fc6db2b3b2e8a05f3b3/charset_normalizer-3.4.4-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:0d3d8f15c07f86e9ff82319b3d9ef6f4bf907608f53fe9d92b28ea9ae3d1fd89", size = 149383, upload-time = "2025-10-14T04:40:46.018Z" }, + { url = "https://files.pythonhosted.org/packages/4f/e8/b289173b4edae05c0dde07f69f8db476a0b511eac556dfe0d6bda3c43384/charset_normalizer-3.4.4-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:9f7fcd74d410a36883701fafa2482a6af2ff5ba96b9a620e9e0721e28ead5569", size = 159098, upload-time = "2025-10-14T04:40:47.081Z" }, + { url = "https://files.pythonhosted.org/packages/d8/df/fe699727754cae3f8478493c7f45f777b17c3ef0600e28abfec8619eb49c/charset_normalizer-3.4.4-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:ebf3e58c7ec8a8bed6d66a75d7fb37b55e5015b03ceae72a8e7c74495551e224", size = 152991, upload-time = "2025-10-14T04:40:48.246Z" }, + { url = "https://files.pythonhosted.org/packages/1a/86/584869fe4ddb6ffa3bd9f491b87a01568797fb9bd8933f557dba9771beaf/charset_normalizer-3.4.4-cp311-cp311-win32.whl", hash = "sha256:eecbc200c7fd5ddb9a7f16c7decb07b566c29fa2161a16cf67b8d068bd21690a", size = 99456, upload-time = "2025-10-14T04:40:49.376Z" }, + { url = "https://files.pythonhosted.org/packages/65/f6/62fdd5feb60530f50f7e38b4f6a1d5203f4d16ff4f9f0952962c044e919a/charset_normalizer-3.4.4-cp311-cp311-win_amd64.whl", hash = "sha256:5ae497466c7901d54b639cf42d5b8c1b6a4fead55215500d2f486d34db48d016", size = 106978, upload-time = "2025-10-14T04:40:50.844Z" }, + { url = "https://files.pythonhosted.org/packages/7a/9d/0710916e6c82948b3be62d9d398cb4fcf4e97b56d6a6aeccd66c4b2f2bd5/charset_normalizer-3.4.4-cp311-cp311-win_arm64.whl", hash = "sha256:65e2befcd84bc6f37095f5961e68a6f077bf44946771354a28ad434c2cce0ae1", size = 99969, upload-time = "2025-10-14T04:40:52.272Z" }, + { url = "https://files.pythonhosted.org/packages/f3/85/1637cd4af66fa687396e757dec650f28025f2a2f5a5531a3208dc0ec43f2/charset_normalizer-3.4.4-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:0a98e6759f854bd25a58a73fa88833fba3b7c491169f86ce1180c948ab3fd394", size = 208425, upload-time = "2025-10-14T04:40:53.353Z" }, + { url = "https://files.pythonhosted.org/packages/9d/6a/04130023fef2a0d9c62d0bae2649b69f7b7d8d24ea5536feef50551029df/charset_normalizer-3.4.4-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b5b290ccc2a263e8d185130284f8501e3e36c5e02750fc6b6bdeb2e9e96f1e25", size = 148162, upload-time = "2025-10-14T04:40:54.558Z" }, + { url = "https://files.pythonhosted.org/packages/78/29/62328d79aa60da22c9e0b9a66539feae06ca0f5a4171ac4f7dc285b83688/charset_normalizer-3.4.4-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:74bb723680f9f7a6234dcf67aea57e708ec1fbdf5699fb91dfd6f511b0a320ef", size = 144558, upload-time = "2025-10-14T04:40:55.677Z" }, + { url = "https://files.pythonhosted.org/packages/86/bb/b32194a4bf15b88403537c2e120b817c61cd4ecffa9b6876e941c3ee38fe/charset_normalizer-3.4.4-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:f1e34719c6ed0b92f418c7c780480b26b5d9c50349e9a9af7d76bf757530350d", size = 161497, upload-time = "2025-10-14T04:40:57.217Z" }, + { url = "https://files.pythonhosted.org/packages/19/89/a54c82b253d5b9b111dc74aca196ba5ccfcca8242d0fb64146d4d3183ff1/charset_normalizer-3.4.4-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2437418e20515acec67d86e12bf70056a33abdacb5cb1655042f6538d6b085a8", size = 159240, upload-time = "2025-10-14T04:40:58.358Z" }, + { url = "https://files.pythonhosted.org/packages/c0/10/d20b513afe03acc89ec33948320a5544d31f21b05368436d580dec4e234d/charset_normalizer-3.4.4-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:11d694519d7f29d6cd09f6ac70028dba10f92f6cdd059096db198c283794ac86", size = 153471, upload-time = "2025-10-14T04:40:59.468Z" }, + { url = "https://files.pythonhosted.org/packages/61/fa/fbf177b55bdd727010f9c0a3c49eefa1d10f960e5f09d1d887bf93c2e698/charset_normalizer-3.4.4-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:ac1c4a689edcc530fc9d9aa11f5774b9e2f33f9a0c6a57864e90908f5208d30a", size = 150864, upload-time = "2025-10-14T04:41:00.623Z" }, + { url = "https://files.pythonhosted.org/packages/05/12/9fbc6a4d39c0198adeebbde20b619790e9236557ca59fc40e0e3cebe6f40/charset_normalizer-3.4.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:21d142cc6c0ec30d2efee5068ca36c128a30b0f2c53c1c07bd78cb6bc1d3be5f", size = 150647, upload-time = "2025-10-14T04:41:01.754Z" }, + { url = "https://files.pythonhosted.org/packages/ad/1f/6a9a593d52e3e8c5d2b167daf8c6b968808efb57ef4c210acb907c365bc4/charset_normalizer-3.4.4-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:5dbe56a36425d26d6cfb40ce79c314a2e4dd6211d51d6d2191c00bed34f354cc", size = 145110, upload-time = "2025-10-14T04:41:03.231Z" }, + { url = "https://files.pythonhosted.org/packages/30/42/9a52c609e72471b0fc54386dc63c3781a387bb4fe61c20231a4ebcd58bdd/charset_normalizer-3.4.4-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:5bfbb1b9acf3334612667b61bd3002196fe2a1eb4dd74d247e0f2a4d50ec9bbf", size = 162839, upload-time = "2025-10-14T04:41:04.715Z" }, + { url = "https://files.pythonhosted.org/packages/c4/5b/c0682bbf9f11597073052628ddd38344a3d673fda35a36773f7d19344b23/charset_normalizer-3.4.4-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:d055ec1e26e441f6187acf818b73564e6e6282709e9bcb5b63f5b23068356a15", size = 150667, upload-time = "2025-10-14T04:41:05.827Z" }, + { url = "https://files.pythonhosted.org/packages/e4/24/a41afeab6f990cf2daf6cb8c67419b63b48cf518e4f56022230840c9bfb2/charset_normalizer-3.4.4-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:af2d8c67d8e573d6de5bc30cdb27e9b95e49115cd9baad5ddbd1a6207aaa82a9", size = 160535, upload-time = "2025-10-14T04:41:06.938Z" }, + { url = "https://files.pythonhosted.org/packages/2a/e5/6a4ce77ed243c4a50a1fecca6aaaab419628c818a49434be428fe24c9957/charset_normalizer-3.4.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:780236ac706e66881f3b7f2f32dfe90507a09e67d1d454c762cf642e6e1586e0", size = 154816, upload-time = "2025-10-14T04:41:08.101Z" }, + { url = "https://files.pythonhosted.org/packages/a8/ef/89297262b8092b312d29cdb2517cb1237e51db8ecef2e9af5edbe7b683b1/charset_normalizer-3.4.4-cp312-cp312-win32.whl", hash = "sha256:5833d2c39d8896e4e19b689ffc198f08ea58116bee26dea51e362ecc7cd3ed26", size = 99694, upload-time = "2025-10-14T04:41:09.23Z" }, + { url = "https://files.pythonhosted.org/packages/3d/2d/1e5ed9dd3b3803994c155cd9aacb60c82c331bad84daf75bcb9c91b3295e/charset_normalizer-3.4.4-cp312-cp312-win_amd64.whl", hash = "sha256:a79cfe37875f822425b89a82333404539ae63dbdddf97f84dcbc3d339aae9525", size = 107131, upload-time = "2025-10-14T04:41:10.467Z" }, + { url = "https://files.pythonhosted.org/packages/d0/d9/0ed4c7098a861482a7b6a95603edce4c0d9db2311af23da1fb2b75ec26fc/charset_normalizer-3.4.4-cp312-cp312-win_arm64.whl", hash = "sha256:376bec83a63b8021bb5c8ea75e21c4ccb86e7e45ca4eb81146091b56599b80c3", size = 100390, upload-time = "2025-10-14T04:41:11.915Z" }, + { url = "https://files.pythonhosted.org/packages/97/45/4b3a1239bbacd321068ea6e7ac28875b03ab8bc0aa0966452db17cd36714/charset_normalizer-3.4.4-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:e1f185f86a6f3403aa2420e815904c67b2f9ebc443f045edd0de921108345794", size = 208091, upload-time = "2025-10-14T04:41:13.346Z" }, + { url = "https://files.pythonhosted.org/packages/7d/62/73a6d7450829655a35bb88a88fca7d736f9882a27eacdca2c6d505b57e2e/charset_normalizer-3.4.4-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6b39f987ae8ccdf0d2642338faf2abb1862340facc796048b604ef14919e55ed", size = 147936, upload-time = "2025-10-14T04:41:14.461Z" }, + { url = "https://files.pythonhosted.org/packages/89/c5/adb8c8b3d6625bef6d88b251bbb0d95f8205831b987631ab0c8bb5d937c2/charset_normalizer-3.4.4-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3162d5d8ce1bb98dd51af660f2121c55d0fa541b46dff7bb9b9f86ea1d87de72", size = 144180, upload-time = "2025-10-14T04:41:15.588Z" }, + { url = "https://files.pythonhosted.org/packages/91/ed/9706e4070682d1cc219050b6048bfd293ccf67b3d4f5a4f39207453d4b99/charset_normalizer-3.4.4-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:81d5eb2a312700f4ecaa977a8235b634ce853200e828fbadf3a9c50bab278328", size = 161346, upload-time = "2025-10-14T04:41:16.738Z" }, + { url = "https://files.pythonhosted.org/packages/d5/0d/031f0d95e4972901a2f6f09ef055751805ff541511dc1252ba3ca1f80cf5/charset_normalizer-3.4.4-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:5bd2293095d766545ec1a8f612559f6b40abc0eb18bb2f5d1171872d34036ede", size = 158874, upload-time = "2025-10-14T04:41:17.923Z" }, + { url = "https://files.pythonhosted.org/packages/f5/83/6ab5883f57c9c801ce5e5677242328aa45592be8a00644310a008d04f922/charset_normalizer-3.4.4-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a8a8b89589086a25749f471e6a900d3f662d1d3b6e2e59dcecf787b1cc3a1894", size = 153076, upload-time = "2025-10-14T04:41:19.106Z" }, + { url = "https://files.pythonhosted.org/packages/75/1e/5ff781ddf5260e387d6419959ee89ef13878229732732ee73cdae01800f2/charset_normalizer-3.4.4-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:bc7637e2f80d8530ee4a78e878bce464f70087ce73cf7c1caf142416923b98f1", size = 150601, upload-time = "2025-10-14T04:41:20.245Z" }, + { url = "https://files.pythonhosted.org/packages/d7/57/71be810965493d3510a6ca79b90c19e48696fb1ff964da319334b12677f0/charset_normalizer-3.4.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:f8bf04158c6b607d747e93949aa60618b61312fe647a6369f88ce2ff16043490", size = 150376, upload-time = "2025-10-14T04:41:21.398Z" }, + { url = "https://files.pythonhosted.org/packages/e5/d5/c3d057a78c181d007014feb7e9f2e65905a6c4ef182c0ddf0de2924edd65/charset_normalizer-3.4.4-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:554af85e960429cf30784dd47447d5125aaa3b99a6f0683589dbd27e2f45da44", size = 144825, upload-time = "2025-10-14T04:41:22.583Z" }, + { url = "https://files.pythonhosted.org/packages/e6/8c/d0406294828d4976f275ffbe66f00266c4b3136b7506941d87c00cab5272/charset_normalizer-3.4.4-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:74018750915ee7ad843a774364e13a3db91682f26142baddf775342c3f5b1133", size = 162583, upload-time = "2025-10-14T04:41:23.754Z" }, + { url = "https://files.pythonhosted.org/packages/d7/24/e2aa1f18c8f15c4c0e932d9287b8609dd30ad56dbe41d926bd846e22fb8d/charset_normalizer-3.4.4-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:c0463276121fdee9c49b98908b3a89c39be45d86d1dbaa22957e38f6321d4ce3", size = 150366, upload-time = "2025-10-14T04:41:25.27Z" }, + { url = "https://files.pythonhosted.org/packages/e4/5b/1e6160c7739aad1e2df054300cc618b06bf784a7a164b0f238360721ab86/charset_normalizer-3.4.4-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:362d61fd13843997c1c446760ef36f240cf81d3ebf74ac62652aebaf7838561e", size = 160300, upload-time = "2025-10-14T04:41:26.725Z" }, + { url = "https://files.pythonhosted.org/packages/7a/10/f882167cd207fbdd743e55534d5d9620e095089d176d55cb22d5322f2afd/charset_normalizer-3.4.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:9a26f18905b8dd5d685d6d07b0cdf98a79f3c7a918906af7cc143ea2e164c8bc", size = 154465, upload-time = "2025-10-14T04:41:28.322Z" }, + { url = "https://files.pythonhosted.org/packages/89/66/c7a9e1b7429be72123441bfdbaf2bc13faab3f90b933f664db506dea5915/charset_normalizer-3.4.4-cp313-cp313-win32.whl", hash = "sha256:9b35f4c90079ff2e2edc5b26c0c77925e5d2d255c42c74fdb70fb49b172726ac", size = 99404, upload-time = "2025-10-14T04:41:29.95Z" }, + { url = "https://files.pythonhosted.org/packages/c4/26/b9924fa27db384bdcd97ab83b4f0a8058d96ad9626ead570674d5e737d90/charset_normalizer-3.4.4-cp313-cp313-win_amd64.whl", hash = "sha256:b435cba5f4f750aa6c0a0d92c541fb79f69a387c91e61f1795227e4ed9cece14", size = 107092, upload-time = "2025-10-14T04:41:31.188Z" }, + { url = "https://files.pythonhosted.org/packages/af/8f/3ed4bfa0c0c72a7ca17f0380cd9e4dd842b09f664e780c13cff1dcf2ef1b/charset_normalizer-3.4.4-cp313-cp313-win_arm64.whl", hash = "sha256:542d2cee80be6f80247095cc36c418f7bddd14f4a6de45af91dfad36d817bba2", size = 100408, upload-time = "2025-10-14T04:41:32.624Z" }, + { url = "https://files.pythonhosted.org/packages/2a/35/7051599bd493e62411d6ede36fd5af83a38f37c4767b92884df7301db25d/charset_normalizer-3.4.4-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:da3326d9e65ef63a817ecbcc0df6e94463713b754fe293eaa03da99befb9a5bd", size = 207746, upload-time = "2025-10-14T04:41:33.773Z" }, + { url = "https://files.pythonhosted.org/packages/10/9a/97c8d48ef10d6cd4fcead2415523221624bf58bcf68a802721a6bc807c8f/charset_normalizer-3.4.4-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8af65f14dc14a79b924524b1e7fffe304517b2bff5a58bf64f30b98bbc5079eb", size = 147889, upload-time = "2025-10-14T04:41:34.897Z" }, + { url = "https://files.pythonhosted.org/packages/10/bf/979224a919a1b606c82bd2c5fa49b5c6d5727aa47b4312bb27b1734f53cd/charset_normalizer-3.4.4-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:74664978bb272435107de04e36db5a9735e78232b85b77d45cfb38f758efd33e", size = 143641, upload-time = "2025-10-14T04:41:36.116Z" }, + { url = "https://files.pythonhosted.org/packages/ba/33/0ad65587441fc730dc7bd90e9716b30b4702dc7b617e6ba4997dc8651495/charset_normalizer-3.4.4-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:752944c7ffbfdd10c074dc58ec2d5a8a4cd9493b314d367c14d24c17684ddd14", size = 160779, upload-time = "2025-10-14T04:41:37.229Z" }, + { url = "https://files.pythonhosted.org/packages/67/ed/331d6b249259ee71ddea93f6f2f0a56cfebd46938bde6fcc6f7b9a3d0e09/charset_normalizer-3.4.4-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d1f13550535ad8cff21b8d757a3257963e951d96e20ec82ab44bc64aeb62a191", size = 159035, upload-time = "2025-10-14T04:41:38.368Z" }, + { url = "https://files.pythonhosted.org/packages/67/ff/f6b948ca32e4f2a4576aa129d8bed61f2e0543bf9f5f2b7fc3758ed005c9/charset_normalizer-3.4.4-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ecaae4149d99b1c9e7b88bb03e3221956f68fd6d50be2ef061b2381b61d20838", size = 152542, upload-time = "2025-10-14T04:41:39.862Z" }, + { url = "https://files.pythonhosted.org/packages/16/85/276033dcbcc369eb176594de22728541a925b2632f9716428c851b149e83/charset_normalizer-3.4.4-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:cb6254dc36b47a990e59e1068afacdcd02958bdcce30bb50cc1700a8b9d624a6", size = 149524, upload-time = "2025-10-14T04:41:41.319Z" }, + { url = "https://files.pythonhosted.org/packages/9e/f2/6a2a1f722b6aba37050e626530a46a68f74e63683947a8acff92569f979a/charset_normalizer-3.4.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:c8ae8a0f02f57a6e61203a31428fa1d677cbe50c93622b4149d5c0f319c1d19e", size = 150395, upload-time = "2025-10-14T04:41:42.539Z" }, + { url = "https://files.pythonhosted.org/packages/60/bb/2186cb2f2bbaea6338cad15ce23a67f9b0672929744381e28b0592676824/charset_normalizer-3.4.4-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:47cc91b2f4dd2833fddaedd2893006b0106129d4b94fdb6af1f4ce5a9965577c", size = 143680, upload-time = "2025-10-14T04:41:43.661Z" }, + { url = "https://files.pythonhosted.org/packages/7d/a5/bf6f13b772fbb2a90360eb620d52ed8f796f3c5caee8398c3b2eb7b1c60d/charset_normalizer-3.4.4-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:82004af6c302b5d3ab2cfc4cc5f29db16123b1a8417f2e25f9066f91d4411090", size = 162045, upload-time = "2025-10-14T04:41:44.821Z" }, + { url = "https://files.pythonhosted.org/packages/df/c5/d1be898bf0dc3ef9030c3825e5d3b83f2c528d207d246cbabe245966808d/charset_normalizer-3.4.4-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:2b7d8f6c26245217bd2ad053761201e9f9680f8ce52f0fcd8d0755aeae5b2152", size = 149687, upload-time = "2025-10-14T04:41:46.442Z" }, + { url = "https://files.pythonhosted.org/packages/a5/42/90c1f7b9341eef50c8a1cb3f098ac43b0508413f33affd762855f67a410e/charset_normalizer-3.4.4-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:799a7a5e4fb2d5898c60b640fd4981d6a25f1c11790935a44ce38c54e985f828", size = 160014, upload-time = "2025-10-14T04:41:47.631Z" }, + { url = "https://files.pythonhosted.org/packages/76/be/4d3ee471e8145d12795ab655ece37baed0929462a86e72372fd25859047c/charset_normalizer-3.4.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:99ae2cffebb06e6c22bdc25801d7b30f503cc87dbd283479e7b606f70aff57ec", size = 154044, upload-time = "2025-10-14T04:41:48.81Z" }, + { url = "https://files.pythonhosted.org/packages/b0/6f/8f7af07237c34a1defe7defc565a9bc1807762f672c0fde711a4b22bf9c0/charset_normalizer-3.4.4-cp314-cp314-win32.whl", hash = "sha256:f9d332f8c2a2fcbffe1378594431458ddbef721c1769d78e2cbc06280d8155f9", size = 99940, upload-time = "2025-10-14T04:41:49.946Z" }, + { url = "https://files.pythonhosted.org/packages/4b/51/8ade005e5ca5b0d80fb4aff72a3775b325bdc3d27408c8113811a7cbe640/charset_normalizer-3.4.4-cp314-cp314-win_amd64.whl", hash = "sha256:8a6562c3700cce886c5be75ade4a5db4214fda19fede41d9792d100288d8f94c", size = 107104, upload-time = "2025-10-14T04:41:51.051Z" }, + { url = "https://files.pythonhosted.org/packages/da/5f/6b8f83a55bb8278772c5ae54a577f3099025f9ade59d0136ac24a0df4bde/charset_normalizer-3.4.4-cp314-cp314-win_arm64.whl", hash = "sha256:de00632ca48df9daf77a2c65a484531649261ec9f25489917f09e455cb09ddb2", size = 100743, upload-time = "2025-10-14T04:41:52.122Z" }, + { url = "https://files.pythonhosted.org/packages/0a/4c/925909008ed5a988ccbb72dcc897407e5d6d3bd72410d69e051fc0c14647/charset_normalizer-3.4.4-py3-none-any.whl", hash = "sha256:7a32c560861a02ff789ad905a2fe94e3f840803362c84fecf1851cb4cf3dc37f", size = 53402, upload-time = "2025-10-14T04:42:31.76Z" }, +] + +[[package]] +name = "click" +version = "8.3.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "colorama", marker = "sys_platform == 'win32'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/46/61/de6cd827efad202d7057d93e0fed9294b96952e188f7384832791c7b2254/click-8.3.0.tar.gz", hash = "sha256:e7b8232224eba16f4ebe410c25ced9f7875cb5f3263ffc93cc3e8da705e229c4", size = 276943, upload-time = "2025-09-18T17:32:23.696Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/db/d3/9dcc0f5797f070ec8edf30fbadfb200e71d9db6b84d211e3b2085a7589a0/click-8.3.0-py3-none-any.whl", hash = "sha256:9b9f285302c6e3064f4330c05f05b81945b2a39544279343e6e7c5f27a9baddc", size = 107295, upload-time = "2025-09-18T17:32:22.42Z" }, +] + +[[package]] +name = "cloudpickle" +version = "3.1.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/27/fb/576f067976d320f5f0114a8d9fa1215425441bb35627b1993e5afd8111e5/cloudpickle-3.1.2.tar.gz", hash = "sha256:7fda9eb655c9c230dab534f1983763de5835249750e85fbcef43aaa30a9a2414", size = 22330, upload-time = "2025-11-03T09:25:26.604Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/88/39/799be3f2f0f38cc727ee3b4f1445fe6d5e4133064ec2e4115069418a5bb6/cloudpickle-3.1.2-py3-none-any.whl", hash = "sha256:9acb47f6afd73f60dc1df93bb801b472f05ff42fa6c84167d25cb206be1fbf4a", size = 22228, upload-time = "2025-11-03T09:25:25.534Z" }, +] + +[[package]] +name = "colorama" +version = "0.4.6" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/d8/53/6f443c9a4a8358a93a6792e2acffb9d9d5cb0a5cfd8802644b7b1c9a02e4/colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44", size = 27697, upload-time = "2022-10-25T02:36:22.414Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/d1/d6/3965ed04c63042e047cb6a3e6ed1a63a35087b6a609aa3a15ed8ac56c221/colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6", size = 25335, upload-time = "2022-10-25T02:36:20.889Z" }, +] + +[[package]] +name = "cuda-bindings" +version = "12.9.4" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "cuda-pathfinder", marker = "sys_platform == 'linux'" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/37/31/bfcc870f69c6a017c4ad5c42316207fc7551940db6f3639aa4466ec5faf3/cuda_bindings-12.9.4-cp310-cp310-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a022c96b8bd847e8dc0675523431149a4c3e872f440e3002213dbb9e08f0331a", size = 11800959, upload-time = "2025-10-21T14:51:26.458Z" }, + { url = "https://files.pythonhosted.org/packages/7a/d8/b546104b8da3f562c1ff8ab36d130c8fe1dd6a045ced80b4f6ad74f7d4e1/cuda_bindings-12.9.4-cp310-cp310-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4d3c842c2a4303b2a580fe955018e31aea30278be19795ae05226235268032e5", size = 12148218, upload-time = "2025-10-21T14:51:28.855Z" }, + { url = "https://files.pythonhosted.org/packages/a9/2b/ebcbb60aa6dba830474cd360c42e10282f7a343c0a1f58d24fbd3b7c2d77/cuda_bindings-12.9.4-cp311-cp311-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a6a429dc6c13148ff1e27c44f40a3dd23203823e637b87fd0854205195988306", size = 11840604, upload-time = "2025-10-21T14:51:34.565Z" }, + { url = "https://files.pythonhosted.org/packages/45/e7/b47792cc2d01c7e1d37c32402182524774dadd2d26339bd224e0e913832e/cuda_bindings-12.9.4-cp311-cp311-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c912a3d9e6b6651853eed8eed96d6800d69c08e94052c292fec3f282c5a817c9", size = 12210593, upload-time = "2025-10-21T14:51:36.574Z" }, + { url = "https://files.pythonhosted.org/packages/0c/c2/65bfd79292b8ff18be4dd7f7442cea37bcbc1a228c1886f1dea515c45b67/cuda_bindings-12.9.4-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:694ba35023846625ef471257e6b5a4bc8af690f961d197d77d34b1d1db393f56", size = 11760260, upload-time = "2025-10-21T14:51:40.79Z" }, + { url = "https://files.pythonhosted.org/packages/a9/c1/dabe88f52c3e3760d861401bb994df08f672ec893b8f7592dc91626adcf3/cuda_bindings-12.9.4-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fda147a344e8eaeca0c6ff113d2851ffca8f7dfc0a6c932374ee5c47caa649c8", size = 12151019, upload-time = "2025-10-21T14:51:43.167Z" }, + { url = "https://files.pythonhosted.org/packages/05/8b/b4b2d1c7775fa403b64333e720cfcfccef8dcb9cdeb99947061ca5a77628/cuda_bindings-12.9.4-cp313-cp313-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:cf8bfaedc238f3b115d957d1fd6562b7e8435ba57f6d0e2f87d0e7149ccb2da5", size = 11570071, upload-time = "2025-10-21T14:51:47.472Z" }, + { url = "https://files.pythonhosted.org/packages/63/56/e465c31dc9111be3441a9ba7df1941fe98f4aa6e71e8788a3fb4534ce24d/cuda_bindings-12.9.4-cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:32bdc5a76906be4c61eb98f546a6786c5773a881f3b166486449b5d141e4a39f", size = 11906628, upload-time = "2025-10-21T14:51:49.905Z" }, + { url = "https://files.pythonhosted.org/packages/ec/07/6aff13bc1e977e35aaa6b22f52b172e2890c608c6db22438cf7ed2bf43a6/cuda_bindings-12.9.4-cp313-cp313t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3adf4958dcf68ae7801a59b73fb00a8b37f8d0595060d66ceae111b1002de38d", size = 11566797, upload-time = "2025-10-21T14:51:54.581Z" }, + { url = "https://files.pythonhosted.org/packages/a3/84/1e6be415e37478070aeeee5884c2022713c1ecc735e6d82d744de0252eee/cuda_bindings-12.9.4-cp313-cp313t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:56e0043c457a99ac473ddc926fe0dc4046694d99caef633e92601ab52cbe17eb", size = 11925991, upload-time = "2025-10-21T14:51:56.535Z" }, + { url = "https://files.pythonhosted.org/packages/1e/b5/96a6696e20c4ffd2b327f54c7d0fde2259bdb998d045c25d5dedbbe30290/cuda_bindings-12.9.4-cp314-cp314-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1f53a7f453d4b2643d8663d036bafe29b5ba89eb904c133180f295df6dc151e5", size = 11624530, upload-time = "2025-10-21T14:52:01.539Z" }, + { url = "https://files.pythonhosted.org/packages/d1/af/6dfd8f2ed90b1d4719bc053ff8940e494640fe4212dc3dd72f383e4992da/cuda_bindings-12.9.4-cp314-cp314-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8b72ee72a9cc1b531db31eebaaee5c69a8ec3500e32c6933f2d3b15297b53686", size = 11922703, upload-time = "2025-10-21T14:52:03.585Z" }, + { url = "https://files.pythonhosted.org/packages/39/73/d2fc40c043bac699c3880bf88d3cebe9d88410cd043795382826c93a89f0/cuda_bindings-12.9.4-cp314-cp314t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:20f2699d61d724de3eb3f3369d57e2b245f93085cab44fd37c3bea036cea1a6f", size = 11565056, upload-time = "2025-10-21T14:52:08.338Z" }, + { url = "https://files.pythonhosted.org/packages/6c/19/90ac264acc00f6df8a49378eedec9fd2db3061bf9263bf9f39fd3d8377c3/cuda_bindings-12.9.4-cp314-cp314t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d80bffc357df9988dca279734bc9674c3934a654cab10cadeed27ce17d8635ee", size = 11924658, upload-time = "2025-10-21T14:52:10.411Z" }, +] + +[[package]] +name = "cuda-pathfinder" +version = "1.3.3" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/0b/02/4dbe7568a42e46582248942f54dc64ad094769532adbe21e525e4edf7bc4/cuda_pathfinder-1.3.3-py3-none-any.whl", hash = "sha256:9984b664e404f7c134954a771be8775dfd6180ea1e1aef4a5a37d4be05d9bbb1", size = 27154, upload-time = "2025-12-04T22:35:08.996Z" }, +] + +[[package]] +name = "datasets" +version = "4.0.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "dill" }, + { name = "filelock" }, + { name = "fsspec", extra = ["http"] }, + { name = "huggingface-hub" }, + { name = "multiprocess" }, + { name = "numpy" }, + { name = "packaging" }, + { name = "pandas" }, + { name = "pyarrow" }, + { name = "pyyaml" }, + { name = "requests" }, + { name = "tqdm" }, + { name = "xxhash" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/e3/9d/348ed92110ba5f9b70b51ca1078d4809767a835aa2b7ce7e74ad2b98323d/datasets-4.0.0.tar.gz", hash = "sha256:9657e7140a9050db13443ba21cb5de185af8af944479b00e7ff1e00a61c8dbf1", size = 569566, upload-time = "2025-07-09T14:35:52.431Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/eb/62/eb8157afb21bd229c864521c1ab4fa8e9b4f1b06bafdd8c4668a7a31b5dd/datasets-4.0.0-py3-none-any.whl", hash = "sha256:7ef95e62025fd122882dbce6cb904c8cd3fbc829de6669a5eb939c77d50e203d", size = 494825, upload-time = "2025-07-09T14:35:50.658Z" }, +] + +[[package]] +name = "decorator" +version = "5.2.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/43/fa/6d96a0978d19e17b68d634497769987b16c8f4cd0a7a05048bec693caa6b/decorator-5.2.1.tar.gz", hash = "sha256:65f266143752f734b0a7cc83c46f4618af75b8c5911b00ccb61d0ac9b6da0360", size = 56711, upload-time = "2025-02-24T04:41:34.073Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/4e/8c/f3147f5c4b73e7550fe5f9352eaa956ae838d5c51eb58e7a25b9f3e2643b/decorator-5.2.1-py3-none-any.whl", hash = "sha256:d316bb415a2d9e2d2b3abcc4084c6502fc09240e292cd76a76afc106a1c8e04a", size = 9190, upload-time = "2025-02-24T04:41:32.565Z" }, +] + +[[package]] +name = "diffusers" +version = "0.35.2" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "filelock" }, + { name = "huggingface-hub" }, + { name = "importlib-metadata" }, + { name = "numpy" }, + { name = "pillow" }, + { name = "regex" }, + { name = "requests" }, + { name = "safetensors" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/03/68/288ca23c7c05c73e87ffe5efffc282400ac9b017f7a9bb03883f4310ea15/diffusers-0.35.2.tar.gz", hash = "sha256:30ecd552303edfcfe1724573c3918a8462ee3ab4d529bdbd4c0045f763affded", size = 3366711, upload-time = "2025-10-15T04:05:17.213Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/2a/2e/38d9824f8c6bb048c5ba21c6d4da54c29c162a46b58b3ef907a360a76d3e/diffusers-0.35.2-py3-none-any.whl", hash = "sha256:d50d5e74fdd6dcf55e5c1d304bc52cc7c2659abd1752740d736d7b54078b4db5", size = 4121649, upload-time = "2025-10-15T04:05:14.391Z" }, +] + +[[package]] +name = "dill" +version = "0.3.8" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/17/4d/ac7ffa80c69ea1df30a8aa11b3578692a5118e7cd1aa157e3ef73b092d15/dill-0.3.8.tar.gz", hash = "sha256:3ebe3c479ad625c4553aca177444d89b486b1d84982eeacded644afc0cf797ca", size = 184847, upload-time = "2024-01-27T23:42:16.145Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/c9/7a/cef76fd8438a42f96db64ddaa85280485a9c395e7df3db8158cfec1eee34/dill-0.3.8-py3-none-any.whl", hash = "sha256:c36ca9ffb54365bdd2f8eb3eff7d2a21237f8452b57ace88b1ac615b7e815bd7", size = 116252, upload-time = "2024-01-27T23:42:14.239Z" }, +] + +[[package]] +name = "distlib" +version = "0.4.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/96/8e/709914eb2b5749865801041647dc7f4e6d00b549cfe88b65ca192995f07c/distlib-0.4.0.tar.gz", hash = "sha256:feec40075be03a04501a973d81f633735b4b69f98b05450592310c0f401a4e0d", size = 614605, upload-time = "2025-07-17T16:52:00.465Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/33/6b/e0547afaf41bf2c42e52430072fa5658766e3d65bd4b03a563d1b6336f57/distlib-0.4.0-py2.py3-none-any.whl", hash = "sha256:9659f7d87e46584a30b5780e43ac7a2143098441670ff0a49d5f9034c54a6c16", size = 469047, upload-time = "2025-07-17T16:51:58.613Z" }, +] + +[[package]] +name = "einops" +version = "0.8.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/e5/81/df4fbe24dff8ba3934af99044188e20a98ed441ad17a274539b74e82e126/einops-0.8.1.tar.gz", hash = "sha256:de5d960a7a761225532e0f1959e5315ebeafc0cd43394732f103ca44b9837e84", size = 54805, upload-time = "2025-02-09T03:17:00.434Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/87/62/9773de14fe6c45c23649e98b83231fffd7b9892b6cf863251dc2afa73643/einops-0.8.1-py3-none-any.whl", hash = "sha256:919387eb55330f5757c6bea9165c5ff5cfe63a642682ea788a6d472576d81737", size = 64359, upload-time = "2025-02-09T03:17:01.998Z" }, +] + +[[package]] +name = "exceptiongroup" +version = "1.3.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "typing-extensions", marker = "python_full_version < '3.11'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/0b/9f/a65090624ecf468cdca03533906e7c69ed7588582240cfe7cc9e770b50eb/exceptiongroup-1.3.0.tar.gz", hash = "sha256:b241f5885f560bc56a59ee63ca4c6a8bfa46ae4ad651af316d4e81817bb9fd88", size = 29749, upload-time = "2025-05-10T17:42:51.123Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/36/f4/c6e662dade71f56cd2f3735141b265c3c79293c109549c1e6933b0651ffc/exceptiongroup-1.3.0-py3-none-any.whl", hash = "sha256:4d111e6e0c13d0644cad6ddaa7ed0261a0b36971f6d23e7ec9b4b9097da78a10", size = 16674, upload-time = "2025-05-10T17:42:49.33Z" }, +] + +[[package]] +name = "fastapi" +version = "0.121.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "annotated-doc" }, + { name = "pydantic" }, + { name = "starlette" }, + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/6b/a4/29e1b861fc9017488ed02ff1052feffa40940cb355ed632a8845df84ce84/fastapi-0.121.1.tar.gz", hash = "sha256:b6dba0538fd15dab6fe4d3e5493c3957d8a9e1e9257f56446b5859af66f32441", size = 342523, upload-time = "2025-11-08T21:48:14.068Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/94/fd/2e6f7d706899cc08690c5f6641e2ffbfffe019e8f16ce77104caa5730910/fastapi-0.121.1-py3-none-any.whl", hash = "sha256:2c5c7028bc3a58d8f5f09aecd3fd88a000ccc0c5ad627693264181a3c33aa1fc", size = 109192, upload-time = "2025-11-08T21:48:12.458Z" }, +] + +[[package]] +name = "fastvideo" +version = "0.1.7" +source = { editable = "." } +dependencies = [ + { name = "accelerate" }, + { name = "aiohttp" }, + { name = "av" }, + { name = "cloudpickle" }, + { name = "datasets" }, + { name = "diffusers" }, + { name = "einops" }, + { name = "fastvideo-kernel" }, + { name = "flask" }, + { name = "flask-restful" }, + { name = "ftfy" }, + { name = "gpustat" }, + { name = "gradio" }, + { name = "h5py" }, + { name = "huggingface-hub" }, + { name = "imageio" }, + { name = "imageio-ffmpeg" }, + { name = "loguru" }, + { name = "moviepy" }, + { name = "opencv-python" }, + { name = "peft" }, + { name = "pillow" }, + { name = "protobuf" }, + { name = "pyarrow" }, + { name = "pytest" }, + { name = "pyyaml" }, + { name = "ray" }, + { name = "remote-pdb" }, + { name = "requests" }, + { name = "scipy" }, + { name = "sentencepiece" }, + { name = "six" }, + { name = "test-tube" }, + { name = "timm" }, + { name = "tokenizers" }, + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, + { name = "torchaudio" }, + { name = "torchdata" }, + { name = "torchvision", version = "0.25.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torchvision", version = "0.25.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torchvision", version = "0.25.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, + { name = "tqdm" }, + { name = "transformers" }, + { name = "wandb" }, + { name = "watch" }, + { name = "wheel" }, +] + +[package.optional-dependencies] +dev = [ + { name = "av" }, + { name = "pre-commit" }, + { name = "pytest" }, + { name = "pytorch-msssim" }, +] +lint = [ + { name = "pre-commit" }, +] +rocm = [ + { name = "amdsmi" }, +] +test = [ + { name = "av" }, + { name = "pytest" }, + { name = "pytorch-msssim" }, +] + +[package.metadata] +requires-dist = [ + { name = "accelerate", specifier = "==1.0.1" }, + { name = "aiohttp" }, + { name = "amdsmi", marker = "extra == 'rocm'" }, + { name = "av" }, + { name = "av", marker = "extra == 'test'" }, + { name = "cloudpickle" }, + { name = "datasets", specifier = "==4.0.0" }, + { name = "diffusers", specifier = ">=0.33.1" }, + { name = "einops" }, + { name = "fastvideo", extras = ["lint"], marker = "extra == 'dev'" }, + { name = "fastvideo", extras = ["test"], marker = "extra == 'dev'" }, + { name = "fastvideo-kernel", editable = "fastvideo-kernel" }, + { name = "flask" }, + { name = "flask-restful" }, + { name = "ftfy", specifier = "==6.3.1" }, + { name = "gpustat" }, + { name = "gradio", specifier = "==5.32.0" }, + { name = "h5py", specifier = "==3.12.1" }, + { name = "huggingface-hub" }, + { name = "imageio", specifier = "==2.36.0" }, + { name = "imageio-ffmpeg", specifier = "==0.5.1" }, + { name = "loguru" }, + { name = "moviepy", specifier = ">=2.0.0" }, + { name = "opencv-python", specifier = "==4.10.0.84" }, + { name = "peft", specifier = ">=0.15.0" }, + { name = "pillow", specifier = ">=10.3.0" }, + { name = "pre-commit", marker = "extra == 'lint'", specifier = "==4.0.1" }, + { name = "protobuf", specifier = ">=5.28.3" }, + { name = "pyarrow" }, + { name = "pytest" }, + { name = "pytest", marker = "extra == 'test'" }, + { name = "pytorch-msssim", marker = "extra == 'test'", specifier = "==1.0.0" }, + { name = "pyyaml", specifier = "==6.0.1" }, + { name = "ray", specifier = ">=2.49.1" }, + { name = "remote-pdb" }, + { name = "requests", specifier = ">=2.32.2" }, + { name = "scipy", specifier = "==1.14.1" }, + { name = "sentencepiece", specifier = "==0.2.0" }, + { name = "six", specifier = "==1.16.0" }, + { name = "test-tube", specifier = "==0.7.5" }, + { name = "timm", specifier = "==1.0.11" }, + { name = "tokenizers", specifier = ">=0.20.1" }, + { name = "torch", marker = "sys_platform != 'linux'", specifier = ">=2.9.1", index = "https://download.pytorch.org/whl/cpu" }, + { name = "torch", marker = "sys_platform == 'linux'", specifier = ">=2.9.1", index = "https://download.pytorch.org/whl/cu128" }, + { name = "torchaudio" }, + { name = "torchdata" }, + { name = "torchvision", marker = "sys_platform != 'linux'", index = "https://download.pytorch.org/whl/cpu" }, + { name = "torchvision", marker = "sys_platform == 'linux'", index = "https://download.pytorch.org/whl/cu128" }, + { name = "tqdm" }, + { name = "transformers", specifier = "==4.57.3" }, + { name = "wandb", specifier = ">=0.21.0" }, + { name = "watch" }, + { name = "wheel" }, +] +provides-extras = ["lint", "test", "dev", "rocm"] + +[[package]] +name = "fastvideo-kernel" +version = "0.2.4" +source = { editable = "fastvideo-kernel" } +dependencies = [ + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, + { name = "triton", version = "3.5.1", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform != 'linux'" }, + { name = "triton", version = "3.6.0", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform == 'linux'" }, +] + +[package.metadata] +requires-dist = [ + { name = "torch", specifier = ">=2.5.0" }, + { name = "triton", specifier = ">=2.0.0" }, +] + +[[package]] +name = "ffmpy" +version = "1.0.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/7d/d2/1c4c582d71bcc65c76fa69fab85de6257d50fdf6fd4a2317c53917e9a581/ffmpy-1.0.0.tar.gz", hash = "sha256:b12932e95435c8820f1cd041024402765f821971e4bae753b327fc02a6e12f8b", size = 5101, upload-time = "2025-11-11T06:24:23.856Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/55/56/dd3669eccebb6d8ac81e624542ebd53fe6f08e1b8f2f8d50aeb7e3b83f99/ffmpy-1.0.0-py3-none-any.whl", hash = "sha256:5640e5f0fd03fb6236d0e119b16ccf6522db1c826fdf35dcb87087b60fd7504f", size = 5614, upload-time = "2025-11-11T06:24:22.818Z" }, +] + +[[package]] +name = "filelock" +version = "3.20.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/58/46/0028a82567109b5ef6e4d2a1f04a583fb513e6cf9527fcdd09afd817deeb/filelock-3.20.0.tar.gz", hash = "sha256:711e943b4ec6be42e1d4e6690b48dc175c822967466bb31c0c293f34334c13f4", size = 18922, upload-time = "2025-10-08T18:03:50.056Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/76/91/7216b27286936c16f5b4d0c530087e4a54eead683e6b0b73dd0c64844af6/filelock-3.20.0-py3-none-any.whl", hash = "sha256:339b4732ffda5cd79b13f4e2711a31b0365ce445d95d243bb996273d072546a2", size = 16054, upload-time = "2025-10-08T18:03:48.35Z" }, +] + +[[package]] +name = "flask" +version = "3.1.2" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "blinker" }, + { name = "click" }, + { name = "itsdangerous" }, + { name = "jinja2" }, + { name = "markupsafe" }, + { name = "werkzeug" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/dc/6d/cfe3c0fcc5e477df242b98bfe186a4c34357b4847e87ecaef04507332dab/flask-3.1.2.tar.gz", hash = "sha256:bf656c15c80190ed628ad08cdfd3aaa35beb087855e2f494910aa3774cc4fd87", size = 720160, upload-time = "2025-08-19T21:03:21.205Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/ec/f9/7f9263c5695f4bd0023734af91bedb2ff8209e8de6ead162f35d8dc762fd/flask-3.1.2-py3-none-any.whl", hash = "sha256:ca1d8112ec8a6158cc29ea4858963350011b5c846a414cdb7a954aa9e967d03c", size = 103308, upload-time = "2025-08-19T21:03:19.499Z" }, +] + +[[package]] +name = "flask-restful" +version = "0.3.10" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "aniso8601" }, + { name = "flask" }, + { name = "pytz" }, + { name = "six" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/c0/ce/a0a133db616ea47f78a41e15c4c68b9f08cab3df31eb960f61899200a119/Flask-RESTful-0.3.10.tar.gz", hash = "sha256:fe4af2ef0027df8f9b4f797aba20c5566801b6ade995ac63b588abf1a59cec37", size = 110453, upload-time = "2023-05-21T03:58:55.781Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/d7/7b/f0b45f0df7d2978e5ae51804bb5939b7897b2ace24306009da0cc34d8d1f/Flask_RESTful-0.3.10-py2.py3-none-any.whl", hash = "sha256:1cf93c535172f112e080b0d4503a8d15f93a48c88bdd36dd87269bdaf405051b", size = 26217, upload-time = "2023-05-21T03:58:54.004Z" }, +] + +[[package]] +name = "frozenlist" +version = "1.8.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/2d/f5/c831fac6cc817d26fd54c7eaccd04ef7e0288806943f7cc5bbf69f3ac1f0/frozenlist-1.8.0.tar.gz", hash = "sha256:3ede829ed8d842f6cd48fc7081d7a41001a56f1f38603f9d49bf3020d59a31ad", size = 45875, upload-time = "2025-10-06T05:38:17.865Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/83/4a/557715d5047da48d54e659203b9335be7bfaafda2c3f627b7c47e0b3aaf3/frozenlist-1.8.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:b37f6d31b3dcea7deb5e9696e529a6aa4a898adc33db82da12e4c60a7c4d2011", size = 86230, upload-time = "2025-10-06T05:35:23.699Z" }, + { url = "https://files.pythonhosted.org/packages/a2/fb/c85f9fed3ea8fe8740e5b46a59cc141c23b842eca617da8876cfce5f760e/frozenlist-1.8.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ef2b7b394f208233e471abc541cc6991f907ffd47dc72584acee3147899d6565", size = 49621, upload-time = "2025-10-06T05:35:25.341Z" }, + { url = "https://files.pythonhosted.org/packages/63/70/26ca3f06aace16f2352796b08704338d74b6d1a24ca38f2771afbb7ed915/frozenlist-1.8.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:a88f062f072d1589b7b46e951698950e7da00442fc1cacbe17e19e025dc327ad", size = 49889, upload-time = "2025-10-06T05:35:26.797Z" }, + { url = "https://files.pythonhosted.org/packages/5d/ed/c7895fd2fde7f3ee70d248175f9b6cdf792fb741ab92dc59cd9ef3bd241b/frozenlist-1.8.0-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:f57fb59d9f385710aa7060e89410aeb5058b99e62f4d16b08b91986b9a2140c2", size = 219464, upload-time = "2025-10-06T05:35:28.254Z" }, + { url = "https://files.pythonhosted.org/packages/6b/83/4d587dccbfca74cb8b810472392ad62bfa100bf8108c7223eb4c4fa2f7b3/frozenlist-1.8.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:799345ab092bee59f01a915620b5d014698547afd011e691a208637312db9186", size = 221649, upload-time = "2025-10-06T05:35:29.454Z" }, + { url = "https://files.pythonhosted.org/packages/6a/c6/fd3b9cd046ec5fff9dab66831083bc2077006a874a2d3d9247dea93ddf7e/frozenlist-1.8.0-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c23c3ff005322a6e16f71bf8692fcf4d5a304aaafe1e262c98c6d4adc7be863e", size = 219188, upload-time = "2025-10-06T05:35:30.951Z" }, + { url = "https://files.pythonhosted.org/packages/ce/80/6693f55eb2e085fc8afb28cf611448fb5b90e98e068fa1d1b8d8e66e5c7d/frozenlist-1.8.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8a76ea0f0b9dfa06f254ee06053d93a600865b3274358ca48a352ce4f0798450", size = 231748, upload-time = "2025-10-06T05:35:32.101Z" }, + { url = "https://files.pythonhosted.org/packages/97/d6/e9459f7c5183854abd989ba384fe0cc1a0fb795a83c033f0571ec5933ca4/frozenlist-1.8.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:c7366fe1418a6133d5aa824ee53d406550110984de7637d65a178010f759c6ef", size = 236351, upload-time = "2025-10-06T05:35:33.834Z" }, + { url = "https://files.pythonhosted.org/packages/97/92/24e97474b65c0262e9ecd076e826bfd1d3074adcc165a256e42e7b8a7249/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:13d23a45c4cebade99340c4165bd90eeb4a56c6d8a9d8aa49568cac19a6d0dc4", size = 218767, upload-time = "2025-10-06T05:35:35.205Z" }, + { url = "https://files.pythonhosted.org/packages/ee/bf/dc394a097508f15abff383c5108cb8ad880d1f64a725ed3b90d5c2fbf0bb/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:e4a3408834f65da56c83528fb52ce7911484f0d1eaf7b761fc66001db1646eff", size = 235887, upload-time = "2025-10-06T05:35:36.354Z" }, + { url = "https://files.pythonhosted.org/packages/40/90/25b201b9c015dbc999a5baf475a257010471a1fa8c200c843fd4abbee725/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:42145cd2748ca39f32801dad54aeea10039da6f86e303659db90db1c4b614c8c", size = 228785, upload-time = "2025-10-06T05:35:37.949Z" }, + { url = "https://files.pythonhosted.org/packages/84/f4/b5bc148df03082f05d2dd30c089e269acdbe251ac9a9cf4e727b2dbb8a3d/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:e2de870d16a7a53901e41b64ffdf26f2fbb8917b3e6ebf398098d72c5b20bd7f", size = 230312, upload-time = "2025-10-06T05:35:39.178Z" }, + { url = "https://files.pythonhosted.org/packages/db/4b/87e95b5d15097c302430e647136b7d7ab2398a702390cf4c8601975709e7/frozenlist-1.8.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:20e63c9493d33ee48536600d1a5c95eefc870cd71e7ab037763d1fbb89cc51e7", size = 217650, upload-time = "2025-10-06T05:35:40.377Z" }, + { url = "https://files.pythonhosted.org/packages/e5/70/78a0315d1fea97120591a83e0acd644da638c872f142fd72a6cebee825f3/frozenlist-1.8.0-cp310-cp310-win32.whl", hash = "sha256:adbeebaebae3526afc3c96fad434367cafbfd1b25d72369a9e5858453b1bb71a", size = 39659, upload-time = "2025-10-06T05:35:41.863Z" }, + { url = "https://files.pythonhosted.org/packages/66/aa/3f04523fb189a00e147e60c5b2205126118f216b0aa908035c45336e27e4/frozenlist-1.8.0-cp310-cp310-win_amd64.whl", hash = "sha256:667c3777ca571e5dbeb76f331562ff98b957431df140b54c85fd4d52eea8d8f6", size = 43837, upload-time = "2025-10-06T05:35:43.205Z" }, + { url = "https://files.pythonhosted.org/packages/39/75/1135feecdd7c336938bd55b4dc3b0dfc46d85b9be12ef2628574b28de776/frozenlist-1.8.0-cp310-cp310-win_arm64.whl", hash = "sha256:80f85f0a7cc86e7a54c46d99c9e1318ff01f4687c172ede30fd52d19d1da1c8e", size = 39989, upload-time = "2025-10-06T05:35:44.596Z" }, + { url = "https://files.pythonhosted.org/packages/bc/03/077f869d540370db12165c0aa51640a873fb661d8b315d1d4d67b284d7ac/frozenlist-1.8.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:09474e9831bc2b2199fad6da3c14c7b0fbdd377cce9d3d77131be28906cb7d84", size = 86912, upload-time = "2025-10-06T05:35:45.98Z" }, + { url = "https://files.pythonhosted.org/packages/df/b5/7610b6bd13e4ae77b96ba85abea1c8cb249683217ef09ac9e0ae93f25a91/frozenlist-1.8.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:17c883ab0ab67200b5f964d2b9ed6b00971917d5d8a92df149dc2c9779208ee9", size = 50046, upload-time = "2025-10-06T05:35:47.009Z" }, + { url = "https://files.pythonhosted.org/packages/6e/ef/0e8f1fe32f8a53dd26bdd1f9347efe0778b0fddf62789ea683f4cc7d787d/frozenlist-1.8.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:fa47e444b8ba08fffd1c18e8cdb9a75db1b6a27f17507522834ad13ed5922b93", size = 50119, upload-time = "2025-10-06T05:35:48.38Z" }, + { url = "https://files.pythonhosted.org/packages/11/b1/71a477adc7c36e5fb628245dfbdea2166feae310757dea848d02bd0689fd/frozenlist-1.8.0-cp311-cp311-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:2552f44204b744fba866e573be4c1f9048d6a324dfe14475103fd51613eb1d1f", size = 231067, upload-time = "2025-10-06T05:35:49.97Z" }, + { url = "https://files.pythonhosted.org/packages/45/7e/afe40eca3a2dc19b9904c0f5d7edfe82b5304cb831391edec0ac04af94c2/frozenlist-1.8.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:957e7c38f250991e48a9a73e6423db1bb9dd14e722a10f6b8bb8e16a0f55f695", size = 233160, upload-time = "2025-10-06T05:35:51.729Z" }, + { url = "https://files.pythonhosted.org/packages/a6/aa/7416eac95603ce428679d273255ffc7c998d4132cfae200103f164b108aa/frozenlist-1.8.0-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:8585e3bb2cdea02fc88ffa245069c36555557ad3609e83be0ec71f54fd4abb52", size = 228544, upload-time = "2025-10-06T05:35:53.246Z" }, + { url = "https://files.pythonhosted.org/packages/8b/3d/2a2d1f683d55ac7e3875e4263d28410063e738384d3adc294f5ff3d7105e/frozenlist-1.8.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:edee74874ce20a373d62dc28b0b18b93f645633c2943fd90ee9d898550770581", size = 243797, upload-time = "2025-10-06T05:35:54.497Z" }, + { url = "https://files.pythonhosted.org/packages/78/1e/2d5565b589e580c296d3bb54da08d206e797d941a83a6fdea42af23be79c/frozenlist-1.8.0-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:c9a63152fe95756b85f31186bddf42e4c02c6321207fd6601a1c89ebac4fe567", size = 247923, upload-time = "2025-10-06T05:35:55.861Z" }, + { url = "https://files.pythonhosted.org/packages/aa/c3/65872fcf1d326a7f101ad4d86285c403c87be7d832b7470b77f6d2ed5ddc/frozenlist-1.8.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:b6db2185db9be0a04fecf2f241c70b63b1a242e2805be291855078f2b404dd6b", size = 230886, upload-time = "2025-10-06T05:35:57.399Z" }, + { url = "https://files.pythonhosted.org/packages/a0/76/ac9ced601d62f6956f03cc794f9e04c81719509f85255abf96e2510f4265/frozenlist-1.8.0-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:f4be2e3d8bc8aabd566f8d5b8ba7ecc09249d74ba3c9ed52e54dc23a293f0b92", size = 245731, upload-time = "2025-10-06T05:35:58.563Z" }, + { url = "https://files.pythonhosted.org/packages/b9/49/ecccb5f2598daf0b4a1415497eba4c33c1e8ce07495eb07d2860c731b8d5/frozenlist-1.8.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:c8d1634419f39ea6f5c427ea2f90ca85126b54b50837f31497f3bf38266e853d", size = 241544, upload-time = "2025-10-06T05:35:59.719Z" }, + { url = "https://files.pythonhosted.org/packages/53/4b/ddf24113323c0bbcc54cb38c8b8916f1da7165e07b8e24a717b4a12cbf10/frozenlist-1.8.0-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:1a7fa382a4a223773ed64242dbe1c9c326ec09457e6b8428efb4118c685c3dfd", size = 241806, upload-time = "2025-10-06T05:36:00.959Z" }, + { url = "https://files.pythonhosted.org/packages/a7/fb/9b9a084d73c67175484ba2789a59f8eebebd0827d186a8102005ce41e1ba/frozenlist-1.8.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:11847b53d722050808926e785df837353bd4d75f1d494377e59b23594d834967", size = 229382, upload-time = "2025-10-06T05:36:02.22Z" }, + { url = "https://files.pythonhosted.org/packages/95/a3/c8fb25aac55bf5e12dae5c5aa6a98f85d436c1dc658f21c3ac73f9fa95e5/frozenlist-1.8.0-cp311-cp311-win32.whl", hash = "sha256:27c6e8077956cf73eadd514be8fb04d77fc946a7fe9f7fe167648b0b9085cc25", size = 39647, upload-time = "2025-10-06T05:36:03.409Z" }, + { url = "https://files.pythonhosted.org/packages/0a/f5/603d0d6a02cfd4c8f2a095a54672b3cf967ad688a60fb9faf04fc4887f65/frozenlist-1.8.0-cp311-cp311-win_amd64.whl", hash = "sha256:ac913f8403b36a2c8610bbfd25b8013488533e71e62b4b4adce9c86c8cea905b", size = 44064, upload-time = "2025-10-06T05:36:04.368Z" }, + { url = "https://files.pythonhosted.org/packages/5d/16/c2c9ab44e181f043a86f9a8f84d5124b62dbcb3a02c0977ec72b9ac1d3e0/frozenlist-1.8.0-cp311-cp311-win_arm64.whl", hash = "sha256:d4d3214a0f8394edfa3e303136d0575eece0745ff2b47bd2cb2e66dd92d4351a", size = 39937, upload-time = "2025-10-06T05:36:05.669Z" }, + { url = "https://files.pythonhosted.org/packages/69/29/948b9aa87e75820a38650af445d2ef2b6b8a6fab1a23b6bb9e4ef0be2d59/frozenlist-1.8.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:78f7b9e5d6f2fdb88cdde9440dc147259b62b9d3b019924def9f6478be254ac1", size = 87782, upload-time = "2025-10-06T05:36:06.649Z" }, + { url = "https://files.pythonhosted.org/packages/64/80/4f6e318ee2a7c0750ed724fa33a4bdf1eacdc5a39a7a24e818a773cd91af/frozenlist-1.8.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:229bf37d2e4acdaf808fd3f06e854a4a7a3661e871b10dc1f8f1896a3b05f18b", size = 50594, upload-time = "2025-10-06T05:36:07.69Z" }, + { url = "https://files.pythonhosted.org/packages/2b/94/5c8a2b50a496b11dd519f4a24cb5496cf125681dd99e94c604ccdea9419a/frozenlist-1.8.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:f833670942247a14eafbb675458b4e61c82e002a148f49e68257b79296e865c4", size = 50448, upload-time = "2025-10-06T05:36:08.78Z" }, + { url = "https://files.pythonhosted.org/packages/6a/bd/d91c5e39f490a49df14320f4e8c80161cfcce09f1e2cde1edd16a551abb3/frozenlist-1.8.0-cp312-cp312-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:494a5952b1c597ba44e0e78113a7266e656b9794eec897b19ead706bd7074383", size = 242411, upload-time = "2025-10-06T05:36:09.801Z" }, + { url = "https://files.pythonhosted.org/packages/8f/83/f61505a05109ef3293dfb1ff594d13d64a2324ac3482be2cedc2be818256/frozenlist-1.8.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:96f423a119f4777a4a056b66ce11527366a8bb92f54e541ade21f2374433f6d4", size = 243014, upload-time = "2025-10-06T05:36:11.394Z" }, + { url = "https://files.pythonhosted.org/packages/d8/cb/cb6c7b0f7d4023ddda30cf56b8b17494eb3a79e3fda666bf735f63118b35/frozenlist-1.8.0-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3462dd9475af2025c31cc61be6652dfa25cbfb56cbbf52f4ccfe029f38decaf8", size = 234909, upload-time = "2025-10-06T05:36:12.598Z" }, + { url = "https://files.pythonhosted.org/packages/31/c5/cd7a1f3b8b34af009fb17d4123c5a778b44ae2804e3ad6b86204255f9ec5/frozenlist-1.8.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c4c800524c9cd9bac5166cd6f55285957fcfc907db323e193f2afcd4d9abd69b", size = 250049, upload-time = "2025-10-06T05:36:14.065Z" }, + { url = "https://files.pythonhosted.org/packages/c0/01/2f95d3b416c584a1e7f0e1d6d31998c4a795f7544069ee2e0962a4b60740/frozenlist-1.8.0-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d6a5df73acd3399d893dafc71663ad22534b5aa4f94e8a2fabfe856c3c1b6a52", size = 256485, upload-time = "2025-10-06T05:36:15.39Z" }, + { url = "https://files.pythonhosted.org/packages/ce/03/024bf7720b3abaebcff6d0793d73c154237b85bdf67b7ed55e5e9596dc9a/frozenlist-1.8.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:405e8fe955c2280ce66428b3ca55e12b3c4e9c336fb2103a4937e891c69a4a29", size = 237619, upload-time = "2025-10-06T05:36:16.558Z" }, + { url = "https://files.pythonhosted.org/packages/69/fa/f8abdfe7d76b731f5d8bd217827cf6764d4f1d9763407e42717b4bed50a0/frozenlist-1.8.0-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:908bd3f6439f2fef9e85031b59fd4f1297af54415fb60e4254a95f75b3cab3f3", size = 250320, upload-time = "2025-10-06T05:36:17.821Z" }, + { url = "https://files.pythonhosted.org/packages/f5/3c/b051329f718b463b22613e269ad72138cc256c540f78a6de89452803a47d/frozenlist-1.8.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:294e487f9ec720bd8ffcebc99d575f7eff3568a08a253d1ee1a0378754b74143", size = 246820, upload-time = "2025-10-06T05:36:19.046Z" }, + { url = "https://files.pythonhosted.org/packages/0f/ae/58282e8f98e444b3f4dd42448ff36fa38bef29e40d40f330b22e7108f565/frozenlist-1.8.0-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:74c51543498289c0c43656701be6b077f4b265868fa7f8a8859c197006efb608", size = 250518, upload-time = "2025-10-06T05:36:20.763Z" }, + { url = "https://files.pythonhosted.org/packages/8f/96/007e5944694d66123183845a106547a15944fbbb7154788cbf7272789536/frozenlist-1.8.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:776f352e8329135506a1d6bf16ac3f87bc25b28e765949282dcc627af36123aa", size = 239096, upload-time = "2025-10-06T05:36:22.129Z" }, + { url = "https://files.pythonhosted.org/packages/66/bb/852b9d6db2fa40be96f29c0d1205c306288f0684df8fd26ca1951d461a56/frozenlist-1.8.0-cp312-cp312-win32.whl", hash = "sha256:433403ae80709741ce34038da08511d4a77062aa924baf411ef73d1146e74faf", size = 39985, upload-time = "2025-10-06T05:36:23.661Z" }, + { url = "https://files.pythonhosted.org/packages/b8/af/38e51a553dd66eb064cdf193841f16f077585d4d28394c2fa6235cb41765/frozenlist-1.8.0-cp312-cp312-win_amd64.whl", hash = "sha256:34187385b08f866104f0c0617404c8eb08165ab1272e884abc89c112e9c00746", size = 44591, upload-time = "2025-10-06T05:36:24.958Z" }, + { url = "https://files.pythonhosted.org/packages/a7/06/1dc65480ab147339fecc70797e9c2f69d9cea9cf38934ce08df070fdb9cb/frozenlist-1.8.0-cp312-cp312-win_arm64.whl", hash = "sha256:fe3c58d2f5db5fbd18c2987cba06d51b0529f52bc3a6cdc33d3f4eab725104bd", size = 40102, upload-time = "2025-10-06T05:36:26.333Z" }, + { url = "https://files.pythonhosted.org/packages/2d/40/0832c31a37d60f60ed79e9dfb5a92e1e2af4f40a16a29abcc7992af9edff/frozenlist-1.8.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:8d92f1a84bb12d9e56f818b3a746f3efba93c1b63c8387a73dde655e1e42282a", size = 85717, upload-time = "2025-10-06T05:36:27.341Z" }, + { url = "https://files.pythonhosted.org/packages/30/ba/b0b3de23f40bc55a7057bd38434e25c34fa48e17f20ee273bbde5e0650f3/frozenlist-1.8.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:96153e77a591c8adc2ee805756c61f59fef4cf4073a9275ee86fe8cba41241f7", size = 49651, upload-time = "2025-10-06T05:36:28.855Z" }, + { url = "https://files.pythonhosted.org/packages/0c/ab/6e5080ee374f875296c4243c381bbdef97a9ac39c6e3ce1d5f7d42cb78d6/frozenlist-1.8.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:f21f00a91358803399890ab167098c131ec2ddd5f8f5fd5fe9c9f2c6fcd91e40", size = 49417, upload-time = "2025-10-06T05:36:29.877Z" }, + { url = "https://files.pythonhosted.org/packages/d5/4e/e4691508f9477ce67da2015d8c00acd751e6287739123113a9fca6f1604e/frozenlist-1.8.0-cp313-cp313-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:fb30f9626572a76dfe4293c7194a09fb1fe93ba94c7d4f720dfae3b646b45027", size = 234391, upload-time = "2025-10-06T05:36:31.301Z" }, + { url = "https://files.pythonhosted.org/packages/40/76/c202df58e3acdf12969a7895fd6f3bc016c642e6726aa63bd3025e0fc71c/frozenlist-1.8.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:eaa352d7047a31d87dafcacbabe89df0aa506abb5b1b85a2fb91bc3faa02d822", size = 233048, upload-time = "2025-10-06T05:36:32.531Z" }, + { url = "https://files.pythonhosted.org/packages/f9/c0/8746afb90f17b73ca5979c7a3958116e105ff796e718575175319b5bb4ce/frozenlist-1.8.0-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:03ae967b4e297f58f8c774c7eabcce57fe3c2434817d4385c50661845a058121", size = 226549, upload-time = "2025-10-06T05:36:33.706Z" }, + { url = "https://files.pythonhosted.org/packages/7e/eb/4c7eefc718ff72f9b6c4893291abaae5fbc0c82226a32dcd8ef4f7a5dbef/frozenlist-1.8.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:f6292f1de555ffcc675941d65fffffb0a5bcd992905015f85d0592201793e0e5", size = 239833, upload-time = "2025-10-06T05:36:34.947Z" }, + { url = "https://files.pythonhosted.org/packages/c2/4e/e5c02187cf704224f8b21bee886f3d713ca379535f16893233b9d672ea71/frozenlist-1.8.0-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:29548f9b5b5e3460ce7378144c3010363d8035cea44bc0bf02d57f5a685e084e", size = 245363, upload-time = "2025-10-06T05:36:36.534Z" }, + { url = "https://files.pythonhosted.org/packages/1f/96/cb85ec608464472e82ad37a17f844889c36100eed57bea094518bf270692/frozenlist-1.8.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:ec3cc8c5d4084591b4237c0a272cc4f50a5b03396a47d9caaf76f5d7b38a4f11", size = 229314, upload-time = "2025-10-06T05:36:38.582Z" }, + { url = "https://files.pythonhosted.org/packages/5d/6f/4ae69c550e4cee66b57887daeebe006fe985917c01d0fff9caab9883f6d0/frozenlist-1.8.0-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:517279f58009d0b1f2e7c1b130b377a349405da3f7621ed6bfae50b10adf20c1", size = 243365, upload-time = "2025-10-06T05:36:40.152Z" }, + { url = "https://files.pythonhosted.org/packages/7a/58/afd56de246cf11780a40a2c28dc7cbabbf06337cc8ddb1c780a2d97e88d8/frozenlist-1.8.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:db1e72ede2d0d7ccb213f218df6a078a9c09a7de257c2fe8fcef16d5925230b1", size = 237763, upload-time = "2025-10-06T05:36:41.355Z" }, + { url = "https://files.pythonhosted.org/packages/cb/36/cdfaf6ed42e2644740d4a10452d8e97fa1c062e2a8006e4b09f1b5fd7d63/frozenlist-1.8.0-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:b4dec9482a65c54a5044486847b8a66bf10c9cb4926d42927ec4e8fd5db7fed8", size = 240110, upload-time = "2025-10-06T05:36:42.716Z" }, + { url = "https://files.pythonhosted.org/packages/03/a8/9ea226fbefad669f11b52e864c55f0bd57d3c8d7eb07e9f2e9a0b39502e1/frozenlist-1.8.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:21900c48ae04d13d416f0e1e0c4d81f7931f73a9dfa0b7a8746fb2fe7dd970ed", size = 233717, upload-time = "2025-10-06T05:36:44.251Z" }, + { url = "https://files.pythonhosted.org/packages/1e/0b/1b5531611e83ba7d13ccc9988967ea1b51186af64c42b7a7af465dcc9568/frozenlist-1.8.0-cp313-cp313-win32.whl", hash = "sha256:8b7b94a067d1c504ee0b16def57ad5738701e4ba10cec90529f13fa03c833496", size = 39628, upload-time = "2025-10-06T05:36:45.423Z" }, + { url = "https://files.pythonhosted.org/packages/d8/cf/174c91dbc9cc49bc7b7aab74d8b734e974d1faa8f191c74af9b7e80848e6/frozenlist-1.8.0-cp313-cp313-win_amd64.whl", hash = "sha256:878be833caa6a3821caf85eb39c5ba92d28e85df26d57afb06b35b2efd937231", size = 43882, upload-time = "2025-10-06T05:36:46.796Z" }, + { url = "https://files.pythonhosted.org/packages/c1/17/502cd212cbfa96eb1388614fe39a3fc9ab87dbbe042b66f97acb57474834/frozenlist-1.8.0-cp313-cp313-win_arm64.whl", hash = "sha256:44389d135b3ff43ba8cc89ff7f51f5a0bb6b63d829c8300f79a2fe4fe61bcc62", size = 39676, upload-time = "2025-10-06T05:36:47.8Z" }, + { url = "https://files.pythonhosted.org/packages/d2/5c/3bbfaa920dfab09e76946a5d2833a7cbdf7b9b4a91c714666ac4855b88b4/frozenlist-1.8.0-cp313-cp313t-macosx_10_13_universal2.whl", hash = "sha256:e25ac20a2ef37e91c1b39938b591457666a0fa835c7783c3a8f33ea42870db94", size = 89235, upload-time = "2025-10-06T05:36:48.78Z" }, + { url = "https://files.pythonhosted.org/packages/d2/d6/f03961ef72166cec1687e84e8925838442b615bd0b8854b54923ce5b7b8a/frozenlist-1.8.0-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:07cdca25a91a4386d2e76ad992916a85038a9b97561bf7a3fd12d5d9ce31870c", size = 50742, upload-time = "2025-10-06T05:36:49.837Z" }, + { url = "https://files.pythonhosted.org/packages/1e/bb/a6d12b7ba4c3337667d0e421f7181c82dda448ce4e7ad7ecd249a16fa806/frozenlist-1.8.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:4e0c11f2cc6717e0a741f84a527c52616140741cd812a50422f83dc31749fb52", size = 51725, upload-time = "2025-10-06T05:36:50.851Z" }, + { url = "https://files.pythonhosted.org/packages/bc/71/d1fed0ffe2c2ccd70b43714c6cab0f4188f09f8a67a7914a6b46ee30f274/frozenlist-1.8.0-cp313-cp313t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:b3210649ee28062ea6099cfda39e147fa1bc039583c8ee4481cb7811e2448c51", size = 284533, upload-time = "2025-10-06T05:36:51.898Z" }, + { url = "https://files.pythonhosted.org/packages/c9/1f/fb1685a7b009d89f9bf78a42d94461bc06581f6e718c39344754a5d9bada/frozenlist-1.8.0-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:581ef5194c48035a7de2aefc72ac6539823bb71508189e5de01d60c9dcd5fa65", size = 292506, upload-time = "2025-10-06T05:36:53.101Z" }, + { url = "https://files.pythonhosted.org/packages/e6/3b/b991fe1612703f7e0d05c0cf734c1b77aaf7c7d321df4572e8d36e7048c8/frozenlist-1.8.0-cp313-cp313t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3ef2d026f16a2b1866e1d86fc4e1291e1ed8a387b2c333809419a2f8b3a77b82", size = 274161, upload-time = "2025-10-06T05:36:54.309Z" }, + { url = "https://files.pythonhosted.org/packages/ca/ec/c5c618767bcdf66e88945ec0157d7f6c4a1322f1473392319b7a2501ded7/frozenlist-1.8.0-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5500ef82073f599ac84d888e3a8c1f77ac831183244bfd7f11eaa0289fb30714", size = 294676, upload-time = "2025-10-06T05:36:55.566Z" }, + { url = "https://files.pythonhosted.org/packages/7c/ce/3934758637d8f8a88d11f0585d6495ef54b2044ed6ec84492a91fa3b27aa/frozenlist-1.8.0-cp313-cp313t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:50066c3997d0091c411a66e710f4e11752251e6d2d73d70d8d5d4c76442a199d", size = 300638, upload-time = "2025-10-06T05:36:56.758Z" }, + { url = "https://files.pythonhosted.org/packages/fc/4f/a7e4d0d467298f42de4b41cbc7ddaf19d3cfeabaf9ff97c20c6c7ee409f9/frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:5c1c8e78426e59b3f8005e9b19f6ff46e5845895adbde20ece9218319eca6506", size = 283067, upload-time = "2025-10-06T05:36:57.965Z" }, + { url = "https://files.pythonhosted.org/packages/dc/48/c7b163063d55a83772b268e6d1affb960771b0e203b632cfe09522d67ea5/frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_armv7l.whl", hash = "sha256:eefdba20de0d938cec6a89bd4d70f346a03108a19b9df4248d3cf0d88f1b0f51", size = 292101, upload-time = "2025-10-06T05:36:59.237Z" }, + { url = "https://files.pythonhosted.org/packages/9f/d0/2366d3c4ecdc2fd391e0afa6e11500bfba0ea772764d631bbf82f0136c9d/frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:cf253e0e1c3ceb4aaff6df637ce033ff6535fb8c70a764a8f46aafd3d6ab798e", size = 289901, upload-time = "2025-10-06T05:37:00.811Z" }, + { url = "https://files.pythonhosted.org/packages/b8/94/daff920e82c1b70e3618a2ac39fbc01ae3e2ff6124e80739ce5d71c9b920/frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_s390x.whl", hash = "sha256:032efa2674356903cd0261c4317a561a6850f3ac864a63fc1583147fb05a79b0", size = 289395, upload-time = "2025-10-06T05:37:02.115Z" }, + { url = "https://files.pythonhosted.org/packages/e3/20/bba307ab4235a09fdcd3cc5508dbabd17c4634a1af4b96e0f69bfe551ebd/frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:6da155091429aeba16851ecb10a9104a108bcd32f6c1642867eadaee401c1c41", size = 283659, upload-time = "2025-10-06T05:37:03.711Z" }, + { url = "https://files.pythonhosted.org/packages/fd/00/04ca1c3a7a124b6de4f8a9a17cc2fcad138b4608e7a3fc5877804b8715d7/frozenlist-1.8.0-cp313-cp313t-win32.whl", hash = "sha256:0f96534f8bfebc1a394209427d0f8a63d343c9779cda6fc25e8e121b5fd8555b", size = 43492, upload-time = "2025-10-06T05:37:04.915Z" }, + { url = "https://files.pythonhosted.org/packages/59/5e/c69f733a86a94ab10f68e496dc6b7e8bc078ebb415281d5698313e3af3a1/frozenlist-1.8.0-cp313-cp313t-win_amd64.whl", hash = "sha256:5d63a068f978fc69421fb0e6eb91a9603187527c86b7cd3f534a5b77a592b888", size = 48034, upload-time = "2025-10-06T05:37:06.343Z" }, + { url = "https://files.pythonhosted.org/packages/16/6c/be9d79775d8abe79b05fa6d23da99ad6e7763a1d080fbae7290b286093fd/frozenlist-1.8.0-cp313-cp313t-win_arm64.whl", hash = "sha256:bf0a7e10b077bf5fb9380ad3ae8ce20ef919a6ad93b4552896419ac7e1d8e042", size = 41749, upload-time = "2025-10-06T05:37:07.431Z" }, + { url = "https://files.pythonhosted.org/packages/f1/c8/85da824b7e7b9b6e7f7705b2ecaf9591ba6f79c1177f324c2735e41d36a2/frozenlist-1.8.0-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:cee686f1f4cadeb2136007ddedd0aaf928ab95216e7691c63e50a8ec066336d0", size = 86127, upload-time = "2025-10-06T05:37:08.438Z" }, + { url = "https://files.pythonhosted.org/packages/8e/e8/a1185e236ec66c20afd72399522f142c3724c785789255202d27ae992818/frozenlist-1.8.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:119fb2a1bd47307e899c2fac7f28e85b9a543864df47aa7ec9d3c1b4545f096f", size = 49698, upload-time = "2025-10-06T05:37:09.48Z" }, + { url = "https://files.pythonhosted.org/packages/a1/93/72b1736d68f03fda5fdf0f2180fb6caaae3894f1b854d006ac61ecc727ee/frozenlist-1.8.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:4970ece02dbc8c3a92fcc5228e36a3e933a01a999f7094ff7c23fbd2beeaa67c", size = 49749, upload-time = "2025-10-06T05:37:10.569Z" }, + { url = "https://files.pythonhosted.org/packages/a7/b2/fabede9fafd976b991e9f1b9c8c873ed86f202889b864756f240ce6dd855/frozenlist-1.8.0-cp314-cp314-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:cba69cb73723c3f329622e34bdbf5ce1f80c21c290ff04256cff1cd3c2036ed2", size = 231298, upload-time = "2025-10-06T05:37:11.993Z" }, + { url = "https://files.pythonhosted.org/packages/3a/3b/d9b1e0b0eed36e70477ffb8360c49c85c8ca8ef9700a4e6711f39a6e8b45/frozenlist-1.8.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:778a11b15673f6f1df23d9586f83c4846c471a8af693a22e066508b77d201ec8", size = 232015, upload-time = "2025-10-06T05:37:13.194Z" }, + { url = "https://files.pythonhosted.org/packages/dc/94/be719d2766c1138148564a3960fc2c06eb688da592bdc25adcf856101be7/frozenlist-1.8.0-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:0325024fe97f94c41c08872db482cf8ac4800d80e79222c6b0b7b162d5b13686", size = 225038, upload-time = "2025-10-06T05:37:14.577Z" }, + { url = "https://files.pythonhosted.org/packages/e4/09/6712b6c5465f083f52f50cf74167b92d4ea2f50e46a9eea0523d658454ae/frozenlist-1.8.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:97260ff46b207a82a7567b581ab4190bd4dfa09f4db8a8b49d1a958f6aa4940e", size = 240130, upload-time = "2025-10-06T05:37:15.781Z" }, + { url = "https://files.pythonhosted.org/packages/f8/d4/cd065cdcf21550b54f3ce6a22e143ac9e4836ca42a0de1022da8498eac89/frozenlist-1.8.0-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:54b2077180eb7f83dd52c40b2750d0a9f175e06a42e3213ce047219de902717a", size = 242845, upload-time = "2025-10-06T05:37:17.037Z" }, + { url = "https://files.pythonhosted.org/packages/62/c3/f57a5c8c70cd1ead3d5d5f776f89d33110b1addae0ab010ad774d9a44fb9/frozenlist-1.8.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:2f05983daecab868a31e1da44462873306d3cbfd76d1f0b5b69c473d21dbb128", size = 229131, upload-time = "2025-10-06T05:37:18.221Z" }, + { url = "https://files.pythonhosted.org/packages/6c/52/232476fe9cb64f0742f3fde2b7d26c1dac18b6d62071c74d4ded55e0ef94/frozenlist-1.8.0-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:33f48f51a446114bc5d251fb2954ab0164d5be02ad3382abcbfe07e2531d650f", size = 240542, upload-time = "2025-10-06T05:37:19.771Z" }, + { url = "https://files.pythonhosted.org/packages/5f/85/07bf3f5d0fb5414aee5f47d33c6f5c77bfe49aac680bfece33d4fdf6a246/frozenlist-1.8.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:154e55ec0655291b5dd1b8731c637ecdb50975a2ae70c606d100750a540082f7", size = 237308, upload-time = "2025-10-06T05:37:20.969Z" }, + { url = "https://files.pythonhosted.org/packages/11/99/ae3a33d5befd41ac0ca2cc7fd3aa707c9c324de2e89db0e0f45db9a64c26/frozenlist-1.8.0-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:4314debad13beb564b708b4a496020e5306c7333fa9a3ab90374169a20ffab30", size = 238210, upload-time = "2025-10-06T05:37:22.252Z" }, + { url = "https://files.pythonhosted.org/packages/b2/60/b1d2da22f4970e7a155f0adde9b1435712ece01b3cd45ba63702aea33938/frozenlist-1.8.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:073f8bf8becba60aa931eb3bc420b217bb7d5b8f4750e6f8b3be7f3da85d38b7", size = 231972, upload-time = "2025-10-06T05:37:23.5Z" }, + { url = "https://files.pythonhosted.org/packages/3f/ab/945b2f32de889993b9c9133216c068b7fcf257d8595a0ac420ac8677cab0/frozenlist-1.8.0-cp314-cp314-win32.whl", hash = "sha256:bac9c42ba2ac65ddc115d930c78d24ab8d4f465fd3fc473cdedfccadb9429806", size = 40536, upload-time = "2025-10-06T05:37:25.581Z" }, + { url = "https://files.pythonhosted.org/packages/59/ad/9caa9b9c836d9ad6f067157a531ac48b7d36499f5036d4141ce78c230b1b/frozenlist-1.8.0-cp314-cp314-win_amd64.whl", hash = "sha256:3e0761f4d1a44f1d1a47996511752cf3dcec5bbdd9cc2b4fe595caf97754b7a0", size = 44330, upload-time = "2025-10-06T05:37:26.928Z" }, + { url = "https://files.pythonhosted.org/packages/82/13/e6950121764f2676f43534c555249f57030150260aee9dcf7d64efda11dd/frozenlist-1.8.0-cp314-cp314-win_arm64.whl", hash = "sha256:d1eaff1d00c7751b7c6662e9c5ba6eb2c17a2306ba5e2a37f24ddf3cc953402b", size = 40627, upload-time = "2025-10-06T05:37:28.075Z" }, + { url = "https://files.pythonhosted.org/packages/c0/c7/43200656ecc4e02d3f8bc248df68256cd9572b3f0017f0a0c4e93440ae23/frozenlist-1.8.0-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:d3bb933317c52d7ea5004a1c442eef86f426886fba134ef8cf4226ea6ee1821d", size = 89238, upload-time = "2025-10-06T05:37:29.373Z" }, + { url = "https://files.pythonhosted.org/packages/d1/29/55c5f0689b9c0fb765055629f472c0de484dcaf0acee2f7707266ae3583c/frozenlist-1.8.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:8009897cdef112072f93a0efdce29cd819e717fd2f649ee3016efd3cd885a7ed", size = 50738, upload-time = "2025-10-06T05:37:30.792Z" }, + { url = "https://files.pythonhosted.org/packages/ba/7d/b7282a445956506fa11da8c2db7d276adcbf2b17d8bb8407a47685263f90/frozenlist-1.8.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:2c5dcbbc55383e5883246d11fd179782a9d07a986c40f49abe89ddf865913930", size = 51739, upload-time = "2025-10-06T05:37:32.127Z" }, + { url = "https://files.pythonhosted.org/packages/62/1c/3d8622e60d0b767a5510d1d3cf21065b9db874696a51ea6d7a43180a259c/frozenlist-1.8.0-cp314-cp314t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:39ecbc32f1390387d2aa4f5a995e465e9e2f79ba3adcac92d68e3e0afae6657c", size = 284186, upload-time = "2025-10-06T05:37:33.21Z" }, + { url = "https://files.pythonhosted.org/packages/2d/14/aa36d5f85a89679a85a1d44cd7a6657e0b1c75f61e7cad987b203d2daca8/frozenlist-1.8.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:92db2bf818d5cc8d9c1f1fc56b897662e24ea5adb36ad1f1d82875bd64e03c24", size = 292196, upload-time = "2025-10-06T05:37:36.107Z" }, + { url = "https://files.pythonhosted.org/packages/05/23/6bde59eb55abd407d34f77d39a5126fb7b4f109a3f611d3929f14b700c66/frozenlist-1.8.0-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:2dc43a022e555de94c3b68a4ef0b11c4f747d12c024a520c7101709a2144fb37", size = 273830, upload-time = "2025-10-06T05:37:37.663Z" }, + { url = "https://files.pythonhosted.org/packages/d2/3f/22cff331bfad7a8afa616289000ba793347fcd7bc275f3b28ecea2a27909/frozenlist-1.8.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:cb89a7f2de3602cfed448095bab3f178399646ab7c61454315089787df07733a", size = 294289, upload-time = "2025-10-06T05:37:39.261Z" }, + { url = "https://files.pythonhosted.org/packages/a4/89/5b057c799de4838b6c69aa82b79705f2027615e01be996d2486a69ca99c4/frozenlist-1.8.0-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:33139dc858c580ea50e7e60a1b0ea003efa1fd42e6ec7fdbad78fff65fad2fd2", size = 300318, upload-time = "2025-10-06T05:37:43.213Z" }, + { url = "https://files.pythonhosted.org/packages/30/de/2c22ab3eb2a8af6d69dc799e48455813bab3690c760de58e1bf43b36da3e/frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:168c0969a329b416119507ba30b9ea13688fafffac1b7822802537569a1cb0ef", size = 282814, upload-time = "2025-10-06T05:37:45.337Z" }, + { url = "https://files.pythonhosted.org/packages/59/f7/970141a6a8dbd7f556d94977858cfb36fa9b66e0892c6dd780d2219d8cd8/frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:28bd570e8e189d7f7b001966435f9dac6718324b5be2990ac496cf1ea9ddb7fe", size = 291762, upload-time = "2025-10-06T05:37:46.657Z" }, + { url = "https://files.pythonhosted.org/packages/c1/15/ca1adae83a719f82df9116d66f5bb28bb95557b3951903d39135620ef157/frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:b2a095d45c5d46e5e79ba1e5b9cb787f541a8dee0433836cea4b96a2c439dcd8", size = 289470, upload-time = "2025-10-06T05:37:47.946Z" }, + { url = "https://files.pythonhosted.org/packages/ac/83/dca6dc53bf657d371fbc88ddeb21b79891e747189c5de990b9dfff2ccba1/frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:eab8145831a0d56ec9c4139b6c3e594c7a83c2c8be25d5bcf2d86136a532287a", size = 289042, upload-time = "2025-10-06T05:37:49.499Z" }, + { url = "https://files.pythonhosted.org/packages/96/52/abddd34ca99be142f354398700536c5bd315880ed0a213812bc491cff5e4/frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:974b28cf63cc99dfb2188d8d222bc6843656188164848c4f679e63dae4b0708e", size = 283148, upload-time = "2025-10-06T05:37:50.745Z" }, + { url = "https://files.pythonhosted.org/packages/af/d3/76bd4ed4317e7119c2b7f57c3f6934aba26d277acc6309f873341640e21f/frozenlist-1.8.0-cp314-cp314t-win32.whl", hash = "sha256:342c97bf697ac5480c0a7ec73cd700ecfa5a8a40ac923bd035484616efecc2df", size = 44676, upload-time = "2025-10-06T05:37:52.222Z" }, + { url = "https://files.pythonhosted.org/packages/89/76/c615883b7b521ead2944bb3480398cbb07e12b7b4e4d073d3752eb721558/frozenlist-1.8.0-cp314-cp314t-win_amd64.whl", hash = "sha256:06be8f67f39c8b1dc671f5d83aaefd3358ae5cdcf8314552c57e7ed3e6475bdd", size = 49451, upload-time = "2025-10-06T05:37:53.425Z" }, + { url = "https://files.pythonhosted.org/packages/e0/a3/5982da14e113d07b325230f95060e2169f5311b1017ea8af2a29b374c289/frozenlist-1.8.0-cp314-cp314t-win_arm64.whl", hash = "sha256:102e6314ca4da683dca92e3b1355490fed5f313b768500084fbe6371fddfdb79", size = 42507, upload-time = "2025-10-06T05:37:54.513Z" }, + { url = "https://files.pythonhosted.org/packages/9a/9a/e35b4a917281c0b8419d4207f4334c8e8c5dbf4f3f5f9ada73958d937dcc/frozenlist-1.8.0-py3-none-any.whl", hash = "sha256:0c18a16eab41e82c295618a77502e17b195883241c563b00f0aa5106fc4eaa0d", size = 13409, upload-time = "2025-10-06T05:38:16.721Z" }, +] + +[[package]] +name = "fsspec" +version = "2025.3.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/34/f4/5721faf47b8c499e776bc34c6a8fc17efdf7fdef0b00f398128bc5dcb4ac/fsspec-2025.3.0.tar.gz", hash = "sha256:a935fd1ea872591f2b5148907d103488fc523295e6c64b835cfad8c3eca44972", size = 298491, upload-time = "2025-03-07T21:47:56.461Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/56/53/eb690efa8513166adef3e0669afd31e95ffde69fb3c52ec2ac7223ed6018/fsspec-2025.3.0-py3-none-any.whl", hash = "sha256:efb87af3efa9103f94ca91a7f8cb7a4df91af9f74fc106c9c7ea0efd7277c1b3", size = 193615, upload-time = "2025-03-07T21:47:54.809Z" }, +] + +[package.optional-dependencies] +http = [ + { name = "aiohttp" }, +] + +[[package]] +name = "ftfy" +version = "6.3.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "wcwidth" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/a5/d3/8650919bc3c7c6e90ee3fa7fd618bf373cbbe55dff043bd67353dbb20cd8/ftfy-6.3.1.tar.gz", hash = "sha256:9b3c3d90f84fb267fe64d375a07b7f8912d817cf86009ae134aa03e1819506ec", size = 308927, upload-time = "2024-10-26T00:50:35.149Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/ab/6e/81d47999aebc1b155f81eca4477a616a70f238a2549848c38983f3c22a82/ftfy-6.3.1-py3-none-any.whl", hash = "sha256:7c70eb532015cd2f9adb53f101fb6c7945988d023a085d127d1573dc49dd0083", size = 44821, upload-time = "2024-10-26T00:50:33.425Z" }, +] + +[[package]] +name = "future" +version = "1.0.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/a7/b2/4140c69c6a66432916b26158687e821ba631a4c9273c474343badf84d3ba/future-1.0.0.tar.gz", hash = "sha256:bd2968309307861edae1458a4f8a4f3598c03be43b97521076aebf5d94c07b05", size = 1228490, upload-time = "2024-02-21T11:52:38.461Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/da/71/ae30dadffc90b9006d77af76b393cb9dfbfc9629f339fc1574a1c52e6806/future-1.0.0-py3-none-any.whl", hash = "sha256:929292d34f5872e70396626ef385ec22355a1fae8ad29e1a734c3e43f9fbc216", size = 491326, upload-time = "2024-02-21T11:52:35.956Z" }, +] + +[[package]] +name = "gitdb" +version = "4.0.12" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "smmap" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/72/94/63b0fc47eb32792c7ba1fe1b694daec9a63620db1e313033d18140c2320a/gitdb-4.0.12.tar.gz", hash = "sha256:5ef71f855d191a3326fcfbc0d5da835f26b13fbcba60c32c21091c349ffdb571", size = 394684, upload-time = "2025-01-02T07:20:46.413Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/a0/61/5c78b91c3143ed5c14207f463aecfc8f9dbb5092fb2869baf37c273b2705/gitdb-4.0.12-py3-none-any.whl", hash = "sha256:67073e15955400952c6565cc3e707c554a4eea2e428946f7a4c162fab9bd9bcf", size = 62794, upload-time = "2025-01-02T07:20:43.624Z" }, +] + +[[package]] +name = "gitpython" +version = "3.1.45" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "gitdb" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/9a/c8/dd58967d119baab745caec2f9d853297cec1989ec1d63f677d3880632b88/gitpython-3.1.45.tar.gz", hash = "sha256:85b0ee964ceddf211c41b9f27a49086010a190fd8132a24e21f362a4b36a791c", size = 215076, upload-time = "2025-07-24T03:45:54.871Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/01/61/d4b89fec821f72385526e1b9d9a3a0385dda4a72b206d28049e2c7cd39b8/gitpython-3.1.45-py3-none-any.whl", hash = "sha256:8908cb2e02fb3b93b7eb0f2827125cb699869470432cc885f019b8fd0fccff77", size = 208168, upload-time = "2025-07-24T03:45:52.517Z" }, +] + +[[package]] +name = "gpustat" +version = "1.1.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "blessed" }, + { name = "nvidia-ml-py" }, + { name = "psutil" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/79/c4/46d005aec3bf911cb030467d91e062a5386ff4a03e51874424cacc0f60c1/gpustat-1.1.1.tar.gz", hash = "sha256:c18d3ed5518fc16300c42d694debc70aebb3be55cae91f1db64d63b5fa8af9d8", size = 98052, upload-time = "2023-08-22T19:39:06.062Z" } + +[[package]] +name = "gradio" +version = "5.32.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "aiofiles" }, + { name = "anyio" }, + { name = "audioop-lts", marker = "python_full_version >= '3.13'" }, + { name = "fastapi" }, + { name = "ffmpy" }, + { name = "gradio-client" }, + { name = "groovy" }, + { name = "httpx" }, + { name = "huggingface-hub" }, + { name = "jinja2" }, + { name = "markupsafe" }, + { name = "numpy" }, + { name = "orjson" }, + { name = "packaging" }, + { name = "pandas" }, + { name = "pillow" }, + { name = "pydantic" }, + { name = "pydub" }, + { name = "python-multipart" }, + { name = "pyyaml" }, + { name = "ruff", marker = "sys_platform != 'emscripten'" }, + { name = "safehttpx" }, + { name = "semantic-version" }, + { name = "starlette", marker = "sys_platform != 'emscripten'" }, + { name = "tomlkit" }, + { name = "typer", marker = "sys_platform != 'emscripten'" }, + { name = "typing-extensions" }, + { name = "urllib3", marker = "sys_platform == 'emscripten'" }, + { name = "uvicorn", marker = "sys_platform != 'emscripten'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/8f/ff/e5bd3863bc25397319547e1a0de7b562fadeb671cd7e6a5d048b0c88be31/gradio-5.32.0.tar.gz", hash = "sha256:c0b151ac039567123b8f9e058aa3256aadc906ecf7b79c26534aa81dcc7ba313", size = 64783727, upload-time = "2025-05-30T13:59:46.265Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/e2/4d/52a719a5e9a70022438d38238f2a9a3297b864c8ceaa61d77e3f1c1b472a/gradio-5.32.0-py3-none-any.whl", hash = "sha256:45fdb15784f4be19eca2beb1d45d107238ca614c177b20b0e3d4f2d6aee81bae", size = 54201078, upload-time = "2025-05-30T13:59:40.764Z" }, +] + +[[package]] +name = "gradio-client" +version = "1.10.2" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "fsspec" }, + { name = "httpx" }, + { name = "huggingface-hub" }, + { name = "packaging" }, + { name = "typing-extensions" }, + { name = "websockets" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/d2/86/6684afe8691b024200fdc8983924f04b5f76bb401b9c700e5752a23595a0/gradio_client-1.10.2.tar.gz", hash = "sha256:bf71ba95714784fa77ca0cfb20189ad91c55e563c2dc71722d023a97f1815d7f", size = 321294, upload-time = "2025-05-30T13:59:55.756Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/9b/1b/b372308c263379ae3ebc440512432979458330113bdee26cef86c89bf48e/gradio_client-1.10.2-py3-none-any.whl", hash = "sha256:6de67b6224123d264c7887caa0586b2a9e2c369ec32ca38927cf8a841694edcd", size = 323311, upload-time = "2025-05-30T13:59:54.555Z" }, +] + +[[package]] +name = "groovy" +version = "0.1.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/52/36/bbdede67400277bef33d3ec0e6a31750da972c469f75966b4930c753218f/groovy-0.1.2.tar.gz", hash = "sha256:25c1dc09b3f9d7e292458aa762c6beb96ea037071bf5e917fc81fb78d2231083", size = 17325, upload-time = "2025-02-28T20:24:56.068Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/28/27/3d6dcadc8a3214d8522c1e7f6a19554e33659be44546d44a2f7572ac7d2a/groovy-0.1.2-py3-none-any.whl", hash = "sha256:7f7975bab18c729a257a8b1ae9dcd70b7cafb1720481beae47719af57c35fa64", size = 14090, upload-time = "2025-02-28T20:24:55.152Z" }, +] + +[[package]] +name = "grpcio" +version = "1.76.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/b6/e0/318c1ce3ae5a17894d5791e87aea147587c9e702f24122cc7a5c8bbaeeb1/grpcio-1.76.0.tar.gz", hash = "sha256:7be78388d6da1a25c0d5ec506523db58b18be22d9c37d8d3a32c08be4987bd73", size = 12785182, upload-time = "2025-10-21T16:23:12.106Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/88/17/ff4795dc9a34b6aee6ec379f1b66438a3789cd1315aac0cbab60d92f74b3/grpcio-1.76.0-cp310-cp310-linux_armv7l.whl", hash = "sha256:65a20de41e85648e00305c1bb09a3598f840422e522277641145a32d42dcefcc", size = 5840037, upload-time = "2025-10-21T16:20:25.069Z" }, + { url = "https://files.pythonhosted.org/packages/4e/ff/35f9b96e3fa2f12e1dcd58a4513a2e2294a001d64dec81677361b7040c9a/grpcio-1.76.0-cp310-cp310-macosx_11_0_universal2.whl", hash = "sha256:40ad3afe81676fd9ec6d9d406eda00933f218038433980aa19d401490e46ecde", size = 11836482, upload-time = "2025-10-21T16:20:30.113Z" }, + { url = "https://files.pythonhosted.org/packages/3e/1c/8374990f9545e99462caacea5413ed783014b3b66ace49e35c533f07507b/grpcio-1.76.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:035d90bc79eaa4bed83f524331d55e35820725c9fbb00ffa1904d5550ed7ede3", size = 6407178, upload-time = "2025-10-21T16:20:32.733Z" }, + { url = "https://files.pythonhosted.org/packages/1e/77/36fd7d7c75a6c12542c90a6d647a27935a1ecaad03e0ffdb7c42db6b04d2/grpcio-1.76.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:4215d3a102bd95e2e11b5395c78562967959824156af11fa93d18fdd18050990", size = 7075684, upload-time = "2025-10-21T16:20:35.435Z" }, + { url = "https://files.pythonhosted.org/packages/38/f7/e3cdb252492278e004722306c5a8935eae91e64ea11f0af3437a7de2e2b7/grpcio-1.76.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:49ce47231818806067aea3324d4bf13825b658ad662d3b25fada0bdad9b8a6af", size = 6611133, upload-time = "2025-10-21T16:20:37.541Z" }, + { url = "https://files.pythonhosted.org/packages/7e/20/340db7af162ccd20a0893b5f3c4a5d676af7b71105517e62279b5b61d95a/grpcio-1.76.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:8cc3309d8e08fd79089e13ed4819d0af72aa935dd8f435a195fd152796752ff2", size = 7195507, upload-time = "2025-10-21T16:20:39.643Z" }, + { url = "https://files.pythonhosted.org/packages/10/f0/b2160addc1487bd8fa4810857a27132fb4ce35c1b330c2f3ac45d697b106/grpcio-1.76.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:971fd5a1d6e62e00d945423a567e42eb1fa678ba89072832185ca836a94daaa6", size = 8160651, upload-time = "2025-10-21T16:20:42.492Z" }, + { url = "https://files.pythonhosted.org/packages/2c/2c/ac6f98aa113c6ef111b3f347854e99ebb7fb9d8f7bb3af1491d438f62af4/grpcio-1.76.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:9d9adda641db7207e800a7f089068f6f645959f2df27e870ee81d44701dd9db3", size = 7620568, upload-time = "2025-10-21T16:20:45.995Z" }, + { url = "https://files.pythonhosted.org/packages/90/84/7852f7e087285e3ac17a2703bc4129fafee52d77c6c82af97d905566857e/grpcio-1.76.0-cp310-cp310-win32.whl", hash = "sha256:063065249d9e7e0782d03d2bca50787f53bd0fb89a67de9a7b521c4a01f1989b", size = 3998879, upload-time = "2025-10-21T16:20:48.592Z" }, + { url = "https://files.pythonhosted.org/packages/10/30/d3d2adcbb6dd3ff59d6ac3df6ef830e02b437fb5c90990429fd180e52f30/grpcio-1.76.0-cp310-cp310-win_amd64.whl", hash = "sha256:a6ae758eb08088d36812dd5d9af7a9859c05b1e0f714470ea243694b49278e7b", size = 4706892, upload-time = "2025-10-21T16:20:50.697Z" }, + { url = "https://files.pythonhosted.org/packages/a0/00/8163a1beeb6971f66b4bbe6ac9457b97948beba8dd2fc8e1281dce7f79ec/grpcio-1.76.0-cp311-cp311-linux_armv7l.whl", hash = "sha256:2e1743fbd7f5fa713a1b0a8ac8ebabf0ec980b5d8809ec358d488e273b9cf02a", size = 5843567, upload-time = "2025-10-21T16:20:52.829Z" }, + { url = "https://files.pythonhosted.org/packages/10/c1/934202f5cf335e6d852530ce14ddb0fef21be612ba9ecbbcbd4d748ca32d/grpcio-1.76.0-cp311-cp311-macosx_11_0_universal2.whl", hash = "sha256:a8c2cf1209497cf659a667d7dea88985e834c24b7c3b605e6254cbb5076d985c", size = 11848017, upload-time = "2025-10-21T16:20:56.705Z" }, + { url = "https://files.pythonhosted.org/packages/11/0b/8dec16b1863d74af6eb3543928600ec2195af49ca58b16334972f6775663/grpcio-1.76.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:08caea849a9d3c71a542827d6df9d5a69067b0a1efbea8a855633ff5d9571465", size = 6412027, upload-time = "2025-10-21T16:20:59.3Z" }, + { url = "https://files.pythonhosted.org/packages/d7/64/7b9e6e7ab910bea9d46f2c090380bab274a0b91fb0a2fe9b0cd399fffa12/grpcio-1.76.0-cp311-cp311-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:f0e34c2079d47ae9f6188211db9e777c619a21d4faba6977774e8fa43b085e48", size = 7075913, upload-time = "2025-10-21T16:21:01.645Z" }, + { url = "https://files.pythonhosted.org/packages/68/86/093c46e9546073cefa789bd76d44c5cb2abc824ca62af0c18be590ff13ba/grpcio-1.76.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:8843114c0cfce61b40ad48df65abcfc00d4dba82eae8718fab5352390848c5da", size = 6615417, upload-time = "2025-10-21T16:21:03.844Z" }, + { url = "https://files.pythonhosted.org/packages/f7/b6/5709a3a68500a9c03da6fb71740dcdd5ef245e39266461a03f31a57036d8/grpcio-1.76.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:8eddfb4d203a237da6f3cc8a540dad0517d274b5a1e9e636fd8d2c79b5c1d397", size = 7199683, upload-time = "2025-10-21T16:21:06.195Z" }, + { url = "https://files.pythonhosted.org/packages/91/d3/4b1f2bf16ed52ce0b508161df3a2d186e4935379a159a834cb4a7d687429/grpcio-1.76.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:32483fe2aab2c3794101c2a159070584e5db11d0aa091b2c0ea9c4fc43d0d749", size = 8163109, upload-time = "2025-10-21T16:21:08.498Z" }, + { url = "https://files.pythonhosted.org/packages/5c/61/d9043f95f5f4cf085ac5dd6137b469d41befb04bd80280952ffa2a4c3f12/grpcio-1.76.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:dcfe41187da8992c5f40aa8c5ec086fa3672834d2be57a32384c08d5a05b4c00", size = 7626676, upload-time = "2025-10-21T16:21:10.693Z" }, + { url = "https://files.pythonhosted.org/packages/36/95/fd9a5152ca02d8881e4dd419cdd790e11805979f499a2e5b96488b85cf27/grpcio-1.76.0-cp311-cp311-win32.whl", hash = "sha256:2107b0c024d1b35f4083f11245c0e23846ae64d02f40b2b226684840260ed054", size = 3997688, upload-time = "2025-10-21T16:21:12.746Z" }, + { url = "https://files.pythonhosted.org/packages/60/9c/5c359c8d4c9176cfa3c61ecd4efe5affe1f38d9bae81e81ac7186b4c9cc8/grpcio-1.76.0-cp311-cp311-win_amd64.whl", hash = "sha256:522175aba7af9113c48ec10cc471b9b9bd4f6ceb36aeb4544a8e2c80ed9d252d", size = 4709315, upload-time = "2025-10-21T16:21:15.26Z" }, + { url = "https://files.pythonhosted.org/packages/bf/05/8e29121994b8d959ffa0afd28996d452f291b48cfc0875619de0bde2c50c/grpcio-1.76.0-cp312-cp312-linux_armv7l.whl", hash = "sha256:81fd9652b37b36f16138611c7e884eb82e0cec137c40d3ef7c3f9b3ed00f6ed8", size = 5799718, upload-time = "2025-10-21T16:21:17.939Z" }, + { url = "https://files.pythonhosted.org/packages/d9/75/11d0e66b3cdf998c996489581bdad8900db79ebd83513e45c19548f1cba4/grpcio-1.76.0-cp312-cp312-macosx_11_0_universal2.whl", hash = "sha256:04bbe1bfe3a68bbfd4e52402ab7d4eb59d72d02647ae2042204326cf4bbad280", size = 11825627, upload-time = "2025-10-21T16:21:20.466Z" }, + { url = "https://files.pythonhosted.org/packages/28/50/2f0aa0498bc188048f5d9504dcc5c2c24f2eb1a9337cd0fa09a61a2e75f0/grpcio-1.76.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d388087771c837cdb6515539f43b9d4bf0b0f23593a24054ac16f7a960be16f4", size = 6359167, upload-time = "2025-10-21T16:21:23.122Z" }, + { url = "https://files.pythonhosted.org/packages/66/e5/bbf0bb97d29ede1d59d6588af40018cfc345b17ce979b7b45424628dc8bb/grpcio-1.76.0-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:9f8f757bebaaea112c00dba718fc0d3260052ce714e25804a03f93f5d1c6cc11", size = 7044267, upload-time = "2025-10-21T16:21:25.995Z" }, + { url = "https://files.pythonhosted.org/packages/f5/86/f6ec2164f743d9609691115ae8ece098c76b894ebe4f7c94a655c6b03e98/grpcio-1.76.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:980a846182ce88c4f2f7e2c22c56aefd515daeb36149d1c897f83cf57999e0b6", size = 6573963, upload-time = "2025-10-21T16:21:28.631Z" }, + { url = "https://files.pythonhosted.org/packages/60/bc/8d9d0d8505feccfdf38a766d262c71e73639c165b311c9457208b56d92ae/grpcio-1.76.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:f92f88e6c033db65a5ae3d97905c8fea9c725b63e28d5a75cb73b49bda5024d8", size = 7164484, upload-time = "2025-10-21T16:21:30.837Z" }, + { url = "https://files.pythonhosted.org/packages/67/e6/5d6c2fc10b95edf6df9b8f19cf10a34263b7fd48493936fffd5085521292/grpcio-1.76.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:4baf3cbe2f0be3289eb68ac8ae771156971848bb8aaff60bad42005539431980", size = 8127777, upload-time = "2025-10-21T16:21:33.577Z" }, + { url = "https://files.pythonhosted.org/packages/3f/c8/dce8ff21c86abe025efe304d9e31fdb0deaaa3b502b6a78141080f206da0/grpcio-1.76.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:615ba64c208aaceb5ec83bfdce7728b80bfeb8be97562944836a7a0a9647d882", size = 7594014, upload-time = "2025-10-21T16:21:41.882Z" }, + { url = "https://files.pythonhosted.org/packages/e0/42/ad28191ebf983a5d0ecef90bab66baa5a6b18f2bfdef9d0a63b1973d9f75/grpcio-1.76.0-cp312-cp312-win32.whl", hash = "sha256:45d59a649a82df5718fd9527ce775fd66d1af35e6d31abdcdc906a49c6822958", size = 3984750, upload-time = "2025-10-21T16:21:44.006Z" }, + { url = "https://files.pythonhosted.org/packages/9e/00/7bd478cbb851c04a48baccaa49b75abaa8e4122f7d86da797500cccdd771/grpcio-1.76.0-cp312-cp312-win_amd64.whl", hash = "sha256:c088e7a90b6017307f423efbb9d1ba97a22aa2170876223f9709e9d1de0b5347", size = 4704003, upload-time = "2025-10-21T16:21:46.244Z" }, + { url = "https://files.pythonhosted.org/packages/fc/ed/71467ab770effc9e8cef5f2e7388beb2be26ed642d567697bb103a790c72/grpcio-1.76.0-cp313-cp313-linux_armv7l.whl", hash = "sha256:26ef06c73eb53267c2b319f43e6634c7556ea37672029241a056629af27c10e2", size = 5807716, upload-time = "2025-10-21T16:21:48.475Z" }, + { url = "https://files.pythonhosted.org/packages/2c/85/c6ed56f9817fab03fa8a111ca91469941fb514e3e3ce6d793cb8f1e1347b/grpcio-1.76.0-cp313-cp313-macosx_11_0_universal2.whl", hash = "sha256:45e0111e73f43f735d70786557dc38141185072d7ff8dc1829d6a77ac1471468", size = 11821522, upload-time = "2025-10-21T16:21:51.142Z" }, + { url = "https://files.pythonhosted.org/packages/ac/31/2b8a235ab40c39cbc141ef647f8a6eb7b0028f023015a4842933bc0d6831/grpcio-1.76.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:83d57312a58dcfe2a3a0f9d1389b299438909a02db60e2f2ea2ae2d8034909d3", size = 6362558, upload-time = "2025-10-21T16:21:54.213Z" }, + { url = "https://files.pythonhosted.org/packages/bd/64/9784eab483358e08847498ee56faf8ff6ea8e0a4592568d9f68edc97e9e9/grpcio-1.76.0-cp313-cp313-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:3e2a27c89eb9ac3d81ec8835e12414d73536c6e620355d65102503064a4ed6eb", size = 7049990, upload-time = "2025-10-21T16:21:56.476Z" }, + { url = "https://files.pythonhosted.org/packages/2b/94/8c12319a6369434e7a184b987e8e9f3b49a114c489b8315f029e24de4837/grpcio-1.76.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:61f69297cba3950a524f61c7c8ee12e55c486cb5f7db47ff9dcee33da6f0d3ae", size = 6575387, upload-time = "2025-10-21T16:21:59.051Z" }, + { url = "https://files.pythonhosted.org/packages/15/0f/f12c32b03f731f4a6242f771f63039df182c8b8e2cf8075b245b409259d4/grpcio-1.76.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:6a15c17af8839b6801d554263c546c69c4d7718ad4321e3166175b37eaacca77", size = 7166668, upload-time = "2025-10-21T16:22:02.049Z" }, + { url = "https://files.pythonhosted.org/packages/ff/2d/3ec9ce0c2b1d92dd59d1c3264aaec9f0f7c817d6e8ac683b97198a36ed5a/grpcio-1.76.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:25a18e9810fbc7e7f03ec2516addc116a957f8cbb8cbc95ccc80faa072743d03", size = 8124928, upload-time = "2025-10-21T16:22:04.984Z" }, + { url = "https://files.pythonhosted.org/packages/1a/74/fd3317be5672f4856bcdd1a9e7b5e17554692d3db9a3b273879dc02d657d/grpcio-1.76.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:931091142fd8cc14edccc0845a79248bc155425eee9a98b2db2ea4f00a235a42", size = 7589983, upload-time = "2025-10-21T16:22:07.881Z" }, + { url = "https://files.pythonhosted.org/packages/45/bb/ca038cf420f405971f19821c8c15bcbc875505f6ffadafe9ffd77871dc4c/grpcio-1.76.0-cp313-cp313-win32.whl", hash = "sha256:5e8571632780e08526f118f74170ad8d50fb0a48c23a746bef2a6ebade3abd6f", size = 3984727, upload-time = "2025-10-21T16:22:10.032Z" }, + { url = "https://files.pythonhosted.org/packages/41/80/84087dc56437ced7cdd4b13d7875e7439a52a261e3ab4e06488ba6173b0a/grpcio-1.76.0-cp313-cp313-win_amd64.whl", hash = "sha256:f9f7bd5faab55f47231ad8dba7787866b69f5e93bc306e3915606779bbfb4ba8", size = 4702799, upload-time = "2025-10-21T16:22:12.709Z" }, + { url = "https://files.pythonhosted.org/packages/b4/46/39adac80de49d678e6e073b70204091e76631e03e94928b9ea4ecf0f6e0e/grpcio-1.76.0-cp314-cp314-linux_armv7l.whl", hash = "sha256:ff8a59ea85a1f2191a0ffcc61298c571bc566332f82e5f5be1b83c9d8e668a62", size = 5808417, upload-time = "2025-10-21T16:22:15.02Z" }, + { url = "https://files.pythonhosted.org/packages/9c/f5/a4531f7fb8b4e2a60b94e39d5d924469b7a6988176b3422487be61fe2998/grpcio-1.76.0-cp314-cp314-macosx_11_0_universal2.whl", hash = "sha256:06c3d6b076e7b593905d04fdba6a0525711b3466f43b3400266f04ff735de0cd", size = 11828219, upload-time = "2025-10-21T16:22:17.954Z" }, + { url = "https://files.pythonhosted.org/packages/4b/1c/de55d868ed7a8bd6acc6b1d6ddc4aa36d07a9f31d33c912c804adb1b971b/grpcio-1.76.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:fd5ef5932f6475c436c4a55e4336ebbe47bd3272be04964a03d316bbf4afbcbc", size = 6367826, upload-time = "2025-10-21T16:22:20.721Z" }, + { url = "https://files.pythonhosted.org/packages/59/64/99e44c02b5adb0ad13ab3adc89cb33cb54bfa90c74770f2607eea629b86f/grpcio-1.76.0-cp314-cp314-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:b331680e46239e090f5b3cead313cc772f6caa7d0fc8de349337563125361a4a", size = 7049550, upload-time = "2025-10-21T16:22:23.637Z" }, + { url = "https://files.pythonhosted.org/packages/43/28/40a5be3f9a86949b83e7d6a2ad6011d993cbe9b6bd27bea881f61c7788b6/grpcio-1.76.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:2229ae655ec4e8999599469559e97630185fdd53ae1e8997d147b7c9b2b72cba", size = 6575564, upload-time = "2025-10-21T16:22:26.016Z" }, + { url = "https://files.pythonhosted.org/packages/4b/a9/1be18e6055b64467440208a8559afac243c66a8b904213af6f392dc2212f/grpcio-1.76.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:490fa6d203992c47c7b9e4a9d39003a0c2bcc1c9aa3c058730884bbbb0ee9f09", size = 7176236, upload-time = "2025-10-21T16:22:28.362Z" }, + { url = "https://files.pythonhosted.org/packages/0f/55/dba05d3fcc151ce6e81327541d2cc8394f442f6b350fead67401661bf041/grpcio-1.76.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:479496325ce554792dba6548fae3df31a72cef7bad71ca2e12b0e58f9b336bfc", size = 8125795, upload-time = "2025-10-21T16:22:31.075Z" }, + { url = "https://files.pythonhosted.org/packages/4a/45/122df922d05655f63930cf42c9e3f72ba20aadb26c100ee105cad4ce4257/grpcio-1.76.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:1c9b93f79f48b03ada57ea24725d83a30284a012ec27eab2cf7e50a550cbbbcc", size = 7592214, upload-time = "2025-10-21T16:22:33.831Z" }, + { url = "https://files.pythonhosted.org/packages/4a/6e/0b899b7f6b66e5af39e377055fb4a6675c9ee28431df5708139df2e93233/grpcio-1.76.0-cp314-cp314-win32.whl", hash = "sha256:747fa73efa9b8b1488a95d0ba1039c8e2dca0f741612d80415b1e1c560febf4e", size = 4062961, upload-time = "2025-10-21T16:22:36.468Z" }, + { url = "https://files.pythonhosted.org/packages/19/41/0b430b01a2eb38ee887f88c1f07644a1df8e289353b78e82b37ef988fb64/grpcio-1.76.0-cp314-cp314-win_amd64.whl", hash = "sha256:922fa70ba549fce362d2e2871ab542082d66e2aaf0c19480ea453905b01f384e", size = 4834462, upload-time = "2025-10-21T16:22:39.772Z" }, +] + +[[package]] +name = "h11" +version = "0.16.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/01/ee/02a2c011bdab74c6fb3c75474d40b3052059d95df7e73351460c8588d963/h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1", size = 101250, upload-time = "2025-04-24T03:35:25.427Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/04/4b/29cac41a4d98d144bf5f6d33995617b185d14b22401f75ca86f384e87ff1/h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86", size = 37515, upload-time = "2025-04-24T03:35:24.344Z" }, +] + +[[package]] +name = "h5py" +version = "3.12.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "numpy" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/cc/0c/5c2b0a88158682aeafb10c1c2b735df5bc31f165bfe192f2ee9f2a23b5f1/h5py-3.12.1.tar.gz", hash = "sha256:326d70b53d31baa61f00b8aa5f95c2fcb9621a3ee8365d770c551a13dbbcbfdf", size = 411457, upload-time = "2024-09-26T16:41:39.883Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/df/7d/b21045fbb004ad8bb6fb3be4e6ca903841722706f7130b9bba31ef2f88e3/h5py-3.12.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:2f0f1a382cbf494679c07b4371f90c70391dedb027d517ac94fa2c05299dacda", size = 3402133, upload-time = "2024-09-26T16:39:27.937Z" }, + { url = "https://files.pythonhosted.org/packages/29/a7/3c2a33fba1da64a0846744726fd067a92fb8abb887875a0dd8e3bac8b45d/h5py-3.12.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:cb65f619dfbdd15e662423e8d257780f9a66677eae5b4b3fc9dca70b5fd2d2a3", size = 2866436, upload-time = "2024-09-26T16:39:32.495Z" }, + { url = "https://files.pythonhosted.org/packages/1e/d0/4bf67c3937a2437c20844165766ddd1a1817ae6b9544c3743050d8e0f403/h5py-3.12.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3b15d8dbd912c97541312c0e07438864d27dbca857c5ad634de68110c6beb1c2", size = 5168596, upload-time = "2024-09-26T16:39:39.107Z" }, + { url = "https://files.pythonhosted.org/packages/85/bc/e76f4b2096e0859225f5441d1b7f5e2041fffa19fc2c16756c67078417aa/h5py-3.12.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:59685fe40d8c1fbbee088c88cd4da415a2f8bee5c270337dc5a1c4aa634e3307", size = 5341537, upload-time = "2024-09-26T16:39:46.037Z" }, + { url = "https://files.pythonhosted.org/packages/99/bd/fb8ed45308bb97e04c02bd7aed324ba11e6a4bf9ed73967ca2a168e9cf92/h5py-3.12.1-cp310-cp310-win_amd64.whl", hash = "sha256:577d618d6b6dea3da07d13cc903ef9634cde5596b13e832476dd861aaf651f3e", size = 2990575, upload-time = "2024-09-26T16:39:50.903Z" }, + { url = "https://files.pythonhosted.org/packages/33/61/c463dc5fc02fbe019566d067a9d18746cd3c664f29c9b8b3c3f9ed025365/h5py-3.12.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ccd9006d92232727d23f784795191bfd02294a4f2ba68708825cb1da39511a93", size = 3410828, upload-time = "2024-09-26T16:39:56.19Z" }, + { url = "https://files.pythonhosted.org/packages/95/9d/eb91a9076aa998bb2179d6b1788055ea09cdf9d6619cd967f1d3321ed056/h5py-3.12.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:ad8a76557880aed5234cfe7279805f4ab5ce16b17954606cca90d578d3e713ef", size = 2872586, upload-time = "2024-09-26T16:40:00.204Z" }, + { url = "https://files.pythonhosted.org/packages/b0/62/e2b1f9723ff713e3bd3c16dfeceec7017eadc21ef063d8b7080c0fcdc58a/h5py-3.12.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1473348139b885393125126258ae2d70753ef7e9cec8e7848434f385ae72069e", size = 5273038, upload-time = "2024-09-26T16:40:06.444Z" }, + { url = "https://files.pythonhosted.org/packages/e1/89/118c3255d6ff2db33b062ec996a762d99ae50c21f54a8a6047ae8eda1b9f/h5py-3.12.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:018a4597f35092ae3fb28ee851fdc756d2b88c96336b8480e124ce1ac6fb9166", size = 5452688, upload-time = "2024-09-26T16:40:13.054Z" }, + { url = "https://files.pythonhosted.org/packages/1d/4d/cbd3014eb78d1e449b29beba1f3293a841aa8086c6f7968c383c2c7ff076/h5py-3.12.1-cp311-cp311-win_amd64.whl", hash = "sha256:3fdf95092d60e8130ba6ae0ef7a9bd4ade8edbe3569c13ebbaf39baefffc5ba4", size = 3006095, upload-time = "2024-09-26T16:40:17.822Z" }, + { url = "https://files.pythonhosted.org/packages/d4/e1/ea9bfe18a3075cdc873f0588ff26ce394726047653557876d7101bf0c74e/h5py-3.12.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:06a903a4e4e9e3ebbc8b548959c3c2552ca2d70dac14fcfa650d9261c66939ed", size = 3372538, upload-time = "2024-09-26T16:40:22.796Z" }, + { url = "https://files.pythonhosted.org/packages/0d/74/1009b663387c025e8fa5f3ee3cf3cd0d99b1ad5c72eeb70e75366b1ce878/h5py-3.12.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:7b3b8f3b48717e46c6a790e3128d39c61ab595ae0a7237f06dfad6a3b51d5351", size = 2868104, upload-time = "2024-09-26T16:40:26.817Z" }, + { url = "https://files.pythonhosted.org/packages/af/52/c604adc06280c15a29037d4aa79a24fe54d8d0b51085e81ed24b2fa995f7/h5py-3.12.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:050a4f2c9126054515169c49cb900949814987f0c7ae74c341b0c9f9b5056834", size = 5194606, upload-time = "2024-09-26T16:40:32.847Z" }, + { url = "https://files.pythonhosted.org/packages/fa/63/eeaacff417b393491beebabb8a3dc5342950409eb6d7b39d437289abdbae/h5py-3.12.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5c4b41d1019322a5afc5082864dfd6359f8935ecd37c11ac0029be78c5d112c9", size = 5413256, upload-time = "2024-09-26T16:40:39.188Z" }, + { url = "https://files.pythonhosted.org/packages/86/f7/bb465dcb92ca3521a15cbe1031f6d18234dbf1fb52a6796a00bfaa846ebf/h5py-3.12.1-cp312-cp312-win_amd64.whl", hash = "sha256:e4d51919110a030913201422fb07987db4338eba5ec8c5a15d6fab8e03d443fc", size = 2993055, upload-time = "2024-09-26T16:40:44.278Z" }, + { url = "https://files.pythonhosted.org/packages/23/1c/ecdd0efab52c24f2a9bf2324289828b860e8dd1e3c5ada3cf0889e14fdc1/h5py-3.12.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:513171e90ed92236fc2ca363ce7a2fc6f2827375efcbb0cc7fbdd7fe11fecafc", size = 3346239, upload-time = "2024-09-26T16:40:48.735Z" }, + { url = "https://files.pythonhosted.org/packages/93/cd/5b6f574bf3e318bbe305bc93ba45181676550eb44ba35e006d2e98004eaa/h5py-3.12.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:59400f88343b79655a242068a9c900001a34b63e3afb040bd7cdf717e440f653", size = 2843416, upload-time = "2024-09-26T16:40:53.424Z" }, + { url = "https://files.pythonhosted.org/packages/8a/4f/b74332f313bfbe94ba03fff784219b9db385e6139708e55b11490149f90a/h5py-3.12.1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d3e465aee0ec353949f0f46bf6c6f9790a2006af896cee7c178a8c3e5090aa32", size = 5154390, upload-time = "2024-09-26T16:40:59.787Z" }, + { url = "https://files.pythonhosted.org/packages/1a/57/93ea9e10a6457ea8d3b867207deb29a527e966a08a84c57ffd954e32152a/h5py-3.12.1-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ba51c0c5e029bb5420a343586ff79d56e7455d496d18a30309616fdbeed1068f", size = 5378244, upload-time = "2024-09-26T16:41:06.22Z" }, + { url = "https://files.pythonhosted.org/packages/50/51/0bbf3663062b2eeee78aa51da71e065f8a0a6e3cb950cc7020b4444999e6/h5py-3.12.1-cp313-cp313-win_amd64.whl", hash = "sha256:52ab036c6c97055b85b2a242cb540ff9590bacfda0c03dd0cf0661b311f522f8", size = 2979760, upload-time = "2024-09-26T16:41:10.425Z" }, +] + +[[package]] +name = "hf-xet" +version = "1.2.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/5e/6e/0f11bacf08a67f7fb5ee09740f2ca54163863b07b70d579356e9222ce5d8/hf_xet-1.2.0.tar.gz", hash = "sha256:a8c27070ca547293b6890c4bf389f713f80e8c478631432962bb7f4bc0bd7d7f", size = 506020, upload-time = "2025-10-24T19:04:32.129Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/9e/a5/85ef910a0aa034a2abcfadc360ab5ac6f6bc4e9112349bd40ca97551cff0/hf_xet-1.2.0-cp313-cp313t-macosx_10_12_x86_64.whl", hash = "sha256:ceeefcd1b7aed4956ae8499e2199607765fbd1c60510752003b6cc0b8413b649", size = 2861870, upload-time = "2025-10-24T19:04:11.422Z" }, + { url = "https://files.pythonhosted.org/packages/ea/40/e2e0a7eb9a51fe8828ba2d47fe22a7e74914ea8a0db68a18c3aa7449c767/hf_xet-1.2.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:b70218dd548e9840224df5638fdc94bd033552963cfa97f9170829381179c813", size = 2717584, upload-time = "2025-10-24T19:04:09.586Z" }, + { url = "https://files.pythonhosted.org/packages/a5/7d/daf7f8bc4594fdd59a8a596f9e3886133fdc68e675292218a5e4c1b7e834/hf_xet-1.2.0-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7d40b18769bb9a8bc82a9ede575ce1a44c75eb80e7375a01d76259089529b5dc", size = 3315004, upload-time = "2025-10-24T19:04:00.314Z" }, + { url = "https://files.pythonhosted.org/packages/b1/ba/45ea2f605fbf6d81c8b21e4d970b168b18a53515923010c312c06cd83164/hf_xet-1.2.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:cd3a6027d59cfb60177c12d6424e31f4b5ff13d8e3a1247b3a584bf8977e6df5", size = 3222636, upload-time = "2025-10-24T19:03:58.111Z" }, + { url = "https://files.pythonhosted.org/packages/4a/1d/04513e3cab8f29ab8c109d309ddd21a2705afab9d52f2ba1151e0c14f086/hf_xet-1.2.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:6de1fc44f58f6dd937956c8d304d8c2dea264c80680bcfa61ca4a15e7b76780f", size = 3408448, upload-time = "2025-10-24T19:04:20.951Z" }, + { url = "https://files.pythonhosted.org/packages/f0/7c/60a2756d7feec7387db3a1176c632357632fbe7849fce576c5559d4520c7/hf_xet-1.2.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:f182f264ed2acd566c514e45da9f2119110e48a87a327ca271027904c70c5832", size = 3503401, upload-time = "2025-10-24T19:04:22.549Z" }, + { url = "https://files.pythonhosted.org/packages/4e/64/48fffbd67fb418ab07451e4ce641a70de1c40c10a13e25325e24858ebe5a/hf_xet-1.2.0-cp313-cp313t-win_amd64.whl", hash = "sha256:293a7a3787e5c95d7be1857358a9130694a9c6021de3f27fa233f37267174382", size = 2900866, upload-time = "2025-10-24T19:04:33.461Z" }, + { url = "https://files.pythonhosted.org/packages/e2/51/f7e2caae42f80af886db414d4e9885fac959330509089f97cccb339c6b87/hf_xet-1.2.0-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:10bfab528b968c70e062607f663e21e34e2bba349e8038db546646875495179e", size = 2861861, upload-time = "2025-10-24T19:04:19.01Z" }, + { url = "https://files.pythonhosted.org/packages/6e/1d/a641a88b69994f9371bd347f1dd35e5d1e2e2460a2e350c8d5165fc62005/hf_xet-1.2.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:2a212e842647b02eb6a911187dc878e79c4aa0aa397e88dd3b26761676e8c1f8", size = 2717699, upload-time = "2025-10-24T19:04:17.306Z" }, + { url = "https://files.pythonhosted.org/packages/df/e0/e5e9bba7d15f0318955f7ec3f4af13f92e773fbb368c0b8008a5acbcb12f/hf_xet-1.2.0-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:30e06daccb3a7d4c065f34fc26c14c74f4653069bb2b194e7f18f17cbe9939c0", size = 3314885, upload-time = "2025-10-24T19:04:07.642Z" }, + { url = "https://files.pythonhosted.org/packages/21/90/b7fe5ff6f2b7b8cbdf1bd56145f863c90a5807d9758a549bf3d916aa4dec/hf_xet-1.2.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:29c8fc913a529ec0a91867ce3d119ac1aac966e098cf49501800c870328cc090", size = 3221550, upload-time = "2025-10-24T19:04:05.55Z" }, + { url = "https://files.pythonhosted.org/packages/6f/cb/73f276f0a7ce46cc6a6ec7d6c7d61cbfe5f2e107123d9bbd0193c355f106/hf_xet-1.2.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:66e159cbfcfbb29f920db2c09ed8b660eb894640d284f102ada929b6e3dc410a", size = 3408010, upload-time = "2025-10-24T19:04:28.598Z" }, + { url = "https://files.pythonhosted.org/packages/b8/1e/d642a12caa78171f4be64f7cd9c40e3ca5279d055d0873188a58c0f5fbb9/hf_xet-1.2.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:9c91d5ae931510107f148874e9e2de8a16052b6f1b3ca3c1b12f15ccb491390f", size = 3503264, upload-time = "2025-10-24T19:04:30.397Z" }, + { url = "https://files.pythonhosted.org/packages/17/b5/33764714923fa1ff922770f7ed18c2daae034d21ae6e10dbf4347c854154/hf_xet-1.2.0-cp314-cp314t-win_amd64.whl", hash = "sha256:210d577732b519ac6ede149d2f2f34049d44e8622bf14eb3d63bbcd2d4b332dc", size = 2901071, upload-time = "2025-10-24T19:04:37.463Z" }, + { url = "https://files.pythonhosted.org/packages/96/2d/22338486473df5923a9ab7107d375dbef9173c338ebef5098ef593d2b560/hf_xet-1.2.0-cp37-abi3-macosx_10_12_x86_64.whl", hash = "sha256:46740d4ac024a7ca9b22bebf77460ff43332868b661186a8e46c227fdae01848", size = 2866099, upload-time = "2025-10-24T19:04:15.366Z" }, + { url = "https://files.pythonhosted.org/packages/7f/8c/c5becfa53234299bc2210ba314eaaae36c2875e0045809b82e40a9544f0c/hf_xet-1.2.0-cp37-abi3-macosx_11_0_arm64.whl", hash = "sha256:27df617a076420d8845bea087f59303da8be17ed7ec0cd7ee3b9b9f579dff0e4", size = 2722178, upload-time = "2025-10-24T19:04:13.695Z" }, + { url = "https://files.pythonhosted.org/packages/9a/92/cf3ab0b652b082e66876d08da57fcc6fa2f0e6c70dfbbafbd470bb73eb47/hf_xet-1.2.0-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3651fd5bfe0281951b988c0facbe726aa5e347b103a675f49a3fa8144c7968fd", size = 3320214, upload-time = "2025-10-24T19:04:03.596Z" }, + { url = "https://files.pythonhosted.org/packages/46/92/3f7ec4a1b6a65bf45b059b6d4a5d38988f63e193056de2f420137e3c3244/hf_xet-1.2.0-cp37-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:d06fa97c8562fb3ee7a378dd9b51e343bc5bc8190254202c9771029152f5e08c", size = 3229054, upload-time = "2025-10-24T19:04:01.949Z" }, + { url = "https://files.pythonhosted.org/packages/0b/dd/7ac658d54b9fb7999a0ccb07ad863b413cbaf5cf172f48ebcd9497ec7263/hf_xet-1.2.0-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:4c1428c9ae73ec0939410ec73023c4f842927f39db09b063b9482dac5a3bb737", size = 3413812, upload-time = "2025-10-24T19:04:24.585Z" }, + { url = "https://files.pythonhosted.org/packages/92/68/89ac4e5b12a9ff6286a12174c8538a5930e2ed662091dd2572bbe0a18c8a/hf_xet-1.2.0-cp37-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:a55558084c16b09b5ed32ab9ed38421e2d87cf3f1f89815764d1177081b99865", size = 3508920, upload-time = "2025-10-24T19:04:26.927Z" }, + { url = "https://files.pythonhosted.org/packages/cb/44/870d44b30e1dcfb6a65932e3e1506c103a8a5aea9103c337e7a53180322c/hf_xet-1.2.0-cp37-abi3-win_amd64.whl", hash = "sha256:e6584a52253f72c9f52f9e549d5895ca7a471608495c4ecaa6cc73dba2b24d69", size = 2905735, upload-time = "2025-10-24T19:04:35.928Z" }, +] + +[[package]] +name = "httpcore" +version = "1.0.9" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "certifi" }, + { name = "h11" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/06/94/82699a10bca87a5556c9c59b5963f2d039dbd239f25bc2a63907a05a14cb/httpcore-1.0.9.tar.gz", hash = "sha256:6e34463af53fd2ab5d807f399a9b45ea31c3dfa2276f15a2c3f00afff6e176e8", size = 85484, upload-time = "2025-04-24T22:06:22.219Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/7e/f5/f66802a942d491edb555dd61e3a9961140fd64c90bce1eafd741609d334d/httpcore-1.0.9-py3-none-any.whl", hash = "sha256:2d400746a40668fc9dec9810239072b40b4484b640a8c38fd654a024c7a1bf55", size = 78784, upload-time = "2025-04-24T22:06:20.566Z" }, +] + +[[package]] +name = "httpx" +version = "0.28.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "anyio" }, + { name = "certifi" }, + { name = "httpcore" }, + { name = "idna" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/b1/df/48c586a5fe32a0f01324ee087459e112ebb7224f646c0b5023f5e79e9956/httpx-0.28.1.tar.gz", hash = "sha256:75e98c5f16b0f35b567856f597f06ff2270a374470a5c2392242528e3e3e42fc", size = 141406, upload-time = "2024-12-06T15:37:23.222Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/2a/39/e50c7c3a983047577ee07d2a9e53faf5a69493943ec3f6a384bdc792deb2/httpx-0.28.1-py3-none-any.whl", hash = "sha256:d909fcccc110f8c7faf814ca82a9a4d816bc5a6dbfea25d6591d6985b8ba59ad", size = 73517, upload-time = "2024-12-06T15:37:21.509Z" }, +] + +[[package]] +name = "huggingface-hub" +version = "0.36.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "filelock" }, + { name = "fsspec" }, + { name = "hf-xet", marker = "platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64'" }, + { name = "packaging" }, + { name = "pyyaml" }, + { name = "requests" }, + { name = "tqdm" }, + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/98/63/4910c5fa9128fdadf6a9c5ac138e8b1b6cee4ca44bf7915bbfbce4e355ee/huggingface_hub-0.36.0.tar.gz", hash = "sha256:47b3f0e2539c39bf5cde015d63b72ec49baff67b6931c3d97f3f84532e2b8d25", size = 463358, upload-time = "2025-10-23T12:12:01.413Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/cb/bd/1a875e0d592d447cbc02805fd3fe0f497714d6a2583f59d14fa9ebad96eb/huggingface_hub-0.36.0-py3-none-any.whl", hash = "sha256:7bcc9ad17d5b3f07b57c78e79d527102d08313caa278a641993acddcb894548d", size = 566094, upload-time = "2025-10-23T12:11:59.557Z" }, +] + +[[package]] +name = "identify" +version = "2.6.15" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/ff/e7/685de97986c916a6d93b3876139e00eef26ad5bbbd61925d670ae8013449/identify-2.6.15.tar.gz", hash = "sha256:e4f4864b96c6557ef2a1e1c951771838f4edc9df3a72ec7118b338801b11c7bf", size = 99311, upload-time = "2025-10-02T17:43:40.631Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/0f/1c/e5fd8f973d4f375adb21565739498e2e9a1e54c858a97b9a8ccfdc81da9b/identify-2.6.15-py2.py3-none-any.whl", hash = "sha256:1181ef7608e00704db228516541eb83a88a9f94433a8c80bb9b5bd54b1d81757", size = 99183, upload-time = "2025-10-02T17:43:39.137Z" }, +] + +[[package]] +name = "idna" +version = "3.11" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/6f/6d/0703ccc57f3a7233505399edb88de3cbd678da106337b9fcde432b65ed60/idna-3.11.tar.gz", hash = "sha256:795dafcc9c04ed0c1fb032c2aa73654d8e8c5023a7df64a53f39190ada629902", size = 194582, upload-time = "2025-10-12T14:55:20.501Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/0e/61/66938bbb5fc52dbdf84594873d5b51fb1f7c7794e9c0f5bd885f30bc507b/idna-3.11-py3-none-any.whl", hash = "sha256:771a87f49d9defaf64091e6e6fe9c18d4833f140bd19464795bc32d966ca37ea", size = 71008, upload-time = "2025-10-12T14:55:18.883Z" }, +] + +[[package]] +name = "imageio" +version = "2.36.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "numpy" }, + { name = "pillow" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/4f/34/a714fd354f5f7fe650477072d4da21446849b20c02045dcf7ac827495121/imageio-2.36.0.tar.gz", hash = "sha256:1c8f294db862c256e9562354d65aa54725b8dafed7f10f02bb3ec20ec1678850", size = 389492, upload-time = "2024-10-14T02:49:01.896Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/4e/e7/26045404a30c8a200e960fb54fbaf4b73d12e58cd28e03b306b084253f4f/imageio-2.36.0-py3-none-any.whl", hash = "sha256:471f1eda55618ee44a3c9960911c35e647d9284c68f077e868df633398f137f0", size = 315414, upload-time = "2024-10-14T02:48:59.564Z" }, +] + +[[package]] +name = "imageio-ffmpeg" +version = "0.5.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "setuptools" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/73/22/b0a0d96ecdbd4a8493c6cd7914a8b2bfbc39f8660f81c20e3bde847182e0/imageio-ffmpeg-0.5.1.tar.gz", hash = "sha256:0ed7a9b31f560b0c9d929c5291cd430edeb9bed3ce9a497480e536dd4326484c", size = 17704, upload-time = "2024-06-03T15:12:23.284Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/be/3d/df5dc571520f495ba2152215cd26deebd46e1530eae0261f503bfd137e99/imageio_ffmpeg-0.5.1-py3-none-macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl", hash = "sha256:1460e84712b9d06910c1f7bb524096b0341d4b7844cea6c20e099d0a24e795b1", size = 22532925, upload-time = "2024-06-03T15:12:06.451Z" }, + { url = "https://files.pythonhosted.org/packages/fe/44/c9e18a73dfd939b5b0ec843870ac72f1f6c17e31a03149b687a85465bff7/imageio_ffmpeg-0.5.1-py3-none-manylinux2010_x86_64.whl", hash = "sha256:5289f75c7f755b499653f3209fea4efd1430cba0e39831c381aad2d458f7a316", size = 26900394, upload-time = "2024-06-03T15:12:16.376Z" }, + { url = "https://files.pythonhosted.org/packages/cd/ca/8537cdbf1a6852912cb293fa23dc7adf256cec793113485447f0cbf0fe79/imageio_ffmpeg-0.5.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:7fa9132a291d5eb28c44553550deb40cbdab831f2a614e55360301a6582eb205", size = 22880461, upload-time = "2024-06-03T15:12:20.744Z" }, + { url = "https://files.pythonhosted.org/packages/a9/97/ff7de8ace4425fffc6e8d7646017500c9d5435df608b13cc34de4835ad4f/imageio_ffmpeg-0.5.1-py3-none-win32.whl", hash = "sha256:89efe2c79979d8174ba8476deb7f74d74c331caee3fb2b65ba2883bec0737625", size = 19652102, upload-time = "2024-06-03T15:11:59.16Z" }, + { url = "https://files.pythonhosted.org/packages/a9/1c/1b9c72bf839def47626436ea5ebaf643404f7850482c5fafd71a3deeaa94/imageio_ffmpeg-0.5.1-py3-none-win_amd64.whl", hash = "sha256:1521e79e253bedbdd36a547e0cbd94a025ba0b558e17f08fea687d805a0e4698", size = 22619891, upload-time = "2024-06-03T15:12:11.248Z" }, +] + +[[package]] +name = "importlib-metadata" +version = "8.7.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "zipp" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/76/66/650a33bd90f786193e4de4b3ad86ea60b53c89b669a5c7be931fac31cdb0/importlib_metadata-8.7.0.tar.gz", hash = "sha256:d13b81ad223b890aa16c5471f2ac3056cf76c5f10f82d6f9292f0b415f389000", size = 56641, upload-time = "2025-04-27T15:29:01.736Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/20/b0/36bd937216ec521246249be3bf9855081de4c5e06a0c9b4219dbeda50373/importlib_metadata-8.7.0-py3-none-any.whl", hash = "sha256:e5dd1551894c77868a30651cef00984d50e1002d06942a7101d34870c5f02afd", size = 27656, upload-time = "2025-04-27T15:29:00.214Z" }, +] + +[[package]] +name = "iniconfig" +version = "2.3.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/72/34/14ca021ce8e5dfedc35312d08ba8bf51fdd999c576889fc2c24cb97f4f10/iniconfig-2.3.0.tar.gz", hash = "sha256:c76315c77db068650d49c5b56314774a7804df16fee4402c1f19d6d15d8c4730", size = 20503, upload-time = "2025-10-18T21:55:43.219Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/cb/b1/3846dd7f199d53cb17f49cba7e651e9ce294d8497c8c150530ed11865bb8/iniconfig-2.3.0-py3-none-any.whl", hash = "sha256:f631c04d2c48c52b84d0d0549c99ff3859c98df65b3101406327ecc7d53fbf12", size = 7484, upload-time = "2025-10-18T21:55:41.639Z" }, +] + +[[package]] +name = "itsdangerous" +version = "2.2.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/9c/cb/8ac0172223afbccb63986cc25049b154ecfb5e85932587206f42317be31d/itsdangerous-2.2.0.tar.gz", hash = "sha256:e0050c0b7da1eea53ffaf149c0cfbb5c6e2e2b69c4bef22c81fa6eb73e5f6173", size = 54410, upload-time = "2024-04-16T21:28:15.614Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/04/96/92447566d16df59b2a776c0fb82dbc4d9e07cd95062562af01e408583fc4/itsdangerous-2.2.0-py3-none-any.whl", hash = "sha256:c6242fc49e35958c8b15141343aa660db5fc54d4f13a1db01a3f5891b98700ef", size = 16234, upload-time = "2024-04-16T21:28:14.499Z" }, +] + +[[package]] +name = "jinja2" +version = "3.1.6" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "markupsafe" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/df/bf/f7da0350254c0ed7c72f3e33cef02e048281fec7ecec5f032d4aac52226b/jinja2-3.1.6.tar.gz", hash = "sha256:0137fb05990d35f1275a587e9aee6d56da821fc83491a0fb838183be43f66d6d", size = 245115, upload-time = "2025-03-05T20:05:02.478Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/62/a1/3d680cbfd5f4b8f15abc1d571870c5fc3e594bb582bc3b64ea099db13e56/jinja2-3.1.6-py3-none-any.whl", hash = "sha256:85ece4451f492d0c13c5dd7c13a64681a86afae63a5f347908daf103ce6d2f67", size = 134899, upload-time = "2025-03-05T20:05:00.369Z" }, +] + +[[package]] +name = "jinxed" +version = "1.3.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "ansicon", marker = "sys_platform == 'win32'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/20/d0/59b2b80e7a52d255f9e0ad040d2e826342d05580c4b1d7d7747cfb8db731/jinxed-1.3.0.tar.gz", hash = "sha256:1593124b18a41b7a3da3b078471442e51dbad3d77b4d4f2b0c26ab6f7d660dbf", size = 80981, upload-time = "2024-07-31T22:39:18.854Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/27/e3/0e0014d6ab159d48189e92044ace13b1e1fe9aa3024ba9f4e8cf172aa7c2/jinxed-1.3.0-py2.py3-none-any.whl", hash = "sha256:b993189f39dc2d7504d802152671535b06d380b26d78070559551cbf92df4fc5", size = 33085, upload-time = "2024-07-31T22:39:17.426Z" }, +] + +[[package]] +name = "jsonschema" +version = "4.26.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "attrs" }, + { name = "jsonschema-specifications" }, + { name = "referencing" }, + { name = "rpds-py" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/b3/fc/e067678238fa451312d4c62bf6e6cf5ec56375422aee02f9cb5f909b3047/jsonschema-4.26.0.tar.gz", hash = "sha256:0c26707e2efad8aa1bfc5b7ce170f3fccc2e4918ff85989ba9ffa9facb2be326", size = 366583, upload-time = "2026-01-07T13:41:07.246Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/69/90/f63fb5873511e014207a475e2bb4e8b2e570d655b00ac19a9a0ca0a385ee/jsonschema-4.26.0-py3-none-any.whl", hash = "sha256:d489f15263b8d200f8387e64b4c3a75f06629559fb73deb8fdfb525f2dab50ce", size = 90630, upload-time = "2026-01-07T13:41:05.306Z" }, +] + +[[package]] +name = "jsonschema-specifications" +version = "2025.9.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "referencing" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/19/74/a633ee74eb36c44aa6d1095e7cc5569bebf04342ee146178e2d36600708b/jsonschema_specifications-2025.9.1.tar.gz", hash = "sha256:b540987f239e745613c7a9176f3edb72b832a4ac465cf02712288397832b5e8d", size = 32855, upload-time = "2025-09-08T01:34:59.186Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/41/45/1a4ed80516f02155c51f51e8cedb3c1902296743db0bbc66608a0db2814f/jsonschema_specifications-2025.9.1-py3-none-any.whl", hash = "sha256:98802fee3a11ee76ecaca44429fda8a41bff98b00a0f2838151b113f210cc6fe", size = 18437, upload-time = "2025-09-08T01:34:57.871Z" }, +] + +[[package]] +name = "loguru" +version = "0.7.3" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "colorama", marker = "sys_platform == 'win32'" }, + { name = "win32-setctime", marker = "sys_platform == 'win32'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/3a/05/a1dae3dffd1116099471c643b8924f5aa6524411dc6c63fdae648c4f1aca/loguru-0.7.3.tar.gz", hash = "sha256:19480589e77d47b8d85b2c827ad95d49bf31b0dcde16593892eb51dd18706eb6", size = 63559, upload-time = "2024-12-06T11:20:56.608Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/0c/29/0348de65b8cc732daa3e33e67806420b2ae89bdce2b04af740289c5c6c8c/loguru-0.7.3-py3-none-any.whl", hash = "sha256:31a33c10c8e1e10422bfd431aeb5d351c7cf7fa671e3c4df004162264b28220c", size = 61595, upload-time = "2024-12-06T11:20:54.538Z" }, +] + +[[package]] +name = "markdown" +version = "3.10" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/7d/ab/7dd27d9d863b3376fcf23a5a13cb5d024aed1db46f963f1b5735ae43b3be/markdown-3.10.tar.gz", hash = "sha256:37062d4f2aa4b2b6b32aefb80faa300f82cc790cb949a35b8caede34f2b68c0e", size = 364931, upload-time = "2025-11-03T19:51:15.007Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/70/81/54e3ce63502cd085a0c556652a4e1b919c45a446bd1e5300e10c44c8c521/markdown-3.10-py3-none-any.whl", hash = "sha256:b5b99d6951e2e4948d939255596523444c0e677c669700b1d17aa4a8a464cb7c", size = 107678, upload-time = "2025-11-03T19:51:13.887Z" }, +] + +[[package]] +name = "markdown-it-py" +version = "4.0.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "mdurl" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/5b/f5/4ec618ed16cc4f8fb3b701563655a69816155e79e24a17b651541804721d/markdown_it_py-4.0.0.tar.gz", hash = "sha256:cb0a2b4aa34f932c007117b194e945bd74e0ec24133ceb5bac59009cda1cb9f3", size = 73070, upload-time = "2025-08-11T12:57:52.854Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/94/54/e7d793b573f298e1c9013b8c4dade17d481164aa517d1d7148619c2cedbf/markdown_it_py-4.0.0-py3-none-any.whl", hash = "sha256:87327c59b172c5011896038353a81343b6754500a08cd7a4973bb48c6d578147", size = 87321, upload-time = "2025-08-11T12:57:51.923Z" }, +] + +[[package]] +name = "markupsafe" +version = "3.0.3" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/7e/99/7690b6d4034fffd95959cbe0c02de8deb3098cc577c67bb6a24fe5d7caa7/markupsafe-3.0.3.tar.gz", hash = "sha256:722695808f4b6457b320fdc131280796bdceb04ab50fe1795cd540799ebe1698", size = 80313, upload-time = "2025-09-27T18:37:40.426Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/e8/4b/3541d44f3937ba468b75da9eebcae497dcf67adb65caa16760b0a6807ebb/markupsafe-3.0.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:2f981d352f04553a7171b8e44369f2af4055f888dfb147d55e42d29e29e74559", size = 11631, upload-time = "2025-09-27T18:36:05.558Z" }, + { url = "https://files.pythonhosted.org/packages/98/1b/fbd8eed11021cabd9226c37342fa6ca4e8a98d8188a8d9b66740494960e4/markupsafe-3.0.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:e1c1493fb6e50ab01d20a22826e57520f1284df32f2d8601fdd90b6304601419", size = 12057, upload-time = "2025-09-27T18:36:07.165Z" }, + { url = "https://files.pythonhosted.org/packages/40/01/e560d658dc0bb8ab762670ece35281dec7b6c1b33f5fbc09ebb57a185519/markupsafe-3.0.3-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1ba88449deb3de88bd40044603fafffb7bc2b055d626a330323a9ed736661695", size = 22050, upload-time = "2025-09-27T18:36:08.005Z" }, + { url = "https://files.pythonhosted.org/packages/af/cd/ce6e848bbf2c32314c9b237839119c5a564a59725b53157c856e90937b7a/markupsafe-3.0.3-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f42d0984e947b8adf7dd6dde396e720934d12c506ce84eea8476409563607591", size = 20681, upload-time = "2025-09-27T18:36:08.881Z" }, + { url = "https://files.pythonhosted.org/packages/c9/2a/b5c12c809f1c3045c4d580b035a743d12fcde53cf685dbc44660826308da/markupsafe-3.0.3-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c0c0b3ade1c0b13b936d7970b1d37a57acde9199dc2aecc4c336773e1d86049c", size = 20705, upload-time = "2025-09-27T18:36:10.131Z" }, + { url = "https://files.pythonhosted.org/packages/cf/e3/9427a68c82728d0a88c50f890d0fc072a1484de2f3ac1ad0bfc1a7214fd5/markupsafe-3.0.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:0303439a41979d9e74d18ff5e2dd8c43ed6c6001fd40e5bf2e43f7bd9bbc523f", size = 21524, upload-time = "2025-09-27T18:36:11.324Z" }, + { url = "https://files.pythonhosted.org/packages/bc/36/23578f29e9e582a4d0278e009b38081dbe363c5e7165113fad546918a232/markupsafe-3.0.3-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:d2ee202e79d8ed691ceebae8e0486bd9a2cd4794cec4824e1c99b6f5009502f6", size = 20282, upload-time = "2025-09-27T18:36:12.573Z" }, + { url = "https://files.pythonhosted.org/packages/56/21/dca11354e756ebd03e036bd8ad58d6d7168c80ce1fe5e75218e4945cbab7/markupsafe-3.0.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:177b5253b2834fe3678cb4a5f0059808258584c559193998be2601324fdeafb1", size = 20745, upload-time = "2025-09-27T18:36:13.504Z" }, + { url = "https://files.pythonhosted.org/packages/87/99/faba9369a7ad6e4d10b6a5fbf71fa2a188fe4a593b15f0963b73859a1bbd/markupsafe-3.0.3-cp310-cp310-win32.whl", hash = "sha256:2a15a08b17dd94c53a1da0438822d70ebcd13f8c3a95abe3a9ef9f11a94830aa", size = 14571, upload-time = "2025-09-27T18:36:14.779Z" }, + { url = "https://files.pythonhosted.org/packages/d6/25/55dc3ab959917602c96985cb1253efaa4ff42f71194bddeb61eb7278b8be/markupsafe-3.0.3-cp310-cp310-win_amd64.whl", hash = "sha256:c4ffb7ebf07cfe8931028e3e4c85f0357459a3f9f9490886198848f4fa002ec8", size = 15056, upload-time = "2025-09-27T18:36:16.125Z" }, + { url = "https://files.pythonhosted.org/packages/d0/9e/0a02226640c255d1da0b8d12e24ac2aa6734da68bff14c05dd53b94a0fc3/markupsafe-3.0.3-cp310-cp310-win_arm64.whl", hash = "sha256:e2103a929dfa2fcaf9bb4e7c091983a49c9ac3b19c9061b6d5427dd7d14d81a1", size = 13932, upload-time = "2025-09-27T18:36:17.311Z" }, + { url = "https://files.pythonhosted.org/packages/08/db/fefacb2136439fc8dd20e797950e749aa1f4997ed584c62cfb8ef7c2be0e/markupsafe-3.0.3-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:1cc7ea17a6824959616c525620e387f6dd30fec8cb44f649e31712db02123dad", size = 11631, upload-time = "2025-09-27T18:36:18.185Z" }, + { url = "https://files.pythonhosted.org/packages/e1/2e/5898933336b61975ce9dc04decbc0a7f2fee78c30353c5efba7f2d6ff27a/markupsafe-3.0.3-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:4bd4cd07944443f5a265608cc6aab442e4f74dff8088b0dfc8238647b8f6ae9a", size = 12058, upload-time = "2025-09-27T18:36:19.444Z" }, + { url = "https://files.pythonhosted.org/packages/1d/09/adf2df3699d87d1d8184038df46a9c80d78c0148492323f4693df54e17bb/markupsafe-3.0.3-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6b5420a1d9450023228968e7e6a9ce57f65d148ab56d2313fcd589eee96a7a50", size = 24287, upload-time = "2025-09-27T18:36:20.768Z" }, + { url = "https://files.pythonhosted.org/packages/30/ac/0273f6fcb5f42e314c6d8cd99effae6a5354604d461b8d392b5ec9530a54/markupsafe-3.0.3-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0bf2a864d67e76e5c9a34dc26ec616a66b9888e25e7b9460e1c76d3293bd9dbf", size = 22940, upload-time = "2025-09-27T18:36:22.249Z" }, + { url = "https://files.pythonhosted.org/packages/19/ae/31c1be199ef767124c042c6c3e904da327a2f7f0cd63a0337e1eca2967a8/markupsafe-3.0.3-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:bc51efed119bc9cfdf792cdeaa4d67e8f6fcccab66ed4bfdd6bde3e59bfcbb2f", size = 21887, upload-time = "2025-09-27T18:36:23.535Z" }, + { url = "https://files.pythonhosted.org/packages/b2/76/7edcab99d5349a4532a459e1fe64f0b0467a3365056ae550d3bcf3f79e1e/markupsafe-3.0.3-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:068f375c472b3e7acbe2d5318dea141359e6900156b5b2ba06a30b169086b91a", size = 23692, upload-time = "2025-09-27T18:36:24.823Z" }, + { url = "https://files.pythonhosted.org/packages/a4/28/6e74cdd26d7514849143d69f0bf2399f929c37dc2b31e6829fd2045b2765/markupsafe-3.0.3-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:7be7b61bb172e1ed687f1754f8e7484f1c8019780f6f6b0786e76bb01c2ae115", size = 21471, upload-time = "2025-09-27T18:36:25.95Z" }, + { url = "https://files.pythonhosted.org/packages/62/7e/a145f36a5c2945673e590850a6f8014318d5577ed7e5920a4b3448e0865d/markupsafe-3.0.3-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:f9e130248f4462aaa8e2552d547f36ddadbeaa573879158d721bbd33dfe4743a", size = 22923, upload-time = "2025-09-27T18:36:27.109Z" }, + { url = "https://files.pythonhosted.org/packages/0f/62/d9c46a7f5c9adbeeeda52f5b8d802e1094e9717705a645efc71b0913a0a8/markupsafe-3.0.3-cp311-cp311-win32.whl", hash = "sha256:0db14f5dafddbb6d9208827849fad01f1a2609380add406671a26386cdf15a19", size = 14572, upload-time = "2025-09-27T18:36:28.045Z" }, + { url = "https://files.pythonhosted.org/packages/83/8a/4414c03d3f891739326e1783338e48fb49781cc915b2e0ee052aa490d586/markupsafe-3.0.3-cp311-cp311-win_amd64.whl", hash = "sha256:de8a88e63464af587c950061a5e6a67d3632e36df62b986892331d4620a35c01", size = 15077, upload-time = "2025-09-27T18:36:29.025Z" }, + { url = "https://files.pythonhosted.org/packages/35/73/893072b42e6862f319b5207adc9ae06070f095b358655f077f69a35601f0/markupsafe-3.0.3-cp311-cp311-win_arm64.whl", hash = "sha256:3b562dd9e9ea93f13d53989d23a7e775fdfd1066c33494ff43f5418bc8c58a5c", size = 13876, upload-time = "2025-09-27T18:36:29.954Z" }, + { url = "https://files.pythonhosted.org/packages/5a/72/147da192e38635ada20e0a2e1a51cf8823d2119ce8883f7053879c2199b5/markupsafe-3.0.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:d53197da72cc091b024dd97249dfc7794d6a56530370992a5e1a08983ad9230e", size = 11615, upload-time = "2025-09-27T18:36:30.854Z" }, + { url = "https://files.pythonhosted.org/packages/9a/81/7e4e08678a1f98521201c3079f77db69fb552acd56067661f8c2f534a718/markupsafe-3.0.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:1872df69a4de6aead3491198eaf13810b565bdbeec3ae2dc8780f14458ec73ce", size = 12020, upload-time = "2025-09-27T18:36:31.971Z" }, + { url = "https://files.pythonhosted.org/packages/1e/2c/799f4742efc39633a1b54a92eec4082e4f815314869865d876824c257c1e/markupsafe-3.0.3-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3a7e8ae81ae39e62a41ec302f972ba6ae23a5c5396c8e60113e9066ef893da0d", size = 24332, upload-time = "2025-09-27T18:36:32.813Z" }, + { url = "https://files.pythonhosted.org/packages/3c/2e/8d0c2ab90a8c1d9a24f0399058ab8519a3279d1bd4289511d74e909f060e/markupsafe-3.0.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d6dd0be5b5b189d31db7cda48b91d7e0a9795f31430b7f271219ab30f1d3ac9d", size = 22947, upload-time = "2025-09-27T18:36:33.86Z" }, + { url = "https://files.pythonhosted.org/packages/2c/54/887f3092a85238093a0b2154bd629c89444f395618842e8b0c41783898ea/markupsafe-3.0.3-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:94c6f0bb423f739146aec64595853541634bde58b2135f27f61c1ffd1cd4d16a", size = 21962, upload-time = "2025-09-27T18:36:35.099Z" }, + { url = "https://files.pythonhosted.org/packages/c9/2f/336b8c7b6f4a4d95e91119dc8521402461b74a485558d8f238a68312f11c/markupsafe-3.0.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:be8813b57049a7dc738189df53d69395eba14fb99345e0a5994914a3864c8a4b", size = 23760, upload-time = "2025-09-27T18:36:36.001Z" }, + { url = "https://files.pythonhosted.org/packages/32/43/67935f2b7e4982ffb50a4d169b724d74b62a3964bc1a9a527f5ac4f1ee2b/markupsafe-3.0.3-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:83891d0e9fb81a825d9a6d61e3f07550ca70a076484292a70fde82c4b807286f", size = 21529, upload-time = "2025-09-27T18:36:36.906Z" }, + { url = "https://files.pythonhosted.org/packages/89/e0/4486f11e51bbba8b0c041098859e869e304d1c261e59244baa3d295d47b7/markupsafe-3.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:77f0643abe7495da77fb436f50f8dab76dbc6e5fd25d39589a0f1fe6548bfa2b", size = 23015, upload-time = "2025-09-27T18:36:37.868Z" }, + { url = "https://files.pythonhosted.org/packages/2f/e1/78ee7a023dac597a5825441ebd17170785a9dab23de95d2c7508ade94e0e/markupsafe-3.0.3-cp312-cp312-win32.whl", hash = "sha256:d88b440e37a16e651bda4c7c2b930eb586fd15ca7406cb39e211fcff3bf3017d", size = 14540, upload-time = "2025-09-27T18:36:38.761Z" }, + { url = "https://files.pythonhosted.org/packages/aa/5b/bec5aa9bbbb2c946ca2733ef9c4ca91c91b6a24580193e891b5f7dbe8e1e/markupsafe-3.0.3-cp312-cp312-win_amd64.whl", hash = "sha256:26a5784ded40c9e318cfc2bdb30fe164bdb8665ded9cd64d500a34fb42067b1c", size = 15105, upload-time = "2025-09-27T18:36:39.701Z" }, + { url = "https://files.pythonhosted.org/packages/e5/f1/216fc1bbfd74011693a4fd837e7026152e89c4bcf3e77b6692fba9923123/markupsafe-3.0.3-cp312-cp312-win_arm64.whl", hash = "sha256:35add3b638a5d900e807944a078b51922212fb3dedb01633a8defc4b01a3c85f", size = 13906, upload-time = "2025-09-27T18:36:40.689Z" }, + { url = "https://files.pythonhosted.org/packages/38/2f/907b9c7bbba283e68f20259574b13d005c121a0fa4c175f9bed27c4597ff/markupsafe-3.0.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:e1cf1972137e83c5d4c136c43ced9ac51d0e124706ee1c8aa8532c1287fa8795", size = 11622, upload-time = "2025-09-27T18:36:41.777Z" }, + { url = "https://files.pythonhosted.org/packages/9c/d9/5f7756922cdd676869eca1c4e3c0cd0df60ed30199ffd775e319089cb3ed/markupsafe-3.0.3-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:116bb52f642a37c115f517494ea5feb03889e04df47eeff5b130b1808ce7c219", size = 12029, upload-time = "2025-09-27T18:36:43.257Z" }, + { url = "https://files.pythonhosted.org/packages/00/07/575a68c754943058c78f30db02ee03a64b3c638586fba6a6dd56830b30a3/markupsafe-3.0.3-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:133a43e73a802c5562be9bbcd03d090aa5a1fe899db609c29e8c8d815c5f6de6", size = 24374, upload-time = "2025-09-27T18:36:44.508Z" }, + { url = "https://files.pythonhosted.org/packages/a9/21/9b05698b46f218fc0e118e1f8168395c65c8a2c750ae2bab54fc4bd4e0e8/markupsafe-3.0.3-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ccfcd093f13f0f0b7fdd0f198b90053bf7b2f02a3927a30e63f3ccc9df56b676", size = 22980, upload-time = "2025-09-27T18:36:45.385Z" }, + { url = "https://files.pythonhosted.org/packages/7f/71/544260864f893f18b6827315b988c146b559391e6e7e8f7252839b1b846a/markupsafe-3.0.3-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:509fa21c6deb7a7a273d629cf5ec029bc209d1a51178615ddf718f5918992ab9", size = 21990, upload-time = "2025-09-27T18:36:46.916Z" }, + { url = "https://files.pythonhosted.org/packages/c2/28/b50fc2f74d1ad761af2f5dcce7492648b983d00a65b8c0e0cb457c82ebbe/markupsafe-3.0.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:a4afe79fb3de0b7097d81da19090f4df4f8d3a2b3adaa8764138aac2e44f3af1", size = 23784, upload-time = "2025-09-27T18:36:47.884Z" }, + { url = "https://files.pythonhosted.org/packages/ed/76/104b2aa106a208da8b17a2fb72e033a5a9d7073c68f7e508b94916ed47a9/markupsafe-3.0.3-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:795e7751525cae078558e679d646ae45574b47ed6e7771863fcc079a6171a0fc", size = 21588, upload-time = "2025-09-27T18:36:48.82Z" }, + { url = "https://files.pythonhosted.org/packages/b5/99/16a5eb2d140087ebd97180d95249b00a03aa87e29cc224056274f2e45fd6/markupsafe-3.0.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:8485f406a96febb5140bfeca44a73e3ce5116b2501ac54fe953e488fb1d03b12", size = 23041, upload-time = "2025-09-27T18:36:49.797Z" }, + { url = "https://files.pythonhosted.org/packages/19/bc/e7140ed90c5d61d77cea142eed9f9c303f4c4806f60a1044c13e3f1471d0/markupsafe-3.0.3-cp313-cp313-win32.whl", hash = "sha256:bdd37121970bfd8be76c5fb069c7751683bdf373db1ed6c010162b2a130248ed", size = 14543, upload-time = "2025-09-27T18:36:51.584Z" }, + { url = "https://files.pythonhosted.org/packages/05/73/c4abe620b841b6b791f2edc248f556900667a5a1cf023a6646967ae98335/markupsafe-3.0.3-cp313-cp313-win_amd64.whl", hash = "sha256:9a1abfdc021a164803f4d485104931fb8f8c1efd55bc6b748d2f5774e78b62c5", size = 15113, upload-time = "2025-09-27T18:36:52.537Z" }, + { url = "https://files.pythonhosted.org/packages/f0/3a/fa34a0f7cfef23cf9500d68cb7c32dd64ffd58a12b09225fb03dd37d5b80/markupsafe-3.0.3-cp313-cp313-win_arm64.whl", hash = "sha256:7e68f88e5b8799aa49c85cd116c932a1ac15caaa3f5db09087854d218359e485", size = 13911, upload-time = "2025-09-27T18:36:53.513Z" }, + { url = "https://files.pythonhosted.org/packages/e4/d7/e05cd7efe43a88a17a37b3ae96e79a19e846f3f456fe79c57ca61356ef01/markupsafe-3.0.3-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:218551f6df4868a8d527e3062d0fb968682fe92054e89978594c28e642c43a73", size = 11658, upload-time = "2025-09-27T18:36:54.819Z" }, + { url = "https://files.pythonhosted.org/packages/99/9e/e412117548182ce2148bdeacdda3bb494260c0b0184360fe0d56389b523b/markupsafe-3.0.3-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:3524b778fe5cfb3452a09d31e7b5adefeea8c5be1d43c4f810ba09f2ceb29d37", size = 12066, upload-time = "2025-09-27T18:36:55.714Z" }, + { url = "https://files.pythonhosted.org/packages/bc/e6/fa0ffcda717ef64a5108eaa7b4f5ed28d56122c9a6d70ab8b72f9f715c80/markupsafe-3.0.3-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4e885a3d1efa2eadc93c894a21770e4bc67899e3543680313b09f139e149ab19", size = 25639, upload-time = "2025-09-27T18:36:56.908Z" }, + { url = "https://files.pythonhosted.org/packages/96/ec/2102e881fe9d25fc16cb4b25d5f5cde50970967ffa5dddafdb771237062d/markupsafe-3.0.3-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8709b08f4a89aa7586de0aadc8da56180242ee0ada3999749b183aa23df95025", size = 23569, upload-time = "2025-09-27T18:36:57.913Z" }, + { url = "https://files.pythonhosted.org/packages/4b/30/6f2fce1f1f205fc9323255b216ca8a235b15860c34b6798f810f05828e32/markupsafe-3.0.3-cp313-cp313t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:b8512a91625c9b3da6f127803b166b629725e68af71f8184ae7e7d54686a56d6", size = 23284, upload-time = "2025-09-27T18:36:58.833Z" }, + { url = "https://files.pythonhosted.org/packages/58/47/4a0ccea4ab9f5dcb6f79c0236d954acb382202721e704223a8aafa38b5c8/markupsafe-3.0.3-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:9b79b7a16f7fedff2495d684f2b59b0457c3b493778c9eed31111be64d58279f", size = 24801, upload-time = "2025-09-27T18:36:59.739Z" }, + { url = "https://files.pythonhosted.org/packages/6a/70/3780e9b72180b6fecb83a4814d84c3bf4b4ae4bf0b19c27196104149734c/markupsafe-3.0.3-cp313-cp313t-musllinux_1_2_riscv64.whl", hash = "sha256:12c63dfb4a98206f045aa9563db46507995f7ef6d83b2f68eda65c307c6829eb", size = 22769, upload-time = "2025-09-27T18:37:00.719Z" }, + { url = "https://files.pythonhosted.org/packages/98/c5/c03c7f4125180fc215220c035beac6b9cb684bc7a067c84fc69414d315f5/markupsafe-3.0.3-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:8f71bc33915be5186016f675cd83a1e08523649b0e33efdb898db577ef5bb009", size = 23642, upload-time = "2025-09-27T18:37:01.673Z" }, + { url = "https://files.pythonhosted.org/packages/80/d6/2d1b89f6ca4bff1036499b1e29a1d02d282259f3681540e16563f27ebc23/markupsafe-3.0.3-cp313-cp313t-win32.whl", hash = "sha256:69c0b73548bc525c8cb9a251cddf1931d1db4d2258e9599c28c07ef3580ef354", size = 14612, upload-time = "2025-09-27T18:37:02.639Z" }, + { url = "https://files.pythonhosted.org/packages/2b/98/e48a4bfba0a0ffcf9925fe2d69240bfaa19c6f7507b8cd09c70684a53c1e/markupsafe-3.0.3-cp313-cp313t-win_amd64.whl", hash = "sha256:1b4b79e8ebf6b55351f0d91fe80f893b4743f104bff22e90697db1590e47a218", size = 15200, upload-time = "2025-09-27T18:37:03.582Z" }, + { url = "https://files.pythonhosted.org/packages/0e/72/e3cc540f351f316e9ed0f092757459afbc595824ca724cbc5a5d4263713f/markupsafe-3.0.3-cp313-cp313t-win_arm64.whl", hash = "sha256:ad2cf8aa28b8c020ab2fc8287b0f823d0a7d8630784c31e9ee5edea20f406287", size = 13973, upload-time = "2025-09-27T18:37:04.929Z" }, + { url = "https://files.pythonhosted.org/packages/33/8a/8e42d4838cd89b7dde187011e97fe6c3af66d8c044997d2183fbd6d31352/markupsafe-3.0.3-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:eaa9599de571d72e2daf60164784109f19978b327a3910d3e9de8c97b5b70cfe", size = 11619, upload-time = "2025-09-27T18:37:06.342Z" }, + { url = "https://files.pythonhosted.org/packages/b5/64/7660f8a4a8e53c924d0fa05dc3a55c9cee10bbd82b11c5afb27d44b096ce/markupsafe-3.0.3-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:c47a551199eb8eb2121d4f0f15ae0f923d31350ab9280078d1e5f12b249e0026", size = 12029, upload-time = "2025-09-27T18:37:07.213Z" }, + { url = "https://files.pythonhosted.org/packages/da/ef/e648bfd021127bef5fa12e1720ffed0c6cbb8310c8d9bea7266337ff06de/markupsafe-3.0.3-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f34c41761022dd093b4b6896d4810782ffbabe30f2d443ff5f083e0cbbb8c737", size = 24408, upload-time = "2025-09-27T18:37:09.572Z" }, + { url = "https://files.pythonhosted.org/packages/41/3c/a36c2450754618e62008bf7435ccb0f88053e07592e6028a34776213d877/markupsafe-3.0.3-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:457a69a9577064c05a97c41f4e65148652db078a3a509039e64d3467b9e7ef97", size = 23005, upload-time = "2025-09-27T18:37:10.58Z" }, + { url = "https://files.pythonhosted.org/packages/bc/20/b7fdf89a8456b099837cd1dc21974632a02a999ec9bf7ca3e490aacd98e7/markupsafe-3.0.3-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:e8afc3f2ccfa24215f8cb28dcf43f0113ac3c37c2f0f0806d8c70e4228c5cf4d", size = 22048, upload-time = "2025-09-27T18:37:11.547Z" }, + { url = "https://files.pythonhosted.org/packages/9a/a7/591f592afdc734f47db08a75793a55d7fbcc6902a723ae4cfbab61010cc5/markupsafe-3.0.3-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:ec15a59cf5af7be74194f7ab02d0f59a62bdcf1a537677ce67a2537c9b87fcda", size = 23821, upload-time = "2025-09-27T18:37:12.48Z" }, + { url = "https://files.pythonhosted.org/packages/7d/33/45b24e4f44195b26521bc6f1a82197118f74df348556594bd2262bda1038/markupsafe-3.0.3-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:0eb9ff8191e8498cca014656ae6b8d61f39da5f95b488805da4bb029cccbfbaf", size = 21606, upload-time = "2025-09-27T18:37:13.485Z" }, + { url = "https://files.pythonhosted.org/packages/ff/0e/53dfaca23a69fbfbbf17a4b64072090e70717344c52eaaaa9c5ddff1e5f0/markupsafe-3.0.3-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:2713baf880df847f2bece4230d4d094280f4e67b1e813eec43b4c0e144a34ffe", size = 23043, upload-time = "2025-09-27T18:37:14.408Z" }, + { url = "https://files.pythonhosted.org/packages/46/11/f333a06fc16236d5238bfe74daccbca41459dcd8d1fa952e8fbd5dccfb70/markupsafe-3.0.3-cp314-cp314-win32.whl", hash = "sha256:729586769a26dbceff69f7a7dbbf59ab6572b99d94576a5592625d5b411576b9", size = 14747, upload-time = "2025-09-27T18:37:15.36Z" }, + { url = "https://files.pythonhosted.org/packages/28/52/182836104b33b444e400b14f797212f720cbc9ed6ba34c800639d154e821/markupsafe-3.0.3-cp314-cp314-win_amd64.whl", hash = "sha256:bdc919ead48f234740ad807933cdf545180bfbe9342c2bb451556db2ed958581", size = 15341, upload-time = "2025-09-27T18:37:16.496Z" }, + { url = "https://files.pythonhosted.org/packages/6f/18/acf23e91bd94fd7b3031558b1f013adfa21a8e407a3fdb32745538730382/markupsafe-3.0.3-cp314-cp314-win_arm64.whl", hash = "sha256:5a7d5dc5140555cf21a6fefbdbf8723f06fcd2f63ef108f2854de715e4422cb4", size = 14073, upload-time = "2025-09-27T18:37:17.476Z" }, + { url = "https://files.pythonhosted.org/packages/3c/f0/57689aa4076e1b43b15fdfa646b04653969d50cf30c32a102762be2485da/markupsafe-3.0.3-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:1353ef0c1b138e1907ae78e2f6c63ff67501122006b0f9abad68fda5f4ffc6ab", size = 11661, upload-time = "2025-09-27T18:37:18.453Z" }, + { url = "https://files.pythonhosted.org/packages/89/c3/2e67a7ca217c6912985ec766c6393b636fb0c2344443ff9d91404dc4c79f/markupsafe-3.0.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:1085e7fbddd3be5f89cc898938f42c0b3c711fdcb37d75221de2666af647c175", size = 12069, upload-time = "2025-09-27T18:37:19.332Z" }, + { url = "https://files.pythonhosted.org/packages/f0/00/be561dce4e6ca66b15276e184ce4b8aec61fe83662cce2f7d72bd3249d28/markupsafe-3.0.3-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1b52b4fb9df4eb9ae465f8d0c228a00624de2334f216f178a995ccdcf82c4634", size = 25670, upload-time = "2025-09-27T18:37:20.245Z" }, + { url = "https://files.pythonhosted.org/packages/50/09/c419f6f5a92e5fadde27efd190eca90f05e1261b10dbd8cbcb39cd8ea1dc/markupsafe-3.0.3-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fed51ac40f757d41b7c48425901843666a6677e3e8eb0abcff09e4ba6e664f50", size = 23598, upload-time = "2025-09-27T18:37:21.177Z" }, + { url = "https://files.pythonhosted.org/packages/22/44/a0681611106e0b2921b3033fc19bc53323e0b50bc70cffdd19f7d679bb66/markupsafe-3.0.3-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:f190daf01f13c72eac4efd5c430a8de82489d9cff23c364c3ea822545032993e", size = 23261, upload-time = "2025-09-27T18:37:22.167Z" }, + { url = "https://files.pythonhosted.org/packages/5f/57/1b0b3f100259dc9fffe780cfb60d4be71375510e435efec3d116b6436d43/markupsafe-3.0.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:e56b7d45a839a697b5eb268c82a71bd8c7f6c94d6fd50c3d577fa39a9f1409f5", size = 24835, upload-time = "2025-09-27T18:37:23.296Z" }, + { url = "https://files.pythonhosted.org/packages/26/6a/4bf6d0c97c4920f1597cc14dd720705eca0bf7c787aebc6bb4d1bead5388/markupsafe-3.0.3-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:f3e98bb3798ead92273dc0e5fd0f31ade220f59a266ffd8a4f6065e0a3ce0523", size = 22733, upload-time = "2025-09-27T18:37:24.237Z" }, + { url = "https://files.pythonhosted.org/packages/14/c7/ca723101509b518797fedc2fdf79ba57f886b4aca8a7d31857ba3ee8281f/markupsafe-3.0.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:5678211cb9333a6468fb8d8be0305520aa073f50d17f089b5b4b477ea6e67fdc", size = 23672, upload-time = "2025-09-27T18:37:25.271Z" }, + { url = "https://files.pythonhosted.org/packages/fb/df/5bd7a48c256faecd1d36edc13133e51397e41b73bb77e1a69deab746ebac/markupsafe-3.0.3-cp314-cp314t-win32.whl", hash = "sha256:915c04ba3851909ce68ccc2b8e2cd691618c4dc4c4232fb7982bca3f41fd8c3d", size = 14819, upload-time = "2025-09-27T18:37:26.285Z" }, + { url = "https://files.pythonhosted.org/packages/1a/8a/0402ba61a2f16038b48b39bccca271134be00c5c9f0f623208399333c448/markupsafe-3.0.3-cp314-cp314t-win_amd64.whl", hash = "sha256:4faffd047e07c38848ce017e8725090413cd80cbc23d86e55c587bf979e579c9", size = 15426, upload-time = "2025-09-27T18:37:27.316Z" }, + { url = "https://files.pythonhosted.org/packages/70/bc/6f1c2f612465f5fa89b95bead1f44dcb607670fd42891d8fdcd5d039f4f4/markupsafe-3.0.3-cp314-cp314t-win_arm64.whl", hash = "sha256:32001d6a8fc98c8cb5c947787c5d08b0a50663d139f1305bac5885d98d9b40fa", size = 14146, upload-time = "2025-09-27T18:37:28.327Z" }, +] + +[[package]] +name = "mdurl" +version = "0.1.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/d6/54/cfe61301667036ec958cb99bd3efefba235e65cdeb9c84d24a8293ba1d90/mdurl-0.1.2.tar.gz", hash = "sha256:bb413d29f5eea38f31dd4754dd7377d4465116fb207585f97bf925588687c1ba", size = 8729, upload-time = "2022-08-14T12:40:10.846Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/b3/38/89ba8ad64ae25be8de66a6d463314cf1eb366222074cfda9ee839c56a4b4/mdurl-0.1.2-py3-none-any.whl", hash = "sha256:84008a41e51615a49fc9966191ff91509e3c40b939176e643fd50a5c2196b8f8", size = 9979, upload-time = "2022-08-14T12:40:09.779Z" }, +] + +[[package]] +name = "moviepy" +version = "2.2.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "decorator" }, + { name = "imageio" }, + { name = "imageio-ffmpeg" }, + { name = "numpy" }, + { name = "pillow" }, + { name = "proglog" }, + { name = "python-dotenv" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/de/61/15f9476e270f64c78a834e7459ca045d669f869cec24eed26807b8cd479d/moviepy-2.2.1.tar.gz", hash = "sha256:c80cb56815ece94e5e3e2d361aa40070eeb30a09d23a24c4e684d03e16deacb1", size = 58431438, upload-time = "2025-05-21T19:31:52.601Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/9a/73/7d3b2010baa0b5eb1e4dfa9e4385e89b6716be76f2fa21a6c0fe34b68e5a/moviepy-2.2.1-py3-none-any.whl", hash = "sha256:6b56803fec2ac54b557404126ac1160e65448e03798fa282bd23e8fab3795060", size = 129871, upload-time = "2025-05-21T19:31:50.11Z" }, +] + +[[package]] +name = "mpmath" +version = "1.3.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/e0/47/dd32fa426cc72114383ac549964eecb20ecfd886d1e5ccf5340b55b02f57/mpmath-1.3.0.tar.gz", hash = "sha256:7a28eb2a9774d00c7bc92411c19a89209d5da7c4c9a9e227be8330a23a25b91f", size = 508106, upload-time = "2023-03-07T16:47:11.061Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/43/e3/7d92a15f894aa0c9c4b49b8ee9ac9850d6e63b03c9c32c0367a13ae62209/mpmath-1.3.0-py3-none-any.whl", hash = "sha256:a0b2b9fe80bbcd81a6647ff13108738cfb482d481d826cc0e02f5b35e5c88d2c", size = 536198, upload-time = "2023-03-07T16:47:09.197Z" }, +] + +[[package]] +name = "msgpack" +version = "1.1.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/4d/f2/bfb55a6236ed8725a96b0aa3acbd0ec17588e6a2c3b62a93eb513ed8783f/msgpack-1.1.2.tar.gz", hash = "sha256:3b60763c1373dd60f398488069bcdc703cd08a711477b5d480eecc9f9626f47e", size = 173581, upload-time = "2025-10-08T09:15:56.596Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/f5/a2/3b68a9e769db68668b25c6108444a35f9bd163bb848c0650d516761a59c0/msgpack-1.1.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:0051fffef5a37ca2cd16978ae4f0aef92f164df86823871b5162812bebecd8e2", size = 81318, upload-time = "2025-10-08T09:14:38.722Z" }, + { url = "https://files.pythonhosted.org/packages/5b/e1/2b720cc341325c00be44e1ed59e7cfeae2678329fbf5aa68f5bda57fe728/msgpack-1.1.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:a605409040f2da88676e9c9e5853b3449ba8011973616189ea5ee55ddbc5bc87", size = 83786, upload-time = "2025-10-08T09:14:40.082Z" }, + { url = "https://files.pythonhosted.org/packages/71/e5/c2241de64bfceac456b140737812a2ab310b10538a7b34a1d393b748e095/msgpack-1.1.2-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8b696e83c9f1532b4af884045ba7f3aa741a63b2bc22617293a2c6a7c645f251", size = 398240, upload-time = "2025-10-08T09:14:41.151Z" }, + { url = "https://files.pythonhosted.org/packages/b7/09/2a06956383c0fdebaef5aa9246e2356776f12ea6f2a44bd1368abf0e46c4/msgpack-1.1.2-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:365c0bbe981a27d8932da71af63ef86acc59ed5c01ad929e09a0b88c6294e28a", size = 406070, upload-time = "2025-10-08T09:14:42.821Z" }, + { url = "https://files.pythonhosted.org/packages/0e/74/2957703f0e1ef20637d6aead4fbb314330c26f39aa046b348c7edcf6ca6b/msgpack-1.1.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:41d1a5d875680166d3ac5c38573896453bbbea7092936d2e107214daf43b1d4f", size = 393403, upload-time = "2025-10-08T09:14:44.38Z" }, + { url = "https://files.pythonhosted.org/packages/a5/09/3bfc12aa90f77b37322fc33e7a8a7c29ba7c8edeadfa27664451801b9860/msgpack-1.1.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:354e81bcdebaab427c3df4281187edc765d5d76bfb3a7c125af9da7a27e8458f", size = 398947, upload-time = "2025-10-08T09:14:45.56Z" }, + { url = "https://files.pythonhosted.org/packages/4b/4f/05fcebd3b4977cb3d840f7ef6b77c51f8582086de5e642f3fefee35c86fc/msgpack-1.1.2-cp310-cp310-win32.whl", hash = "sha256:e64c8d2f5e5d5fda7b842f55dec6133260ea8f53c4257d64494c534f306bf7a9", size = 64769, upload-time = "2025-10-08T09:14:47.334Z" }, + { url = "https://files.pythonhosted.org/packages/d0/3e/b4547e3a34210956382eed1c85935fff7e0f9b98be3106b3745d7dec9c5e/msgpack-1.1.2-cp310-cp310-win_amd64.whl", hash = "sha256:db6192777d943bdaaafb6ba66d44bf65aa0e9c5616fa1d2da9bb08828c6b39aa", size = 71293, upload-time = "2025-10-08T09:14:48.665Z" }, + { url = "https://files.pythonhosted.org/packages/2c/97/560d11202bcd537abca693fd85d81cebe2107ba17301de42b01ac1677b69/msgpack-1.1.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:2e86a607e558d22985d856948c12a3fa7b42efad264dca8a3ebbcfa2735d786c", size = 82271, upload-time = "2025-10-08T09:14:49.967Z" }, + { url = "https://files.pythonhosted.org/packages/83/04/28a41024ccbd67467380b6fb440ae916c1e4f25e2cd4c63abe6835ac566e/msgpack-1.1.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:283ae72fc89da59aa004ba147e8fc2f766647b1251500182fac0350d8af299c0", size = 84914, upload-time = "2025-10-08T09:14:50.958Z" }, + { url = "https://files.pythonhosted.org/packages/71/46/b817349db6886d79e57a966346cf0902a426375aadc1e8e7a86a75e22f19/msgpack-1.1.2-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:61c8aa3bd513d87c72ed0b37b53dd5c5a0f58f2ff9f26e1555d3bd7948fb7296", size = 416962, upload-time = "2025-10-08T09:14:51.997Z" }, + { url = "https://files.pythonhosted.org/packages/da/e0/6cc2e852837cd6086fe7d8406af4294e66827a60a4cf60b86575a4a65ca8/msgpack-1.1.2-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:454e29e186285d2ebe65be34629fa0e8605202c60fbc7c4c650ccd41870896ef", size = 426183, upload-time = "2025-10-08T09:14:53.477Z" }, + { url = "https://files.pythonhosted.org/packages/25/98/6a19f030b3d2ea906696cedd1eb251708e50a5891d0978b012cb6107234c/msgpack-1.1.2-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:7bc8813f88417599564fafa59fd6f95be417179f76b40325b500b3c98409757c", size = 411454, upload-time = "2025-10-08T09:14:54.648Z" }, + { url = "https://files.pythonhosted.org/packages/b7/cd/9098fcb6adb32187a70b7ecaabf6339da50553351558f37600e53a4a2a23/msgpack-1.1.2-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:bafca952dc13907bdfdedfc6a5f579bf4f292bdd506fadb38389afa3ac5b208e", size = 422341, upload-time = "2025-10-08T09:14:56.328Z" }, + { url = "https://files.pythonhosted.org/packages/e6/ae/270cecbcf36c1dc85ec086b33a51a4d7d08fc4f404bdbc15b582255d05ff/msgpack-1.1.2-cp311-cp311-win32.whl", hash = "sha256:602b6740e95ffc55bfb078172d279de3773d7b7db1f703b2f1323566b878b90e", size = 64747, upload-time = "2025-10-08T09:14:57.882Z" }, + { url = "https://files.pythonhosted.org/packages/2a/79/309d0e637f6f37e83c711f547308b91af02b72d2326ddd860b966080ef29/msgpack-1.1.2-cp311-cp311-win_amd64.whl", hash = "sha256:d198d275222dc54244bf3327eb8cbe00307d220241d9cec4d306d49a44e85f68", size = 71633, upload-time = "2025-10-08T09:14:59.177Z" }, + { url = "https://files.pythonhosted.org/packages/73/4d/7c4e2b3d9b1106cd0aa6cb56cc57c6267f59fa8bfab7d91df5adc802c847/msgpack-1.1.2-cp311-cp311-win_arm64.whl", hash = "sha256:86f8136dfa5c116365a8a651a7d7484b65b13339731dd6faebb9a0242151c406", size = 64755, upload-time = "2025-10-08T09:15:00.48Z" }, + { url = "https://files.pythonhosted.org/packages/ad/bd/8b0d01c756203fbab65d265859749860682ccd2a59594609aeec3a144efa/msgpack-1.1.2-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:70a0dff9d1f8da25179ffcf880e10cf1aad55fdb63cd59c9a49a1b82290062aa", size = 81939, upload-time = "2025-10-08T09:15:01.472Z" }, + { url = "https://files.pythonhosted.org/packages/34/68/ba4f155f793a74c1483d4bdef136e1023f7bcba557f0db4ef3db3c665cf1/msgpack-1.1.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:446abdd8b94b55c800ac34b102dffd2f6aa0ce643c55dfc017ad89347db3dbdb", size = 85064, upload-time = "2025-10-08T09:15:03.764Z" }, + { url = "https://files.pythonhosted.org/packages/f2/60/a064b0345fc36c4c3d2c743c82d9100c40388d77f0b48b2f04d6041dbec1/msgpack-1.1.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c63eea553c69ab05b6747901b97d620bb2a690633c77f23feb0c6a947a8a7b8f", size = 417131, upload-time = "2025-10-08T09:15:05.136Z" }, + { url = "https://files.pythonhosted.org/packages/65/92/a5100f7185a800a5d29f8d14041f61475b9de465ffcc0f3b9fba606e4505/msgpack-1.1.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:372839311ccf6bdaf39b00b61288e0557916c3729529b301c52c2d88842add42", size = 427556, upload-time = "2025-10-08T09:15:06.837Z" }, + { url = "https://files.pythonhosted.org/packages/f5/87/ffe21d1bf7d9991354ad93949286f643b2bb6ddbeab66373922b44c3b8cc/msgpack-1.1.2-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:2929af52106ca73fcb28576218476ffbb531a036c2adbcf54a3664de124303e9", size = 404920, upload-time = "2025-10-08T09:15:08.179Z" }, + { url = "https://files.pythonhosted.org/packages/ff/41/8543ed2b8604f7c0d89ce066f42007faac1eaa7d79a81555f206a5cdb889/msgpack-1.1.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:be52a8fc79e45b0364210eef5234a7cf8d330836d0a64dfbb878efa903d84620", size = 415013, upload-time = "2025-10-08T09:15:09.83Z" }, + { url = "https://files.pythonhosted.org/packages/41/0d/2ddfaa8b7e1cee6c490d46cb0a39742b19e2481600a7a0e96537e9c22f43/msgpack-1.1.2-cp312-cp312-win32.whl", hash = "sha256:1fff3d825d7859ac888b0fbda39a42d59193543920eda9d9bea44d958a878029", size = 65096, upload-time = "2025-10-08T09:15:11.11Z" }, + { url = "https://files.pythonhosted.org/packages/8c/ec/d431eb7941fb55a31dd6ca3404d41fbb52d99172df2e7707754488390910/msgpack-1.1.2-cp312-cp312-win_amd64.whl", hash = "sha256:1de460f0403172cff81169a30b9a92b260cb809c4cb7e2fc79ae8d0510c78b6b", size = 72708, upload-time = "2025-10-08T09:15:12.554Z" }, + { url = "https://files.pythonhosted.org/packages/c5/31/5b1a1f70eb0e87d1678e9624908f86317787b536060641d6798e3cf70ace/msgpack-1.1.2-cp312-cp312-win_arm64.whl", hash = "sha256:be5980f3ee0e6bd44f3a9e9dea01054f175b50c3e6cdb692bc9424c0bbb8bf69", size = 64119, upload-time = "2025-10-08T09:15:13.589Z" }, + { url = "https://files.pythonhosted.org/packages/6b/31/b46518ecc604d7edf3a4f94cb3bf021fc62aa301f0cb849936968164ef23/msgpack-1.1.2-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:4efd7b5979ccb539c221a4c4e16aac1a533efc97f3b759bb5a5ac9f6d10383bf", size = 81212, upload-time = "2025-10-08T09:15:14.552Z" }, + { url = "https://files.pythonhosted.org/packages/92/dc/c385f38f2c2433333345a82926c6bfa5ecfff3ef787201614317b58dd8be/msgpack-1.1.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:42eefe2c3e2af97ed470eec850facbe1b5ad1d6eacdbadc42ec98e7dcf68b4b7", size = 84315, upload-time = "2025-10-08T09:15:15.543Z" }, + { url = "https://files.pythonhosted.org/packages/d3/68/93180dce57f684a61a88a45ed13047558ded2be46f03acb8dec6d7c513af/msgpack-1.1.2-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1fdf7d83102bf09e7ce3357de96c59b627395352a4024f6e2458501f158bf999", size = 412721, upload-time = "2025-10-08T09:15:16.567Z" }, + { url = "https://files.pythonhosted.org/packages/5d/ba/459f18c16f2b3fc1a1ca871f72f07d70c07bf768ad0a507a698b8052ac58/msgpack-1.1.2-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fac4be746328f90caa3cd4bc67e6fe36ca2bf61d5c6eb6d895b6527e3f05071e", size = 424657, upload-time = "2025-10-08T09:15:17.825Z" }, + { url = "https://files.pythonhosted.org/packages/38/f8/4398c46863b093252fe67368b44edc6c13b17f4e6b0e4929dbf0bdb13f23/msgpack-1.1.2-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:fffee09044073e69f2bad787071aeec727183e7580443dfeb8556cbf1978d162", size = 402668, upload-time = "2025-10-08T09:15:19.003Z" }, + { url = "https://files.pythonhosted.org/packages/28/ce/698c1eff75626e4124b4d78e21cca0b4cc90043afb80a507626ea354ab52/msgpack-1.1.2-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:5928604de9b032bc17f5099496417f113c45bc6bc21b5c6920caf34b3c428794", size = 419040, upload-time = "2025-10-08T09:15:20.183Z" }, + { url = "https://files.pythonhosted.org/packages/67/32/f3cd1667028424fa7001d82e10ee35386eea1408b93d399b09fb0aa7875f/msgpack-1.1.2-cp313-cp313-win32.whl", hash = "sha256:a7787d353595c7c7e145e2331abf8b7ff1e6673a6b974ded96e6d4ec09f00c8c", size = 65037, upload-time = "2025-10-08T09:15:21.416Z" }, + { url = "https://files.pythonhosted.org/packages/74/07/1ed8277f8653c40ebc65985180b007879f6a836c525b3885dcc6448ae6cb/msgpack-1.1.2-cp313-cp313-win_amd64.whl", hash = "sha256:a465f0dceb8e13a487e54c07d04ae3ba131c7c5b95e2612596eafde1dccf64a9", size = 72631, upload-time = "2025-10-08T09:15:22.431Z" }, + { url = "https://files.pythonhosted.org/packages/e5/db/0314e4e2db56ebcf450f277904ffd84a7988b9e5da8d0d61ab2d057df2b6/msgpack-1.1.2-cp313-cp313-win_arm64.whl", hash = "sha256:e69b39f8c0aa5ec24b57737ebee40be647035158f14ed4b40e6f150077e21a84", size = 64118, upload-time = "2025-10-08T09:15:23.402Z" }, + { url = "https://files.pythonhosted.org/packages/22/71/201105712d0a2ff07b7873ed3c220292fb2ea5120603c00c4b634bcdafb3/msgpack-1.1.2-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:e23ce8d5f7aa6ea6d2a2b326b4ba46c985dbb204523759984430db7114f8aa00", size = 81127, upload-time = "2025-10-08T09:15:24.408Z" }, + { url = "https://files.pythonhosted.org/packages/1b/9f/38ff9e57a2eade7bf9dfee5eae17f39fc0e998658050279cbb14d97d36d9/msgpack-1.1.2-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:6c15b7d74c939ebe620dd8e559384be806204d73b4f9356320632d783d1f7939", size = 84981, upload-time = "2025-10-08T09:15:25.812Z" }, + { url = "https://files.pythonhosted.org/packages/8e/a9/3536e385167b88c2cc8f4424c49e28d49a6fc35206d4a8060f136e71f94c/msgpack-1.1.2-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:99e2cb7b9031568a2a5c73aa077180f93dd2e95b4f8d3b8e14a73ae94a9e667e", size = 411885, upload-time = "2025-10-08T09:15:27.22Z" }, + { url = "https://files.pythonhosted.org/packages/2f/40/dc34d1a8d5f1e51fc64640b62b191684da52ca469da9cd74e84936ffa4a6/msgpack-1.1.2-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:180759d89a057eab503cf62eeec0aa61c4ea1200dee709f3a8e9397dbb3b6931", size = 419658, upload-time = "2025-10-08T09:15:28.4Z" }, + { url = "https://files.pythonhosted.org/packages/3b/ef/2b92e286366500a09a67e03496ee8b8ba00562797a52f3c117aa2b29514b/msgpack-1.1.2-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:04fb995247a6e83830b62f0b07bf36540c213f6eac8e851166d8d86d83cbd014", size = 403290, upload-time = "2025-10-08T09:15:29.764Z" }, + { url = "https://files.pythonhosted.org/packages/78/90/e0ea7990abea5764e4655b8177aa7c63cdfa89945b6e7641055800f6c16b/msgpack-1.1.2-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:8e22ab046fa7ede9e36eeb4cfad44d46450f37bb05d5ec482b02868f451c95e2", size = 415234, upload-time = "2025-10-08T09:15:31.022Z" }, + { url = "https://files.pythonhosted.org/packages/72/4e/9390aed5db983a2310818cd7d3ec0aecad45e1f7007e0cda79c79507bb0d/msgpack-1.1.2-cp314-cp314-win32.whl", hash = "sha256:80a0ff7d4abf5fecb995fcf235d4064b9a9a8a40a3ab80999e6ac1e30b702717", size = 66391, upload-time = "2025-10-08T09:15:32.265Z" }, + { url = "https://files.pythonhosted.org/packages/6e/f1/abd09c2ae91228c5f3998dbd7f41353def9eac64253de3c8105efa2082f7/msgpack-1.1.2-cp314-cp314-win_amd64.whl", hash = "sha256:9ade919fac6a3e7260b7f64cea89df6bec59104987cbea34d34a2fa15d74310b", size = 73787, upload-time = "2025-10-08T09:15:33.219Z" }, + { url = "https://files.pythonhosted.org/packages/6a/b0/9d9f667ab48b16ad4115c1935d94023b82b3198064cb84a123e97f7466c1/msgpack-1.1.2-cp314-cp314-win_arm64.whl", hash = "sha256:59415c6076b1e30e563eb732e23b994a61c159cec44deaf584e5cc1dd662f2af", size = 66453, upload-time = "2025-10-08T09:15:34.225Z" }, + { url = "https://files.pythonhosted.org/packages/16/67/93f80545eb1792b61a217fa7f06d5e5cb9e0055bed867f43e2b8e012e137/msgpack-1.1.2-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:897c478140877e5307760b0ea66e0932738879e7aa68144d9b78ea4c8302a84a", size = 85264, upload-time = "2025-10-08T09:15:35.61Z" }, + { url = "https://files.pythonhosted.org/packages/87/1c/33c8a24959cf193966ef11a6f6a2995a65eb066bd681fd085afd519a57ce/msgpack-1.1.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:a668204fa43e6d02f89dbe79a30b0d67238d9ec4c5bd8a940fc3a004a47b721b", size = 89076, upload-time = "2025-10-08T09:15:36.619Z" }, + { url = "https://files.pythonhosted.org/packages/fc/6b/62e85ff7193663fbea5c0254ef32f0c77134b4059f8da89b958beb7696f3/msgpack-1.1.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5559d03930d3aa0f3aacb4c42c776af1a2ace2611871c84a75afe436695e6245", size = 435242, upload-time = "2025-10-08T09:15:37.647Z" }, + { url = "https://files.pythonhosted.org/packages/c1/47/5c74ecb4cc277cf09f64e913947871682ffa82b3b93c8dad68083112f412/msgpack-1.1.2-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:70c5a7a9fea7f036b716191c29047374c10721c389c21e9ffafad04df8c52c90", size = 432509, upload-time = "2025-10-08T09:15:38.794Z" }, + { url = "https://files.pythonhosted.org/packages/24/a4/e98ccdb56dc4e98c929a3f150de1799831c0a800583cde9fa022fa90602d/msgpack-1.1.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:f2cb069d8b981abc72b41aea1c580ce92d57c673ec61af4c500153a626cb9e20", size = 415957, upload-time = "2025-10-08T09:15:40.238Z" }, + { url = "https://files.pythonhosted.org/packages/da/28/6951f7fb67bc0a4e184a6b38ab71a92d9ba58080b27a77d3e2fb0be5998f/msgpack-1.1.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:d62ce1f483f355f61adb5433ebfd8868c5f078d1a52d042b0a998682b4fa8c27", size = 422910, upload-time = "2025-10-08T09:15:41.505Z" }, + { url = "https://files.pythonhosted.org/packages/f0/03/42106dcded51f0a0b5284d3ce30a671e7bd3f7318d122b2ead66ad289fed/msgpack-1.1.2-cp314-cp314t-win32.whl", hash = "sha256:1d1418482b1ee984625d88aa9585db570180c286d942da463533b238b98b812b", size = 75197, upload-time = "2025-10-08T09:15:42.954Z" }, + { url = "https://files.pythonhosted.org/packages/15/86/d0071e94987f8db59d4eeb386ddc64d0bb9b10820a8d82bcd3e53eeb2da6/msgpack-1.1.2-cp314-cp314t-win_amd64.whl", hash = "sha256:5a46bf7e831d09470ad92dff02b8b1ac92175ca36b087f904a0519857c6be3ff", size = 85772, upload-time = "2025-10-08T09:15:43.954Z" }, + { url = "https://files.pythonhosted.org/packages/81/f2/08ace4142eb281c12701fc3b93a10795e4d4dc7f753911d836675050f886/msgpack-1.1.2-cp314-cp314t-win_arm64.whl", hash = "sha256:d99ef64f349d5ec3293688e91486c5fdb925ed03807f64d98d205d2713c60b46", size = 70868, upload-time = "2025-10-08T09:15:44.959Z" }, +] + +[[package]] +name = "multidict" +version = "6.7.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "typing-extensions", marker = "python_full_version < '3.11'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/80/1e/5492c365f222f907de1039b91f922b93fa4f764c713ee858d235495d8f50/multidict-6.7.0.tar.gz", hash = "sha256:c6e99d9a65ca282e578dfea819cfa9c0a62b2499d8677392e09feaf305e9e6f5", size = 101834, upload-time = "2025-10-06T14:52:30.657Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/a9/63/7bdd4adc330abcca54c85728db2327130e49e52e8c3ce685cec44e0f2e9f/multidict-6.7.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:9f474ad5acda359c8758c8accc22032c6abe6dc87a8be2440d097785e27a9349", size = 77153, upload-time = "2025-10-06T14:48:26.409Z" }, + { url = "https://files.pythonhosted.org/packages/3f/bb/b6c35ff175ed1a3142222b78455ee31be71a8396ed3ab5280fbe3ebe4e85/multidict-6.7.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:4b7a9db5a870f780220e931d0002bbfd88fb53aceb6293251e2c839415c1b20e", size = 44993, upload-time = "2025-10-06T14:48:28.4Z" }, + { url = "https://files.pythonhosted.org/packages/e0/1f/064c77877c5fa6df6d346e68075c0f6998547afe952d6471b4c5f6a7345d/multidict-6.7.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:03ca744319864e92721195fa28c7a3b2bc7b686246b35e4078c1e4d0eb5466d3", size = 44607, upload-time = "2025-10-06T14:48:29.581Z" }, + { url = "https://files.pythonhosted.org/packages/04/7a/bf6aa92065dd47f287690000b3d7d332edfccb2277634cadf6a810463c6a/multidict-6.7.0-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:f0e77e3c0008bc9316e662624535b88d360c3a5d3f81e15cf12c139a75250046", size = 241847, upload-time = "2025-10-06T14:48:32.107Z" }, + { url = "https://files.pythonhosted.org/packages/94/39/297a8de920f76eda343e4ce05f3b489f0ab3f9504f2576dfb37b7c08ca08/multidict-6.7.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:08325c9e5367aa379a3496aa9a022fe8837ff22e00b94db256d3a1378c76ab32", size = 242616, upload-time = "2025-10-06T14:48:34.054Z" }, + { url = "https://files.pythonhosted.org/packages/39/3a/d0eee2898cfd9d654aea6cb8c4addc2f9756e9a7e09391cfe55541f917f7/multidict-6.7.0-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e2862408c99f84aa571ab462d25236ef9cb12a602ea959ba9c9009a54902fc73", size = 222333, upload-time = "2025-10-06T14:48:35.9Z" }, + { url = "https://files.pythonhosted.org/packages/05/48/3b328851193c7a4240815b71eea165b49248867bbb6153a0aee227a0bb47/multidict-6.7.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:4d72a9a2d885f5c208b0cb91ff2ed43636bb7e345ec839ff64708e04f69a13cc", size = 253239, upload-time = "2025-10-06T14:48:37.302Z" }, + { url = "https://files.pythonhosted.org/packages/b1/ca/0706a98c8d126a89245413225ca4a3fefc8435014de309cf8b30acb68841/multidict-6.7.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:478cc36476687bac1514d651cbbaa94b86b0732fb6855c60c673794c7dd2da62", size = 251618, upload-time = "2025-10-06T14:48:38.963Z" }, + { url = "https://files.pythonhosted.org/packages/5e/4f/9c7992f245554d8b173f6f0a048ad24b3e645d883f096857ec2c0822b8bd/multidict-6.7.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6843b28b0364dc605f21481c90fadb5f60d9123b442eb8a726bb74feef588a84", size = 241655, upload-time = "2025-10-06T14:48:40.312Z" }, + { url = "https://files.pythonhosted.org/packages/31/79/26a85991ae67efd1c0b1fc2e0c275b8a6aceeb155a68861f63f87a798f16/multidict-6.7.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:23bfeee5316266e5ee2d625df2d2c602b829435fc3a235c2ba2131495706e4a0", size = 239245, upload-time = "2025-10-06T14:48:41.848Z" }, + { url = "https://files.pythonhosted.org/packages/14/1e/75fa96394478930b79d0302eaf9a6c69f34005a1a5251ac8b9c336486ec9/multidict-6.7.0-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:680878b9f3d45c31e1f730eef731f9b0bc1da456155688c6745ee84eb818e90e", size = 233523, upload-time = "2025-10-06T14:48:43.749Z" }, + { url = "https://files.pythonhosted.org/packages/b2/5e/085544cb9f9c4ad2b5d97467c15f856df8d9bac410cffd5c43991a5d878b/multidict-6.7.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:eb866162ef2f45063acc7a53a88ef6fe8bf121d45c30ea3c9cd87ce7e191a8d4", size = 243129, upload-time = "2025-10-06T14:48:45.225Z" }, + { url = "https://files.pythonhosted.org/packages/b9/c3/e9d9e2f20c9474e7a8fcef28f863c5cbd29bb5adce6b70cebe8bdad0039d/multidict-6.7.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:df0e3bf7993bdbeca5ac25aa859cf40d39019e015c9c91809ba7093967f7a648", size = 248999, upload-time = "2025-10-06T14:48:46.703Z" }, + { url = "https://files.pythonhosted.org/packages/b5/3f/df171b6efa3239ae33b97b887e42671cd1d94d460614bfb2c30ffdab3b95/multidict-6.7.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:661709cdcd919a2ece2234f9bae7174e5220c80b034585d7d8a755632d3e2111", size = 243711, upload-time = "2025-10-06T14:48:48.146Z" }, + { url = "https://files.pythonhosted.org/packages/3c/2f/9b5564888c4e14b9af64c54acf149263721a283aaf4aa0ae89b091d5d8c1/multidict-6.7.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:096f52730c3fb8ed419db2d44391932b63891b2c5ed14850a7e215c0ba9ade36", size = 237504, upload-time = "2025-10-06T14:48:49.447Z" }, + { url = "https://files.pythonhosted.org/packages/6c/3a/0bd6ca0f7d96d790542d591c8c3354c1e1b6bfd2024d4d92dc3d87485ec7/multidict-6.7.0-cp310-cp310-win32.whl", hash = "sha256:afa8a2978ec65d2336305550535c9c4ff50ee527914328c8677b3973ade52b85", size = 41422, upload-time = "2025-10-06T14:48:50.789Z" }, + { url = "https://files.pythonhosted.org/packages/00/35/f6a637ea2c75f0d3b7c7d41b1189189acff0d9deeb8b8f35536bb30f5e33/multidict-6.7.0-cp310-cp310-win_amd64.whl", hash = "sha256:b15b3afff74f707b9275d5ba6a91ae8f6429c3ffb29bbfd216b0b375a56f13d7", size = 46050, upload-time = "2025-10-06T14:48:51.938Z" }, + { url = "https://files.pythonhosted.org/packages/e7/b8/f7bf8329b39893d02d9d95cf610c75885d12fc0f402b1c894e1c8e01c916/multidict-6.7.0-cp310-cp310-win_arm64.whl", hash = "sha256:4b73189894398d59131a66ff157837b1fafea9974be486d036bb3d32331fdbf0", size = 43153, upload-time = "2025-10-06T14:48:53.146Z" }, + { url = "https://files.pythonhosted.org/packages/34/9e/5c727587644d67b2ed479041e4b1c58e30afc011e3d45d25bbe35781217c/multidict-6.7.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:4d409aa42a94c0b3fa617708ef5276dfe81012ba6753a0370fcc9d0195d0a1fc", size = 76604, upload-time = "2025-10-06T14:48:54.277Z" }, + { url = "https://files.pythonhosted.org/packages/17/e4/67b5c27bd17c085a5ea8f1ec05b8a3e5cba0ca734bfcad5560fb129e70ca/multidict-6.7.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:14c9e076eede3b54c636f8ce1c9c252b5f057c62131211f0ceeec273810c9721", size = 44715, upload-time = "2025-10-06T14:48:55.445Z" }, + { url = "https://files.pythonhosted.org/packages/4d/e1/866a5d77be6ea435711bef2a4291eed11032679b6b28b56b4776ab06ba3e/multidict-6.7.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:4c09703000a9d0fa3c3404b27041e574cc7f4df4c6563873246d0e11812a94b6", size = 44332, upload-time = "2025-10-06T14:48:56.706Z" }, + { url = "https://files.pythonhosted.org/packages/31/61/0c2d50241ada71ff61a79518db85ada85fdabfcf395d5968dae1cbda04e5/multidict-6.7.0-cp311-cp311-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:a265acbb7bb33a3a2d626afbe756371dce0279e7b17f4f4eda406459c2b5ff1c", size = 245212, upload-time = "2025-10-06T14:48:58.042Z" }, + { url = "https://files.pythonhosted.org/packages/ac/e0/919666a4e4b57fff1b57f279be1c9316e6cdc5de8a8b525d76f6598fefc7/multidict-6.7.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:51cb455de290ae462593e5b1cb1118c5c22ea7f0d3620d9940bf695cea5a4bd7", size = 246671, upload-time = "2025-10-06T14:49:00.004Z" }, + { url = "https://files.pythonhosted.org/packages/a1/cc/d027d9c5a520f3321b65adea289b965e7bcbd2c34402663f482648c716ce/multidict-6.7.0-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:db99677b4457c7a5c5a949353e125ba72d62b35f74e26da141530fbb012218a7", size = 225491, upload-time = "2025-10-06T14:49:01.393Z" }, + { url = "https://files.pythonhosted.org/packages/75/c4/bbd633980ce6155a28ff04e6a6492dd3335858394d7bb752d8b108708558/multidict-6.7.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:f470f68adc395e0183b92a2f4689264d1ea4b40504a24d9882c27375e6662bb9", size = 257322, upload-time = "2025-10-06T14:49:02.745Z" }, + { url = "https://files.pythonhosted.org/packages/4c/6d/d622322d344f1f053eae47e033b0b3f965af01212de21b10bcf91be991fb/multidict-6.7.0-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0db4956f82723cc1c270de9c6e799b4c341d327762ec78ef82bb962f79cc07d8", size = 254694, upload-time = "2025-10-06T14:49:04.15Z" }, + { url = "https://files.pythonhosted.org/packages/a8/9f/78f8761c2705d4c6d7516faed63c0ebdac569f6db1bef95e0d5218fdc146/multidict-6.7.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3e56d780c238f9e1ae66a22d2adf8d16f485381878250db8d496623cd38b22bd", size = 246715, upload-time = "2025-10-06T14:49:05.967Z" }, + { url = "https://files.pythonhosted.org/packages/78/59/950818e04f91b9c2b95aab3d923d9eabd01689d0dcd889563988e9ea0fd8/multidict-6.7.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:9d14baca2ee12c1a64740d4531356ba50b82543017f3ad6de0deb943c5979abb", size = 243189, upload-time = "2025-10-06T14:49:07.37Z" }, + { url = "https://files.pythonhosted.org/packages/7a/3d/77c79e1934cad2ee74991840f8a0110966d9599b3af95964c0cd79bb905b/multidict-6.7.0-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:295a92a76188917c7f99cda95858c822f9e4aae5824246bba9b6b44004ddd0a6", size = 237845, upload-time = "2025-10-06T14:49:08.759Z" }, + { url = "https://files.pythonhosted.org/packages/63/1b/834ce32a0a97a3b70f86437f685f880136677ac00d8bce0027e9fd9c2db7/multidict-6.7.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:39f1719f57adbb767ef592a50ae5ebb794220d1188f9ca93de471336401c34d2", size = 246374, upload-time = "2025-10-06T14:49:10.574Z" }, + { url = "https://files.pythonhosted.org/packages/23/ef/43d1c3ba205b5dec93dc97f3fba179dfa47910fc73aaaea4f7ceb41cec2a/multidict-6.7.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:0a13fb8e748dfc94749f622de065dd5c1def7e0d2216dba72b1d8069a389c6ff", size = 253345, upload-time = "2025-10-06T14:49:12.331Z" }, + { url = "https://files.pythonhosted.org/packages/6b/03/eaf95bcc2d19ead522001f6a650ef32811aa9e3624ff0ad37c445c7a588c/multidict-6.7.0-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:e3aa16de190d29a0ea1b48253c57d99a68492c8dd8948638073ab9e74dc9410b", size = 246940, upload-time = "2025-10-06T14:49:13.821Z" }, + { url = "https://files.pythonhosted.org/packages/e8/df/ec8a5fd66ea6cd6f525b1fcbb23511b033c3e9bc42b81384834ffa484a62/multidict-6.7.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:a048ce45dcdaaf1defb76b2e684f997fb5abf74437b6cb7b22ddad934a964e34", size = 242229, upload-time = "2025-10-06T14:49:15.603Z" }, + { url = "https://files.pythonhosted.org/packages/8a/a2/59b405d59fd39ec86d1142630e9049243015a5f5291ba49cadf3c090c541/multidict-6.7.0-cp311-cp311-win32.whl", hash = "sha256:a90af66facec4cebe4181b9e62a68be65e45ac9b52b67de9eec118701856e7ff", size = 41308, upload-time = "2025-10-06T14:49:16.871Z" }, + { url = "https://files.pythonhosted.org/packages/32/0f/13228f26f8b882c34da36efa776c3b7348455ec383bab4a66390e42963ae/multidict-6.7.0-cp311-cp311-win_amd64.whl", hash = "sha256:95b5ffa4349df2887518bb839409bcf22caa72d82beec453216802f475b23c81", size = 46037, upload-time = "2025-10-06T14:49:18.457Z" }, + { url = "https://files.pythonhosted.org/packages/84/1f/68588e31b000535a3207fd3c909ebeec4fb36b52c442107499c18a896a2a/multidict-6.7.0-cp311-cp311-win_arm64.whl", hash = "sha256:329aa225b085b6f004a4955271a7ba9f1087e39dcb7e65f6284a988264a63912", size = 43023, upload-time = "2025-10-06T14:49:19.648Z" }, + { url = "https://files.pythonhosted.org/packages/c2/9e/9f61ac18d9c8b475889f32ccfa91c9f59363480613fc807b6e3023d6f60b/multidict-6.7.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:8a3862568a36d26e650a19bb5cbbba14b71789032aebc0423f8cc5f150730184", size = 76877, upload-time = "2025-10-06T14:49:20.884Z" }, + { url = "https://files.pythonhosted.org/packages/38/6f/614f09a04e6184f8824268fce4bc925e9849edfa654ddd59f0b64508c595/multidict-6.7.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:960c60b5849b9b4f9dcc9bea6e3626143c252c74113df2c1540aebce70209b45", size = 45467, upload-time = "2025-10-06T14:49:22.054Z" }, + { url = "https://files.pythonhosted.org/packages/b3/93/c4f67a436dd026f2e780c433277fff72be79152894d9fc36f44569cab1a6/multidict-6.7.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:2049be98fb57a31b4ccf870bf377af2504d4ae35646a19037ec271e4c07998aa", size = 43834, upload-time = "2025-10-06T14:49:23.566Z" }, + { url = "https://files.pythonhosted.org/packages/7f/f5/013798161ca665e4a422afbc5e2d9e4070142a9ff8905e482139cd09e4d0/multidict-6.7.0-cp312-cp312-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:0934f3843a1860dd465d38895c17fce1f1cb37295149ab05cd1b9a03afacb2a7", size = 250545, upload-time = "2025-10-06T14:49:24.882Z" }, + { url = "https://files.pythonhosted.org/packages/71/2f/91dbac13e0ba94669ea5119ba267c9a832f0cb65419aca75549fcf09a3dc/multidict-6.7.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b3e34f3a1b8131ba06f1a73adab24f30934d148afcd5f5de9a73565a4404384e", size = 258305, upload-time = "2025-10-06T14:49:26.778Z" }, + { url = "https://files.pythonhosted.org/packages/ef/b0/754038b26f6e04488b48ac621f779c341338d78503fb45403755af2df477/multidict-6.7.0-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:efbb54e98446892590dc2458c19c10344ee9a883a79b5cec4bc34d6656e8d546", size = 242363, upload-time = "2025-10-06T14:49:28.562Z" }, + { url = "https://files.pythonhosted.org/packages/87/15/9da40b9336a7c9fa606c4cf2ed80a649dffeb42b905d4f63a1d7eb17d746/multidict-6.7.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a35c5fc61d4f51eb045061e7967cfe3123d622cd500e8868e7c0c592a09fedc4", size = 268375, upload-time = "2025-10-06T14:49:29.96Z" }, + { url = "https://files.pythonhosted.org/packages/82/72/c53fcade0cc94dfaad583105fd92b3a783af2091eddcb41a6d5a52474000/multidict-6.7.0-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:29fe6740ebccba4175af1b9b87bf553e9c15cd5868ee967e010efcf94e4fd0f1", size = 269346, upload-time = "2025-10-06T14:49:31.404Z" }, + { url = "https://files.pythonhosted.org/packages/0d/e2/9baffdae21a76f77ef8447f1a05a96ec4bc0a24dae08767abc0a2fe680b8/multidict-6.7.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:123e2a72e20537add2f33a79e605f6191fba2afda4cbb876e35c1a7074298a7d", size = 256107, upload-time = "2025-10-06T14:49:32.974Z" }, + { url = "https://files.pythonhosted.org/packages/3c/06/3f06f611087dc60d65ef775f1fb5aca7c6d61c6db4990e7cda0cef9b1651/multidict-6.7.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:b284e319754366c1aee2267a2036248b24eeb17ecd5dc16022095e747f2f4304", size = 253592, upload-time = "2025-10-06T14:49:34.52Z" }, + { url = "https://files.pythonhosted.org/packages/20/24/54e804ec7945b6023b340c412ce9c3f81e91b3bf5fa5ce65558740141bee/multidict-6.7.0-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:803d685de7be4303b5a657b76e2f6d1240e7e0a8aa2968ad5811fa2285553a12", size = 251024, upload-time = "2025-10-06T14:49:35.956Z" }, + { url = "https://files.pythonhosted.org/packages/14/48/011cba467ea0b17ceb938315d219391d3e421dfd35928e5dbdc3f4ae76ef/multidict-6.7.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:c04a328260dfd5db8c39538f999f02779012268f54614902d0afc775d44e0a62", size = 251484, upload-time = "2025-10-06T14:49:37.631Z" }, + { url = "https://files.pythonhosted.org/packages/0d/2f/919258b43bb35b99fa127435cfb2d91798eb3a943396631ef43e3720dcf4/multidict-6.7.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:8a19cdb57cd3df4cd865849d93ee14920fb97224300c88501f16ecfa2604b4e0", size = 263579, upload-time = "2025-10-06T14:49:39.502Z" }, + { url = "https://files.pythonhosted.org/packages/31/22/a0e884d86b5242b5a74cf08e876bdf299e413016b66e55511f7a804a366e/multidict-6.7.0-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:9b2fd74c52accced7e75de26023b7dccee62511a600e62311b918ec5c168fc2a", size = 259654, upload-time = "2025-10-06T14:49:41.32Z" }, + { url = "https://files.pythonhosted.org/packages/b2/e5/17e10e1b5c5f5a40f2fcbb45953c9b215f8a4098003915e46a93f5fcaa8f/multidict-6.7.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:3e8bfdd0e487acf992407a140d2589fe598238eaeffa3da8448d63a63cd363f8", size = 251511, upload-time = "2025-10-06T14:49:46.021Z" }, + { url = "https://files.pythonhosted.org/packages/e3/9a/201bb1e17e7af53139597069c375e7b0dcbd47594604f65c2d5359508566/multidict-6.7.0-cp312-cp312-win32.whl", hash = "sha256:dd32a49400a2c3d52088e120ee00c1e3576cbff7e10b98467962c74fdb762ed4", size = 41895, upload-time = "2025-10-06T14:49:48.718Z" }, + { url = "https://files.pythonhosted.org/packages/46/e2/348cd32faad84eaf1d20cce80e2bb0ef8d312c55bca1f7fa9865e7770aaf/multidict-6.7.0-cp312-cp312-win_amd64.whl", hash = "sha256:92abb658ef2d7ef22ac9f8bb88e8b6c3e571671534e029359b6d9e845923eb1b", size = 46073, upload-time = "2025-10-06T14:49:50.28Z" }, + { url = "https://files.pythonhosted.org/packages/25/ec/aad2613c1910dce907480e0c3aa306905830f25df2e54ccc9dea450cb5aa/multidict-6.7.0-cp312-cp312-win_arm64.whl", hash = "sha256:490dab541a6a642ce1a9d61a4781656b346a55c13038f0b1244653828e3a83ec", size = 43226, upload-time = "2025-10-06T14:49:52.304Z" }, + { url = "https://files.pythonhosted.org/packages/d2/86/33272a544eeb36d66e4d9a920602d1a2f57d4ebea4ef3cdfe5a912574c95/multidict-6.7.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:bee7c0588aa0076ce77c0ea5d19a68d76ad81fcd9fe8501003b9a24f9d4000f6", size = 76135, upload-time = "2025-10-06T14:49:54.26Z" }, + { url = "https://files.pythonhosted.org/packages/91/1c/eb97db117a1ebe46d457a3d235a7b9d2e6dcab174f42d1b67663dd9e5371/multidict-6.7.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:7ef6b61cad77091056ce0e7ce69814ef72afacb150b7ac6a3e9470def2198159", size = 45117, upload-time = "2025-10-06T14:49:55.82Z" }, + { url = "https://files.pythonhosted.org/packages/f1/d8/6c3442322e41fb1dd4de8bd67bfd11cd72352ac131f6368315617de752f1/multidict-6.7.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:9c0359b1ec12b1d6849c59f9d319610b7f20ef990a6d454ab151aa0e3b9f78ca", size = 43472, upload-time = "2025-10-06T14:49:57.048Z" }, + { url = "https://files.pythonhosted.org/packages/75/3f/e2639e80325af0b6c6febdf8e57cc07043ff15f57fa1ef808f4ccb5ac4cd/multidict-6.7.0-cp313-cp313-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:cd240939f71c64bd658f186330603aac1a9a81bf6273f523fca63673cb7378a8", size = 249342, upload-time = "2025-10-06T14:49:58.368Z" }, + { url = "https://files.pythonhosted.org/packages/5d/cc/84e0585f805cbeaa9cbdaa95f9a3d6aed745b9d25700623ac89a6ecff400/multidict-6.7.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a60a4d75718a5efa473ebd5ab685786ba0c67b8381f781d1be14da49f1a2dc60", size = 257082, upload-time = "2025-10-06T14:49:59.89Z" }, + { url = "https://files.pythonhosted.org/packages/b0/9c/ac851c107c92289acbbf5cfb485694084690c1b17e555f44952c26ddc5bd/multidict-6.7.0-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:53a42d364f323275126aff81fb67c5ca1b7a04fda0546245730a55c8c5f24bc4", size = 240704, upload-time = "2025-10-06T14:50:01.485Z" }, + { url = "https://files.pythonhosted.org/packages/50/cc/5f93e99427248c09da95b62d64b25748a5f5c98c7c2ab09825a1d6af0e15/multidict-6.7.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:3b29b980d0ddbecb736735ee5bef69bb2ddca56eff603c86f3f29a1128299b4f", size = 266355, upload-time = "2025-10-06T14:50:02.955Z" }, + { url = "https://files.pythonhosted.org/packages/ec/0c/2ec1d883ceb79c6f7f6d7ad90c919c898f5d1c6ea96d322751420211e072/multidict-6.7.0-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:f8a93b1c0ed2d04b97a5e9336fd2d33371b9a6e29ab7dd6503d63407c20ffbaf", size = 267259, upload-time = "2025-10-06T14:50:04.446Z" }, + { url = "https://files.pythonhosted.org/packages/c6/2d/f0b184fa88d6630aa267680bdb8623fb69cb0d024b8c6f0d23f9a0f406d3/multidict-6.7.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9ff96e8815eecacc6645da76c413eb3b3d34cfca256c70b16b286a687d013c32", size = 254903, upload-time = "2025-10-06T14:50:05.98Z" }, + { url = "https://files.pythonhosted.org/packages/06/c9/11ea263ad0df7dfabcad404feb3c0dd40b131bc7f232d5537f2fb1356951/multidict-6.7.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:7516c579652f6a6be0e266aec0acd0db80829ca305c3d771ed898538804c2036", size = 252365, upload-time = "2025-10-06T14:50:07.511Z" }, + { url = "https://files.pythonhosted.org/packages/41/88/d714b86ee2c17d6e09850c70c9d310abac3d808ab49dfa16b43aba9d53fd/multidict-6.7.0-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:040f393368e63fb0f3330e70c26bfd336656bed925e5cbe17c9da839a6ab13ec", size = 250062, upload-time = "2025-10-06T14:50:09.074Z" }, + { url = "https://files.pythonhosted.org/packages/15/fe/ad407bb9e818c2b31383f6131ca19ea7e35ce93cf1310fce69f12e89de75/multidict-6.7.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:b3bc26a951007b1057a1c543af845f1c7e3e71cc240ed1ace7bf4484aa99196e", size = 249683, upload-time = "2025-10-06T14:50:10.714Z" }, + { url = "https://files.pythonhosted.org/packages/8c/a4/a89abdb0229e533fb925e7c6e5c40201c2873efebc9abaf14046a4536ee6/multidict-6.7.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:7b022717c748dd1992a83e219587aabe45980d88969f01b316e78683e6285f64", size = 261254, upload-time = "2025-10-06T14:50:12.28Z" }, + { url = "https://files.pythonhosted.org/packages/8d/aa/0e2b27bd88b40a4fb8dc53dd74eecac70edaa4c1dd0707eb2164da3675b3/multidict-6.7.0-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:9600082733859f00d79dee64effc7aef1beb26adb297416a4ad2116fd61374bd", size = 257967, upload-time = "2025-10-06T14:50:14.16Z" }, + { url = "https://files.pythonhosted.org/packages/d0/8e/0c67b7120d5d5f6d874ed85a085f9dc770a7f9d8813e80f44a9fec820bb7/multidict-6.7.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:94218fcec4d72bc61df51c198d098ce2b378e0ccbac41ddbed5ef44092913288", size = 250085, upload-time = "2025-10-06T14:50:15.639Z" }, + { url = "https://files.pythonhosted.org/packages/ba/55/b73e1d624ea4b8fd4dd07a3bb70f6e4c7c6c5d9d640a41c6ffe5cdbd2a55/multidict-6.7.0-cp313-cp313-win32.whl", hash = "sha256:a37bd74c3fa9d00be2d7b8eca074dc56bd8077ddd2917a839bd989612671ed17", size = 41713, upload-time = "2025-10-06T14:50:17.066Z" }, + { url = "https://files.pythonhosted.org/packages/32/31/75c59e7d3b4205075b4c183fa4ca398a2daf2303ddf616b04ae6ef55cffe/multidict-6.7.0-cp313-cp313-win_amd64.whl", hash = "sha256:30d193c6cc6d559db42b6bcec8a5d395d34d60c9877a0b71ecd7c204fcf15390", size = 45915, upload-time = "2025-10-06T14:50:18.264Z" }, + { url = "https://files.pythonhosted.org/packages/31/2a/8987831e811f1184c22bc2e45844934385363ee61c0a2dcfa8f71b87e608/multidict-6.7.0-cp313-cp313-win_arm64.whl", hash = "sha256:ea3334cabe4d41b7ccd01e4d349828678794edbc2d3ae97fc162a3312095092e", size = 43077, upload-time = "2025-10-06T14:50:19.853Z" }, + { url = "https://files.pythonhosted.org/packages/e8/68/7b3a5170a382a340147337b300b9eb25a9ddb573bcdfff19c0fa3f31ffba/multidict-6.7.0-cp313-cp313t-macosx_10_13_universal2.whl", hash = "sha256:ad9ce259f50abd98a1ca0aa6e490b58c316a0fce0617f609723e40804add2c00", size = 83114, upload-time = "2025-10-06T14:50:21.223Z" }, + { url = "https://files.pythonhosted.org/packages/55/5c/3fa2d07c84df4e302060f555bbf539310980362236ad49f50eeb0a1c1eb9/multidict-6.7.0-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:07f5594ac6d084cbb5de2df218d78baf55ef150b91f0ff8a21cc7a2e3a5a58eb", size = 48442, upload-time = "2025-10-06T14:50:22.871Z" }, + { url = "https://files.pythonhosted.org/packages/fc/56/67212d33239797f9bd91962bb899d72bb0f4c35a8652dcdb8ed049bef878/multidict-6.7.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:0591b48acf279821a579282444814a2d8d0af624ae0bc600aa4d1b920b6e924b", size = 46885, upload-time = "2025-10-06T14:50:24.258Z" }, + { url = "https://files.pythonhosted.org/packages/46/d1/908f896224290350721597a61a69cd19b89ad8ee0ae1f38b3f5cd12ea2ac/multidict-6.7.0-cp313-cp313t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:749a72584761531d2b9467cfbdfd29487ee21124c304c4b6cb760d8777b27f9c", size = 242588, upload-time = "2025-10-06T14:50:25.716Z" }, + { url = "https://files.pythonhosted.org/packages/ab/67/8604288bbd68680eee0ab568fdcb56171d8b23a01bcd5cb0c8fedf6e5d99/multidict-6.7.0-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6b4c3d199f953acd5b446bf7c0de1fe25d94e09e79086f8dc2f48a11a129cdf1", size = 249966, upload-time = "2025-10-06T14:50:28.192Z" }, + { url = "https://files.pythonhosted.org/packages/20/33/9228d76339f1ba51e3efef7da3ebd91964d3006217aae13211653193c3ff/multidict-6.7.0-cp313-cp313t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:9fb0211dfc3b51efea2f349ec92c114d7754dd62c01f81c3e32b765b70c45c9b", size = 228618, upload-time = "2025-10-06T14:50:29.82Z" }, + { url = "https://files.pythonhosted.org/packages/f8/2d/25d9b566d10cab1c42b3b9e5b11ef79c9111eaf4463b8c257a3bd89e0ead/multidict-6.7.0-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a027ec240fe73a8d6281872690b988eed307cd7d91b23998ff35ff577ca688b5", size = 257539, upload-time = "2025-10-06T14:50:31.731Z" }, + { url = "https://files.pythonhosted.org/packages/b6/b1/8d1a965e6637fc33de3c0d8f414485c2b7e4af00f42cab3d84e7b955c222/multidict-6.7.0-cp313-cp313t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d1d964afecdf3a8288789df2f5751dc0a8261138c3768d9af117ed384e538fad", size = 256345, upload-time = "2025-10-06T14:50:33.26Z" }, + { url = "https://files.pythonhosted.org/packages/ba/0c/06b5a8adbdeedada6f4fb8d8f193d44a347223b11939b42953eeb6530b6b/multidict-6.7.0-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:caf53b15b1b7df9fbd0709aa01409000a2b4dd03a5f6f5cc548183c7c8f8b63c", size = 247934, upload-time = "2025-10-06T14:50:34.808Z" }, + { url = "https://files.pythonhosted.org/packages/8f/31/b2491b5fe167ca044c6eb4b8f2c9f3b8a00b24c432c365358eadac5d7625/multidict-6.7.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:654030da3197d927f05a536a66186070e98765aa5142794c9904555d3a9d8fb5", size = 245243, upload-time = "2025-10-06T14:50:36.436Z" }, + { url = "https://files.pythonhosted.org/packages/61/1a/982913957cb90406c8c94f53001abd9eafc271cb3e70ff6371590bec478e/multidict-6.7.0-cp313-cp313t-musllinux_1_2_armv7l.whl", hash = "sha256:2090d3718829d1e484706a2f525e50c892237b2bf9b17a79b059cb98cddc2f10", size = 235878, upload-time = "2025-10-06T14:50:37.953Z" }, + { url = "https://files.pythonhosted.org/packages/be/c0/21435d804c1a1cf7a2608593f4d19bca5bcbd7a81a70b253fdd1c12af9c0/multidict-6.7.0-cp313-cp313t-musllinux_1_2_i686.whl", hash = "sha256:2d2cfeec3f6f45651b3d408c4acec0ebf3daa9bc8a112a084206f5db5d05b754", size = 243452, upload-time = "2025-10-06T14:50:39.574Z" }, + { url = "https://files.pythonhosted.org/packages/54/0a/4349d540d4a883863191be6eb9a928846d4ec0ea007d3dcd36323bb058ac/multidict-6.7.0-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:4ef089f985b8c194d341eb2c24ae6e7408c9a0e2e5658699c92f497437d88c3c", size = 252312, upload-time = "2025-10-06T14:50:41.612Z" }, + { url = "https://files.pythonhosted.org/packages/26/64/d5416038dbda1488daf16b676e4dbfd9674dde10a0cc8f4fc2b502d8125d/multidict-6.7.0-cp313-cp313t-musllinux_1_2_s390x.whl", hash = "sha256:e93a0617cd16998784bf4414c7e40f17a35d2350e5c6f0bd900d3a8e02bd3762", size = 246935, upload-time = "2025-10-06T14:50:43.972Z" }, + { url = "https://files.pythonhosted.org/packages/9f/8c/8290c50d14e49f35e0bd4abc25e1bc7711149ca9588ab7d04f886cdf03d9/multidict-6.7.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:f0feece2ef8ebc42ed9e2e8c78fc4aa3cf455733b507c09ef7406364c94376c6", size = 243385, upload-time = "2025-10-06T14:50:45.648Z" }, + { url = "https://files.pythonhosted.org/packages/ef/a0/f83ae75e42d694b3fbad3e047670e511c138be747bc713cf1b10d5096416/multidict-6.7.0-cp313-cp313t-win32.whl", hash = "sha256:19a1d55338ec1be74ef62440ca9e04a2f001a04d0cc49a4983dc320ff0f3212d", size = 47777, upload-time = "2025-10-06T14:50:47.154Z" }, + { url = "https://files.pythonhosted.org/packages/dc/80/9b174a92814a3830b7357307a792300f42c9e94664b01dee8e457551fa66/multidict-6.7.0-cp313-cp313t-win_amd64.whl", hash = "sha256:3da4fb467498df97e986af166b12d01f05d2e04f978a9c1c680ea1988e0bc4b6", size = 53104, upload-time = "2025-10-06T14:50:48.851Z" }, + { url = "https://files.pythonhosted.org/packages/cc/28/04baeaf0428d95bb7a7bea0e691ba2f31394338ba424fb0679a9ed0f4c09/multidict-6.7.0-cp313-cp313t-win_arm64.whl", hash = "sha256:b4121773c49a0776461f4a904cdf6264c88e42218aaa8407e803ca8025872792", size = 45503, upload-time = "2025-10-06T14:50:50.16Z" }, + { url = "https://files.pythonhosted.org/packages/e2/b1/3da6934455dd4b261d4c72f897e3a5728eba81db59959f3a639245891baa/multidict-6.7.0-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:3bab1e4aff7adaa34410f93b1f8e57c4b36b9af0426a76003f441ee1d3c7e842", size = 75128, upload-time = "2025-10-06T14:50:51.92Z" }, + { url = "https://files.pythonhosted.org/packages/14/2c/f069cab5b51d175a1a2cb4ccdf7a2c2dabd58aa5bd933fa036a8d15e2404/multidict-6.7.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:b8512bac933afc3e45fb2b18da8e59b78d4f408399a960339598374d4ae3b56b", size = 44410, upload-time = "2025-10-06T14:50:53.275Z" }, + { url = "https://files.pythonhosted.org/packages/42/e2/64bb41266427af6642b6b128e8774ed84c11b80a90702c13ac0a86bb10cc/multidict-6.7.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:79dcf9e477bc65414ebfea98ffd013cb39552b5ecd62908752e0e413d6d06e38", size = 43205, upload-time = "2025-10-06T14:50:54.911Z" }, + { url = "https://files.pythonhosted.org/packages/02/68/6b086fef8a3f1a8541b9236c594f0c9245617c29841f2e0395d979485cde/multidict-6.7.0-cp314-cp314-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:31bae522710064b5cbeddaf2e9f32b1abab70ac6ac91d42572502299e9953128", size = 245084, upload-time = "2025-10-06T14:50:56.369Z" }, + { url = "https://files.pythonhosted.org/packages/15/ee/f524093232007cd7a75c1d132df70f235cfd590a7c9eaccd7ff422ef4ae8/multidict-6.7.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4a0df7ff02397bb63e2fd22af2c87dfa39e8c7f12947bc524dbdc528282c7e34", size = 252667, upload-time = "2025-10-06T14:50:57.991Z" }, + { url = "https://files.pythonhosted.org/packages/02/a5/eeb3f43ab45878f1895118c3ef157a480db58ede3f248e29b5354139c2c9/multidict-6.7.0-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:7a0222514e8e4c514660e182d5156a415c13ef0aabbd71682fc714e327b95e99", size = 233590, upload-time = "2025-10-06T14:50:59.589Z" }, + { url = "https://files.pythonhosted.org/packages/6a/1e/76d02f8270b97269d7e3dbd45644b1785bda457b474315f8cf999525a193/multidict-6.7.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:2397ab4daaf2698eb51a76721e98db21ce4f52339e535725de03ea962b5a3202", size = 264112, upload-time = "2025-10-06T14:51:01.183Z" }, + { url = "https://files.pythonhosted.org/packages/76/0b/c28a70ecb58963847c2a8efe334904cd254812b10e535aefb3bcce513918/multidict-6.7.0-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:8891681594162635948a636c9fe0ff21746aeb3dd5463f6e25d9bea3a8a39ca1", size = 261194, upload-time = "2025-10-06T14:51:02.794Z" }, + { url = "https://files.pythonhosted.org/packages/b4/63/2ab26e4209773223159b83aa32721b4021ffb08102f8ac7d689c943fded1/multidict-6.7.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:18706cc31dbf402a7945916dd5cddf160251b6dab8a2c5f3d6d5a55949f676b3", size = 248510, upload-time = "2025-10-06T14:51:04.724Z" }, + { url = "https://files.pythonhosted.org/packages/93/cd/06c1fa8282af1d1c46fd55c10a7930af652afdce43999501d4d68664170c/multidict-6.7.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:f844a1bbf1d207dd311a56f383f7eda2d0e134921d45751842d8235e7778965d", size = 248395, upload-time = "2025-10-06T14:51:06.306Z" }, + { url = "https://files.pythonhosted.org/packages/99/ac/82cb419dd6b04ccf9e7e61befc00c77614fc8134362488b553402ecd55ce/multidict-6.7.0-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:d4393e3581e84e5645506923816b9cc81f5609a778c7e7534054091acc64d1c6", size = 239520, upload-time = "2025-10-06T14:51:08.091Z" }, + { url = "https://files.pythonhosted.org/packages/fa/f3/a0f9bf09493421bd8716a362e0cd1d244f5a6550f5beffdd6b47e885b331/multidict-6.7.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:fbd18dc82d7bf274b37aa48d664534330af744e03bccf696d6f4c6042e7d19e7", size = 245479, upload-time = "2025-10-06T14:51:10.365Z" }, + { url = "https://files.pythonhosted.org/packages/8d/01/476d38fc73a212843f43c852b0eee266b6971f0e28329c2184a8df90c376/multidict-6.7.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:b6234e14f9314731ec45c42fc4554b88133ad53a09092cc48a88e771c125dadb", size = 258903, upload-time = "2025-10-06T14:51:12.466Z" }, + { url = "https://files.pythonhosted.org/packages/49/6d/23faeb0868adba613b817d0e69c5f15531b24d462af8012c4f6de4fa8dc3/multidict-6.7.0-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:08d4379f9744d8f78d98c8673c06e202ffa88296f009c71bbafe8a6bf847d01f", size = 252333, upload-time = "2025-10-06T14:51:14.48Z" }, + { url = "https://files.pythonhosted.org/packages/1e/cc/48d02ac22b30fa247f7dad82866e4b1015431092f4ba6ebc7e77596e0b18/multidict-6.7.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:9fe04da3f79387f450fd0061d4dd2e45a72749d31bf634aecc9e27f24fdc4b3f", size = 243411, upload-time = "2025-10-06T14:51:16.072Z" }, + { url = "https://files.pythonhosted.org/packages/4a/03/29a8bf5a18abf1fe34535c88adbdfa88c9fb869b5a3b120692c64abe8284/multidict-6.7.0-cp314-cp314-win32.whl", hash = "sha256:fbafe31d191dfa7c4c51f7a6149c9fb7e914dcf9ffead27dcfd9f1ae382b3885", size = 40940, upload-time = "2025-10-06T14:51:17.544Z" }, + { url = "https://files.pythonhosted.org/packages/82/16/7ed27b680791b939de138f906d5cf2b4657b0d45ca6f5dd6236fdddafb1a/multidict-6.7.0-cp314-cp314-win_amd64.whl", hash = "sha256:2f67396ec0310764b9222a1728ced1ab638f61aadc6226f17a71dd9324f9a99c", size = 45087, upload-time = "2025-10-06T14:51:18.875Z" }, + { url = "https://files.pythonhosted.org/packages/cd/3c/e3e62eb35a1950292fe39315d3c89941e30a9d07d5d2df42965ab041da43/multidict-6.7.0-cp314-cp314-win_arm64.whl", hash = "sha256:ba672b26069957ee369cfa7fc180dde1fc6f176eaf1e6beaf61fbebbd3d9c000", size = 42368, upload-time = "2025-10-06T14:51:20.225Z" }, + { url = "https://files.pythonhosted.org/packages/8b/40/cd499bd0dbc5f1136726db3153042a735fffd0d77268e2ee20d5f33c010f/multidict-6.7.0-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:c1dcc7524066fa918c6a27d61444d4ee7900ec635779058571f70d042d86ed63", size = 82326, upload-time = "2025-10-06T14:51:21.588Z" }, + { url = "https://files.pythonhosted.org/packages/13/8a/18e031eca251c8df76daf0288e6790561806e439f5ce99a170b4af30676b/multidict-6.7.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:27e0b36c2d388dc7b6ced3406671b401e84ad7eb0656b8f3a2f46ed0ce483718", size = 48065, upload-time = "2025-10-06T14:51:22.93Z" }, + { url = "https://files.pythonhosted.org/packages/40/71/5e6701277470a87d234e433fb0a3a7deaf3bcd92566e421e7ae9776319de/multidict-6.7.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:2a7baa46a22e77f0988e3b23d4ede5513ebec1929e34ee9495be535662c0dfe2", size = 46475, upload-time = "2025-10-06T14:51:24.352Z" }, + { url = "https://files.pythonhosted.org/packages/fe/6a/bab00cbab6d9cfb57afe1663318f72ec28289ea03fd4e8236bb78429893a/multidict-6.7.0-cp314-cp314t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:7bf77f54997a9166a2f5675d1201520586439424c2511723a7312bdb4bcc034e", size = 239324, upload-time = "2025-10-06T14:51:25.822Z" }, + { url = "https://files.pythonhosted.org/packages/2a/5f/8de95f629fc22a7769ade8b41028e3e5a822c1f8904f618d175945a81ad3/multidict-6.7.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e011555abada53f1578d63389610ac8a5400fc70ce71156b0aa30d326f1a5064", size = 246877, upload-time = "2025-10-06T14:51:27.604Z" }, + { url = "https://files.pythonhosted.org/packages/23/b4/38881a960458f25b89e9f4a4fdcb02ac101cfa710190db6e5528841e67de/multidict-6.7.0-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:28b37063541b897fd6a318007373930a75ca6d6ac7c940dbe14731ffdd8d498e", size = 225824, upload-time = "2025-10-06T14:51:29.664Z" }, + { url = "https://files.pythonhosted.org/packages/1e/39/6566210c83f8a261575f18e7144736059f0c460b362e96e9cf797a24b8e7/multidict-6.7.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:05047ada7a2fde2631a0ed706f1fd68b169a681dfe5e4cf0f8e4cb6618bbc2cd", size = 253558, upload-time = "2025-10-06T14:51:31.684Z" }, + { url = "https://files.pythonhosted.org/packages/00/a3/67f18315100f64c269f46e6c0319fa87ba68f0f64f2b8e7fd7c72b913a0b/multidict-6.7.0-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:716133f7d1d946a4e1b91b1756b23c088881e70ff180c24e864c26192ad7534a", size = 252339, upload-time = "2025-10-06T14:51:33.699Z" }, + { url = "https://files.pythonhosted.org/packages/c8/2a/1cb77266afee2458d82f50da41beba02159b1d6b1f7973afc9a1cad1499b/multidict-6.7.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d1bed1b467ef657f2a0ae62844a607909ef1c6889562de5e1d505f74457d0b96", size = 244895, upload-time = "2025-10-06T14:51:36.189Z" }, + { url = "https://files.pythonhosted.org/packages/dd/72/09fa7dd487f119b2eb9524946ddd36e2067c08510576d43ff68469563b3b/multidict-6.7.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:ca43bdfa5d37bd6aee89d85e1d0831fb86e25541be7e9d376ead1b28974f8e5e", size = 241862, upload-time = "2025-10-06T14:51:41.291Z" }, + { url = "https://files.pythonhosted.org/packages/65/92/bc1f8bd0853d8669300f732c801974dfc3702c3eeadae2f60cef54dc69d7/multidict-6.7.0-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:44b546bd3eb645fd26fb949e43c02a25a2e632e2ca21a35e2e132c8105dc8599", size = 232376, upload-time = "2025-10-06T14:51:43.55Z" }, + { url = "https://files.pythonhosted.org/packages/09/86/ac39399e5cb9d0c2ac8ef6e10a768e4d3bc933ac808d49c41f9dc23337eb/multidict-6.7.0-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:a6ef16328011d3f468e7ebc326f24c1445f001ca1dec335b2f8e66bed3006394", size = 240272, upload-time = "2025-10-06T14:51:45.265Z" }, + { url = "https://files.pythonhosted.org/packages/3d/b6/fed5ac6b8563ec72df6cb1ea8dac6d17f0a4a1f65045f66b6d3bf1497c02/multidict-6.7.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:5aa873cbc8e593d361ae65c68f85faadd755c3295ea2c12040ee146802f23b38", size = 248774, upload-time = "2025-10-06T14:51:46.836Z" }, + { url = "https://files.pythonhosted.org/packages/6b/8d/b954d8c0dc132b68f760aefd45870978deec6818897389dace00fcde32ff/multidict-6.7.0-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:3d7b6ccce016e29df4b7ca819659f516f0bc7a4b3efa3bb2012ba06431b044f9", size = 242731, upload-time = "2025-10-06T14:51:48.541Z" }, + { url = "https://files.pythonhosted.org/packages/16/9d/a2dac7009125d3540c2f54e194829ea18ac53716c61b655d8ed300120b0f/multidict-6.7.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:171b73bd4ee683d307599b66793ac80981b06f069b62eea1c9e29c9241aa66b0", size = 240193, upload-time = "2025-10-06T14:51:50.355Z" }, + { url = "https://files.pythonhosted.org/packages/39/ca/c05f144128ea232ae2178b008d5011d4e2cea86e4ee8c85c2631b1b94802/multidict-6.7.0-cp314-cp314t-win32.whl", hash = "sha256:b2d7f80c4e1fd010b07cb26820aae86b7e73b681ee4889684fb8d2d4537aab13", size = 48023, upload-time = "2025-10-06T14:51:51.883Z" }, + { url = "https://files.pythonhosted.org/packages/ba/8f/0a60e501584145588be1af5cc829265701ba3c35a64aec8e07cbb71d39bb/multidict-6.7.0-cp314-cp314t-win_amd64.whl", hash = "sha256:09929cab6fcb68122776d575e03c6cc64ee0b8fca48d17e135474b042ce515cd", size = 53507, upload-time = "2025-10-06T14:51:53.672Z" }, + { url = "https://files.pythonhosted.org/packages/7f/ae/3148b988a9c6239903e786eac19c889fab607c31d6efa7fb2147e5680f23/multidict-6.7.0-cp314-cp314t-win_arm64.whl", hash = "sha256:cc41db090ed742f32bd2d2c721861725e6109681eddf835d0a82bd3a5c382827", size = 44804, upload-time = "2025-10-06T14:51:55.415Z" }, + { url = "https://files.pythonhosted.org/packages/b7/da/7d22601b625e241d4f23ef1ebff8acfc60da633c9e7e7922e24d10f592b3/multidict-6.7.0-py3-none-any.whl", hash = "sha256:394fc5c42a333c9ffc3e421a4c85e08580d990e08b99f6bf35b4132114c5dcb3", size = 12317, upload-time = "2025-10-06T14:52:29.272Z" }, +] + +[[package]] +name = "multiprocess" +version = "0.70.16" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "dill" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/b5/ae/04f39c5d0d0def03247c2893d6f2b83c136bf3320a2154d7b8858f2ba72d/multiprocess-0.70.16.tar.gz", hash = "sha256:161af703d4652a0e1410be6abccecde4a7ddffd19341be0a7011b94aeb171ac1", size = 1772603, upload-time = "2024-01-28T18:52:34.85Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/ef/76/6e712a2623d146d314f17598df5de7224c85c0060ef63fd95cc15a25b3fa/multiprocess-0.70.16-pp310-pypy310_pp73-macosx_10_13_x86_64.whl", hash = "sha256:476887be10e2f59ff183c006af746cb6f1fd0eadcfd4ef49e605cbe2659920ee", size = 134980, upload-time = "2024-01-28T18:52:15.731Z" }, + { url = "https://files.pythonhosted.org/packages/0f/ab/1e6e8009e380e22254ff539ebe117861e5bdb3bff1fc977920972237c6c7/multiprocess-0.70.16-pp310-pypy310_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:d951bed82c8f73929ac82c61f01a7b5ce8f3e5ef40f5b52553b4f547ce2b08ec", size = 134982, upload-time = "2024-01-28T18:52:17.783Z" }, + { url = "https://files.pythonhosted.org/packages/bc/f7/7ec7fddc92e50714ea3745631f79bd9c96424cb2702632521028e57d3a36/multiprocess-0.70.16-py310-none-any.whl", hash = "sha256:c4a9944c67bd49f823687463660a2d6daae94c289adff97e0f9d696ba6371d02", size = 134824, upload-time = "2024-01-28T18:52:26.062Z" }, + { url = "https://files.pythonhosted.org/packages/50/15/b56e50e8debaf439f44befec5b2af11db85f6e0f344c3113ae0be0593a91/multiprocess-0.70.16-py311-none-any.whl", hash = "sha256:af4cabb0dac72abfb1e794fa7855c325fd2b55a10a44628a3c1ad3311c04127a", size = 143519, upload-time = "2024-01-28T18:52:28.115Z" }, + { url = "https://files.pythonhosted.org/packages/0a/7d/a988f258104dcd2ccf1ed40fdc97e26c4ac351eeaf81d76e266c52d84e2f/multiprocess-0.70.16-py312-none-any.whl", hash = "sha256:fc0544c531920dde3b00c29863377f87e1632601092ea2daca74e4beb40faa2e", size = 146741, upload-time = "2024-01-28T18:52:29.395Z" }, + { url = "https://files.pythonhosted.org/packages/ea/89/38df130f2c799090c978b366cfdf5b96d08de5b29a4a293df7f7429fa50b/multiprocess-0.70.16-py38-none-any.whl", hash = "sha256:a71d82033454891091a226dfc319d0cfa8019a4e888ef9ca910372a446de4435", size = 132628, upload-time = "2024-01-28T18:52:30.853Z" }, + { url = "https://files.pythonhosted.org/packages/da/d9/f7f9379981e39b8c2511c9e0326d212accacb82f12fbfdc1aa2ce2a7b2b6/multiprocess-0.70.16-py39-none-any.whl", hash = "sha256:a0bafd3ae1b732eac64be2e72038231c1ba97724b60b09400d68f229fcc2fbf3", size = 133351, upload-time = "2024-01-28T18:52:31.981Z" }, +] + +[[package]] +name = "networkx" +version = "3.4.2" +source = { registry = "https://pypi.org/simple" } +resolution-markers = [ + "python_full_version < '3.11' and sys_platform == 'darwin'", + "python_full_version < '3.11' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux'", +] +sdist = { url = "https://files.pythonhosted.org/packages/fd/1d/06475e1cd5264c0b870ea2cc6fdb3e37177c1e565c43f56ff17a10e3937f/networkx-3.4.2.tar.gz", hash = "sha256:307c3669428c5362aab27c8a1260aa8f47c4e91d3891f48be0141738d8d053e1", size = 2151368, upload-time = "2024-10-21T12:39:38.695Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/b9/54/dd730b32ea14ea797530a4479b2ed46a6fb250f682a9cfb997e968bf0261/networkx-3.4.2-py3-none-any.whl", hash = "sha256:df5d4365b724cf81b8c6a7312509d0c22386097011ad1abe274afd5e9d3bbc5f", size = 1723263, upload-time = "2024-10-21T12:39:36.247Z" }, +] + +[[package]] +name = "networkx" +version = "3.6rc0" +source = { registry = "https://pypi.org/simple" } +resolution-markers = [ + "python_full_version >= '3.13' and sys_platform == 'darwin'", + "python_full_version >= '3.13' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version >= '3.13' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version >= '3.13' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.12.*' and sys_platform == 'darwin'", + "python_full_version == '3.12.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.11.*' and sys_platform == 'darwin'", + "python_full_version == '3.11.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and sys_platform != 'darwin' and sys_platform != 'linux'", +] +sdist = { url = "https://files.pythonhosted.org/packages/94/9a/36478bc8f4ec62071657ab52017b9884ef75c934f5a3596ae2cdff269f01/networkx-3.6rc0.tar.gz", hash = "sha256:025d33515a05844d3bece4cbda3a96485aa41aaf1d03fda937ba6e7d6118ebdd", size = 2508712, upload-time = "2025-11-04T15:20:51.025Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/b8/4d/41a32aefff47c4a7e17869cfa823368d6c8f8be825cfd7b8611b1365b0c3/networkx-3.6rc0-py3-none-any.whl", hash = "sha256:86ea7545ac5b7e9a3f5ec35b35fb428d511e84cfb432499148a490024099aa84", size = 2061866, upload-time = "2025-11-04T15:20:48.9Z" }, +] + +[[package]] +name = "nodeenv" +version = "1.9.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/43/16/fc88b08840de0e0a72a2f9d8c6bae36be573e475a6326ae854bcc549fc45/nodeenv-1.9.1.tar.gz", hash = "sha256:6ec12890a2dab7946721edbfbcd91f3319c6ccc9aec47be7c7e6b7011ee6645f", size = 47437, upload-time = "2024-06-04T18:44:11.171Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/d2/1d/1b658dbd2b9fa9c4c9f32accbfc0205d532c8c6194dc0f2a4c0428e7128a/nodeenv-1.9.1-py2.py3-none-any.whl", hash = "sha256:ba11c9782d29c27c70ffbdda2d7415098754709be8a7056d79a737cd901155c9", size = 22314, upload-time = "2024-06-04T18:44:08.352Z" }, +] + +[[package]] +name = "numpy" +version = "2.2.6" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/76/21/7d2a95e4bba9dc13d043ee156a356c0a8f0c6309dff6b21b4d71a073b8a8/numpy-2.2.6.tar.gz", hash = "sha256:e29554e2bef54a90aa5cc07da6ce955accb83f21ab5de01a62c8478897b264fd", size = 20276440, upload-time = "2025-05-17T22:38:04.611Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/9a/3e/ed6db5be21ce87955c0cbd3009f2803f59fa08df21b5df06862e2d8e2bdd/numpy-2.2.6-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:b412caa66f72040e6d268491a59f2c43bf03eb6c96dd8f0307829feb7fa2b6fb", size = 21165245, upload-time = "2025-05-17T21:27:58.555Z" }, + { url = "https://files.pythonhosted.org/packages/22/c2/4b9221495b2a132cc9d2eb862e21d42a009f5a60e45fc44b00118c174bff/numpy-2.2.6-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:8e41fd67c52b86603a91c1a505ebaef50b3314de0213461c7a6e99c9a3beff90", size = 14360048, upload-time = "2025-05-17T21:28:21.406Z" }, + { url = "https://files.pythonhosted.org/packages/fd/77/dc2fcfc66943c6410e2bf598062f5959372735ffda175b39906d54f02349/numpy-2.2.6-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:37e990a01ae6ec7fe7fa1c26c55ecb672dd98b19c3d0e1d1f326fa13cb38d163", size = 5340542, upload-time = "2025-05-17T21:28:30.931Z" }, + { url = "https://files.pythonhosted.org/packages/7a/4f/1cb5fdc353a5f5cc7feb692db9b8ec2c3d6405453f982435efc52561df58/numpy-2.2.6-cp310-cp310-macosx_14_0_x86_64.whl", hash = "sha256:5a6429d4be8ca66d889b7cf70f536a397dc45ba6faeb5f8c5427935d9592e9cf", size = 6878301, upload-time = "2025-05-17T21:28:41.613Z" }, + { url = "https://files.pythonhosted.org/packages/eb/17/96a3acd228cec142fcb8723bd3cc39c2a474f7dcf0a5d16731980bcafa95/numpy-2.2.6-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:efd28d4e9cd7d7a8d39074a4d44c63eda73401580c5c76acda2ce969e0a38e83", size = 14297320, upload-time = "2025-05-17T21:29:02.78Z" }, + { url = "https://files.pythonhosted.org/packages/b4/63/3de6a34ad7ad6646ac7d2f55ebc6ad439dbbf9c4370017c50cf403fb19b5/numpy-2.2.6-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fc7b73d02efb0e18c000e9ad8b83480dfcd5dfd11065997ed4c6747470ae8915", size = 16801050, upload-time = "2025-05-17T21:29:27.675Z" }, + { url = "https://files.pythonhosted.org/packages/07/b6/89d837eddef52b3d0cec5c6ba0456c1bf1b9ef6a6672fc2b7873c3ec4e2e/numpy-2.2.6-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:74d4531beb257d2c3f4b261bfb0fc09e0f9ebb8842d82a7b4209415896adc680", size = 15807034, upload-time = "2025-05-17T21:29:51.102Z" }, + { url = "https://files.pythonhosted.org/packages/01/c8/dc6ae86e3c61cfec1f178e5c9f7858584049b6093f843bca541f94120920/numpy-2.2.6-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:8fc377d995680230e83241d8a96def29f204b5782f371c532579b4f20607a289", size = 18614185, upload-time = "2025-05-17T21:30:18.703Z" }, + { url = "https://files.pythonhosted.org/packages/5b/c5/0064b1b7e7c89137b471ccec1fd2282fceaae0ab3a9550f2568782d80357/numpy-2.2.6-cp310-cp310-win32.whl", hash = "sha256:b093dd74e50a8cba3e873868d9e93a85b78e0daf2e98c6797566ad8044e8363d", size = 6527149, upload-time = "2025-05-17T21:30:29.788Z" }, + { url = "https://files.pythonhosted.org/packages/a3/dd/4b822569d6b96c39d1215dbae0582fd99954dcbcf0c1a13c61783feaca3f/numpy-2.2.6-cp310-cp310-win_amd64.whl", hash = "sha256:f0fd6321b839904e15c46e0d257fdd101dd7f530fe03fd6359c1ea63738703f3", size = 12904620, upload-time = "2025-05-17T21:30:48.994Z" }, + { url = "https://files.pythonhosted.org/packages/da/a8/4f83e2aa666a9fbf56d6118faaaf5f1974d456b1823fda0a176eff722839/numpy-2.2.6-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:f9f1adb22318e121c5c69a09142811a201ef17ab257a1e66ca3025065b7f53ae", size = 21176963, upload-time = "2025-05-17T21:31:19.36Z" }, + { url = "https://files.pythonhosted.org/packages/b3/2b/64e1affc7972decb74c9e29e5649fac940514910960ba25cd9af4488b66c/numpy-2.2.6-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:c820a93b0255bc360f53eca31a0e676fd1101f673dda8da93454a12e23fc5f7a", size = 14406743, upload-time = "2025-05-17T21:31:41.087Z" }, + { url = "https://files.pythonhosted.org/packages/4a/9f/0121e375000b5e50ffdd8b25bf78d8e1a5aa4cca3f185d41265198c7b834/numpy-2.2.6-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:3d70692235e759f260c3d837193090014aebdf026dfd167834bcba43e30c2a42", size = 5352616, upload-time = "2025-05-17T21:31:50.072Z" }, + { url = "https://files.pythonhosted.org/packages/31/0d/b48c405c91693635fbe2dcd7bc84a33a602add5f63286e024d3b6741411c/numpy-2.2.6-cp311-cp311-macosx_14_0_x86_64.whl", hash = "sha256:481b49095335f8eed42e39e8041327c05b0f6f4780488f61286ed3c01368d491", size = 6889579, upload-time = "2025-05-17T21:32:01.712Z" }, + { url = "https://files.pythonhosted.org/packages/52/b8/7f0554d49b565d0171eab6e99001846882000883998e7b7d9f0d98b1f934/numpy-2.2.6-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b64d8d4d17135e00c8e346e0a738deb17e754230d7e0810ac5012750bbd85a5a", size = 14312005, upload-time = "2025-05-17T21:32:23.332Z" }, + { url = "https://files.pythonhosted.org/packages/b3/dd/2238b898e51bd6d389b7389ffb20d7f4c10066d80351187ec8e303a5a475/numpy-2.2.6-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ba10f8411898fc418a521833e014a77d3ca01c15b0c6cdcce6a0d2897e6dbbdf", size = 16821570, upload-time = "2025-05-17T21:32:47.991Z" }, + { url = "https://files.pythonhosted.org/packages/83/6c/44d0325722cf644f191042bf47eedad61c1e6df2432ed65cbe28509d404e/numpy-2.2.6-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:bd48227a919f1bafbdda0583705e547892342c26fb127219d60a5c36882609d1", size = 15818548, upload-time = "2025-05-17T21:33:11.728Z" }, + { url = "https://files.pythonhosted.org/packages/ae/9d/81e8216030ce66be25279098789b665d49ff19eef08bfa8cb96d4957f422/numpy-2.2.6-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:9551a499bf125c1d4f9e250377c1ee2eddd02e01eac6644c080162c0c51778ab", size = 18620521, upload-time = "2025-05-17T21:33:39.139Z" }, + { url = "https://files.pythonhosted.org/packages/6a/fd/e19617b9530b031db51b0926eed5345ce8ddc669bb3bc0044b23e275ebe8/numpy-2.2.6-cp311-cp311-win32.whl", hash = "sha256:0678000bb9ac1475cd454c6b8c799206af8107e310843532b04d49649c717a47", size = 6525866, upload-time = "2025-05-17T21:33:50.273Z" }, + { url = "https://files.pythonhosted.org/packages/31/0a/f354fb7176b81747d870f7991dc763e157a934c717b67b58456bc63da3df/numpy-2.2.6-cp311-cp311-win_amd64.whl", hash = "sha256:e8213002e427c69c45a52bbd94163084025f533a55a59d6f9c5b820774ef3303", size = 12907455, upload-time = "2025-05-17T21:34:09.135Z" }, + { url = "https://files.pythonhosted.org/packages/82/5d/c00588b6cf18e1da539b45d3598d3557084990dcc4331960c15ee776ee41/numpy-2.2.6-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:41c5a21f4a04fa86436124d388f6ed60a9343a6f767fced1a8a71c3fbca038ff", size = 20875348, upload-time = "2025-05-17T21:34:39.648Z" }, + { url = "https://files.pythonhosted.org/packages/66/ee/560deadcdde6c2f90200450d5938f63a34b37e27ebff162810f716f6a230/numpy-2.2.6-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:de749064336d37e340f640b05f24e9e3dd678c57318c7289d222a8a2f543e90c", size = 14119362, upload-time = "2025-05-17T21:35:01.241Z" }, + { url = "https://files.pythonhosted.org/packages/3c/65/4baa99f1c53b30adf0acd9a5519078871ddde8d2339dc5a7fde80d9d87da/numpy-2.2.6-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:894b3a42502226a1cac872f840030665f33326fc3dac8e57c607905773cdcde3", size = 5084103, upload-time = "2025-05-17T21:35:10.622Z" }, + { url = "https://files.pythonhosted.org/packages/cc/89/e5a34c071a0570cc40c9a54eb472d113eea6d002e9ae12bb3a8407fb912e/numpy-2.2.6-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:71594f7c51a18e728451bb50cc60a3ce4e6538822731b2933209a1f3614e9282", size = 6625382, upload-time = "2025-05-17T21:35:21.414Z" }, + { url = "https://files.pythonhosted.org/packages/f8/35/8c80729f1ff76b3921d5c9487c7ac3de9b2a103b1cd05e905b3090513510/numpy-2.2.6-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f2618db89be1b4e05f7a1a847a9c1c0abd63e63a1607d892dd54668dd92faf87", size = 14018462, upload-time = "2025-05-17T21:35:42.174Z" }, + { url = "https://files.pythonhosted.org/packages/8c/3d/1e1db36cfd41f895d266b103df00ca5b3cbe965184df824dec5c08c6b803/numpy-2.2.6-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fd83c01228a688733f1ded5201c678f0c53ecc1006ffbc404db9f7a899ac6249", size = 16527618, upload-time = "2025-05-17T21:36:06.711Z" }, + { url = "https://files.pythonhosted.org/packages/61/c6/03ed30992602c85aa3cd95b9070a514f8b3c33e31124694438d88809ae36/numpy-2.2.6-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:37c0ca431f82cd5fa716eca9506aefcabc247fb27ba69c5062a6d3ade8cf8f49", size = 15505511, upload-time = "2025-05-17T21:36:29.965Z" }, + { url = "https://files.pythonhosted.org/packages/b7/25/5761d832a81df431e260719ec45de696414266613c9ee268394dd5ad8236/numpy-2.2.6-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:fe27749d33bb772c80dcd84ae7e8df2adc920ae8297400dabec45f0dedb3f6de", size = 18313783, upload-time = "2025-05-17T21:36:56.883Z" }, + { url = "https://files.pythonhosted.org/packages/57/0a/72d5a3527c5ebffcd47bde9162c39fae1f90138c961e5296491ce778e682/numpy-2.2.6-cp312-cp312-win32.whl", hash = "sha256:4eeaae00d789f66c7a25ac5f34b71a7035bb474e679f410e5e1a94deb24cf2d4", size = 6246506, upload-time = "2025-05-17T21:37:07.368Z" }, + { url = "https://files.pythonhosted.org/packages/36/fa/8c9210162ca1b88529ab76b41ba02d433fd54fecaf6feb70ef9f124683f1/numpy-2.2.6-cp312-cp312-win_amd64.whl", hash = "sha256:c1f9540be57940698ed329904db803cf7a402f3fc200bfe599334c9bd84a40b2", size = 12614190, upload-time = "2025-05-17T21:37:26.213Z" }, + { url = "https://files.pythonhosted.org/packages/f9/5c/6657823f4f594f72b5471f1db1ab12e26e890bb2e41897522d134d2a3e81/numpy-2.2.6-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:0811bb762109d9708cca4d0b13c4f67146e3c3b7cf8d34018c722adb2d957c84", size = 20867828, upload-time = "2025-05-17T21:37:56.699Z" }, + { url = "https://files.pythonhosted.org/packages/dc/9e/14520dc3dadf3c803473bd07e9b2bd1b69bc583cb2497b47000fed2fa92f/numpy-2.2.6-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:287cc3162b6f01463ccd86be154f284d0893d2b3ed7292439ea97eafa8170e0b", size = 14143006, upload-time = "2025-05-17T21:38:18.291Z" }, + { url = "https://files.pythonhosted.org/packages/4f/06/7e96c57d90bebdce9918412087fc22ca9851cceaf5567a45c1f404480e9e/numpy-2.2.6-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:f1372f041402e37e5e633e586f62aa53de2eac8d98cbfb822806ce4bbefcb74d", size = 5076765, upload-time = "2025-05-17T21:38:27.319Z" }, + { url = "https://files.pythonhosted.org/packages/73/ed/63d920c23b4289fdac96ddbdd6132e9427790977d5457cd132f18e76eae0/numpy-2.2.6-cp313-cp313-macosx_14_0_x86_64.whl", hash = "sha256:55a4d33fa519660d69614a9fad433be87e5252f4b03850642f88993f7b2ca566", size = 6617736, upload-time = "2025-05-17T21:38:38.141Z" }, + { url = "https://files.pythonhosted.org/packages/85/c5/e19c8f99d83fd377ec8c7e0cf627a8049746da54afc24ef0a0cb73d5dfb5/numpy-2.2.6-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f92729c95468a2f4f15e9bb94c432a9229d0d50de67304399627a943201baa2f", size = 14010719, upload-time = "2025-05-17T21:38:58.433Z" }, + { url = "https://files.pythonhosted.org/packages/19/49/4df9123aafa7b539317bf6d342cb6d227e49f7a35b99c287a6109b13dd93/numpy-2.2.6-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1bc23a79bfabc5d056d106f9befb8d50c31ced2fbc70eedb8155aec74a45798f", size = 16526072, upload-time = "2025-05-17T21:39:22.638Z" }, + { url = "https://files.pythonhosted.org/packages/b2/6c/04b5f47f4f32f7c2b0e7260442a8cbcf8168b0e1a41ff1495da42f42a14f/numpy-2.2.6-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:e3143e4451880bed956e706a3220b4e5cf6172ef05fcc397f6f36a550b1dd868", size = 15503213, upload-time = "2025-05-17T21:39:45.865Z" }, + { url = "https://files.pythonhosted.org/packages/17/0a/5cd92e352c1307640d5b6fec1b2ffb06cd0dabe7d7b8227f97933d378422/numpy-2.2.6-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:b4f13750ce79751586ae2eb824ba7e1e8dba64784086c98cdbbcc6a42112ce0d", size = 18316632, upload-time = "2025-05-17T21:40:13.331Z" }, + { url = "https://files.pythonhosted.org/packages/f0/3b/5cba2b1d88760ef86596ad0f3d484b1cbff7c115ae2429678465057c5155/numpy-2.2.6-cp313-cp313-win32.whl", hash = "sha256:5beb72339d9d4fa36522fc63802f469b13cdbe4fdab4a288f0c441b74272ebfd", size = 6244532, upload-time = "2025-05-17T21:43:46.099Z" }, + { url = "https://files.pythonhosted.org/packages/cb/3b/d58c12eafcb298d4e6d0d40216866ab15f59e55d148a5658bb3132311fcf/numpy-2.2.6-cp313-cp313-win_amd64.whl", hash = "sha256:b0544343a702fa80c95ad5d3d608ea3599dd54d4632df855e4c8d24eb6ecfa1c", size = 12610885, upload-time = "2025-05-17T21:44:05.145Z" }, + { url = "https://files.pythonhosted.org/packages/6b/9e/4bf918b818e516322db999ac25d00c75788ddfd2d2ade4fa66f1f38097e1/numpy-2.2.6-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:0bca768cd85ae743b2affdc762d617eddf3bcf8724435498a1e80132d04879e6", size = 20963467, upload-time = "2025-05-17T21:40:44Z" }, + { url = "https://files.pythonhosted.org/packages/61/66/d2de6b291507517ff2e438e13ff7b1e2cdbdb7cb40b3ed475377aece69f9/numpy-2.2.6-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:fc0c5673685c508a142ca65209b4e79ed6740a4ed6b2267dbba90f34b0b3cfda", size = 14225144, upload-time = "2025-05-17T21:41:05.695Z" }, + { url = "https://files.pythonhosted.org/packages/e4/25/480387655407ead912e28ba3a820bc69af9adf13bcbe40b299d454ec011f/numpy-2.2.6-cp313-cp313t-macosx_14_0_arm64.whl", hash = "sha256:5bd4fc3ac8926b3819797a7c0e2631eb889b4118a9898c84f585a54d475b7e40", size = 5200217, upload-time = "2025-05-17T21:41:15.903Z" }, + { url = "https://files.pythonhosted.org/packages/aa/4a/6e313b5108f53dcbf3aca0c0f3e9c92f4c10ce57a0a721851f9785872895/numpy-2.2.6-cp313-cp313t-macosx_14_0_x86_64.whl", hash = "sha256:fee4236c876c4e8369388054d02d0e9bb84821feb1a64dd59e137e6511a551f8", size = 6712014, upload-time = "2025-05-17T21:41:27.321Z" }, + { url = "https://files.pythonhosted.org/packages/b7/30/172c2d5c4be71fdf476e9de553443cf8e25feddbe185e0bd88b096915bcc/numpy-2.2.6-cp313-cp313t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e1dda9c7e08dc141e0247a5b8f49cf05984955246a327d4c48bda16821947b2f", size = 14077935, upload-time = "2025-05-17T21:41:49.738Z" }, + { url = "https://files.pythonhosted.org/packages/12/fb/9e743f8d4e4d3c710902cf87af3512082ae3d43b945d5d16563f26ec251d/numpy-2.2.6-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f447e6acb680fd307f40d3da4852208af94afdfab89cf850986c3ca00562f4fa", size = 16600122, upload-time = "2025-05-17T21:42:14.046Z" }, + { url = "https://files.pythonhosted.org/packages/12/75/ee20da0e58d3a66f204f38916757e01e33a9737d0b22373b3eb5a27358f9/numpy-2.2.6-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:389d771b1623ec92636b0786bc4ae56abafad4a4c513d36a55dce14bd9ce8571", size = 15586143, upload-time = "2025-05-17T21:42:37.464Z" }, + { url = "https://files.pythonhosted.org/packages/76/95/bef5b37f29fc5e739947e9ce5179ad402875633308504a52d188302319c8/numpy-2.2.6-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:8e9ace4a37db23421249ed236fdcdd457d671e25146786dfc96835cd951aa7c1", size = 18385260, upload-time = "2025-05-17T21:43:05.189Z" }, + { url = "https://files.pythonhosted.org/packages/09/04/f2f83279d287407cf36a7a8053a5abe7be3622a4363337338f2585e4afda/numpy-2.2.6-cp313-cp313t-win32.whl", hash = "sha256:038613e9fb8c72b0a41f025a7e4c3f0b7a1b5d768ece4796b674c8f3fe13efff", size = 6377225, upload-time = "2025-05-17T21:43:16.254Z" }, + { url = "https://files.pythonhosted.org/packages/67/0e/35082d13c09c02c011cf21570543d202ad929d961c02a147493cb0c2bdf5/numpy-2.2.6-cp313-cp313t-win_amd64.whl", hash = "sha256:6031dd6dfecc0cf9f668681a37648373bddd6421fff6c66ec1624eed0180ee06", size = 12771374, upload-time = "2025-05-17T21:43:35.479Z" }, + { url = "https://files.pythonhosted.org/packages/9e/3b/d94a75f4dbf1ef5d321523ecac21ef23a3cd2ac8b78ae2aac40873590229/numpy-2.2.6-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:0b605b275d7bd0c640cad4e5d30fa701a8d59302e127e5f79138ad62762c3e3d", size = 21040391, upload-time = "2025-05-17T21:44:35.948Z" }, + { url = "https://files.pythonhosted.org/packages/17/f4/09b2fa1b58f0fb4f7c7963a1649c64c4d315752240377ed74d9cd878f7b5/numpy-2.2.6-pp310-pypy310_pp73-macosx_14_0_x86_64.whl", hash = "sha256:7befc596a7dc9da8a337f79802ee8adb30a552a94f792b9c9d18c840055907db", size = 6786754, upload-time = "2025-05-17T21:44:47.446Z" }, + { url = "https://files.pythonhosted.org/packages/af/30/feba75f143bdc868a1cc3f44ccfa6c4b9ec522b36458e738cd00f67b573f/numpy-2.2.6-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ce47521a4754c8f4593837384bd3424880629f718d87c5d44f8ed763edd63543", size = 16643476, upload-time = "2025-05-17T21:45:11.871Z" }, + { url = "https://files.pythonhosted.org/packages/37/48/ac2a9584402fb6c0cd5b5d1a91dcf176b15760130dd386bbafdbfe3640bf/numpy-2.2.6-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:d042d24c90c41b54fd506da306759e06e568864df8ec17ccc17e9e884634fd00", size = 12812666, upload-time = "2025-05-17T21:45:31.426Z" }, +] + +[[package]] +name = "nvidia-cublas-cu12" +version = "12.8.4.1" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/29/99/db44d685f0e257ff0e213ade1964fc459b4a690a73293220e98feb3307cf/nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:b86f6dd8935884615a0683b663891d43781b819ac4f2ba2b0c9604676af346d0", size = 590537124, upload-time = "2025-03-07T01:43:53.556Z" }, + { url = "https://files.pythonhosted.org/packages/dc/61/e24b560ab2e2eaeb3c839129175fb330dfcfc29e5203196e5541a4c44682/nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:8ac4e771d5a348c551b2a426eda6193c19aa630236b418086020df5ba9667142", size = 594346921, upload-time = "2025-03-07T01:44:31.254Z" }, +] + +[[package]] +name = "nvidia-cuda-cupti-cu12" +version = "12.8.90" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/d5/1f/b3bd73445e5cb342727fd24fe1f7b748f690b460acadc27ea22f904502c8/nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:4412396548808ddfed3f17a467b104ba7751e6b58678a4b840675c56d21cf7ed", size = 9533318, upload-time = "2025-03-07T01:40:10.421Z" }, + { url = "https://files.pythonhosted.org/packages/f8/02/2adcaa145158bf1a8295d83591d22e4103dbfd821bcaf6f3f53151ca4ffa/nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ea0cb07ebda26bb9b29ba82cda34849e73c166c18162d3913575b0c9db9a6182", size = 10248621, upload-time = "2025-03-07T01:40:21.213Z" }, +] + +[[package]] +name = "nvidia-cuda-nvrtc-cu12" +version = "12.8.93" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/05/6b/32f747947df2da6994e999492ab306a903659555dddc0fbdeb9d71f75e52/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:a7756528852ef889772a84c6cd89d41dfa74667e24cca16bb31f8f061e3e9994", size = 88040029, upload-time = "2025-03-07T01:42:13.562Z" }, + { url = "https://files.pythonhosted.org/packages/eb/d1/e50d0acaab360482034b84b6e27ee83c6738f7d32182b987f9c7a4e32962/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:fc1fec1e1637854b4c0a65fb9a8346b51dd9ee69e61ebaccc82058441f15bce8", size = 43106076, upload-time = "2025-03-07T01:41:59.817Z" }, +] + +[[package]] +name = "nvidia-cuda-runtime-cu12" +version = "12.8.90" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/7c/75/f865a3b236e4647605ea34cc450900854ba123834a5f1598e160b9530c3a/nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:52bf7bbee900262ffefe5e9d5a2a69a30d97e2bc5bb6cc866688caa976966e3d", size = 965265, upload-time = "2025-03-07T01:39:43.533Z" }, + { url = "https://files.pythonhosted.org/packages/0d/9b/a997b638fcd068ad6e4d53b8551a7d30fe8b404d6f1804abf1df69838932/nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adade8dcbd0edf427b7204d480d6066d33902cab2a4707dcfc48a2d0fd44ab90", size = 954765, upload-time = "2025-03-07T01:40:01.615Z" }, +] + +[[package]] +name = "nvidia-cudnn-cu12" +version = "9.10.2.21" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/fa/41/e79269ce215c857c935fd86bcfe91a451a584dfc27f1e068f568b9ad1ab7/nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:c9132cc3f8958447b4910a1720036d9eff5928cc3179b0a51fb6d167c6cc87d8", size = 705026878, upload-time = "2025-06-06T21:52:51.348Z" }, + { url = "https://files.pythonhosted.org/packages/ba/51/e123d997aa098c61d029f76663dedbfb9bc8dcf8c60cbd6adbe42f76d049/nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:949452be657fa16687d0930933f032835951ef0892b37d2d53824d1a84dc97a8", size = 706758467, upload-time = "2025-06-06T21:54:08.597Z" }, +] + +[[package]] +name = "nvidia-cufft-cu12" +version = "11.3.3.83" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/60/bc/7771846d3a0272026c416fbb7e5f4c1f146d6d80704534d0b187dd6f4800/nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:848ef7224d6305cdb2a4df928759dca7b1201874787083b6e7550dd6765ce69a", size = 193109211, upload-time = "2025-03-07T01:44:56.873Z" }, + { url = "https://files.pythonhosted.org/packages/1f/13/ee4e00f30e676b66ae65b4f08cb5bcbb8392c03f54f2d5413ea99a5d1c80/nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4d2dd21ec0b88cf61b62e6b43564355e5222e4a3fb394cac0db101f2dd0d4f74", size = 193118695, upload-time = "2025-03-07T01:45:27.821Z" }, +] + +[[package]] +name = "nvidia-cufile-cu12" +version = "1.13.1.3" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/bb/fe/1bcba1dfbfb8d01be8d93f07bfc502c93fa23afa6fd5ab3fc7c1df71038a/nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1d069003be650e131b21c932ec3d8969c1715379251f8d23a1860554b1cb24fc", size = 1197834, upload-time = "2025-03-07T01:45:50.723Z" }, + { url = "https://files.pythonhosted.org/packages/1e/f5/5607710447a6fe9fd9b3283956fceeee8a06cda1d2f56ce31371f595db2a/nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:4beb6d4cce47c1a0f1013d72e02b0994730359e17801d395bdcbf20cfb3bb00a", size = 1120705, upload-time = "2025-03-07T01:45:41.434Z" }, +] + +[[package]] +name = "nvidia-curand-cu12" +version = "10.3.9.90" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/45/5e/92aa15eca622a388b80fbf8375d4760738df6285b1e92c43d37390a33a9a/nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:dfab99248034673b779bc6decafdc3404a8a6f502462201f2f31f11354204acd", size = 63625754, upload-time = "2025-03-07T01:46:10.735Z" }, + { url = "https://files.pythonhosted.org/packages/fb/aa/6584b56dc84ebe9cf93226a5cde4d99080c8e90ab40f0c27bda7a0f29aa1/nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:b32331d4f4df5d6eefa0554c565b626c7216f87a06a4f56fab27c3b68a830ec9", size = 63619976, upload-time = "2025-03-07T01:46:23.323Z" }, +] + +[[package]] +name = "nvidia-cusolver-cu12" +version = "11.7.3.90" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cusparse-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/c8/32/f7cd6ce8a7690544d084ea21c26e910a97e077c9b7f07bf5de623ee19981/nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:db9ed69dbef9715071232caa9b69c52ac7de3a95773c2db65bdba85916e4e5c0", size = 267229841, upload-time = "2025-03-07T01:46:54.356Z" }, + { url = "https://files.pythonhosted.org/packages/85/48/9a13d2975803e8cf2777d5ed57b87a0b6ca2cc795f9a4f59796a910bfb80/nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:4376c11ad263152bd50ea295c05370360776f8c3427b30991df774f9fb26c450", size = 267506905, upload-time = "2025-03-07T01:47:16.273Z" }, +] + +[[package]] +name = "nvidia-cusparse-cu12" +version = "12.5.8.93" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/bc/f7/cd777c4109681367721b00a106f491e0d0d15cfa1fd59672ce580ce42a97/nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:9b6c161cb130be1a07a27ea6923df8141f3c295852f4b260c65f18f3e0a091dc", size = 288117129, upload-time = "2025-03-07T01:47:40.407Z" }, + { url = "https://files.pythonhosted.org/packages/c2/f5/e1854cb2f2bcd4280c44736c93550cc300ff4b8c95ebe370d0aa7d2b473d/nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1ec05d76bbbd8b61b06a80e1eaf8cf4959c3d4ce8e711b65ebd0443bb0ebb13b", size = 288216466, upload-time = "2025-03-07T01:48:13.779Z" }, +] + +[[package]] +name = "nvidia-cusparselt-cu12" +version = "0.7.1" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/73/b9/598f6ff36faaece4b3c50d26f50e38661499ff34346f00e057760b35cc9d/nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:8878dce784d0fac90131b6817b607e803c36e629ba34dc5b433471382196b6a5", size = 283835557, upload-time = "2025-02-26T00:16:54.265Z" }, + { url = "https://files.pythonhosted.org/packages/56/79/12978b96bd44274fe38b5dde5cfb660b1d114f70a65ef962bcbbed99b549/nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:f1bb701d6b930d5a7cea44c19ceb973311500847f81b634d802b7b539dc55623", size = 287193691, upload-time = "2025-02-26T00:15:44.104Z" }, +] + +[[package]] +name = "nvidia-ml-py" +version = "13.580.82" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/dd/6c/4a533f2c0185027c465adb6063086bc3728301e95f483665bfa9ebafb2d3/nvidia_ml_py-13.580.82.tar.gz", hash = "sha256:0c028805dc53a0e2a6985ea801888197765ac2ef8f1c9e29a7bf0d3616a5efc7", size = 47999, upload-time = "2025-09-11T16:44:56.267Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/7f/96/d6d25a4c307d6645f4a9b91d620c0151c544ad38b5e371313a87d2761004/nvidia_ml_py-13.580.82-py3-none-any.whl", hash = "sha256:4361db337b0c551e2d101936dae2e9a60f957af26818e8c0c3a1f32b8db8d0a7", size = 49008, upload-time = "2025-09-11T16:44:54.915Z" }, +] + +[[package]] +name = "nvidia-nccl-cu12" +version = "2.27.5" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/bb/1c/857979db0ef194ca5e21478a0612bcdbbe59458d7694361882279947b349/nvidia_nccl_cu12-2.27.5-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:31432ad4d1fb1004eb0c56203dc9bc2178a1ba69d1d9e02d64a6938ab5e40e7a", size = 322400625, upload-time = "2025-06-26T04:11:04.496Z" }, + { url = "https://files.pythonhosted.org/packages/6e/89/f7a07dc961b60645dbbf42e80f2bc85ade7feb9a491b11a1e973aa00071f/nvidia_nccl_cu12-2.27.5-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ad730cf15cb5d25fe849c6e6ca9eb5b76db16a80f13f425ac68d8e2e55624457", size = 322348229, upload-time = "2025-06-26T04:11:28.385Z" }, +] + +[[package]] +name = "nvidia-nvjitlink-cu12" +version = "12.8.93" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/f6/74/86a07f1d0f42998ca31312f998bd3b9a7eff7f52378f4f270c8679c77fb9/nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:81ff63371a7ebd6e6451970684f916be2eab07321b73c9d244dc2b4da7f73b88", size = 39254836, upload-time = "2025-03-07T01:49:55.661Z" }, + { url = "https://files.pythonhosted.org/packages/2a/a2/8cee5da30d13430e87bf99bb33455d2724d0a4a9cb5d7926d80ccb96d008/nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:adccd7161ace7261e01bb91e44e88da350895c270d23f744f0820c818b7229e7", size = 38386204, upload-time = "2025-03-07T01:49:43.612Z" }, +] + +[[package]] +name = "nvidia-nvshmem-cu12" +version = "3.4.5" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/1d/6a/03aa43cc9bd3ad91553a88b5f6fb25ed6a3752ae86ce2180221962bc2aa5/nvidia_nvshmem_cu12-3.4.5-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:0b48363fc6964dede448029434c6abed6c5e37f823cb43c3bcde7ecfc0457e15", size = 138936938, upload-time = "2025-09-06T00:32:05.589Z" }, + { url = "https://files.pythonhosted.org/packages/b5/09/6ea3ea725f82e1e76684f0708bbedd871fc96da89945adeba65c3835a64c/nvidia_nvshmem_cu12-3.4.5-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:042f2500f24c021db8a06c5eec2539027d57460e1c1a762055a6554f72c369bd", size = 139103095, upload-time = "2025-09-06T00:32:31.266Z" }, +] + +[[package]] +name = "nvidia-nvtx-cu12" +version = "12.8.90" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/10/c0/1b303feea90d296f6176f32a2a70b5ef230f9bdeb3a72bddb0dc922dc137/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d7ad891da111ebafbf7e015d34879f7112832fc239ff0d7d776b6cb685274615", size = 91161, upload-time = "2025-03-07T01:42:23.922Z" }, + { url = "https://files.pythonhosted.org/packages/a2/eb/86626c1bbc2edb86323022371c39aa48df6fd8b0a1647bc274577f72e90b/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5b17e2001cc0d751a5bc2c6ec6d26ad95913324a4adb86788c944f8ce9ba441f", size = 89954, upload-time = "2025-03-07T01:42:44.131Z" }, +] + +[[package]] +name = "opencv-python" +version = "4.10.0.84" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "numpy" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/4a/e7/b70a2d9ab205110d715906fc8ec83fbb00404aeb3a37a0654fdb68eb0c8c/opencv-python-4.10.0.84.tar.gz", hash = "sha256:72d234e4582e9658ffea8e9cae5b63d488ad06994ef12d81dc303b17472f3526", size = 95103981, upload-time = "2024-06-17T18:29:56.757Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/66/82/564168a349148298aca281e342551404ef5521f33fba17b388ead0a84dc5/opencv_python-4.10.0.84-cp37-abi3-macosx_11_0_arm64.whl", hash = "sha256:fc182f8f4cda51b45f01c64e4cbedfc2f00aff799debebc305d8d0210c43f251", size = 54835524, upload-time = "2024-06-18T04:57:32.973Z" }, + { url = "https://files.pythonhosted.org/packages/64/4a/016cda9ad7cf18c58ba074628a4eaae8aa55f3fd06a266398cef8831a5b9/opencv_python-4.10.0.84-cp37-abi3-macosx_12_0_x86_64.whl", hash = "sha256:71e575744f1d23f79741450254660442785f45a0797212852ee5199ef12eed98", size = 56475426, upload-time = "2024-06-17T19:34:10.927Z" }, + { url = "https://files.pythonhosted.org/packages/81/e4/7a987ebecfe5ceaf32db413b67ff18eb3092c598408862fff4d7cc3fd19b/opencv_python-4.10.0.84-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:09a332b50488e2dda866a6c5573ee192fe3583239fb26ff2f7f9ceb0bc119ea6", size = 41746971, upload-time = "2024-06-17T20:00:25.211Z" }, + { url = "https://files.pythonhosted.org/packages/3f/a4/d2537f47fd7fcfba966bd806e3ec18e7ee1681056d4b0a9c8d983983e4d5/opencv_python-4.10.0.84-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9ace140fc6d647fbe1c692bcb2abce768973491222c067c131d80957c595b71f", size = 62548253, upload-time = "2024-06-17T18:29:43.659Z" }, + { url = "https://files.pythonhosted.org/packages/1e/39/bbf57e7b9dab623e8773f6ff36385456b7ae7fa9357a5e53db732c347eac/opencv_python-4.10.0.84-cp37-abi3-win32.whl", hash = "sha256:2db02bb7e50b703f0a2d50c50ced72e95c574e1e5a0bb35a8a86d0b35c98c236", size = 28737688, upload-time = "2024-06-17T18:28:13.177Z" }, + { url = "https://files.pythonhosted.org/packages/ec/6c/fab8113424af5049f85717e8e527ca3773299a3c6b02506e66436e19874f/opencv_python-4.10.0.84-cp37-abi3-win_amd64.whl", hash = "sha256:32dbbd94c26f611dc5cc6979e6b7aa1f55a64d6b463cc1dcd3c95505a63e48fe", size = 38842521, upload-time = "2024-06-17T18:28:21.813Z" }, +] + +[[package]] +name = "opentelemetry-api" +version = "1.38.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "importlib-metadata" }, + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/08/d8/0f354c375628e048bd0570645b310797299754730079853095bf000fba69/opentelemetry_api-1.38.0.tar.gz", hash = "sha256:f4c193b5e8acb0912b06ac5b16321908dd0843d75049c091487322284a3eea12", size = 65242, upload-time = "2025-10-16T08:35:50.25Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/ae/a2/d86e01c28300bd41bab8f18afd613676e2bd63515417b77636fc1add426f/opentelemetry_api-1.38.0-py3-none-any.whl", hash = "sha256:2891b0197f47124454ab9f0cf58f3be33faca394457ac3e09daba13ff50aa582", size = 65947, upload-time = "2025-10-16T08:35:30.23Z" }, +] + +[[package]] +name = "opentelemetry-sdk" +version = "1.38.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "opentelemetry-api" }, + { name = "opentelemetry-semantic-conventions" }, + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/85/cb/f0eee1445161faf4c9af3ba7b848cc22a50a3d3e2515051ad8628c35ff80/opentelemetry_sdk-1.38.0.tar.gz", hash = "sha256:93df5d4d871ed09cb4272305be4d996236eedb232253e3ab864c8620f051cebe", size = 171942, upload-time = "2025-10-16T08:36:02.257Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/2f/2e/e93777a95d7d9c40d270a371392b6d6f1ff170c2a3cb32d6176741b5b723/opentelemetry_sdk-1.38.0-py3-none-any.whl", hash = "sha256:1c66af6564ecc1553d72d811a01df063ff097cdc82ce188da9951f93b8d10f6b", size = 132349, upload-time = "2025-10-16T08:35:46.995Z" }, +] + +[[package]] +name = "opentelemetry-semantic-conventions" +version = "0.59b0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "opentelemetry-api" }, + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/40/bc/8b9ad3802cd8ac6583a4eb7de7e5d7db004e89cb7efe7008f9c8a537ee75/opentelemetry_semantic_conventions-0.59b0.tar.gz", hash = "sha256:7a6db3f30d70202d5bf9fa4b69bc866ca6a30437287de6c510fb594878aed6b0", size = 129861, upload-time = "2025-10-16T08:36:03.346Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/24/7d/c88d7b15ba8fe5c6b8f93be50fc11795e9fc05386c44afaf6b76fe191f9b/opentelemetry_semantic_conventions-0.59b0-py3-none-any.whl", hash = "sha256:35d3b8833ef97d614136e253c1da9342b4c3c083bbaf29ce31d572a1c3825eed", size = 207954, upload-time = "2025-10-16T08:35:48.054Z" }, +] + +[[package]] +name = "orjson" +version = "3.11.4" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/c6/fe/ed708782d6709cc60eb4c2d8a361a440661f74134675c72990f2c48c785f/orjson-3.11.4.tar.gz", hash = "sha256:39485f4ab4c9b30a3943cfe99e1a213c4776fb69e8abd68f66b83d5a0b0fdc6d", size = 5945188, upload-time = "2025-10-24T15:50:38.027Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/e0/30/5aed63d5af1c8b02fbd2a8d83e2a6c8455e30504c50dbf08c8b51403d873/orjson-3.11.4-cp310-cp310-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:e3aa2118a3ece0d25489cbe48498de8a5d580e42e8d9979f65bf47900a15aba1", size = 243870, upload-time = "2025-10-24T15:48:28.908Z" }, + { url = "https://files.pythonhosted.org/packages/44/1f/da46563c08bef33c41fd63c660abcd2184b4d2b950c8686317d03b9f5f0c/orjson-3.11.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a69ab657a4e6733133a3dca82768f2f8b884043714e8d2b9ba9f52b6efef5c44", size = 130622, upload-time = "2025-10-24T15:48:31.361Z" }, + { url = "https://files.pythonhosted.org/packages/02/bd/b551a05d0090eab0bf8008a13a14edc0f3c3e0236aa6f5b697760dd2817b/orjson-3.11.4-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:3740bffd9816fc0326ddc406098a3a8f387e42223f5f455f2a02a9f834ead80c", size = 129344, upload-time = "2025-10-24T15:48:32.71Z" }, + { url = "https://files.pythonhosted.org/packages/87/6c/9ddd5e609f443b2548c5e7df3c44d0e86df2c68587a0e20c50018cdec535/orjson-3.11.4-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:65fd2f5730b1bf7f350c6dc896173d3460d235c4be007af73986d7cd9a2acd23", size = 136633, upload-time = "2025-10-24T15:48:34.128Z" }, + { url = "https://files.pythonhosted.org/packages/95/f2/9f04f2874c625a9fb60f6918c33542320661255323c272e66f7dcce14df2/orjson-3.11.4-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9fdc3ae730541086158d549c97852e2eea6820665d4faf0f41bf99df41bc11ea", size = 137695, upload-time = "2025-10-24T15:48:35.654Z" }, + { url = "https://files.pythonhosted.org/packages/d2/c2/c7302afcbdfe8a891baae0e2cee091583a30e6fa613e8bdf33b0e9c8a8c7/orjson-3.11.4-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e10b4d65901da88845516ce9f7f9736f9638d19a1d483b3883dc0182e6e5edba", size = 136879, upload-time = "2025-10-24T15:48:37.483Z" }, + { url = "https://files.pythonhosted.org/packages/c6/3a/b31c8f0182a3e27f48e703f46e61bb769666cd0dac4700a73912d07a1417/orjson-3.11.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fb6a03a678085f64b97f9d4a9ae69376ce91a3a9e9b56a82b1580d8e1d501aff", size = 136374, upload-time = "2025-10-24T15:48:38.624Z" }, + { url = "https://files.pythonhosted.org/packages/29/d0/fd9ab96841b090d281c46df566b7f97bc6c8cd9aff3f3ebe99755895c406/orjson-3.11.4-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:2c82e4f0b1c712477317434761fbc28b044c838b6b1240d895607441412371ac", size = 140519, upload-time = "2025-10-24T15:48:39.756Z" }, + { url = "https://files.pythonhosted.org/packages/d6/ce/36eb0f15978bb88e33a3480e1a3fb891caa0f189ba61ce7713e0ccdadabf/orjson-3.11.4-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:d58c166a18f44cc9e2bad03a327dc2d1a3d2e85b847133cfbafd6bfc6719bd79", size = 406522, upload-time = "2025-10-24T15:48:41.198Z" }, + { url = "https://files.pythonhosted.org/packages/85/11/e8af3161a288f5c6a00c188fc729c7ba193b0cbc07309a1a29c004347c30/orjson-3.11.4-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:94f206766bf1ea30e1382e4890f763bd1eefddc580e08fec1ccdc20ddd95c827", size = 149790, upload-time = "2025-10-24T15:48:42.664Z" }, + { url = "https://files.pythonhosted.org/packages/ea/96/209d52db0cf1e10ed48d8c194841e383e23c2ced5a2ee766649fe0e32d02/orjson-3.11.4-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:41bf25fb39a34cf8edb4398818523277ee7096689db352036a9e8437f2f3ee6b", size = 140040, upload-time = "2025-10-24T15:48:44.042Z" }, + { url = "https://files.pythonhosted.org/packages/ef/0e/526db1395ccb74c3d59ac1660b9a325017096dc5643086b38f27662b4add/orjson-3.11.4-cp310-cp310-win32.whl", hash = "sha256:fa9627eba4e82f99ca6d29bc967f09aba446ee2b5a1ea728949ede73d313f5d3", size = 135955, upload-time = "2025-10-24T15:48:45.495Z" }, + { url = "https://files.pythonhosted.org/packages/e6/69/18a778c9de3702b19880e73c9866b91cc85f904b885d816ba1ab318b223c/orjson-3.11.4-cp310-cp310-win_amd64.whl", hash = "sha256:23ef7abc7fca96632d8174ac115e668c1e931b8fe4dde586e92a500bf1914dcc", size = 131577, upload-time = "2025-10-24T15:48:46.609Z" }, + { url = "https://files.pythonhosted.org/packages/63/1d/1ea6005fffb56715fd48f632611e163d1604e8316a5bad2288bee9a1c9eb/orjson-3.11.4-cp311-cp311-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:5e59d23cd93ada23ec59a96f215139753fbfe3a4d989549bcb390f8c00370b39", size = 243498, upload-time = "2025-10-24T15:48:48.101Z" }, + { url = "https://files.pythonhosted.org/packages/37/d7/ffed10c7da677f2a9da307d491b9eb1d0125b0307019c4ad3d665fd31f4f/orjson-3.11.4-cp311-cp311-macosx_15_0_arm64.whl", hash = "sha256:5c3aedecfc1beb988c27c79d52ebefab93b6c3921dbec361167e6559aba2d36d", size = 128961, upload-time = "2025-10-24T15:48:49.571Z" }, + { url = "https://files.pythonhosted.org/packages/a2/96/3e4d10a18866d1368f73c8c44b7fe37cc8a15c32f2a7620be3877d4c55a3/orjson-3.11.4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:da9e5301f1c2caa2a9a4a303480d79c9ad73560b2e7761de742ab39fe59d9175", size = 130321, upload-time = "2025-10-24T15:48:50.713Z" }, + { url = "https://files.pythonhosted.org/packages/eb/1f/465f66e93f434f968dd74d5b623eb62c657bdba2332f5a8be9f118bb74c7/orjson-3.11.4-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8873812c164a90a79f65368f8f96817e59e35d0cc02786a5356f0e2abed78040", size = 129207, upload-time = "2025-10-24T15:48:52.193Z" }, + { url = "https://files.pythonhosted.org/packages/28/43/d1e94837543321c119dff277ae8e348562fe8c0fafbb648ef7cb0c67e521/orjson-3.11.4-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:5d7feb0741ebb15204e748f26c9638e6665a5fa93c37a2c73d64f1669b0ddc63", size = 136323, upload-time = "2025-10-24T15:48:54.806Z" }, + { url = "https://files.pythonhosted.org/packages/bf/04/93303776c8890e422a5847dd012b4853cdd88206b8bbd3edc292c90102d1/orjson-3.11.4-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:01ee5487fefee21e6910da4c2ee9eef005bee568a0879834df86f888d2ffbdd9", size = 137440, upload-time = "2025-10-24T15:48:56.326Z" }, + { url = "https://files.pythonhosted.org/packages/1e/ef/75519d039e5ae6b0f34d0336854d55544ba903e21bf56c83adc51cd8bf82/orjson-3.11.4-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3d40d46f348c0321df01507f92b95a377240c4ec31985225a6668f10e2676f9a", size = 136680, upload-time = "2025-10-24T15:48:57.476Z" }, + { url = "https://files.pythonhosted.org/packages/b5/18/bf8581eaae0b941b44efe14fee7b7862c3382fbc9a0842132cfc7cf5ecf4/orjson-3.11.4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:95713e5fc8af84d8edc75b785d2386f653b63d62b16d681687746734b4dfc0be", size = 136160, upload-time = "2025-10-24T15:48:59.631Z" }, + { url = "https://files.pythonhosted.org/packages/c4/35/a6d582766d351f87fc0a22ad740a641b0a8e6fc47515e8614d2e4790ae10/orjson-3.11.4-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:ad73ede24f9083614d6c4ca9a85fe70e33be7bf047ec586ee2363bc7418fe4d7", size = 140318, upload-time = "2025-10-24T15:49:00.834Z" }, + { url = "https://files.pythonhosted.org/packages/76/b3/5a4801803ab2e2e2d703bce1a56540d9f99a9143fbec7bf63d225044fef8/orjson-3.11.4-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:842289889de515421f3f224ef9c1f1efb199a32d76d8d2ca2706fa8afe749549", size = 406330, upload-time = "2025-10-24T15:49:02.327Z" }, + { url = "https://files.pythonhosted.org/packages/80/55/a8f682f64833e3a649f620eafefee175cbfeb9854fc5b710b90c3bca45df/orjson-3.11.4-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:3b2427ed5791619851c52a1261b45c233930977e7de8cf36de05636c708fa905", size = 149580, upload-time = "2025-10-24T15:49:03.517Z" }, + { url = "https://files.pythonhosted.org/packages/ad/e4/c132fa0c67afbb3eb88274fa98df9ac1f631a675e7877037c611805a4413/orjson-3.11.4-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3c36e524af1d29982e9b190573677ea02781456b2e537d5840e4538a5ec41907", size = 139846, upload-time = "2025-10-24T15:49:04.761Z" }, + { url = "https://files.pythonhosted.org/packages/54/06/dc3491489efd651fef99c5908e13951abd1aead1257c67f16135f95ce209/orjson-3.11.4-cp311-cp311-win32.whl", hash = "sha256:87255b88756eab4a68ec61837ca754e5d10fa8bc47dc57f75cedfeaec358d54c", size = 135781, upload-time = "2025-10-24T15:49:05.969Z" }, + { url = "https://files.pythonhosted.org/packages/79/b7/5e5e8d77bd4ea02a6ac54c42c818afb01dd31961be8a574eb79f1d2cfb1e/orjson-3.11.4-cp311-cp311-win_amd64.whl", hash = "sha256:e2d5d5d798aba9a0e1fede8d853fa899ce2cb930ec0857365f700dffc2c7af6a", size = 131391, upload-time = "2025-10-24T15:49:07.355Z" }, + { url = "https://files.pythonhosted.org/packages/0f/dc/9484127cc1aa213be398ed735f5f270eedcb0c0977303a6f6ddc46b60204/orjson-3.11.4-cp311-cp311-win_arm64.whl", hash = "sha256:6bb6bb41b14c95d4f2702bce9975fda4516f1db48e500102fc4d8119032ff045", size = 126252, upload-time = "2025-10-24T15:49:08.869Z" }, + { url = "https://files.pythonhosted.org/packages/63/51/6b556192a04595b93e277a9ff71cd0cc06c21a7df98bcce5963fa0f5e36f/orjson-3.11.4-cp312-cp312-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:d4371de39319d05d3f482f372720b841c841b52f5385bd99c61ed69d55d9ab50", size = 243571, upload-time = "2025-10-24T15:49:10.008Z" }, + { url = "https://files.pythonhosted.org/packages/1c/2c/2602392ddf2601d538ff11848b98621cd465d1a1ceb9db9e8043181f2f7b/orjson-3.11.4-cp312-cp312-macosx_15_0_arm64.whl", hash = "sha256:e41fd3b3cac850eaae78232f37325ed7d7436e11c471246b87b2cd294ec94853", size = 128891, upload-time = "2025-10-24T15:49:11.297Z" }, + { url = "https://files.pythonhosted.org/packages/4e/47/bf85dcf95f7a3a12bf223394a4f849430acd82633848d52def09fa3f46ad/orjson-3.11.4-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:600e0e9ca042878c7fdf189cf1b028fe2c1418cc9195f6cb9824eb6ed99cb938", size = 130137, upload-time = "2025-10-24T15:49:12.544Z" }, + { url = "https://files.pythonhosted.org/packages/b4/4d/a0cb31007f3ab6f1fd2a1b17057c7c349bc2baf8921a85c0180cc7be8011/orjson-3.11.4-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7bbf9b333f1568ef5da42bc96e18bf30fd7f8d54e9ae066d711056add508e415", size = 129152, upload-time = "2025-10-24T15:49:13.754Z" }, + { url = "https://files.pythonhosted.org/packages/f7/ef/2811def7ce3d8576b19e3929fff8f8f0d44bc5eb2e0fdecb2e6e6cc6c720/orjson-3.11.4-cp312-cp312-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4806363144bb6e7297b8e95870e78d30a649fdc4e23fc84daa80c8ebd366ce44", size = 136834, upload-time = "2025-10-24T15:49:15.307Z" }, + { url = "https://files.pythonhosted.org/packages/00/d4/9aee9e54f1809cec8ed5abd9bc31e8a9631d19460e3b8470145d25140106/orjson-3.11.4-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ad355e8308493f527d41154e9053b86a5be892b3b359a5c6d5d95cda23601cb2", size = 137519, upload-time = "2025-10-24T15:49:16.557Z" }, + { url = "https://files.pythonhosted.org/packages/db/ea/67bfdb5465d5679e8ae8d68c11753aaf4f47e3e7264bad66dc2f2249e643/orjson-3.11.4-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:c8a7517482667fb9f0ff1b2f16fe5829296ed7a655d04d68cd9711a4d8a4e708", size = 136749, upload-time = "2025-10-24T15:49:17.796Z" }, + { url = "https://files.pythonhosted.org/packages/01/7e/62517dddcfce6d53a39543cd74d0dccfcbdf53967017c58af68822100272/orjson-3.11.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:97eb5942c7395a171cbfecc4ef6701fc3c403e762194683772df4c54cfbb2210", size = 136325, upload-time = "2025-10-24T15:49:19.347Z" }, + { url = "https://files.pythonhosted.org/packages/18/ae/40516739f99ab4c7ec3aaa5cc242d341fcb03a45d89edeeaabc5f69cb2cf/orjson-3.11.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:149d95d5e018bdd822e3f38c103b1a7c91f88d38a88aada5c4e9b3a73a244241", size = 140204, upload-time = "2025-10-24T15:49:20.545Z" }, + { url = "https://files.pythonhosted.org/packages/82/18/ff5734365623a8916e3a4037fcef1cd1782bfc14cf0992afe7940c5320bf/orjson-3.11.4-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:624f3951181eb46fc47dea3d221554e98784c823e7069edb5dbd0dc826ac909b", size = 406242, upload-time = "2025-10-24T15:49:21.884Z" }, + { url = "https://files.pythonhosted.org/packages/e1/43/96436041f0a0c8c8deca6a05ebeaf529bf1de04839f93ac5e7c479807aec/orjson-3.11.4-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:03bfa548cf35e3f8b3a96c4e8e41f753c686ff3d8e182ce275b1751deddab58c", size = 150013, upload-time = "2025-10-24T15:49:23.185Z" }, + { url = "https://files.pythonhosted.org/packages/1b/48/78302d98423ed8780479a1e682b9aecb869e8404545d999d34fa486e573e/orjson-3.11.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:525021896afef44a68148f6ed8a8bf8375553d6066c7f48537657f64823565b9", size = 139951, upload-time = "2025-10-24T15:49:24.428Z" }, + { url = "https://files.pythonhosted.org/packages/4a/7b/ad613fdcdaa812f075ec0875143c3d37f8654457d2af17703905425981bf/orjson-3.11.4-cp312-cp312-win32.whl", hash = "sha256:b58430396687ce0f7d9eeb3dd47761ca7d8fda8e9eb92b3077a7a353a75efefa", size = 136049, upload-time = "2025-10-24T15:49:25.973Z" }, + { url = "https://files.pythonhosted.org/packages/b9/3c/9cf47c3ff5f39b8350fb21ba65d789b6a1129d4cbb3033ba36c8a9023520/orjson-3.11.4-cp312-cp312-win_amd64.whl", hash = "sha256:c6dbf422894e1e3c80a177133c0dda260f81428f9de16d61041949f6a2e5c140", size = 131461, upload-time = "2025-10-24T15:49:27.259Z" }, + { url = "https://files.pythonhosted.org/packages/c6/3b/e2425f61e5825dc5b08c2a5a2b3af387eaaca22a12b9c8c01504f8614c36/orjson-3.11.4-cp312-cp312-win_arm64.whl", hash = "sha256:d38d2bc06d6415852224fcc9c0bfa834c25431e466dc319f0edd56cca81aa96e", size = 126167, upload-time = "2025-10-24T15:49:28.511Z" }, + { url = "https://files.pythonhosted.org/packages/23/15/c52aa7112006b0f3d6180386c3a46ae057f932ab3425bc6f6ac50431cca1/orjson-3.11.4-cp313-cp313-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:2d6737d0e616a6e053c8b4acc9eccea6b6cce078533666f32d140e4f85002534", size = 243525, upload-time = "2025-10-24T15:49:29.737Z" }, + { url = "https://files.pythonhosted.org/packages/ec/38/05340734c33b933fd114f161f25a04e651b0c7c33ab95e9416ade5cb44b8/orjson-3.11.4-cp313-cp313-macosx_15_0_arm64.whl", hash = "sha256:afb14052690aa328cc118a8e09f07c651d301a72e44920b887c519b313d892ff", size = 128871, upload-time = "2025-10-24T15:49:31.109Z" }, + { url = "https://files.pythonhosted.org/packages/55/b9/ae8d34899ff0c012039b5a7cb96a389b2476e917733294e498586b45472d/orjson-3.11.4-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:38aa9e65c591febb1b0aed8da4d469eba239d434c218562df179885c94e1a3ad", size = 130055, upload-time = "2025-10-24T15:49:33.382Z" }, + { url = "https://files.pythonhosted.org/packages/33/aa/6346dd5073730451bee3681d901e3c337e7ec17342fb79659ec9794fc023/orjson-3.11.4-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:f2cf4dfaf9163b0728d061bebc1e08631875c51cd30bf47cb9e3293bfbd7dcd5", size = 129061, upload-time = "2025-10-24T15:49:34.935Z" }, + { url = "https://files.pythonhosted.org/packages/39/e4/8eea51598f66a6c853c380979912d17ec510e8e66b280d968602e680b942/orjson-3.11.4-cp313-cp313-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:89216ff3dfdde0e4070932e126320a1752c9d9a758d6a32ec54b3b9334991a6a", size = 136541, upload-time = "2025-10-24T15:49:36.923Z" }, + { url = "https://files.pythonhosted.org/packages/9a/47/cb8c654fa9adcc60e99580e17c32b9e633290e6239a99efa6b885aba9dbc/orjson-3.11.4-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9daa26ca8e97fae0ce8aa5d80606ef8f7914e9b129b6b5df9104266f764ce436", size = 137535, upload-time = "2025-10-24T15:49:38.307Z" }, + { url = "https://files.pythonhosted.org/packages/43/92/04b8cc5c2b729f3437ee013ce14a60ab3d3001465d95c184758f19362f23/orjson-3.11.4-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:5c8b2769dc31883c44a9cd126560327767f848eb95f99c36c9932f51090bfce9", size = 136703, upload-time = "2025-10-24T15:49:40.795Z" }, + { url = "https://files.pythonhosted.org/packages/aa/fd/d0733fcb9086b8be4ebcfcda2d0312865d17d0d9884378b7cffb29d0763f/orjson-3.11.4-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1469d254b9884f984026bd9b0fa5bbab477a4bfe558bba6848086f6d43eb5e73", size = 136293, upload-time = "2025-10-24T15:49:42.347Z" }, + { url = "https://files.pythonhosted.org/packages/c2/d7/3c5514e806837c210492d72ae30ccf050ce3f940f45bf085bab272699ef4/orjson-3.11.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:68e44722541983614e37117209a194e8c3ad07838ccb3127d96863c95ec7f1e0", size = 140131, upload-time = "2025-10-24T15:49:43.638Z" }, + { url = "https://files.pythonhosted.org/packages/9c/dd/ba9d32a53207babf65bd510ac4d0faaa818bd0df9a9c6f472fe7c254f2e3/orjson-3.11.4-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:8e7805fda9672c12be2f22ae124dcd7b03928d6c197544fe12174b86553f3196", size = 406164, upload-time = "2025-10-24T15:49:45.498Z" }, + { url = "https://files.pythonhosted.org/packages/8e/f9/f68ad68f4af7c7bde57cd514eaa2c785e500477a8bc8f834838eb696a685/orjson-3.11.4-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:04b69c14615fb4434ab867bf6f38b2d649f6f300af30a6705397e895f7aec67a", size = 149859, upload-time = "2025-10-24T15:49:46.981Z" }, + { url = "https://files.pythonhosted.org/packages/b6/d2/7f847761d0c26818395b3d6b21fb6bc2305d94612a35b0a30eae65a22728/orjson-3.11.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:639c3735b8ae7f970066930e58cf0ed39a852d417c24acd4a25fc0b3da3c39a6", size = 139926, upload-time = "2025-10-24T15:49:48.321Z" }, + { url = "https://files.pythonhosted.org/packages/9f/37/acd14b12dc62db9a0e1d12386271b8661faae270b22492580d5258808975/orjson-3.11.4-cp313-cp313-win32.whl", hash = "sha256:6c13879c0d2964335491463302a6ca5ad98105fc5db3565499dcb80b1b4bd839", size = 136007, upload-time = "2025-10-24T15:49:49.938Z" }, + { url = "https://files.pythonhosted.org/packages/c0/a9/967be009ddf0a1fffd7a67de9c36656b28c763659ef91352acc02cbe364c/orjson-3.11.4-cp313-cp313-win_amd64.whl", hash = "sha256:09bf242a4af98732db9f9a1ec57ca2604848e16f132e3f72edfd3c5c96de009a", size = 131314, upload-time = "2025-10-24T15:49:51.248Z" }, + { url = "https://files.pythonhosted.org/packages/cb/db/399abd6950fbd94ce125cb8cd1a968def95174792e127b0642781e040ed4/orjson-3.11.4-cp313-cp313-win_arm64.whl", hash = "sha256:a85f0adf63319d6c1ba06fb0dbf997fced64a01179cf17939a6caca662bf92de", size = 126152, upload-time = "2025-10-24T15:49:52.922Z" }, + { url = "https://files.pythonhosted.org/packages/25/e3/54ff63c093cc1697e758e4fceb53164dd2661a7d1bcd522260ba09f54533/orjson-3.11.4-cp314-cp314-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:42d43a1f552be1a112af0b21c10a5f553983c2a0938d2bbb8ecd8bc9fb572803", size = 243501, upload-time = "2025-10-24T15:49:54.288Z" }, + { url = "https://files.pythonhosted.org/packages/ac/7d/e2d1076ed2e8e0ae9badca65bf7ef22710f93887b29eaa37f09850604e09/orjson-3.11.4-cp314-cp314-macosx_15_0_arm64.whl", hash = "sha256:26a20f3fbc6c7ff2cb8e89c4c5897762c9d88cf37330c6a117312365d6781d54", size = 128862, upload-time = "2025-10-24T15:49:55.961Z" }, + { url = "https://files.pythonhosted.org/packages/9f/37/ca2eb40b90621faddfa9517dfe96e25f5ae4d8057a7c0cdd613c17e07b2c/orjson-3.11.4-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6e3f20be9048941c7ffa8fc523ccbd17f82e24df1549d1d1fe9317712d19938e", size = 130047, upload-time = "2025-10-24T15:49:57.406Z" }, + { url = "https://files.pythonhosted.org/packages/c7/62/1021ed35a1f2bad9040f05fa4cc4f9893410df0ba3eaa323ccf899b1c90a/orjson-3.11.4-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:aac364c758dc87a52e68e349924d7e4ded348dedff553889e4d9f22f74785316", size = 129073, upload-time = "2025-10-24T15:49:58.782Z" }, + { url = "https://files.pythonhosted.org/packages/e8/3f/f84d966ec2a6fd5f73b1a707e7cd876813422ae4bf9f0145c55c9c6a0f57/orjson-3.11.4-cp314-cp314-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d5c54a6d76e3d741dcc3f2707f8eeb9ba2a791d3adbf18f900219b62942803b1", size = 136597, upload-time = "2025-10-24T15:50:00.12Z" }, + { url = "https://files.pythonhosted.org/packages/32/78/4fa0aeca65ee82bbabb49e055bd03fa4edea33f7c080c5c7b9601661ef72/orjson-3.11.4-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f28485bdca8617b79d44627f5fb04336897041dfd9fa66d383a49d09d86798bc", size = 137515, upload-time = "2025-10-24T15:50:01.57Z" }, + { url = "https://files.pythonhosted.org/packages/c1/9d/0c102e26e7fde40c4c98470796d050a2ec1953897e2c8ab0cb95b0759fa2/orjson-3.11.4-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:bfc2a484cad3585e4ba61985a6062a4c2ed5c7925db6d39f1fa267c9d166487f", size = 136703, upload-time = "2025-10-24T15:50:02.944Z" }, + { url = "https://files.pythonhosted.org/packages/df/ac/2de7188705b4cdfaf0b6c97d2f7849c17d2003232f6e70df98602173f788/orjson-3.11.4-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e34dbd508cb91c54f9c9788923daca129fe5b55c5b4eebe713bf5ed3791280cf", size = 136311, upload-time = "2025-10-24T15:50:04.441Z" }, + { url = "https://files.pythonhosted.org/packages/e0/52/847fcd1a98407154e944feeb12e3b4d487a0e264c40191fb44d1269cbaa1/orjson-3.11.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:b13c478fa413d4b4ee606ec8e11c3b2e52683a640b006bb586b3041c2ca5f606", size = 140127, upload-time = "2025-10-24T15:50:07.398Z" }, + { url = "https://files.pythonhosted.org/packages/c1/ae/21d208f58bdb847dd4d0d9407e2929862561841baa22bdab7aea10ca088e/orjson-3.11.4-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:724ca721ecc8a831b319dcd72cfa370cc380db0bf94537f08f7edd0a7d4e1780", size = 406201, upload-time = "2025-10-24T15:50:08.796Z" }, + { url = "https://files.pythonhosted.org/packages/8d/55/0789d6de386c8366059db098a628e2ad8798069e94409b0d8935934cbcb9/orjson-3.11.4-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:977c393f2e44845ce1b540e19a786e9643221b3323dae190668a98672d43fb23", size = 149872, upload-time = "2025-10-24T15:50:10.234Z" }, + { url = "https://files.pythonhosted.org/packages/cc/1d/7ff81ea23310e086c17b41d78a72270d9de04481e6113dbe2ac19118f7fb/orjson-3.11.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:1e539e382cf46edec157ad66b0b0872a90d829a6b71f17cb633d6c160a223155", size = 139931, upload-time = "2025-10-24T15:50:11.623Z" }, + { url = "https://files.pythonhosted.org/packages/77/92/25b886252c50ed64be68c937b562b2f2333b45afe72d53d719e46a565a50/orjson-3.11.4-cp314-cp314-win32.whl", hash = "sha256:d63076d625babab9db5e7836118bdfa086e60f37d8a174194ae720161eb12394", size = 136065, upload-time = "2025-10-24T15:50:13.025Z" }, + { url = "https://files.pythonhosted.org/packages/63/b8/718eecf0bb7e9d64e4956afaafd23db9f04c776d445f59fe94f54bdae8f0/orjson-3.11.4-cp314-cp314-win_amd64.whl", hash = "sha256:0a54d6635fa3aaa438ae32e8570b9f0de36f3f6562c308d2a2a452e8b0592db1", size = 131310, upload-time = "2025-10-24T15:50:14.46Z" }, + { url = "https://files.pythonhosted.org/packages/1a/bf/def5e25d4d8bfce296a9a7c8248109bf58622c21618b590678f945a2c59c/orjson-3.11.4-cp314-cp314-win_arm64.whl", hash = "sha256:78b999999039db3cf58f6d230f524f04f75f129ba3d1ca2ed121f8657e575d3d", size = 126151, upload-time = "2025-10-24T15:50:15.878Z" }, +] + +[[package]] +name = "packaging" +version = "25.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/a1/d4/1fc4078c65507b51b96ca8f8c3ba19e6a61c8253c72794544580a7b6c24d/packaging-25.0.tar.gz", hash = "sha256:d443872c98d677bf60f6a1f2f8c1cb748e8fe762d2bf9d3148b5599295b0fc4f", size = 165727, upload-time = "2025-04-19T11:48:59.673Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/20/12/38679034af332785aac8774540895e234f4d07f7545804097de4b666afd8/packaging-25.0-py3-none-any.whl", hash = "sha256:29572ef2b1f17581046b3a2227d5c611fb25ec70ca1ba8554b24b0e69331a484", size = 66469, upload-time = "2025-04-19T11:48:57.875Z" }, +] + +[[package]] +name = "pandas" +version = "2.3.3" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "numpy" }, + { name = "python-dateutil" }, + { name = "pytz" }, + { name = "tzdata" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/33/01/d40b85317f86cf08d853a4f495195c73815fdf205eef3993821720274518/pandas-2.3.3.tar.gz", hash = "sha256:e05e1af93b977f7eafa636d043f9f94c7ee3ac81af99c13508215942e64c993b", size = 4495223, upload-time = "2025-09-29T23:34:51.853Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/3d/f7/f425a00df4fcc22b292c6895c6831c0c8ae1d9fac1e024d16f98a9ce8749/pandas-2.3.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:376c6446ae31770764215a6c937f72d917f214b43560603cd60da6408f183b6c", size = 11555763, upload-time = "2025-09-29T23:16:53.287Z" }, + { url = "https://files.pythonhosted.org/packages/13/4f/66d99628ff8ce7857aca52fed8f0066ce209f96be2fede6cef9f84e8d04f/pandas-2.3.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:e19d192383eab2f4ceb30b412b22ea30690c9e618f78870357ae1d682912015a", size = 10801217, upload-time = "2025-09-29T23:17:04.522Z" }, + { url = "https://files.pythonhosted.org/packages/1d/03/3fc4a529a7710f890a239cc496fc6d50ad4a0995657dccc1d64695adb9f4/pandas-2.3.3-cp310-cp310-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5caf26f64126b6c7aec964f74266f435afef1c1b13da3b0636c7518a1fa3e2b1", size = 12148791, upload-time = "2025-09-29T23:17:18.444Z" }, + { url = "https://files.pythonhosted.org/packages/40/a8/4dac1f8f8235e5d25b9955d02ff6f29396191d4e665d71122c3722ca83c5/pandas-2.3.3-cp310-cp310-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:dd7478f1463441ae4ca7308a70e90b33470fa593429f9d4c578dd00d1fa78838", size = 12769373, upload-time = "2025-09-29T23:17:35.846Z" }, + { url = "https://files.pythonhosted.org/packages/df/91/82cc5169b6b25440a7fc0ef3a694582418d875c8e3ebf796a6d6470aa578/pandas-2.3.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:4793891684806ae50d1288c9bae9330293ab4e083ccd1c5e383c34549c6e4250", size = 13200444, upload-time = "2025-09-29T23:17:49.341Z" }, + { url = "https://files.pythonhosted.org/packages/10/ae/89b3283800ab58f7af2952704078555fa60c807fff764395bb57ea0b0dbd/pandas-2.3.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:28083c648d9a99a5dd035ec125d42439c6c1c525098c58af0fc38dd1a7a1b3d4", size = 13858459, upload-time = "2025-09-29T23:18:03.722Z" }, + { url = "https://files.pythonhosted.org/packages/85/72/530900610650f54a35a19476eca5104f38555afccda1aa11a92ee14cb21d/pandas-2.3.3-cp310-cp310-win_amd64.whl", hash = "sha256:503cf027cf9940d2ceaa1a93cfb5f8c8c7e6e90720a2850378f0b3f3b1e06826", size = 11346086, upload-time = "2025-09-29T23:18:18.505Z" }, + { url = "https://files.pythonhosted.org/packages/c1/fa/7ac648108144a095b4fb6aa3de1954689f7af60a14cf25583f4960ecb878/pandas-2.3.3-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:602b8615ebcc4a0c1751e71840428ddebeb142ec02c786e8ad6b1ce3c8dec523", size = 11578790, upload-time = "2025-09-29T23:18:30.065Z" }, + { url = "https://files.pythonhosted.org/packages/9b/35/74442388c6cf008882d4d4bdfc4109be87e9b8b7ccd097ad1e7f006e2e95/pandas-2.3.3-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:8fe25fc7b623b0ef6b5009149627e34d2a4657e880948ec3c840e9402e5c1b45", size = 10833831, upload-time = "2025-09-29T23:38:56.071Z" }, + { url = "https://files.pythonhosted.org/packages/fe/e4/de154cbfeee13383ad58d23017da99390b91d73f8c11856f2095e813201b/pandas-2.3.3-cp311-cp311-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b468d3dad6ff947df92dcb32ede5b7bd41a9b3cceef0a30ed925f6d01fb8fa66", size = 12199267, upload-time = "2025-09-29T23:18:41.627Z" }, + { url = "https://files.pythonhosted.org/packages/bf/c9/63f8d545568d9ab91476b1818b4741f521646cbdd151c6efebf40d6de6f7/pandas-2.3.3-cp311-cp311-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b98560e98cb334799c0b07ca7967ac361a47326e9b4e5a7dfb5ab2b1c9d35a1b", size = 12789281, upload-time = "2025-09-29T23:18:56.834Z" }, + { url = "https://files.pythonhosted.org/packages/f2/00/a5ac8c7a0e67fd1a6059e40aa08fa1c52cc00709077d2300e210c3ce0322/pandas-2.3.3-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:1d37b5848ba49824e5c30bedb9c830ab9b7751fd049bc7914533e01c65f79791", size = 13240453, upload-time = "2025-09-29T23:19:09.247Z" }, + { url = "https://files.pythonhosted.org/packages/27/4d/5c23a5bc7bd209231618dd9e606ce076272c9bc4f12023a70e03a86b4067/pandas-2.3.3-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:db4301b2d1f926ae677a751eb2bd0e8c5f5319c9cb3f88b0becbbb0b07b34151", size = 13890361, upload-time = "2025-09-29T23:19:25.342Z" }, + { url = "https://files.pythonhosted.org/packages/8e/59/712db1d7040520de7a4965df15b774348980e6df45c129b8c64d0dbe74ef/pandas-2.3.3-cp311-cp311-win_amd64.whl", hash = "sha256:f086f6fe114e19d92014a1966f43a3e62285109afe874f067f5abbdcbb10e59c", size = 11348702, upload-time = "2025-09-29T23:19:38.296Z" }, + { url = "https://files.pythonhosted.org/packages/9c/fb/231d89e8637c808b997d172b18e9d4a4bc7bf31296196c260526055d1ea0/pandas-2.3.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:6d21f6d74eb1725c2efaa71a2bfc661a0689579b58e9c0ca58a739ff0b002b53", size = 11597846, upload-time = "2025-09-29T23:19:48.856Z" }, + { url = "https://files.pythonhosted.org/packages/5c/bd/bf8064d9cfa214294356c2d6702b716d3cf3bb24be59287a6a21e24cae6b/pandas-2.3.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:3fd2f887589c7aa868e02632612ba39acb0b8948faf5cc58f0850e165bd46f35", size = 10729618, upload-time = "2025-09-29T23:39:08.659Z" }, + { url = "https://files.pythonhosted.org/packages/57/56/cf2dbe1a3f5271370669475ead12ce77c61726ffd19a35546e31aa8edf4e/pandas-2.3.3-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ecaf1e12bdc03c86ad4a7ea848d66c685cb6851d807a26aa245ca3d2017a1908", size = 11737212, upload-time = "2025-09-29T23:19:59.765Z" }, + { url = "https://files.pythonhosted.org/packages/e5/63/cd7d615331b328e287d8233ba9fdf191a9c2d11b6af0c7a59cfcec23de68/pandas-2.3.3-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b3d11d2fda7eb164ef27ffc14b4fcab16a80e1ce67e9f57e19ec0afaf715ba89", size = 12362693, upload-time = "2025-09-29T23:20:14.098Z" }, + { url = "https://files.pythonhosted.org/packages/a6/de/8b1895b107277d52f2b42d3a6806e69cfef0d5cf1d0ba343470b9d8e0a04/pandas-2.3.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:a68e15f780eddf2b07d242e17a04aa187a7ee12b40b930bfdd78070556550e98", size = 12771002, upload-time = "2025-09-29T23:20:26.76Z" }, + { url = "https://files.pythonhosted.org/packages/87/21/84072af3187a677c5893b170ba2c8fbe450a6ff911234916da889b698220/pandas-2.3.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:371a4ab48e950033bcf52b6527eccb564f52dc826c02afd9a1bc0ab731bba084", size = 13450971, upload-time = "2025-09-29T23:20:41.344Z" }, + { url = "https://files.pythonhosted.org/packages/86/41/585a168330ff063014880a80d744219dbf1dd7a1c706e75ab3425a987384/pandas-2.3.3-cp312-cp312-win_amd64.whl", hash = "sha256:a16dcec078a01eeef8ee61bf64074b4e524a2a3f4b3be9326420cabe59c4778b", size = 10992722, upload-time = "2025-09-29T23:20:54.139Z" }, + { url = "https://files.pythonhosted.org/packages/cd/4b/18b035ee18f97c1040d94debd8f2e737000ad70ccc8f5513f4eefad75f4b/pandas-2.3.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:56851a737e3470de7fa88e6131f41281ed440d29a9268dcbf0002da5ac366713", size = 11544671, upload-time = "2025-09-29T23:21:05.024Z" }, + { url = "https://files.pythonhosted.org/packages/31/94/72fac03573102779920099bcac1c3b05975c2cb5f01eac609faf34bed1ca/pandas-2.3.3-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:bdcd9d1167f4885211e401b3036c0c8d9e274eee67ea8d0758a256d60704cfe8", size = 10680807, upload-time = "2025-09-29T23:21:15.979Z" }, + { url = "https://files.pythonhosted.org/packages/16/87/9472cf4a487d848476865321de18cc8c920b8cab98453ab79dbbc98db63a/pandas-2.3.3-cp313-cp313-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e32e7cc9af0f1cc15548288a51a3b681cc2a219faa838e995f7dc53dbab1062d", size = 11709872, upload-time = "2025-09-29T23:21:27.165Z" }, + { url = "https://files.pythonhosted.org/packages/15/07/284f757f63f8a8d69ed4472bfd85122bd086e637bf4ed09de572d575a693/pandas-2.3.3-cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:318d77e0e42a628c04dc56bcef4b40de67918f7041c2b061af1da41dcff670ac", size = 12306371, upload-time = "2025-09-29T23:21:40.532Z" }, + { url = "https://files.pythonhosted.org/packages/33/81/a3afc88fca4aa925804a27d2676d22dcd2031c2ebe08aabd0ae55b9ff282/pandas-2.3.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:4e0a175408804d566144e170d0476b15d78458795bb18f1304fb94160cabf40c", size = 12765333, upload-time = "2025-09-29T23:21:55.77Z" }, + { url = "https://files.pythonhosted.org/packages/8d/0f/b4d4ae743a83742f1153464cf1a8ecfafc3ac59722a0b5c8602310cb7158/pandas-2.3.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:93c2d9ab0fc11822b5eece72ec9587e172f63cff87c00b062f6e37448ced4493", size = 13418120, upload-time = "2025-09-29T23:22:10.109Z" }, + { url = "https://files.pythonhosted.org/packages/4f/c7/e54682c96a895d0c808453269e0b5928a07a127a15704fedb643e9b0a4c8/pandas-2.3.3-cp313-cp313-win_amd64.whl", hash = "sha256:f8bfc0e12dc78f777f323f55c58649591b2cd0c43534e8355c51d3fede5f4dee", size = 10993991, upload-time = "2025-09-29T23:25:04.889Z" }, + { url = "https://files.pythonhosted.org/packages/f9/ca/3f8d4f49740799189e1395812f3bf23b5e8fc7c190827d55a610da72ce55/pandas-2.3.3-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:75ea25f9529fdec2d2e93a42c523962261e567d250b0013b16210e1d40d7c2e5", size = 12048227, upload-time = "2025-09-29T23:22:24.343Z" }, + { url = "https://files.pythonhosted.org/packages/0e/5a/f43efec3e8c0cc92c4663ccad372dbdff72b60bdb56b2749f04aa1d07d7e/pandas-2.3.3-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:74ecdf1d301e812db96a465a525952f4dde225fdb6d8e5a521d47e1f42041e21", size = 11411056, upload-time = "2025-09-29T23:22:37.762Z" }, + { url = "https://files.pythonhosted.org/packages/46/b1/85331edfc591208c9d1a63a06baa67b21d332e63b7a591a5ba42a10bb507/pandas-2.3.3-cp313-cp313t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6435cb949cb34ec11cc9860246ccb2fdc9ecd742c12d3304989017d53f039a78", size = 11645189, upload-time = "2025-09-29T23:22:51.688Z" }, + { url = "https://files.pythonhosted.org/packages/44/23/78d645adc35d94d1ac4f2a3c4112ab6f5b8999f4898b8cdf01252f8df4a9/pandas-2.3.3-cp313-cp313t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:900f47d8f20860de523a1ac881c4c36d65efcb2eb850e6948140fa781736e110", size = 12121912, upload-time = "2025-09-29T23:23:05.042Z" }, + { url = "https://files.pythonhosted.org/packages/53/da/d10013df5e6aaef6b425aa0c32e1fc1f3e431e4bcabd420517dceadce354/pandas-2.3.3-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:a45c765238e2ed7d7c608fc5bc4a6f88b642f2f01e70c0c23d2224dd21829d86", size = 12712160, upload-time = "2025-09-29T23:23:28.57Z" }, + { url = "https://files.pythonhosted.org/packages/bd/17/e756653095a083d8a37cbd816cb87148debcfcd920129b25f99dd8d04271/pandas-2.3.3-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:c4fc4c21971a1a9f4bdb4c73978c7f7256caa3e62b323f70d6cb80db583350bc", size = 13199233, upload-time = "2025-09-29T23:24:24.876Z" }, + { url = "https://files.pythonhosted.org/packages/04/fd/74903979833db8390b73b3a8a7d30d146d710bd32703724dd9083950386f/pandas-2.3.3-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:ee15f284898e7b246df8087fc82b87b01686f98ee67d85a17b7ab44143a3a9a0", size = 11540635, upload-time = "2025-09-29T23:25:52.486Z" }, + { url = "https://files.pythonhosted.org/packages/21/00/266d6b357ad5e6d3ad55093a7e8efc7dd245f5a842b584db9f30b0f0a287/pandas-2.3.3-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1611aedd912e1ff81ff41c745822980c49ce4a7907537be8692c8dbc31924593", size = 10759079, upload-time = "2025-09-29T23:26:33.204Z" }, + { url = "https://files.pythonhosted.org/packages/ca/05/d01ef80a7a3a12b2f8bbf16daba1e17c98a2f039cbc8e2f77a2c5a63d382/pandas-2.3.3-cp314-cp314-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6d2cefc361461662ac48810cb14365a365ce864afe85ef1f447ff5a1e99ea81c", size = 11814049, upload-time = "2025-09-29T23:27:15.384Z" }, + { url = "https://files.pythonhosted.org/packages/15/b2/0e62f78c0c5ba7e3d2c5945a82456f4fac76c480940f805e0b97fcbc2f65/pandas-2.3.3-cp314-cp314-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ee67acbbf05014ea6c763beb097e03cd629961c8a632075eeb34247120abcb4b", size = 12332638, upload-time = "2025-09-29T23:27:51.625Z" }, + { url = "https://files.pythonhosted.org/packages/c5/33/dd70400631b62b9b29c3c93d2feee1d0964dc2bae2e5ad7a6c73a7f25325/pandas-2.3.3-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:c46467899aaa4da076d5abc11084634e2d197e9460643dd455ac3db5856b24d6", size = 12886834, upload-time = "2025-09-29T23:28:21.289Z" }, + { url = "https://files.pythonhosted.org/packages/d3/18/b5d48f55821228d0d2692b34fd5034bb185e854bdb592e9c640f6290e012/pandas-2.3.3-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:6253c72c6a1d990a410bc7de641d34053364ef8bcd3126f7e7450125887dffe3", size = 13409925, upload-time = "2025-09-29T23:28:58.261Z" }, + { url = "https://files.pythonhosted.org/packages/a6/3d/124ac75fcd0ecc09b8fdccb0246ef65e35b012030defb0e0eba2cbbbe948/pandas-2.3.3-cp314-cp314-win_amd64.whl", hash = "sha256:1b07204a219b3b7350abaae088f451860223a52cfb8a6c53358e7948735158e5", size = 11109071, upload-time = "2025-09-29T23:32:27.484Z" }, + { url = "https://files.pythonhosted.org/packages/89/9c/0e21c895c38a157e0faa1fb64587a9226d6dd46452cac4532d80c3c4a244/pandas-2.3.3-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:2462b1a365b6109d275250baaae7b760fd25c726aaca0054649286bcfbb3e8ec", size = 12048504, upload-time = "2025-09-29T23:29:31.47Z" }, + { url = "https://files.pythonhosted.org/packages/d7/82/b69a1c95df796858777b68fbe6a81d37443a33319761d7c652ce77797475/pandas-2.3.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:0242fe9a49aa8b4d78a4fa03acb397a58833ef6199e9aa40a95f027bb3a1b6e7", size = 11410702, upload-time = "2025-09-29T23:29:54.591Z" }, + { url = "https://files.pythonhosted.org/packages/f9/88/702bde3ba0a94b8c73a0181e05144b10f13f29ebfc2150c3a79062a8195d/pandas-2.3.3-cp314-cp314t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a21d830e78df0a515db2b3d2f5570610f5e6bd2e27749770e8bb7b524b89b450", size = 11634535, upload-time = "2025-09-29T23:30:21.003Z" }, + { url = "https://files.pythonhosted.org/packages/a4/1e/1bac1a839d12e6a82ec6cb40cda2edde64a2013a66963293696bbf31fbbb/pandas-2.3.3-cp314-cp314t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2e3ebdb170b5ef78f19bfb71b0dc5dc58775032361fa188e814959b74d726dd5", size = 12121582, upload-time = "2025-09-29T23:30:43.391Z" }, + { url = "https://files.pythonhosted.org/packages/44/91/483de934193e12a3b1d6ae7c8645d083ff88dec75f46e827562f1e4b4da6/pandas-2.3.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:d051c0e065b94b7a3cea50eb1ec32e912cd96dba41647eb24104b6c6c14c5788", size = 12699963, upload-time = "2025-09-29T23:31:10.009Z" }, + { url = "https://files.pythonhosted.org/packages/70/44/5191d2e4026f86a2a109053e194d3ba7a31a2d10a9c2348368c63ed4e85a/pandas-2.3.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:3869faf4bd07b3b66a9f462417d0ca3a9df29a9f6abd5d0d0dbab15dac7abe87", size = 13202175, upload-time = "2025-09-29T23:31:59.173Z" }, +] + +[[package]] +name = "peft" +version = "0.18.0rc0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "accelerate" }, + { name = "huggingface-hub" }, + { name = "numpy" }, + { name = "packaging" }, + { name = "psutil" }, + { name = "pyyaml" }, + { name = "safetensors" }, + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, + { name = "tqdm" }, + { name = "transformers" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/a4/cf/7aaf059e3f22ff8fd1aebd4fb955b0a32daa7b4a9953e7918f95781ded9c/peft-0.18.0rc0.tar.gz", hash = "sha256:a7a91da57beaf10ca1e96cc32d4513a66176ce1cdc2a4e13899d8b9a6f0bb2d7", size = 629152, upload-time = "2025-10-28T08:43:59.254Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/76/b2/1a3a12b945d0f6f94bf12ef065c26daaf994c36455b5bb699179d668da6f/peft-0.18.0rc0-py3-none-any.whl", hash = "sha256:8632323e36f566fc1987b8f306c68238b11f9b173e894d88a6911e35cbe0f758", size = 554871, upload-time = "2025-10-28T08:43:57.415Z" }, +] + +[[package]] +name = "pillow" +version = "11.3.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/f3/0d/d0d6dea55cd152ce3d6767bb38a8fc10e33796ba4ba210cbab9354b6d238/pillow-11.3.0.tar.gz", hash = "sha256:3828ee7586cd0b2091b6209e5ad53e20d0649bbe87164a459d0676e035e8f523", size = 47113069, upload-time = "2025-07-01T09:16:30.666Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/4c/5d/45a3553a253ac8763f3561371432a90bdbe6000fbdcf1397ffe502aa206c/pillow-11.3.0-cp310-cp310-macosx_10_10_x86_64.whl", hash = "sha256:1b9c17fd4ace828b3003dfd1e30bff24863e0eb59b535e8f80194d9cc7ecf860", size = 5316554, upload-time = "2025-07-01T09:13:39.342Z" }, + { url = "https://files.pythonhosted.org/packages/7c/c8/67c12ab069ef586a25a4a79ced553586748fad100c77c0ce59bb4983ac98/pillow-11.3.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:65dc69160114cdd0ca0f35cb434633c75e8e7fad4cf855177a05bf38678f73ad", size = 4686548, upload-time = "2025-07-01T09:13:41.835Z" }, + { url = "https://files.pythonhosted.org/packages/2f/bd/6741ebd56263390b382ae4c5de02979af7f8bd9807346d068700dd6d5cf9/pillow-11.3.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7107195ddc914f656c7fc8e4a5e1c25f32e9236ea3ea860f257b0436011fddd0", size = 5859742, upload-time = "2025-07-03T13:09:47.439Z" }, + { url = "https://files.pythonhosted.org/packages/ca/0b/c412a9e27e1e6a829e6ab6c2dca52dd563efbedf4c9c6aa453d9a9b77359/pillow-11.3.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:cc3e831b563b3114baac7ec2ee86819eb03caa1a2cef0b481a5675b59c4fe23b", size = 7633087, upload-time = "2025-07-03T13:09:51.796Z" }, + { url = "https://files.pythonhosted.org/packages/59/9d/9b7076aaf30f5dd17e5e5589b2d2f5a5d7e30ff67a171eb686e4eecc2adf/pillow-11.3.0-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f1f182ebd2303acf8c380a54f615ec883322593320a9b00438eb842c1f37ae50", size = 5963350, upload-time = "2025-07-01T09:13:43.865Z" }, + { url = "https://files.pythonhosted.org/packages/f0/16/1a6bf01fb622fb9cf5c91683823f073f053005c849b1f52ed613afcf8dae/pillow-11.3.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4445fa62e15936a028672fd48c4c11a66d641d2c05726c7ec1f8ba6a572036ae", size = 6631840, upload-time = "2025-07-01T09:13:46.161Z" }, + { url = "https://files.pythonhosted.org/packages/7b/e6/6ff7077077eb47fde78739e7d570bdcd7c10495666b6afcd23ab56b19a43/pillow-11.3.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:71f511f6b3b91dd543282477be45a033e4845a40278fa8dcdbfdb07109bf18f9", size = 6074005, upload-time = "2025-07-01T09:13:47.829Z" }, + { url = "https://files.pythonhosted.org/packages/c3/3a/b13f36832ea6d279a697231658199e0a03cd87ef12048016bdcc84131601/pillow-11.3.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:040a5b691b0713e1f6cbe222e0f4f74cd233421e105850ae3b3c0ceda520f42e", size = 6708372, upload-time = "2025-07-01T09:13:52.145Z" }, + { url = "https://files.pythonhosted.org/packages/6c/e4/61b2e1a7528740efbc70b3d581f33937e38e98ef3d50b05007267a55bcb2/pillow-11.3.0-cp310-cp310-win32.whl", hash = "sha256:89bd777bc6624fe4115e9fac3352c79ed60f3bb18651420635f26e643e3dd1f6", size = 6277090, upload-time = "2025-07-01T09:13:53.915Z" }, + { url = "https://files.pythonhosted.org/packages/a9/d3/60c781c83a785d6afbd6a326ed4d759d141de43aa7365725cbcd65ce5e54/pillow-11.3.0-cp310-cp310-win_amd64.whl", hash = "sha256:19d2ff547c75b8e3ff46f4d9ef969a06c30ab2d4263a9e287733aa8b2429ce8f", size = 6985988, upload-time = "2025-07-01T09:13:55.699Z" }, + { url = "https://files.pythonhosted.org/packages/9f/28/4f4a0203165eefb3763939c6789ba31013a2e90adffb456610f30f613850/pillow-11.3.0-cp310-cp310-win_arm64.whl", hash = "sha256:819931d25e57b513242859ce1876c58c59dc31587847bf74cfe06b2e0cb22d2f", size = 2422899, upload-time = "2025-07-01T09:13:57.497Z" }, + { url = "https://files.pythonhosted.org/packages/db/26/77f8ed17ca4ffd60e1dcd220a6ec6d71210ba398cfa33a13a1cd614c5613/pillow-11.3.0-cp311-cp311-macosx_10_10_x86_64.whl", hash = "sha256:1cd110edf822773368b396281a2293aeb91c90a2db00d78ea43e7e861631b722", size = 5316531, upload-time = "2025-07-01T09:13:59.203Z" }, + { url = "https://files.pythonhosted.org/packages/cb/39/ee475903197ce709322a17a866892efb560f57900d9af2e55f86db51b0a5/pillow-11.3.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:9c412fddd1b77a75aa904615ebaa6001f169b26fd467b4be93aded278266b288", size = 4686560, upload-time = "2025-07-01T09:14:01.101Z" }, + { url = "https://files.pythonhosted.org/packages/d5/90/442068a160fd179938ba55ec8c97050a612426fae5ec0a764e345839f76d/pillow-11.3.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7d1aa4de119a0ecac0a34a9c8bde33f34022e2e8f99104e47a3ca392fd60e37d", size = 5870978, upload-time = "2025-07-03T13:09:55.638Z" }, + { url = "https://files.pythonhosted.org/packages/13/92/dcdd147ab02daf405387f0218dcf792dc6dd5b14d2573d40b4caeef01059/pillow-11.3.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:91da1d88226663594e3f6b4b8c3c8d85bd504117d043740a8e0ec449087cc494", size = 7641168, upload-time = "2025-07-03T13:10:00.37Z" }, + { url = "https://files.pythonhosted.org/packages/6e/db/839d6ba7fd38b51af641aa904e2960e7a5644d60ec754c046b7d2aee00e5/pillow-11.3.0-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:643f189248837533073c405ec2f0bb250ba54598cf80e8c1e043381a60632f58", size = 5973053, upload-time = "2025-07-01T09:14:04.491Z" }, + { url = "https://files.pythonhosted.org/packages/f2/2f/d7675ecae6c43e9f12aa8d58b6012683b20b6edfbdac7abcb4e6af7a3784/pillow-11.3.0-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:106064daa23a745510dabce1d84f29137a37224831d88eb4ce94bb187b1d7e5f", size = 6640273, upload-time = "2025-07-01T09:14:06.235Z" }, + { url = "https://files.pythonhosted.org/packages/45/ad/931694675ede172e15b2ff03c8144a0ddaea1d87adb72bb07655eaffb654/pillow-11.3.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:cd8ff254faf15591e724dc7c4ddb6bf4793efcbe13802a4ae3e863cd300b493e", size = 6082043, upload-time = "2025-07-01T09:14:07.978Z" }, + { url = "https://files.pythonhosted.org/packages/3a/04/ba8f2b11fc80d2dd462d7abec16351b45ec99cbbaea4387648a44190351a/pillow-11.3.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:932c754c2d51ad2b2271fd01c3d121daaa35e27efae2a616f77bf164bc0b3e94", size = 6715516, upload-time = "2025-07-01T09:14:10.233Z" }, + { url = "https://files.pythonhosted.org/packages/48/59/8cd06d7f3944cc7d892e8533c56b0acb68399f640786313275faec1e3b6f/pillow-11.3.0-cp311-cp311-win32.whl", hash = "sha256:b4b8f3efc8d530a1544e5962bd6b403d5f7fe8b9e08227c6b255f98ad82b4ba0", size = 6274768, upload-time = "2025-07-01T09:14:11.921Z" }, + { url = "https://files.pythonhosted.org/packages/f1/cc/29c0f5d64ab8eae20f3232da8f8571660aa0ab4b8f1331da5c2f5f9a938e/pillow-11.3.0-cp311-cp311-win_amd64.whl", hash = "sha256:1a992e86b0dd7aeb1f053cd506508c0999d710a8f07b4c791c63843fc6a807ac", size = 6986055, upload-time = "2025-07-01T09:14:13.623Z" }, + { url = "https://files.pythonhosted.org/packages/c6/df/90bd886fabd544c25addd63e5ca6932c86f2b701d5da6c7839387a076b4a/pillow-11.3.0-cp311-cp311-win_arm64.whl", hash = "sha256:30807c931ff7c095620fe04448e2c2fc673fcbb1ffe2a7da3fb39613489b1ddd", size = 2423079, upload-time = "2025-07-01T09:14:15.268Z" }, + { url = "https://files.pythonhosted.org/packages/40/fe/1bc9b3ee13f68487a99ac9529968035cca2f0a51ec36892060edcc51d06a/pillow-11.3.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:fdae223722da47b024b867c1ea0be64e0df702c5e0a60e27daad39bf960dd1e4", size = 5278800, upload-time = "2025-07-01T09:14:17.648Z" }, + { url = "https://files.pythonhosted.org/packages/2c/32/7e2ac19b5713657384cec55f89065fb306b06af008cfd87e572035b27119/pillow-11.3.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:921bd305b10e82b4d1f5e802b6850677f965d8394203d182f078873851dada69", size = 4686296, upload-time = "2025-07-01T09:14:19.828Z" }, + { url = "https://files.pythonhosted.org/packages/8e/1e/b9e12bbe6e4c2220effebc09ea0923a07a6da1e1f1bfbc8d7d29a01ce32b/pillow-11.3.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:eb76541cba2f958032d79d143b98a3a6b3ea87f0959bbe256c0b5e416599fd5d", size = 5871726, upload-time = "2025-07-03T13:10:04.448Z" }, + { url = "https://files.pythonhosted.org/packages/8d/33/e9200d2bd7ba00dc3ddb78df1198a6e80d7669cce6c2bdbeb2530a74ec58/pillow-11.3.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:67172f2944ebba3d4a7b54f2e95c786a3a50c21b88456329314caaa28cda70f6", size = 7644652, upload-time = "2025-07-03T13:10:10.391Z" }, + { url = "https://files.pythonhosted.org/packages/41/f1/6f2427a26fc683e00d985bc391bdd76d8dd4e92fac33d841127eb8fb2313/pillow-11.3.0-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:97f07ed9f56a3b9b5f49d3661dc9607484e85c67e27f3e8be2c7d28ca032fec7", size = 5977787, upload-time = "2025-07-01T09:14:21.63Z" }, + { url = "https://files.pythonhosted.org/packages/e4/c9/06dd4a38974e24f932ff5f98ea3c546ce3f8c995d3f0985f8e5ba48bba19/pillow-11.3.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:676b2815362456b5b3216b4fd5bd89d362100dc6f4945154ff172e206a22c024", size = 6645236, upload-time = "2025-07-01T09:14:23.321Z" }, + { url = "https://files.pythonhosted.org/packages/40/e7/848f69fb79843b3d91241bad658e9c14f39a32f71a301bcd1d139416d1be/pillow-11.3.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:3e184b2f26ff146363dd07bde8b711833d7b0202e27d13540bfe2e35a323a809", size = 6086950, upload-time = "2025-07-01T09:14:25.237Z" }, + { url = "https://files.pythonhosted.org/packages/0b/1a/7cff92e695a2a29ac1958c2a0fe4c0b2393b60aac13b04a4fe2735cad52d/pillow-11.3.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:6be31e3fc9a621e071bc17bb7de63b85cbe0bfae91bb0363c893cbe67247780d", size = 6723358, upload-time = "2025-07-01T09:14:27.053Z" }, + { url = "https://files.pythonhosted.org/packages/26/7d/73699ad77895f69edff76b0f332acc3d497f22f5d75e5360f78cbcaff248/pillow-11.3.0-cp312-cp312-win32.whl", hash = "sha256:7b161756381f0918e05e7cb8a371fff367e807770f8fe92ecb20d905d0e1c149", size = 6275079, upload-time = "2025-07-01T09:14:30.104Z" }, + { url = "https://files.pythonhosted.org/packages/8c/ce/e7dfc873bdd9828f3b6e5c2bbb74e47a98ec23cc5c74fc4e54462f0d9204/pillow-11.3.0-cp312-cp312-win_amd64.whl", hash = "sha256:a6444696fce635783440b7f7a9fc24b3ad10a9ea3f0ab66c5905be1c19ccf17d", size = 6986324, upload-time = "2025-07-01T09:14:31.899Z" }, + { url = "https://files.pythonhosted.org/packages/16/8f/b13447d1bf0b1f7467ce7d86f6e6edf66c0ad7cf44cf5c87a37f9bed9936/pillow-11.3.0-cp312-cp312-win_arm64.whl", hash = "sha256:2aceea54f957dd4448264f9bf40875da0415c83eb85f55069d89c0ed436e3542", size = 2423067, upload-time = "2025-07-01T09:14:33.709Z" }, + { url = "https://files.pythonhosted.org/packages/1e/93/0952f2ed8db3a5a4c7a11f91965d6184ebc8cd7cbb7941a260d5f018cd2d/pillow-11.3.0-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:1c627742b539bba4309df89171356fcb3cc5a9178355b2727d1b74a6cf155fbd", size = 2128328, upload-time = "2025-07-01T09:14:35.276Z" }, + { url = "https://files.pythonhosted.org/packages/4b/e8/100c3d114b1a0bf4042f27e0f87d2f25e857e838034e98ca98fe7b8c0a9c/pillow-11.3.0-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:30b7c02f3899d10f13d7a48163c8969e4e653f8b43416d23d13d1bbfdc93b9f8", size = 2170652, upload-time = "2025-07-01T09:14:37.203Z" }, + { url = "https://files.pythonhosted.org/packages/aa/86/3f758a28a6e381758545f7cdb4942e1cb79abd271bea932998fc0db93cb6/pillow-11.3.0-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:7859a4cc7c9295f5838015d8cc0a9c215b77e43d07a25e460f35cf516df8626f", size = 2227443, upload-time = "2025-07-01T09:14:39.344Z" }, + { url = "https://files.pythonhosted.org/packages/01/f4/91d5b3ffa718df2f53b0dc109877993e511f4fd055d7e9508682e8aba092/pillow-11.3.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:ec1ee50470b0d050984394423d96325b744d55c701a439d2bd66089bff963d3c", size = 5278474, upload-time = "2025-07-01T09:14:41.843Z" }, + { url = "https://files.pythonhosted.org/packages/f9/0e/37d7d3eca6c879fbd9dba21268427dffda1ab00d4eb05b32923d4fbe3b12/pillow-11.3.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7db51d222548ccfd274e4572fdbf3e810a5e66b00608862f947b163e613b67dd", size = 4686038, upload-time = "2025-07-01T09:14:44.008Z" }, + { url = "https://files.pythonhosted.org/packages/ff/b0/3426e5c7f6565e752d81221af9d3676fdbb4f352317ceafd42899aaf5d8a/pillow-11.3.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:2d6fcc902a24ac74495df63faad1884282239265c6839a0a6416d33faedfae7e", size = 5864407, upload-time = "2025-07-03T13:10:15.628Z" }, + { url = "https://files.pythonhosted.org/packages/fc/c1/c6c423134229f2a221ee53f838d4be9d82bab86f7e2f8e75e47b6bf6cd77/pillow-11.3.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:f0f5d8f4a08090c6d6d578351a2b91acf519a54986c055af27e7a93feae6d3f1", size = 7639094, upload-time = "2025-07-03T13:10:21.857Z" }, + { url = "https://files.pythonhosted.org/packages/ba/c9/09e6746630fe6372c67c648ff9deae52a2bc20897d51fa293571977ceb5d/pillow-11.3.0-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c37d8ba9411d6003bba9e518db0db0c58a680ab9fe5179f040b0463644bc9805", size = 5973503, upload-time = "2025-07-01T09:14:45.698Z" }, + { url = "https://files.pythonhosted.org/packages/d5/1c/a2a29649c0b1983d3ef57ee87a66487fdeb45132df66ab30dd37f7dbe162/pillow-11.3.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:13f87d581e71d9189ab21fe0efb5a23e9f28552d5be6979e84001d3b8505abe8", size = 6642574, upload-time = "2025-07-01T09:14:47.415Z" }, + { url = "https://files.pythonhosted.org/packages/36/de/d5cc31cc4b055b6c6fd990e3e7f0f8aaf36229a2698501bcb0cdf67c7146/pillow-11.3.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:023f6d2d11784a465f09fd09a34b150ea4672e85fb3d05931d89f373ab14abb2", size = 6084060, upload-time = "2025-07-01T09:14:49.636Z" }, + { url = "https://files.pythonhosted.org/packages/d5/ea/502d938cbaeec836ac28a9b730193716f0114c41325db428e6b280513f09/pillow-11.3.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:45dfc51ac5975b938e9809451c51734124e73b04d0f0ac621649821a63852e7b", size = 6721407, upload-time = "2025-07-01T09:14:51.962Z" }, + { url = "https://files.pythonhosted.org/packages/45/9c/9c5e2a73f125f6cbc59cc7087c8f2d649a7ae453f83bd0362ff7c9e2aee2/pillow-11.3.0-cp313-cp313-win32.whl", hash = "sha256:a4d336baed65d50d37b88ca5b60c0fa9d81e3a87d4a7930d3880d1624d5b31f3", size = 6273841, upload-time = "2025-07-01T09:14:54.142Z" }, + { url = "https://files.pythonhosted.org/packages/23/85/397c73524e0cd212067e0c969aa245b01d50183439550d24d9f55781b776/pillow-11.3.0-cp313-cp313-win_amd64.whl", hash = "sha256:0bce5c4fd0921f99d2e858dc4d4d64193407e1b99478bc5cacecba2311abde51", size = 6978450, upload-time = "2025-07-01T09:14:56.436Z" }, + { url = "https://files.pythonhosted.org/packages/17/d2/622f4547f69cd173955194b78e4d19ca4935a1b0f03a302d655c9f6aae65/pillow-11.3.0-cp313-cp313-win_arm64.whl", hash = "sha256:1904e1264881f682f02b7f8167935cce37bc97db457f8e7849dc3a6a52b99580", size = 2423055, upload-time = "2025-07-01T09:14:58.072Z" }, + { url = "https://files.pythonhosted.org/packages/dd/80/a8a2ac21dda2e82480852978416cfacd439a4b490a501a288ecf4fe2532d/pillow-11.3.0-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:4c834a3921375c48ee6b9624061076bc0a32a60b5532b322cc0ea64e639dd50e", size = 5281110, upload-time = "2025-07-01T09:14:59.79Z" }, + { url = "https://files.pythonhosted.org/packages/44/d6/b79754ca790f315918732e18f82a8146d33bcd7f4494380457ea89eb883d/pillow-11.3.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:5e05688ccef30ea69b9317a9ead994b93975104a677a36a8ed8106be9260aa6d", size = 4689547, upload-time = "2025-07-01T09:15:01.648Z" }, + { url = "https://files.pythonhosted.org/packages/49/20/716b8717d331150cb00f7fdd78169c01e8e0c219732a78b0e59b6bdb2fd6/pillow-11.3.0-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:1019b04af07fc0163e2810167918cb5add8d74674b6267616021ab558dc98ced", size = 5901554, upload-time = "2025-07-03T13:10:27.018Z" }, + { url = "https://files.pythonhosted.org/packages/74/cf/a9f3a2514a65bb071075063a96f0a5cf949c2f2fce683c15ccc83b1c1cab/pillow-11.3.0-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:f944255db153ebb2b19c51fe85dd99ef0ce494123f21b9db4877ffdfc5590c7c", size = 7669132, upload-time = "2025-07-03T13:10:33.01Z" }, + { url = "https://files.pythonhosted.org/packages/98/3c/da78805cbdbee9cb43efe8261dd7cc0b4b93f2ac79b676c03159e9db2187/pillow-11.3.0-cp313-cp313t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1f85acb69adf2aaee8b7da124efebbdb959a104db34d3a2cb0f3793dbae422a8", size = 6005001, upload-time = "2025-07-01T09:15:03.365Z" }, + { url = "https://files.pythonhosted.org/packages/6c/fa/ce044b91faecf30e635321351bba32bab5a7e034c60187fe9698191aef4f/pillow-11.3.0-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:05f6ecbeff5005399bb48d198f098a9b4b6bdf27b8487c7f38ca16eeb070cd59", size = 6668814, upload-time = "2025-07-01T09:15:05.655Z" }, + { url = "https://files.pythonhosted.org/packages/7b/51/90f9291406d09bf93686434f9183aba27b831c10c87746ff49f127ee80cb/pillow-11.3.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:a7bc6e6fd0395bc052f16b1a8670859964dbd7003bd0af2ff08342eb6e442cfe", size = 6113124, upload-time = "2025-07-01T09:15:07.358Z" }, + { url = "https://files.pythonhosted.org/packages/cd/5a/6fec59b1dfb619234f7636d4157d11fb4e196caeee220232a8d2ec48488d/pillow-11.3.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:83e1b0161c9d148125083a35c1c5a89db5b7054834fd4387499e06552035236c", size = 6747186, upload-time = "2025-07-01T09:15:09.317Z" }, + { url = "https://files.pythonhosted.org/packages/49/6b/00187a044f98255225f172de653941e61da37104a9ea60e4f6887717e2b5/pillow-11.3.0-cp313-cp313t-win32.whl", hash = "sha256:2a3117c06b8fb646639dce83694f2f9eac405472713fcb1ae887469c0d4f6788", size = 6277546, upload-time = "2025-07-01T09:15:11.311Z" }, + { url = "https://files.pythonhosted.org/packages/e8/5c/6caaba7e261c0d75bab23be79f1d06b5ad2a2ae49f028ccec801b0e853d6/pillow-11.3.0-cp313-cp313t-win_amd64.whl", hash = "sha256:857844335c95bea93fb39e0fa2726b4d9d758850b34075a7e3ff4f4fa3aa3b31", size = 6985102, upload-time = "2025-07-01T09:15:13.164Z" }, + { url = "https://files.pythonhosted.org/packages/f3/7e/b623008460c09a0cb38263c93b828c666493caee2eb34ff67f778b87e58c/pillow-11.3.0-cp313-cp313t-win_arm64.whl", hash = "sha256:8797edc41f3e8536ae4b10897ee2f637235c94f27404cac7297f7b607dd0716e", size = 2424803, upload-time = "2025-07-01T09:15:15.695Z" }, + { url = "https://files.pythonhosted.org/packages/73/f4/04905af42837292ed86cb1b1dabe03dce1edc008ef14c473c5c7e1443c5d/pillow-11.3.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:d9da3df5f9ea2a89b81bb6087177fb1f4d1c7146d583a3fe5c672c0d94e55e12", size = 5278520, upload-time = "2025-07-01T09:15:17.429Z" }, + { url = "https://files.pythonhosted.org/packages/41/b0/33d79e377a336247df6348a54e6d2a2b85d644ca202555e3faa0cf811ecc/pillow-11.3.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:0b275ff9b04df7b640c59ec5a3cb113eefd3795a8df80bac69646ef699c6981a", size = 4686116, upload-time = "2025-07-01T09:15:19.423Z" }, + { url = "https://files.pythonhosted.org/packages/49/2d/ed8bc0ab219ae8768f529597d9509d184fe8a6c4741a6864fea334d25f3f/pillow-11.3.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:0743841cabd3dba6a83f38a92672cccbd69af56e3e91777b0ee7f4dba4385632", size = 5864597, upload-time = "2025-07-03T13:10:38.404Z" }, + { url = "https://files.pythonhosted.org/packages/b5/3d/b932bb4225c80b58dfadaca9d42d08d0b7064d2d1791b6a237f87f661834/pillow-11.3.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:2465a69cf967b8b49ee1b96d76718cd98c4e925414ead59fdf75cf0fd07df673", size = 7638246, upload-time = "2025-07-03T13:10:44.987Z" }, + { url = "https://files.pythonhosted.org/packages/09/b5/0487044b7c096f1b48f0d7ad416472c02e0e4bf6919541b111efd3cae690/pillow-11.3.0-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:41742638139424703b4d01665b807c6468e23e699e8e90cffefe291c5832b027", size = 5973336, upload-time = "2025-07-01T09:15:21.237Z" }, + { url = "https://files.pythonhosted.org/packages/a8/2d/524f9318f6cbfcc79fbc004801ea6b607ec3f843977652fdee4857a7568b/pillow-11.3.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:93efb0b4de7e340d99057415c749175e24c8864302369e05914682ba642e5d77", size = 6642699, upload-time = "2025-07-01T09:15:23.186Z" }, + { url = "https://files.pythonhosted.org/packages/6f/d2/a9a4f280c6aefedce1e8f615baaa5474e0701d86dd6f1dede66726462bbd/pillow-11.3.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:7966e38dcd0fa11ca390aed7c6f20454443581d758242023cf36fcb319b1a874", size = 6083789, upload-time = "2025-07-01T09:15:25.1Z" }, + { url = "https://files.pythonhosted.org/packages/fe/54/86b0cd9dbb683a9d5e960b66c7379e821a19be4ac5810e2e5a715c09a0c0/pillow-11.3.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:98a9afa7b9007c67ed84c57c9e0ad86a6000da96eaa638e4f8abe5b65ff83f0a", size = 6720386, upload-time = "2025-07-01T09:15:27.378Z" }, + { url = "https://files.pythonhosted.org/packages/e7/95/88efcaf384c3588e24259c4203b909cbe3e3c2d887af9e938c2022c9dd48/pillow-11.3.0-cp314-cp314-win32.whl", hash = "sha256:02a723e6bf909e7cea0dac1b0e0310be9d7650cd66222a5f1c571455c0a45214", size = 6370911, upload-time = "2025-07-01T09:15:29.294Z" }, + { url = "https://files.pythonhosted.org/packages/2e/cc/934e5820850ec5eb107e7b1a72dd278140731c669f396110ebc326f2a503/pillow-11.3.0-cp314-cp314-win_amd64.whl", hash = "sha256:a418486160228f64dd9e9efcd132679b7a02a5f22c982c78b6fc7dab3fefb635", size = 7117383, upload-time = "2025-07-01T09:15:31.128Z" }, + { url = "https://files.pythonhosted.org/packages/d6/e9/9c0a616a71da2a5d163aa37405e8aced9a906d574b4a214bede134e731bc/pillow-11.3.0-cp314-cp314-win_arm64.whl", hash = "sha256:155658efb5e044669c08896c0c44231c5e9abcaadbc5cd3648df2f7c0b96b9a6", size = 2511385, upload-time = "2025-07-01T09:15:33.328Z" }, + { url = "https://files.pythonhosted.org/packages/1a/33/c88376898aff369658b225262cd4f2659b13e8178e7534df9e6e1fa289f6/pillow-11.3.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:59a03cdf019efbfeeed910bf79c7c93255c3d54bc45898ac2a4140071b02b4ae", size = 5281129, upload-time = "2025-07-01T09:15:35.194Z" }, + { url = "https://files.pythonhosted.org/packages/1f/70/d376247fb36f1844b42910911c83a02d5544ebd2a8bad9efcc0f707ea774/pillow-11.3.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:f8a5827f84d973d8636e9dc5764af4f0cf2318d26744b3d902931701b0d46653", size = 4689580, upload-time = "2025-07-01T09:15:37.114Z" }, + { url = "https://files.pythonhosted.org/packages/eb/1c/537e930496149fbac69efd2fc4329035bbe2e5475b4165439e3be9cb183b/pillow-11.3.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:ee92f2fd10f4adc4b43d07ec5e779932b4eb3dbfbc34790ada5a6669bc095aa6", size = 5902860, upload-time = "2025-07-03T13:10:50.248Z" }, + { url = "https://files.pythonhosted.org/packages/bd/57/80f53264954dcefeebcf9dae6e3eb1daea1b488f0be8b8fef12f79a3eb10/pillow-11.3.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:c96d333dcf42d01f47b37e0979b6bd73ec91eae18614864622d9b87bbd5bbf36", size = 7670694, upload-time = "2025-07-03T13:10:56.432Z" }, + { url = "https://files.pythonhosted.org/packages/70/ff/4727d3b71a8578b4587d9c276e90efad2d6fe0335fd76742a6da08132e8c/pillow-11.3.0-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4c96f993ab8c98460cd0c001447bff6194403e8b1d7e149ade5f00594918128b", size = 6005888, upload-time = "2025-07-01T09:15:39.436Z" }, + { url = "https://files.pythonhosted.org/packages/05/ae/716592277934f85d3be51d7256f3636672d7b1abfafdc42cf3f8cbd4b4c8/pillow-11.3.0-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:41342b64afeba938edb034d122b2dda5db2139b9a4af999729ba8818e0056477", size = 6670330, upload-time = "2025-07-01T09:15:41.269Z" }, + { url = "https://files.pythonhosted.org/packages/e7/bb/7fe6cddcc8827b01b1a9766f5fdeb7418680744f9082035bdbabecf1d57f/pillow-11.3.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:068d9c39a2d1b358eb9f245ce7ab1b5c3246c7c8c7d9ba58cfa5b43146c06e50", size = 6114089, upload-time = "2025-07-01T09:15:43.13Z" }, + { url = "https://files.pythonhosted.org/packages/8b/f5/06bfaa444c8e80f1a8e4bff98da9c83b37b5be3b1deaa43d27a0db37ef84/pillow-11.3.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:a1bc6ba083b145187f648b667e05a2534ecc4b9f2784c2cbe3089e44868f2b9b", size = 6748206, upload-time = "2025-07-01T09:15:44.937Z" }, + { url = "https://files.pythonhosted.org/packages/f0/77/bc6f92a3e8e6e46c0ca78abfffec0037845800ea38c73483760362804c41/pillow-11.3.0-cp314-cp314t-win32.whl", hash = "sha256:118ca10c0d60b06d006be10a501fd6bbdfef559251ed31b794668ed569c87e12", size = 6377370, upload-time = "2025-07-01T09:15:46.673Z" }, + { url = "https://files.pythonhosted.org/packages/4a/82/3a721f7d69dca802befb8af08b7c79ebcab461007ce1c18bd91a5d5896f9/pillow-11.3.0-cp314-cp314t-win_amd64.whl", hash = "sha256:8924748b688aa210d79883357d102cd64690e56b923a186f35a82cbc10f997db", size = 7121500, upload-time = "2025-07-01T09:15:48.512Z" }, + { url = "https://files.pythonhosted.org/packages/89/c7/5572fa4a3f45740eaab6ae86fcdf7195b55beac1371ac8c619d880cfe948/pillow-11.3.0-cp314-cp314t-win_arm64.whl", hash = "sha256:79ea0d14d3ebad43ec77ad5272e6ff9bba5b679ef73375ea760261207fa8e0aa", size = 2512835, upload-time = "2025-07-01T09:15:50.399Z" }, + { url = "https://files.pythonhosted.org/packages/6f/8b/209bd6b62ce8367f47e68a218bffac88888fdf2c9fcf1ecadc6c3ec1ebc7/pillow-11.3.0-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:3cee80663f29e3843b68199b9d6f4f54bd1d4a6b59bdd91bceefc51238bcb967", size = 5270556, upload-time = "2025-07-01T09:16:09.961Z" }, + { url = "https://files.pythonhosted.org/packages/2e/e6/231a0b76070c2cfd9e260a7a5b504fb72da0a95279410fa7afd99d9751d6/pillow-11.3.0-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:b5f56c3f344f2ccaf0dd875d3e180f631dc60a51b314295a3e681fe8cf851fbe", size = 4654625, upload-time = "2025-07-01T09:16:11.913Z" }, + { url = "https://files.pythonhosted.org/packages/13/f4/10cf94fda33cb12765f2397fc285fa6d8eb9c29de7f3185165b702fc7386/pillow-11.3.0-pp310-pypy310_pp73-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:e67d793d180c9df62f1f40aee3accca4829d3794c95098887edc18af4b8b780c", size = 4874207, upload-time = "2025-07-03T13:11:10.201Z" }, + { url = "https://files.pythonhosted.org/packages/72/c9/583821097dc691880c92892e8e2d41fe0a5a3d6021f4963371d2f6d57250/pillow-11.3.0-pp310-pypy310_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:d000f46e2917c705e9fb93a3606ee4a819d1e3aa7a9b442f6444f07e77cf5e25", size = 6583939, upload-time = "2025-07-03T13:11:15.68Z" }, + { url = "https://files.pythonhosted.org/packages/3b/8e/5c9d410f9217b12320efc7c413e72693f48468979a013ad17fd690397b9a/pillow-11.3.0-pp310-pypy310_pp73-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:527b37216b6ac3a12d7838dc3bd75208ec57c1c6d11ef01902266a5a0c14fc27", size = 4957166, upload-time = "2025-07-01T09:16:13.74Z" }, + { url = "https://files.pythonhosted.org/packages/62/bb/78347dbe13219991877ffb3a91bf09da8317fbfcd4b5f9140aeae020ad71/pillow-11.3.0-pp310-pypy310_pp73-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:be5463ac478b623b9dd3937afd7fb7ab3d79dd290a28e2b6df292dc75063eb8a", size = 5581482, upload-time = "2025-07-01T09:16:16.107Z" }, + { url = "https://files.pythonhosted.org/packages/d9/28/1000353d5e61498aaeaaf7f1e4b49ddb05f2c6575f9d4f9f914a3538b6e1/pillow-11.3.0-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:8dc70ca24c110503e16918a658b869019126ecfe03109b754c402daff12b3d9f", size = 6984596, upload-time = "2025-07-01T09:16:18.07Z" }, + { url = "https://files.pythonhosted.org/packages/9e/e3/6fa84033758276fb31da12e5fb66ad747ae83b93c67af17f8c6ff4cc8f34/pillow-11.3.0-pp311-pypy311_pp73-macosx_10_15_x86_64.whl", hash = "sha256:7c8ec7a017ad1bd562f93dbd8505763e688d388cde6e4a010ae1486916e713e6", size = 5270566, upload-time = "2025-07-01T09:16:19.801Z" }, + { url = "https://files.pythonhosted.org/packages/5b/ee/e8d2e1ab4892970b561e1ba96cbd59c0d28cf66737fc44abb2aec3795a4e/pillow-11.3.0-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:9ab6ae226de48019caa8074894544af5b53a117ccb9d3b3dcb2871464c829438", size = 4654618, upload-time = "2025-07-01T09:16:21.818Z" }, + { url = "https://files.pythonhosted.org/packages/f2/6d/17f80f4e1f0761f02160fc433abd4109fa1548dcfdca46cfdadaf9efa565/pillow-11.3.0-pp311-pypy311_pp73-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:fe27fb049cdcca11f11a7bfda64043c37b30e6b91f10cb5bab275806c32f6ab3", size = 4874248, upload-time = "2025-07-03T13:11:20.738Z" }, + { url = "https://files.pythonhosted.org/packages/de/5f/c22340acd61cef960130585bbe2120e2fd8434c214802f07e8c03596b17e/pillow-11.3.0-pp311-pypy311_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:465b9e8844e3c3519a983d58b80be3f668e2a7a5db97f2784e7079fbc9f9822c", size = 6583963, upload-time = "2025-07-03T13:11:26.283Z" }, + { url = "https://files.pythonhosted.org/packages/31/5e/03966aedfbfcbb4d5f8aa042452d3361f325b963ebbadddac05b122e47dd/pillow-11.3.0-pp311-pypy311_pp73-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5418b53c0d59b3824d05e029669efa023bbef0f3e92e75ec8428f3799487f361", size = 4957170, upload-time = "2025-07-01T09:16:23.762Z" }, + { url = "https://files.pythonhosted.org/packages/cc/2d/e082982aacc927fc2cab48e1e731bdb1643a1406acace8bed0900a61464e/pillow-11.3.0-pp311-pypy311_pp73-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:504b6f59505f08ae014f724b6207ff6222662aab5cc9542577fb084ed0676ac7", size = 5581505, upload-time = "2025-07-01T09:16:25.593Z" }, + { url = "https://files.pythonhosted.org/packages/34/e7/ae39f538fd6844e982063c3a5e4598b8ced43b9633baa3a85ef33af8c05c/pillow-11.3.0-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:c84d689db21a1c397d001aa08241044aa2069e7587b398c8cc63020390b1c1b8", size = 6984598, upload-time = "2025-07-01T09:16:27.732Z" }, +] + +[[package]] +name = "platformdirs" +version = "4.5.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/61/33/9611380c2bdb1225fdef633e2a9610622310fed35ab11dac9620972ee088/platformdirs-4.5.0.tar.gz", hash = "sha256:70ddccdd7c99fc5942e9fc25636a8b34d04c24b335100223152c2803e4063312", size = 21632, upload-time = "2025-10-08T17:44:48.791Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/73/cb/ac7874b3e5d58441674fb70742e6c374b28b0c7cb988d37d991cde47166c/platformdirs-4.5.0-py3-none-any.whl", hash = "sha256:e578a81bb873cbb89a41fcc904c7ef523cc18284b7e3b3ccf06aca1403b7ebd3", size = 18651, upload-time = "2025-10-08T17:44:47.223Z" }, +] + +[[package]] +name = "pluggy" +version = "1.6.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/f9/e2/3e91f31a7d2b083fe6ef3fa267035b518369d9511ffab804f839851d2779/pluggy-1.6.0.tar.gz", hash = "sha256:7dcc130b76258d33b90f61b658791dede3486c3e6bfb003ee5c9bfb396dd22f3", size = 69412, upload-time = "2025-05-15T12:30:07.975Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/54/20/4d324d65cc6d9205fabedc306948156824eb9f0ee1633355a8f7ec5c66bf/pluggy-1.6.0-py3-none-any.whl", hash = "sha256:e920276dd6813095e9377c0bc5566d94c932c33b27a3e3945d8389c374dd4746", size = 20538, upload-time = "2025-05-15T12:30:06.134Z" }, +] + +[[package]] +name = "pre-commit" +version = "4.0.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "cfgv" }, + { name = "identify" }, + { name = "nodeenv" }, + { name = "pyyaml" }, + { name = "virtualenv" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/2e/c8/e22c292035f1bac8b9f5237a2622305bc0304e776080b246f3df57c4ff9f/pre_commit-4.0.1.tar.gz", hash = "sha256:80905ac375958c0444c65e9cebebd948b3cdb518f335a091a670a89d652139d2", size = 191678, upload-time = "2024-10-08T16:09:37.641Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/16/8f/496e10d51edd6671ebe0432e33ff800aa86775d2d147ce7d43389324a525/pre_commit-4.0.1-py2.py3-none-any.whl", hash = "sha256:efde913840816312445dc98787724647c65473daefe420785f885e8ed9a06878", size = 218713, upload-time = "2024-10-08T16:09:35.726Z" }, +] + +[[package]] +name = "proglog" +version = "0.1.12" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "tqdm" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/c2/af/c108866c452eda1132f3d6b3cb6be2ae8430c97e9309f38ca9dbd430af37/proglog-0.1.12.tar.gz", hash = "sha256:361ee074721c277b89b75c061336cb8c5f287c92b043efa562ccf7866cda931c", size = 8794, upload-time = "2025-05-09T14:36:18.316Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/c1/1b/f7ea6cde25621cd9236541c66ff018f4268012a534ec31032bcb187dc5e7/proglog-0.1.12-py3-none-any.whl", hash = "sha256:ccaafce51e80a81c65dc907a460c07ccb8ec1f78dc660cfd8f9ec3a22f01b84c", size = 6337, upload-time = "2025-05-09T14:36:16.798Z" }, +] + +[[package]] +name = "propcache" +version = "0.4.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/9e/da/e9fc233cf63743258bff22b3dfa7ea5baef7b5bc324af47a0ad89b8ffc6f/propcache-0.4.1.tar.gz", hash = "sha256:f48107a8c637e80362555f37ecf49abe20370e557cc4ab374f04ec4423c97c3d", size = 46442, upload-time = "2025-10-08T19:49:02.291Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/3c/0e/934b541323035566a9af292dba85a195f7b78179114f2c6ebb24551118a9/propcache-0.4.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:7c2d1fa3201efaf55d730400d945b5b3ab6e672e100ba0f9a409d950ab25d7db", size = 79534, upload-time = "2025-10-08T19:46:02.083Z" }, + { url = "https://files.pythonhosted.org/packages/a1/6b/db0d03d96726d995dc7171286c6ba9d8d14251f37433890f88368951a44e/propcache-0.4.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:1eb2994229cc8ce7fe9b3db88f5465f5fd8651672840b2e426b88cdb1a30aac8", size = 45526, upload-time = "2025-10-08T19:46:03.884Z" }, + { url = "https://files.pythonhosted.org/packages/e4/c3/82728404aea669e1600f304f2609cde9e665c18df5a11cdd57ed73c1dceb/propcache-0.4.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:66c1f011f45a3b33d7bcb22daed4b29c0c9e2224758b6be00686731e1b46f925", size = 47263, upload-time = "2025-10-08T19:46:05.405Z" }, + { url = "https://files.pythonhosted.org/packages/df/1b/39313ddad2bf9187a1432654c38249bab4562ef535ef07f5eb6eb04d0b1b/propcache-0.4.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9a52009f2adffe195d0b605c25ec929d26b36ef986ba85244891dee3b294df21", size = 201012, upload-time = "2025-10-08T19:46:07.165Z" }, + { url = "https://files.pythonhosted.org/packages/5b/01/f1d0b57d136f294a142acf97f4ed58c8e5b974c21e543000968357115011/propcache-0.4.1-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5d4e2366a9c7b837555cf02fb9be2e3167d333aff716332ef1b7c3a142ec40c5", size = 209491, upload-time = "2025-10-08T19:46:08.909Z" }, + { url = "https://files.pythonhosted.org/packages/a1/c8/038d909c61c5bb039070b3fb02ad5cccdb1dde0d714792e251cdb17c9c05/propcache-0.4.1-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:9d2b6caef873b4f09e26ea7e33d65f42b944837563a47a94719cc3544319a0db", size = 215319, upload-time = "2025-10-08T19:46:10.7Z" }, + { url = "https://files.pythonhosted.org/packages/08/57/8c87e93142b2c1fa2408e45695205a7ba05fb5db458c0bf5c06ba0e09ea6/propcache-0.4.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2b16ec437a8c8a965ecf95739448dd938b5c7f56e67ea009f4300d8df05f32b7", size = 196856, upload-time = "2025-10-08T19:46:12.003Z" }, + { url = "https://files.pythonhosted.org/packages/42/df/5615fec76aa561987a534759b3686008a288e73107faa49a8ae5795a9f7a/propcache-0.4.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:296f4c8ed03ca7476813fe666c9ea97869a8d7aec972618671b33a38a5182ef4", size = 193241, upload-time = "2025-10-08T19:46:13.495Z" }, + { url = "https://files.pythonhosted.org/packages/d5/21/62949eb3a7a54afe8327011c90aca7e03547787a88fb8bd9726806482fea/propcache-0.4.1-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:1f0978529a418ebd1f49dad413a2b68af33f85d5c5ca5c6ca2a3bed375a7ac60", size = 190552, upload-time = "2025-10-08T19:46:14.938Z" }, + { url = "https://files.pythonhosted.org/packages/30/ee/ab4d727dd70806e5b4de96a798ae7ac6e4d42516f030ee60522474b6b332/propcache-0.4.1-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:fd138803047fb4c062b1c1dd95462f5209456bfab55c734458f15d11da288f8f", size = 200113, upload-time = "2025-10-08T19:46:16.695Z" }, + { url = "https://files.pythonhosted.org/packages/8a/0b/38b46208e6711b016aa8966a3ac793eee0d05c7159d8342aa27fc0bc365e/propcache-0.4.1-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:8c9b3cbe4584636d72ff556d9036e0c9317fa27b3ac1f0f558e7e84d1c9c5900", size = 200778, upload-time = "2025-10-08T19:46:18.023Z" }, + { url = "https://files.pythonhosted.org/packages/cf/81/5abec54355ed344476bee711e9f04815d4b00a311ab0535599204eecc257/propcache-0.4.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:f93243fdc5657247533273ac4f86ae106cc6445a0efacb9a1bfe982fcfefd90c", size = 193047, upload-time = "2025-10-08T19:46:19.449Z" }, + { url = "https://files.pythonhosted.org/packages/ec/b6/1f237c04e32063cb034acd5f6ef34ef3a394f75502e72703545631ab1ef6/propcache-0.4.1-cp310-cp310-win32.whl", hash = "sha256:a0ee98db9c5f80785b266eb805016e36058ac72c51a064040f2bc43b61101cdb", size = 38093, upload-time = "2025-10-08T19:46:20.643Z" }, + { url = "https://files.pythonhosted.org/packages/a6/67/354aac4e0603a15f76439caf0427781bcd6797f370377f75a642133bc954/propcache-0.4.1-cp310-cp310-win_amd64.whl", hash = "sha256:1cdb7988c4e5ac7f6d175a28a9aa0c94cb6f2ebe52756a3c0cda98d2809a9e37", size = 41638, upload-time = "2025-10-08T19:46:21.935Z" }, + { url = "https://files.pythonhosted.org/packages/e0/e1/74e55b9fd1a4c209ff1a9a824bf6c8b3d1fc5a1ac3eabe23462637466785/propcache-0.4.1-cp310-cp310-win_arm64.whl", hash = "sha256:d82ad62b19645419fe79dd63b3f9253e15b30e955c0170e5cebc350c1844e581", size = 38229, upload-time = "2025-10-08T19:46:23.368Z" }, + { url = "https://files.pythonhosted.org/packages/8c/d4/4e2c9aaf7ac2242b9358f98dccd8f90f2605402f5afeff6c578682c2c491/propcache-0.4.1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:60a8fda9644b7dfd5dece8c61d8a85e271cb958075bfc4e01083c148b61a7caf", size = 80208, upload-time = "2025-10-08T19:46:24.597Z" }, + { url = "https://files.pythonhosted.org/packages/c2/21/d7b68e911f9c8e18e4ae43bdbc1e1e9bbd971f8866eb81608947b6f585ff/propcache-0.4.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:c30b53e7e6bda1d547cabb47c825f3843a0a1a42b0496087bb58d8fedf9f41b5", size = 45777, upload-time = "2025-10-08T19:46:25.733Z" }, + { url = "https://files.pythonhosted.org/packages/d3/1d/11605e99ac8ea9435651ee71ab4cb4bf03f0949586246476a25aadfec54a/propcache-0.4.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:6918ecbd897443087a3b7cd978d56546a812517dcaaca51b49526720571fa93e", size = 47647, upload-time = "2025-10-08T19:46:27.304Z" }, + { url = "https://files.pythonhosted.org/packages/58/1a/3c62c127a8466c9c843bccb503d40a273e5cc69838805f322e2826509e0d/propcache-0.4.1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3d902a36df4e5989763425a8ab9e98cd8ad5c52c823b34ee7ef307fd50582566", size = 214929, upload-time = "2025-10-08T19:46:28.62Z" }, + { url = "https://files.pythonhosted.org/packages/56/b9/8fa98f850960b367c4b8fe0592e7fc341daa7a9462e925228f10a60cf74f/propcache-0.4.1-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a9695397f85973bb40427dedddf70d8dc4a44b22f1650dd4af9eedf443d45165", size = 221778, upload-time = "2025-10-08T19:46:30.358Z" }, + { url = "https://files.pythonhosted.org/packages/46/a6/0ab4f660eb59649d14b3d3d65c439421cf2f87fe5dd68591cbe3c1e78a89/propcache-0.4.1-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2bb07ffd7eaad486576430c89f9b215f9e4be68c4866a96e97db9e97fead85dc", size = 228144, upload-time = "2025-10-08T19:46:32.607Z" }, + { url = "https://files.pythonhosted.org/packages/52/6a/57f43e054fb3d3a56ac9fc532bc684fc6169a26c75c353e65425b3e56eef/propcache-0.4.1-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fd6f30fdcf9ae2a70abd34da54f18da086160e4d7d9251f81f3da0ff84fc5a48", size = 210030, upload-time = "2025-10-08T19:46:33.969Z" }, + { url = "https://files.pythonhosted.org/packages/40/e2/27e6feebb5f6b8408fa29f5efbb765cd54c153ac77314d27e457a3e993b7/propcache-0.4.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:fc38cba02d1acba4e2869eef1a57a43dfbd3d49a59bf90dda7444ec2be6a5570", size = 208252, upload-time = "2025-10-08T19:46:35.309Z" }, + { url = "https://files.pythonhosted.org/packages/9e/f8/91c27b22ccda1dbc7967f921c42825564fa5336a01ecd72eb78a9f4f53c2/propcache-0.4.1-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:67fad6162281e80e882fb3ec355398cf72864a54069d060321f6cd0ade95fe85", size = 202064, upload-time = "2025-10-08T19:46:36.993Z" }, + { url = "https://files.pythonhosted.org/packages/f2/26/7f00bd6bd1adba5aafe5f4a66390f243acab58eab24ff1a08bebb2ef9d40/propcache-0.4.1-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:f10207adf04d08bec185bae14d9606a1444715bc99180f9331c9c02093e1959e", size = 212429, upload-time = "2025-10-08T19:46:38.398Z" }, + { url = "https://files.pythonhosted.org/packages/84/89/fd108ba7815c1117ddca79c228f3f8a15fc82a73bca8b142eb5de13b2785/propcache-0.4.1-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:e9b0d8d0845bbc4cfcdcbcdbf5086886bc8157aa963c31c777ceff7846c77757", size = 216727, upload-time = "2025-10-08T19:46:39.732Z" }, + { url = "https://files.pythonhosted.org/packages/79/37/3ec3f7e3173e73f1d600495d8b545b53802cbf35506e5732dd8578db3724/propcache-0.4.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:981333cb2f4c1896a12f4ab92a9cc8f09ea664e9b7dbdc4eff74627af3a11c0f", size = 205097, upload-time = "2025-10-08T19:46:41.025Z" }, + { url = "https://files.pythonhosted.org/packages/61/b0/b2631c19793f869d35f47d5a3a56fb19e9160d3c119f15ac7344fc3ccae7/propcache-0.4.1-cp311-cp311-win32.whl", hash = "sha256:f1d2f90aeec838a52f1c1a32fe9a619fefd5e411721a9117fbf82aea638fe8a1", size = 38084, upload-time = "2025-10-08T19:46:42.693Z" }, + { url = "https://files.pythonhosted.org/packages/f4/78/6cce448e2098e9f3bfc91bb877f06aa24b6ccace872e39c53b2f707c4648/propcache-0.4.1-cp311-cp311-win_amd64.whl", hash = "sha256:364426a62660f3f699949ac8c621aad6977be7126c5807ce48c0aeb8e7333ea6", size = 41637, upload-time = "2025-10-08T19:46:43.778Z" }, + { url = "https://files.pythonhosted.org/packages/9c/e9/754f180cccd7f51a39913782c74717c581b9cc8177ad0e949f4d51812383/propcache-0.4.1-cp311-cp311-win_arm64.whl", hash = "sha256:e53f3a38d3510c11953f3e6a33f205c6d1b001129f972805ca9b42fc308bc239", size = 38064, upload-time = "2025-10-08T19:46:44.872Z" }, + { url = "https://files.pythonhosted.org/packages/a2/0f/f17b1b2b221d5ca28b4b876e8bb046ac40466513960646bda8e1853cdfa2/propcache-0.4.1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:e153e9cd40cc8945138822807139367f256f89c6810c2634a4f6902b52d3b4e2", size = 80061, upload-time = "2025-10-08T19:46:46.075Z" }, + { url = "https://files.pythonhosted.org/packages/76/47/8ccf75935f51448ba9a16a71b783eb7ef6b9ee60f5d14c7f8a8a79fbeed7/propcache-0.4.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:cd547953428f7abb73c5ad82cbb32109566204260d98e41e5dfdc682eb7f8403", size = 46037, upload-time = "2025-10-08T19:46:47.23Z" }, + { url = "https://files.pythonhosted.org/packages/0a/b6/5c9a0e42df4d00bfb4a3cbbe5cf9f54260300c88a0e9af1f47ca5ce17ac0/propcache-0.4.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:f048da1b4f243fc44f205dfd320933a951b8d89e0afd4c7cacc762a8b9165207", size = 47324, upload-time = "2025-10-08T19:46:48.384Z" }, + { url = "https://files.pythonhosted.org/packages/9e/d3/6c7ee328b39a81ee877c962469f1e795f9db87f925251efeb0545e0020d0/propcache-0.4.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ec17c65562a827bba85e3872ead335f95405ea1674860d96483a02f5c698fa72", size = 225505, upload-time = "2025-10-08T19:46:50.055Z" }, + { url = "https://files.pythonhosted.org/packages/01/5d/1c53f4563490b1d06a684742cc6076ef944bc6457df6051b7d1a877c057b/propcache-0.4.1-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:405aac25c6394ef275dee4c709be43745d36674b223ba4eb7144bf4d691b7367", size = 230242, upload-time = "2025-10-08T19:46:51.815Z" }, + { url = "https://files.pythonhosted.org/packages/20/e1/ce4620633b0e2422207c3cb774a0ee61cac13abc6217763a7b9e2e3f4a12/propcache-0.4.1-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0013cb6f8dde4b2a2f66903b8ba740bdfe378c943c4377a200551ceb27f379e4", size = 238474, upload-time = "2025-10-08T19:46:53.208Z" }, + { url = "https://files.pythonhosted.org/packages/46/4b/3aae6835b8e5f44ea6a68348ad90f78134047b503765087be2f9912140ea/propcache-0.4.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:15932ab57837c3368b024473a525e25d316d8353016e7cc0e5ba9eb343fbb1cf", size = 221575, upload-time = "2025-10-08T19:46:54.511Z" }, + { url = "https://files.pythonhosted.org/packages/6e/a5/8a5e8678bcc9d3a1a15b9a29165640d64762d424a16af543f00629c87338/propcache-0.4.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:031dce78b9dc099f4c29785d9cf5577a3faf9ebf74ecbd3c856a7b92768c3df3", size = 216736, upload-time = "2025-10-08T19:46:56.212Z" }, + { url = "https://files.pythonhosted.org/packages/f1/63/b7b215eddeac83ca1c6b934f89d09a625aa9ee4ba158338854c87210cc36/propcache-0.4.1-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:ab08df6c9a035bee56e31af99be621526bd237bea9f32def431c656b29e41778", size = 213019, upload-time = "2025-10-08T19:46:57.595Z" }, + { url = "https://files.pythonhosted.org/packages/57/74/f580099a58c8af587cac7ba19ee7cb418506342fbbe2d4a4401661cca886/propcache-0.4.1-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:4d7af63f9f93fe593afbf104c21b3b15868efb2c21d07d8732c0c4287e66b6a6", size = 220376, upload-time = "2025-10-08T19:46:59.067Z" }, + { url = "https://files.pythonhosted.org/packages/c4/ee/542f1313aff7eaf19c2bb758c5d0560d2683dac001a1c96d0774af799843/propcache-0.4.1-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:cfc27c945f422e8b5071b6e93169679e4eb5bf73bbcbf1ba3ae3a83d2f78ebd9", size = 226988, upload-time = "2025-10-08T19:47:00.544Z" }, + { url = "https://files.pythonhosted.org/packages/8f/18/9c6b015dd9c6930f6ce2229e1f02fb35298b847f2087ea2b436a5bfa7287/propcache-0.4.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:35c3277624a080cc6ec6f847cbbbb5b49affa3598c4535a0a4682a697aaa5c75", size = 215615, upload-time = "2025-10-08T19:47:01.968Z" }, + { url = "https://files.pythonhosted.org/packages/80/9e/e7b85720b98c45a45e1fca6a177024934dc9bc5f4d5dd04207f216fc33ed/propcache-0.4.1-cp312-cp312-win32.whl", hash = "sha256:671538c2262dadb5ba6395e26c1731e1d52534bfe9ae56d0b5573ce539266aa8", size = 38066, upload-time = "2025-10-08T19:47:03.503Z" }, + { url = "https://files.pythonhosted.org/packages/54/09/d19cff2a5aaac632ec8fc03737b223597b1e347416934c1b3a7df079784c/propcache-0.4.1-cp312-cp312-win_amd64.whl", hash = "sha256:cb2d222e72399fcf5890d1d5cc1060857b9b236adff2792ff48ca2dfd46c81db", size = 41655, upload-time = "2025-10-08T19:47:04.973Z" }, + { url = "https://files.pythonhosted.org/packages/68/ab/6b5c191bb5de08036a8c697b265d4ca76148efb10fa162f14af14fb5f076/propcache-0.4.1-cp312-cp312-win_arm64.whl", hash = "sha256:204483131fb222bdaaeeea9f9e6c6ed0cac32731f75dfc1d4a567fc1926477c1", size = 37789, upload-time = "2025-10-08T19:47:06.077Z" }, + { url = "https://files.pythonhosted.org/packages/bf/df/6d9c1b6ac12b003837dde8a10231a7344512186e87b36e855bef32241942/propcache-0.4.1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:43eedf29202c08550aac1d14e0ee619b0430aaef78f85864c1a892294fbc28cf", size = 77750, upload-time = "2025-10-08T19:47:07.648Z" }, + { url = "https://files.pythonhosted.org/packages/8b/e8/677a0025e8a2acf07d3418a2e7ba529c9c33caf09d3c1f25513023c1db56/propcache-0.4.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:d62cdfcfd89ccb8de04e0eda998535c406bf5e060ffd56be6c586cbcc05b3311", size = 44780, upload-time = "2025-10-08T19:47:08.851Z" }, + { url = "https://files.pythonhosted.org/packages/89/a4/92380f7ca60f99ebae761936bc48a72a639e8a47b29050615eef757cb2a7/propcache-0.4.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:cae65ad55793da34db5f54e4029b89d3b9b9490d8abe1b4c7ab5d4b8ec7ebf74", size = 46308, upload-time = "2025-10-08T19:47:09.982Z" }, + { url = "https://files.pythonhosted.org/packages/2d/48/c5ac64dee5262044348d1d78a5f85dd1a57464a60d30daee946699963eb3/propcache-0.4.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:333ddb9031d2704a301ee3e506dc46b1fe5f294ec198ed6435ad5b6a085facfe", size = 208182, upload-time = "2025-10-08T19:47:11.319Z" }, + { url = "https://files.pythonhosted.org/packages/c6/0c/cd762dd011a9287389a6a3eb43aa30207bde253610cca06824aeabfe9653/propcache-0.4.1-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:fd0858c20f078a32cf55f7e81473d96dcf3b93fd2ccdb3d40fdf54b8573df3af", size = 211215, upload-time = "2025-10-08T19:47:13.146Z" }, + { url = "https://files.pythonhosted.org/packages/30/3e/49861e90233ba36890ae0ca4c660e95df565b2cd15d4a68556ab5865974e/propcache-0.4.1-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:678ae89ebc632c5c204c794f8dab2837c5f159aeb59e6ed0539500400577298c", size = 218112, upload-time = "2025-10-08T19:47:14.913Z" }, + { url = "https://files.pythonhosted.org/packages/f1/8b/544bc867e24e1bd48f3118cecd3b05c694e160a168478fa28770f22fd094/propcache-0.4.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d472aeb4fbf9865e0c6d622d7f4d54a4e101a89715d8904282bb5f9a2f476c3f", size = 204442, upload-time = "2025-10-08T19:47:16.277Z" }, + { url = "https://files.pythonhosted.org/packages/50/a6/4282772fd016a76d3e5c0df58380a5ea64900afd836cec2c2f662d1b9bb3/propcache-0.4.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:4d3df5fa7e36b3225954fba85589da77a0fe6a53e3976de39caf04a0db4c36f1", size = 199398, upload-time = "2025-10-08T19:47:17.962Z" }, + { url = "https://files.pythonhosted.org/packages/3e/ec/d8a7cd406ee1ddb705db2139f8a10a8a427100347bd698e7014351c7af09/propcache-0.4.1-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:ee17f18d2498f2673e432faaa71698032b0127ebf23ae5974eeaf806c279df24", size = 196920, upload-time = "2025-10-08T19:47:19.355Z" }, + { url = "https://files.pythonhosted.org/packages/f6/6c/f38ab64af3764f431e359f8baf9e0a21013e24329e8b85d2da32e8ed07ca/propcache-0.4.1-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:580e97762b950f993ae618e167e7be9256b8353c2dcd8b99ec100eb50f5286aa", size = 203748, upload-time = "2025-10-08T19:47:21.338Z" }, + { url = "https://files.pythonhosted.org/packages/d6/e3/fa846bd70f6534d647886621388f0a265254d30e3ce47e5c8e6e27dbf153/propcache-0.4.1-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:501d20b891688eb8e7aa903021f0b72d5a55db40ffaab27edefd1027caaafa61", size = 205877, upload-time = "2025-10-08T19:47:23.059Z" }, + { url = "https://files.pythonhosted.org/packages/e2/39/8163fc6f3133fea7b5f2827e8eba2029a0277ab2c5beee6c1db7b10fc23d/propcache-0.4.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:9a0bd56e5b100aef69bd8562b74b46254e7c8812918d3baa700c8a8009b0af66", size = 199437, upload-time = "2025-10-08T19:47:24.445Z" }, + { url = "https://files.pythonhosted.org/packages/93/89/caa9089970ca49c7c01662bd0eeedfe85494e863e8043565aeb6472ce8fe/propcache-0.4.1-cp313-cp313-win32.whl", hash = "sha256:bcc9aaa5d80322bc2fb24bb7accb4a30f81e90ab8d6ba187aec0744bc302ad81", size = 37586, upload-time = "2025-10-08T19:47:25.736Z" }, + { url = "https://files.pythonhosted.org/packages/f5/ab/f76ec3c3627c883215b5c8080debb4394ef5a7a29be811f786415fc1e6fd/propcache-0.4.1-cp313-cp313-win_amd64.whl", hash = "sha256:381914df18634f5494334d201e98245c0596067504b9372d8cf93f4bb23e025e", size = 40790, upload-time = "2025-10-08T19:47:26.847Z" }, + { url = "https://files.pythonhosted.org/packages/59/1b/e71ae98235f8e2ba5004d8cb19765a74877abf189bc53fc0c80d799e56c3/propcache-0.4.1-cp313-cp313-win_arm64.whl", hash = "sha256:8873eb4460fd55333ea49b7d189749ecf6e55bf85080f11b1c4530ed3034cba1", size = 37158, upload-time = "2025-10-08T19:47:27.961Z" }, + { url = "https://files.pythonhosted.org/packages/83/ce/a31bbdfc24ee0dcbba458c8175ed26089cf109a55bbe7b7640ed2470cfe9/propcache-0.4.1-cp313-cp313t-macosx_10_13_universal2.whl", hash = "sha256:92d1935ee1f8d7442da9c0c4fa7ac20d07e94064184811b685f5c4fada64553b", size = 81451, upload-time = "2025-10-08T19:47:29.445Z" }, + { url = "https://files.pythonhosted.org/packages/25/9c/442a45a470a68456e710d96cacd3573ef26a1d0a60067e6a7d5e655621ed/propcache-0.4.1-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:473c61b39e1460d386479b9b2f337da492042447c9b685f28be4f74d3529e566", size = 46374, upload-time = "2025-10-08T19:47:30.579Z" }, + { url = "https://files.pythonhosted.org/packages/f4/bf/b1d5e21dbc3b2e889ea4327044fb16312a736d97640fb8b6aa3f9c7b3b65/propcache-0.4.1-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:c0ef0aaafc66fbd87842a3fe3902fd889825646bc21149eafe47be6072725835", size = 48396, upload-time = "2025-10-08T19:47:31.79Z" }, + { url = "https://files.pythonhosted.org/packages/f4/04/5b4c54a103d480e978d3c8a76073502b18db0c4bc17ab91b3cb5092ad949/propcache-0.4.1-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f95393b4d66bfae908c3ca8d169d5f79cd65636ae15b5e7a4f6e67af675adb0e", size = 275950, upload-time = "2025-10-08T19:47:33.481Z" }, + { url = "https://files.pythonhosted.org/packages/b4/c1/86f846827fb969c4b78b0af79bba1d1ea2156492e1b83dea8b8a6ae27395/propcache-0.4.1-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c07fda85708bc48578467e85099645167a955ba093be0a2dcba962195676e859", size = 273856, upload-time = "2025-10-08T19:47:34.906Z" }, + { url = "https://files.pythonhosted.org/packages/36/1d/fc272a63c8d3bbad6878c336c7a7dea15e8f2d23a544bda43205dfa83ada/propcache-0.4.1-cp313-cp313t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:af223b406d6d000830c6f65f1e6431783fc3f713ba3e6cc8c024d5ee96170a4b", size = 280420, upload-time = "2025-10-08T19:47:36.338Z" }, + { url = "https://files.pythonhosted.org/packages/07/0c/01f2219d39f7e53d52e5173bcb09c976609ba30209912a0680adfb8c593a/propcache-0.4.1-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a78372c932c90ee474559c5ddfffd718238e8673c340dc21fe45c5b8b54559a0", size = 263254, upload-time = "2025-10-08T19:47:37.692Z" }, + { url = "https://files.pythonhosted.org/packages/2d/18/cd28081658ce597898f0c4d174d4d0f3c5b6d4dc27ffafeef835c95eb359/propcache-0.4.1-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:564d9f0d4d9509e1a870c920a89b2fec951b44bf5ba7d537a9e7c1ccec2c18af", size = 261205, upload-time = "2025-10-08T19:47:39.659Z" }, + { url = "https://files.pythonhosted.org/packages/7a/71/1f9e22eb8b8316701c2a19fa1f388c8a3185082607da8e406a803c9b954e/propcache-0.4.1-cp313-cp313t-musllinux_1_2_armv7l.whl", hash = "sha256:17612831fda0138059cc5546f4d12a2aacfb9e47068c06af35c400ba58ba7393", size = 247873, upload-time = "2025-10-08T19:47:41.084Z" }, + { url = "https://files.pythonhosted.org/packages/4a/65/3d4b61f36af2b4eddba9def857959f1016a51066b4f1ce348e0cf7881f58/propcache-0.4.1-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:41a89040cb10bd345b3c1a873b2bf36413d48da1def52f268a055f7398514874", size = 262739, upload-time = "2025-10-08T19:47:42.51Z" }, + { url = "https://files.pythonhosted.org/packages/2a/42/26746ab087faa77c1c68079b228810436ccd9a5ce9ac85e2b7307195fd06/propcache-0.4.1-cp313-cp313t-musllinux_1_2_s390x.whl", hash = "sha256:e35b88984e7fa64aacecea39236cee32dd9bd8c55f57ba8a75cf2399553f9bd7", size = 263514, upload-time = "2025-10-08T19:47:43.927Z" }, + { url = "https://files.pythonhosted.org/packages/94/13/630690fe201f5502d2403dd3cfd451ed8858fe3c738ee88d095ad2ff407b/propcache-0.4.1-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:6f8b465489f927b0df505cbe26ffbeed4d6d8a2bbc61ce90eb074ff129ef0ab1", size = 257781, upload-time = "2025-10-08T19:47:45.448Z" }, + { url = "https://files.pythonhosted.org/packages/92/f7/1d4ec5841505f423469efbfc381d64b7b467438cd5a4bbcbb063f3b73d27/propcache-0.4.1-cp313-cp313t-win32.whl", hash = "sha256:2ad890caa1d928c7c2965b48f3a3815c853180831d0e5503d35cf00c472f4717", size = 41396, upload-time = "2025-10-08T19:47:47.202Z" }, + { url = "https://files.pythonhosted.org/packages/48/f0/615c30622316496d2cbbc29f5985f7777d3ada70f23370608c1d3e081c1f/propcache-0.4.1-cp313-cp313t-win_amd64.whl", hash = "sha256:f7ee0e597f495cf415bcbd3da3caa3bd7e816b74d0d52b8145954c5e6fd3ff37", size = 44897, upload-time = "2025-10-08T19:47:48.336Z" }, + { url = "https://files.pythonhosted.org/packages/fd/ca/6002e46eccbe0e33dcd4069ef32f7f1c9e243736e07adca37ae8c4830ec3/propcache-0.4.1-cp313-cp313t-win_arm64.whl", hash = "sha256:929d7cbe1f01bb7baffb33dc14eb5691c95831450a26354cd210a8155170c93a", size = 39789, upload-time = "2025-10-08T19:47:49.876Z" }, + { url = "https://files.pythonhosted.org/packages/8e/5c/bca52d654a896f831b8256683457ceddd490ec18d9ec50e97dfd8fc726a8/propcache-0.4.1-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:3f7124c9d820ba5548d431afb4632301acf965db49e666aa21c305cbe8c6de12", size = 78152, upload-time = "2025-10-08T19:47:51.051Z" }, + { url = "https://files.pythonhosted.org/packages/65/9b/03b04e7d82a5f54fb16113d839f5ea1ede58a61e90edf515f6577c66fa8f/propcache-0.4.1-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:c0d4b719b7da33599dfe3b22d3db1ef789210a0597bc650b7cee9c77c2be8c5c", size = 44869, upload-time = "2025-10-08T19:47:52.594Z" }, + { url = "https://files.pythonhosted.org/packages/b2/fa/89a8ef0468d5833a23fff277b143d0573897cf75bd56670a6d28126c7d68/propcache-0.4.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:9f302f4783709a78240ebc311b793f123328716a60911d667e0c036bc5dcbded", size = 46596, upload-time = "2025-10-08T19:47:54.073Z" }, + { url = "https://files.pythonhosted.org/packages/86/bd/47816020d337f4a746edc42fe8d53669965138f39ee117414c7d7a340cfe/propcache-0.4.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c80ee5802e3fb9ea37938e7eecc307fb984837091d5fd262bb37238b1ae97641", size = 206981, upload-time = "2025-10-08T19:47:55.715Z" }, + { url = "https://files.pythonhosted.org/packages/df/f6/c5fa1357cc9748510ee55f37173eb31bfde6d94e98ccd9e6f033f2fc06e1/propcache-0.4.1-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ed5a841e8bb29a55fb8159ed526b26adc5bdd7e8bd7bf793ce647cb08656cdf4", size = 211490, upload-time = "2025-10-08T19:47:57.499Z" }, + { url = "https://files.pythonhosted.org/packages/80/1e/e5889652a7c4a3846683401a48f0f2e5083ce0ec1a8a5221d8058fbd1adf/propcache-0.4.1-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:55c72fd6ea2da4c318e74ffdf93c4fe4e926051133657459131a95c846d16d44", size = 215371, upload-time = "2025-10-08T19:47:59.317Z" }, + { url = "https://files.pythonhosted.org/packages/b2/f2/889ad4b2408f72fe1a4f6a19491177b30ea7bf1a0fd5f17050ca08cfc882/propcache-0.4.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8326e144341460402713f91df60ade3c999d601e7eb5ff8f6f7862d54de0610d", size = 201424, upload-time = "2025-10-08T19:48:00.67Z" }, + { url = "https://files.pythonhosted.org/packages/27/73/033d63069b57b0812c8bd19f311faebeceb6ba31b8f32b73432d12a0b826/propcache-0.4.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:060b16ae65bc098da7f6d25bf359f1f31f688384858204fe5d652979e0015e5b", size = 197566, upload-time = "2025-10-08T19:48:02.604Z" }, + { url = "https://files.pythonhosted.org/packages/dc/89/ce24f3dc182630b4e07aa6d15f0ff4b14ed4b9955fae95a0b54c58d66c05/propcache-0.4.1-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:89eb3fa9524f7bec9de6e83cf3faed9d79bffa560672c118a96a171a6f55831e", size = 193130, upload-time = "2025-10-08T19:48:04.499Z" }, + { url = "https://files.pythonhosted.org/packages/a9/24/ef0d5fd1a811fb5c609278d0209c9f10c35f20581fcc16f818da959fc5b4/propcache-0.4.1-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:dee69d7015dc235f526fe80a9c90d65eb0039103fe565776250881731f06349f", size = 202625, upload-time = "2025-10-08T19:48:06.213Z" }, + { url = "https://files.pythonhosted.org/packages/f5/02/98ec20ff5546f68d673df2f7a69e8c0d076b5abd05ca882dc7ee3a83653d/propcache-0.4.1-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:5558992a00dfd54ccbc64a32726a3357ec93825a418a401f5cc67df0ac5d9e49", size = 204209, upload-time = "2025-10-08T19:48:08.432Z" }, + { url = "https://files.pythonhosted.org/packages/a0/87/492694f76759b15f0467a2a93ab68d32859672b646aa8a04ce4864e7932d/propcache-0.4.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:c9b822a577f560fbd9554812526831712c1436d2c046cedee4c3796d3543b144", size = 197797, upload-time = "2025-10-08T19:48:09.968Z" }, + { url = "https://files.pythonhosted.org/packages/ee/36/66367de3575db1d2d3f3d177432bd14ee577a39d3f5d1b3d5df8afe3b6e2/propcache-0.4.1-cp314-cp314-win32.whl", hash = "sha256:ab4c29b49d560fe48b696cdcb127dd36e0bc2472548f3bf56cc5cb3da2b2984f", size = 38140, upload-time = "2025-10-08T19:48:11.232Z" }, + { url = "https://files.pythonhosted.org/packages/0c/2a/a758b47de253636e1b8aef181c0b4f4f204bf0dd964914fb2af90a95b49b/propcache-0.4.1-cp314-cp314-win_amd64.whl", hash = "sha256:5a103c3eb905fcea0ab98be99c3a9a5ab2de60228aa5aceedc614c0281cf6153", size = 41257, upload-time = "2025-10-08T19:48:12.707Z" }, + { url = "https://files.pythonhosted.org/packages/34/5e/63bd5896c3fec12edcbd6f12508d4890d23c265df28c74b175e1ef9f4f3b/propcache-0.4.1-cp314-cp314-win_arm64.whl", hash = "sha256:74c1fb26515153e482e00177a1ad654721bf9207da8a494a0c05e797ad27b992", size = 38097, upload-time = "2025-10-08T19:48:13.923Z" }, + { url = "https://files.pythonhosted.org/packages/99/85/9ff785d787ccf9bbb3f3106f79884a130951436f58392000231b4c737c80/propcache-0.4.1-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:824e908bce90fb2743bd6b59db36eb4f45cd350a39637c9f73b1c1ea66f5b75f", size = 81455, upload-time = "2025-10-08T19:48:15.16Z" }, + { url = "https://files.pythonhosted.org/packages/90/85/2431c10c8e7ddb1445c1f7c4b54d886e8ad20e3c6307e7218f05922cad67/propcache-0.4.1-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:c2b5e7db5328427c57c8e8831abda175421b709672f6cfc3d630c3b7e2146393", size = 46372, upload-time = "2025-10-08T19:48:16.424Z" }, + { url = "https://files.pythonhosted.org/packages/01/20/b0972d902472da9bcb683fa595099911f4d2e86e5683bcc45de60dd05dc3/propcache-0.4.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:6f6ff873ed40292cd4969ef5310179afd5db59fdf055897e282485043fc80ad0", size = 48411, upload-time = "2025-10-08T19:48:17.577Z" }, + { url = "https://files.pythonhosted.org/packages/e2/e3/7dc89f4f21e8f99bad3d5ddb3a3389afcf9da4ac69e3deb2dcdc96e74169/propcache-0.4.1-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:49a2dc67c154db2c1463013594c458881a069fcf98940e61a0569016a583020a", size = 275712, upload-time = "2025-10-08T19:48:18.901Z" }, + { url = "https://files.pythonhosted.org/packages/20/67/89800c8352489b21a8047c773067644e3897f02ecbbd610f4d46b7f08612/propcache-0.4.1-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:005f08e6a0529984491e37d8dbc3dd86f84bd78a8ceb5fa9a021f4c48d4984be", size = 273557, upload-time = "2025-10-08T19:48:20.762Z" }, + { url = "https://files.pythonhosted.org/packages/e2/a1/b52b055c766a54ce6d9c16d9aca0cad8059acd9637cdf8aa0222f4a026ef/propcache-0.4.1-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:5c3310452e0d31390da9035c348633b43d7e7feb2e37be252be6da45abd1abcc", size = 280015, upload-time = "2025-10-08T19:48:22.592Z" }, + { url = "https://files.pythonhosted.org/packages/48/c8/33cee30bd890672c63743049f3c9e4be087e6780906bfc3ec58528be59c1/propcache-0.4.1-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4c3c70630930447f9ef1caac7728c8ad1c56bc5015338b20fed0d08ea2480b3a", size = 262880, upload-time = "2025-10-08T19:48:23.947Z" }, + { url = "https://files.pythonhosted.org/packages/0c/b1/8f08a143b204b418285c88b83d00edbd61afbc2c6415ffafc8905da7038b/propcache-0.4.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:8e57061305815dfc910a3634dcf584f08168a8836e6999983569f51a8544cd89", size = 260938, upload-time = "2025-10-08T19:48:25.656Z" }, + { url = "https://files.pythonhosted.org/packages/cf/12/96e4664c82ca2f31e1c8dff86afb867348979eb78d3cb8546a680287a1e9/propcache-0.4.1-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:521a463429ef54143092c11a77e04056dd00636f72e8c45b70aaa3140d639726", size = 247641, upload-time = "2025-10-08T19:48:27.207Z" }, + { url = "https://files.pythonhosted.org/packages/18/ed/e7a9cfca28133386ba52278136d42209d3125db08d0a6395f0cba0c0285c/propcache-0.4.1-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:120c964da3fdc75e3731aa392527136d4ad35868cc556fd09bb6d09172d9a367", size = 262510, upload-time = "2025-10-08T19:48:28.65Z" }, + { url = "https://files.pythonhosted.org/packages/f5/76/16d8bf65e8845dd62b4e2b57444ab81f07f40caa5652b8969b87ddcf2ef6/propcache-0.4.1-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:d8f353eb14ee3441ee844ade4277d560cdd68288838673273b978e3d6d2c8f36", size = 263161, upload-time = "2025-10-08T19:48:30.133Z" }, + { url = "https://files.pythonhosted.org/packages/e7/70/c99e9edb5d91d5ad8a49fa3c1e8285ba64f1476782fed10ab251ff413ba1/propcache-0.4.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:ab2943be7c652f09638800905ee1bab2c544e537edb57d527997a24c13dc1455", size = 257393, upload-time = "2025-10-08T19:48:31.567Z" }, + { url = "https://files.pythonhosted.org/packages/08/02/87b25304249a35c0915d236575bc3574a323f60b47939a2262b77632a3ee/propcache-0.4.1-cp314-cp314t-win32.whl", hash = "sha256:05674a162469f31358c30bcaa8883cb7829fa3110bf9c0991fe27d7896c42d85", size = 42546, upload-time = "2025-10-08T19:48:32.872Z" }, + { url = "https://files.pythonhosted.org/packages/cb/ef/3c6ecf8b317aa982f309835e8f96987466123c6e596646d4e6a1dfcd080f/propcache-0.4.1-cp314-cp314t-win_amd64.whl", hash = "sha256:990f6b3e2a27d683cb7602ed6c86f15ee6b43b1194736f9baaeb93d0016633b1", size = 46259, upload-time = "2025-10-08T19:48:34.226Z" }, + { url = "https://files.pythonhosted.org/packages/c4/2d/346e946d4951f37eca1e4f55be0f0174c52cd70720f84029b02f296f4a38/propcache-0.4.1-cp314-cp314t-win_arm64.whl", hash = "sha256:ecef2343af4cc68e05131e45024ba34f6095821988a9d0a02aa7c73fcc448aa9", size = 40428, upload-time = "2025-10-08T19:48:35.441Z" }, + { url = "https://files.pythonhosted.org/packages/5b/5a/bc7b4a4ef808fa59a816c17b20c4bef6884daebbdf627ff2a161da67da19/propcache-0.4.1-py3-none-any.whl", hash = "sha256:af2a6052aeb6cf17d3e46ee169099044fd8224cbaf75c76a2ef596e8163e2237", size = 13305, upload-time = "2025-10-08T19:49:00.792Z" }, +] + +[[package]] +name = "protobuf" +version = "6.33.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/19/ff/64a6c8f420818bb873713988ca5492cba3a7946be57e027ac63495157d97/protobuf-6.33.0.tar.gz", hash = "sha256:140303d5c8d2037730c548f8c7b93b20bb1dc301be280c378b82b8894589c954", size = 443463, upload-time = "2025-10-15T20:39:52.159Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/7e/ee/52b3fa8feb6db4a833dfea4943e175ce645144532e8a90f72571ad85df4e/protobuf-6.33.0-cp310-abi3-win32.whl", hash = "sha256:d6101ded078042a8f17959eccd9236fb7a9ca20d3b0098bbcb91533a5680d035", size = 425593, upload-time = "2025-10-15T20:39:40.29Z" }, + { url = "https://files.pythonhosted.org/packages/7b/c6/7a465f1825872c55e0341ff4a80198743f73b69ce5d43ab18043699d1d81/protobuf-6.33.0-cp310-abi3-win_amd64.whl", hash = "sha256:9a031d10f703f03768f2743a1c403af050b6ae1f3480e9c140f39c45f81b13ee", size = 436882, upload-time = "2025-10-15T20:39:42.841Z" }, + { url = "https://files.pythonhosted.org/packages/e1/a9/b6eee662a6951b9c3640e8e452ab3e09f117d99fc10baa32d1581a0d4099/protobuf-6.33.0-cp39-abi3-macosx_10_9_universal2.whl", hash = "sha256:905b07a65f1a4b72412314082c7dbfae91a9e8b68a0cc1577515f8df58ecf455", size = 427521, upload-time = "2025-10-15T20:39:43.803Z" }, + { url = "https://files.pythonhosted.org/packages/10/35/16d31e0f92c6d2f0e77c2a3ba93185130ea13053dd16200a57434c882f2b/protobuf-6.33.0-cp39-abi3-manylinux2014_aarch64.whl", hash = "sha256:e0697ece353e6239b90ee43a9231318302ad8353c70e6e45499fa52396debf90", size = 324445, upload-time = "2025-10-15T20:39:44.932Z" }, + { url = "https://files.pythonhosted.org/packages/e6/eb/2a981a13e35cda8b75b5585aaffae2eb904f8f351bdd3870769692acbd8a/protobuf-6.33.0-cp39-abi3-manylinux2014_s390x.whl", hash = "sha256:e0a1715e4f27355afd9570f3ea369735afc853a6c3951a6afe1f80d8569ad298", size = 339159, upload-time = "2025-10-15T20:39:46.186Z" }, + { url = "https://files.pythonhosted.org/packages/21/51/0b1cbad62074439b867b4e04cc09b93f6699d78fd191bed2bbb44562e077/protobuf-6.33.0-cp39-abi3-manylinux2014_x86_64.whl", hash = "sha256:35be49fd3f4fefa4e6e2aacc35e8b837d6703c37a2168a55ac21e9b1bc7559ef", size = 323172, upload-time = "2025-10-15T20:39:47.465Z" }, + { url = "https://files.pythonhosted.org/packages/07/d1/0a28c21707807c6aacd5dc9c3704b2aa1effbf37adebd8caeaf68b17a636/protobuf-6.33.0-py3-none-any.whl", hash = "sha256:25c9e1963c6734448ea2d308cfa610e692b801304ba0908d7bfa564ac5132995", size = 170477, upload-time = "2025-10-15T20:39:51.311Z" }, +] + +[[package]] +name = "psutil" +version = "7.1.3" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/e1/88/bdd0a41e5857d5d703287598cbf08dad90aed56774ea52ae071bae9071b6/psutil-7.1.3.tar.gz", hash = "sha256:6c86281738d77335af7aec228328e944b30930899ea760ecf33a4dba66be5e74", size = 489059, upload-time = "2025-11-02T12:25:54.619Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/bd/93/0c49e776b8734fef56ec9c5c57f923922f2cf0497d62e0f419465f28f3d0/psutil-7.1.3-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:0005da714eee687b4b8decd3d6cc7c6db36215c9e74e5ad2264b90c3df7d92dc", size = 239751, upload-time = "2025-11-02T12:25:58.161Z" }, + { url = "https://files.pythonhosted.org/packages/6f/8d/b31e39c769e70780f007969815195a55c81a63efebdd4dbe9e7a113adb2f/psutil-7.1.3-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:19644c85dcb987e35eeeaefdc3915d059dac7bd1167cdcdbf27e0ce2df0c08c0", size = 240368, upload-time = "2025-11-02T12:26:00.491Z" }, + { url = "https://files.pythonhosted.org/packages/62/61/23fd4acc3c9eebbf6b6c78bcd89e5d020cfde4acf0a9233e9d4e3fa698b4/psutil-7.1.3-cp313-cp313t-manylinux2010_x86_64.manylinux_2_12_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:95ef04cf2e5ba0ab9eaafc4a11eaae91b44f4ef5541acd2ee91d9108d00d59a7", size = 287134, upload-time = "2025-11-02T12:26:02.613Z" }, + { url = "https://files.pythonhosted.org/packages/30/1c/f921a009ea9ceb51aa355cb0cc118f68d354db36eae18174bab63affb3e6/psutil-7.1.3-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1068c303be3a72f8e18e412c5b2a8f6d31750fb152f9cb106b54090296c9d251", size = 289904, upload-time = "2025-11-02T12:26:05.207Z" }, + { url = "https://files.pythonhosted.org/packages/a6/82/62d68066e13e46a5116df187d319d1724b3f437ddd0f958756fc052677f4/psutil-7.1.3-cp313-cp313t-win_amd64.whl", hash = "sha256:18349c5c24b06ac5612c0428ec2a0331c26443d259e2a0144a9b24b4395b58fa", size = 249642, upload-time = "2025-11-02T12:26:07.447Z" }, + { url = "https://files.pythonhosted.org/packages/df/ad/c1cd5fe965c14a0392112f68362cfceb5230819dbb5b1888950d18a11d9f/psutil-7.1.3-cp313-cp313t-win_arm64.whl", hash = "sha256:c525ffa774fe4496282fb0b1187725793de3e7c6b29e41562733cae9ada151ee", size = 245518, upload-time = "2025-11-02T12:26:09.719Z" }, + { url = "https://files.pythonhosted.org/packages/2e/bb/6670bded3e3236eb4287c7bcdc167e9fae6e1e9286e437f7111caed2f909/psutil-7.1.3-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:b403da1df4d6d43973dc004d19cee3b848e998ae3154cc8097d139b77156c353", size = 239843, upload-time = "2025-11-02T12:26:11.968Z" }, + { url = "https://files.pythonhosted.org/packages/b8/66/853d50e75a38c9a7370ddbeefabdd3d3116b9c31ef94dc92c6729bc36bec/psutil-7.1.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:ad81425efc5e75da3f39b3e636293360ad8d0b49bed7df824c79764fb4ba9b8b", size = 240369, upload-time = "2025-11-02T12:26:14.358Z" }, + { url = "https://files.pythonhosted.org/packages/41/bd/313aba97cb5bfb26916dc29cf0646cbe4dd6a89ca69e8c6edce654876d39/psutil-7.1.3-cp314-cp314t-manylinux2010_x86_64.manylinux_2_12_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8f33a3702e167783a9213db10ad29650ebf383946e91bc77f28a5eb083496bc9", size = 288210, upload-time = "2025-11-02T12:26:16.699Z" }, + { url = "https://files.pythonhosted.org/packages/c2/fa/76e3c06e760927a0cfb5705eb38164254de34e9bd86db656d4dbaa228b04/psutil-7.1.3-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fac9cd332c67f4422504297889da5ab7e05fd11e3c4392140f7370f4208ded1f", size = 291182, upload-time = "2025-11-02T12:26:18.848Z" }, + { url = "https://files.pythonhosted.org/packages/0f/1d/5774a91607035ee5078b8fd747686ebec28a962f178712de100d00b78a32/psutil-7.1.3-cp314-cp314t-win_amd64.whl", hash = "sha256:3792983e23b69843aea49c8f5b8f115572c5ab64c153bada5270086a2123c7e7", size = 250466, upload-time = "2025-11-02T12:26:21.183Z" }, + { url = "https://files.pythonhosted.org/packages/00/ca/e426584bacb43a5cb1ac91fae1937f478cd8fbe5e4ff96574e698a2c77cd/psutil-7.1.3-cp314-cp314t-win_arm64.whl", hash = "sha256:31d77fcedb7529f27bb3a0472bea9334349f9a04160e8e6e5020f22c59893264", size = 245756, upload-time = "2025-11-02T12:26:23.148Z" }, + { url = "https://files.pythonhosted.org/packages/ef/94/46b9154a800253e7ecff5aaacdf8ebf43db99de4a2dfa18575b02548654e/psutil-7.1.3-cp36-abi3-macosx_10_9_x86_64.whl", hash = "sha256:2bdbcd0e58ca14996a42adf3621a6244f1bb2e2e528886959c72cf1e326677ab", size = 238359, upload-time = "2025-11-02T12:26:25.284Z" }, + { url = "https://files.pythonhosted.org/packages/68/3a/9f93cff5c025029a36d9a92fef47220ab4692ee7f2be0fba9f92813d0cb8/psutil-7.1.3-cp36-abi3-macosx_11_0_arm64.whl", hash = "sha256:bc31fa00f1fbc3c3802141eede66f3a2d51d89716a194bf2cd6fc68310a19880", size = 239171, upload-time = "2025-11-02T12:26:27.23Z" }, + { url = "https://files.pythonhosted.org/packages/ce/b1/5f49af514f76431ba4eea935b8ad3725cdeb397e9245ab919dbc1d1dc20f/psutil-7.1.3-cp36-abi3-manylinux2010_x86_64.manylinux_2_12_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3bb428f9f05c1225a558f53e30ccbad9930b11c3fc206836242de1091d3e7dd3", size = 263261, upload-time = "2025-11-02T12:26:29.48Z" }, + { url = "https://files.pythonhosted.org/packages/e0/95/992c8816a74016eb095e73585d747e0a8ea21a061ed3689474fabb29a395/psutil-7.1.3-cp36-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:56d974e02ca2c8eb4812c3f76c30e28836fffc311d55d979f1465c1feeb2b68b", size = 264635, upload-time = "2025-11-02T12:26:31.74Z" }, + { url = "https://files.pythonhosted.org/packages/55/4c/c3ed1a622b6ae2fd3c945a366e64eb35247a31e4db16cf5095e269e8eb3c/psutil-7.1.3-cp37-abi3-win_amd64.whl", hash = "sha256:f39c2c19fe824b47484b96f9692932248a54c43799a84282cfe58d05a6449efd", size = 247633, upload-time = "2025-11-02T12:26:33.887Z" }, + { url = "https://files.pythonhosted.org/packages/c9/ad/33b2ccec09bf96c2b2ef3f9a6f66baac8253d7565d8839e024a6b905d45d/psutil-7.1.3-cp37-abi3-win_arm64.whl", hash = "sha256:bd0d69cee829226a761e92f28140bec9a5ee9d5b4fb4b0cc589068dbfff559b1", size = 244608, upload-time = "2025-11-02T12:26:36.136Z" }, +] + +[[package]] +name = "pyarrow" +version = "22.0.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/30/53/04a7fdc63e6056116c9ddc8b43bc28c12cdd181b85cbeadb79278475f3ae/pyarrow-22.0.0.tar.gz", hash = "sha256:3d600dc583260d845c7d8a6db540339dd883081925da2bd1c5cb808f720b3cd9", size = 1151151, upload-time = "2025-10-24T12:30:00.762Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/d9/9b/cb3f7e0a345353def531ca879053e9ef6b9f38ed91aebcf68b09ba54dec0/pyarrow-22.0.0-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:77718810bd3066158db1e95a63c160ad7ce08c6b0710bc656055033e39cdad88", size = 34223968, upload-time = "2025-10-24T10:03:31.21Z" }, + { url = "https://files.pythonhosted.org/packages/6c/41/3184b8192a120306270c5307f105b70320fdaa592c99843c5ef78aaefdcf/pyarrow-22.0.0-cp310-cp310-macosx_12_0_x86_64.whl", hash = "sha256:44d2d26cda26d18f7af7db71453b7b783788322d756e81730acb98f24eb90ace", size = 35942085, upload-time = "2025-10-24T10:03:38.146Z" }, + { url = "https://files.pythonhosted.org/packages/d9/3d/a1eab2f6f08001f9fb714b8ed5cfb045e2fe3e3e3c0c221f2c9ed1e6d67d/pyarrow-22.0.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:b9d71701ce97c95480fecb0039ec5bb889e75f110da72005743451339262f4ce", size = 44964613, upload-time = "2025-10-24T10:03:46.516Z" }, + { url = "https://files.pythonhosted.org/packages/46/46/a1d9c24baf21cfd9ce994ac820a24608decf2710521b29223d4334985127/pyarrow-22.0.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:710624ab925dc2b05a6229d47f6f0dac1c1155e6ed559be7109f684eba048a48", size = 47627059, upload-time = "2025-10-24T10:03:55.353Z" }, + { url = "https://files.pythonhosted.org/packages/3a/4c/f711acb13075c1391fd54bc17e078587672c575f8de2a6e62509af026dcf/pyarrow-22.0.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:f963ba8c3b0199f9d6b794c90ec77545e05eadc83973897a4523c9e8d84e9340", size = 47947043, upload-time = "2025-10-24T10:04:05.408Z" }, + { url = "https://files.pythonhosted.org/packages/4e/70/1f3180dd7c2eab35c2aca2b29ace6c519f827dcd4cfeb8e0dca41612cf7a/pyarrow-22.0.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:bd0d42297ace400d8febe55f13fdf46e86754842b860c978dfec16f081e5c653", size = 50206505, upload-time = "2025-10-24T10:04:15.786Z" }, + { url = "https://files.pythonhosted.org/packages/80/07/fea6578112c8c60ffde55883a571e4c4c6bc7049f119d6b09333b5cc6f73/pyarrow-22.0.0-cp310-cp310-win_amd64.whl", hash = "sha256:00626d9dc0f5ef3a75fe63fd68b9c7c8302d2b5bbc7f74ecaedba83447a24f84", size = 28101641, upload-time = "2025-10-24T10:04:22.57Z" }, + { url = "https://files.pythonhosted.org/packages/2e/b7/18f611a8cdc43417f9394a3ccd3eace2f32183c08b9eddc3d17681819f37/pyarrow-22.0.0-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:3e294c5eadfb93d78b0763e859a0c16d4051fc1c5231ae8956d61cb0b5666f5a", size = 34272022, upload-time = "2025-10-24T10:04:28.973Z" }, + { url = "https://files.pythonhosted.org/packages/26/5c/f259e2526c67eb4b9e511741b19870a02363a47a35edbebc55c3178db22d/pyarrow-22.0.0-cp311-cp311-macosx_12_0_x86_64.whl", hash = "sha256:69763ab2445f632d90b504a815a2a033f74332997052b721002298ed6de40f2e", size = 35995834, upload-time = "2025-10-24T10:04:35.467Z" }, + { url = "https://files.pythonhosted.org/packages/50/8d/281f0f9b9376d4b7f146913b26fac0aa2829cd1ee7e997f53a27411bbb92/pyarrow-22.0.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:b41f37cabfe2463232684de44bad753d6be08a7a072f6a83447eeaf0e4d2a215", size = 45030348, upload-time = "2025-10-24T10:04:43.366Z" }, + { url = "https://files.pythonhosted.org/packages/f5/e5/53c0a1c428f0976bf22f513d79c73000926cb00b9c138d8e02daf2102e18/pyarrow-22.0.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:35ad0f0378c9359b3f297299c3309778bb03b8612f987399a0333a560b43862d", size = 47699480, upload-time = "2025-10-24T10:04:51.486Z" }, + { url = "https://files.pythonhosted.org/packages/95/e1/9dbe4c465c3365959d183e6345d0a8d1dc5b02ca3f8db4760b3bc834cf25/pyarrow-22.0.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:8382ad21458075c2e66a82a29d650f963ce51c7708c7c0ff313a8c206c4fd5e8", size = 48011148, upload-time = "2025-10-24T10:04:59.585Z" }, + { url = "https://files.pythonhosted.org/packages/c5/b4/7caf5d21930061444c3cf4fa7535c82faf5263e22ce43af7c2759ceb5b8b/pyarrow-22.0.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:1a812a5b727bc09c3d7ea072c4eebf657c2f7066155506ba31ebf4792f88f016", size = 50276964, upload-time = "2025-10-24T10:05:08.175Z" }, + { url = "https://files.pythonhosted.org/packages/ae/f3/cec89bd99fa3abf826f14d4e53d3d11340ce6f6af4d14bdcd54cd83b6576/pyarrow-22.0.0-cp311-cp311-win_amd64.whl", hash = "sha256:ec5d40dd494882704fb876c16fa7261a69791e784ae34e6b5992e977bd2e238c", size = 28106517, upload-time = "2025-10-24T10:05:14.314Z" }, + { url = "https://files.pythonhosted.org/packages/af/63/ba23862d69652f85b615ca14ad14f3bcfc5bf1b99ef3f0cd04ff93fdad5a/pyarrow-22.0.0-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:bea79263d55c24a32b0d79c00a1c58bb2ee5f0757ed95656b01c0fb310c5af3d", size = 34211578, upload-time = "2025-10-24T10:05:21.583Z" }, + { url = "https://files.pythonhosted.org/packages/b1/d0/f9ad86fe809efd2bcc8be32032fa72e8b0d112b01ae56a053006376c5930/pyarrow-22.0.0-cp312-cp312-macosx_12_0_x86_64.whl", hash = "sha256:12fe549c9b10ac98c91cf791d2945e878875d95508e1a5d14091a7aaa66d9cf8", size = 35989906, upload-time = "2025-10-24T10:05:29.485Z" }, + { url = "https://files.pythonhosted.org/packages/b4/a8/f910afcb14630e64d673f15904ec27dd31f1e009b77033c365c84e8c1e1d/pyarrow-22.0.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:334f900ff08ce0423407af97e6c26ad5d4e3b0763645559ece6fbf3747d6a8f5", size = 45021677, upload-time = "2025-10-24T10:05:38.274Z" }, + { url = "https://files.pythonhosted.org/packages/13/95/aec81f781c75cd10554dc17a25849c720d54feafb6f7847690478dcf5ef8/pyarrow-22.0.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:c6c791b09c57ed76a18b03f2631753a4960eefbbca80f846da8baefc6491fcfe", size = 47726315, upload-time = "2025-10-24T10:05:47.314Z" }, + { url = "https://files.pythonhosted.org/packages/bb/d4/74ac9f7a54cfde12ee42734ea25d5a3c9a45db78f9def949307a92720d37/pyarrow-22.0.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:c3200cb41cdbc65156e5f8c908d739b0dfed57e890329413da2748d1a2cd1a4e", size = 47990906, upload-time = "2025-10-24T10:05:58.254Z" }, + { url = "https://files.pythonhosted.org/packages/2e/71/fedf2499bf7a95062eafc989ace56572f3343432570e1c54e6599d5b88da/pyarrow-22.0.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:ac93252226cf288753d8b46280f4edf3433bf9508b6977f8dd8526b521a1bbb9", size = 50306783, upload-time = "2025-10-24T10:06:08.08Z" }, + { url = "https://files.pythonhosted.org/packages/68/ed/b202abd5a5b78f519722f3d29063dda03c114711093c1995a33b8e2e0f4b/pyarrow-22.0.0-cp312-cp312-win_amd64.whl", hash = "sha256:44729980b6c50a5f2bfcc2668d36c569ce17f8b17bccaf470c4313dcbbf13c9d", size = 27972883, upload-time = "2025-10-24T10:06:14.204Z" }, + { url = "https://files.pythonhosted.org/packages/a6/d6/d0fac16a2963002fc22c8fa75180a838737203d558f0ed3b564c4a54eef5/pyarrow-22.0.0-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:e6e95176209257803a8b3d0394f21604e796dadb643d2f7ca21b66c9c0b30c9a", size = 34204629, upload-time = "2025-10-24T10:06:20.274Z" }, + { url = "https://files.pythonhosted.org/packages/c6/9c/1d6357347fbae062ad3f17082f9ebc29cc733321e892c0d2085f42a2212b/pyarrow-22.0.0-cp313-cp313-macosx_12_0_x86_64.whl", hash = "sha256:001ea83a58024818826a9e3f89bf9310a114f7e26dfe404a4c32686f97bd7901", size = 35985783, upload-time = "2025-10-24T10:06:27.301Z" }, + { url = "https://files.pythonhosted.org/packages/ff/c0/782344c2ce58afbea010150df07e3a2f5fdad299cd631697ae7bd3bac6e3/pyarrow-22.0.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:ce20fe000754f477c8a9125543f1936ea5b8867c5406757c224d745ed033e691", size = 45020999, upload-time = "2025-10-24T10:06:35.387Z" }, + { url = "https://files.pythonhosted.org/packages/1b/8b/5362443737a5307a7b67c1017c42cd104213189b4970bf607e05faf9c525/pyarrow-22.0.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:e0a15757fccb38c410947df156f9749ae4a3c89b2393741a50521f39a8cf202a", size = 47724601, upload-time = "2025-10-24T10:06:43.551Z" }, + { url = "https://files.pythonhosted.org/packages/69/4d/76e567a4fc2e190ee6072967cb4672b7d9249ac59ae65af2d7e3047afa3b/pyarrow-22.0.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:cedb9dd9358e4ea1d9bce3665ce0797f6adf97ff142c8e25b46ba9cdd508e9b6", size = 48001050, upload-time = "2025-10-24T10:06:52.284Z" }, + { url = "https://files.pythonhosted.org/packages/01/5e/5653f0535d2a1aef8223cee9d92944cb6bccfee5cf1cd3f462d7cb022790/pyarrow-22.0.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:252be4a05f9d9185bb8c18e83764ebcfea7185076c07a7a662253af3a8c07941", size = 50307877, upload-time = "2025-10-24T10:07:02.405Z" }, + { url = "https://files.pythonhosted.org/packages/2d/f8/1d0bd75bf9328a3b826e24a16e5517cd7f9fbf8d34a3184a4566ef5a7f29/pyarrow-22.0.0-cp313-cp313-win_amd64.whl", hash = "sha256:a4893d31e5ef780b6edcaf63122df0f8d321088bb0dee4c8c06eccb1ca28d145", size = 27977099, upload-time = "2025-10-24T10:08:07.259Z" }, + { url = "https://files.pythonhosted.org/packages/90/81/db56870c997805bf2b0f6eeeb2d68458bf4654652dccdcf1bf7a42d80903/pyarrow-22.0.0-cp313-cp313t-macosx_12_0_arm64.whl", hash = "sha256:f7fe3dbe871294ba70d789be16b6e7e52b418311e166e0e3cba9522f0f437fb1", size = 34336685, upload-time = "2025-10-24T10:07:11.47Z" }, + { url = "https://files.pythonhosted.org/packages/1c/98/0727947f199aba8a120f47dfc229eeb05df15bcd7a6f1b669e9f882afc58/pyarrow-22.0.0-cp313-cp313t-macosx_12_0_x86_64.whl", hash = "sha256:ba95112d15fd4f1105fb2402c4eab9068f0554435e9b7085924bcfaac2cc306f", size = 36032158, upload-time = "2025-10-24T10:07:18.626Z" }, + { url = "https://files.pythonhosted.org/packages/96/b4/9babdef9c01720a0785945c7cf550e4acd0ebcd7bdd2e6f0aa7981fa85e2/pyarrow-22.0.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:c064e28361c05d72eed8e744c9605cbd6d2bb7481a511c74071fd9b24bc65d7d", size = 44892060, upload-time = "2025-10-24T10:07:26.002Z" }, + { url = "https://files.pythonhosted.org/packages/f8/ca/2f8804edd6279f78a37062d813de3f16f29183874447ef6d1aadbb4efa0f/pyarrow-22.0.0-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:6f9762274496c244d951c819348afbcf212714902742225f649cf02823a6a10f", size = 47504395, upload-time = "2025-10-24T10:07:34.09Z" }, + { url = "https://files.pythonhosted.org/packages/b9/f0/77aa5198fd3943682b2e4faaf179a674f0edea0d55d326d83cb2277d9363/pyarrow-22.0.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:a9d9ffdc2ab696f6b15b4d1f7cec6658e1d788124418cb30030afbae31c64746", size = 48066216, upload-time = "2025-10-24T10:07:43.528Z" }, + { url = "https://files.pythonhosted.org/packages/79/87/a1937b6e78b2aff18b706d738c9e46ade5bfcf11b294e39c87706a0089ac/pyarrow-22.0.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:ec1a15968a9d80da01e1d30349b2b0d7cc91e96588ee324ce1b5228175043e95", size = 50288552, upload-time = "2025-10-24T10:07:53.519Z" }, + { url = "https://files.pythonhosted.org/packages/60/ae/b5a5811e11f25788ccfdaa8f26b6791c9807119dffcf80514505527c384c/pyarrow-22.0.0-cp313-cp313t-win_amd64.whl", hash = "sha256:bba208d9c7decf9961998edf5c65e3ea4355d5818dd6cd0f6809bec1afb951cc", size = 28262504, upload-time = "2025-10-24T10:08:00.932Z" }, + { url = "https://files.pythonhosted.org/packages/bd/b0/0fa4d28a8edb42b0a7144edd20befd04173ac79819547216f8a9f36f9e50/pyarrow-22.0.0-cp314-cp314-macosx_12_0_arm64.whl", hash = "sha256:9bddc2cade6561f6820d4cd73f99a0243532ad506bc510a75a5a65a522b2d74d", size = 34224062, upload-time = "2025-10-24T10:08:14.101Z" }, + { url = "https://files.pythonhosted.org/packages/0f/a8/7a719076b3c1be0acef56a07220c586f25cd24de0e3f3102b438d18ae5df/pyarrow-22.0.0-cp314-cp314-macosx_12_0_x86_64.whl", hash = "sha256:e70ff90c64419709d38c8932ea9fe1cc98415c4f87ea8da81719e43f02534bc9", size = 35990057, upload-time = "2025-10-24T10:08:21.842Z" }, + { url = "https://files.pythonhosted.org/packages/89/3c/359ed54c93b47fb6fe30ed16cdf50e3f0e8b9ccfb11b86218c3619ae50a8/pyarrow-22.0.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:92843c305330aa94a36e706c16209cd4df274693e777ca47112617db7d0ef3d7", size = 45068002, upload-time = "2025-10-24T10:08:29.034Z" }, + { url = "https://files.pythonhosted.org/packages/55/fc/4945896cc8638536ee787a3bd6ce7cec8ec9acf452d78ec39ab328efa0a1/pyarrow-22.0.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:6dda1ddac033d27421c20d7a7943eec60be44e0db4e079f33cc5af3b8280ccde", size = 47737765, upload-time = "2025-10-24T10:08:38.559Z" }, + { url = "https://files.pythonhosted.org/packages/cd/5e/7cb7edeb2abfaa1f79b5d5eb89432356155c8426f75d3753cbcb9592c0fd/pyarrow-22.0.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:84378110dd9a6c06323b41b56e129c504d157d1a983ce8f5443761eb5256bafc", size = 48048139, upload-time = "2025-10-24T10:08:46.784Z" }, + { url = "https://files.pythonhosted.org/packages/88/c6/546baa7c48185f5e9d6e59277c4b19f30f48c94d9dd938c2a80d4d6b067c/pyarrow-22.0.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:854794239111d2b88b40b6ef92aa478024d1e5074f364033e73e21e3f76b25e0", size = 50314244, upload-time = "2025-10-24T10:08:55.771Z" }, + { url = "https://files.pythonhosted.org/packages/3c/79/755ff2d145aafec8d347bf18f95e4e81c00127f06d080135dfc86aea417c/pyarrow-22.0.0-cp314-cp314-win_amd64.whl", hash = "sha256:b883fe6fd85adad7932b3271c38ac289c65b7337c2c132e9569f9d3940620730", size = 28757501, upload-time = "2025-10-24T10:09:59.891Z" }, + { url = "https://files.pythonhosted.org/packages/0e/d2/237d75ac28ced3147912954e3c1a174df43a95f4f88e467809118a8165e0/pyarrow-22.0.0-cp314-cp314t-macosx_12_0_arm64.whl", hash = "sha256:7a820d8ae11facf32585507c11f04e3f38343c1e784c9b5a8b1da5c930547fe2", size = 34355506, upload-time = "2025-10-24T10:09:02.953Z" }, + { url = "https://files.pythonhosted.org/packages/1e/2c/733dfffe6d3069740f98e57ff81007809067d68626c5faef293434d11bd6/pyarrow-22.0.0-cp314-cp314t-macosx_12_0_x86_64.whl", hash = "sha256:c6ec3675d98915bf1ec8b3c7986422682f7232ea76cad276f4c8abd5b7319b70", size = 36047312, upload-time = "2025-10-24T10:09:10.334Z" }, + { url = "https://files.pythonhosted.org/packages/7c/2b/29d6e3782dc1f299727462c1543af357a0f2c1d3c160ce199950d9ca51eb/pyarrow-22.0.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:3e739edd001b04f654b166204fc7a9de896cf6007eaff33409ee9e50ceaff754", size = 45081609, upload-time = "2025-10-24T10:09:18.61Z" }, + { url = "https://files.pythonhosted.org/packages/8d/42/aa9355ecc05997915af1b7b947a7f66c02dcaa927f3203b87871c114ba10/pyarrow-22.0.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:7388ac685cab5b279a41dfe0a6ccd99e4dbf322edfb63e02fc0443bf24134e91", size = 47703663, upload-time = "2025-10-24T10:09:27.369Z" }, + { url = "https://files.pythonhosted.org/packages/ee/62/45abedde480168e83a1de005b7b7043fd553321c1e8c5a9a114425f64842/pyarrow-22.0.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:f633074f36dbc33d5c05b5dc75371e5660f1dbf9c8b1d95669def05e5425989c", size = 48066543, upload-time = "2025-10-24T10:09:34.908Z" }, + { url = "https://files.pythonhosted.org/packages/84/e9/7878940a5b072e4f3bf998770acafeae13b267f9893af5f6d4ab3904b67e/pyarrow-22.0.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:4c19236ae2402a8663a2c8f21f1870a03cc57f0bef7e4b6eb3238cc82944de80", size = 50288838, upload-time = "2025-10-24T10:09:44.394Z" }, + { url = "https://files.pythonhosted.org/packages/7b/03/f335d6c52b4a4761bcc83499789a1e2e16d9d201a58c327a9b5cc9a41bd9/pyarrow-22.0.0-cp314-cp314t-win_amd64.whl", hash = "sha256:0c34fe18094686194f204a3b1787a27456897d8a2d62caf84b61e8dfbc0252ae", size = 29185594, upload-time = "2025-10-24T10:09:53.111Z" }, +] + +[[package]] +name = "pydantic" +version = "2.11.10" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "annotated-types" }, + { name = "pydantic-core" }, + { name = "typing-extensions" }, + { name = "typing-inspection" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/ae/54/ecab642b3bed45f7d5f59b38443dcb36ef50f85af192e6ece103dbfe9587/pydantic-2.11.10.tar.gz", hash = "sha256:dc280f0982fbda6c38fada4e476dc0a4f3aeaf9c6ad4c28df68a666ec3c61423", size = 788494, upload-time = "2025-10-04T10:40:41.338Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/bd/1f/73c53fcbfb0b5a78f91176df41945ca466e71e9d9d836e5c522abda39ee7/pydantic-2.11.10-py3-none-any.whl", hash = "sha256:802a655709d49bd004c31e865ef37da30b540786a46bfce02333e0e24b5fe29a", size = 444823, upload-time = "2025-10-04T10:40:39.055Z" }, +] + +[[package]] +name = "pydantic-core" +version = "2.33.2" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/ad/88/5f2260bdfae97aabf98f1778d43f69574390ad787afb646292a638c923d4/pydantic_core-2.33.2.tar.gz", hash = "sha256:7cb8bc3605c29176e1b105350d2e6474142d7c1bd1d9327c4a9bdb46bf827acc", size = 435195, upload-time = "2025-04-23T18:33:52.104Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/e5/92/b31726561b5dae176c2d2c2dc43a9c5bfba5d32f96f8b4c0a600dd492447/pydantic_core-2.33.2-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:2b3d326aaef0c0399d9afffeb6367d5e26ddc24d351dbc9c636840ac355dc5d8", size = 2028817, upload-time = "2025-04-23T18:30:43.919Z" }, + { url = "https://files.pythonhosted.org/packages/a3/44/3f0b95fafdaca04a483c4e685fe437c6891001bf3ce8b2fded82b9ea3aa1/pydantic_core-2.33.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:0e5b2671f05ba48b94cb90ce55d8bdcaaedb8ba00cc5359f6810fc918713983d", size = 1861357, upload-time = "2025-04-23T18:30:46.372Z" }, + { url = "https://files.pythonhosted.org/packages/30/97/e8f13b55766234caae05372826e8e4b3b96e7b248be3157f53237682e43c/pydantic_core-2.33.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0069c9acc3f3981b9ff4cdfaf088e98d83440a4c7ea1bc07460af3d4dc22e72d", size = 1898011, upload-time = "2025-04-23T18:30:47.591Z" }, + { url = "https://files.pythonhosted.org/packages/9b/a3/99c48cf7bafc991cc3ee66fd544c0aae8dc907b752f1dad2d79b1b5a471f/pydantic_core-2.33.2-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:d53b22f2032c42eaaf025f7c40c2e3b94568ae077a606f006d206a463bc69572", size = 1982730, upload-time = "2025-04-23T18:30:49.328Z" }, + { url = "https://files.pythonhosted.org/packages/de/8e/a5b882ec4307010a840fb8b58bd9bf65d1840c92eae7534c7441709bf54b/pydantic_core-2.33.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:0405262705a123b7ce9f0b92f123334d67b70fd1f20a9372b907ce1080c7ba02", size = 2136178, upload-time = "2025-04-23T18:30:50.907Z" }, + { url = "https://files.pythonhosted.org/packages/e4/bb/71e35fc3ed05af6834e890edb75968e2802fe98778971ab5cba20a162315/pydantic_core-2.33.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:4b25d91e288e2c4e0662b8038a28c6a07eaac3e196cfc4ff69de4ea3db992a1b", size = 2736462, upload-time = "2025-04-23T18:30:52.083Z" }, + { url = "https://files.pythonhosted.org/packages/31/0d/c8f7593e6bc7066289bbc366f2235701dcbebcd1ff0ef8e64f6f239fb47d/pydantic_core-2.33.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6bdfe4b3789761f3bcb4b1ddf33355a71079858958e3a552f16d5af19768fef2", size = 2005652, upload-time = "2025-04-23T18:30:53.389Z" }, + { url = "https://files.pythonhosted.org/packages/d2/7a/996d8bd75f3eda405e3dd219ff5ff0a283cd8e34add39d8ef9157e722867/pydantic_core-2.33.2-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:efec8db3266b76ef9607c2c4c419bdb06bf335ae433b80816089ea7585816f6a", size = 2113306, upload-time = "2025-04-23T18:30:54.661Z" }, + { url = "https://files.pythonhosted.org/packages/ff/84/daf2a6fb2db40ffda6578a7e8c5a6e9c8affb251a05c233ae37098118788/pydantic_core-2.33.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:031c57d67ca86902726e0fae2214ce6770bbe2f710dc33063187a68744a5ecac", size = 2073720, upload-time = "2025-04-23T18:30:56.11Z" }, + { url = "https://files.pythonhosted.org/packages/77/fb/2258da019f4825128445ae79456a5499c032b55849dbd5bed78c95ccf163/pydantic_core-2.33.2-cp310-cp310-musllinux_1_1_armv7l.whl", hash = "sha256:f8de619080e944347f5f20de29a975c2d815d9ddd8be9b9b7268e2e3ef68605a", size = 2244915, upload-time = "2025-04-23T18:30:57.501Z" }, + { url = "https://files.pythonhosted.org/packages/d8/7a/925ff73756031289468326e355b6fa8316960d0d65f8b5d6b3a3e7866de7/pydantic_core-2.33.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:73662edf539e72a9440129f231ed3757faab89630d291b784ca99237fb94db2b", size = 2241884, upload-time = "2025-04-23T18:30:58.867Z" }, + { url = "https://files.pythonhosted.org/packages/0b/b0/249ee6d2646f1cdadcb813805fe76265745c4010cf20a8eba7b0e639d9b2/pydantic_core-2.33.2-cp310-cp310-win32.whl", hash = "sha256:0a39979dcbb70998b0e505fb1556a1d550a0781463ce84ebf915ba293ccb7e22", size = 1910496, upload-time = "2025-04-23T18:31:00.078Z" }, + { url = "https://files.pythonhosted.org/packages/66/ff/172ba8f12a42d4b552917aa65d1f2328990d3ccfc01d5b7c943ec084299f/pydantic_core-2.33.2-cp310-cp310-win_amd64.whl", hash = "sha256:b0379a2b24882fef529ec3b4987cb5d003b9cda32256024e6fe1586ac45fc640", size = 1955019, upload-time = "2025-04-23T18:31:01.335Z" }, + { url = "https://files.pythonhosted.org/packages/3f/8d/71db63483d518cbbf290261a1fc2839d17ff89fce7089e08cad07ccfce67/pydantic_core-2.33.2-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:4c5b0a576fb381edd6d27f0a85915c6daf2f8138dc5c267a57c08a62900758c7", size = 2028584, upload-time = "2025-04-23T18:31:03.106Z" }, + { url = "https://files.pythonhosted.org/packages/24/2f/3cfa7244ae292dd850989f328722d2aef313f74ffc471184dc509e1e4e5a/pydantic_core-2.33.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:e799c050df38a639db758c617ec771fd8fb7a5f8eaaa4b27b101f266b216a246", size = 1855071, upload-time = "2025-04-23T18:31:04.621Z" }, + { url = "https://files.pythonhosted.org/packages/b3/d3/4ae42d33f5e3f50dd467761304be2fa0a9417fbf09735bc2cce003480f2a/pydantic_core-2.33.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:dc46a01bf8d62f227d5ecee74178ffc448ff4e5197c756331f71efcc66dc980f", size = 1897823, upload-time = "2025-04-23T18:31:06.377Z" }, + { url = "https://files.pythonhosted.org/packages/f4/f3/aa5976e8352b7695ff808599794b1fba2a9ae2ee954a3426855935799488/pydantic_core-2.33.2-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:a144d4f717285c6d9234a66778059f33a89096dfb9b39117663fd8413d582dcc", size = 1983792, upload-time = "2025-04-23T18:31:07.93Z" }, + { url = "https://files.pythonhosted.org/packages/d5/7a/cda9b5a23c552037717f2b2a5257e9b2bfe45e687386df9591eff7b46d28/pydantic_core-2.33.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:73cf6373c21bc80b2e0dc88444f41ae60b2f070ed02095754eb5a01df12256de", size = 2136338, upload-time = "2025-04-23T18:31:09.283Z" }, + { url = "https://files.pythonhosted.org/packages/2b/9f/b8f9ec8dd1417eb9da784e91e1667d58a2a4a7b7b34cf4af765ef663a7e5/pydantic_core-2.33.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3dc625f4aa79713512d1976fe9f0bc99f706a9dee21dfd1810b4bbbf228d0e8a", size = 2730998, upload-time = "2025-04-23T18:31:11.7Z" }, + { url = "https://files.pythonhosted.org/packages/47/bc/cd720e078576bdb8255d5032c5d63ee5c0bf4b7173dd955185a1d658c456/pydantic_core-2.33.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:881b21b5549499972441da4758d662aeea93f1923f953e9cbaff14b8b9565aef", size = 2003200, upload-time = "2025-04-23T18:31:13.536Z" }, + { url = "https://files.pythonhosted.org/packages/ca/22/3602b895ee2cd29d11a2b349372446ae9727c32e78a94b3d588a40fdf187/pydantic_core-2.33.2-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:bdc25f3681f7b78572699569514036afe3c243bc3059d3942624e936ec93450e", size = 2113890, upload-time = "2025-04-23T18:31:15.011Z" }, + { url = "https://files.pythonhosted.org/packages/ff/e6/e3c5908c03cf00d629eb38393a98fccc38ee0ce8ecce32f69fc7d7b558a7/pydantic_core-2.33.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:fe5b32187cbc0c862ee201ad66c30cf218e5ed468ec8dc1cf49dec66e160cc4d", size = 2073359, upload-time = "2025-04-23T18:31:16.393Z" }, + { url = "https://files.pythonhosted.org/packages/12/e7/6a36a07c59ebefc8777d1ffdaf5ae71b06b21952582e4b07eba88a421c79/pydantic_core-2.33.2-cp311-cp311-musllinux_1_1_armv7l.whl", hash = "sha256:bc7aee6f634a6f4a95676fcb5d6559a2c2a390330098dba5e5a5f28a2e4ada30", size = 2245883, upload-time = "2025-04-23T18:31:17.892Z" }, + { url = "https://files.pythonhosted.org/packages/16/3f/59b3187aaa6cc0c1e6616e8045b284de2b6a87b027cce2ffcea073adf1d2/pydantic_core-2.33.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:235f45e5dbcccf6bd99f9f472858849f73d11120d76ea8707115415f8e5ebebf", size = 2241074, upload-time = "2025-04-23T18:31:19.205Z" }, + { url = "https://files.pythonhosted.org/packages/e0/ed/55532bb88f674d5d8f67ab121a2a13c385df382de2a1677f30ad385f7438/pydantic_core-2.33.2-cp311-cp311-win32.whl", hash = "sha256:6368900c2d3ef09b69cb0b913f9f8263b03786e5b2a387706c5afb66800efd51", size = 1910538, upload-time = "2025-04-23T18:31:20.541Z" }, + { url = "https://files.pythonhosted.org/packages/fe/1b/25b7cccd4519c0b23c2dd636ad39d381abf113085ce4f7bec2b0dc755eb1/pydantic_core-2.33.2-cp311-cp311-win_amd64.whl", hash = "sha256:1e063337ef9e9820c77acc768546325ebe04ee38b08703244c1309cccc4f1bab", size = 1952909, upload-time = "2025-04-23T18:31:22.371Z" }, + { url = "https://files.pythonhosted.org/packages/49/a9/d809358e49126438055884c4366a1f6227f0f84f635a9014e2deb9b9de54/pydantic_core-2.33.2-cp311-cp311-win_arm64.whl", hash = "sha256:6b99022f1d19bc32a4c2a0d544fc9a76e3be90f0b3f4af413f87d38749300e65", size = 1897786, upload-time = "2025-04-23T18:31:24.161Z" }, + { url = "https://files.pythonhosted.org/packages/18/8a/2b41c97f554ec8c71f2a8a5f85cb56a8b0956addfe8b0efb5b3d77e8bdc3/pydantic_core-2.33.2-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:a7ec89dc587667f22b6a0b6579c249fca9026ce7c333fc142ba42411fa243cdc", size = 2009000, upload-time = "2025-04-23T18:31:25.863Z" }, + { url = "https://files.pythonhosted.org/packages/a1/02/6224312aacb3c8ecbaa959897af57181fb6cf3a3d7917fd44d0f2917e6f2/pydantic_core-2.33.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:3c6db6e52c6d70aa0d00d45cdb9b40f0433b96380071ea80b09277dba021ddf7", size = 1847996, upload-time = "2025-04-23T18:31:27.341Z" }, + { url = "https://files.pythonhosted.org/packages/d6/46/6dcdf084a523dbe0a0be59d054734b86a981726f221f4562aed313dbcb49/pydantic_core-2.33.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4e61206137cbc65e6d5256e1166f88331d3b6238e082d9f74613b9b765fb9025", size = 1880957, upload-time = "2025-04-23T18:31:28.956Z" }, + { url = "https://files.pythonhosted.org/packages/ec/6b/1ec2c03837ac00886ba8160ce041ce4e325b41d06a034adbef11339ae422/pydantic_core-2.33.2-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:eb8c529b2819c37140eb51b914153063d27ed88e3bdc31b71198a198e921e011", size = 1964199, upload-time = "2025-04-23T18:31:31.025Z" }, + { url = "https://files.pythonhosted.org/packages/2d/1d/6bf34d6adb9debd9136bd197ca72642203ce9aaaa85cfcbfcf20f9696e83/pydantic_core-2.33.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:c52b02ad8b4e2cf14ca7b3d918f3eb0ee91e63b3167c32591e57c4317e134f8f", size = 2120296, upload-time = "2025-04-23T18:31:32.514Z" }, + { url = "https://files.pythonhosted.org/packages/e0/94/2bd0aaf5a591e974b32a9f7123f16637776c304471a0ab33cf263cf5591a/pydantic_core-2.33.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:96081f1605125ba0855dfda83f6f3df5ec90c61195421ba72223de35ccfb2f88", size = 2676109, upload-time = "2025-04-23T18:31:33.958Z" }, + { url = "https://files.pythonhosted.org/packages/f9/41/4b043778cf9c4285d59742281a769eac371b9e47e35f98ad321349cc5d61/pydantic_core-2.33.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8f57a69461af2a5fa6e6bbd7a5f60d3b7e6cebb687f55106933188e79ad155c1", size = 2002028, upload-time = "2025-04-23T18:31:39.095Z" }, + { url = "https://files.pythonhosted.org/packages/cb/d5/7bb781bf2748ce3d03af04d5c969fa1308880e1dca35a9bd94e1a96a922e/pydantic_core-2.33.2-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:572c7e6c8bb4774d2ac88929e3d1f12bc45714ae5ee6d9a788a9fb35e60bb04b", size = 2100044, upload-time = "2025-04-23T18:31:41.034Z" }, + { url = "https://files.pythonhosted.org/packages/fe/36/def5e53e1eb0ad896785702a5bbfd25eed546cdcf4087ad285021a90ed53/pydantic_core-2.33.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:db4b41f9bd95fbe5acd76d89920336ba96f03e149097365afe1cb092fceb89a1", size = 2058881, upload-time = "2025-04-23T18:31:42.757Z" }, + { url = "https://files.pythonhosted.org/packages/01/6c/57f8d70b2ee57fc3dc8b9610315949837fa8c11d86927b9bb044f8705419/pydantic_core-2.33.2-cp312-cp312-musllinux_1_1_armv7l.whl", hash = "sha256:fa854f5cf7e33842a892e5c73f45327760bc7bc516339fda888c75ae60edaeb6", size = 2227034, upload-time = "2025-04-23T18:31:44.304Z" }, + { url = "https://files.pythonhosted.org/packages/27/b9/9c17f0396a82b3d5cbea4c24d742083422639e7bb1d5bf600e12cb176a13/pydantic_core-2.33.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:5f483cfb75ff703095c59e365360cb73e00185e01aaea067cd19acffd2ab20ea", size = 2234187, upload-time = "2025-04-23T18:31:45.891Z" }, + { url = "https://files.pythonhosted.org/packages/b0/6a/adf5734ffd52bf86d865093ad70b2ce543415e0e356f6cacabbc0d9ad910/pydantic_core-2.33.2-cp312-cp312-win32.whl", hash = "sha256:9cb1da0f5a471435a7bc7e439b8a728e8b61e59784b2af70d7c169f8dd8ae290", size = 1892628, upload-time = "2025-04-23T18:31:47.819Z" }, + { url = "https://files.pythonhosted.org/packages/43/e4/5479fecb3606c1368d496a825d8411e126133c41224c1e7238be58b87d7e/pydantic_core-2.33.2-cp312-cp312-win_amd64.whl", hash = "sha256:f941635f2a3d96b2973e867144fde513665c87f13fe0e193c158ac51bfaaa7b2", size = 1955866, upload-time = "2025-04-23T18:31:49.635Z" }, + { url = "https://files.pythonhosted.org/packages/0d/24/8b11e8b3e2be9dd82df4b11408a67c61bb4dc4f8e11b5b0fc888b38118b5/pydantic_core-2.33.2-cp312-cp312-win_arm64.whl", hash = "sha256:cca3868ddfaccfbc4bfb1d608e2ccaaebe0ae628e1416aeb9c4d88c001bb45ab", size = 1888894, upload-time = "2025-04-23T18:31:51.609Z" }, + { url = "https://files.pythonhosted.org/packages/46/8c/99040727b41f56616573a28771b1bfa08a3d3fe74d3d513f01251f79f172/pydantic_core-2.33.2-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:1082dd3e2d7109ad8b7da48e1d4710c8d06c253cbc4a27c1cff4fbcaa97a9e3f", size = 2015688, upload-time = "2025-04-23T18:31:53.175Z" }, + { url = "https://files.pythonhosted.org/packages/3a/cc/5999d1eb705a6cefc31f0b4a90e9f7fc400539b1a1030529700cc1b51838/pydantic_core-2.33.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:f517ca031dfc037a9c07e748cefd8d96235088b83b4f4ba8939105d20fa1dcd6", size = 1844808, upload-time = "2025-04-23T18:31:54.79Z" }, + { url = "https://files.pythonhosted.org/packages/6f/5e/a0a7b8885c98889a18b6e376f344da1ef323d270b44edf8174d6bce4d622/pydantic_core-2.33.2-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0a9f2c9dd19656823cb8250b0724ee9c60a82f3cdf68a080979d13092a3b0fef", size = 1885580, upload-time = "2025-04-23T18:31:57.393Z" }, + { url = "https://files.pythonhosted.org/packages/3b/2a/953581f343c7d11a304581156618c3f592435523dd9d79865903272c256a/pydantic_core-2.33.2-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:2b0a451c263b01acebe51895bfb0e1cc842a5c666efe06cdf13846c7418caa9a", size = 1973859, upload-time = "2025-04-23T18:31:59.065Z" }, + { url = "https://files.pythonhosted.org/packages/e6/55/f1a813904771c03a3f97f676c62cca0c0a4138654107c1b61f19c644868b/pydantic_core-2.33.2-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1ea40a64d23faa25e62a70ad163571c0b342b8bf66d5fa612ac0dec4f069d916", size = 2120810, upload-time = "2025-04-23T18:32:00.78Z" }, + { url = "https://files.pythonhosted.org/packages/aa/c3/053389835a996e18853ba107a63caae0b9deb4a276c6b472931ea9ae6e48/pydantic_core-2.33.2-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:0fb2d542b4d66f9470e8065c5469ec676978d625a8b7a363f07d9a501a9cb36a", size = 2676498, upload-time = "2025-04-23T18:32:02.418Z" }, + { url = "https://files.pythonhosted.org/packages/eb/3c/f4abd740877a35abade05e437245b192f9d0ffb48bbbbd708df33d3cda37/pydantic_core-2.33.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9fdac5d6ffa1b5a83bca06ffe7583f5576555e6c8b3a91fbd25ea7780f825f7d", size = 2000611, upload-time = "2025-04-23T18:32:04.152Z" }, + { url = "https://files.pythonhosted.org/packages/59/a7/63ef2fed1837d1121a894d0ce88439fe3e3b3e48c7543b2a4479eb99c2bd/pydantic_core-2.33.2-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:04a1a413977ab517154eebb2d326da71638271477d6ad87a769102f7c2488c56", size = 2107924, upload-time = "2025-04-23T18:32:06.129Z" }, + { url = "https://files.pythonhosted.org/packages/04/8f/2551964ef045669801675f1cfc3b0d74147f4901c3ffa42be2ddb1f0efc4/pydantic_core-2.33.2-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:c8e7af2f4e0194c22b5b37205bfb293d166a7344a5b0d0eaccebc376546d77d5", size = 2063196, upload-time = "2025-04-23T18:32:08.178Z" }, + { url = "https://files.pythonhosted.org/packages/26/bd/d9602777e77fc6dbb0c7db9ad356e9a985825547dce5ad1d30ee04903918/pydantic_core-2.33.2-cp313-cp313-musllinux_1_1_armv7l.whl", hash = "sha256:5c92edd15cd58b3c2d34873597a1e20f13094f59cf88068adb18947df5455b4e", size = 2236389, upload-time = "2025-04-23T18:32:10.242Z" }, + { url = "https://files.pythonhosted.org/packages/42/db/0e950daa7e2230423ab342ae918a794964b053bec24ba8af013fc7c94846/pydantic_core-2.33.2-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:65132b7b4a1c0beded5e057324b7e16e10910c106d43675d9bd87d4f38dde162", size = 2239223, upload-time = "2025-04-23T18:32:12.382Z" }, + { url = "https://files.pythonhosted.org/packages/58/4d/4f937099c545a8a17eb52cb67fe0447fd9a373b348ccfa9a87f141eeb00f/pydantic_core-2.33.2-cp313-cp313-win32.whl", hash = "sha256:52fb90784e0a242bb96ec53f42196a17278855b0f31ac7c3cc6f5c1ec4811849", size = 1900473, upload-time = "2025-04-23T18:32:14.034Z" }, + { url = "https://files.pythonhosted.org/packages/a0/75/4a0a9bac998d78d889def5e4ef2b065acba8cae8c93696906c3a91f310ca/pydantic_core-2.33.2-cp313-cp313-win_amd64.whl", hash = "sha256:c083a3bdd5a93dfe480f1125926afcdbf2917ae714bdb80b36d34318b2bec5d9", size = 1955269, upload-time = "2025-04-23T18:32:15.783Z" }, + { url = "https://files.pythonhosted.org/packages/f9/86/1beda0576969592f1497b4ce8e7bc8cbdf614c352426271b1b10d5f0aa64/pydantic_core-2.33.2-cp313-cp313-win_arm64.whl", hash = "sha256:e80b087132752f6b3d714f041ccf74403799d3b23a72722ea2e6ba2e892555b9", size = 1893921, upload-time = "2025-04-23T18:32:18.473Z" }, + { url = "https://files.pythonhosted.org/packages/a4/7d/e09391c2eebeab681df2b74bfe6c43422fffede8dc74187b2b0bf6fd7571/pydantic_core-2.33.2-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:61c18fba8e5e9db3ab908620af374db0ac1baa69f0f32df4f61ae23f15e586ac", size = 1806162, upload-time = "2025-04-23T18:32:20.188Z" }, + { url = "https://files.pythonhosted.org/packages/f1/3d/847b6b1fed9f8ed3bb95a9ad04fbd0b212e832d4f0f50ff4d9ee5a9f15cf/pydantic_core-2.33.2-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:95237e53bb015f67b63c91af7518a62a8660376a6a0db19b89acc77a4d6199f5", size = 1981560, upload-time = "2025-04-23T18:32:22.354Z" }, + { url = "https://files.pythonhosted.org/packages/6f/9a/e73262f6c6656262b5fdd723ad90f518f579b7bc8622e43a942eec53c938/pydantic_core-2.33.2-cp313-cp313t-win_amd64.whl", hash = "sha256:c2fc0a768ef76c15ab9238afa6da7f69895bb5d1ee83aeea2e3509af4472d0b9", size = 1935777, upload-time = "2025-04-23T18:32:25.088Z" }, + { url = "https://files.pythonhosted.org/packages/30/68/373d55e58b7e83ce371691f6eaa7175e3a24b956c44628eb25d7da007917/pydantic_core-2.33.2-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:5c4aa4e82353f65e548c476b37e64189783aa5384903bfea4f41580f255fddfa", size = 2023982, upload-time = "2025-04-23T18:32:53.14Z" }, + { url = "https://files.pythonhosted.org/packages/a4/16/145f54ac08c96a63d8ed6442f9dec17b2773d19920b627b18d4f10a061ea/pydantic_core-2.33.2-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:d946c8bf0d5c24bf4fe333af284c59a19358aa3ec18cb3dc4370080da1e8ad29", size = 1858412, upload-time = "2025-04-23T18:32:55.52Z" }, + { url = "https://files.pythonhosted.org/packages/41/b1/c6dc6c3e2de4516c0bb2c46f6a373b91b5660312342a0cf5826e38ad82fa/pydantic_core-2.33.2-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:87b31b6846e361ef83fedb187bb5b4372d0da3f7e28d85415efa92d6125d6e6d", size = 1892749, upload-time = "2025-04-23T18:32:57.546Z" }, + { url = "https://files.pythonhosted.org/packages/12/73/8cd57e20afba760b21b742106f9dbdfa6697f1570b189c7457a1af4cd8a0/pydantic_core-2.33.2-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:aa9d91b338f2df0508606f7009fde642391425189bba6d8c653afd80fd6bb64e", size = 2067527, upload-time = "2025-04-23T18:32:59.771Z" }, + { url = "https://files.pythonhosted.org/packages/e3/d5/0bb5d988cc019b3cba4a78f2d4b3854427fc47ee8ec8e9eaabf787da239c/pydantic_core-2.33.2-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:2058a32994f1fde4ca0480ab9d1e75a0e8c87c22b53a3ae66554f9af78f2fe8c", size = 2108225, upload-time = "2025-04-23T18:33:04.51Z" }, + { url = "https://files.pythonhosted.org/packages/f1/c5/00c02d1571913d496aabf146106ad8239dc132485ee22efe08085084ff7c/pydantic_core-2.33.2-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:0e03262ab796d986f978f79c943fc5f620381be7287148b8010b4097f79a39ec", size = 2069490, upload-time = "2025-04-23T18:33:06.391Z" }, + { url = "https://files.pythonhosted.org/packages/22/a8/dccc38768274d3ed3a59b5d06f59ccb845778687652daa71df0cab4040d7/pydantic_core-2.33.2-pp310-pypy310_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:1a8695a8d00c73e50bff9dfda4d540b7dee29ff9b8053e38380426a85ef10052", size = 2237525, upload-time = "2025-04-23T18:33:08.44Z" }, + { url = "https://files.pythonhosted.org/packages/d4/e7/4f98c0b125dda7cf7ccd14ba936218397b44f50a56dd8c16a3091df116c3/pydantic_core-2.33.2-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:fa754d1850735a0b0e03bcffd9d4b4343eb417e47196e4485d9cca326073a42c", size = 2238446, upload-time = "2025-04-23T18:33:10.313Z" }, + { url = "https://files.pythonhosted.org/packages/ce/91/2ec36480fdb0b783cd9ef6795753c1dea13882f2e68e73bce76ae8c21e6a/pydantic_core-2.33.2-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:a11c8d26a50bfab49002947d3d237abe4d9e4b5bdc8846a63537b6488e197808", size = 2066678, upload-time = "2025-04-23T18:33:12.224Z" }, + { url = "https://files.pythonhosted.org/packages/7b/27/d4ae6487d73948d6f20dddcd94be4ea43e74349b56eba82e9bdee2d7494c/pydantic_core-2.33.2-pp311-pypy311_pp73-macosx_10_12_x86_64.whl", hash = "sha256:dd14041875d09cc0f9308e37a6f8b65f5585cf2598a53aa0123df8b129d481f8", size = 2025200, upload-time = "2025-04-23T18:33:14.199Z" }, + { url = "https://files.pythonhosted.org/packages/f1/b8/b3cb95375f05d33801024079b9392a5ab45267a63400bf1866e7ce0f0de4/pydantic_core-2.33.2-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:d87c561733f66531dced0da6e864f44ebf89a8fba55f31407b00c2f7f9449593", size = 1859123, upload-time = "2025-04-23T18:33:16.555Z" }, + { url = "https://files.pythonhosted.org/packages/05/bc/0d0b5adeda59a261cd30a1235a445bf55c7e46ae44aea28f7bd6ed46e091/pydantic_core-2.33.2-pp311-pypy311_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2f82865531efd18d6e07a04a17331af02cb7a651583c418df8266f17a63c6612", size = 1892852, upload-time = "2025-04-23T18:33:18.513Z" }, + { url = "https://files.pythonhosted.org/packages/3e/11/d37bdebbda2e449cb3f519f6ce950927b56d62f0b84fd9cb9e372a26a3d5/pydantic_core-2.33.2-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2bfb5112df54209d820d7bf9317c7a6c9025ea52e49f46b6a2060104bba37de7", size = 2067484, upload-time = "2025-04-23T18:33:20.475Z" }, + { url = "https://files.pythonhosted.org/packages/8c/55/1f95f0a05ce72ecb02a8a8a1c3be0579bbc29b1d5ab68f1378b7bebc5057/pydantic_core-2.33.2-pp311-pypy311_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:64632ff9d614e5eecfb495796ad51b0ed98c453e447a76bcbeeb69615079fc7e", size = 2108896, upload-time = "2025-04-23T18:33:22.501Z" }, + { url = "https://files.pythonhosted.org/packages/53/89/2b2de6c81fa131f423246a9109d7b2a375e83968ad0800d6e57d0574629b/pydantic_core-2.33.2-pp311-pypy311_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:f889f7a40498cc077332c7ab6b4608d296d852182211787d4f3ee377aaae66e8", size = 2069475, upload-time = "2025-04-23T18:33:24.528Z" }, + { url = "https://files.pythonhosted.org/packages/b8/e9/1f7efbe20d0b2b10f6718944b5d8ece9152390904f29a78e68d4e7961159/pydantic_core-2.33.2-pp311-pypy311_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:de4b83bb311557e439b9e186f733f6c645b9417c84e2eb8203f3f820a4b988bf", size = 2239013, upload-time = "2025-04-23T18:33:26.621Z" }, + { url = "https://files.pythonhosted.org/packages/3c/b2/5309c905a93811524a49b4e031e9851a6b00ff0fb668794472ea7746b448/pydantic_core-2.33.2-pp311-pypy311_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:82f68293f055f51b51ea42fafc74b6aad03e70e191799430b90c13d643059ebb", size = 2238715, upload-time = "2025-04-23T18:33:28.656Z" }, + { url = "https://files.pythonhosted.org/packages/32/56/8a7ca5d2cd2cda1d245d34b1c9a942920a718082ae8e54e5f3e5a58b7add/pydantic_core-2.33.2-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:329467cecfb529c925cf2bbd4d60d2c509bc2fb52a20c1045bf09bb70971a9c1", size = 2066757, upload-time = "2025-04-23T18:33:30.645Z" }, +] + +[[package]] +name = "pydub" +version = "0.25.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/fe/9a/e6bca0eed82db26562c73b5076539a4a08d3cffd19c3cc5913a3e61145fd/pydub-0.25.1.tar.gz", hash = "sha256:980a33ce9949cab2a569606b65674d748ecbca4f0796887fd6f46173a7b0d30f", size = 38326, upload-time = "2021-03-10T02:09:54.659Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/a6/53/d78dc063216e62fc55f6b2eebb447f6a4b0a59f55c8406376f76bf959b08/pydub-0.25.1-py2.py3-none-any.whl", hash = "sha256:65617e33033874b59d87db603aa1ed450633288aefead953b30bded59cb599a6", size = 32327, upload-time = "2021-03-10T02:09:53.503Z" }, +] + +[[package]] +name = "pygments" +version = "2.19.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/b0/77/a5b8c569bf593b0140bde72ea885a803b82086995367bf2037de0159d924/pygments-2.19.2.tar.gz", hash = "sha256:636cb2477cec7f8952536970bc533bc43743542f70392ae026374600add5b887", size = 4968631, upload-time = "2025-06-21T13:39:12.283Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/c7/21/705964c7812476f378728bdf590ca4b771ec72385c533964653c68e86bdc/pygments-2.19.2-py3-none-any.whl", hash = "sha256:86540386c03d588bb81d44bc3928634ff26449851e99741617ecb9037ee5ec0b", size = 1225217, upload-time = "2025-06-21T13:39:07.939Z" }, +] + +[[package]] +name = "pytest" +version = "9.0.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "colorama", marker = "sys_platform == 'win32'" }, + { name = "exceptiongroup", marker = "python_full_version < '3.11'" }, + { name = "iniconfig" }, + { name = "packaging" }, + { name = "pluggy" }, + { name = "pygments" }, + { name = "tomli", marker = "python_full_version < '3.11'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/07/56/f013048ac4bc4c1d9be45afd4ab209ea62822fb1598f40687e6bf45dcea4/pytest-9.0.1.tar.gz", hash = "sha256:3e9c069ea73583e255c3b21cf46b8d3c56f6e3a1a8f6da94ccb0fcf57b9d73c8", size = 1564125, upload-time = "2025-11-12T13:05:09.333Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/0b/8b/6300fb80f858cda1c51ffa17075df5d846757081d11ab4aa35cef9e6258b/pytest-9.0.1-py3-none-any.whl", hash = "sha256:67be0030d194df2dfa7b556f2e56fb3c3315bd5c8822c6951162b92b32ce7dad", size = 373668, upload-time = "2025-11-12T13:05:07.379Z" }, +] + +[[package]] +name = "python-dateutil" +version = "2.9.0.post0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "six" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/66/c0/0c8b6ad9f17a802ee498c46e004a0eb49bc148f2fd230864601a86dcf6db/python-dateutil-2.9.0.post0.tar.gz", hash = "sha256:37dd54208da7e1cd875388217d5e00ebd4179249f90fb72437e91a35459a0ad3", size = 342432, upload-time = "2024-03-01T18:36:20.211Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/ec/57/56b9bcc3c9c6a792fcbaf139543cee77261f3651ca9da0c93f5c1221264b/python_dateutil-2.9.0.post0-py2.py3-none-any.whl", hash = "sha256:a8b2bc7bffae282281c8140a97d3aa9c14da0b136dfe83f850eea9a5f7470427", size = 229892, upload-time = "2024-03-01T18:36:18.57Z" }, +] + +[[package]] +name = "python-dotenv" +version = "1.2.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/f0/26/19cadc79a718c5edbec86fd4919a6b6d3f681039a2f6d66d14be94e75fb9/python_dotenv-1.2.1.tar.gz", hash = "sha256:42667e897e16ab0d66954af0e60a9caa94f0fd4ecf3aaf6d2d260eec1aa36ad6", size = 44221, upload-time = "2025-10-26T15:12:10.434Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/14/1b/a298b06749107c305e1fe0f814c6c74aea7b2f1e10989cb30f544a1b3253/python_dotenv-1.2.1-py3-none-any.whl", hash = "sha256:b81ee9561e9ca4004139c6cbba3a238c32b03e4894671e181b671e8cb8425d61", size = 21230, upload-time = "2025-10-26T15:12:09.109Z" }, +] + +[[package]] +name = "python-multipart" +version = "0.0.20" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/f3/87/f44d7c9f274c7ee665a29b885ec97089ec5dc034c7f3fafa03da9e39a09e/python_multipart-0.0.20.tar.gz", hash = "sha256:8dd0cab45b8e23064ae09147625994d090fa46f5b0d1e13af944c331a7fa9d13", size = 37158, upload-time = "2024-12-16T19:45:46.972Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/45/58/38b5afbc1a800eeea951b9285d3912613f2603bdf897a4ab0f4bd7f405fc/python_multipart-0.0.20-py3-none-any.whl", hash = "sha256:8a62d3a8335e06589fe01f2a3e178cdcc632f3fbe0d492ad9ee0ec35aab1f104", size = 24546, upload-time = "2024-12-16T19:45:44.423Z" }, +] + +[[package]] +name = "pytorch-msssim" +version = "1.0.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/69/43/fd181aef46e96ed025f5056efa1c45e63c93d377fbe7ea784a21ce058d03/pytorch_msssim-1.0.0.tar.gz", hash = "sha256:53273f03ce5c435f706f49feb5110a8afe90b2519484b0fe0df9639670106d0b", size = 10113, upload-time = "2023-05-25T17:15:57.146Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/e2/8c/856047f955acc30179e9255fdc488059ca22f0938519523d53494f7cfee8/pytorch_msssim-1.0.0-py3-none-any.whl", hash = "sha256:0b4b7bbf7035fe9dc8084244237aac13b1f104852c45b63a7e9fab4363bede54", size = 7744, upload-time = "2023-05-25T17:15:55.809Z" }, +] + +[[package]] +name = "pytz" +version = "2025.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/f8/bf/abbd3cdfb8fbc7fb3d4d38d320f2441b1e7cbe29be4f23797b4a2b5d8aac/pytz-2025.2.tar.gz", hash = "sha256:360b9e3dbb49a209c21ad61809c7fb453643e048b38924c765813546746e81c3", size = 320884, upload-time = "2025-03-25T02:25:00.538Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/81/c4/34e93fe5f5429d7570ec1fa436f1986fb1f00c3e0f43a589fe2bbcd22c3f/pytz-2025.2-py2.py3-none-any.whl", hash = "sha256:5ddf76296dd8c44c26eb8f4b6f35488f3ccbf6fbbd7adee0b7262d43f0ec2f00", size = 509225, upload-time = "2025-03-25T02:24:58.468Z" }, +] + +[[package]] +name = "pyyaml" +version = "6.0.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/cd/e5/af35f7ea75cf72f2cd079c95ee16797de7cd71f29ea7c68ae5ce7be1eda0/PyYAML-6.0.1.tar.gz", hash = "sha256:bfdf460b1736c775f2ba9f6a92bca30bc2095067b8a9d77876d1fad6cc3b4a43", size = 125201, upload-time = "2023-07-18T00:00:23.308Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/96/06/4beb652c0fe16834032e54f0956443d4cc797fe645527acee59e7deaa0a2/PyYAML-6.0.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:d858aa552c999bc8a8d57426ed01e40bef403cd8ccdd0fc5f6f04a00414cac2a", size = 189447, upload-time = "2023-07-17T23:57:04.325Z" }, + { url = "https://files.pythonhosted.org/packages/5b/07/10033a403b23405a8fc48975444463d3d10a5c2736b7eb2550b07b367429/PyYAML-6.0.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:fd66fc5d0da6d9815ba2cebeb4205f95818ff4b79c3ebe268e75d961704af52f", size = 169264, upload-time = "2023-07-17T23:57:07.787Z" }, + { url = "https://files.pythonhosted.org/packages/f1/26/55e4f21db1f72eaef092015d9017c11510e7e6301c62a6cfee91295d13c6/PyYAML-6.0.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:69b023b2b4daa7548bcfbd4aa3da05b3a74b772db9e23b982788168117739938", size = 677003, upload-time = "2023-07-17T23:57:13.144Z" }, + { url = "https://files.pythonhosted.org/packages/ba/91/090818dfa62e85181f3ae23dd1e8b7ea7f09684864a900cab72d29c57346/PyYAML-6.0.1-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:81e0b275a9ecc9c0c0c07b4b90ba548307583c125f54d5b6946cfee6360c733d", size = 699070, upload-time = "2023-07-17T23:57:19.402Z" }, + { url = "https://files.pythonhosted.org/packages/29/61/bf33c6c85c55bc45a29eee3195848ff2d518d84735eb0e2d8cb42e0d285e/PyYAML-6.0.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ba336e390cd8e4d1739f42dfe9bb83a3cc2e80f567d8805e11b46f4a943f5515", size = 705525, upload-time = "2023-07-17T23:57:25.272Z" }, + { url = "https://files.pythonhosted.org/packages/07/91/45dfd0ef821a7f41d9d0136ea3608bb5b1653e42fd56a7970532cb5c003f/PyYAML-6.0.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:326c013efe8048858a6d312ddd31d56e468118ad4cdeda36c719bf5bb6192290", size = 707514, upload-time = "2023-08-28T18:43:20.945Z" }, + { url = "https://files.pythonhosted.org/packages/b6/a0/b6700da5d49e9fed49dc3243d3771b598dad07abb37cc32e524607f96adc/PyYAML-6.0.1-cp310-cp310-win32.whl", hash = "sha256:bd4af7373a854424dabd882decdc5579653d7868b8fb26dc7d0e99f823aa5924", size = 130488, upload-time = "2023-07-17T23:57:28.144Z" }, + { url = "https://files.pythonhosted.org/packages/24/97/9b59b43431f98d01806b288532da38099cc6f2fea0f3d712e21e269c0279/PyYAML-6.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:fd1592b3fdf65fff2ad0004b5e363300ef59ced41c2e6b3a99d4089fa8c5435d", size = 145338, upload-time = "2023-07-17T23:57:31.118Z" }, + { url = "https://files.pythonhosted.org/packages/ec/0d/26fb23e8863e0aeaac0c64e03fd27367ad2ae3f3cccf3798ee98ce160368/PyYAML-6.0.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:6965a7bc3cf88e5a1c3bd2e0b5c22f8d677dc88a455344035f03399034eb3007", size = 187867, upload-time = "2023-07-17T23:57:34.35Z" }, + { url = "https://files.pythonhosted.org/packages/28/09/55f715ddbf95a054b764b547f617e22f1d5e45d83905660e9a088078fe67/PyYAML-6.0.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:f003ed9ad21d6a4713f0a9b5a7a0a79e08dd0f221aff4525a2be4c346ee60aab", size = 167530, upload-time = "2023-07-17T23:57:36.975Z" }, + { url = "https://files.pythonhosted.org/packages/5e/94/7d5ee059dfb92ca9e62f4057dcdec9ac08a9e42679644854dc01177f8145/PyYAML-6.0.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:42f8152b8dbc4fe7d96729ec2b99c7097d656dc1213a3229ca5383f973a5ed6d", size = 732244, upload-time = "2023-07-17T23:57:43.774Z" }, + { url = "https://files.pythonhosted.org/packages/06/92/e0224aa6ebf9dc54a06a4609da37da40bb08d126f5535d81bff6b417b2ae/PyYAML-6.0.1-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:062582fca9fabdd2c8b54a3ef1c978d786e0f6b3a1510e0ac93ef59e0ddae2bc", size = 752871, upload-time = "2023-07-17T23:57:51.921Z" }, + { url = "https://files.pythonhosted.org/packages/7b/5e/efd033ab7199a0b2044dab3b9f7a4f6670e6a52c089de572e928d2873b06/PyYAML-6.0.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d2b04aac4d386b172d5b9692e2d2da8de7bfb6c387fa4f801fbf6fb2e6ba4673", size = 757729, upload-time = "2023-07-17T23:57:59.865Z" }, + { url = "https://files.pythonhosted.org/packages/03/5c/c4671451b2f1d76ebe352c0945d4cd13500adb5d05f5a51ee296d80152f7/PyYAML-6.0.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:e7d73685e87afe9f3b36c799222440d6cf362062f78be1013661b00c5c6f678b", size = 748528, upload-time = "2023-08-28T18:43:23.207Z" }, + { url = "https://files.pythonhosted.org/packages/73/9c/766e78d1efc0d1fca637a6b62cea1b4510a7fb93617eb805223294fef681/PyYAML-6.0.1-cp311-cp311-win32.whl", hash = "sha256:1635fd110e8d85d55237ab316b5b011de701ea0f29d07611174a1b42f1444741", size = 130286, upload-time = "2023-07-17T23:58:02.964Z" }, + { url = "https://files.pythonhosted.org/packages/b3/34/65bb4b2d7908044963ebf614fe0fdb080773fc7030d7e39c8d3eddcd4257/PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34", size = 144699, upload-time = "2023-07-17T23:58:05.586Z" }, + { url = "https://files.pythonhosted.org/packages/bc/06/1b305bf6aa704343be85444c9d011f626c763abb40c0edc1cad13bfd7f86/PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28", size = 178692, upload-time = "2023-08-28T18:43:24.924Z" }, + { url = "https://files.pythonhosted.org/packages/84/02/404de95ced348b73dd84f70e15a41843d817ff8c1744516bf78358f2ffd2/PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9", size = 165622, upload-time = "2023-08-28T18:43:26.54Z" }, + { url = "https://files.pythonhosted.org/packages/c7/4c/4a2908632fc980da6d918b9de9c1d9d7d7e70b2672b1ad5166ed27841ef7/PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef", size = 696937, upload-time = "2024-01-18T20:40:22.92Z" }, + { url = "https://files.pythonhosted.org/packages/b4/33/720548182ffa8344418126017aa1d4ab4aeec9a2275f04ce3f3573d8ace8/PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0", size = 724969, upload-time = "2023-08-28T18:43:28.56Z" }, + { url = "https://files.pythonhosted.org/packages/4f/78/77b40157b6cb5f2d3d31a3d9b2efd1ba3505371f76730d267e8b32cf4b7f/PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4", size = 712604, upload-time = "2023-08-28T18:43:30.206Z" }, + { url = "https://files.pythonhosted.org/packages/2e/97/3e0e089ee85e840f4b15bfa00e4e63d84a3691ababbfea92d6f820ea6f21/PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54", size = 126098, upload-time = "2023-08-28T18:43:31.835Z" }, + { url = "https://files.pythonhosted.org/packages/2b/9f/fbade56564ad486809c27b322d0f7e6a89c01f6b4fe208402e90d4443a99/PyYAML-6.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:0d3304d8c0adc42be59c5f8a4d9e3d7379e6955ad754aa9d6ab7a398b59dd1df", size = 138675, upload-time = "2023-08-28T18:43:33.613Z" }, +] + +[[package]] +name = "ray" +version = "2.53.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "click" }, + { name = "filelock" }, + { name = "jsonschema" }, + { name = "msgpack" }, + { name = "packaging" }, + { name = "protobuf" }, + { name = "pyyaml" }, + { name = "requests" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/2f/99/21986c7f8135dafbf7c49229c52faaa9d2d365db7d86fffe978dde8ee967/ray-2.53.0-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:4db914a0a6dd608fa49c066929a1282745a2dbd73caee67d7b80fe684ca65bdd", size = 69473649, upload-time = "2025-12-20T16:05:40.58Z" }, + { url = "https://files.pythonhosted.org/packages/70/d9/58b5426a3f11993851db3c93841358cebdddd948153481d355b720f31f9d/ray-2.53.0-cp310-cp310-manylinux2014_aarch64.whl", hash = "sha256:4108280d8a1cb90d7d68e5c954c35e63b8bb9a4ba15f88c5e7da0e2025647712", size = 71342662, upload-time = "2025-12-20T16:05:46.936Z" }, + { url = "https://files.pythonhosted.org/packages/c5/05/4aa32370b313481c2d1d41cb53ec786daebdb2ef665b01ef2ac43d9cf457/ray-2.53.0-cp310-cp310-manylinux2014_x86_64.whl", hash = "sha256:4dbb5fce1364763f29741055f50abe33cf726397141f9cc0e845dd3cc963e455", size = 72188620, upload-time = "2025-12-20T16:05:52.817Z" }, + { url = "https://files.pythonhosted.org/packages/f7/c6/21efe5886898421df20078a333b0984eade7d7aa4bdc68a336f0c66db27e/ray-2.53.0-cp310-cp310-win_amd64.whl", hash = "sha256:90faf630d20b6abf3135997fb3edb5842134aff92e04ee709865db04816d97ef", size = 27200553, upload-time = "2025-12-20T16:05:57.655Z" }, + { url = "https://files.pythonhosted.org/packages/bf/64/d5c29a4b014d8b9a624203a88b67630072c1d6960425dbf7a1f0fa5d6b74/ray-2.53.0-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:bd3ec4c342776ddac23ae2b108c64f5939f417ccc4875900d586c7c978463269", size = 69479296, upload-time = "2025-12-20T16:06:05.111Z" }, + { url = "https://files.pythonhosted.org/packages/c6/41/9e19d1e5d9458a5ba157c36642e2874bcb22fddbd7c1e77b668e5afc3f3d/ray-2.53.0-cp311-cp311-manylinux2014_aarch64.whl", hash = "sha256:a0bbb98b0b0f25a3ee075ca10171e1260e70b6bc690cd509ecd7ce1228af854d", size = 71463449, upload-time = "2025-12-20T16:06:10.983Z" }, + { url = "https://files.pythonhosted.org/packages/63/de/58c19906b0dd16ea06b4f2465b7327f5f180e6b6e1c8c9b610d7c589ea5f/ray-2.53.0-cp311-cp311-manylinux2014_x86_64.whl", hash = "sha256:eb000c17f7301071fdd15c44c4cd3ac0f7953bb4c7c227e61719fe7048195bcd", size = 72305102, upload-time = "2025-12-20T16:06:17.989Z" }, + { url = "https://files.pythonhosted.org/packages/b1/43/72cc1cfe17d26abe62a793eab10445f9546dce24192b85a6cd0cdc47ed86/ray-2.53.0-cp311-cp311-win_amd64.whl", hash = "sha256:4a1bb3fe09ab4cd0d16ddc96b9f60c9ed83b3f93b87aa8506e0d3b746fd4e825", size = 27194174, upload-time = "2025-12-20T16:06:23.042Z" }, + { url = "https://files.pythonhosted.org/packages/b2/44/562718a634e63e8ef7985285288a167d4af62bc2a7decce3300cf937776a/ray-2.53.0-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:d8b95d047d947493803fb8417aea31225dcacdab15afdc75b8a238901949d457", size = 69463763, upload-time = "2025-12-20T16:06:28.685Z" }, + { url = "https://files.pythonhosted.org/packages/38/68/8e59b8413f3751fe7ce8b98ee8787d13964b47a4043587950790a9dd2151/ray-2.53.0-cp312-cp312-manylinux2014_aarch64.whl", hash = "sha256:65e2ce58d3dc6baa3cf45824d889c1968ebde565ee54dfd80a98af8f31af8e4a", size = 71504450, upload-time = "2025-12-20T16:06:34.922Z" }, + { url = "https://files.pythonhosted.org/packages/2a/db/978a50d264565ca42e2a4bf115ec9a1f04f19ca5e620e6aa2f280747b644/ray-2.53.0-cp312-cp312-manylinux2014_x86_64.whl", hash = "sha256:14f46363e9b4cf0c1c8b4d8623ec337c5bd408377831b5e5b50067930137bbca", size = 72370424, upload-time = "2025-12-20T16:06:40.821Z" }, + { url = "https://files.pythonhosted.org/packages/8d/6c/bba6f22a9d83ee8f236000ba315f0c197bdc79888b4fa42fd762f729cbbd/ray-2.53.0-cp312-cp312-win_amd64.whl", hash = "sha256:b828c147f9ff2f277b1d254e4fe9a746fdfaee7e313a93a97c7edf4dae9b81a4", size = 27178106, upload-time = "2025-12-20T16:06:45.594Z" }, + { url = "https://files.pythonhosted.org/packages/3d/38/450cf9cf3c490fa4cc6d470597f819444da60f85579d2b34b95ee79fcb6f/ray-2.53.0-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:85b472ab6fb8f1189f8cef81913fd91b24dd69b3fa7dcca7e144827bd924f6c0", size = 69409819, upload-time = "2025-12-20T16:06:50.668Z" }, + { url = "https://files.pythonhosted.org/packages/71/5e/d452970b07174d5e4f8688abae889d01321b51ced827db1f1d1cb7d56d44/ray-2.53.0-cp313-cp313-manylinux2014_aarch64.whl", hash = "sha256:7196e5358dfcc8211be864f45e6dfe4827202df294af3c7a76ff8fbc080e0522", size = 71409529, upload-time = "2025-12-20T16:06:56.2Z" }, + { url = "https://files.pythonhosted.org/packages/cb/84/50b317a125617a638a64694c12f56183edd5df01828a35fa4c55c7b13c66/ray-2.53.0-cp313-cp313-manylinux2014_x86_64.whl", hash = "sha256:73dbbaa7962a7f5e38aa8cf9483e0e9817205e989aa3dc859c738c2af1ae01df", size = 72283961, upload-time = "2025-12-20T16:07:05.831Z" }, +] + +[[package]] +name = "referencing" +version = "0.37.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "attrs" }, + { name = "rpds-py" }, + { name = "typing-extensions", marker = "python_full_version < '3.13'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/22/f5/df4e9027acead3ecc63e50fe1e36aca1523e1719559c499951bb4b53188f/referencing-0.37.0.tar.gz", hash = "sha256:44aefc3142c5b842538163acb373e24cce6632bd54bdb01b21ad5863489f50d8", size = 78036, upload-time = "2025-10-13T15:30:48.871Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/2c/58/ca301544e1fa93ed4f80d724bf5b194f6e4b945841c5bfd555878eea9fcb/referencing-0.37.0-py3-none-any.whl", hash = "sha256:381329a9f99628c9069361716891d34ad94af76e461dcb0335825aecc7692231", size = 26766, upload-time = "2025-10-13T15:30:47.625Z" }, +] + +[[package]] +name = "regex" +version = "2025.11.3" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/cc/a9/546676f25e573a4cf00fe8e119b78a37b6a8fe2dc95cda877b30889c9c45/regex-2025.11.3.tar.gz", hash = "sha256:1fedc720f9bb2494ce31a58a1631f9c82df6a09b49c19517ea5cc280b4541e01", size = 414669, upload-time = "2025-11-03T21:34:22.089Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/8a/d6/d788d52da01280a30a3f6268aef2aa71043bff359c618fea4c5b536654d5/regex-2025.11.3-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:2b441a4ae2c8049106e8b39973bfbddfb25a179dda2bdb99b0eeb60c40a6a3af", size = 488087, upload-time = "2025-11-03T21:30:47.317Z" }, + { url = "https://files.pythonhosted.org/packages/69/39/abec3bd688ec9bbea3562de0fd764ff802976185f5ff22807bf0a2697992/regex-2025.11.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:2fa2eed3f76677777345d2f81ee89f5de2f5745910e805f7af7386a920fa7313", size = 290544, upload-time = "2025-11-03T21:30:49.912Z" }, + { url = "https://files.pythonhosted.org/packages/39/b3/9a231475d5653e60002508f41205c61684bb2ffbf2401351ae2186897fc4/regex-2025.11.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:d8b4a27eebd684319bdf473d39f1d79eed36bf2cd34bd4465cdb4618d82b3d56", size = 288408, upload-time = "2025-11-03T21:30:51.344Z" }, + { url = "https://files.pythonhosted.org/packages/c3/c5/1929a0491bd5ac2d1539a866768b88965fa8c405f3e16a8cef84313098d6/regex-2025.11.3-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5cf77eac15bd264986c4a2c63353212c095b40f3affb2bc6b4ef80c4776c1a28", size = 781584, upload-time = "2025-11-03T21:30:52.596Z" }, + { url = "https://files.pythonhosted.org/packages/ce/fd/16aa16cf5d497ef727ec966f74164fbe75d6516d3d58ac9aa989bc9cdaad/regex-2025.11.3-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b7f9ee819f94c6abfa56ec7b1dbab586f41ebbdc0a57e6524bd5e7f487a878c7", size = 850733, upload-time = "2025-11-03T21:30:53.825Z" }, + { url = "https://files.pythonhosted.org/packages/e6/49/3294b988855a221cb6565189edf5dc43239957427df2d81d4a6b15244f64/regex-2025.11.3-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:838441333bc90b829406d4a03cb4b8bf7656231b84358628b0406d803931ef32", size = 898691, upload-time = "2025-11-03T21:30:55.575Z" }, + { url = "https://files.pythonhosted.org/packages/14/62/b56d29e70b03666193369bdbdedfdc23946dbe9f81dd78ce262c74d988ab/regex-2025.11.3-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:cfe6d3f0c9e3b7e8c0c694b24d25e677776f5ca26dce46fd6b0489f9c8339391", size = 791662, upload-time = "2025-11-03T21:30:57.262Z" }, + { url = "https://files.pythonhosted.org/packages/15/fc/e4c31d061eced63fbf1ce9d853975f912c61a7d406ea14eda2dd355f48e7/regex-2025.11.3-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:2ab815eb8a96379a27c3b6157fcb127c8f59c36f043c1678110cea492868f1d5", size = 782587, upload-time = "2025-11-03T21:30:58.788Z" }, + { url = "https://files.pythonhosted.org/packages/b2/bb/5e30c7394bcf63f0537121c23e796be67b55a8847c3956ae6068f4c70702/regex-2025.11.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:728a9d2d173a65b62bdc380b7932dd8e74ed4295279a8fe1021204ce210803e7", size = 774709, upload-time = "2025-11-03T21:31:00.081Z" }, + { url = "https://files.pythonhosted.org/packages/c5/c4/fce773710af81b0cb37cb4ff0947e75d5d17dee304b93d940b87a67fc2f4/regex-2025.11.3-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:509dc827f89c15c66a0c216331260d777dd6c81e9a4e4f830e662b0bb296c313", size = 845773, upload-time = "2025-11-03T21:31:01.583Z" }, + { url = "https://files.pythonhosted.org/packages/7b/5e/9466a7ec4b8ec282077095c6eb50a12a389d2e036581134d4919e8ca518c/regex-2025.11.3-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:849202cd789e5f3cf5dcc7822c34b502181b4824a65ff20ce82da5524e45e8e9", size = 836164, upload-time = "2025-11-03T21:31:03.244Z" }, + { url = "https://files.pythonhosted.org/packages/95/18/82980a60e8ed1594eb3c89eb814fb276ef51b9af7caeab1340bfd8564af6/regex-2025.11.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:b6f78f98741dcc89607c16b1e9426ee46ce4bf31ac5e6b0d40e81c89f3481ea5", size = 779832, upload-time = "2025-11-03T21:31:04.876Z" }, + { url = "https://files.pythonhosted.org/packages/03/cc/90ab0fdbe6dce064a42015433f9152710139fb04a8b81b4fb57a1cb63ffa/regex-2025.11.3-cp310-cp310-win32.whl", hash = "sha256:149eb0bba95231fb4f6d37c8f760ec9fa6fabf65bab555e128dde5f2475193ec", size = 265802, upload-time = "2025-11-03T21:31:06.581Z" }, + { url = "https://files.pythonhosted.org/packages/34/9d/e9e8493a85f3b1ddc4a5014465f5c2b78c3ea1cbf238dcfde78956378041/regex-2025.11.3-cp310-cp310-win_amd64.whl", hash = "sha256:ee3a83ce492074c35a74cc76cf8235d49e77b757193a5365ff86e3f2f93db9fd", size = 277722, upload-time = "2025-11-03T21:31:08.144Z" }, + { url = "https://files.pythonhosted.org/packages/15/c4/b54b24f553966564506dbf873a3e080aef47b356a3b39b5d5aba992b50db/regex-2025.11.3-cp310-cp310-win_arm64.whl", hash = "sha256:38af559ad934a7b35147716655d4a2f79fcef2d695ddfe06a06ba40ae631fa7e", size = 270289, upload-time = "2025-11-03T21:31:10.267Z" }, + { url = "https://files.pythonhosted.org/packages/f7/90/4fb5056e5f03a7048abd2b11f598d464f0c167de4f2a51aa868c376b8c70/regex-2025.11.3-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:eadade04221641516fa25139273505a1c19f9bf97589a05bc4cfcd8b4a618031", size = 488081, upload-time = "2025-11-03T21:31:11.946Z" }, + { url = "https://files.pythonhosted.org/packages/85/23/63e481293fac8b069d84fba0299b6666df720d875110efd0338406b5d360/regex-2025.11.3-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:feff9e54ec0dd3833d659257f5c3f5322a12eee58ffa360984b716f8b92983f4", size = 290554, upload-time = "2025-11-03T21:31:13.387Z" }, + { url = "https://files.pythonhosted.org/packages/2b/9d/b101d0262ea293a0066b4522dfb722eb6a8785a8c3e084396a5f2c431a46/regex-2025.11.3-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:3b30bc921d50365775c09a7ed446359e5c0179e9e2512beec4a60cbcef6ddd50", size = 288407, upload-time = "2025-11-03T21:31:14.809Z" }, + { url = "https://files.pythonhosted.org/packages/0c/64/79241c8209d5b7e00577ec9dca35cd493cc6be35b7d147eda367d6179f6d/regex-2025.11.3-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f99be08cfead2020c7ca6e396c13543baea32343b7a9a5780c462e323bd8872f", size = 793418, upload-time = "2025-11-03T21:31:16.556Z" }, + { url = "https://files.pythonhosted.org/packages/3d/e2/23cd5d3573901ce8f9757c92ca4db4d09600b865919b6d3e7f69f03b1afd/regex-2025.11.3-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:6dd329a1b61c0ee95ba95385fb0c07ea0d3fe1a21e1349fa2bec272636217118", size = 860448, upload-time = "2025-11-03T21:31:18.12Z" }, + { url = "https://files.pythonhosted.org/packages/2a/4c/aecf31beeaa416d0ae4ecb852148d38db35391aac19c687b5d56aedf3a8b/regex-2025.11.3-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4c5238d32f3c5269d9e87be0cf096437b7622b6920f5eac4fd202468aaeb34d2", size = 907139, upload-time = "2025-11-03T21:31:20.753Z" }, + { url = "https://files.pythonhosted.org/packages/61/22/b8cb00df7d2b5e0875f60628594d44dba283e951b1ae17c12f99e332cc0a/regex-2025.11.3-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:10483eefbfb0adb18ee9474498c9a32fcf4e594fbca0543bb94c48bac6183e2e", size = 800439, upload-time = "2025-11-03T21:31:22.069Z" }, + { url = "https://files.pythonhosted.org/packages/02/a8/c4b20330a5cdc7a8eb265f9ce593f389a6a88a0c5f280cf4d978f33966bc/regex-2025.11.3-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:78c2d02bb6e1da0720eedc0bad578049cad3f71050ef8cd065ecc87691bed2b0", size = 782965, upload-time = "2025-11-03T21:31:23.598Z" }, + { url = "https://files.pythonhosted.org/packages/b4/4c/ae3e52988ae74af4b04d2af32fee4e8077f26e51b62ec2d12d246876bea2/regex-2025.11.3-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:e6b49cd2aad93a1790ce9cffb18964f6d3a4b0b3dbdbd5de094b65296fce6e58", size = 854398, upload-time = "2025-11-03T21:31:25.008Z" }, + { url = "https://files.pythonhosted.org/packages/06/d1/a8b9cf45874eda14b2e275157ce3b304c87e10fb38d9fc26a6e14eb18227/regex-2025.11.3-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:885b26aa3ee56433b630502dc3d36ba78d186a00cc535d3806e6bfd9ed3c70ab", size = 845897, upload-time = "2025-11-03T21:31:26.427Z" }, + { url = "https://files.pythonhosted.org/packages/ea/fe/1830eb0236be93d9b145e0bd8ab499f31602fe0999b1f19e99955aa8fe20/regex-2025.11.3-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:ddd76a9f58e6a00f8772e72cff8ebcff78e022be95edf018766707c730593e1e", size = 788906, upload-time = "2025-11-03T21:31:28.078Z" }, + { url = "https://files.pythonhosted.org/packages/66/47/dc2577c1f95f188c1e13e2e69d8825a5ac582ac709942f8a03af42ed6e93/regex-2025.11.3-cp311-cp311-win32.whl", hash = "sha256:3e816cc9aac1cd3cc9a4ec4d860f06d40f994b5c7b4d03b93345f44e08cc68bf", size = 265812, upload-time = "2025-11-03T21:31:29.72Z" }, + { url = "https://files.pythonhosted.org/packages/50/1e/15f08b2f82a9bbb510621ec9042547b54d11e83cb620643ebb54e4eb7d71/regex-2025.11.3-cp311-cp311-win_amd64.whl", hash = "sha256:087511f5c8b7dfbe3a03f5d5ad0c2a33861b1fc387f21f6f60825a44865a385a", size = 277737, upload-time = "2025-11-03T21:31:31.422Z" }, + { url = "https://files.pythonhosted.org/packages/f4/fc/6500eb39f5f76c5e47a398df82e6b535a5e345f839581012a418b16f9cc3/regex-2025.11.3-cp311-cp311-win_arm64.whl", hash = "sha256:1ff0d190c7f68ae7769cd0313fe45820ba07ffebfddfaa89cc1eb70827ba0ddc", size = 270290, upload-time = "2025-11-03T21:31:33.041Z" }, + { url = "https://files.pythonhosted.org/packages/e8/74/18f04cb53e58e3fb107439699bd8375cf5a835eec81084e0bddbd122e4c2/regex-2025.11.3-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:bc8ab71e2e31b16e40868a40a69007bc305e1109bd4658eb6cad007e0bf67c41", size = 489312, upload-time = "2025-11-03T21:31:34.343Z" }, + { url = "https://files.pythonhosted.org/packages/78/3f/37fcdd0d2b1e78909108a876580485ea37c91e1acf66d3bb8e736348f441/regex-2025.11.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:22b29dda7e1f7062a52359fca6e58e548e28c6686f205e780b02ad8ef710de36", size = 291256, upload-time = "2025-11-03T21:31:35.675Z" }, + { url = "https://files.pythonhosted.org/packages/bf/26/0a575f58eb23b7ebd67a45fccbc02ac030b737b896b7e7a909ffe43ffd6a/regex-2025.11.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:3a91e4a29938bc1a082cc28fdea44be420bf2bebe2665343029723892eb073e1", size = 288921, upload-time = "2025-11-03T21:31:37.07Z" }, + { url = "https://files.pythonhosted.org/packages/ea/98/6a8dff667d1af907150432cf5abc05a17ccd32c72a3615410d5365ac167a/regex-2025.11.3-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:08b884f4226602ad40c5d55f52bf91a9df30f513864e0054bad40c0e9cf1afb7", size = 798568, upload-time = "2025-11-03T21:31:38.784Z" }, + { url = "https://files.pythonhosted.org/packages/64/15/92c1db4fa4e12733dd5a526c2dd2b6edcbfe13257e135fc0f6c57f34c173/regex-2025.11.3-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:3e0b11b2b2433d1c39c7c7a30e3f3d0aeeea44c2a8d0bae28f6b95f639927a69", size = 864165, upload-time = "2025-11-03T21:31:40.559Z" }, + { url = "https://files.pythonhosted.org/packages/f9/e7/3ad7da8cdee1ce66c7cd37ab5ab05c463a86ffeb52b1a25fe7bd9293b36c/regex-2025.11.3-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:87eb52a81ef58c7ba4d45c3ca74e12aa4b4e77816f72ca25258a85b3ea96cb48", size = 912182, upload-time = "2025-11-03T21:31:42.002Z" }, + { url = "https://files.pythonhosted.org/packages/84/bd/9ce9f629fcb714ffc2c3faf62b6766ecb7a585e1e885eb699bcf130a5209/regex-2025.11.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a12ab1f5c29b4e93db518f5e3872116b7e9b1646c9f9f426f777b50d44a09e8c", size = 803501, upload-time = "2025-11-03T21:31:43.815Z" }, + { url = "https://files.pythonhosted.org/packages/7c/0f/8dc2e4349d8e877283e6edd6c12bdcebc20f03744e86f197ab6e4492bf08/regex-2025.11.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:7521684c8c7c4f6e88e35ec89680ee1aa8358d3f09d27dfbdf62c446f5d4c695", size = 787842, upload-time = "2025-11-03T21:31:45.353Z" }, + { url = "https://files.pythonhosted.org/packages/f9/73/cff02702960bc185164d5619c0c62a2f598a6abff6695d391b096237d4ab/regex-2025.11.3-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:7fe6e5440584e94cc4b3f5f4d98a25e29ca12dccf8873679a635638349831b98", size = 858519, upload-time = "2025-11-03T21:31:46.814Z" }, + { url = "https://files.pythonhosted.org/packages/61/83/0e8d1ae71e15bc1dc36231c90b46ee35f9d52fab2e226b0e039e7ea9c10a/regex-2025.11.3-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:8e026094aa12b43f4fd74576714e987803a315c76edb6b098b9809db5de58f74", size = 850611, upload-time = "2025-11-03T21:31:48.289Z" }, + { url = "https://files.pythonhosted.org/packages/c8/f5/70a5cdd781dcfaa12556f2955bf170cd603cb1c96a1827479f8faea2df97/regex-2025.11.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:435bbad13e57eb5606a68443af62bed3556de2f46deb9f7d4237bc2f1c9fb3a0", size = 789759, upload-time = "2025-11-03T21:31:49.759Z" }, + { url = "https://files.pythonhosted.org/packages/59/9b/7c29be7903c318488983e7d97abcf8ebd3830e4c956c4c540005fcfb0462/regex-2025.11.3-cp312-cp312-win32.whl", hash = "sha256:3839967cf4dc4b985e1570fd8d91078f0c519f30491c60f9ac42a8db039be204", size = 266194, upload-time = "2025-11-03T21:31:51.53Z" }, + { url = "https://files.pythonhosted.org/packages/1a/67/3b92df89f179d7c367be654ab5626ae311cb28f7d5c237b6bb976cd5fbbb/regex-2025.11.3-cp312-cp312-win_amd64.whl", hash = "sha256:e721d1b46e25c481dc5ded6f4b3f66c897c58d2e8cfdf77bbced84339108b0b9", size = 277069, upload-time = "2025-11-03T21:31:53.151Z" }, + { url = "https://files.pythonhosted.org/packages/d7/55/85ba4c066fe5094d35b249c3ce8df0ba623cfd35afb22d6764f23a52a1c5/regex-2025.11.3-cp312-cp312-win_arm64.whl", hash = "sha256:64350685ff08b1d3a6fff33f45a9ca183dc1d58bbfe4981604e70ec9801bbc26", size = 270330, upload-time = "2025-11-03T21:31:54.514Z" }, + { url = "https://files.pythonhosted.org/packages/e1/a7/dda24ebd49da46a197436ad96378f17df30ceb40e52e859fc42cac45b850/regex-2025.11.3-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:c1e448051717a334891f2b9a620fe36776ebf3dd8ec46a0b877c8ae69575feb4", size = 489081, upload-time = "2025-11-03T21:31:55.9Z" }, + { url = "https://files.pythonhosted.org/packages/19/22/af2dc751aacf88089836aa088a1a11c4f21a04707eb1b0478e8e8fb32847/regex-2025.11.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:9b5aca4d5dfd7fbfbfbdaf44850fcc7709a01146a797536a8f84952e940cca76", size = 291123, upload-time = "2025-11-03T21:31:57.758Z" }, + { url = "https://files.pythonhosted.org/packages/a3/88/1a3ea5672f4b0a84802ee9891b86743438e7c04eb0b8f8c4e16a42375327/regex-2025.11.3-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:04d2765516395cf7dda331a244a3282c0f5ae96075f728629287dfa6f76ba70a", size = 288814, upload-time = "2025-11-03T21:32:01.12Z" }, + { url = "https://files.pythonhosted.org/packages/fb/8c/f5987895bf42b8ddeea1b315c9fedcfe07cadee28b9c98cf50d00adcb14d/regex-2025.11.3-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5d9903ca42bfeec4cebedba8022a7c97ad2aab22e09573ce9976ba01b65e4361", size = 798592, upload-time = "2025-11-03T21:32:03.006Z" }, + { url = "https://files.pythonhosted.org/packages/99/2a/6591ebeede78203fa77ee46a1c36649e02df9eaa77a033d1ccdf2fcd5d4e/regex-2025.11.3-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:639431bdc89d6429f6721625e8129413980ccd62e9d3f496be618a41d205f160", size = 864122, upload-time = "2025-11-03T21:32:04.553Z" }, + { url = "https://files.pythonhosted.org/packages/94/d6/be32a87cf28cf8ed064ff281cfbd49aefd90242a83e4b08b5a86b38e8eb4/regex-2025.11.3-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:f117efad42068f9715677c8523ed2be1518116d1c49b1dd17987716695181efe", size = 912272, upload-time = "2025-11-03T21:32:06.148Z" }, + { url = "https://files.pythonhosted.org/packages/62/11/9bcef2d1445665b180ac7f230406ad80671f0fc2a6ffb93493b5dd8cd64c/regex-2025.11.3-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4aecb6f461316adf9f1f0f6a4a1a3d79e045f9b71ec76055a791affa3b285850", size = 803497, upload-time = "2025-11-03T21:32:08.162Z" }, + { url = "https://files.pythonhosted.org/packages/e5/a7/da0dc273d57f560399aa16d8a68ae7f9b57679476fc7ace46501d455fe84/regex-2025.11.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:3b3a5f320136873cc5561098dfab677eea139521cb9a9e8db98b7e64aef44cbc", size = 787892, upload-time = "2025-11-03T21:32:09.769Z" }, + { url = "https://files.pythonhosted.org/packages/da/4b/732a0c5a9736a0b8d6d720d4945a2f1e6f38f87f48f3173559f53e8d5d82/regex-2025.11.3-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:75fa6f0056e7efb1f42a1c34e58be24072cb9e61a601340cc1196ae92326a4f9", size = 858462, upload-time = "2025-11-03T21:32:11.769Z" }, + { url = "https://files.pythonhosted.org/packages/0c/f5/a2a03df27dc4c2d0c769220f5110ba8c4084b0bfa9ab0f9b4fcfa3d2b0fc/regex-2025.11.3-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:dbe6095001465294f13f1adcd3311e50dd84e5a71525f20a10bd16689c61ce0b", size = 850528, upload-time = "2025-11-03T21:32:13.906Z" }, + { url = "https://files.pythonhosted.org/packages/d6/09/e1cd5bee3841c7f6eb37d95ca91cdee7100b8f88b81e41c2ef426910891a/regex-2025.11.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:454d9b4ae7881afbc25015b8627c16d88a597479b9dea82b8c6e7e2e07240dc7", size = 789866, upload-time = "2025-11-03T21:32:15.748Z" }, + { url = "https://files.pythonhosted.org/packages/eb/51/702f5ea74e2a9c13d855a6a85b7f80c30f9e72a95493260193c07f3f8d74/regex-2025.11.3-cp313-cp313-win32.whl", hash = "sha256:28ba4d69171fc6e9896337d4fc63a43660002b7da53fc15ac992abcf3410917c", size = 266189, upload-time = "2025-11-03T21:32:17.493Z" }, + { url = "https://files.pythonhosted.org/packages/8b/00/6e29bb314e271a743170e53649db0fdb8e8ff0b64b4f425f5602f4eb9014/regex-2025.11.3-cp313-cp313-win_amd64.whl", hash = "sha256:bac4200befe50c670c405dc33af26dad5a3b6b255dd6c000d92fe4629f9ed6a5", size = 277054, upload-time = "2025-11-03T21:32:19.042Z" }, + { url = "https://files.pythonhosted.org/packages/25/f1/b156ff9f2ec9ac441710764dda95e4edaf5f36aca48246d1eea3f1fd96ec/regex-2025.11.3-cp313-cp313-win_arm64.whl", hash = "sha256:2292cd5a90dab247f9abe892ac584cb24f0f54680c73fcb4a7493c66c2bf2467", size = 270325, upload-time = "2025-11-03T21:32:21.338Z" }, + { url = "https://files.pythonhosted.org/packages/20/28/fd0c63357caefe5680b8ea052131acbd7f456893b69cc2a90cc3e0dc90d4/regex-2025.11.3-cp313-cp313t-macosx_10_13_universal2.whl", hash = "sha256:1eb1ebf6822b756c723e09f5186473d93236c06c579d2cc0671a722d2ab14281", size = 491984, upload-time = "2025-11-03T21:32:23.466Z" }, + { url = "https://files.pythonhosted.org/packages/df/ec/7014c15626ab46b902b3bcc4b28a7bae46d8f281fc7ea9c95e22fcaaa917/regex-2025.11.3-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:1e00ec2970aab10dc5db34af535f21fcf32b4a31d99e34963419636e2f85ae39", size = 292673, upload-time = "2025-11-03T21:32:25.034Z" }, + { url = "https://files.pythonhosted.org/packages/23/ab/3b952ff7239f20d05f1f99e9e20188513905f218c81d52fb5e78d2bf7634/regex-2025.11.3-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:a4cb042b615245d5ff9b3794f56be4138b5adc35a4166014d31d1814744148c7", size = 291029, upload-time = "2025-11-03T21:32:26.528Z" }, + { url = "https://files.pythonhosted.org/packages/21/7e/3dc2749fc684f455f162dcafb8a187b559e2614f3826877d3844a131f37b/regex-2025.11.3-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:44f264d4bf02f3176467d90b294d59bf1db9fe53c141ff772f27a8b456b2a9ed", size = 807437, upload-time = "2025-11-03T21:32:28.363Z" }, + { url = "https://files.pythonhosted.org/packages/1b/0b/d529a85ab349c6a25d1ca783235b6e3eedf187247eab536797021f7126c6/regex-2025.11.3-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:7be0277469bf3bd7a34a9c57c1b6a724532a0d235cd0dc4e7f4316f982c28b19", size = 873368, upload-time = "2025-11-03T21:32:30.4Z" }, + { url = "https://files.pythonhosted.org/packages/7d/18/2d868155f8c9e3e9d8f9e10c64e9a9f496bb8f7e037a88a8bed26b435af6/regex-2025.11.3-cp313-cp313t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0d31e08426ff4b5b650f68839f5af51a92a5b51abd8554a60c2fbc7c71f25d0b", size = 914921, upload-time = "2025-11-03T21:32:32.123Z" }, + { url = "https://files.pythonhosted.org/packages/2d/71/9d72ff0f354fa783fe2ba913c8734c3b433b86406117a8db4ea2bf1c7a2f/regex-2025.11.3-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e43586ce5bd28f9f285a6e729466841368c4a0353f6fd08d4ce4630843d3648a", size = 812708, upload-time = "2025-11-03T21:32:34.305Z" }, + { url = "https://files.pythonhosted.org/packages/e7/19/ce4bf7f5575c97f82b6e804ffb5c4e940c62609ab2a0d9538d47a7fdf7d4/regex-2025.11.3-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:0f9397d561a4c16829d4e6ff75202c1c08b68a3bdbfe29dbfcdb31c9830907c6", size = 795472, upload-time = "2025-11-03T21:32:36.364Z" }, + { url = "https://files.pythonhosted.org/packages/03/86/fd1063a176ffb7b2315f9a1b08d17b18118b28d9df163132615b835a26ee/regex-2025.11.3-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:dd16e78eb18ffdb25ee33a0682d17912e8cc8a770e885aeee95020046128f1ce", size = 868341, upload-time = "2025-11-03T21:32:38.042Z" }, + { url = "https://files.pythonhosted.org/packages/12/43/103fb2e9811205e7386366501bc866a164a0430c79dd59eac886a2822950/regex-2025.11.3-cp313-cp313t-musllinux_1_2_s390x.whl", hash = "sha256:ffcca5b9efe948ba0661e9df0fa50d2bc4b097c70b9810212d6b62f05d83b2dd", size = 854666, upload-time = "2025-11-03T21:32:40.079Z" }, + { url = "https://files.pythonhosted.org/packages/7d/22/e392e53f3869b75804762c7c848bd2dd2abf2b70fb0e526f58724638bd35/regex-2025.11.3-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:c56b4d162ca2b43318ac671c65bd4d563e841a694ac70e1a976ac38fcf4ca1d2", size = 799473, upload-time = "2025-11-03T21:32:42.148Z" }, + { url = "https://files.pythonhosted.org/packages/4f/f9/8bd6b656592f925b6845fcbb4d57603a3ac2fb2373344ffa1ed70aa6820a/regex-2025.11.3-cp313-cp313t-win32.whl", hash = "sha256:9ddc42e68114e161e51e272f667d640f97e84a2b9ef14b7477c53aac20c2d59a", size = 268792, upload-time = "2025-11-03T21:32:44.13Z" }, + { url = "https://files.pythonhosted.org/packages/e5/87/0e7d603467775ff65cd2aeabf1b5b50cc1c3708556a8b849a2fa4dd1542b/regex-2025.11.3-cp313-cp313t-win_amd64.whl", hash = "sha256:7a7c7fdf755032ffdd72c77e3d8096bdcb0eb92e89e17571a196f03d88b11b3c", size = 280214, upload-time = "2025-11-03T21:32:45.853Z" }, + { url = "https://files.pythonhosted.org/packages/8d/d0/2afc6f8e94e2b64bfb738a7c2b6387ac1699f09f032d363ed9447fd2bb57/regex-2025.11.3-cp313-cp313t-win_arm64.whl", hash = "sha256:df9eb838c44f570283712e7cff14c16329a9f0fb19ca492d21d4b7528ee6821e", size = 271469, upload-time = "2025-11-03T21:32:48.026Z" }, + { url = "https://files.pythonhosted.org/packages/31/e9/f6e13de7e0983837f7b6d238ad9458800a874bf37c264f7923e63409944c/regex-2025.11.3-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:9697a52e57576c83139d7c6f213d64485d3df5bf84807c35fa409e6c970801c6", size = 489089, upload-time = "2025-11-03T21:32:50.027Z" }, + { url = "https://files.pythonhosted.org/packages/a3/5c/261f4a262f1fa65141c1b74b255988bd2fa020cc599e53b080667d591cfc/regex-2025.11.3-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:e18bc3f73bd41243c9b38a6d9f2366cd0e0137a9aebe2d8ff76c5b67d4c0a3f4", size = 291059, upload-time = "2025-11-03T21:32:51.682Z" }, + { url = "https://files.pythonhosted.org/packages/8e/57/f14eeb7f072b0e9a5a090d1712741fd8f214ec193dba773cf5410108bb7d/regex-2025.11.3-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:61a08bcb0ec14ff4e0ed2044aad948d0659604f824cbd50b55e30b0ec6f09c73", size = 288900, upload-time = "2025-11-03T21:32:53.569Z" }, + { url = "https://files.pythonhosted.org/packages/3c/6b/1d650c45e99a9b327586739d926a1cd4e94666b1bd4af90428b36af66dc7/regex-2025.11.3-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c9c30003b9347c24bcc210958c5d167b9e4f9be786cb380a7d32f14f9b84674f", size = 799010, upload-time = "2025-11-03T21:32:55.222Z" }, + { url = "https://files.pythonhosted.org/packages/99/ee/d66dcbc6b628ce4e3f7f0cbbb84603aa2fc0ffc878babc857726b8aab2e9/regex-2025.11.3-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:4e1e592789704459900728d88d41a46fe3969b82ab62945560a31732ffc19a6d", size = 864893, upload-time = "2025-11-03T21:32:57.239Z" }, + { url = "https://files.pythonhosted.org/packages/bf/2d/f238229f1caba7ac87a6c4153d79947fb0261415827ae0f77c304260c7d3/regex-2025.11.3-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:6538241f45eb5a25aa575dbba1069ad786f68a4f2773a29a2bd3dd1f9de787be", size = 911522, upload-time = "2025-11-03T21:32:59.274Z" }, + { url = "https://files.pythonhosted.org/packages/bd/3d/22a4eaba214a917c80e04f6025d26143690f0419511e0116508e24b11c9b/regex-2025.11.3-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bce22519c989bb72a7e6b36a199384c53db7722fe669ba891da75907fe3587db", size = 803272, upload-time = "2025-11-03T21:33:01.393Z" }, + { url = "https://files.pythonhosted.org/packages/84/b1/03188f634a409353a84b5ef49754b97dbcc0c0f6fd6c8ede505a8960a0a4/regex-2025.11.3-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:66d559b21d3640203ab9075797a55165d79017520685fb407b9234d72ab63c62", size = 787958, upload-time = "2025-11-03T21:33:03.379Z" }, + { url = "https://files.pythonhosted.org/packages/99/6a/27d072f7fbf6fadd59c64d210305e1ff865cc3b78b526fd147db768c553b/regex-2025.11.3-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:669dcfb2e38f9e8c69507bace46f4889e3abbfd9b0c29719202883c0a603598f", size = 859289, upload-time = "2025-11-03T21:33:05.374Z" }, + { url = "https://files.pythonhosted.org/packages/9a/70/1b3878f648e0b6abe023172dacb02157e685564853cc363d9961bcccde4e/regex-2025.11.3-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:32f74f35ff0f25a5021373ac61442edcb150731fbaa28286bbc8bb1582c89d02", size = 850026, upload-time = "2025-11-03T21:33:07.131Z" }, + { url = "https://files.pythonhosted.org/packages/dd/d5/68e25559b526b8baab8e66839304ede68ff6727237a47727d240006bd0ff/regex-2025.11.3-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:e6c7a21dffba883234baefe91bc3388e629779582038f75d2a5be918e250f0ed", size = 789499, upload-time = "2025-11-03T21:33:09.141Z" }, + { url = "https://files.pythonhosted.org/packages/fc/df/43971264857140a350910d4e33df725e8c94dd9dee8d2e4729fa0d63d49e/regex-2025.11.3-cp314-cp314-win32.whl", hash = "sha256:795ea137b1d809eb6836b43748b12634291c0ed55ad50a7d72d21edf1cd565c4", size = 271604, upload-time = "2025-11-03T21:33:10.9Z" }, + { url = "https://files.pythonhosted.org/packages/01/6f/9711b57dc6894a55faf80a4c1b5aa4f8649805cb9c7aef46f7d27e2b9206/regex-2025.11.3-cp314-cp314-win_amd64.whl", hash = "sha256:9f95fbaa0ee1610ec0fc6b26668e9917a582ba80c52cc6d9ada15e30aa9ab9ad", size = 280320, upload-time = "2025-11-03T21:33:12.572Z" }, + { url = "https://files.pythonhosted.org/packages/f1/7e/f6eaa207d4377481f5e1775cdeb5a443b5a59b392d0065f3417d31d80f87/regex-2025.11.3-cp314-cp314-win_arm64.whl", hash = "sha256:dfec44d532be4c07088c3de2876130ff0fbeeacaa89a137decbbb5f665855a0f", size = 273372, upload-time = "2025-11-03T21:33:14.219Z" }, + { url = "https://files.pythonhosted.org/packages/c3/06/49b198550ee0f5e4184271cee87ba4dfd9692c91ec55289e6282f0f86ccf/regex-2025.11.3-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:ba0d8a5d7f04f73ee7d01d974d47c5834f8a1b0224390e4fe7c12a3a92a78ecc", size = 491985, upload-time = "2025-11-03T21:33:16.555Z" }, + { url = "https://files.pythonhosted.org/packages/ce/bf/abdafade008f0b1c9da10d934034cb670432d6cf6cbe38bbb53a1cfd6cf8/regex-2025.11.3-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:442d86cf1cfe4faabf97db7d901ef58347efd004934da045c745e7b5bd57ac49", size = 292669, upload-time = "2025-11-03T21:33:18.32Z" }, + { url = "https://files.pythonhosted.org/packages/f9/ef/0c357bb8edbd2ad8e273fcb9e1761bc37b8acbc6e1be050bebd6475f19c1/regex-2025.11.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:fd0a5e563c756de210bb964789b5abe4f114dacae9104a47e1a649b910361536", size = 291030, upload-time = "2025-11-03T21:33:20.048Z" }, + { url = "https://files.pythonhosted.org/packages/79/06/edbb67257596649b8fb088d6aeacbcb248ac195714b18a65e018bf4c0b50/regex-2025.11.3-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:bf3490bcbb985a1ae97b2ce9ad1c0f06a852d5b19dde9b07bdf25bf224248c95", size = 807674, upload-time = "2025-11-03T21:33:21.797Z" }, + { url = "https://files.pythonhosted.org/packages/f4/d9/ad4deccfce0ea336296bd087f1a191543bb99ee1c53093dcd4c64d951d00/regex-2025.11.3-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:3809988f0a8b8c9dcc0f92478d6501fac7200b9ec56aecf0ec21f4a2ec4b6009", size = 873451, upload-time = "2025-11-03T21:33:23.741Z" }, + { url = "https://files.pythonhosted.org/packages/13/75/a55a4724c56ef13e3e04acaab29df26582f6978c000ac9cd6810ad1f341f/regex-2025.11.3-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:f4ff94e58e84aedb9c9fce66d4ef9f27a190285b451420f297c9a09f2b9abee9", size = 914980, upload-time = "2025-11-03T21:33:25.999Z" }, + { url = "https://files.pythonhosted.org/packages/67/1e/a1657ee15bd9116f70d4a530c736983eed997b361e20ecd8f5ca3759d5c5/regex-2025.11.3-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7eb542fd347ce61e1321b0a6b945d5701528dca0cd9759c2e3bb8bd57e47964d", size = 812852, upload-time = "2025-11-03T21:33:27.852Z" }, + { url = "https://files.pythonhosted.org/packages/b8/6f/f7516dde5506a588a561d296b2d0044839de06035bb486b326065b4c101e/regex-2025.11.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:d6c2d5919075a1f2e413c00b056ea0c2f065b3f5fe83c3d07d325ab92dce51d6", size = 795566, upload-time = "2025-11-03T21:33:32.364Z" }, + { url = "https://files.pythonhosted.org/packages/d9/dd/3d10b9e170cc16fb34cb2cef91513cf3df65f440b3366030631b2984a264/regex-2025.11.3-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:3f8bf11a4827cc7ce5a53d4ef6cddd5ad25595d3c1435ef08f76825851343154", size = 868463, upload-time = "2025-11-03T21:33:34.459Z" }, + { url = "https://files.pythonhosted.org/packages/f5/8e/935e6beff1695aa9085ff83195daccd72acc82c81793df480f34569330de/regex-2025.11.3-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:22c12d837298651e5550ac1d964e4ff57c3f56965fc1812c90c9fb2028eaf267", size = 854694, upload-time = "2025-11-03T21:33:36.793Z" }, + { url = "https://files.pythonhosted.org/packages/92/12/10650181a040978b2f5720a6a74d44f841371a3d984c2083fc1752e4acf6/regex-2025.11.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:62ba394a3dda9ad41c7c780f60f6e4a70988741415ae96f6d1bf6c239cf01379", size = 799691, upload-time = "2025-11-03T21:33:39.079Z" }, + { url = "https://files.pythonhosted.org/packages/67/90/8f37138181c9a7690e7e4cb388debbd389342db3c7381d636d2875940752/regex-2025.11.3-cp314-cp314t-win32.whl", hash = "sha256:4bf146dca15cdd53224a1bf46d628bd7590e4a07fbb69e720d561aea43a32b38", size = 274583, upload-time = "2025-11-03T21:33:41.302Z" }, + { url = "https://files.pythonhosted.org/packages/8f/cd/867f5ec442d56beb56f5f854f40abcfc75e11d10b11fdb1869dd39c63aaf/regex-2025.11.3-cp314-cp314t-win_amd64.whl", hash = "sha256:adad1a1bcf1c9e76346e091d22d23ac54ef28e1365117d99521631078dfec9de", size = 284286, upload-time = "2025-11-03T21:33:43.324Z" }, + { url = "https://files.pythonhosted.org/packages/20/31/32c0c4610cbc070362bf1d2e4ea86d1ea29014d400a6d6c2486fcfd57766/regex-2025.11.3-cp314-cp314t-win_arm64.whl", hash = "sha256:c54f768482cef41e219720013cd05933b6f971d9562544d691c68699bf2b6801", size = 274741, upload-time = "2025-11-03T21:33:45.557Z" }, +] + +[[package]] +name = "remote-pdb" +version = "2.1.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/e4/b5/4944cac06fd9fc4a2e168313ec220aa25ed96ce83947b63eea5b4045b22d/remote-pdb-2.1.0.tar.gz", hash = "sha256:2d70c6f41e0eabf0165e8f1be58f82aa7a605aaeab8f2aefeb9ce246431091c1", size = 22295, upload-time = "2020-07-24T13:31:32.985Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/71/c5/d208c66344bb785d800adb61aef512290d3473052b9e7697890f0547aff2/remote_pdb-2.1.0-py2.py3-none-any.whl", hash = "sha256:94f73a92ac1248cf16189211011f97096bdada8a7baac8c79372663bbb57b5d0", size = 6304, upload-time = "2020-07-24T13:31:31.535Z" }, +] + +[[package]] +name = "requests" +version = "2.32.5" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "certifi" }, + { name = "charset-normalizer" }, + { name = "idna" }, + { name = "urllib3" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/c9/74/b3ff8e6c8446842c3f5c837e9c3dfcfe2018ea6ecef224c710c85ef728f4/requests-2.32.5.tar.gz", hash = "sha256:dbba0bac56e100853db0ea71b82b4dfd5fe2bf6d3754a8893c3af500cec7d7cf", size = 134517, upload-time = "2025-08-18T20:46:02.573Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/1e/db/4254e3eabe8020b458f1a747140d32277ec7a271daf1d235b70dc0b4e6e3/requests-2.32.5-py3-none-any.whl", hash = "sha256:2462f94637a34fd532264295e186976db0f5d453d1cdd31473c85a6a161affb6", size = 64738, upload-time = "2025-08-18T20:46:00.542Z" }, +] + +[[package]] +name = "rich" +version = "14.2.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "markdown-it-py" }, + { name = "pygments" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/fb/d2/8920e102050a0de7bfabeb4c4614a49248cf8d5d7a8d01885fbb24dc767a/rich-14.2.0.tar.gz", hash = "sha256:73ff50c7c0c1c77c8243079283f4edb376f0f6442433aecb8ce7e6d0b92d1fe4", size = 219990, upload-time = "2025-10-09T14:16:53.064Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/25/7a/b0178788f8dc6cafce37a212c99565fa1fe7872c70c6c9c1e1a372d9d88f/rich-14.2.0-py3-none-any.whl", hash = "sha256:76bc51fe2e57d2b1be1f96c524b890b816e334ab4c1e45888799bfaab0021edd", size = 243393, upload-time = "2025-10-09T14:16:51.245Z" }, +] + +[[package]] +name = "rpds-py" +version = "0.30.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/20/af/3f2f423103f1113b36230496629986e0ef7e199d2aa8392452b484b38ced/rpds_py-0.30.0.tar.gz", hash = "sha256:dd8ff7cf90014af0c0f787eea34794ebf6415242ee1d6fa91eaba725cc441e84", size = 69469, upload-time = "2025-11-30T20:24:38.837Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/06/0c/0c411a0ec64ccb6d104dcabe0e713e05e153a9a2c3c2bd2b32ce412166fe/rpds_py-0.30.0-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:679ae98e00c0e8d68a7fda324e16b90fd5260945b45d3b824c892cec9eea3288", size = 370490, upload-time = "2025-11-30T20:21:33.256Z" }, + { url = "https://files.pythonhosted.org/packages/19/6a/4ba3d0fb7297ebae71171822554abe48d7cab29c28b8f9f2c04b79988c05/rpds_py-0.30.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:4cc2206b76b4f576934f0ed374b10d7ca5f457858b157ca52064bdfc26b9fc00", size = 359751, upload-time = "2025-11-30T20:21:34.591Z" }, + { url = "https://files.pythonhosted.org/packages/cd/7c/e4933565ef7f7a0818985d87c15d9d273f1a649afa6a52ea35ad011195ea/rpds_py-0.30.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:389a2d49eded1896c3d48b0136ead37c48e221b391c052fba3f4055c367f60a6", size = 389696, upload-time = "2025-11-30T20:21:36.122Z" }, + { url = "https://files.pythonhosted.org/packages/5e/01/6271a2511ad0815f00f7ed4390cf2567bec1d4b1da39e2c27a41e6e3b4de/rpds_py-0.30.0-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:32c8528634e1bf7121f3de08fa85b138f4e0dc47657866630611b03967f041d7", size = 403136, upload-time = "2025-11-30T20:21:37.728Z" }, + { url = "https://files.pythonhosted.org/packages/55/64/c857eb7cd7541e9b4eee9d49c196e833128a55b89a9850a9c9ac33ccf897/rpds_py-0.30.0-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f207f69853edd6f6700b86efb84999651baf3789e78a466431df1331608e5324", size = 524699, upload-time = "2025-11-30T20:21:38.92Z" }, + { url = "https://files.pythonhosted.org/packages/9c/ed/94816543404078af9ab26159c44f9e98e20fe47e2126d5d32c9d9948d10a/rpds_py-0.30.0-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:67b02ec25ba7a9e8fa74c63b6ca44cf5707f2fbfadae3ee8e7494297d56aa9df", size = 412022, upload-time = "2025-11-30T20:21:40.407Z" }, + { url = "https://files.pythonhosted.org/packages/61/b5/707f6cf0066a6412aacc11d17920ea2e19e5b2f04081c64526eb35b5c6e7/rpds_py-0.30.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0c0e95f6819a19965ff420f65578bacb0b00f251fefe2c8b23347c37174271f3", size = 390522, upload-time = "2025-11-30T20:21:42.17Z" }, + { url = "https://files.pythonhosted.org/packages/13/4e/57a85fda37a229ff4226f8cbcf09f2a455d1ed20e802ce5b2b4a7f5ed053/rpds_py-0.30.0-cp310-cp310-manylinux_2_31_riscv64.whl", hash = "sha256:a452763cc5198f2f98898eb98f7569649fe5da666c2dc6b5ddb10fde5a574221", size = 404579, upload-time = "2025-11-30T20:21:43.769Z" }, + { url = "https://files.pythonhosted.org/packages/f9/da/c9339293513ec680a721e0e16bf2bac3db6e5d7e922488de471308349bba/rpds_py-0.30.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:e0b65193a413ccc930671c55153a03ee57cecb49e6227204b04fae512eb657a7", size = 421305, upload-time = "2025-11-30T20:21:44.994Z" }, + { url = "https://files.pythonhosted.org/packages/f9/be/522cb84751114f4ad9d822ff5a1aa3c98006341895d5f084779b99596e5c/rpds_py-0.30.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:858738e9c32147f78b3ac24dc0edb6610000e56dc0f700fd5f651d0a0f0eb9ff", size = 572503, upload-time = "2025-11-30T20:21:46.91Z" }, + { url = "https://files.pythonhosted.org/packages/a2/9b/de879f7e7ceddc973ea6e4629e9b380213a6938a249e94b0cdbcc325bb66/rpds_py-0.30.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:da279aa314f00acbb803da1e76fa18666778e8a8f83484fba94526da5de2cba7", size = 598322, upload-time = "2025-11-30T20:21:48.709Z" }, + { url = "https://files.pythonhosted.org/packages/48/ac/f01fc22efec3f37d8a914fc1b2fb9bcafd56a299edbe96406f3053edea5a/rpds_py-0.30.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:7c64d38fb49b6cdeda16ab49e35fe0da2e1e9b34bc38bd78386530f218b37139", size = 560792, upload-time = "2025-11-30T20:21:50.024Z" }, + { url = "https://files.pythonhosted.org/packages/e2/da/4e2b19d0f131f35b6146425f846563d0ce036763e38913d917187307a671/rpds_py-0.30.0-cp310-cp310-win32.whl", hash = "sha256:6de2a32a1665b93233cde140ff8b3467bdb9e2af2b91079f0333a0974d12d464", size = 221901, upload-time = "2025-11-30T20:21:51.32Z" }, + { url = "https://files.pythonhosted.org/packages/96/cb/156d7a5cf4f78a7cc571465d8aec7a3c447c94f6749c5123f08438bcf7bc/rpds_py-0.30.0-cp310-cp310-win_amd64.whl", hash = "sha256:1726859cd0de969f88dc8673bdd954185b9104e05806be64bcd87badbe313169", size = 235823, upload-time = "2025-11-30T20:21:52.505Z" }, + { url = "https://files.pythonhosted.org/packages/4d/6e/f964e88b3d2abee2a82c1ac8366da848fce1c6d834dc2132c3fda3970290/rpds_py-0.30.0-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:a2bffea6a4ca9f01b3f8e548302470306689684e61602aa3d141e34da06cf425", size = 370157, upload-time = "2025-11-30T20:21:53.789Z" }, + { url = "https://files.pythonhosted.org/packages/94/ba/24e5ebb7c1c82e74c4e4f33b2112a5573ddc703915b13a073737b59b86e0/rpds_py-0.30.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:dc4f992dfe1e2bc3ebc7444f6c7051b4bc13cd8e33e43511e8ffd13bf407010d", size = 359676, upload-time = "2025-11-30T20:21:55.475Z" }, + { url = "https://files.pythonhosted.org/packages/84/86/04dbba1b087227747d64d80c3b74df946b986c57af0a9f0c98726d4d7a3b/rpds_py-0.30.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:422c3cb9856d80b09d30d2eb255d0754b23e090034e1deb4083f8004bd0761e4", size = 389938, upload-time = "2025-11-30T20:21:57.079Z" }, + { url = "https://files.pythonhosted.org/packages/42/bb/1463f0b1722b7f45431bdd468301991d1328b16cffe0b1c2918eba2c4eee/rpds_py-0.30.0-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:07ae8a593e1c3c6b82ca3292efbe73c30b61332fd612e05abee07c79359f292f", size = 402932, upload-time = "2025-11-30T20:21:58.47Z" }, + { url = "https://files.pythonhosted.org/packages/99/ee/2520700a5c1f2d76631f948b0736cdf9b0acb25abd0ca8e889b5c62ac2e3/rpds_py-0.30.0-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:12f90dd7557b6bd57f40abe7747e81e0c0b119bef015ea7726e69fe550e394a4", size = 525830, upload-time = "2025-11-30T20:21:59.699Z" }, + { url = "https://files.pythonhosted.org/packages/e0/ad/bd0331f740f5705cc555a5e17fdf334671262160270962e69a2bdef3bf76/rpds_py-0.30.0-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:99b47d6ad9a6da00bec6aabe5a6279ecd3c06a329d4aa4771034a21e335c3a97", size = 412033, upload-time = "2025-11-30T20:22:00.991Z" }, + { url = "https://files.pythonhosted.org/packages/f8/1e/372195d326549bb51f0ba0f2ecb9874579906b97e08880e7a65c3bef1a99/rpds_py-0.30.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:33f559f3104504506a44bb666b93a33f5d33133765b0c216a5bf2f1e1503af89", size = 390828, upload-time = "2025-11-30T20:22:02.723Z" }, + { url = "https://files.pythonhosted.org/packages/ab/2b/d88bb33294e3e0c76bc8f351a3721212713629ffca1700fa94979cb3eae8/rpds_py-0.30.0-cp311-cp311-manylinux_2_31_riscv64.whl", hash = "sha256:946fe926af6e44f3697abbc305ea168c2c31d3e3ef1058cf68f379bf0335a78d", size = 404683, upload-time = "2025-11-30T20:22:04.367Z" }, + { url = "https://files.pythonhosted.org/packages/50/32/c759a8d42bcb5289c1fac697cd92f6fe01a018dd937e62ae77e0e7f15702/rpds_py-0.30.0-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:495aeca4b93d465efde585977365187149e75383ad2684f81519f504f5c13038", size = 421583, upload-time = "2025-11-30T20:22:05.814Z" }, + { url = "https://files.pythonhosted.org/packages/2b/81/e729761dbd55ddf5d84ec4ff1f47857f4374b0f19bdabfcf929164da3e24/rpds_py-0.30.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:d9a0ca5da0386dee0655b4ccdf46119df60e0f10da268d04fe7cc87886872ba7", size = 572496, upload-time = "2025-11-30T20:22:07.713Z" }, + { url = "https://files.pythonhosted.org/packages/14/f6/69066a924c3557c9c30baa6ec3a0aa07526305684c6f86c696b08860726c/rpds_py-0.30.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:8d6d1cc13664ec13c1b84241204ff3b12f9bb82464b8ad6e7a5d3486975c2eed", size = 598669, upload-time = "2025-11-30T20:22:09.312Z" }, + { url = "https://files.pythonhosted.org/packages/5f/48/905896b1eb8a05630d20333d1d8ffd162394127b74ce0b0784ae04498d32/rpds_py-0.30.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3896fa1be39912cf0757753826bc8bdc8ca331a28a7c4ae46b7a21280b06bb85", size = 561011, upload-time = "2025-11-30T20:22:11.309Z" }, + { url = "https://files.pythonhosted.org/packages/22/16/cd3027c7e279d22e5eb431dd3c0fbc677bed58797fe7581e148f3f68818b/rpds_py-0.30.0-cp311-cp311-win32.whl", hash = "sha256:55f66022632205940f1827effeff17c4fa7ae1953d2b74a8581baaefb7d16f8c", size = 221406, upload-time = "2025-11-30T20:22:13.101Z" }, + { url = "https://files.pythonhosted.org/packages/fa/5b/e7b7aa136f28462b344e652ee010d4de26ee9fd16f1bfd5811f5153ccf89/rpds_py-0.30.0-cp311-cp311-win_amd64.whl", hash = "sha256:a51033ff701fca756439d641c0ad09a41d9242fa69121c7d8769604a0a629825", size = 236024, upload-time = "2025-11-30T20:22:14.853Z" }, + { url = "https://files.pythonhosted.org/packages/14/a6/364bba985e4c13658edb156640608f2c9e1d3ea3c81b27aa9d889fff0e31/rpds_py-0.30.0-cp311-cp311-win_arm64.whl", hash = "sha256:47b0ef6231c58f506ef0b74d44e330405caa8428e770fec25329ed2cb971a229", size = 229069, upload-time = "2025-11-30T20:22:16.577Z" }, + { url = "https://files.pythonhosted.org/packages/03/e7/98a2f4ac921d82f33e03f3835f5bf3a4a40aa1bfdc57975e74a97b2b4bdd/rpds_py-0.30.0-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:a161f20d9a43006833cd7068375a94d035714d73a172b681d8881820600abfad", size = 375086, upload-time = "2025-11-30T20:22:17.93Z" }, + { url = "https://files.pythonhosted.org/packages/4d/a1/bca7fd3d452b272e13335db8d6b0b3ecde0f90ad6f16f3328c6fb150c889/rpds_py-0.30.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:6abc8880d9d036ecaafe709079969f56e876fcf107f7a8e9920ba6d5a3878d05", size = 359053, upload-time = "2025-11-30T20:22:19.297Z" }, + { url = "https://files.pythonhosted.org/packages/65/1c/ae157e83a6357eceff62ba7e52113e3ec4834a84cfe07fa4b0757a7d105f/rpds_py-0.30.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ca28829ae5f5d569bb62a79512c842a03a12576375d5ece7d2cadf8abe96ec28", size = 390763, upload-time = "2025-11-30T20:22:21.661Z" }, + { url = "https://files.pythonhosted.org/packages/d4/36/eb2eb8515e2ad24c0bd43c3ee9cd74c33f7ca6430755ccdb240fd3144c44/rpds_py-0.30.0-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:a1010ed9524c73b94d15919ca4d41d8780980e1765babf85f9a2f90d247153dd", size = 408951, upload-time = "2025-11-30T20:22:23.408Z" }, + { url = "https://files.pythonhosted.org/packages/d6/65/ad8dc1784a331fabbd740ef6f71ce2198c7ed0890dab595adb9ea2d775a1/rpds_py-0.30.0-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f8d1736cfb49381ba528cd5baa46f82fdc65c06e843dab24dd70b63d09121b3f", size = 514622, upload-time = "2025-11-30T20:22:25.16Z" }, + { url = "https://files.pythonhosted.org/packages/63/8e/0cfa7ae158e15e143fe03993b5bcd743a59f541f5952e1546b1ac1b5fd45/rpds_py-0.30.0-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:d948b135c4693daff7bc2dcfc4ec57237a29bd37e60c2fabf5aff2bbacf3e2f1", size = 414492, upload-time = "2025-11-30T20:22:26.505Z" }, + { url = "https://files.pythonhosted.org/packages/60/1b/6f8f29f3f995c7ffdde46a626ddccd7c63aefc0efae881dc13b6e5d5bb16/rpds_py-0.30.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:47f236970bccb2233267d89173d3ad2703cd36a0e2a6e92d0560d333871a3d23", size = 394080, upload-time = "2025-11-30T20:22:27.934Z" }, + { url = "https://files.pythonhosted.org/packages/6d/d5/a266341051a7a3ca2f4b750a3aa4abc986378431fc2da508c5034d081b70/rpds_py-0.30.0-cp312-cp312-manylinux_2_31_riscv64.whl", hash = "sha256:2e6ecb5a5bcacf59c3f912155044479af1d0b6681280048b338b28e364aca1f6", size = 408680, upload-time = "2025-11-30T20:22:29.341Z" }, + { url = "https://files.pythonhosted.org/packages/10/3b/71b725851df9ab7a7a4e33cf36d241933da66040d195a84781f49c50490c/rpds_py-0.30.0-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:a8fa71a2e078c527c3e9dc9fc5a98c9db40bcc8a92b4e8858e36d329f8684b51", size = 423589, upload-time = "2025-11-30T20:22:31.469Z" }, + { url = "https://files.pythonhosted.org/packages/00/2b/e59e58c544dc9bd8bd8384ecdb8ea91f6727f0e37a7131baeff8d6f51661/rpds_py-0.30.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:73c67f2db7bc334e518d097c6d1e6fed021bbc9b7d678d6cc433478365d1d5f5", size = 573289, upload-time = "2025-11-30T20:22:32.997Z" }, + { url = "https://files.pythonhosted.org/packages/da/3e/a18e6f5b460893172a7d6a680e86d3b6bc87a54c1f0b03446a3c8c7b588f/rpds_py-0.30.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:5ba103fb455be00f3b1c2076c9d4264bfcb037c976167a6047ed82f23153f02e", size = 599737, upload-time = "2025-11-30T20:22:34.419Z" }, + { url = "https://files.pythonhosted.org/packages/5c/e2/714694e4b87b85a18e2c243614974413c60aa107fd815b8cbc42b873d1d7/rpds_py-0.30.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:7cee9c752c0364588353e627da8a7e808a66873672bcb5f52890c33fd965b394", size = 563120, upload-time = "2025-11-30T20:22:35.903Z" }, + { url = "https://files.pythonhosted.org/packages/6f/ab/d5d5e3bcedb0a77f4f613706b750e50a5a3ba1c15ccd3665ecc636c968fd/rpds_py-0.30.0-cp312-cp312-win32.whl", hash = "sha256:1ab5b83dbcf55acc8b08fc62b796ef672c457b17dbd7820a11d6c52c06839bdf", size = 223782, upload-time = "2025-11-30T20:22:37.271Z" }, + { url = "https://files.pythonhosted.org/packages/39/3b/f786af9957306fdc38a74cef405b7b93180f481fb48453a114bb6465744a/rpds_py-0.30.0-cp312-cp312-win_amd64.whl", hash = "sha256:a090322ca841abd453d43456ac34db46e8b05fd9b3b4ac0c78bcde8b089f959b", size = 240463, upload-time = "2025-11-30T20:22:39.021Z" }, + { url = "https://files.pythonhosted.org/packages/f3/d2/b91dc748126c1559042cfe41990deb92c4ee3e2b415f6b5234969ffaf0cc/rpds_py-0.30.0-cp312-cp312-win_arm64.whl", hash = "sha256:669b1805bd639dd2989b281be2cfd951c6121b65e729d9b843e9639ef1fd555e", size = 230868, upload-time = "2025-11-30T20:22:40.493Z" }, + { url = "https://files.pythonhosted.org/packages/ed/dc/d61221eb88ff410de3c49143407f6f3147acf2538c86f2ab7ce65ae7d5f9/rpds_py-0.30.0-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:f83424d738204d9770830d35290ff3273fbb02b41f919870479fab14b9d303b2", size = 374887, upload-time = "2025-11-30T20:22:41.812Z" }, + { url = "https://files.pythonhosted.org/packages/fd/32/55fb50ae104061dbc564ef15cc43c013dc4a9f4527a1f4d99baddf56fe5f/rpds_py-0.30.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:e7536cd91353c5273434b4e003cbda89034d67e7710eab8761fd918ec6c69cf8", size = 358904, upload-time = "2025-11-30T20:22:43.479Z" }, + { url = "https://files.pythonhosted.org/packages/58/70/faed8186300e3b9bdd138d0273109784eea2396c68458ed580f885dfe7ad/rpds_py-0.30.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2771c6c15973347f50fece41fc447c054b7ac2ae0502388ce3b6738cd366e3d4", size = 389945, upload-time = "2025-11-30T20:22:44.819Z" }, + { url = "https://files.pythonhosted.org/packages/bd/a8/073cac3ed2c6387df38f71296d002ab43496a96b92c823e76f46b8af0543/rpds_py-0.30.0-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:0a59119fc6e3f460315fe9d08149f8102aa322299deaa5cab5b40092345c2136", size = 407783, upload-time = "2025-11-30T20:22:46.103Z" }, + { url = "https://files.pythonhosted.org/packages/77/57/5999eb8c58671f1c11eba084115e77a8899d6e694d2a18f69f0ba471ec8b/rpds_py-0.30.0-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:76fec018282b4ead0364022e3c54b60bf368b9d926877957a8624b58419169b7", size = 515021, upload-time = "2025-11-30T20:22:47.458Z" }, + { url = "https://files.pythonhosted.org/packages/e0/af/5ab4833eadc36c0a8ed2bc5c0de0493c04f6c06de223170bd0798ff98ced/rpds_py-0.30.0-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:692bef75a5525db97318e8cd061542b5a79812d711ea03dbc1f6f8dbb0c5f0d2", size = 414589, upload-time = "2025-11-30T20:22:48.872Z" }, + { url = "https://files.pythonhosted.org/packages/b7/de/f7192e12b21b9e9a68a6d0f249b4af3fdcdff8418be0767a627564afa1f1/rpds_py-0.30.0-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9027da1ce107104c50c81383cae773ef5c24d296dd11c99e2629dbd7967a20c6", size = 394025, upload-time = "2025-11-30T20:22:50.196Z" }, + { url = "https://files.pythonhosted.org/packages/91/c4/fc70cd0249496493500e7cc2de87504f5aa6509de1e88623431fec76d4b6/rpds_py-0.30.0-cp313-cp313-manylinux_2_31_riscv64.whl", hash = "sha256:9cf69cdda1f5968a30a359aba2f7f9aa648a9ce4b580d6826437f2b291cfc86e", size = 408895, upload-time = "2025-11-30T20:22:51.87Z" }, + { url = "https://files.pythonhosted.org/packages/58/95/d9275b05ab96556fefff73a385813eb66032e4c99f411d0795372d9abcea/rpds_py-0.30.0-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:a4796a717bf12b9da9d3ad002519a86063dcac8988b030e405704ef7d74d2d9d", size = 422799, upload-time = "2025-11-30T20:22:53.341Z" }, + { url = "https://files.pythonhosted.org/packages/06/c1/3088fc04b6624eb12a57eb814f0d4997a44b0d208d6cace713033ff1a6ba/rpds_py-0.30.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:5d4c2aa7c50ad4728a094ebd5eb46c452e9cb7edbfdb18f9e1221f597a73e1e7", size = 572731, upload-time = "2025-11-30T20:22:54.778Z" }, + { url = "https://files.pythonhosted.org/packages/d8/42/c612a833183b39774e8ac8fecae81263a68b9583ee343db33ab571a7ce55/rpds_py-0.30.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:ba81a9203d07805435eb06f536d95a266c21e5b2dfbf6517748ca40c98d19e31", size = 599027, upload-time = "2025-11-30T20:22:56.212Z" }, + { url = "https://files.pythonhosted.org/packages/5f/60/525a50f45b01d70005403ae0e25f43c0384369ad24ffe46e8d9068b50086/rpds_py-0.30.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:945dccface01af02675628334f7cf49c2af4c1c904748efc5cf7bbdf0b579f95", size = 563020, upload-time = "2025-11-30T20:22:58.2Z" }, + { url = "https://files.pythonhosted.org/packages/0b/5d/47c4655e9bcd5ca907148535c10e7d489044243cc9941c16ed7cd53be91d/rpds_py-0.30.0-cp313-cp313-win32.whl", hash = "sha256:b40fb160a2db369a194cb27943582b38f79fc4887291417685f3ad693c5a1d5d", size = 223139, upload-time = "2025-11-30T20:23:00.209Z" }, + { url = "https://files.pythonhosted.org/packages/f2/e1/485132437d20aa4d3e1d8b3fb5a5e65aa8139f1e097080c2a8443201742c/rpds_py-0.30.0-cp313-cp313-win_amd64.whl", hash = "sha256:806f36b1b605e2d6a72716f321f20036b9489d29c51c91f4dd29a3e3afb73b15", size = 240224, upload-time = "2025-11-30T20:23:02.008Z" }, + { url = "https://files.pythonhosted.org/packages/24/95/ffd128ed1146a153d928617b0ef673960130be0009c77d8fbf0abe306713/rpds_py-0.30.0-cp313-cp313-win_arm64.whl", hash = "sha256:d96c2086587c7c30d44f31f42eae4eac89b60dabbac18c7669be3700f13c3ce1", size = 230645, upload-time = "2025-11-30T20:23:03.43Z" }, + { url = "https://files.pythonhosted.org/packages/ff/1b/b10de890a0def2a319a2626334a7f0ae388215eb60914dbac8a3bae54435/rpds_py-0.30.0-cp313-cp313t-macosx_10_12_x86_64.whl", hash = "sha256:eb0b93f2e5c2189ee831ee43f156ed34e2a89a78a66b98cadad955972548be5a", size = 364443, upload-time = "2025-11-30T20:23:04.878Z" }, + { url = "https://files.pythonhosted.org/packages/0d/bf/27e39f5971dc4f305a4fb9c672ca06f290f7c4e261c568f3dea16a410d47/rpds_py-0.30.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:922e10f31f303c7c920da8981051ff6d8c1a56207dbdf330d9047f6d30b70e5e", size = 353375, upload-time = "2025-11-30T20:23:06.342Z" }, + { url = "https://files.pythonhosted.org/packages/40/58/442ada3bba6e8e6615fc00483135c14a7538d2ffac30e2d933ccf6852232/rpds_py-0.30.0-cp313-cp313t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cdc62c8286ba9bf7f47befdcea13ea0e26bf294bda99758fd90535cbaf408000", size = 383850, upload-time = "2025-11-30T20:23:07.825Z" }, + { url = "https://files.pythonhosted.org/packages/14/14/f59b0127409a33c6ef6f5c1ebd5ad8e32d7861c9c7adfa9a624fc3889f6c/rpds_py-0.30.0-cp313-cp313t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:47f9a91efc418b54fb8190a6b4aa7813a23fb79c51f4bb84e418f5476c38b8db", size = 392812, upload-time = "2025-11-30T20:23:09.228Z" }, + { url = "https://files.pythonhosted.org/packages/b3/66/e0be3e162ac299b3a22527e8913767d869e6cc75c46bd844aa43fb81ab62/rpds_py-0.30.0-cp313-cp313t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1f3587eb9b17f3789ad50824084fa6f81921bbf9a795826570bda82cb3ed91f2", size = 517841, upload-time = "2025-11-30T20:23:11.186Z" }, + { url = "https://files.pythonhosted.org/packages/3d/55/fa3b9cf31d0c963ecf1ba777f7cf4b2a2c976795ac430d24a1f43d25a6ba/rpds_py-0.30.0-cp313-cp313t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:39c02563fc592411c2c61d26b6c5fe1e51eaa44a75aa2c8735ca88b0d9599daa", size = 408149, upload-time = "2025-11-30T20:23:12.864Z" }, + { url = "https://files.pythonhosted.org/packages/60/ca/780cf3b1a32b18c0f05c441958d3758f02544f1d613abf9488cd78876378/rpds_py-0.30.0-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:51a1234d8febafdfd33a42d97da7a43f5dcb120c1060e352a3fbc0c6d36e2083", size = 383843, upload-time = "2025-11-30T20:23:14.638Z" }, + { url = "https://files.pythonhosted.org/packages/82/86/d5f2e04f2aa6247c613da0c1dd87fcd08fa17107e858193566048a1e2f0a/rpds_py-0.30.0-cp313-cp313t-manylinux_2_31_riscv64.whl", hash = "sha256:eb2c4071ab598733724c08221091e8d80e89064cd472819285a9ab0f24bcedb9", size = 396507, upload-time = "2025-11-30T20:23:16.105Z" }, + { url = "https://files.pythonhosted.org/packages/4b/9a/453255d2f769fe44e07ea9785c8347edaf867f7026872e76c1ad9f7bed92/rpds_py-0.30.0-cp313-cp313t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:6bdfdb946967d816e6adf9a3d8201bfad269c67efe6cefd7093ef959683c8de0", size = 414949, upload-time = "2025-11-30T20:23:17.539Z" }, + { url = "https://files.pythonhosted.org/packages/a3/31/622a86cdc0c45d6df0e9ccb6becdba5074735e7033c20e401a6d9d0e2ca0/rpds_py-0.30.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:c77afbd5f5250bf27bf516c7c4a016813eb2d3e116139aed0096940c5982da94", size = 565790, upload-time = "2025-11-30T20:23:19.029Z" }, + { url = "https://files.pythonhosted.org/packages/1c/5d/15bbf0fb4a3f58a3b1c67855ec1efcc4ceaef4e86644665fff03e1b66d8d/rpds_py-0.30.0-cp313-cp313t-musllinux_1_2_i686.whl", hash = "sha256:61046904275472a76c8c90c9ccee9013d70a6d0f73eecefd38c1ae7c39045a08", size = 590217, upload-time = "2025-11-30T20:23:20.885Z" }, + { url = "https://files.pythonhosted.org/packages/6d/61/21b8c41f68e60c8cc3b2e25644f0e3681926020f11d06ab0b78e3c6bbff1/rpds_py-0.30.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:4c5f36a861bc4b7da6516dbdf302c55313afa09b81931e8280361a4f6c9a2d27", size = 555806, upload-time = "2025-11-30T20:23:22.488Z" }, + { url = "https://files.pythonhosted.org/packages/f9/39/7e067bb06c31de48de3eb200f9fc7c58982a4d3db44b07e73963e10d3be9/rpds_py-0.30.0-cp313-cp313t-win32.whl", hash = "sha256:3d4a69de7a3e50ffc214ae16d79d8fbb0922972da0356dcf4d0fdca2878559c6", size = 211341, upload-time = "2025-11-30T20:23:24.449Z" }, + { url = "https://files.pythonhosted.org/packages/0a/4d/222ef0b46443cf4cf46764d9c630f3fe4abaa7245be9417e56e9f52b8f65/rpds_py-0.30.0-cp313-cp313t-win_amd64.whl", hash = "sha256:f14fc5df50a716f7ece6a80b6c78bb35ea2ca47c499e422aa4463455dd96d56d", size = 225768, upload-time = "2025-11-30T20:23:25.908Z" }, + { url = "https://files.pythonhosted.org/packages/86/81/dad16382ebbd3d0e0328776d8fd7ca94220e4fa0798d1dc5e7da48cb3201/rpds_py-0.30.0-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:68f19c879420aa08f61203801423f6cd5ac5f0ac4ac82a2368a9fcd6a9a075e0", size = 362099, upload-time = "2025-11-30T20:23:27.316Z" }, + { url = "https://files.pythonhosted.org/packages/2b/60/19f7884db5d5603edf3c6bce35408f45ad3e97e10007df0e17dd57af18f8/rpds_py-0.30.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:ec7c4490c672c1a0389d319b3a9cfcd098dcdc4783991553c332a15acf7249be", size = 353192, upload-time = "2025-11-30T20:23:29.151Z" }, + { url = "https://files.pythonhosted.org/packages/bf/c4/76eb0e1e72d1a9c4703c69607cec123c29028bff28ce41588792417098ac/rpds_py-0.30.0-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f251c812357a3fed308d684a5079ddfb9d933860fc6de89f2b7ab00da481e65f", size = 384080, upload-time = "2025-11-30T20:23:30.785Z" }, + { url = "https://files.pythonhosted.org/packages/72/87/87ea665e92f3298d1b26d78814721dc39ed8d2c74b86e83348d6b48a6f31/rpds_py-0.30.0-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:ac98b175585ecf4c0348fd7b29c3864bda53b805c773cbf7bfdaffc8070c976f", size = 394841, upload-time = "2025-11-30T20:23:32.209Z" }, + { url = "https://files.pythonhosted.org/packages/77/ad/7783a89ca0587c15dcbf139b4a8364a872a25f861bdb88ed99f9b0dec985/rpds_py-0.30.0-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:3e62880792319dbeb7eb866547f2e35973289e7d5696c6e295476448f5b63c87", size = 516670, upload-time = "2025-11-30T20:23:33.742Z" }, + { url = "https://files.pythonhosted.org/packages/5b/3c/2882bdac942bd2172f3da574eab16f309ae10a3925644e969536553cb4ee/rpds_py-0.30.0-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:4e7fc54e0900ab35d041b0601431b0a0eb495f0851a0639b6ef90f7741b39a18", size = 408005, upload-time = "2025-11-30T20:23:35.253Z" }, + { url = "https://files.pythonhosted.org/packages/ce/81/9a91c0111ce1758c92516a3e44776920b579d9a7c09b2b06b642d4de3f0f/rpds_py-0.30.0-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:47e77dc9822d3ad616c3d5759ea5631a75e5809d5a28707744ef79d7a1bcfcad", size = 382112, upload-time = "2025-11-30T20:23:36.842Z" }, + { url = "https://files.pythonhosted.org/packages/cf/8e/1da49d4a107027e5fbc64daeab96a0706361a2918da10cb41769244b805d/rpds_py-0.30.0-cp314-cp314-manylinux_2_31_riscv64.whl", hash = "sha256:b4dc1a6ff022ff85ecafef7979a2c6eb423430e05f1165d6688234e62ba99a07", size = 399049, upload-time = "2025-11-30T20:23:38.343Z" }, + { url = "https://files.pythonhosted.org/packages/df/5a/7ee239b1aa48a127570ec03becbb29c9d5a9eb092febbd1699d567cae859/rpds_py-0.30.0-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:4559c972db3a360808309e06a74628b95eaccbf961c335c8fe0d590cf587456f", size = 415661, upload-time = "2025-11-30T20:23:40.263Z" }, + { url = "https://files.pythonhosted.org/packages/70/ea/caa143cf6b772f823bc7929a45da1fa83569ee49b11d18d0ada7f5ee6fd6/rpds_py-0.30.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:0ed177ed9bded28f8deb6ab40c183cd1192aa0de40c12f38be4d59cd33cb5c65", size = 565606, upload-time = "2025-11-30T20:23:42.186Z" }, + { url = "https://files.pythonhosted.org/packages/64/91/ac20ba2d69303f961ad8cf55bf7dbdb4763f627291ba3d0d7d67333cced9/rpds_py-0.30.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:ad1fa8db769b76ea911cb4e10f049d80bf518c104f15b3edb2371cc65375c46f", size = 591126, upload-time = "2025-11-30T20:23:44.086Z" }, + { url = "https://files.pythonhosted.org/packages/21/20/7ff5f3c8b00c8a95f75985128c26ba44503fb35b8e0259d812766ea966c7/rpds_py-0.30.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:46e83c697b1f1c72b50e5ee5adb4353eef7406fb3f2043d64c33f20ad1c2fc53", size = 553371, upload-time = "2025-11-30T20:23:46.004Z" }, + { url = "https://files.pythonhosted.org/packages/72/c7/81dadd7b27c8ee391c132a6b192111ca58d866577ce2d9b0ca157552cce0/rpds_py-0.30.0-cp314-cp314-win32.whl", hash = "sha256:ee454b2a007d57363c2dfd5b6ca4a5d7e2c518938f8ed3b706e37e5d470801ed", size = 215298, upload-time = "2025-11-30T20:23:47.696Z" }, + { url = "https://files.pythonhosted.org/packages/3e/d2/1aaac33287e8cfb07aab2e6b8ac1deca62f6f65411344f1433c55e6f3eb8/rpds_py-0.30.0-cp314-cp314-win_amd64.whl", hash = "sha256:95f0802447ac2d10bcc69f6dc28fe95fdf17940367b21d34e34c737870758950", size = 228604, upload-time = "2025-11-30T20:23:49.501Z" }, + { url = "https://files.pythonhosted.org/packages/e8/95/ab005315818cc519ad074cb7784dae60d939163108bd2b394e60dc7b5461/rpds_py-0.30.0-cp314-cp314-win_arm64.whl", hash = "sha256:613aa4771c99f03346e54c3f038e4cc574ac09a3ddfb0e8878487335e96dead6", size = 222391, upload-time = "2025-11-30T20:23:50.96Z" }, + { url = "https://files.pythonhosted.org/packages/9e/68/154fe0194d83b973cdedcdcc88947a2752411165930182ae41d983dcefa6/rpds_py-0.30.0-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:7e6ecfcb62edfd632e56983964e6884851786443739dbfe3582947e87274f7cb", size = 364868, upload-time = "2025-11-30T20:23:52.494Z" }, + { url = "https://files.pythonhosted.org/packages/83/69/8bbc8b07ec854d92a8b75668c24d2abcb1719ebf890f5604c61c9369a16f/rpds_py-0.30.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:a1d0bc22a7cdc173fedebb73ef81e07faef93692b8c1ad3733b67e31e1b6e1b8", size = 353747, upload-time = "2025-11-30T20:23:54.036Z" }, + { url = "https://files.pythonhosted.org/packages/ab/00/ba2e50183dbd9abcce9497fa5149c62b4ff3e22d338a30d690f9af970561/rpds_py-0.30.0-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0d08f00679177226c4cb8c5265012eea897c8ca3b93f429e546600c971bcbae7", size = 383795, upload-time = "2025-11-30T20:23:55.556Z" }, + { url = "https://files.pythonhosted.org/packages/05/6f/86f0272b84926bcb0e4c972262f54223e8ecc556b3224d281e6598fc9268/rpds_py-0.30.0-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:5965af57d5848192c13534f90f9dd16464f3c37aaf166cc1da1cae1fd5a34898", size = 393330, upload-time = "2025-11-30T20:23:57.033Z" }, + { url = "https://files.pythonhosted.org/packages/cb/e9/0e02bb2e6dc63d212641da45df2b0bf29699d01715913e0d0f017ee29438/rpds_py-0.30.0-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9a4e86e34e9ab6b667c27f3211ca48f73dba7cd3d90f8d5b11be56e5dbc3fb4e", size = 518194, upload-time = "2025-11-30T20:23:58.637Z" }, + { url = "https://files.pythonhosted.org/packages/ee/ca/be7bca14cf21513bdf9c0606aba17d1f389ea2b6987035eb4f62bd923f25/rpds_py-0.30.0-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e5d3e6b26f2c785d65cc25ef1e5267ccbe1b069c5c21b8cc724efee290554419", size = 408340, upload-time = "2025-11-30T20:24:00.2Z" }, + { url = "https://files.pythonhosted.org/packages/c2/c7/736e00ebf39ed81d75544c0da6ef7b0998f8201b369acf842f9a90dc8fce/rpds_py-0.30.0-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:626a7433c34566535b6e56a1b39a7b17ba961e97ce3b80ec62e6f1312c025551", size = 383765, upload-time = "2025-11-30T20:24:01.759Z" }, + { url = "https://files.pythonhosted.org/packages/4a/3f/da50dfde9956aaf365c4adc9533b100008ed31aea635f2b8d7b627e25b49/rpds_py-0.30.0-cp314-cp314t-manylinux_2_31_riscv64.whl", hash = "sha256:acd7eb3f4471577b9b5a41baf02a978e8bdeb08b4b355273994f8b87032000a8", size = 396834, upload-time = "2025-11-30T20:24:03.687Z" }, + { url = "https://files.pythonhosted.org/packages/4e/00/34bcc2565b6020eab2623349efbdec810676ad571995911f1abdae62a3a0/rpds_py-0.30.0-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:fe5fa731a1fa8a0a56b0977413f8cacac1768dad38d16b3a296712709476fbd5", size = 415470, upload-time = "2025-11-30T20:24:05.232Z" }, + { url = "https://files.pythonhosted.org/packages/8c/28/882e72b5b3e6f718d5453bd4d0d9cf8df36fddeb4ddbbab17869d5868616/rpds_py-0.30.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:74a3243a411126362712ee1524dfc90c650a503502f135d54d1b352bd01f2404", size = 565630, upload-time = "2025-11-30T20:24:06.878Z" }, + { url = "https://files.pythonhosted.org/packages/3b/97/04a65539c17692de5b85c6e293520fd01317fd878ea1995f0367d4532fb1/rpds_py-0.30.0-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:3e8eeb0544f2eb0d2581774be4c3410356eba189529a6b3e36bbbf9696175856", size = 591148, upload-time = "2025-11-30T20:24:08.445Z" }, + { url = "https://files.pythonhosted.org/packages/85/70/92482ccffb96f5441aab93e26c4d66489eb599efdcf96fad90c14bbfb976/rpds_py-0.30.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:dbd936cde57abfee19ab3213cf9c26be06d60750e60a8e4dd85d1ab12c8b1f40", size = 556030, upload-time = "2025-11-30T20:24:10.956Z" }, + { url = "https://files.pythonhosted.org/packages/20/53/7c7e784abfa500a2b6b583b147ee4bb5a2b3747a9166bab52fec4b5b5e7d/rpds_py-0.30.0-cp314-cp314t-win32.whl", hash = "sha256:dc824125c72246d924f7f796b4f63c1e9dc810c7d9e2355864b3c3a73d59ade0", size = 211570, upload-time = "2025-11-30T20:24:12.735Z" }, + { url = "https://files.pythonhosted.org/packages/d0/02/fa464cdfbe6b26e0600b62c528b72d8608f5cc49f96b8d6e38c95d60c676/rpds_py-0.30.0-cp314-cp314t-win_amd64.whl", hash = "sha256:27f4b0e92de5bfbc6f86e43959e6edd1425c33b5e69aab0984a72047f2bcf1e3", size = 226532, upload-time = "2025-11-30T20:24:14.634Z" }, + { url = "https://files.pythonhosted.org/packages/69/71/3f34339ee70521864411f8b6992e7ab13ac30d8e4e3309e07c7361767d91/rpds_py-0.30.0-pp311-pypy311_pp73-macosx_10_12_x86_64.whl", hash = "sha256:c2262bdba0ad4fc6fb5545660673925c2d2a5d9e2e0fb603aad545427be0fc58", size = 372292, upload-time = "2025-11-30T20:24:16.537Z" }, + { url = "https://files.pythonhosted.org/packages/57/09/f183df9b8f2d66720d2ef71075c59f7e1b336bec7ee4c48f0a2b06857653/rpds_py-0.30.0-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:ee6af14263f25eedc3bb918a3c04245106a42dfd4f5c2285ea6f997b1fc3f89a", size = 362128, upload-time = "2025-11-30T20:24:18.086Z" }, + { url = "https://files.pythonhosted.org/packages/7a/68/5c2594e937253457342e078f0cc1ded3dd7b2ad59afdbf2d354869110a02/rpds_py-0.30.0-pp311-pypy311_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3adbb8179ce342d235c31ab8ec511e66c73faa27a47e076ccc92421add53e2bb", size = 391542, upload-time = "2025-11-30T20:24:20.092Z" }, + { url = "https://files.pythonhosted.org/packages/49/5c/31ef1afd70b4b4fbdb2800249f34c57c64beb687495b10aec0365f53dfc4/rpds_py-0.30.0-pp311-pypy311_pp73-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:250fa00e9543ac9b97ac258bd37367ff5256666122c2d0f2bc97577c60a1818c", size = 404004, upload-time = "2025-11-30T20:24:22.231Z" }, + { url = "https://files.pythonhosted.org/packages/e3/63/0cfbea38d05756f3440ce6534d51a491d26176ac045e2707adc99bb6e60a/rpds_py-0.30.0-pp311-pypy311_pp73-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9854cf4f488b3d57b9aaeb105f06d78e5529d3145b1e4a41750167e8c213c6d3", size = 527063, upload-time = "2025-11-30T20:24:24.302Z" }, + { url = "https://files.pythonhosted.org/packages/42/e6/01e1f72a2456678b0f618fc9a1a13f882061690893c192fcad9f2926553a/rpds_py-0.30.0-pp311-pypy311_pp73-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:993914b8e560023bc0a8bf742c5f303551992dcb85e247b1e5c7f4a7d145bda5", size = 413099, upload-time = "2025-11-30T20:24:25.916Z" }, + { url = "https://files.pythonhosted.org/packages/b8/25/8df56677f209003dcbb180765520c544525e3ef21ea72279c98b9aa7c7fb/rpds_py-0.30.0-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:58edca431fb9b29950807e301826586e5bbf24163677732429770a697ffe6738", size = 392177, upload-time = "2025-11-30T20:24:27.834Z" }, + { url = "https://files.pythonhosted.org/packages/4a/b4/0a771378c5f16f8115f796d1f437950158679bcd2a7c68cf251cfb00ed5b/rpds_py-0.30.0-pp311-pypy311_pp73-manylinux_2_31_riscv64.whl", hash = "sha256:dea5b552272a944763b34394d04577cf0f9bd013207bc32323b5a89a53cf9c2f", size = 406015, upload-time = "2025-11-30T20:24:29.457Z" }, + { url = "https://files.pythonhosted.org/packages/36/d8/456dbba0af75049dc6f63ff295a2f92766b9d521fa00de67a2bd6427d57a/rpds_py-0.30.0-pp311-pypy311_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:ba3af48635eb83d03f6c9735dfb21785303e73d22ad03d489e88adae6eab8877", size = 423736, upload-time = "2025-11-30T20:24:31.22Z" }, + { url = "https://files.pythonhosted.org/packages/13/64/b4d76f227d5c45a7e0b796c674fd81b0a6c4fbd48dc29271857d8219571c/rpds_py-0.30.0-pp311-pypy311_pp73-musllinux_1_2_aarch64.whl", hash = "sha256:dff13836529b921e22f15cb099751209a60009731a68519630a24d61f0b1b30a", size = 573981, upload-time = "2025-11-30T20:24:32.934Z" }, + { url = "https://files.pythonhosted.org/packages/20/91/092bacadeda3edf92bf743cc96a7be133e13a39cdbfd7b5082e7ab638406/rpds_py-0.30.0-pp311-pypy311_pp73-musllinux_1_2_i686.whl", hash = "sha256:1b151685b23929ab7beec71080a8889d4d6d9fa9a983d213f07121205d48e2c4", size = 599782, upload-time = "2025-11-30T20:24:35.169Z" }, + { url = "https://files.pythonhosted.org/packages/d1/b7/b95708304cd49b7b6f82fdd039f1748b66ec2b21d6a45180910802f1abf1/rpds_py-0.30.0-pp311-pypy311_pp73-musllinux_1_2_x86_64.whl", hash = "sha256:ac37f9f516c51e5753f27dfdef11a88330f04de2d564be3991384b2f3535d02e", size = 562191, upload-time = "2025-11-30T20:24:36.853Z" }, +] + +[[package]] +name = "ruff" +version = "0.14.4" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/df/55/cccfca45157a2031dcbb5a462a67f7cf27f8b37d4b3b1cd7438f0f5c1df6/ruff-0.14.4.tar.gz", hash = "sha256:f459a49fe1085a749f15414ca76f61595f1a2cc8778ed7c279b6ca2e1fd19df3", size = 5587844, upload-time = "2025-11-06T22:07:45.033Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/17/b9/67240254166ae1eaa38dec32265e9153ac53645a6c6670ed36ad00722af8/ruff-0.14.4-py3-none-linux_armv6l.whl", hash = "sha256:e6604613ffbcf2297cd5dcba0e0ac9bd0c11dc026442dfbb614504e87c349518", size = 12606781, upload-time = "2025-11-06T22:07:01.841Z" }, + { url = "https://files.pythonhosted.org/packages/46/c8/09b3ab245d8652eafe5256ab59718641429f68681ee713ff06c5c549f156/ruff-0.14.4-py3-none-macosx_10_12_x86_64.whl", hash = "sha256:d99c0b52b6f0598acede45ee78288e5e9b4409d1ce7f661f0fa36d4cbeadf9a4", size = 12946765, upload-time = "2025-11-06T22:07:05.858Z" }, + { url = "https://files.pythonhosted.org/packages/14/bb/1564b000219144bf5eed2359edc94c3590dd49d510751dad26202c18a17d/ruff-0.14.4-py3-none-macosx_11_0_arm64.whl", hash = "sha256:9358d490ec030f1b51d048a7fd6ead418ed0826daf6149e95e30aa67c168af33", size = 11928120, upload-time = "2025-11-06T22:07:08.023Z" }, + { url = "https://files.pythonhosted.org/packages/a3/92/d5f1770e9988cc0742fefaa351e840d9aef04ec24ae1be36f333f96d5704/ruff-0.14.4-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:81b40d27924f1f02dfa827b9c0712a13c0e4b108421665322218fc38caf615c2", size = 12370877, upload-time = "2025-11-06T22:07:10.015Z" }, + { url = "https://files.pythonhosted.org/packages/e2/29/e9282efa55f1973d109faf839a63235575519c8ad278cc87a182a366810e/ruff-0.14.4-py3-none-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:f5e649052a294fe00818650712083cddc6cc02744afaf37202c65df9ea52efa5", size = 12408538, upload-time = "2025-11-06T22:07:13.085Z" }, + { url = "https://files.pythonhosted.org/packages/8e/01/930ed6ecfce130144b32d77d8d69f5c610e6d23e6857927150adf5d7379a/ruff-0.14.4-py3-none-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:aa082a8f878deeba955531f975881828fd6afd90dfa757c2b0808aadb437136e", size = 13141942, upload-time = "2025-11-06T22:07:15.386Z" }, + { url = "https://files.pythonhosted.org/packages/6a/46/a9c89b42b231a9f487233f17a89cbef9d5acd538d9488687a02ad288fa6b/ruff-0.14.4-py3-none-manylinux_2_17_ppc64.manylinux2014_ppc64.whl", hash = "sha256:1043c6811c2419e39011890f14d0a30470f19d47d197c4858b2787dfa698f6c8", size = 14544306, upload-time = "2025-11-06T22:07:17.631Z" }, + { url = "https://files.pythonhosted.org/packages/78/96/9c6cf86491f2a6d52758b830b89b78c2ae61e8ca66b86bf5a20af73d20e6/ruff-0.14.4-py3-none-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a9f3a936ac27fb7c2a93e4f4b943a662775879ac579a433291a6f69428722649", size = 14210427, upload-time = "2025-11-06T22:07:19.832Z" }, + { url = "https://files.pythonhosted.org/packages/71/f4/0666fe7769a54f63e66404e8ff698de1dcde733e12e2fd1c9c6efb689cb5/ruff-0.14.4-py3-none-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:95643ffd209ce78bc113266b88fba3d39e0461f0cbc8b55fb92505030fb4a850", size = 13658488, upload-time = "2025-11-06T22:07:22.32Z" }, + { url = "https://files.pythonhosted.org/packages/ee/79/6ad4dda2cfd55e41ac9ed6d73ef9ab9475b1eef69f3a85957210c74ba12c/ruff-0.14.4-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:456daa2fa1021bc86ca857f43fe29d5d8b3f0e55e9f90c58c317c1dcc2afc7b5", size = 13354908, upload-time = "2025-11-06T22:07:24.347Z" }, + { url = "https://files.pythonhosted.org/packages/b5/60/f0b6990f740bb15c1588601d19d21bcc1bd5de4330a07222041678a8e04f/ruff-0.14.4-py3-none-manylinux_2_31_riscv64.whl", hash = "sha256:f911bba769e4a9f51af6e70037bb72b70b45a16db5ce73e1f72aefe6f6d62132", size = 13587803, upload-time = "2025-11-06T22:07:26.327Z" }, + { url = "https://files.pythonhosted.org/packages/c9/da/eaaada586f80068728338e0ef7f29ab3e4a08a692f92eb901a4f06bbff24/ruff-0.14.4-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:76158a7369b3979fa878612c623a7e5430c18b2fd1c73b214945c2d06337db67", size = 12279654, upload-time = "2025-11-06T22:07:28.46Z" }, + { url = "https://files.pythonhosted.org/packages/66/d4/b1d0e82cf9bf8aed10a6d45be47b3f402730aa2c438164424783ac88c0ed/ruff-0.14.4-py3-none-musllinux_1_2_armv7l.whl", hash = "sha256:f3b8f3b442d2b14c246e7aeca2e75915159e06a3540e2f4bed9f50d062d24469", size = 12357520, upload-time = "2025-11-06T22:07:31.468Z" }, + { url = "https://files.pythonhosted.org/packages/04/f4/53e2b42cc82804617e5c7950b7079d79996c27e99c4652131c6a1100657f/ruff-0.14.4-py3-none-musllinux_1_2_i686.whl", hash = "sha256:c62da9a06779deecf4d17ed04939ae8b31b517643b26370c3be1d26f3ef7dbde", size = 12719431, upload-time = "2025-11-06T22:07:33.831Z" }, + { url = "https://files.pythonhosted.org/packages/a2/94/80e3d74ed9a72d64e94a7b7706b1c1ebaa315ef2076fd33581f6a1cd2f95/ruff-0.14.4-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:5a443a83a1506c684e98acb8cb55abaf3ef725078be40237463dae4463366349", size = 13464394, upload-time = "2025-11-06T22:07:35.905Z" }, + { url = "https://files.pythonhosted.org/packages/54/1a/a49f071f04c42345c793d22f6cf5e0920095e286119ee53a64a3a3004825/ruff-0.14.4-py3-none-win32.whl", hash = "sha256:643b69cb63cd996f1fc7229da726d07ac307eae442dd8974dbc7cf22c1e18fff", size = 12493429, upload-time = "2025-11-06T22:07:38.43Z" }, + { url = "https://files.pythonhosted.org/packages/bc/22/e58c43e641145a2b670328fb98bc384e20679b5774258b1e540207580266/ruff-0.14.4-py3-none-win_amd64.whl", hash = "sha256:26673da283b96fe35fa0c939bf8411abec47111644aa9f7cfbd3c573fb125d2c", size = 13635380, upload-time = "2025-11-06T22:07:40.496Z" }, + { url = "https://files.pythonhosted.org/packages/30/bd/4168a751ddbbf43e86544b4de8b5c3b7be8d7167a2a5cb977d274e04f0a1/ruff-0.14.4-py3-none-win_arm64.whl", hash = "sha256:dd09c292479596b0e6fec8cd95c65c3a6dc68e9ad17b8f2382130f87ff6a75bb", size = 12663065, upload-time = "2025-11-06T22:07:42.603Z" }, +] + +[[package]] +name = "safehttpx" +version = "0.1.7" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "httpx" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/89/d1/4282284d9cf1ee873607a46442da977fc3c985059315ab23610be31d5885/safehttpx-0.1.7.tar.gz", hash = "sha256:db201c0978c41eddb8bb480f3eee59dd67304fdd91646035e9d9a720049a9d23", size = 10385, upload-time = "2025-10-24T18:30:09.783Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/2e/a3/0f0b7d78e2f1eb9e8e1afbff1d2bff8d60144aee17aca51c065b516743dd/safehttpx-0.1.7-py3-none-any.whl", hash = "sha256:c4f4a162db6993464d7ca3d7cc4af0ffc6515a606dfd220b9f82c6945d869cde", size = 8959, upload-time = "2025-10-24T18:30:08.733Z" }, +] + +[[package]] +name = "safetensors" +version = "0.7.0rc0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/88/44/8c66cad2fa18757c28e7164ccf664f281648148f03addfa0f8106213a0ee/safetensors-0.7.0rc0.tar.gz", hash = "sha256:8036070784341d33ef3f647a21822b332b7136a23cce7aa08b9f65604cc71da1", size = 208862, upload-time = "2025-10-30T14:09:27.599Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/48/67/c1471e8c84bfb5f66c14a2e36b9005dfc93b63f70a3754ad333c7b8512c1/safetensors-0.7.0rc0-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:8494ca841c8616fba2da59c22ac021ca5202dd6552c16d31e9129d8b85ea523a", size = 463526, upload-time = "2025-10-30T14:09:21.039Z" }, + { url = "https://files.pythonhosted.org/packages/94/c9/2fee63fb38bcf0f4fa0d1a615b103abf1c8519c6a660a02fe7b9eed13f36/safetensors-0.7.0rc0-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:1203c4b0a497dc70b76a11bb716b086cecfb426916a6a5103c3fef62ed73640e", size = 437244, upload-time = "2025-10-30T14:09:19.551Z" }, + { url = "https://files.pythonhosted.org/packages/89/bf/62acaa0ba7995c119de19feb65ed192769180d0390043c6223d41f057282/safetensors-0.7.0rc0-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c6194e95ae70a84d32e7fef7a9c877331166ca99dc6c8f8b1881de63ab57db80", size = 481695, upload-time = "2025-10-30T14:09:11.384Z" }, + { url = "https://files.pythonhosted.org/packages/10/15/b1cb7feb23c39d6df372a2ca8c9552cb928a06df2a8abea26932fec523d1/safetensors-0.7.0rc0-cp38-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4aff0b6f78ee14e42afe8f0bf02c71c8318f3000d00be5e7ed76521afd533dbc", size = 496610, upload-time = "2025-10-30T14:09:13.435Z" }, + { url = "https://files.pythonhosted.org/packages/b3/86/554e918d3779f0563471a2040886abc8f83d68219cd7f73ca68455b301c6/safetensors-0.7.0rc0-cp38-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1f61efc01a1118422a8cb3361b2fcd0ad6b7a862ed77242c6d628b8eb5555ab8", size = 625590, upload-time = "2025-10-30T14:09:14.875Z" }, + { url = "https://files.pythonhosted.org/packages/b6/67/321e7b2a4bbea618335aacb1152d5743f0b4c8526bd37fb7179f0eefc51c/safetensors-0.7.0rc0-cp38-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ab1534d6c49761dc1e8eeba0d4c0f489edb62c83416aada96f9d983c6281e1da", size = 529683, upload-time = "2025-10-30T14:09:15.915Z" }, + { url = "https://files.pythonhosted.org/packages/a0/9d/1ca95020814732c191709b73f454932e1ab939c4a1d716465c467d3f1057/safetensors-0.7.0rc0-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8767b077d5d1c8748cd80fdf6e72574cbcf1c5a52e47bfb8e278a5ea3f6a60d9", size = 502087, upload-time = "2025-10-30T14:09:18.555Z" }, + { url = "https://files.pythonhosted.org/packages/5d/c3/e24651bf7b4e1b71d7ed005ef5595882872834367a57bcd7021454438c61/safetensors-0.7.0rc0-cp38-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c96d3810437c58840fe4d73fa1f9873cc089d2ef7ced20d80d080561a0bde2cd", size = 529884, upload-time = "2025-10-30T14:09:17.478Z" }, + { url = "https://files.pythonhosted.org/packages/6f/05/d4d91d700f40cbba03cfa898bba405358e431f8ea5fe70eeafcaf1523964/safetensors-0.7.0rc0-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:0a77f87ab5ae2b6e1b1fae6bc298ebf542500e59a83bb1429c9945e74d2958d0", size = 661453, upload-time = "2025-10-30T14:09:22.392Z" }, + { url = "https://files.pythonhosted.org/packages/9a/24/c5e02bfffb52c19d0930c89e5a7368b4e1f901876521c316b66495729c25/safetensors-0.7.0rc0-cp38-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:8ab1608cf068aa07dbf395cfe92057efa5a23ccf3fe9e47bdae483516b03ea5a", size = 759928, upload-time = "2025-10-30T14:09:23.679Z" }, + { url = "https://files.pythonhosted.org/packages/f5/a8/b1f68f0b7b27f0420c673aaca22b59f30dd4946e91aae75d7136ee596a4f/safetensors-0.7.0rc0-cp38-abi3-musllinux_1_2_i686.whl", hash = "sha256:23af05b4d8563add047a1b65eb16fae65330d957b8635cc98ecd2864170fa196", size = 700440, upload-time = "2025-10-30T14:09:24.72Z" }, + { url = "https://files.pythonhosted.org/packages/d6/65/91afe9b066e7cd23f7486ca799b212183300418c02a95c4cc2d4a4d4a224/safetensors-0.7.0rc0-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:25f6719d213d6e6cbe1ca60079136469403de9593e4c3f3d39f349ffcfbc0bb3", size = 666522, upload-time = "2025-10-30T14:09:26.403Z" }, + { url = "https://files.pythonhosted.org/packages/68/83/4aa2f65f5116587b8168400c4f6ff711700cefa7351bebc345d1e5406c2d/safetensors-0.7.0rc0-cp38-abi3-win32.whl", hash = "sha256:e161d72c9e0823be1cdcf501fba652cd52fd066456bf5a9b36098a8c57bcbeb0", size = 319140, upload-time = "2025-10-30T14:09:30.244Z" }, + { url = "https://files.pythonhosted.org/packages/a0/10/ccd58fad34fe1cafd4ed3bc7f75706a12a2e0291b2454e738646f5153432/safetensors-0.7.0rc0-cp38-abi3-win_amd64.whl", hash = "sha256:5a48f71f153ebb9ddc52a8978dcbe9e374f765d182f503c87b2d6b4bbfd5fef9", size = 331266, upload-time = "2025-10-30T14:09:28.769Z" }, +] + +[[package]] +name = "scipy" +version = "1.14.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "numpy" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/62/11/4d44a1f274e002784e4dbdb81e0ea96d2de2d1045b2132d5af62cc31fd28/scipy-1.14.1.tar.gz", hash = "sha256:5a275584e726026a5699459aa72f828a610821006228e841b94275c4a7c08417", size = 58620554, upload-time = "2024-08-21T00:09:20.662Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/64/68/3bc0cfaf64ff507d82b1e5d5b64521df4c8bf7e22bc0b897827cbee9872c/scipy-1.14.1-cp310-cp310-macosx_10_13_x86_64.whl", hash = "sha256:b28d2ca4add7ac16ae8bb6632a3c86e4b9e4d52d3e34267f6e1b0c1f8d87e389", size = 39069598, upload-time = "2024-08-21T00:03:32.896Z" }, + { url = "https://files.pythonhosted.org/packages/43/a5/8d02f9c372790326ad405d94f04d4339482ec082455b9e6e288f7100513b/scipy-1.14.1-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:d0d2821003174de06b69e58cef2316a6622b60ee613121199cb2852a873f8cf3", size = 29879676, upload-time = "2024-08-21T00:03:38.844Z" }, + { url = "https://files.pythonhosted.org/packages/07/42/0e0bea9666fcbf2cb6ea0205db42c81b1f34d7b729ba251010edf9c80ebd/scipy-1.14.1-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:8bddf15838ba768bb5f5083c1ea012d64c9a444e16192762bd858f1e126196d0", size = 23088696, upload-time = "2024-08-21T00:03:43.583Z" }, + { url = "https://files.pythonhosted.org/packages/15/47/298ab6fef5ebf31b426560e978b8b8548421d4ed0bf99263e1eb44532306/scipy-1.14.1-cp310-cp310-macosx_14_0_x86_64.whl", hash = "sha256:97c5dddd5932bd2a1a31c927ba5e1463a53b87ca96b5c9bdf5dfd6096e27efc3", size = 25470699, upload-time = "2024-08-21T00:03:48.466Z" }, + { url = "https://files.pythonhosted.org/packages/d8/df/cdb6be5274bc694c4c22862ac3438cb04f360ed9df0aecee02ce0b798380/scipy-1.14.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2ff0a7e01e422c15739ecd64432743cf7aae2b03f3084288f399affcefe5222d", size = 35606631, upload-time = "2024-08-21T00:03:54.532Z" }, + { url = "https://files.pythonhosted.org/packages/47/78/b0c2c23880dd1e99e938ad49ccfb011ae353758a2dc5ed7ee59baff684c3/scipy-1.14.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8e32dced201274bf96899e6491d9ba3e9a5f6b336708656466ad0522d8528f69", size = 41178528, upload-time = "2024-08-21T00:04:00.862Z" }, + { url = "https://files.pythonhosted.org/packages/5d/aa/994b45c34b897637b853ec04334afa55a85650a0d11dacfa67232260fb0a/scipy-1.14.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:8426251ad1e4ad903a4514712d2fa8fdd5382c978010d1c6f5f37ef286a713ad", size = 42784535, upload-time = "2024-08-21T00:04:12.65Z" }, + { url = "https://files.pythonhosted.org/packages/e7/1c/8daa6df17a945cb1a2a1e3bae3c49643f7b3b94017ff01a4787064f03f84/scipy-1.14.1-cp310-cp310-win_amd64.whl", hash = "sha256:a49f6ed96f83966f576b33a44257d869756df6cf1ef4934f59dd58b25e0327e5", size = 44772117, upload-time = "2024-08-21T00:04:20.613Z" }, + { url = "https://files.pythonhosted.org/packages/b2/ab/070ccfabe870d9f105b04aee1e2860520460ef7ca0213172abfe871463b9/scipy-1.14.1-cp311-cp311-macosx_10_13_x86_64.whl", hash = "sha256:2da0469a4ef0ecd3693761acbdc20f2fdeafb69e6819cc081308cc978153c675", size = 39076999, upload-time = "2024-08-21T00:04:32.61Z" }, + { url = "https://files.pythonhosted.org/packages/a7/c5/02ac82f9bb8f70818099df7e86c3ad28dae64e1347b421d8e3adf26acab6/scipy-1.14.1-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:c0ee987efa6737242745f347835da2cc5bb9f1b42996a4d97d5c7ff7928cb6f2", size = 29894570, upload-time = "2024-08-21T00:04:37.938Z" }, + { url = "https://files.pythonhosted.org/packages/ed/05/7f03e680cc5249c4f96c9e4e845acde08eb1aee5bc216eff8a089baa4ddb/scipy-1.14.1-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:3a1b111fac6baec1c1d92f27e76511c9e7218f1695d61b59e05e0fe04dc59617", size = 23103567, upload-time = "2024-08-21T00:04:42.582Z" }, + { url = "https://files.pythonhosted.org/packages/5e/fc/9f1413bef53171f379d786aabc104d4abeea48ee84c553a3e3d8c9f96a9c/scipy-1.14.1-cp311-cp311-macosx_14_0_x86_64.whl", hash = "sha256:8475230e55549ab3f207bff11ebfc91c805dc3463ef62eda3ccf593254524ce8", size = 25499102, upload-time = "2024-08-21T00:04:47.467Z" }, + { url = "https://files.pythonhosted.org/packages/c2/4b/b44bee3c2ddc316b0159b3d87a3d467ef8d7edfd525e6f7364a62cd87d90/scipy-1.14.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:278266012eb69f4a720827bdd2dc54b2271c97d84255b2faaa8f161a158c3b37", size = 35586346, upload-time = "2024-08-21T00:04:53.872Z" }, + { url = "https://files.pythonhosted.org/packages/93/6b/701776d4bd6bdd9b629c387b5140f006185bd8ddea16788a44434376b98f/scipy-1.14.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fef8c87f8abfb884dac04e97824b61299880c43f4ce675dd2cbeadd3c9b466d2", size = 41165244, upload-time = "2024-08-21T00:05:00.489Z" }, + { url = "https://files.pythonhosted.org/packages/06/57/e6aa6f55729a8f245d8a6984f2855696c5992113a5dc789065020f8be753/scipy-1.14.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:b05d43735bb2f07d689f56f7b474788a13ed8adc484a85aa65c0fd931cf9ccd2", size = 42817917, upload-time = "2024-08-21T00:05:07.533Z" }, + { url = "https://files.pythonhosted.org/packages/ea/c2/5ecadc5fcccefaece775feadcd795060adf5c3b29a883bff0e678cfe89af/scipy-1.14.1-cp311-cp311-win_amd64.whl", hash = "sha256:716e389b694c4bb564b4fc0c51bc84d381735e0d39d3f26ec1af2556ec6aad94", size = 44781033, upload-time = "2024-08-21T00:05:14.297Z" }, + { url = "https://files.pythonhosted.org/packages/c0/04/2bdacc8ac6387b15db6faa40295f8bd25eccf33f1f13e68a72dc3c60a99e/scipy-1.14.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:631f07b3734d34aced009aaf6fedfd0eb3498a97e581c3b1e5f14a04164a456d", size = 39128781, upload-time = "2024-08-21T04:08:04.15Z" }, + { url = "https://files.pythonhosted.org/packages/c8/53/35b4d41f5fd42f5781dbd0dd6c05d35ba8aa75c84ecddc7d44756cd8da2e/scipy-1.14.1-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:af29a935803cc707ab2ed7791c44288a682f9c8107bc00f0eccc4f92c08d6e07", size = 29939542, upload-time = "2024-08-21T00:05:25.758Z" }, + { url = "https://files.pythonhosted.org/packages/66/67/6ef192e0e4d77b20cc33a01e743b00bc9e68fb83b88e06e636d2619a8767/scipy-1.14.1-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:2843f2d527d9eebec9a43e6b406fb7266f3af25a751aa91d62ff416f54170bc5", size = 23148375, upload-time = "2024-08-21T00:05:30.359Z" }, + { url = "https://files.pythonhosted.org/packages/f6/32/3a6dedd51d68eb7b8e7dc7947d5d841bcb699f1bf4463639554986f4d782/scipy-1.14.1-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:eb58ca0abd96911932f688528977858681a59d61a7ce908ffd355957f7025cfc", size = 25578573, upload-time = "2024-08-21T00:05:35.274Z" }, + { url = "https://files.pythonhosted.org/packages/f0/5a/efa92a58dc3a2898705f1dc9dbaf390ca7d4fba26d6ab8cfffb0c72f656f/scipy-1.14.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:30ac8812c1d2aab7131a79ba62933a2a76f582d5dbbc695192453dae67ad6310", size = 35319299, upload-time = "2024-08-21T00:05:40.956Z" }, + { url = "https://files.pythonhosted.org/packages/8e/ee/8a26858ca517e9c64f84b4c7734b89bda8e63bec85c3d2f432d225bb1886/scipy-1.14.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8f9ea80f2e65bdaa0b7627fb00cbeb2daf163caa015e59b7516395fe3bd1e066", size = 40849331, upload-time = "2024-08-21T00:05:47.53Z" }, + { url = "https://files.pythonhosted.org/packages/a5/cd/06f72bc9187840f1c99e1a8750aad4216fc7dfdd7df46e6280add14b4822/scipy-1.14.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:edaf02b82cd7639db00dbff629995ef185c8df4c3ffa71a5562a595765a06ce1", size = 42544049, upload-time = "2024-08-21T00:05:59.294Z" }, + { url = "https://files.pythonhosted.org/packages/aa/7d/43ab67228ef98c6b5dd42ab386eae2d7877036970a0d7e3dd3eb47a0d530/scipy-1.14.1-cp312-cp312-win_amd64.whl", hash = "sha256:2ff38e22128e6c03ff73b6bb0f85f897d2362f8c052e3b8ad00532198fbdae3f", size = 44521212, upload-time = "2024-08-21T00:06:06.521Z" }, + { url = "https://files.pythonhosted.org/packages/50/ef/ac98346db016ff18a6ad7626a35808f37074d25796fd0234c2bb0ed1e054/scipy-1.14.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:1729560c906963fc8389f6aac023739ff3983e727b1a4d87696b7bf108316a79", size = 39091068, upload-time = "2024-08-21T00:06:13.671Z" }, + { url = "https://files.pythonhosted.org/packages/b9/cc/70948fe9f393b911b4251e96b55bbdeaa8cca41f37c26fd1df0232933b9e/scipy-1.14.1-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:4079b90df244709e675cdc8b93bfd8a395d59af40b72e339c2287c91860deb8e", size = 29875417, upload-time = "2024-08-21T00:06:21.482Z" }, + { url = "https://files.pythonhosted.org/packages/3b/2e/35f549b7d231c1c9f9639f9ef49b815d816bf54dd050da5da1c11517a218/scipy-1.14.1-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:e0cf28db0f24a38b2a0ca33a85a54852586e43cf6fd876365c86e0657cfe7d73", size = 23084508, upload-time = "2024-08-21T00:06:28.064Z" }, + { url = "https://files.pythonhosted.org/packages/3f/d6/b028e3f3e59fae61fb8c0f450db732c43dd1d836223a589a8be9f6377203/scipy-1.14.1-cp313-cp313-macosx_14_0_x86_64.whl", hash = "sha256:0c2f95de3b04e26f5f3ad5bb05e74ba7f68b837133a4492414b3afd79dfe540e", size = 25503364, upload-time = "2024-08-21T00:06:35.25Z" }, + { url = "https://files.pythonhosted.org/packages/a7/2f/6c142b352ac15967744d62b165537a965e95d557085db4beab2a11f7943b/scipy-1.14.1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b99722ea48b7ea25e8e015e8341ae74624f72e5f21fc2abd45f3a93266de4c5d", size = 35292639, upload-time = "2024-08-21T00:06:44.542Z" }, + { url = "https://files.pythonhosted.org/packages/56/46/2449e6e51e0d7c3575f289f6acb7f828938eaab8874dbccfeb0cd2b71a27/scipy-1.14.1-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5149e3fd2d686e42144a093b206aef01932a0059c2a33ddfa67f5f035bdfe13e", size = 40798288, upload-time = "2024-08-21T00:06:54.182Z" }, + { url = "https://files.pythonhosted.org/packages/32/cd/9d86f7ed7f4497c9fd3e39f8918dd93d9f647ba80d7e34e4946c0c2d1a7c/scipy-1.14.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e4f5a7c49323533f9103d4dacf4e4f07078f360743dec7f7596949149efeec06", size = 42524647, upload-time = "2024-08-21T00:07:04.649Z" }, + { url = "https://files.pythonhosted.org/packages/f5/1b/6ee032251bf4cdb0cc50059374e86a9f076308c1512b61c4e003e241efb7/scipy-1.14.1-cp313-cp313-win_amd64.whl", hash = "sha256:baff393942b550823bfce952bb62270ee17504d02a1801d7fd0719534dfb9c84", size = 44469524, upload-time = "2024-08-21T00:07:15.381Z" }, +] + +[[package]] +name = "semantic-version" +version = "2.10.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/7d/31/f2289ce78b9b473d582568c234e104d2a342fd658cc288a7553d83bb8595/semantic_version-2.10.0.tar.gz", hash = "sha256:bdabb6d336998cbb378d4b9db3a4b56a1e3235701dc05ea2690d9a997ed5041c", size = 52289, upload-time = "2022-05-26T13:35:23.454Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/6a/23/8146aad7d88f4fcb3a6218f41a60f6c2d4e3a72de72da1825dc7c8f7877c/semantic_version-2.10.0-py2.py3-none-any.whl", hash = "sha256:de78a3b8e0feda74cabc54aab2da702113e33ac9d9eb9d2389bcf1f58b7d9177", size = 15552, upload-time = "2022-05-26T13:35:21.206Z" }, +] + +[[package]] +name = "sentencepiece" +version = "0.2.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/c9/d2/b9c7ca067c26d8ff085d252c89b5f69609ca93fb85a00ede95f4857865d4/sentencepiece-0.2.0.tar.gz", hash = "sha256:a52c19171daaf2e697dc6cbe67684e0fa341b1248966f6aebb541de654d15843", size = 2632106, upload-time = "2024-02-19T17:06:47.428Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/f6/71/98648c3b64b23edb5403f74bcc906ad21766872a6e1ada26ea3f1eb941ab/sentencepiece-0.2.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:188779e1298a1c8b8253c7d3ad729cb0a9891e5cef5e5d07ce4592c54869e227", size = 2408979, upload-time = "2024-02-19T17:05:34.651Z" }, + { url = "https://files.pythonhosted.org/packages/77/9f/7efbaa6d4c0c718a9affbecc536b03ca62f99f421bdffb531c16030e2d2b/sentencepiece-0.2.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:bed9cf85b296fa2b76fc2547b9cbb691a523864cebaee86304c43a7b4cb1b452", size = 1238845, upload-time = "2024-02-19T17:05:37.371Z" }, + { url = "https://files.pythonhosted.org/packages/1c/e4/c2541027a43ec6962ba9b601805d17ba3f86b38bdeae0e8ac65a2981e248/sentencepiece-0.2.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:d7b67e724bead13f18db6e1d10b6bbdc454af574d70efbb36f27d90387be1ca3", size = 1181472, upload-time = "2024-02-19T17:05:39.775Z" }, + { url = "https://files.pythonhosted.org/packages/fd/46/316c1ba6c52b97de76aff7b9da678f7afbb52136afb2987c474d95630e65/sentencepiece-0.2.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2fde4b08cfe237be4484c6c7c2e2c75fb862cfeab6bd5449ce4caeafd97b767a", size = 1259151, upload-time = "2024-02-19T17:05:42.594Z" }, + { url = "https://files.pythonhosted.org/packages/aa/5a/3c48738a0835d76dd06c62b6ac48d39c923cde78dd0f587353bdcbb99851/sentencepiece-0.2.0-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4c378492056202d1c48a4979650981635fd97875a00eabb1f00c6a236b013b5e", size = 1355931, upload-time = "2024-02-19T17:05:44.695Z" }, + { url = "https://files.pythonhosted.org/packages/a6/27/33019685023221ca8ed98e8ceb7ae5e166032686fa3662c68f1f1edf334e/sentencepiece-0.2.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1380ce6540a368de2ef6d7e6ba14ba8f3258df650d39ba7d833b79ee68a52040", size = 1301537, upload-time = "2024-02-19T17:05:46.713Z" }, + { url = "https://files.pythonhosted.org/packages/ca/e4/55f97cef14293171fef5f96e96999919ab5b4d1ce95b53547ad653d7e3bf/sentencepiece-0.2.0-cp310-cp310-win32.whl", hash = "sha256:a1151d6a6dd4b43e552394aed0edfe9292820272f0194bd56c7c1660a0c06c3d", size = 936747, upload-time = "2024-02-19T17:05:48.705Z" }, + { url = "https://files.pythonhosted.org/packages/85/f4/4ef1a6e0e9dbd8a60780a91df8b7452ada14cfaa0e17b3b8dfa42cecae18/sentencepiece-0.2.0-cp310-cp310-win_amd64.whl", hash = "sha256:d490142b0521ef22bc1085f061d922a2a6666175bb6b42e588ff95c0db6819b2", size = 991525, upload-time = "2024-02-19T17:05:55.145Z" }, + { url = "https://files.pythonhosted.org/packages/32/43/8f8885168a47a02eba1455bd3f4f169f50ad5b8cebd2402d0f5e20854d04/sentencepiece-0.2.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:17982700c4f6dbb55fa3594f3d7e5dd1c8659a274af3738e33c987d2a27c9d5c", size = 2409036, upload-time = "2024-02-19T17:05:58.021Z" }, + { url = "https://files.pythonhosted.org/packages/0f/35/e63ba28062af0a3d688a9f128e407a1a2608544b2f480cb49bf7f4b1cbb9/sentencepiece-0.2.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:7c867012c0e8bcd5bdad0f791609101cb5c66acb303ab3270218d6debc68a65e", size = 1238921, upload-time = "2024-02-19T17:06:06.434Z" }, + { url = "https://files.pythonhosted.org/packages/de/42/ae30952c4a0bd773e90c9bf2579f5533037c886dfc8ec68133d5694f4dd2/sentencepiece-0.2.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:7fd6071249c74f779c5b27183295b9202f8dedb68034e716784364443879eaa6", size = 1181477, upload-time = "2024-02-19T17:06:09.292Z" }, + { url = "https://files.pythonhosted.org/packages/e3/ac/2f2ab1d60bb2d795d054eebe5e3f24b164bc21b5a9b75fba7968b3b91b5a/sentencepiece-0.2.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:27f90c55a65013cbb8f4d7aab0599bf925cde4adc67ae43a0d323677b5a1c6cb", size = 1259182, upload-time = "2024-02-19T17:06:16.459Z" }, + { url = "https://files.pythonhosted.org/packages/45/fb/14633c6ecf262c468759ffcdb55c3a7ee38fe4eda6a70d75ee7c7d63c58b/sentencepiece-0.2.0-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:b293734059ef656dcd65be62ff771507bea8fed0a711b6733976e1ed3add4553", size = 1355537, upload-time = "2024-02-19T17:06:19.274Z" }, + { url = "https://files.pythonhosted.org/packages/fb/12/2f5c8d4764b00033cf1c935b702d3bb878d10be9f0b87f0253495832d85f/sentencepiece-0.2.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e58b47f933aca74c6a60a79dcb21d5b9e47416256c795c2d58d55cec27f9551d", size = 1301464, upload-time = "2024-02-19T17:06:21.796Z" }, + { url = "https://files.pythonhosted.org/packages/4e/b1/67afc0bde24f6dcb3acdea0dd8dcdf4b8b0db240f6bacd39378bd32d09f8/sentencepiece-0.2.0-cp311-cp311-win32.whl", hash = "sha256:c581258cf346b327c62c4f1cebd32691826306f6a41d8c4bec43b010dee08e75", size = 936749, upload-time = "2024-02-19T17:06:24.167Z" }, + { url = "https://files.pythonhosted.org/packages/a2/f6/587c62fd21fc988555b85351f50bbde43a51524caafd63bc69240ded14fd/sentencepiece-0.2.0-cp311-cp311-win_amd64.whl", hash = "sha256:0993dbc665f4113017892f1b87c3904a44d0640eda510abcacdfb07f74286d36", size = 991520, upload-time = "2024-02-19T17:06:26.936Z" }, + { url = "https://files.pythonhosted.org/packages/27/5a/141b227ed54293360a9ffbb7bf8252b4e5efc0400cdeac5809340e5d2b21/sentencepiece-0.2.0-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:ea5f536e32ea8ec96086ee00d7a4a131ce583a1b18d130711707c10e69601cb2", size = 2409370, upload-time = "2024-02-19T17:06:29.315Z" }, + { url = "https://files.pythonhosted.org/packages/2e/08/a4c135ad6fc2ce26798d14ab72790d66e813efc9589fd30a5316a88ca8d5/sentencepiece-0.2.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:d0cb51f53b6aae3c36bafe41e86167c71af8370a039f542c43b0cce5ef24a68c", size = 1239288, upload-time = "2024-02-19T17:06:31.674Z" }, + { url = "https://files.pythonhosted.org/packages/49/0a/2fe387f825ac5aad5a0bfe221904882106cac58e1b693ba7818785a882b6/sentencepiece-0.2.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:3212121805afc58d8b00ab4e7dd1f8f76c203ddb9dc94aa4079618a31cf5da0f", size = 1181597, upload-time = "2024-02-19T17:06:33.763Z" }, + { url = "https://files.pythonhosted.org/packages/cc/38/e4698ee2293fe4835dc033c49796a39b3eebd8752098f6bd0aa53a14af1f/sentencepiece-0.2.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2a3149e3066c2a75e0d68a43eb632d7ae728c7925b517f4c05c40f6f7280ce08", size = 1259220, upload-time = "2024-02-19T17:06:35.85Z" }, + { url = "https://files.pythonhosted.org/packages/12/24/fd7ef967c9dad2f6e6e5386d0cadaf65cda8b7be6e3861a9ab3121035139/sentencepiece-0.2.0-cp312-cp312-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:632f3594d3e7ac8b367bca204cb3fd05a01d5b21455acd097ea4c0e30e2f63d7", size = 1355962, upload-time = "2024-02-19T17:06:38.616Z" }, + { url = "https://files.pythonhosted.org/packages/4f/d2/18246f43ca730bb81918f87b7e886531eda32d835811ad9f4657c54eee35/sentencepiece-0.2.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f295105c6bdbb05bd5e1b0cafbd78ff95036f5d3641e7949455a3f4e5e7c3109", size = 1301706, upload-time = "2024-02-19T17:06:40.712Z" }, + { url = "https://files.pythonhosted.org/packages/8a/47/ca237b562f420044ab56ddb4c278672f7e8c866e183730a20e413b38a989/sentencepiece-0.2.0-cp312-cp312-win32.whl", hash = "sha256:fb89f811e5efd18bab141afc3fea3de141c3f69f3fe9e898f710ae7fe3aab251", size = 936941, upload-time = "2024-02-19T17:06:42.802Z" }, + { url = "https://files.pythonhosted.org/packages/c6/97/d159c32642306ee2b70732077632895438867b3b6df282354bd550cf2a67/sentencepiece-0.2.0-cp312-cp312-win_amd64.whl", hash = "sha256:7a673a72aab81fef5ebe755c6e0cc60087d1f3a4700835d40537183c1703a45f", size = 991994, upload-time = "2024-02-19T17:06:45.01Z" }, +] + +[[package]] +name = "sentry-sdk" +version = "3.0.0a7" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "certifi" }, + { name = "opentelemetry-sdk" }, + { name = "urllib3" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/66/a5/747ab268f2cf3f4cb50af521e1a85a959411f74cf4a2ecf47782e1981cee/sentry_sdk-3.0.0a7.tar.gz", hash = "sha256:439a858d409b84407560df6b7ebb760acd9ad3d14e05cd919ae36b204e411cb2", size = 329999, upload-time = "2025-10-20T13:50:38.131Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/ce/5e/23837eadf18ad4f9d5894c1210e6c2591136fd36cff338bedc90a1222bf2/sentry_sdk-3.0.0a7-py2.py3-none-any.whl", hash = "sha256:ada791a77c2d489e07b04c068e0fcacd6bbcfe1cb40b5f303207738baf38e868", size = 353618, upload-time = "2025-10-20T13:50:36.158Z" }, +] + +[[package]] +name = "setuptools" +version = "80.9.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/18/5d/3bf57dcd21979b887f014ea83c24ae194cfcd12b9e0fda66b957c69d1fca/setuptools-80.9.0.tar.gz", hash = "sha256:f36b47402ecde768dbfafc46e8e4207b4360c654f1f3bb84475f0a28628fb19c", size = 1319958, upload-time = "2025-05-27T00:56:51.443Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/a3/dc/17031897dae0efacfea57dfd3a82fdd2a2aeb58e0ff71b77b87e44edc772/setuptools-80.9.0-py3-none-any.whl", hash = "sha256:062d34222ad13e0cc312a4c02d73f059e86a4acbfbdea8f8f76b28c99f306922", size = 1201486, upload-time = "2025-05-27T00:56:49.664Z" }, +] + +[[package]] +name = "shellingham" +version = "1.5.4" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/58/15/8b3609fd3830ef7b27b655beb4b4e9c62313a4e8da8c676e142cc210d58e/shellingham-1.5.4.tar.gz", hash = "sha256:8dbca0739d487e5bd35ab3ca4b36e11c4078f3a234bfce294b0a0291363404de", size = 10310, upload-time = "2023-10-24T04:13:40.426Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/e0/f9/0595336914c5619e5f28a1fb793285925a8cd4b432c9da0a987836c7f822/shellingham-1.5.4-py2.py3-none-any.whl", hash = "sha256:7ecfff8f2fd72616f7481040475a65b2bf8af90a56c89140852d1120324e8686", size = 9755, upload-time = "2023-10-24T04:13:38.866Z" }, +] + +[[package]] +name = "six" +version = "1.16.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/71/39/171f1c67cd00715f190ba0b100d606d440a28c93c7714febeca8b79af85e/six-1.16.0.tar.gz", hash = "sha256:1e61c37477a1626458e36f7b1d82aa5c9b094fa4802892072e49de9c60c4c926", size = 34041, upload-time = "2021-05-05T14:18:18.379Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/d9/5a/e7c31adbe875f2abbb91bd84cf2dc52d792b5a01506781dbcf25c91daf11/six-1.16.0-py2.py3-none-any.whl", hash = "sha256:8abb2f1d86890a2dfb989f9a77cfcfd3e47c2a354b01111771326f8aa26e0254", size = 11053, upload-time = "2021-05-05T14:18:17.237Z" }, +] + +[[package]] +name = "smmap" +version = "5.0.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/44/cd/a040c4b3119bbe532e5b0732286f805445375489fceaec1f48306068ee3b/smmap-5.0.2.tar.gz", hash = "sha256:26ea65a03958fa0c8a1c7e8c7a58fdc77221b8910f6be2131affade476898ad5", size = 22329, upload-time = "2025-01-02T07:14:40.909Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/04/be/d09147ad1ec7934636ad912901c5fd7667e1c858e19d355237db0d0cd5e4/smmap-5.0.2-py3-none-any.whl", hash = "sha256:b30115f0def7d7531d22a0fb6502488d879e75b260a9db4d0819cfb25403af5e", size = 24303, upload-time = "2025-01-02T07:14:38.724Z" }, +] + +[[package]] +name = "sniffio" +version = "1.3.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/a2/87/a6771e1546d97e7e041b6ae58d80074f81b7d5121207425c964ddf5cfdbd/sniffio-1.3.1.tar.gz", hash = "sha256:f4324edc670a0f49750a81b895f35c3adb843cca46f0530f79fc1babb23789dc", size = 20372, upload-time = "2024-02-25T23:20:04.057Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/e9/44/75a9c9421471a6c4805dbf2356f7c181a29c1879239abab1ea2cc8f38b40/sniffio-1.3.1-py3-none-any.whl", hash = "sha256:2f6da418d1f1e0fddd844478f41680e794e6051915791a034ff65e5f100525a2", size = 10235, upload-time = "2024-02-25T23:20:01.196Z" }, +] + +[[package]] +name = "starlette" +version = "0.49.3" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "anyio" }, + { name = "typing-extensions", marker = "python_full_version < '3.13'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/de/1a/608df0b10b53b0beb96a37854ee05864d182ddd4b1156a22f1ad3860425a/starlette-0.49.3.tar.gz", hash = "sha256:1c14546f299b5901a1ea0e34410575bc33bbd741377a10484a54445588d00284", size = 2655031, upload-time = "2025-11-01T15:12:26.13Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/a3/e0/021c772d6a662f43b63044ab481dc6ac7592447605b5b35a957785363122/starlette-0.49.3-py3-none-any.whl", hash = "sha256:b579b99715fdc2980cf88c8ec96d3bf1ce16f5a8051a7c2b84ef9b1cdecaea2f", size = 74340, upload-time = "2025-11-01T15:12:24.387Z" }, +] + +[[package]] +name = "sympy" +version = "1.14.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "mpmath" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/83/d3/803453b36afefb7c2bb238361cd4ae6125a569b4db67cd9e79846ba2d68c/sympy-1.14.0.tar.gz", hash = "sha256:d3d3fe8df1e5a0b42f0e7bdf50541697dbe7d23746e894990c030e2b05e72517", size = 7793921, upload-time = "2025-04-27T18:05:01.611Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/a2/09/77d55d46fd61b4a135c444fc97158ef34a095e5681d0a6c10b75bf356191/sympy-1.14.0-py3-none-any.whl", hash = "sha256:e091cc3e99d2141a0ba2847328f5479b05d94a6635cb96148ccb3f34671bd8f5", size = 6299353, upload-time = "2025-04-27T18:04:59.103Z" }, +] + +[[package]] +name = "tensorboard" +version = "2.20.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "absl-py" }, + { name = "grpcio" }, + { name = "markdown" }, + { name = "numpy" }, + { name = "packaging" }, + { name = "pillow" }, + { name = "protobuf" }, + { name = "setuptools" }, + { name = "tensorboard-data-server" }, + { name = "werkzeug" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/9c/d9/a5db55f88f258ac669a92858b70a714bbbd5acd993820b41ec4a96a4d77f/tensorboard-2.20.0-py3-none-any.whl", hash = "sha256:9dc9f978cb84c0723acf9a345d96c184f0293d18f166bb8d59ee098e6cfaaba6", size = 5525680, upload-time = "2025-07-17T19:20:49.638Z" }, +] + +[[package]] +name = "tensorboard-data-server" +version = "0.7.2" +source = { registry = "https://pypi.org/simple" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/7a/13/e503968fefabd4c6b2650af21e110aa8466fe21432cd7c43a84577a89438/tensorboard_data_server-0.7.2-py3-none-any.whl", hash = "sha256:7e0610d205889588983836ec05dc098e80f97b7e7bbff7e994ebb78f578d0ddb", size = 2356, upload-time = "2023-10-23T21:23:32.16Z" }, + { url = "https://files.pythonhosted.org/packages/b7/85/dabeaf902892922777492e1d253bb7e1264cadce3cea932f7ff599e53fea/tensorboard_data_server-0.7.2-py3-none-macosx_10_9_x86_64.whl", hash = "sha256:9fe5d24221b29625dbc7328b0436ca7fc1c23de4acf4d272f1180856e32f9f60", size = 4823598, upload-time = "2023-10-23T21:23:33.714Z" }, + { url = "https://files.pythonhosted.org/packages/73/c6/825dab04195756cf8ff2e12698f22513b3db2f64925bdd41671bfb33aaa5/tensorboard_data_server-0.7.2-py3-none-manylinux_2_31_x86_64.whl", hash = "sha256:ef687163c24185ae9754ed5650eb5bc4d84ff257aabdc33f0cc6f74d8ba54530", size = 6590363, upload-time = "2023-10-23T21:23:35.583Z" }, +] + +[[package]] +name = "test-tube" +version = "0.7.5" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "future" }, + { name = "imageio" }, + { name = "numpy" }, + { name = "pandas" }, + { name = "tensorboard" }, + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/91/f0/5c32f2fbd824f32354f7f4632c957163071597bb2c6a4105f507bc9af7c0/test_tube-0.7.5.tar.gz", hash = "sha256:1379c33eb8cde3e9b36610f87da0f16c2e06496b1cfebac473df4e7be2faa124", size = 21849, upload-time = "2019-12-01T01:19:50.642Z" } + +[[package]] +name = "timm" +version = "1.0.11" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "huggingface-hub" }, + { name = "pyyaml" }, + { name = "safetensors" }, + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, + { name = "torchvision", version = "0.25.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torchvision", version = "0.25.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torchvision", version = "0.25.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/5b/53/33b596c79eb911586ca90e16d64fde9738ddd0d34273ae562ea664eacca7/timm-1.0.11.tar.gz", hash = "sha256:a005f72b87e67ed30cdbf405a9ffd4e723360c780a43b1cefe266af8ecc9d151", size = 2195894, upload-time = "2024-10-16T21:17:48.892Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/77/f7/144d69d921fba82baa8498e279d782a12a13694f45ff74d1777bc3231ffe/timm-1.0.11-py3-none-any.whl", hash = "sha256:52a6f895e2cbec35a87d0846870419c2c0aa40de9b205bcda917f38278bf3044", size = 2326571, upload-time = "2024-10-16T21:17:47.129Z" }, +] + +[[package]] +name = "tokenizers" +version = "0.22.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "huggingface-hub" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/1c/46/fb6854cec3278fbfa4a75b50232c77622bc517ac886156e6afbfa4d8fc6e/tokenizers-0.22.1.tar.gz", hash = "sha256:61de6522785310a309b3407bac22d99c4db5dba349935e99e4d15ea2226af2d9", size = 363123, upload-time = "2025-09-19T09:49:23.424Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/bf/33/f4b2d94ada7ab297328fc671fed209368ddb82f965ec2224eb1892674c3a/tokenizers-0.22.1-cp39-abi3-macosx_10_12_x86_64.whl", hash = "sha256:59fdb013df17455e5f950b4b834a7b3ee2e0271e6378ccb33aa74d178b513c73", size = 3069318, upload-time = "2025-09-19T09:49:11.848Z" }, + { url = "https://files.pythonhosted.org/packages/1c/58/2aa8c874d02b974990e89ff95826a4852a8b2a273c7d1b4411cdd45a4565/tokenizers-0.22.1-cp39-abi3-macosx_11_0_arm64.whl", hash = "sha256:8d4e484f7b0827021ac5f9f71d4794aaef62b979ab7608593da22b1d2e3c4edc", size = 2926478, upload-time = "2025-09-19T09:49:09.759Z" }, + { url = "https://files.pythonhosted.org/packages/1e/3b/55e64befa1e7bfea963cf4b787b2cea1011362c4193f5477047532ce127e/tokenizers-0.22.1-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:19d2962dd28bc67c1f205ab180578a78eef89ac60ca7ef7cbe9635a46a56422a", size = 3256994, upload-time = "2025-09-19T09:48:56.701Z" }, + { url = "https://files.pythonhosted.org/packages/71/0b/fbfecf42f67d9b7b80fde4aabb2b3110a97fac6585c9470b5bff103a80cb/tokenizers-0.22.1-cp39-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:38201f15cdb1f8a6843e6563e6e79f4abd053394992b9bbdf5213ea3469b4ae7", size = 3153141, upload-time = "2025-09-19T09:48:59.749Z" }, + { url = "https://files.pythonhosted.org/packages/17/a9/b38f4e74e0817af8f8ef925507c63c6ae8171e3c4cb2d5d4624bf58fca69/tokenizers-0.22.1-cp39-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d1cbe5454c9a15df1b3443c726063d930c16f047a3cc724b9e6e1a91140e5a21", size = 3508049, upload-time = "2025-09-19T09:49:05.868Z" }, + { url = "https://files.pythonhosted.org/packages/d2/48/dd2b3dac46bb9134a88e35d72e1aa4869579eacc1a27238f1577270773ff/tokenizers-0.22.1-cp39-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e7d094ae6312d69cc2a872b54b91b309f4f6fbce871ef28eb27b52a98e4d0214", size = 3710730, upload-time = "2025-09-19T09:49:01.832Z" }, + { url = "https://files.pythonhosted.org/packages/93/0e/ccabc8d16ae4ba84a55d41345207c1e2ea88784651a5a487547d80851398/tokenizers-0.22.1-cp39-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:afd7594a56656ace95cdd6df4cca2e4059d294c5cfb1679c57824b605556cb2f", size = 3412560, upload-time = "2025-09-19T09:49:03.867Z" }, + { url = "https://files.pythonhosted.org/packages/d0/c6/dc3a0db5a6766416c32c034286d7c2d406da1f498e4de04ab1b8959edd00/tokenizers-0.22.1-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e2ef6063d7a84994129732b47e7915e8710f27f99f3a3260b8a38fc7ccd083f4", size = 3250221, upload-time = "2025-09-19T09:49:07.664Z" }, + { url = "https://files.pythonhosted.org/packages/d7/a6/2c8486eef79671601ff57b093889a345dd3d576713ef047776015dc66de7/tokenizers-0.22.1-cp39-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:ba0a64f450b9ef412c98f6bcd2a50c6df6e2443b560024a09fa6a03189726879", size = 9345569, upload-time = "2025-09-19T09:49:14.214Z" }, + { url = "https://files.pythonhosted.org/packages/6b/16/32ce667f14c35537f5f605fe9bea3e415ea1b0a646389d2295ec348d5657/tokenizers-0.22.1-cp39-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:331d6d149fa9c7d632cde4490fb8bbb12337fa3a0232e77892be656464f4b446", size = 9271599, upload-time = "2025-09-19T09:49:16.639Z" }, + { url = "https://files.pythonhosted.org/packages/51/7c/a5f7898a3f6baa3fc2685c705e04c98c1094c523051c805cdd9306b8f87e/tokenizers-0.22.1-cp39-abi3-musllinux_1_2_i686.whl", hash = "sha256:607989f2ea68a46cb1dfbaf3e3aabdf3f21d8748312dbeb6263d1b3b66c5010a", size = 9533862, upload-time = "2025-09-19T09:49:19.146Z" }, + { url = "https://files.pythonhosted.org/packages/36/65/7e75caea90bc73c1dd8d40438adf1a7bc26af3b8d0a6705ea190462506e1/tokenizers-0.22.1-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:a0f307d490295717726598ef6fa4f24af9d484809223bbc253b201c740a06390", size = 9681250, upload-time = "2025-09-19T09:49:21.501Z" }, + { url = "https://files.pythonhosted.org/packages/30/2c/959dddef581b46e6209da82df3b78471e96260e2bc463f89d23b1bf0e52a/tokenizers-0.22.1-cp39-abi3-win32.whl", hash = "sha256:b5120eed1442765cd90b903bb6cfef781fd8fe64e34ccaecbae4c619b7b12a82", size = 2472003, upload-time = "2025-09-19T09:49:27.089Z" }, + { url = "https://files.pythonhosted.org/packages/b3/46/e33a8c93907b631a99377ef4c5f817ab453d0b34f93529421f42ff559671/tokenizers-0.22.1-cp39-abi3-win_amd64.whl", hash = "sha256:65fd6e3fb11ca1e78a6a93602490f134d1fdeb13bcef99389d5102ea318ed138", size = 2674684, upload-time = "2025-09-19T09:49:24.953Z" }, +] + +[[package]] +name = "tomli" +version = "2.3.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/52/ed/3f73f72945444548f33eba9a87fc7a6e969915e7b1acc8260b30e1f76a2f/tomli-2.3.0.tar.gz", hash = "sha256:64be704a875d2a59753d80ee8a533c3fe183e3f06807ff7dc2232938ccb01549", size = 17392, upload-time = "2025-10-08T22:01:47.119Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/b3/2e/299f62b401438d5fe1624119c723f5d877acc86a4c2492da405626665f12/tomli-2.3.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:88bd15eb972f3664f5ed4b57c1634a97153b4bac4479dcb6a495f41921eb7f45", size = 153236, upload-time = "2025-10-08T22:01:00.137Z" }, + { url = "https://files.pythonhosted.org/packages/86/7f/d8fffe6a7aefdb61bced88fcb5e280cfd71e08939da5894161bd71bea022/tomli-2.3.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:883b1c0d6398a6a9d29b508c331fa56adbcdff647f6ace4dfca0f50e90dfd0ba", size = 148084, upload-time = "2025-10-08T22:01:01.63Z" }, + { url = "https://files.pythonhosted.org/packages/47/5c/24935fb6a2ee63e86d80e4d3b58b222dafaf438c416752c8b58537c8b89a/tomli-2.3.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d1381caf13ab9f300e30dd8feadb3de072aeb86f1d34a8569453ff32a7dea4bf", size = 234832, upload-time = "2025-10-08T22:01:02.543Z" }, + { url = "https://files.pythonhosted.org/packages/89/da/75dfd804fc11e6612846758a23f13271b76d577e299592b4371a4ca4cd09/tomli-2.3.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a0e285d2649b78c0d9027570d4da3425bdb49830a6156121360b3f8511ea3441", size = 242052, upload-time = "2025-10-08T22:01:03.836Z" }, + { url = "https://files.pythonhosted.org/packages/70/8c/f48ac899f7b3ca7eb13af73bacbc93aec37f9c954df3c08ad96991c8c373/tomli-2.3.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:0a154a9ae14bfcf5d8917a59b51ffd5a3ac1fd149b71b47a3a104ca4edcfa845", size = 239555, upload-time = "2025-10-08T22:01:04.834Z" }, + { url = "https://files.pythonhosted.org/packages/ba/28/72f8afd73f1d0e7829bfc093f4cb98ce0a40ffc0cc997009ee1ed94ba705/tomli-2.3.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:74bf8464ff93e413514fefd2be591c3b0b23231a77f901db1eb30d6f712fc42c", size = 245128, upload-time = "2025-10-08T22:01:05.84Z" }, + { url = "https://files.pythonhosted.org/packages/b6/eb/a7679c8ac85208706d27436e8d421dfa39d4c914dcf5fa8083a9305f58d9/tomli-2.3.0-cp311-cp311-win32.whl", hash = "sha256:00b5f5d95bbfc7d12f91ad8c593a1659b6387b43f054104cda404be6bda62456", size = 96445, upload-time = "2025-10-08T22:01:06.896Z" }, + { url = "https://files.pythonhosted.org/packages/0a/fe/3d3420c4cb1ad9cb462fb52967080575f15898da97e21cb6f1361d505383/tomli-2.3.0-cp311-cp311-win_amd64.whl", hash = "sha256:4dc4ce8483a5d429ab602f111a93a6ab1ed425eae3122032db7e9acf449451be", size = 107165, upload-time = "2025-10-08T22:01:08.107Z" }, + { url = "https://files.pythonhosted.org/packages/ff/b7/40f36368fcabc518bb11c8f06379a0fd631985046c038aca08c6d6a43c6e/tomli-2.3.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:d7d86942e56ded512a594786a5ba0a5e521d02529b3826e7761a05138341a2ac", size = 154891, upload-time = "2025-10-08T22:01:09.082Z" }, + { url = "https://files.pythonhosted.org/packages/f9/3f/d9dd692199e3b3aab2e4e4dd948abd0f790d9ded8cd10cbaae276a898434/tomli-2.3.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:73ee0b47d4dad1c5e996e3cd33b8a76a50167ae5f96a2607cbe8cc773506ab22", size = 148796, upload-time = "2025-10-08T22:01:10.266Z" }, + { url = "https://files.pythonhosted.org/packages/60/83/59bff4996c2cf9f9387a0f5a3394629c7efa5ef16142076a23a90f1955fa/tomli-2.3.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:792262b94d5d0a466afb5bc63c7daa9d75520110971ee269152083270998316f", size = 242121, upload-time = "2025-10-08T22:01:11.332Z" }, + { url = "https://files.pythonhosted.org/packages/45/e5/7c5119ff39de8693d6baab6c0b6dcb556d192c165596e9fc231ea1052041/tomli-2.3.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4f195fe57ecceac95a66a75ac24d9d5fbc98ef0962e09b2eddec5d39375aae52", size = 250070, upload-time = "2025-10-08T22:01:12.498Z" }, + { url = "https://files.pythonhosted.org/packages/45/12/ad5126d3a278f27e6701abde51d342aa78d06e27ce2bb596a01f7709a5a2/tomli-2.3.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:e31d432427dcbf4d86958c184b9bfd1e96b5b71f8eb17e6d02531f434fd335b8", size = 245859, upload-time = "2025-10-08T22:01:13.551Z" }, + { url = "https://files.pythonhosted.org/packages/fb/a1/4d6865da6a71c603cfe6ad0e6556c73c76548557a8d658f9e3b142df245f/tomli-2.3.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:7b0882799624980785240ab732537fcfc372601015c00f7fc367c55308c186f6", size = 250296, upload-time = "2025-10-08T22:01:14.614Z" }, + { url = "https://files.pythonhosted.org/packages/a0/b7/a7a7042715d55c9ba6e8b196d65d2cb662578b4d8cd17d882d45322b0d78/tomli-2.3.0-cp312-cp312-win32.whl", hash = "sha256:ff72b71b5d10d22ecb084d345fc26f42b5143c5533db5e2eaba7d2d335358876", size = 97124, upload-time = "2025-10-08T22:01:15.629Z" }, + { url = "https://files.pythonhosted.org/packages/06/1e/f22f100db15a68b520664eb3328fb0ae4e90530887928558112c8d1f4515/tomli-2.3.0-cp312-cp312-win_amd64.whl", hash = "sha256:1cb4ed918939151a03f33d4242ccd0aa5f11b3547d0cf30f7c74a408a5b99878", size = 107698, upload-time = "2025-10-08T22:01:16.51Z" }, + { url = "https://files.pythonhosted.org/packages/89/48/06ee6eabe4fdd9ecd48bf488f4ac783844fd777f547b8d1b61c11939974e/tomli-2.3.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:5192f562738228945d7b13d4930baffda67b69425a7f0da96d360b0a3888136b", size = 154819, upload-time = "2025-10-08T22:01:17.964Z" }, + { url = "https://files.pythonhosted.org/packages/f1/01/88793757d54d8937015c75dcdfb673c65471945f6be98e6a0410fba167ed/tomli-2.3.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:be71c93a63d738597996be9528f4abe628d1adf5e6eb11607bc8fe1a510b5dae", size = 148766, upload-time = "2025-10-08T22:01:18.959Z" }, + { url = "https://files.pythonhosted.org/packages/42/17/5e2c956f0144b812e7e107f94f1cc54af734eb17b5191c0bbfb72de5e93e/tomli-2.3.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c4665508bcbac83a31ff8ab08f424b665200c0e1e645d2bd9ab3d3e557b6185b", size = 240771, upload-time = "2025-10-08T22:01:20.106Z" }, + { url = "https://files.pythonhosted.org/packages/d5/f4/0fbd014909748706c01d16824eadb0307115f9562a15cbb012cd9b3512c5/tomli-2.3.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4021923f97266babc6ccab9f5068642a0095faa0a51a246a6a02fccbb3514eaf", size = 248586, upload-time = "2025-10-08T22:01:21.164Z" }, + { url = "https://files.pythonhosted.org/packages/30/77/fed85e114bde5e81ecf9bc5da0cc69f2914b38f4708c80ae67d0c10180c5/tomli-2.3.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:a4ea38c40145a357d513bffad0ed869f13c1773716cf71ccaa83b0fa0cc4e42f", size = 244792, upload-time = "2025-10-08T22:01:22.417Z" }, + { url = "https://files.pythonhosted.org/packages/55/92/afed3d497f7c186dc71e6ee6d4fcb0acfa5f7d0a1a2878f8beae379ae0cc/tomli-2.3.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:ad805ea85eda330dbad64c7ea7a4556259665bdf9d2672f5dccc740eb9d3ca05", size = 248909, upload-time = "2025-10-08T22:01:23.859Z" }, + { url = "https://files.pythonhosted.org/packages/f8/84/ef50c51b5a9472e7265ce1ffc7f24cd4023d289e109f669bdb1553f6a7c2/tomli-2.3.0-cp313-cp313-win32.whl", hash = "sha256:97d5eec30149fd3294270e889b4234023f2c69747e555a27bd708828353ab606", size = 96946, upload-time = "2025-10-08T22:01:24.893Z" }, + { url = "https://files.pythonhosted.org/packages/b2/b7/718cd1da0884f281f95ccfa3a6cc572d30053cba64603f79d431d3c9b61b/tomli-2.3.0-cp313-cp313-win_amd64.whl", hash = "sha256:0c95ca56fbe89e065c6ead5b593ee64b84a26fca063b5d71a1122bf26e533999", size = 107705, upload-time = "2025-10-08T22:01:26.153Z" }, + { url = "https://files.pythonhosted.org/packages/19/94/aeafa14a52e16163008060506fcb6aa1949d13548d13752171a755c65611/tomli-2.3.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:cebc6fe843e0733ee827a282aca4999b596241195f43b4cc371d64fc6639da9e", size = 154244, upload-time = "2025-10-08T22:01:27.06Z" }, + { url = "https://files.pythonhosted.org/packages/db/e4/1e58409aa78eefa47ccd19779fc6f36787edbe7d4cd330eeeedb33a4515b/tomli-2.3.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:4c2ef0244c75aba9355561272009d934953817c49f47d768070c3c94355c2aa3", size = 148637, upload-time = "2025-10-08T22:01:28.059Z" }, + { url = "https://files.pythonhosted.org/packages/26/b6/d1eccb62f665e44359226811064596dd6a366ea1f985839c566cd61525ae/tomli-2.3.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c22a8bf253bacc0cf11f35ad9808b6cb75ada2631c2d97c971122583b129afbc", size = 241925, upload-time = "2025-10-08T22:01:29.066Z" }, + { url = "https://files.pythonhosted.org/packages/70/91/7cdab9a03e6d3d2bb11beae108da5bdc1c34bdeb06e21163482544ddcc90/tomli-2.3.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0eea8cc5c5e9f89c9b90c4896a8deefc74f518db5927d0e0e8d4a80953d774d0", size = 249045, upload-time = "2025-10-08T22:01:31.98Z" }, + { url = "https://files.pythonhosted.org/packages/15/1b/8c26874ed1f6e4f1fcfeb868db8a794cbe9f227299402db58cfcc858766c/tomli-2.3.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:b74a0e59ec5d15127acdabd75ea17726ac4c5178ae51b85bfe39c4f8a278e879", size = 245835, upload-time = "2025-10-08T22:01:32.989Z" }, + { url = "https://files.pythonhosted.org/packages/fd/42/8e3c6a9a4b1a1360c1a2a39f0b972cef2cc9ebd56025168c4137192a9321/tomli-2.3.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:b5870b50c9db823c595983571d1296a6ff3e1b88f734a4c8f6fc6188397de005", size = 253109, upload-time = "2025-10-08T22:01:34.052Z" }, + { url = "https://files.pythonhosted.org/packages/22/0c/b4da635000a71b5f80130937eeac12e686eefb376b8dee113b4a582bba42/tomli-2.3.0-cp314-cp314-win32.whl", hash = "sha256:feb0dacc61170ed7ab602d3d972a58f14ee3ee60494292d384649a3dc38ef463", size = 97930, upload-time = "2025-10-08T22:01:35.082Z" }, + { url = "https://files.pythonhosted.org/packages/b9/74/cb1abc870a418ae99cd5c9547d6bce30701a954e0e721821df483ef7223c/tomli-2.3.0-cp314-cp314-win_amd64.whl", hash = "sha256:b273fcbd7fc64dc3600c098e39136522650c49bca95df2d11cf3b626422392c8", size = 107964, upload-time = "2025-10-08T22:01:36.057Z" }, + { url = "https://files.pythonhosted.org/packages/54/78/5c46fff6432a712af9f792944f4fcd7067d8823157949f4e40c56b8b3c83/tomli-2.3.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:940d56ee0410fa17ee1f12b817b37a4d4e4dc4d27340863cc67236c74f582e77", size = 163065, upload-time = "2025-10-08T22:01:37.27Z" }, + { url = "https://files.pythonhosted.org/packages/39/67/f85d9bd23182f45eca8939cd2bc7050e1f90c41f4a2ecbbd5963a1d1c486/tomli-2.3.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:f85209946d1fe94416debbb88d00eb92ce9cd5266775424ff81bc959e001acaf", size = 159088, upload-time = "2025-10-08T22:01:38.235Z" }, + { url = "https://files.pythonhosted.org/packages/26/5a/4b546a0405b9cc0659b399f12b6adb750757baf04250b148d3c5059fc4eb/tomli-2.3.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a56212bdcce682e56b0aaf79e869ba5d15a6163f88d5451cbde388d48b13f530", size = 268193, upload-time = "2025-10-08T22:01:39.712Z" }, + { url = "https://files.pythonhosted.org/packages/42/4f/2c12a72ae22cf7b59a7fe75b3465b7aba40ea9145d026ba41cb382075b0e/tomli-2.3.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c5f3ffd1e098dfc032d4d3af5c0ac64f6d286d98bc148698356847b80fa4de1b", size = 275488, upload-time = "2025-10-08T22:01:40.773Z" }, + { url = "https://files.pythonhosted.org/packages/92/04/a038d65dbe160c3aa5a624e93ad98111090f6804027d474ba9c37c8ae186/tomli-2.3.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:5e01decd096b1530d97d5d85cb4dff4af2d8347bd35686654a004f8dea20fc67", size = 272669, upload-time = "2025-10-08T22:01:41.824Z" }, + { url = "https://files.pythonhosted.org/packages/be/2f/8b7c60a9d1612a7cbc39ffcca4f21a73bf368a80fc25bccf8253e2563267/tomli-2.3.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:8a35dd0e643bb2610f156cca8db95d213a90015c11fee76c946aa62b7ae7e02f", size = 279709, upload-time = "2025-10-08T22:01:43.177Z" }, + { url = "https://files.pythonhosted.org/packages/7e/46/cc36c679f09f27ded940281c38607716c86cf8ba4a518d524e349c8b4874/tomli-2.3.0-cp314-cp314t-win32.whl", hash = "sha256:a1f7f282fe248311650081faafa5f4732bdbfef5d45fe3f2e702fbc6f2d496e0", size = 107563, upload-time = "2025-10-08T22:01:44.233Z" }, + { url = "https://files.pythonhosted.org/packages/84/ff/426ca8683cf7b753614480484f6437f568fd2fda2edbdf57a2d3d8b27a0b/tomli-2.3.0-cp314-cp314t-win_amd64.whl", hash = "sha256:70a251f8d4ba2d9ac2542eecf008b3c8a9fc5c3f9f02c56a9d7952612be2fdba", size = 119756, upload-time = "2025-10-08T22:01:45.234Z" }, + { url = "https://files.pythonhosted.org/packages/77/b8/0135fadc89e73be292b473cb820b4f5a08197779206b33191e801feeae40/tomli-2.3.0-py3-none-any.whl", hash = "sha256:e95b1af3c5b07d9e643909b5abbec77cd9f1217e6d0bca72b0234736b9fb1f1b", size = 14408, upload-time = "2025-10-08T22:01:46.04Z" }, +] + +[[package]] +name = "tomlkit" +version = "0.13.3" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/cc/18/0bbf3884e9eaa38819ebe46a7bd25dcd56b67434402b66a58c4b8e552575/tomlkit-0.13.3.tar.gz", hash = "sha256:430cf247ee57df2b94ee3fbe588e71d362a941ebb545dec29b53961d61add2a1", size = 185207, upload-time = "2025-06-05T07:13:44.947Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/bd/75/8539d011f6be8e29f339c42e633aae3cb73bffa95dd0f9adec09b9c58e85/tomlkit-0.13.3-py3-none-any.whl", hash = "sha256:c89c649d79ee40629a9fda55f8ace8c6a1b42deb912b2a8fd8d942ddadb606b0", size = 38901, upload-time = "2025-06-05T07:13:43.546Z" }, +] + +[[package]] +name = "torch" +version = "2.10.0" +source = { registry = "https://download.pytorch.org/whl/cpu" } +resolution-markers = [ + "python_full_version >= '3.13' and sys_platform == 'darwin'", + "python_full_version == '3.12.*' and sys_platform == 'darwin'", + "python_full_version == '3.11.*' and sys_platform == 'darwin'", + "python_full_version < '3.11' and sys_platform == 'darwin'", +] +dependencies = [ + { name = "filelock", marker = "sys_platform == 'darwin'" }, + { name = "fsspec", marker = "sys_platform == 'darwin'" }, + { name = "jinja2", marker = "sys_platform == 'darwin'" }, + { name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11' and sys_platform == 'darwin'" }, + { name = "networkx", version = "3.6rc0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11' and sys_platform == 'darwin'" }, + { name = "setuptools", marker = "python_full_version >= '3.12' and sys_platform == 'darwin'" }, + { name = "sympy", marker = "sys_platform == 'darwin'" }, + { name = "typing-extensions", marker = "sys_platform == 'darwin'" }, +] +wheels = [ + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0-cp310-none-macosx_11_0_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0-cp311-none-macosx_11_0_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0-cp312-none-macosx_11_0_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0-cp313-cp313t-macosx_14_0_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0-cp313-none-macosx_11_0_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0-cp314-cp314-macosx_14_0_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0-cp314-cp314t-macosx_14_0_arm64.whl" }, +] + +[[package]] +name = "torch" +version = "2.10.0+cpu" +source = { registry = "https://download.pytorch.org/whl/cpu" } +resolution-markers = [ + "python_full_version >= '3.13' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.12.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.11.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux'", +] +dependencies = [ + { name = "filelock", marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "fsspec", marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "jinja2", marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "networkx", version = "3.6rc0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11' and sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "setuptools", marker = "python_full_version >= '3.12' and sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "sympy", marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "typing-extensions", marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, +] +wheels = [ + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp310-cp310-win_amd64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp311-cp311-win_amd64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp311-cp311-win_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp312-cp312-win_amd64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp312-cp312-win_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp313-cp313-win_amd64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp313-cp313-win_arm64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp313-cp313t-win_amd64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp314-cp314-win_amd64.whl" }, + { url = "https://download.pytorch.org/whl/cpu/torch-2.10.0%2Bcpu-cp314-cp314t-win_amd64.whl" }, +] + +[[package]] +name = "torch" +version = "2.10.0+cu128" +source = { registry = "https://download.pytorch.org/whl/cu128" } +resolution-markers = [ + "python_full_version >= '3.13' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version >= '3.13' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and platform_machine != 'aarch64' and sys_platform == 'linux'", +] +dependencies = [ + { name = "cuda-bindings", marker = "sys_platform == 'linux'" }, + { name = "filelock", marker = "sys_platform == 'linux'" }, + { name = "fsspec", marker = "sys_platform == 'linux'" }, + { name = "jinja2", marker = "sys_platform == 'linux'" }, + { name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11' and sys_platform == 'linux'" }, + { name = "networkx", version = "3.6rc0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11' and sys_platform == 'linux'" }, + { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cuda-cupti-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cuda-nvrtc-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cuda-runtime-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cudnn-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cufft-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cufile-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-curand-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cusolver-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cusparse-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-cusparselt-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-nccl-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-nvshmem-cu12", marker = "sys_platform == 'linux'" }, + { name = "nvidia-nvtx-cu12", marker = "sys_platform == 'linux'" }, + { name = "setuptools", marker = "python_full_version >= '3.12' and sys_platform == 'linux'" }, + { name = "sympy", marker = "sys_platform == 'linux'" }, + { name = "triton", version = "3.6.0", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform == 'linux'" }, + { name = "typing-extensions", marker = "sys_platform == 'linux'" }, +] +wheels = [ + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:e186f57ef1de1aa877943259819468fc6f27efb583b4a91f9215ada7b7f4e6cc" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:36368507b56eaa51acbd3c96ac8893bb9a86991ffcd0699fea3a1a74a2b8bdcb" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:85ed7944655ea6fd69377692e9cbfd7bba28d99696ceae79985e7caa99cf0a95" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:1d01ffaebf64715c0f507a39463149cb19e596ff702bd4bcf862601f2881dabc" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:6f09cdf2415516be028ae82e6b985bcfc3eac37bc52ab401142689f6224516ca" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:628e89bd5110ced7debee2a57c69959725b7fbc64eab81a39dd70e46c7e28ba5" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:bdbcc703382f948e951c063448c9406bf38ce66c41dd698d9e2733fcf96c037a" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:7b4bd23ed63de97456fcc81c26fea9f02ee02ce1112111c4dac0d8cfe574b23e" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:f1f8b840c64b645a4bc61a393db48effb9c92b2dc26c8373873911f0750d1ea7" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:23f58258012bcf1c349cb22af387e33aadca7f83ea617b080e774eb41e4fe8ff" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:c42377bc2607e3e1c60da71b792fb507c3938c87fd6edab8b21c59c91473c36d" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:37d71feea068776855686a1512058df3f19f6f040a151f055aa746601678744f" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:777461f50b2daf77e4bdd8e2ad34bdfc5a993bf1bdf2ab9ef39f5edfe4e9c12b" }, + { url = "https://download.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:7bcba6a7c5f0987a13298b1ca843155dcceceac758fa3c7ccd5c7af4059a1080" }, +] + +[[package]] +name = "torchaudio" +version = "2.10.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/04/59/88ab8ebff9d91f1f1365088b30f1b9ccce07c5eeac666038a5dee5e2f9b1/torchaudio-2.10.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:4cde383582a6240c1315443df5c5638863e96b03acf1cb44a298aff07a72d373", size = 734944, upload-time = "2026-01-21T16:28:49.535Z" }, + { url = "https://files.pythonhosted.org/packages/9b/d6/41f25f9ae9b37c191bed4cd474e403626685d2be8f7d20d011e6601fede1/torchaudio-2.10.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:cfb2ad4b7847d81931989127d803487263c8284f21156e9000daec1ac16c0831", size = 390449, upload-time = "2026-01-21T16:28:48.585Z" }, + { url = "https://files.pythonhosted.org/packages/43/ac/a14425fddd1cf56bb052a3bfd38880258008f8c3cd17f37bba55b3a88ce7/torchaudio-2.10.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:316cdb15fb37290fca89894b095d97b4dc14a90c4c61148ae5c96bb334d962cd", size = 1891070, upload-time = "2026-01-21T16:28:47.323Z" }, + { url = "https://files.pythonhosted.org/packages/6e/03/d1898db1bf7ecd47ca9b4e1b70927597d236cf721e3736d953d555901832/torchaudio-2.10.0-cp310-cp310-win_amd64.whl", hash = "sha256:013079d1ba2a652184703e671b8339cbc7991f17e4ed927071fe7635f908a4a1", size = 474045, upload-time = "2026-01-21T16:28:46.191Z" }, + { url = "https://files.pythonhosted.org/packages/5c/e7/401fe1d024bf9352371d854be6f339ad9928669e6bc8a5ba08e9dbce81cf/torchaudio-2.10.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:bcab0e39eb18da84cba1a0c87f600abb6ce97c882200cb46e841caea106f037f", size = 736373, upload-time = "2026-01-21T16:28:41.589Z" }, + { url = "https://files.pythonhosted.org/packages/6f/b7/c66dc34a27441d78997e20d0ffe2f5ad73db9f7b1267511be255bb94ac9b/torchaudio-2.10.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:87c841a21e82703ebd4a29170c4e60c25a2b47312dc212930087ad58965ac0c8", size = 391843, upload-time = "2026-01-21T16:28:43.093Z" }, + { url = "https://files.pythonhosted.org/packages/13/ae/a2a34a64947c4fa4a61b4c86d8f36fbcb4ebfec30fdde140267db260f96c/torchaudio-2.10.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:b2c77fb9114dd463dc805560bf55a1ac2a52e219794cc32b7b32cf2aeffd2826", size = 1894140, upload-time = "2026-01-21T16:28:35.892Z" }, + { url = "https://files.pythonhosted.org/packages/69/26/cd2aec609b4f8918e4e85e5c6a3f569bc7b5f72a7ecba3f784077102749c/torchaudio-2.10.0-cp311-cp311-win_amd64.whl", hash = "sha256:4c6e9609046143b30a30183893d23ff1ce5de603dbe914b3cce5cc29f5aa5a9c", size = 474792, upload-time = "2026-01-21T16:28:45.254Z" }, + { url = "https://files.pythonhosted.org/packages/0f/36/28a6f3e857616cf7576bdbf8170e483b8c5d0a1f8d349ecb2b75921236aa/torchaudio-2.10.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:9d0fbdbfd2f621c51d28571050d6d0c7287791034e5c7303b31480af1258f33f", size = 737144, upload-time = "2026-01-21T16:28:44.189Z" }, + { url = "https://files.pythonhosted.org/packages/ea/3f/df620439a76ece170472d41438d11a1545d5db5dc9f1eaeab8c6e055a328/torchaudio-2.10.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:42b148a0921a3721abd1f6ae098b1ec9f89703e555c4f7a0d44da87b8decbcb9", size = 391973, upload-time = "2026-01-21T16:28:39.732Z" }, + { url = "https://files.pythonhosted.org/packages/98/25/e55a30d7138f8fe56ed006df25b0a3c27681f0ec7bc9989e1778e6d559c3/torchaudio-2.10.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:0e77b2956448d63790a99beed0b74ac8b8cd3a94dcdd9ad01974411078f46278", size = 1895234, upload-time = "2026-01-21T16:28:37.034Z" }, + { url = "https://files.pythonhosted.org/packages/be/a0/da53c7d20fac15f66f8838653b91162de1bf21fb40fee88cf839e4ef5174/torchaudio-2.10.0-cp312-cp312-win_amd64.whl", hash = "sha256:7f76a01ecebf1869e1f2c50a261f1cf07e5fccb24402b4e9bbb82d6725b9c7dd", size = 475470, upload-time = "2026-01-21T16:28:40.615Z" }, + { url = "https://files.pythonhosted.org/packages/b6/02/341e7bd588355f82c5180103cb2f8070a72ab1be920ab27553a1135d4aa6/torchaudio-2.10.0-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:8fd38d28ee150c584d3ee3b05f39e021f0ad8a8ec8fec1f26dfe150c9db9b2f5", size = 737164, upload-time = "2026-01-21T16:28:38.354Z" }, + { url = "https://files.pythonhosted.org/packages/49/fd/831c2595c81b17141180ca11ab3c0836cc544ef13e15aa0e7b2cb619e582/torchaudio-2.10.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:5bc39ff3ea341097ce1ab023dd88c9dd8ca5f96ebf48821e7d23766137bb55d7", size = 392757, upload-time = "2026-01-21T16:28:33.631Z" }, + { url = "https://files.pythonhosted.org/packages/8e/d8/405c80c57dc68ca5855bddfaae57c3d84ea7397bf1eb2aa5d59c9fa1d3a9/torchaudio-2.10.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:3057c4286db5673d266124a2a10ca54e19f516772e9057f44573a7da5b85e328", size = 1897099, upload-time = "2026-01-21T16:28:24.793Z" }, + { url = "https://files.pythonhosted.org/packages/73/cf/0e48d67788c935e3b3d00e6f55a930a54a67f432e04c33ef80a38cb764fd/torchaudio-2.10.0-cp313-cp313-win_amd64.whl", hash = "sha256:99e74d1901742bc10961d807fe75c0dd9496f4a4a4ff4bb317c5de4a0b6f24e6", size = 475476, upload-time = "2026-01-21T16:28:28.249Z" }, + { url = "https://files.pythonhosted.org/packages/48/29/30bcce0f17a8279b051b09250993691a828f89a03278306b23571c18df04/torchaudio-2.10.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:6cfe98ef0ea9bee6d6297493ce67ce0c54a38d80caf6535a3ae48900fd5f3769", size = 742449, upload-time = "2026-01-21T16:28:29.556Z" }, + { url = "https://files.pythonhosted.org/packages/43/8c/653e7f67855424bf3b7cbb48335f8316f7fb02bb01a6cab38f6bf9555676/torchaudio-2.10.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:b41b254d958632dc00dc7768431cadda516c91641d798775cbb19bcd4f0d2be4", size = 393430, upload-time = "2026-01-21T16:28:34.855Z" }, + { url = "https://files.pythonhosted.org/packages/8e/1f/f91fcb9dd47a19b720fb48042a2f6f023651948e73726e98fff60d5ed5c7/torchaudio-2.10.0-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:da1081d1018a1e95f5a13947402aeb037cf5ac8861219a6164df004898a96bb1", size = 1897271, upload-time = "2026-01-21T16:28:23.519Z" }, + { url = "https://files.pythonhosted.org/packages/57/27/270c26890f43838e8faa5d3e52f079bd9d9d09f9a535a11cf6b94e20ed21/torchaudio-2.10.0-cp313-cp313t-win_amd64.whl", hash = "sha256:f1afa53146a5655258d3a86e689c6879dfe78581d9bee9ef611ace98722f86bb", size = 478966, upload-time = "2026-01-21T16:28:32.491Z" }, + { url = "https://files.pythonhosted.org/packages/cc/5c/0e54b162bd0d1ec2f87b545553af839f906b940888d0122cdef04b965385/torchaudio-2.10.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1f2897fbf776d55afcb5f6d9b7bdfaea850ca7a129c8f5e4b3a4b025c431130d", size = 739544, upload-time = "2026-01-21T16:28:26.947Z" }, + { url = "https://files.pythonhosted.org/packages/57/a1/ef5571406858f4ea89c18d6ad844d21cb9858708149e6bbd9a789ee30ea5/torchaudio-2.10.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:b2d5e11a2bec08f02a4f5fb7d1902ff82d48c533a27ceedc21e6ade650cf65b3", size = 393061, upload-time = "2026-01-21T16:28:25.802Z" }, + { url = "https://files.pythonhosted.org/packages/9d/0f/a0cf0ebc6f71b1868ea056dd4cd4f1a2244b8da8bc38372a1adc984a7c1f/torchaudio-2.10.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:77f6cf11a3b61af1b0967cd642368ecd30a86d70f622b22410ae6cb42d980b72", size = 1897137, upload-time = "2026-01-21T16:28:15.366Z" }, + { url = "https://files.pythonhosted.org/packages/7f/48/98e6710a4601e190bc923c3683629c29d41fb18a818a9328515541f023ed/torchaudio-2.10.0-cp314-cp314-win_amd64.whl", hash = "sha256:4711c2a86a005685ca3b5da135b2f370d81ac354e3dcb142ef45fe2c78b9c9c4", size = 475154, upload-time = "2026-01-21T16:28:22.438Z" }, + { url = "https://files.pythonhosted.org/packages/c1/9b/cd02f8add38bd98761548b0821a5e54c564117a9bbeafaf95f665ab0fd72/torchaudio-2.10.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:13bdc1bde0c88e999699d1503304a56fc9dea6401b76bc08a5f268368129d46c", size = 742453, upload-time = "2026-01-21T16:28:20.989Z" }, + { url = "https://files.pythonhosted.org/packages/53/8a/946aa07393845b918d318b5e34b3bd0359fd27fc9fac10a85fae2bb86382/torchaudio-2.10.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:ed912de8ec1b400e17a5172badcfcddc601a9cd4e02d200f3a9504fc8e54961c", size = 393434, upload-time = "2026-01-21T16:28:18.668Z" }, + { url = "https://files.pythonhosted.org/packages/e1/68/e37e8fbbae986afa80f8851e08fc017eb8ae5f7b398ee28ed92303da163e/torchaudio-2.10.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:f7aa33a8198e87949896e16ea245ea731906445becdf10130e8823c68494a94a", size = 1897289, upload-time = "2026-01-21T16:28:17.059Z" }, + { url = "https://files.pythonhosted.org/packages/5d/61/0e1f464463b85bc677036faffdfd23493aa17e8c3fc3a649abca8c019701/torchaudio-2.10.0-cp314-cp314t-win_amd64.whl", hash = "sha256:e49f6a18a8552620c4394f8529b7551eda9312d46dfdd3500bd2be459c86aea4", size = 478968, upload-time = "2026-01-21T16:28:19.542Z" }, +] + +[[package]] +name = "torchdata" +version = "0.11.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "requests" }, + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, + { name = "urllib3" }, +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/95/d4/af694ef718aedbe95a72760ab9ff7a6a7a44ace2d7f70c27bfeb67c5c503/torchdata-0.11.0-py3-none-any.whl", hash = "sha256:52b940fbbe0e00fb21cabddf528449d1bec5bfb0d0823b7487b15f951658ee33", size = 61968, upload-time = "2025-02-20T22:26:30.666Z" }, +] + +[[package]] +name = "torchvision" +version = "0.25.0" +source = { registry = "https://download.pytorch.org/whl/cpu" } +resolution-markers = [ + "python_full_version >= '3.13' and sys_platform == 'darwin'", + "python_full_version == '3.12.*' and sys_platform == 'darwin'", + "python_full_version == '3.11.*' and sys_platform == 'darwin'", + "python_full_version < '3.11' and sys_platform == 'darwin'", +] +dependencies = [ + { name = "numpy", marker = "sys_platform == 'darwin'" }, + { name = "pillow", marker = "sys_platform == 'darwin'" }, + { name = "torch", version = "2.10.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" }, +] +wheels = [ + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:7810841916a2a870cbfb581f4084b59479908fb9db6edfdf78139931392d8677" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:a76ce7b8d4fce291a25721ee2f921c783acc6dbd4fc32dc741ed2a1d5a8dde2f" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:724f212a58a0d0d758649ce288601056b5f46a01de545702f42bccc5b25cb0cc" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:6c444119c6af8fa48b79c59f1529fc15a45a2bbf823cf85f851e7203d84f727f" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:80895a40faa5783ce19ed5a692a54062c7888a385490e866324355f8d59b2eb3" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:a64ae543fe0a057a93954af94c239629723db196ed65090e6775f136726b22f8" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:b7df71942e31b74a5161dea3452da35e5fb1d36c752aedf61f2dd3c7be35739a" }, +] + +[[package]] +name = "torchvision" +version = "0.25.0+cpu" +source = { registry = "https://download.pytorch.org/whl/cpu" } +resolution-markers = [ + "python_full_version >= '3.13' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.12.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.11.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux'", +] +dependencies = [ + { name = "numpy", marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "pillow", marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, + { name = "torch", version = "2.10.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin' and sys_platform != 'linux'" }, +] +wheels = [ + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0%2Bcpu-cp310-cp310-win_amd64.whl", hash = "sha256:3e2ae9981e32a5b9db685659d5c7af0f04b159ff20394650a90124baf6ada51a" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0%2Bcpu-cp311-cp311-win_amd64.whl", hash = "sha256:c7eb5f219fdfaf1f65e68c00eb81172ab4fa08a9874dae9dad2bca360da34d0f" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0%2Bcpu-cp312-cp312-win_amd64.whl", hash = "sha256:2d444009c0956669ada149f61ed78f257c1cc96d259efa6acf3929ca96ceb3f0" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0%2Bcpu-cp313-cp313-win_amd64.whl", hash = "sha256:783c8fc580bbfc159bff52f4f72cdd538e42b32956e70dffa42b940db114e151" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0%2Bcpu-cp313-cp313t-win_amd64.whl", hash = "sha256:9212210f417888e6261c040495180f053084812cf873dedba9fc51ff4b24b2d3" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0%2Bcpu-cp314-cp314-win_amd64.whl", hash = "sha256:499eae1e535766391b6ee2d1e6e841239c20e2e6d88203a15b8f9f8d60a1f8bd" }, + { url = "https://download.pytorch.org/whl/cpu/torchvision-0.25.0%2Bcpu-cp314-cp314t-win_amd64.whl", hash = "sha256:fb9f07f6a10f0ac24ac482ae68c6df99110b74a0d80a4c64fddc9753267d8815" }, +] + +[[package]] +name = "torchvision" +version = "0.25.0+cu128" +source = { registry = "https://download.pytorch.org/whl/cu128" } +resolution-markers = [ + "python_full_version >= '3.13' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version >= '3.13' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and platform_machine != 'aarch64' and sys_platform == 'linux'", +] +dependencies = [ + { name = "numpy", marker = "sys_platform == 'linux'" }, + { name = "pillow", marker = "sys_platform == 'linux'" }, + { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux'" }, +] +wheels = [ + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:7fe593ec669dcb74c3749548c21ae302edbb610d1844bf2bae622accdef76dbc" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:43350ca113e9f224dedb90b0dbc31853fde4f322794cbe4a7589fccab5768f0c" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:5d576c65d40198627e0fad03bddeb0ef536371312f2bdfcc804c22fd28fa6018" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:ebf2b495c76097796b9a2eac9290efbcae96e0fd9e5ae52c40eff188610bb440" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:8623e534ef6a815bd6407d4b52dd70c7154e2eda626ad4b9cb895d36c5a3305b" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:1255a0ca2bf987acf9f103b96c5c4cfe3415fc4a1eef17fa08af527a04a4f573" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:12c253520a26483fe3c614f63ff16eca6d9b0b4ebe510699b7d15d88e6c0cd35" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:a9c0de893dce9c2913c9c7ae88a916910f92d02b99da149678806d18e8079f29" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:58b2971b55c761f1d2491bd80fcc4618ea97d363d387a9dd3aff23220cbee264" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:1b6878b043513ea3dea1b90bfb5193455d9b248b8c4d5e66ea9f5d1643a43f13" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:e2dbf9ea9f4b2416822249e96ff3ad873d9a84e51285d6b9967732be3015c523" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:5b7ad3fb6cf03ef2a2fd617cb4b4e41efa9bb0143c67f506c2a3e6765c7b12ad" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:687987fbcb074fd7f7a61cf2b407b1eac07588ace8351a3a36978546a00adc52" }, + { url = "https://download.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:84c5e2cb699235339b8a5c295e974a795244a45d1104ecee658d9d19600cdc75" }, +] + +[[package]] +name = "tqdm" +version = "4.67.1" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "colorama", marker = "sys_platform == 'win32'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/a8/4b/29b4ef32e036bb34e4ab51796dd745cdba7ed47ad142a9f4a1eb8e0c744d/tqdm-4.67.1.tar.gz", hash = "sha256:f8aef9c52c08c13a65f30ea34f4e5aac3fd1a34959879d7e59e63027286627f2", size = 169737, upload-time = "2024-11-24T20:12:22.481Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/d0/30/dc54f88dd4a2b5dc8a0279bdd7270e735851848b762aeb1c1184ed1f6b14/tqdm-4.67.1-py3-none-any.whl", hash = "sha256:26445eca388f82e72884e0d580d5464cd801a3ea01e63e5601bdff9ba6a48de2", size = 78540, upload-time = "2024-11-24T20:12:19.698Z" }, +] + +[[package]] +name = "transformers" +version = "4.57.3" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "filelock" }, + { name = "huggingface-hub" }, + { name = "numpy" }, + { name = "packaging" }, + { name = "pyyaml" }, + { name = "regex" }, + { name = "requests" }, + { name = "safetensors" }, + { name = "tokenizers" }, + { name = "tqdm" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/dd/70/d42a739e8dfde3d92bb2fff5819cbf331fe9657323221e79415cd5eb65ee/transformers-4.57.3.tar.gz", hash = "sha256:df4945029aaddd7c09eec5cad851f30662f8bd1746721b34cc031d70c65afebc", size = 10139680, upload-time = "2025-11-25T15:51:30.139Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/6a/6b/2f416568b3c4c91c96e5a365d164f8a4a4a88030aa8ab4644181fdadce97/transformers-4.57.3-py3-none-any.whl", hash = "sha256:c77d353a4851b1880191603d36acb313411d3577f6e2897814f333841f7003f4", size = 11993463, upload-time = "2025-11-25T15:51:26.493Z" }, +] + +[[package]] +name = "triton" +version = "3.5.1" +source = { registry = "https://pypi.org/simple" } +resolution-markers = [ + "python_full_version >= '3.13' and sys_platform == 'darwin'", + "python_full_version >= '3.13' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.12.*' and sys_platform == 'darwin'", + "python_full_version == '3.12.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version == '3.11.*' and sys_platform == 'darwin'", + "python_full_version == '3.11.*' and sys_platform != 'darwin' and sys_platform != 'linux'", + "python_full_version < '3.11' and sys_platform == 'darwin'", + "python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux'", +] + +[[package]] +name = "triton" +version = "3.6.0" +source = { registry = "https://pypi.org/simple" } +resolution-markers = [ + "python_full_version >= '3.13' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version >= '3.13' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.12.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version == '3.11.*' and platform_machine != 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and platform_machine == 'aarch64' and sys_platform == 'linux'", + "python_full_version < '3.11' and platform_machine != 'aarch64' and sys_platform == 'linux'", +] +wheels = [ + { url = "https://files.pythonhosted.org/packages/44/ba/b1b04f4b291a3205d95ebd24465de0e5bf010a2df27a4e58a9b5f039d8f2/triton-3.6.0-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6c723cfb12f6842a0ae94ac307dba7e7a44741d720a40cf0e270ed4a4e3be781", size = 175972180, upload-time = "2026-01-20T16:15:53.664Z" }, + { url = "https://files.pythonhosted.org/packages/8c/f7/f1c9d3424ab199ac53c2da567b859bcddbb9c9e7154805119f8bd95ec36f/triton-3.6.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a6550fae429e0667e397e5de64b332d1e5695b73650ee75a6146e2e902770bea", size = 188105201, upload-time = "2026-01-20T16:00:29.272Z" }, + { url = "https://files.pythonhosted.org/packages/0f/2c/96f92f3c60387e14cc45aed49487f3486f89ea27106c1b1376913c62abe4/triton-3.6.0-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:49df5ef37379c0c2b5c0012286f80174fcf0e073e5ade1ca9a86c36814553651", size = 176081190, upload-time = "2026-01-20T16:16:00.523Z" }, + { url = "https://files.pythonhosted.org/packages/e0/12/b05ba554d2c623bffa59922b94b0775673de251f468a9609bc9e45de95e9/triton-3.6.0-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e8e323d608e3a9bfcc2d9efcc90ceefb764a82b99dea12a86d643c72539ad5d3", size = 188214640, upload-time = "2026-01-20T16:00:35.869Z" }, + { url = "https://files.pythonhosted.org/packages/17/5d/08201db32823bdf77a0e2b9039540080b2e5c23a20706ddba942924ebcd6/triton-3.6.0-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:374f52c11a711fd062b4bfbb201fd9ac0a5febd28a96fb41b4a0f51dde3157f4", size = 176128243, upload-time = "2026-01-20T16:16:07.857Z" }, + { url = "https://files.pythonhosted.org/packages/ab/a8/cdf8b3e4c98132f965f88c2313a4b493266832ad47fb52f23d14d4f86bb5/triton-3.6.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:74caf5e34b66d9f3a429af689c1c7128daba1d8208df60e81106b115c00d6fca", size = 188266850, upload-time = "2026-01-20T16:00:43.041Z" }, + { url = "https://files.pythonhosted.org/packages/3c/12/34d71b350e89a204c2c7777a9bba0dcf2f19a5bfdd70b57c4dbc5ffd7154/triton-3.6.0-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:448e02fe6dc898e9e5aa89cf0ee5c371e99df5aa5e8ad976a80b93334f3494fd", size = 176133521, upload-time = "2026-01-20T16:16:13.321Z" }, + { url = "https://files.pythonhosted.org/packages/f9/0b/37d991d8c130ce81a8728ae3c25b6e60935838e9be1b58791f5997b24a54/triton-3.6.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:10c7f76c6e72d2ef08df639e3d0d30729112f47a56b0c81672edc05ee5116ac9", size = 188289450, upload-time = "2026-01-20T16:00:49.136Z" }, + { url = "https://files.pythonhosted.org/packages/ce/4e/41b0c8033b503fd3cfcd12392cdd256945026a91ff02452bef40ec34bee7/triton-3.6.0-cp313-cp313t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1722e172d34e32abc3eb7711d0025bb69d7959ebea84e3b7f7a341cd7ed694d6", size = 176276087, upload-time = "2026-01-20T16:16:18.989Z" }, + { url = "https://files.pythonhosted.org/packages/35/f8/9c66bfc55361ec6d0e4040a0337fb5924ceb23de4648b8a81ae9d33b2b38/triton-3.6.0-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d002e07d7180fd65e622134fbd980c9a3d4211fb85224b56a0a0efbd422ab72f", size = 188400296, upload-time = "2026-01-20T16:00:56.042Z" }, + { url = "https://files.pythonhosted.org/packages/49/55/5ecf0dcaa0f2fbbd4420f7ef227ee3cb172e91e5fede9d0ecaddc43363b4/triton-3.6.0-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ef5523241e7d1abca00f1d240949eebdd7c673b005edbbce0aca95b8191f1d43", size = 176138577, upload-time = "2026-01-20T16:16:25.426Z" }, + { url = "https://files.pythonhosted.org/packages/df/3d/9e7eee57b37c80cec63322c0231bb6da3cfe535a91d7a4d64896fcb89357/triton-3.6.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a17a5d5985f0ac494ed8a8e54568f092f7057ef60e1b0fa09d3fd1512064e803", size = 188273063, upload-time = "2026-01-20T16:01:07.278Z" }, + { url = "https://files.pythonhosted.org/packages/48/db/56ee649cab5eaff4757541325aca81f52d02d4a7cd3506776cad2451e060/triton-3.6.0-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0b3a97e8ed304dfa9bd23bb41ca04cdf6b2e617d5e782a8653d616037a5d537d", size = 176274804, upload-time = "2026-01-20T16:16:31.528Z" }, + { url = "https://files.pythonhosted.org/packages/f6/56/6113c23ff46c00aae423333eb58b3e60bdfe9179d542781955a5e1514cb3/triton-3.6.0-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:46bd1c1af4b6704e554cad2eeb3b0a6513a980d470ccfa63189737340c7746a7", size = 188397994, upload-time = "2026-01-20T16:01:14.236Z" }, +] + +[[package]] +name = "typer" +version = "0.20.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "click" }, + { name = "rich" }, + { name = "shellingham" }, + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/8f/28/7c85c8032b91dbe79725b6f17d2fffc595dff06a35c7a30a37bef73a1ab4/typer-0.20.0.tar.gz", hash = "sha256:1aaf6494031793e4876fb0bacfa6a912b551cf43c1e63c800df8b1a866720c37", size = 106492, upload-time = "2025-10-20T17:03:49.445Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/78/64/7713ffe4b5983314e9d436a90d5bd4f63b6054e2aca783a3cfc44cb95bbf/typer-0.20.0-py3-none-any.whl", hash = "sha256:5b463df6793ec1dca6213a3cf4c0f03bc6e322ac5e16e13ddd622a889489784a", size = 47028, upload-time = "2025-10-20T17:03:47.617Z" }, +] + +[[package]] +name = "typing-extensions" +version = "4.15.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/72/94/1a15dd82efb362ac84269196e94cf00f187f7ed21c242792a923cdb1c61f/typing_extensions-4.15.0.tar.gz", hash = "sha256:0cea48d173cc12fa28ecabc3b837ea3cf6f38c6d1136f85cbaaf598984861466", size = 109391, upload-time = "2025-08-25T13:49:26.313Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/18/67/36e9267722cc04a6b9f15c7f3441c2363321a3ea07da7ae0c0707beb2a9c/typing_extensions-4.15.0-py3-none-any.whl", hash = "sha256:f0fa19c6845758ab08074a0cfa8b7aecb71c999ca73d62883bc25cc018c4e548", size = 44614, upload-time = "2025-08-25T13:49:24.86Z" }, +] + +[[package]] +name = "typing-inspection" +version = "0.4.2" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/55/e3/70399cb7dd41c10ac53367ae42139cf4b1ca5f36bb3dc6c9d33acdb43655/typing_inspection-0.4.2.tar.gz", hash = "sha256:ba561c48a67c5958007083d386c3295464928b01faa735ab8547c5692e87f464", size = 75949, upload-time = "2025-10-01T02:14:41.687Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/dc/9b/47798a6c91d8bdb567fe2698fe81e0c6b7cb7ef4d13da4114b41d239f65d/typing_inspection-0.4.2-py3-none-any.whl", hash = "sha256:4ed1cacbdc298c220f1bd249ed5287caa16f34d44ef4e9c3d0cbad5b521545e7", size = 14611, upload-time = "2025-10-01T02:14:40.154Z" }, +] + +[[package]] +name = "tzdata" +version = "2025.2" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/95/32/1a225d6164441be760d75c2c42e2780dc0873fe382da3e98a2e1e48361e5/tzdata-2025.2.tar.gz", hash = "sha256:b60a638fcc0daffadf82fe0f57e53d06bdec2f36c4df66280ae79bce6bd6f2b9", size = 196380, upload-time = "2025-03-23T13:54:43.652Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/5c/23/c7abc0ca0a1526a0774eca151daeb8de62ec457e77262b66b359c3c7679e/tzdata-2025.2-py2.py3-none-any.whl", hash = "sha256:1a403fada01ff9221ca8044d701868fa132215d84beb92242d9acd2147f667a8", size = 347839, upload-time = "2025-03-23T13:54:41.845Z" }, +] + +[[package]] +name = "urllib3" +version = "2.5.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/15/22/9ee70a2574a4f4599c47dd506532914ce044817c7752a79b6a51286319bc/urllib3-2.5.0.tar.gz", hash = "sha256:3fc47733c7e419d4bc3f6b3dc2b4f890bb743906a30d56ba4a5bfa4bbff92760", size = 393185, upload-time = "2025-06-18T14:07:41.644Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/a7/c2/fe1e52489ae3122415c51f387e221dd0773709bad6c6cdaa599e8a2c5185/urllib3-2.5.0-py3-none-any.whl", hash = "sha256:e6b01673c0fa6a13e374b50871808eb3bf7046c4b125b216f6bf1cc604cff0dc", size = 129795, upload-time = "2025-06-18T14:07:40.39Z" }, +] + +[[package]] +name = "uvicorn" +version = "0.38.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "click" }, + { name = "h11" }, + { name = "typing-extensions", marker = "python_full_version < '3.11'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/cb/ce/f06b84e2697fef4688ca63bdb2fdf113ca0a3be33f94488f2cadb690b0cf/uvicorn-0.38.0.tar.gz", hash = "sha256:fd97093bdd120a2609fc0d3afe931d4d4ad688b6e75f0f929fde1bc36fe0e91d", size = 80605, upload-time = "2025-10-18T13:46:44.63Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/ee/d9/d88e73ca598f4f6ff671fb5fde8a32925c2e08a637303a1d12883c7305fa/uvicorn-0.38.0-py3-none-any.whl", hash = "sha256:48c0afd214ceb59340075b4a052ea1ee91c16fbc2a9b1469cca0e54566977b02", size = 68109, upload-time = "2025-10-18T13:46:42.958Z" }, +] + +[[package]] +name = "virtualenv" +version = "20.35.4" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "distlib" }, + { name = "filelock" }, + { name = "platformdirs" }, + { name = "typing-extensions", marker = "python_full_version < '3.11'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/20/28/e6f1a6f655d620846bd9df527390ecc26b3805a0c5989048c210e22c5ca9/virtualenv-20.35.4.tar.gz", hash = "sha256:643d3914d73d3eeb0c552cbb12d7e82adf0e504dbf86a3182f8771a153a1971c", size = 6028799, upload-time = "2025-10-29T06:57:40.511Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/79/0c/c05523fa3181fdf0c9c52a6ba91a23fbf3246cc095f26f6516f9c60e6771/virtualenv-20.35.4-py3-none-any.whl", hash = "sha256:c21c9cede36c9753eeade68ba7d523529f228a403463376cf821eaae2b650f1b", size = 6005095, upload-time = "2025-10-29T06:57:37.598Z" }, +] + +[[package]] +name = "wandb" +version = "0.23.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "click" }, + { name = "gitpython" }, + { name = "packaging" }, + { name = "platformdirs" }, + { name = "protobuf" }, + { name = "pydantic" }, + { name = "pyyaml" }, + { name = "requests" }, + { name = "sentry-sdk" }, + { name = "typing-extensions" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/ef/8b/db2d44395c967cd452517311fd6ede5d1e07310769f448358d4874248512/wandb-0.23.0.tar.gz", hash = "sha256:e5f98c61a8acc3ee84583ca78057f64344162ce026b9f71cb06eea44aec27c93", size = 44413921, upload-time = "2025-11-11T21:06:30.737Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/41/61/a3220c7fa4cadfb2b2a5c09e3fa401787326584ade86d7c1f58bf1cd43bd/wandb-0.23.0-py3-none-macosx_12_0_arm64.whl", hash = "sha256:b682ec5e38fc97bd2e868ac7615a0ab4fc6a15220ee1159e87270a5ebb7a816d", size = 18992250, upload-time = "2025-11-11T21:06:03.412Z" }, + { url = "https://files.pythonhosted.org/packages/90/16/e69333cf3d11e7847f424afc6c8ae325e1f6061b2e5118d7a17f41b6525d/wandb-0.23.0-py3-none-macosx_12_0_x86_64.whl", hash = "sha256:ec094eb71b778e77db8c188da19e52c4f96cb9d5b4421d7dc05028afc66fd7e7", size = 20045616, upload-time = "2025-11-11T21:06:07.109Z" }, + { url = "https://files.pythonhosted.org/packages/62/79/42dc6c7bb0b425775fe77f1a3f1a22d75d392841a06b43e150a3a7f2553a/wandb-0.23.0-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4e43f1f04b98c34f407dcd2744cec0a590abce39bed14a61358287f817514a7b", size = 18758848, upload-time = "2025-11-11T21:06:09.832Z" }, + { url = "https://files.pythonhosted.org/packages/b8/94/d6ddb78334996ccfc1179444bfcfc0f37ffd07ee79bb98940466da6f68f8/wandb-0.23.0-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6e5847f98cbb3175caf5291932374410141f5bb3b7c25f9c5e562c1988ce0bf5", size = 20231493, upload-time = "2025-11-11T21:06:12.323Z" }, + { url = "https://files.pythonhosted.org/packages/52/4d/0ad6df0e750c19dabd24d2cecad0938964f69a072f05fbdab7281bec2b64/wandb-0.23.0-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:6151355fd922539926e870be811474238c9614b96541773b990f1ce53368aef6", size = 18793473, upload-time = "2025-11-11T21:06:14.967Z" }, + { url = "https://files.pythonhosted.org/packages/f8/da/c2ba49c5573dff93dafc0acce691bb1c3d57361bf834b2f2c58e6193439b/wandb-0.23.0-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:df62e426e448ebc44269140deb7240df474e743b12d4b1f53b753afde4aa06d4", size = 20332882, upload-time = "2025-11-11T21:06:17.865Z" }, + { url = "https://files.pythonhosted.org/packages/40/65/21bfb10ee5cd93fbcaf794958863c7e05bac4bbeb1cc1b652094aa3743a5/wandb-0.23.0-py3-none-win32.whl", hash = "sha256:6c21d3eadda17aef7df6febdffdddfb0b4835c7754435fc4fe27631724269f5c", size = 19433198, upload-time = "2025-11-11T21:06:21.913Z" }, + { url = "https://files.pythonhosted.org/packages/f1/33/cbe79e66c171204e32cf940c7fdfb8b5f7d2af7a00f301c632f3a38aa84b/wandb-0.23.0-py3-none-win_amd64.whl", hash = "sha256:b50635fa0e16e528bde25715bf446e9153368428634ca7a5dbd7a22c8ae4e915", size = 19433201, upload-time = "2025-11-11T21:06:24.607Z" }, + { url = "https://files.pythonhosted.org/packages/1c/a0/5ecfae12d78ea036a746c071e4c13b54b28d641efbba61d2947c73b3e6f9/wandb-0.23.0-py3-none-win_arm64.whl", hash = "sha256:fa0181b02ce4d1993588f4a728d8b73ae487eb3cb341e6ce01c156be7a98ec72", size = 17678649, upload-time = "2025-11-11T21:06:27.289Z" }, +] + +[[package]] +name = "watch" +version = "0.2.7" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/4d/30/05e13da086995c5d9fde762ff3cb742884aa76fb6167eda9a73a876aee8b/watch-0.2.7.tar.gz", hash = "sha256:d51f3648091c031e08d0cf0f3177da884078f25538a9d8dc568693ab5ac2dee0", size = 6441, upload-time = "2018-10-02T21:28:53.915Z" } + +[[package]] +name = "wcwidth" +version = "0.2.14" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/24/30/6b0809f4510673dc723187aeaf24c7f5459922d01e2f794277a3dfb90345/wcwidth-0.2.14.tar.gz", hash = "sha256:4d478375d31bc5395a3c55c40ccdf3354688364cd61c4f6adacaa9215d0b3605", size = 102293, upload-time = "2025-09-22T16:29:53.023Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/af/b5/123f13c975e9f27ab9c0770f514345bd406d0e8d3b7a0723af9d43f710af/wcwidth-0.2.14-py2.py3-none-any.whl", hash = "sha256:a7bb560c8aee30f9957e5f9895805edd20602f2d7f720186dfd906e82b4982e1", size = 37286, upload-time = "2025-09-22T16:29:51.641Z" }, +] + +[[package]] +name = "websockets" +version = "15.0.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/21/e6/26d09fab466b7ca9c7737474c52be4f76a40301b08362eb2dbc19dcc16c1/websockets-15.0.1.tar.gz", hash = "sha256:82544de02076bafba038ce055ee6412d68da13ab47f0c60cab827346de828dee", size = 177016, upload-time = "2025-03-05T20:03:41.606Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/1e/da/6462a9f510c0c49837bbc9345aca92d767a56c1fb2939e1579df1e1cdcf7/websockets-15.0.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:d63efaa0cd96cf0c5fe4d581521d9fa87744540d4bc999ae6e08595a1014b45b", size = 175423, upload-time = "2025-03-05T20:01:35.363Z" }, + { url = "https://files.pythonhosted.org/packages/1c/9f/9d11c1a4eb046a9e106483b9ff69bce7ac880443f00e5ce64261b47b07e7/websockets-15.0.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ac60e3b188ec7574cb761b08d50fcedf9d77f1530352db4eef1707fe9dee7205", size = 173080, upload-time = "2025-03-05T20:01:37.304Z" }, + { url = "https://files.pythonhosted.org/packages/d5/4f/b462242432d93ea45f297b6179c7333dd0402b855a912a04e7fc61c0d71f/websockets-15.0.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5756779642579d902eed757b21b0164cd6fe338506a8083eb58af5c372e39d9a", size = 173329, upload-time = "2025-03-05T20:01:39.668Z" }, + { url = "https://files.pythonhosted.org/packages/6e/0c/6afa1f4644d7ed50284ac59cc70ef8abd44ccf7d45850d989ea7310538d0/websockets-15.0.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0fdfe3e2a29e4db3659dbd5bbf04560cea53dd9610273917799f1cde46aa725e", size = 182312, upload-time = "2025-03-05T20:01:41.815Z" }, + { url = "https://files.pythonhosted.org/packages/dd/d4/ffc8bd1350b229ca7a4db2a3e1c482cf87cea1baccd0ef3e72bc720caeec/websockets-15.0.1-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4c2529b320eb9e35af0fa3016c187dffb84a3ecc572bcee7c3ce302bfeba52bf", size = 181319, upload-time = "2025-03-05T20:01:43.967Z" }, + { url = "https://files.pythonhosted.org/packages/97/3a/5323a6bb94917af13bbb34009fac01e55c51dfde354f63692bf2533ffbc2/websockets-15.0.1-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ac1e5c9054fe23226fb11e05a6e630837f074174c4c2f0fe442996112a6de4fb", size = 181631, upload-time = "2025-03-05T20:01:46.104Z" }, + { url = "https://files.pythonhosted.org/packages/a6/cc/1aeb0f7cee59ef065724041bb7ed667b6ab1eeffe5141696cccec2687b66/websockets-15.0.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:5df592cd503496351d6dc14f7cdad49f268d8e618f80dce0cd5a36b93c3fc08d", size = 182016, upload-time = "2025-03-05T20:01:47.603Z" }, + { url = "https://files.pythonhosted.org/packages/79/f9/c86f8f7af208e4161a7f7e02774e9d0a81c632ae76db2ff22549e1718a51/websockets-15.0.1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:0a34631031a8f05657e8e90903e656959234f3a04552259458aac0b0f9ae6fd9", size = 181426, upload-time = "2025-03-05T20:01:48.949Z" }, + { url = "https://files.pythonhosted.org/packages/c7/b9/828b0bc6753db905b91df6ae477c0b14a141090df64fb17f8a9d7e3516cf/websockets-15.0.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3d00075aa65772e7ce9e990cab3ff1de702aa09be3940d1dc88d5abf1ab8a09c", size = 181360, upload-time = "2025-03-05T20:01:50.938Z" }, + { url = "https://files.pythonhosted.org/packages/89/fb/250f5533ec468ba6327055b7d98b9df056fb1ce623b8b6aaafb30b55d02e/websockets-15.0.1-cp310-cp310-win32.whl", hash = "sha256:1234d4ef35db82f5446dca8e35a7da7964d02c127b095e172e54397fb6a6c256", size = 176388, upload-time = "2025-03-05T20:01:52.213Z" }, + { url = "https://files.pythonhosted.org/packages/1c/46/aca7082012768bb98e5608f01658ff3ac8437e563eca41cf068bd5849a5e/websockets-15.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:39c1fec2c11dc8d89bba6b2bf1556af381611a173ac2b511cf7231622058af41", size = 176830, upload-time = "2025-03-05T20:01:53.922Z" }, + { url = "https://files.pythonhosted.org/packages/9f/32/18fcd5919c293a398db67443acd33fde142f283853076049824fc58e6f75/websockets-15.0.1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:823c248b690b2fd9303ba00c4f66cd5e2d8c3ba4aa968b2779be9532a4dad431", size = 175423, upload-time = "2025-03-05T20:01:56.276Z" }, + { url = "https://files.pythonhosted.org/packages/76/70/ba1ad96b07869275ef42e2ce21f07a5b0148936688c2baf7e4a1f60d5058/websockets-15.0.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:678999709e68425ae2593acf2e3ebcbcf2e69885a5ee78f9eb80e6e371f1bf57", size = 173082, upload-time = "2025-03-05T20:01:57.563Z" }, + { url = "https://files.pythonhosted.org/packages/86/f2/10b55821dd40eb696ce4704a87d57774696f9451108cff0d2824c97e0f97/websockets-15.0.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:d50fd1ee42388dcfb2b3676132c78116490976f1300da28eb629272d5d93e905", size = 173330, upload-time = "2025-03-05T20:01:59.063Z" }, + { url = "https://files.pythonhosted.org/packages/a5/90/1c37ae8b8a113d3daf1065222b6af61cc44102da95388ac0018fcb7d93d9/websockets-15.0.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d99e5546bf73dbad5bf3547174cd6cb8ba7273062a23808ffea025ecb1cf8562", size = 182878, upload-time = "2025-03-05T20:02:00.305Z" }, + { url = "https://files.pythonhosted.org/packages/8e/8d/96e8e288b2a41dffafb78e8904ea7367ee4f891dafc2ab8d87e2124cb3d3/websockets-15.0.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:66dd88c918e3287efc22409d426c8f729688d89a0c587c88971a0faa2c2f3792", size = 181883, upload-time = "2025-03-05T20:02:03.148Z" }, + { url = "https://files.pythonhosted.org/packages/93/1f/5d6dbf551766308f6f50f8baf8e9860be6182911e8106da7a7f73785f4c4/websockets-15.0.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8dd8327c795b3e3f219760fa603dcae1dcc148172290a8ab15158cf85a953413", size = 182252, upload-time = "2025-03-05T20:02:05.29Z" }, + { url = "https://files.pythonhosted.org/packages/d4/78/2d4fed9123e6620cbf1706c0de8a1632e1a28e7774d94346d7de1bba2ca3/websockets-15.0.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:8fdc51055e6ff4adeb88d58a11042ec9a5eae317a0a53d12c062c8a8865909e8", size = 182521, upload-time = "2025-03-05T20:02:07.458Z" }, + { url = "https://files.pythonhosted.org/packages/e7/3b/66d4c1b444dd1a9823c4a81f50231b921bab54eee2f69e70319b4e21f1ca/websockets-15.0.1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:693f0192126df6c2327cce3baa7c06f2a117575e32ab2308f7f8216c29d9e2e3", size = 181958, upload-time = "2025-03-05T20:02:09.842Z" }, + { url = "https://files.pythonhosted.org/packages/08/ff/e9eed2ee5fed6f76fdd6032ca5cd38c57ca9661430bb3d5fb2872dc8703c/websockets-15.0.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:54479983bd5fb469c38f2f5c7e3a24f9a4e70594cd68cd1fa6b9340dadaff7cf", size = 181918, upload-time = "2025-03-05T20:02:11.968Z" }, + { url = "https://files.pythonhosted.org/packages/d8/75/994634a49b7e12532be6a42103597b71098fd25900f7437d6055ed39930a/websockets-15.0.1-cp311-cp311-win32.whl", hash = "sha256:16b6c1b3e57799b9d38427dda63edcbe4926352c47cf88588c0be4ace18dac85", size = 176388, upload-time = "2025-03-05T20:02:13.32Z" }, + { url = "https://files.pythonhosted.org/packages/98/93/e36c73f78400a65f5e236cd376713c34182e6663f6889cd45a4a04d8f203/websockets-15.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:27ccee0071a0e75d22cb35849b1db43f2ecd3e161041ac1ee9d2352ddf72f065", size = 176828, upload-time = "2025-03-05T20:02:14.585Z" }, + { url = "https://files.pythonhosted.org/packages/51/6b/4545a0d843594f5d0771e86463606a3988b5a09ca5123136f8a76580dd63/websockets-15.0.1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:3e90baa811a5d73f3ca0bcbf32064d663ed81318ab225ee4f427ad4e26e5aff3", size = 175437, upload-time = "2025-03-05T20:02:16.706Z" }, + { url = "https://files.pythonhosted.org/packages/f4/71/809a0f5f6a06522af902e0f2ea2757f71ead94610010cf570ab5c98e99ed/websockets-15.0.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:592f1a9fe869c778694f0aa806ba0374e97648ab57936f092fd9d87f8bc03665", size = 173096, upload-time = "2025-03-05T20:02:18.832Z" }, + { url = "https://files.pythonhosted.org/packages/3d/69/1a681dd6f02180916f116894181eab8b2e25b31e484c5d0eae637ec01f7c/websockets-15.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0701bc3cfcb9164d04a14b149fd74be7347a530ad3bbf15ab2c678a2cd3dd9a2", size = 173332, upload-time = "2025-03-05T20:02:20.187Z" }, + { url = "https://files.pythonhosted.org/packages/a6/02/0073b3952f5bce97eafbb35757f8d0d54812b6174ed8dd952aa08429bcc3/websockets-15.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e8b56bdcdb4505c8078cb6c7157d9811a85790f2f2b3632c7d1462ab5783d215", size = 183152, upload-time = "2025-03-05T20:02:22.286Z" }, + { url = "https://files.pythonhosted.org/packages/74/45/c205c8480eafd114b428284840da0b1be9ffd0e4f87338dc95dc6ff961a1/websockets-15.0.1-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:0af68c55afbd5f07986df82831c7bff04846928ea8d1fd7f30052638788bc9b5", size = 182096, upload-time = "2025-03-05T20:02:24.368Z" }, + { url = "https://files.pythonhosted.org/packages/14/8f/aa61f528fba38578ec553c145857a181384c72b98156f858ca5c8e82d9d3/websockets-15.0.1-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:64dee438fed052b52e4f98f76c5790513235efaa1ef7f3f2192c392cd7c91b65", size = 182523, upload-time = "2025-03-05T20:02:25.669Z" }, + { url = "https://files.pythonhosted.org/packages/ec/6d/0267396610add5bc0d0d3e77f546d4cd287200804fe02323797de77dbce9/websockets-15.0.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:d5f6b181bb38171a8ad1d6aa58a67a6aa9d4b38d0f8c5f496b9e42561dfc62fe", size = 182790, upload-time = "2025-03-05T20:02:26.99Z" }, + { url = "https://files.pythonhosted.org/packages/02/05/c68c5adbf679cf610ae2f74a9b871ae84564462955d991178f95a1ddb7dd/websockets-15.0.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:5d54b09eba2bada6011aea5375542a157637b91029687eb4fdb2dab11059c1b4", size = 182165, upload-time = "2025-03-05T20:02:30.291Z" }, + { url = "https://files.pythonhosted.org/packages/29/93/bb672df7b2f5faac89761cb5fa34f5cec45a4026c383a4b5761c6cea5c16/websockets-15.0.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:3be571a8b5afed347da347bfcf27ba12b069d9d7f42cb8c7028b5e98bbb12597", size = 182160, upload-time = "2025-03-05T20:02:31.634Z" }, + { url = "https://files.pythonhosted.org/packages/ff/83/de1f7709376dc3ca9b7eeb4b9a07b4526b14876b6d372a4dc62312bebee0/websockets-15.0.1-cp312-cp312-win32.whl", hash = "sha256:c338ffa0520bdb12fbc527265235639fb76e7bc7faafbb93f6ba80d9c06578a9", size = 176395, upload-time = "2025-03-05T20:02:33.017Z" }, + { url = "https://files.pythonhosted.org/packages/7d/71/abf2ebc3bbfa40f391ce1428c7168fb20582d0ff57019b69ea20fa698043/websockets-15.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:fcd5cf9e305d7b8338754470cf69cf81f420459dbae8a3b40cee57417f4614a7", size = 176841, upload-time = "2025-03-05T20:02:34.498Z" }, + { url = "https://files.pythonhosted.org/packages/cb/9f/51f0cf64471a9d2b4d0fc6c534f323b664e7095640c34562f5182e5a7195/websockets-15.0.1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:ee443ef070bb3b6ed74514f5efaa37a252af57c90eb33b956d35c8e9c10a1931", size = 175440, upload-time = "2025-03-05T20:02:36.695Z" }, + { url = "https://files.pythonhosted.org/packages/8a/05/aa116ec9943c718905997412c5989f7ed671bc0188ee2ba89520e8765d7b/websockets-15.0.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:5a939de6b7b4e18ca683218320fc67ea886038265fd1ed30173f5ce3f8e85675", size = 173098, upload-time = "2025-03-05T20:02:37.985Z" }, + { url = "https://files.pythonhosted.org/packages/ff/0b/33cef55ff24f2d92924923c99926dcce78e7bd922d649467f0eda8368923/websockets-15.0.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:746ee8dba912cd6fc889a8147168991d50ed70447bf18bcda7039f7d2e3d9151", size = 173329, upload-time = "2025-03-05T20:02:39.298Z" }, + { url = "https://files.pythonhosted.org/packages/31/1d/063b25dcc01faa8fada1469bdf769de3768b7044eac9d41f734fd7b6ad6d/websockets-15.0.1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:595b6c3969023ecf9041b2936ac3827e4623bfa3ccf007575f04c5a6aa318c22", size = 183111, upload-time = "2025-03-05T20:02:40.595Z" }, + { url = "https://files.pythonhosted.org/packages/93/53/9a87ee494a51bf63e4ec9241c1ccc4f7c2f45fff85d5bde2ff74fcb68b9e/websockets-15.0.1-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:3c714d2fc58b5ca3e285461a4cc0c9a66bd0e24c5da9911e30158286c9b5be7f", size = 182054, upload-time = "2025-03-05T20:02:41.926Z" }, + { url = "https://files.pythonhosted.org/packages/ff/b2/83a6ddf56cdcbad4e3d841fcc55d6ba7d19aeb89c50f24dd7e859ec0805f/websockets-15.0.1-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0f3c1e2ab208db911594ae5b4f79addeb3501604a165019dd221c0bdcabe4db8", size = 182496, upload-time = "2025-03-05T20:02:43.304Z" }, + { url = "https://files.pythonhosted.org/packages/98/41/e7038944ed0abf34c45aa4635ba28136f06052e08fc2168520bb8b25149f/websockets-15.0.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:229cf1d3ca6c1804400b0a9790dc66528e08a6a1feec0d5040e8b9eb14422375", size = 182829, upload-time = "2025-03-05T20:02:48.812Z" }, + { url = "https://files.pythonhosted.org/packages/e0/17/de15b6158680c7623c6ef0db361da965ab25d813ae54fcfeae2e5b9ef910/websockets-15.0.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:756c56e867a90fb00177d530dca4b097dd753cde348448a1012ed6c5131f8b7d", size = 182217, upload-time = "2025-03-05T20:02:50.14Z" }, + { url = "https://files.pythonhosted.org/packages/33/2b/1f168cb6041853eef0362fb9554c3824367c5560cbdaad89ac40f8c2edfc/websockets-15.0.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:558d023b3df0bffe50a04e710bc87742de35060580a293c2a984299ed83bc4e4", size = 182195, upload-time = "2025-03-05T20:02:51.561Z" }, + { url = "https://files.pythonhosted.org/packages/86/eb/20b6cdf273913d0ad05a6a14aed4b9a85591c18a987a3d47f20fa13dcc47/websockets-15.0.1-cp313-cp313-win32.whl", hash = "sha256:ba9e56e8ceeeedb2e080147ba85ffcd5cd0711b89576b83784d8605a7df455fa", size = 176393, upload-time = "2025-03-05T20:02:53.814Z" }, + { url = "https://files.pythonhosted.org/packages/1b/6c/c65773d6cab416a64d191d6ee8a8b1c68a09970ea6909d16965d26bfed1e/websockets-15.0.1-cp313-cp313-win_amd64.whl", hash = "sha256:e09473f095a819042ecb2ab9465aee615bd9c2028e4ef7d933600a8401c79561", size = 176837, upload-time = "2025-03-05T20:02:55.237Z" }, + { url = "https://files.pythonhosted.org/packages/02/9e/d40f779fa16f74d3468357197af8d6ad07e7c5a27ea1ca74ceb38986f77a/websockets-15.0.1-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:0c9e74d766f2818bb95f84c25be4dea09841ac0f734d1966f415e4edfc4ef1c3", size = 173109, upload-time = "2025-03-05T20:03:17.769Z" }, + { url = "https://files.pythonhosted.org/packages/bc/cd/5b887b8585a593073fd92f7c23ecd3985cd2c3175025a91b0d69b0551372/websockets-15.0.1-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:1009ee0c7739c08a0cd59de430d6de452a55e42d6b522de7aa15e6f67db0b8e1", size = 173343, upload-time = "2025-03-05T20:03:19.094Z" }, + { url = "https://files.pythonhosted.org/packages/fe/ae/d34f7556890341e900a95acf4886833646306269f899d58ad62f588bf410/websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:76d1f20b1c7a2fa82367e04982e708723ba0e7b8d43aa643d3dcd404d74f1475", size = 174599, upload-time = "2025-03-05T20:03:21.1Z" }, + { url = "https://files.pythonhosted.org/packages/71/e6/5fd43993a87db364ec60fc1d608273a1a465c0caba69176dd160e197ce42/websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f29d80eb9a9263b8d109135351caf568cc3f80b9928bccde535c235de55c22d9", size = 174207, upload-time = "2025-03-05T20:03:23.221Z" }, + { url = "https://files.pythonhosted.org/packages/2b/fb/c492d6daa5ec067c2988ac80c61359ace5c4c674c532985ac5a123436cec/websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b359ed09954d7c18bbc1680f380c7301f92c60bf924171629c5db97febb12f04", size = 174155, upload-time = "2025-03-05T20:03:25.321Z" }, + { url = "https://files.pythonhosted.org/packages/68/a1/dcb68430b1d00b698ae7a7e0194433bce4f07ded185f0ee5fb21e2a2e91e/websockets-15.0.1-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:cad21560da69f4ce7658ca2cb83138fb4cf695a2ba3e475e0559e05991aa8122", size = 176884, upload-time = "2025-03-05T20:03:27.934Z" }, + { url = "https://files.pythonhosted.org/packages/fa/a8/5b41e0da817d64113292ab1f8247140aac61cbf6cfd085d6a0fa77f4984f/websockets-15.0.1-py3-none-any.whl", hash = "sha256:f7a866fbc1e97b5c617ee4116daaa09b722101d4a3c170c787450ba409f9736f", size = 169743, upload-time = "2025-03-05T20:03:39.41Z" }, +] + +[[package]] +name = "werkzeug" +version = "3.1.3" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "markupsafe" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/9f/69/83029f1f6300c5fb2471d621ab06f6ec6b3324685a2ce0f9777fd4a8b71e/werkzeug-3.1.3.tar.gz", hash = "sha256:60723ce945c19328679790e3282cc758aa4a6040e4bb330f53d30fa546d44746", size = 806925, upload-time = "2024-11-08T15:52:18.093Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/52/24/ab44c871b0f07f491e5d2ad12c9bd7358e527510618cb1b803a88e986db1/werkzeug-3.1.3-py3-none-any.whl", hash = "sha256:54b78bf3716d19a65be4fceccc0d1d7b89e608834989dfae50ea87564639213e", size = 224498, upload-time = "2024-11-08T15:52:16.132Z" }, +] + +[[package]] +name = "wheel" +version = "0.45.1" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/8a/98/2d9906746cdc6a6ef809ae6338005b3f21bb568bea3165cfc6a243fdc25c/wheel-0.45.1.tar.gz", hash = "sha256:661e1abd9198507b1409a20c02106d9670b2576e916d58f520316666abca6729", size = 107545, upload-time = "2024-11-23T00:18:23.513Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/0b/2c/87f3254fd8ffd29e4c02732eee68a83a1d3c346ae39bc6822dcbcb697f2b/wheel-0.45.1-py3-none-any.whl", hash = "sha256:708e7481cc80179af0e556bbf0cc00b8444c7321e2700b8d8580231d13017248", size = 72494, upload-time = "2024-11-23T00:18:21.207Z" }, +] + +[[package]] +name = "win32-setctime" +version = "1.2.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/b3/8f/705086c9d734d3b663af0e9bb3d4de6578d08f46b1b101c2442fd9aecaa2/win32_setctime-1.2.0.tar.gz", hash = "sha256:ae1fdf948f5640aae05c511ade119313fb6a30d7eabe25fef9764dca5873c4c0", size = 4867, upload-time = "2024-12-07T15:28:28.314Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/e1/07/c6fe3ad3e685340704d314d765b7912993bcb8dc198f0e7a89382d37974b/win32_setctime-1.2.0-py3-none-any.whl", hash = "sha256:95d644c4e708aba81dc3704a116d8cbc974d70b3bdb8be1d150e36be6e9d1390", size = 4083, upload-time = "2024-12-07T15:28:26.465Z" }, +] + +[[package]] +name = "xxhash" +version = "3.6.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/02/84/30869e01909fb37a6cc7e18688ee8bf1e42d57e7e0777636bd47524c43c7/xxhash-3.6.0.tar.gz", hash = "sha256:f0162a78b13a0d7617b2845b90c763339d1f1d82bb04a4b07f4ab535cc5e05d6", size = 85160, upload-time = "2025-10-02T14:37:08.097Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/34/ee/f9f1d656ad168681bb0f6b092372c1e533c4416b8069b1896a175c46e484/xxhash-3.6.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:87ff03d7e35c61435976554477a7f4cd1704c3596a89a8300d5ce7fc83874a71", size = 32845, upload-time = "2025-10-02T14:33:51.573Z" }, + { url = "https://files.pythonhosted.org/packages/a3/b1/93508d9460b292c74a09b83d16750c52a0ead89c51eea9951cb97a60d959/xxhash-3.6.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:f572dfd3d0e2eb1a57511831cf6341242f5a9f8298a45862d085f5b93394a27d", size = 30807, upload-time = "2025-10-02T14:33:52.964Z" }, + { url = "https://files.pythonhosted.org/packages/07/55/28c93a3662f2d200c70704efe74aab9640e824f8ce330d8d3943bf7c9b3c/xxhash-3.6.0-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:89952ea539566b9fed2bbd94e589672794b4286f342254fad28b149f9615fef8", size = 193786, upload-time = "2025-10-02T14:33:54.272Z" }, + { url = "https://files.pythonhosted.org/packages/c1/96/fec0be9bb4b8f5d9c57d76380a366f31a1781fb802f76fc7cda6c84893c7/xxhash-3.6.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:48e6f2ffb07a50b52465a1032c3cf1f4a5683f944acaca8a134a2f23674c2058", size = 212830, upload-time = "2025-10-02T14:33:55.706Z" }, + { url = "https://files.pythonhosted.org/packages/c4/a0/c706845ba77b9611f81fd2e93fad9859346b026e8445e76f8c6fd057cc6d/xxhash-3.6.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b5b848ad6c16d308c3ac7ad4ba6bede80ed5df2ba8ed382f8932df63158dd4b2", size = 211606, upload-time = "2025-10-02T14:33:57.133Z" }, + { url = "https://files.pythonhosted.org/packages/67/1e/164126a2999e5045f04a69257eea946c0dc3e86541b400d4385d646b53d7/xxhash-3.6.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a034590a727b44dd8ac5914236a7b8504144447a9682586c3327e935f33ec8cc", size = 444872, upload-time = "2025-10-02T14:33:58.446Z" }, + { url = "https://files.pythonhosted.org/packages/2d/4b/55ab404c56cd70a2cf5ecfe484838865d0fea5627365c6c8ca156bd09c8f/xxhash-3.6.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8a8f1972e75ebdd161d7896743122834fe87378160c20e97f8b09166213bf8cc", size = 193217, upload-time = "2025-10-02T14:33:59.724Z" }, + { url = "https://files.pythonhosted.org/packages/45/e6/52abf06bac316db33aa269091ae7311bd53cfc6f4b120ae77bac1b348091/xxhash-3.6.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:ee34327b187f002a596d7b167ebc59a1b729e963ce645964bbc050d2f1b73d07", size = 210139, upload-time = "2025-10-02T14:34:02.041Z" }, + { url = "https://files.pythonhosted.org/packages/34/37/db94d490b8691236d356bc249c08819cbcef9273a1a30acf1254ff9ce157/xxhash-3.6.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:339f518c3c7a850dd033ab416ea25a692759dc7478a71131fe8869010d2b75e4", size = 197669, upload-time = "2025-10-02T14:34:03.664Z" }, + { url = "https://files.pythonhosted.org/packages/b7/36/c4f219ef4a17a4f7a64ed3569bc2b5a9c8311abdb22249ac96093625b1a4/xxhash-3.6.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:bf48889c9630542d4709192578aebbd836177c9f7a4a2778a7d6340107c65f06", size = 210018, upload-time = "2025-10-02T14:34:05.325Z" }, + { url = "https://files.pythonhosted.org/packages/fd/06/bfac889a374fc2fc439a69223d1750eed2e18a7db8514737ab630534fa08/xxhash-3.6.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:5576b002a56207f640636056b4160a378fe36a58db73ae5c27a7ec8db35f71d4", size = 413058, upload-time = "2025-10-02T14:34:06.925Z" }, + { url = "https://files.pythonhosted.org/packages/c9/d1/555d8447e0dd32ad0930a249a522bb2e289f0d08b6b16204cfa42c1f5a0c/xxhash-3.6.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:af1f3278bd02814d6dedc5dec397993b549d6f16c19379721e5a1d31e132c49b", size = 190628, upload-time = "2025-10-02T14:34:08.669Z" }, + { url = "https://files.pythonhosted.org/packages/d1/15/8751330b5186cedc4ed4b597989882ea05e0408b53fa47bcb46a6125bfc6/xxhash-3.6.0-cp310-cp310-win32.whl", hash = "sha256:aed058764db109dc9052720da65fafe84873b05eb8b07e5e653597951af57c3b", size = 30577, upload-time = "2025-10-02T14:34:10.234Z" }, + { url = "https://files.pythonhosted.org/packages/bb/cc/53f87e8b5871a6eb2ff7e89c48c66093bda2be52315a8161ddc54ea550c4/xxhash-3.6.0-cp310-cp310-win_amd64.whl", hash = "sha256:e82da5670f2d0d98950317f82a0e4a0197150ff19a6df2ba40399c2a3b9ae5fb", size = 31487, upload-time = "2025-10-02T14:34:11.618Z" }, + { url = "https://files.pythonhosted.org/packages/9f/00/60f9ea3bb697667a14314d7269956f58bf56bb73864f8f8d52a3c2535e9a/xxhash-3.6.0-cp310-cp310-win_arm64.whl", hash = "sha256:4a082ffff8c6ac07707fb6b671caf7c6e020c75226c561830b73d862060f281d", size = 27863, upload-time = "2025-10-02T14:34:12.619Z" }, + { url = "https://files.pythonhosted.org/packages/17/d4/cc2f0400e9154df4b9964249da78ebd72f318e35ccc425e9f403c392f22a/xxhash-3.6.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:b47bbd8cf2d72797f3c2772eaaac0ded3d3af26481a26d7d7d41dc2d3c46b04a", size = 32844, upload-time = "2025-10-02T14:34:14.037Z" }, + { url = "https://files.pythonhosted.org/packages/5e/ec/1cc11cd13e26ea8bc3cb4af4eaadd8d46d5014aebb67be3f71fb0b68802a/xxhash-3.6.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:2b6821e94346f96db75abaa6e255706fb06ebd530899ed76d32cd99f20dc52fa", size = 30809, upload-time = "2025-10-02T14:34:15.484Z" }, + { url = "https://files.pythonhosted.org/packages/04/5f/19fe357ea348d98ca22f456f75a30ac0916b51c753e1f8b2e0e6fb884cce/xxhash-3.6.0-cp311-cp311-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:d0a9751f71a1a65ce3584e9cae4467651c7e70c9d31017fa57574583a4540248", size = 194665, upload-time = "2025-10-02T14:34:16.541Z" }, + { url = "https://files.pythonhosted.org/packages/90/3b/d1f1a8f5442a5fd8beedae110c5af7604dc37349a8e16519c13c19a9a2de/xxhash-3.6.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8b29ee68625ab37b04c0b40c3fafdf24d2f75ccd778333cfb698f65f6c463f62", size = 213550, upload-time = "2025-10-02T14:34:17.878Z" }, + { url = "https://files.pythonhosted.org/packages/c4/ef/3a9b05eb527457d5db13a135a2ae1a26c80fecd624d20f3e8dcc4cb170f3/xxhash-3.6.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:6812c25fe0d6c36a46ccb002f40f27ac903bf18af9f6dd8f9669cb4d176ab18f", size = 212384, upload-time = "2025-10-02T14:34:19.182Z" }, + { url = "https://files.pythonhosted.org/packages/0f/18/ccc194ee698c6c623acbf0f8c2969811a8a4b6185af5e824cd27b9e4fd3e/xxhash-3.6.0-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4ccbff013972390b51a18ef1255ef5ac125c92dc9143b2d1909f59abc765540e", size = 445749, upload-time = "2025-10-02T14:34:20.659Z" }, + { url = "https://files.pythonhosted.org/packages/a5/86/cf2c0321dc3940a7aa73076f4fd677a0fb3e405cb297ead7d864fd90847e/xxhash-3.6.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:297b7fbf86c82c550e12e8fb71968b3f033d27b874276ba3624ea868c11165a8", size = 193880, upload-time = "2025-10-02T14:34:22.431Z" }, + { url = "https://files.pythonhosted.org/packages/82/fb/96213c8560e6f948a1ecc9a7613f8032b19ee45f747f4fca4eb31bb6d6ed/xxhash-3.6.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:dea26ae1eb293db089798d3973a5fc928a18fdd97cc8801226fae705b02b14b0", size = 210912, upload-time = "2025-10-02T14:34:23.937Z" }, + { url = "https://files.pythonhosted.org/packages/40/aa/4395e669b0606a096d6788f40dbdf2b819d6773aa290c19e6e83cbfc312f/xxhash-3.6.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:7a0b169aafb98f4284f73635a8e93f0735f9cbde17bd5ec332480484241aaa77", size = 198654, upload-time = "2025-10-02T14:34:25.644Z" }, + { url = "https://files.pythonhosted.org/packages/67/74/b044fcd6b3d89e9b1b665924d85d3f400636c23590226feb1eb09e1176ce/xxhash-3.6.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:08d45aef063a4531b785cd72de4887766d01dc8f362a515693df349fdb825e0c", size = 210867, upload-time = "2025-10-02T14:34:27.203Z" }, + { url = "https://files.pythonhosted.org/packages/bc/fd/3ce73bf753b08cb19daee1eb14aa0d7fe331f8da9c02dd95316ddfe5275e/xxhash-3.6.0-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:929142361a48ee07f09121fe9e96a84950e8d4df3bb298ca5d88061969f34d7b", size = 414012, upload-time = "2025-10-02T14:34:28.409Z" }, + { url = "https://files.pythonhosted.org/packages/ba/b3/5a4241309217c5c876f156b10778f3ab3af7ba7e3259e6d5f5c7d0129eb2/xxhash-3.6.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:51312c768403d8540487dbbfb557454cfc55589bbde6424456951f7fcd4facb3", size = 191409, upload-time = "2025-10-02T14:34:29.696Z" }, + { url = "https://files.pythonhosted.org/packages/c0/01/99bfbc15fb9abb9a72b088c1d95219fc4782b7d01fc835bd5744d66dd0b8/xxhash-3.6.0-cp311-cp311-win32.whl", hash = "sha256:d1927a69feddc24c987b337ce81ac15c4720955b667fe9b588e02254b80446fd", size = 30574, upload-time = "2025-10-02T14:34:31.028Z" }, + { url = "https://files.pythonhosted.org/packages/65/79/9d24d7f53819fe301b231044ea362ce64e86c74f6e8c8e51320de248b3e5/xxhash-3.6.0-cp311-cp311-win_amd64.whl", hash = "sha256:26734cdc2d4ffe449b41d186bbeac416f704a482ed835d375a5c0cb02bc63fef", size = 31481, upload-time = "2025-10-02T14:34:32.062Z" }, + { url = "https://files.pythonhosted.org/packages/30/4e/15cd0e3e8772071344eab2961ce83f6e485111fed8beb491a3f1ce100270/xxhash-3.6.0-cp311-cp311-win_arm64.whl", hash = "sha256:d72f67ef8bf36e05f5b6c65e8524f265bd61071471cd4cf1d36743ebeeeb06b7", size = 27861, upload-time = "2025-10-02T14:34:33.555Z" }, + { url = "https://files.pythonhosted.org/packages/9a/07/d9412f3d7d462347e4511181dea65e47e0d0e16e26fbee2ea86a2aefb657/xxhash-3.6.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:01362c4331775398e7bb34e3ab403bc9ee9f7c497bc7dee6272114055277dd3c", size = 32744, upload-time = "2025-10-02T14:34:34.622Z" }, + { url = "https://files.pythonhosted.org/packages/79/35/0429ee11d035fc33abe32dca1b2b69e8c18d236547b9a9b72c1929189b9a/xxhash-3.6.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:b7b2df81a23f8cb99656378e72501b2cb41b1827c0f5a86f87d6b06b69f9f204", size = 30816, upload-time = "2025-10-02T14:34:36.043Z" }, + { url = "https://files.pythonhosted.org/packages/b7/f2/57eb99aa0f7d98624c0932c5b9a170e1806406cdbcdb510546634a1359e0/xxhash-3.6.0-cp312-cp312-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:dc94790144e66b14f67b10ac8ed75b39ca47536bf8800eb7c24b50271ea0c490", size = 194035, upload-time = "2025-10-02T14:34:37.354Z" }, + { url = "https://files.pythonhosted.org/packages/4c/ed/6224ba353690d73af7a3f1c7cdb1fc1b002e38f783cb991ae338e1eb3d79/xxhash-3.6.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:93f107c673bccf0d592cdba077dedaf52fe7f42dcd7676eba1f6d6f0c3efffd2", size = 212914, upload-time = "2025-10-02T14:34:38.6Z" }, + { url = "https://files.pythonhosted.org/packages/38/86/fb6b6130d8dd6b8942cc17ab4d90e223653a89aa32ad2776f8af7064ed13/xxhash-3.6.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:2aa5ee3444c25b69813663c9f8067dcfaa2e126dc55e8dddf40f4d1c25d7effa", size = 212163, upload-time = "2025-10-02T14:34:39.872Z" }, + { url = "https://files.pythonhosted.org/packages/ee/dc/e84875682b0593e884ad73b2d40767b5790d417bde603cceb6878901d647/xxhash-3.6.0-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:f7f99123f0e1194fa59cc69ad46dbae2e07becec5df50a0509a808f90a0f03f0", size = 445411, upload-time = "2025-10-02T14:34:41.569Z" }, + { url = "https://files.pythonhosted.org/packages/11/4f/426f91b96701ec2f37bb2b8cec664eff4f658a11f3fa9d94f0a887ea6d2b/xxhash-3.6.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:49e03e6fe2cac4a1bc64952dd250cf0dbc5ef4ebb7b8d96bce82e2de163c82a2", size = 193883, upload-time = "2025-10-02T14:34:43.249Z" }, + { url = "https://files.pythonhosted.org/packages/53/5a/ddbb83eee8e28b778eacfc5a85c969673e4023cdeedcfcef61f36731610b/xxhash-3.6.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:bd17fede52a17a4f9a7bc4472a5867cb0b160deeb431795c0e4abe158bc784e9", size = 210392, upload-time = "2025-10-02T14:34:45.042Z" }, + { url = "https://files.pythonhosted.org/packages/1e/c2/ff69efd07c8c074ccdf0a4f36fcdd3d27363665bcdf4ba399abebe643465/xxhash-3.6.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:6fb5f5476bef678f69db04f2bd1efbed3030d2aba305b0fc1773645f187d6a4e", size = 197898, upload-time = "2025-10-02T14:34:46.302Z" }, + { url = "https://files.pythonhosted.org/packages/58/ca/faa05ac19b3b622c7c9317ac3e23954187516298a091eb02c976d0d3dd45/xxhash-3.6.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:843b52f6d88071f87eba1631b684fcb4b2068cd2180a0224122fe4ef011a9374", size = 210655, upload-time = "2025-10-02T14:34:47.571Z" }, + { url = "https://files.pythonhosted.org/packages/d4/7a/06aa7482345480cc0cb597f5c875b11a82c3953f534394f620b0be2f700c/xxhash-3.6.0-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:7d14a6cfaf03b1b6f5f9790f76880601ccc7896aff7ab9cd8978a939c1eb7e0d", size = 414001, upload-time = "2025-10-02T14:34:49.273Z" }, + { url = "https://files.pythonhosted.org/packages/23/07/63ffb386cd47029aa2916b3d2f454e6cc5b9f5c5ada3790377d5430084e7/xxhash-3.6.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:418daf3db71e1413cfe211c2f9a528456936645c17f46b5204705581a45390ae", size = 191431, upload-time = "2025-10-02T14:34:50.798Z" }, + { url = "https://files.pythonhosted.org/packages/0f/93/14fde614cadb4ddf5e7cebf8918b7e8fac5ae7861c1875964f17e678205c/xxhash-3.6.0-cp312-cp312-win32.whl", hash = "sha256:50fc255f39428a27299c20e280d6193d8b63b8ef8028995323bf834a026b4fbb", size = 30617, upload-time = "2025-10-02T14:34:51.954Z" }, + { url = "https://files.pythonhosted.org/packages/13/5d/0d125536cbe7565a83d06e43783389ecae0c0f2ed037b48ede185de477c0/xxhash-3.6.0-cp312-cp312-win_amd64.whl", hash = "sha256:c0f2ab8c715630565ab8991b536ecded9416d615538be8ecddce43ccf26cbc7c", size = 31534, upload-time = "2025-10-02T14:34:53.276Z" }, + { url = "https://files.pythonhosted.org/packages/54/85/6ec269b0952ec7e36ba019125982cf11d91256a778c7c3f98a4c5043d283/xxhash-3.6.0-cp312-cp312-win_arm64.whl", hash = "sha256:eae5c13f3bc455a3bbb68bdc513912dc7356de7e2280363ea235f71f54064829", size = 27876, upload-time = "2025-10-02T14:34:54.371Z" }, + { url = "https://files.pythonhosted.org/packages/33/76/35d05267ac82f53ae9b0e554da7c5e281ee61f3cad44c743f0fcd354f211/xxhash-3.6.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:599e64ba7f67472481ceb6ee80fa3bd828fd61ba59fb11475572cc5ee52b89ec", size = 32738, upload-time = "2025-10-02T14:34:55.839Z" }, + { url = "https://files.pythonhosted.org/packages/31/a8/3fbce1cd96534a95e35d5120637bf29b0d7f5d8fa2f6374e31b4156dd419/xxhash-3.6.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7d8b8aaa30fca4f16f0c84a5c8d7ddee0e25250ec2796c973775373257dde8f1", size = 30821, upload-time = "2025-10-02T14:34:57.219Z" }, + { url = "https://files.pythonhosted.org/packages/0c/ea/d387530ca7ecfa183cb358027f1833297c6ac6098223fd14f9782cd0015c/xxhash-3.6.0-cp313-cp313-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:d597acf8506d6e7101a4a44a5e428977a51c0fadbbfd3c39650cca9253f6e5a6", size = 194127, upload-time = "2025-10-02T14:34:59.21Z" }, + { url = "https://files.pythonhosted.org/packages/ba/0c/71435dcb99874b09a43b8d7c54071e600a7481e42b3e3ce1eb5226a5711a/xxhash-3.6.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:858dc935963a33bc33490128edc1c12b0c14d9c7ebaa4e387a7869ecc4f3e263", size = 212975, upload-time = "2025-10-02T14:35:00.816Z" }, + { url = "https://files.pythonhosted.org/packages/84/7a/c2b3d071e4bb4a90b7057228a99b10d51744878f4a8a6dd643c8bd897620/xxhash-3.6.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ba284920194615cb8edf73bf52236ce2e1664ccd4a38fdb543506413529cc546", size = 212241, upload-time = "2025-10-02T14:35:02.207Z" }, + { url = "https://files.pythonhosted.org/packages/81/5f/640b6eac0128e215f177df99eadcd0f1b7c42c274ab6a394a05059694c5a/xxhash-3.6.0-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4b54219177f6c6674d5378bd862c6aedf64725f70dd29c472eaae154df1a2e89", size = 445471, upload-time = "2025-10-02T14:35:03.61Z" }, + { url = "https://files.pythonhosted.org/packages/5e/1e/3c3d3ef071b051cc3abbe3721ffb8365033a172613c04af2da89d5548a87/xxhash-3.6.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:42c36dd7dbad2f5238950c377fcbf6811b1cdb1c444fab447960030cea60504d", size = 193936, upload-time = "2025-10-02T14:35:05.013Z" }, + { url = "https://files.pythonhosted.org/packages/2c/bd/4a5f68381939219abfe1c22a9e3a5854a4f6f6f3c4983a87d255f21f2e5d/xxhash-3.6.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:f22927652cba98c44639ffdc7aaf35828dccf679b10b31c4ad72a5b530a18eb7", size = 210440, upload-time = "2025-10-02T14:35:06.239Z" }, + { url = "https://files.pythonhosted.org/packages/eb/37/b80fe3d5cfb9faff01a02121a0f4d565eb7237e9e5fc66e73017e74dcd36/xxhash-3.6.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:b45fad44d9c5c119e9c6fbf2e1c656a46dc68e280275007bbfd3d572b21426db", size = 197990, upload-time = "2025-10-02T14:35:07.735Z" }, + { url = "https://files.pythonhosted.org/packages/d7/fd/2c0a00c97b9e18f72e1f240ad4e8f8a90fd9d408289ba9c7c495ed7dc05c/xxhash-3.6.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:6f2580ffab1a8b68ef2b901cde7e55fa8da5e4be0977c68f78fc80f3c143de42", size = 210689, upload-time = "2025-10-02T14:35:09.438Z" }, + { url = "https://files.pythonhosted.org/packages/93/86/5dd8076a926b9a95db3206aba20d89a7fc14dd5aac16e5c4de4b56033140/xxhash-3.6.0-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:40c391dd3cd041ebc3ffe6f2c862f402e306eb571422e0aa918d8070ba31da11", size = 414068, upload-time = "2025-10-02T14:35:11.162Z" }, + { url = "https://files.pythonhosted.org/packages/af/3c/0bb129170ee8f3650f08e993baee550a09593462a5cddd8e44d0011102b1/xxhash-3.6.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:f205badabde7aafd1a31e8ca2a3e5a763107a71c397c4481d6a804eb5063d8bd", size = 191495, upload-time = "2025-10-02T14:35:12.971Z" }, + { url = "https://files.pythonhosted.org/packages/e9/3a/6797e0114c21d1725e2577508e24006fd7ff1d8c0c502d3b52e45c1771d8/xxhash-3.6.0-cp313-cp313-win32.whl", hash = "sha256:2577b276e060b73b73a53042ea5bd5203d3e6347ce0d09f98500f418a9fcf799", size = 30620, upload-time = "2025-10-02T14:35:14.129Z" }, + { url = "https://files.pythonhosted.org/packages/86/15/9bc32671e9a38b413a76d24722a2bf8784a132c043063a8f5152d390b0f9/xxhash-3.6.0-cp313-cp313-win_amd64.whl", hash = "sha256:757320d45d2fbcce8f30c42a6b2f47862967aea7bf458b9625b4bbe7ee390392", size = 31542, upload-time = "2025-10-02T14:35:15.21Z" }, + { url = "https://files.pythonhosted.org/packages/39/c5/cc01e4f6188656e56112d6a8e0dfe298a16934b8c47a247236549a3f7695/xxhash-3.6.0-cp313-cp313-win_arm64.whl", hash = "sha256:457b8f85dec5825eed7b69c11ae86834a018b8e3df5e77783c999663da2f96d6", size = 27880, upload-time = "2025-10-02T14:35:16.315Z" }, + { url = "https://files.pythonhosted.org/packages/f3/30/25e5321c8732759e930c555176d37e24ab84365482d257c3b16362235212/xxhash-3.6.0-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:a42e633d75cdad6d625434e3468126c73f13f7584545a9cf34e883aa1710e702", size = 32956, upload-time = "2025-10-02T14:35:17.413Z" }, + { url = "https://files.pythonhosted.org/packages/9f/3c/0573299560d7d9f8ab1838f1efc021a280b5ae5ae2e849034ef3dee18810/xxhash-3.6.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:568a6d743219e717b07b4e03b0a828ce593833e498c3b64752e0f5df6bfe84db", size = 31072, upload-time = "2025-10-02T14:35:18.844Z" }, + { url = "https://files.pythonhosted.org/packages/7a/1c/52d83a06e417cd9d4137722693424885cc9878249beb3a7c829e74bf7ce9/xxhash-3.6.0-cp313-cp313t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:bec91b562d8012dae276af8025a55811b875baace6af510412a5e58e3121bc54", size = 196409, upload-time = "2025-10-02T14:35:20.31Z" }, + { url = "https://files.pythonhosted.org/packages/e3/8e/c6d158d12a79bbd0b878f8355432075fc82759e356ab5a111463422a239b/xxhash-3.6.0-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:78e7f2f4c521c30ad5e786fdd6bae89d47a32672a80195467b5de0480aa97b1f", size = 215736, upload-time = "2025-10-02T14:35:21.616Z" }, + { url = "https://files.pythonhosted.org/packages/bc/68/c4c80614716345d55071a396cf03d06e34b5f4917a467faf43083c995155/xxhash-3.6.0-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:3ed0df1b11a79856df5ffcab572cbd6b9627034c1c748c5566fa79df9048a7c5", size = 214833, upload-time = "2025-10-02T14:35:23.32Z" }, + { url = "https://files.pythonhosted.org/packages/7e/e9/ae27c8ffec8b953efa84c7c4a6c6802c263d587b9fc0d6e7cea64e08c3af/xxhash-3.6.0-cp313-cp313t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0e4edbfc7d420925b0dd5e792478ed393d6e75ff8fc219a6546fb446b6a417b1", size = 448348, upload-time = "2025-10-02T14:35:25.111Z" }, + { url = "https://files.pythonhosted.org/packages/d7/6b/33e21afb1b5b3f46b74b6bd1913639066af218d704cc0941404ca717fc57/xxhash-3.6.0-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fba27a198363a7ef87f8c0f6b171ec36b674fe9053742c58dd7e3201c1ab30ee", size = 196070, upload-time = "2025-10-02T14:35:26.586Z" }, + { url = "https://files.pythonhosted.org/packages/96/b6/fcabd337bc5fa624e7203aa0fa7d0c49eed22f72e93229431752bddc83d9/xxhash-3.6.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:794fe9145fe60191c6532fa95063765529770edcdd67b3d537793e8004cabbfd", size = 212907, upload-time = "2025-10-02T14:35:28.087Z" }, + { url = "https://files.pythonhosted.org/packages/4b/d3/9ee6160e644d660fcf176c5825e61411c7f62648728f69c79ba237250143/xxhash-3.6.0-cp313-cp313t-musllinux_1_2_i686.whl", hash = "sha256:6105ef7e62b5ac73a837778efc331a591d8442f8ef5c7e102376506cb4ae2729", size = 200839, upload-time = "2025-10-02T14:35:29.857Z" }, + { url = "https://files.pythonhosted.org/packages/0d/98/e8de5baa5109394baf5118f5e72ab21a86387c4f89b0e77ef3e2f6b0327b/xxhash-3.6.0-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:f01375c0e55395b814a679b3eea205db7919ac2af213f4a6682e01220e5fe292", size = 213304, upload-time = "2025-10-02T14:35:31.222Z" }, + { url = "https://files.pythonhosted.org/packages/7b/1d/71056535dec5c3177eeb53e38e3d367dd1d16e024e63b1cee208d572a033/xxhash-3.6.0-cp313-cp313t-musllinux_1_2_s390x.whl", hash = "sha256:d706dca2d24d834a4661619dcacf51a75c16d65985718d6a7d73c1eeeb903ddf", size = 416930, upload-time = "2025-10-02T14:35:32.517Z" }, + { url = "https://files.pythonhosted.org/packages/dc/6c/5cbde9de2cd967c322e651c65c543700b19e7ae3e0aae8ece3469bf9683d/xxhash-3.6.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:5f059d9faeacd49c0215d66f4056e1326c80503f51a1532ca336a385edadd033", size = 193787, upload-time = "2025-10-02T14:35:33.827Z" }, + { url = "https://files.pythonhosted.org/packages/19/fa/0172e350361d61febcea941b0cc541d6e6c8d65d153e85f850a7b256ff8a/xxhash-3.6.0-cp313-cp313t-win32.whl", hash = "sha256:1244460adc3a9be84731d72b8e80625788e5815b68da3da8b83f78115a40a7ec", size = 30916, upload-time = "2025-10-02T14:35:35.107Z" }, + { url = "https://files.pythonhosted.org/packages/ad/e6/e8cf858a2b19d6d45820f072eff1bea413910592ff17157cabc5f1227a16/xxhash-3.6.0-cp313-cp313t-win_amd64.whl", hash = "sha256:b1e420ef35c503869c4064f4a2f2b08ad6431ab7b229a05cce39d74268bca6b8", size = 31799, upload-time = "2025-10-02T14:35:36.165Z" }, + { url = "https://files.pythonhosted.org/packages/56/15/064b197e855bfb7b343210e82490ae672f8bc7cdf3ddb02e92f64304ee8a/xxhash-3.6.0-cp313-cp313t-win_arm64.whl", hash = "sha256:ec44b73a4220623235f67a996c862049f375df3b1052d9899f40a6382c32d746", size = 28044, upload-time = "2025-10-02T14:35:37.195Z" }, + { url = "https://files.pythonhosted.org/packages/7e/5e/0138bc4484ea9b897864d59fce9be9086030825bc778b76cb5a33a906d37/xxhash-3.6.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:a40a3d35b204b7cc7643cbcf8c9976d818cb47befcfac8bbefec8038ac363f3e", size = 32754, upload-time = "2025-10-02T14:35:38.245Z" }, + { url = "https://files.pythonhosted.org/packages/18/d7/5dac2eb2ec75fd771957a13e5dda560efb2176d5203f39502a5fc571f899/xxhash-3.6.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:a54844be970d3fc22630b32d515e79a90d0a3ddb2644d8d7402e3c4c8da61405", size = 30846, upload-time = "2025-10-02T14:35:39.6Z" }, + { url = "https://files.pythonhosted.org/packages/fe/71/8bc5be2bb00deb5682e92e8da955ebe5fa982da13a69da5a40a4c8db12fb/xxhash-3.6.0-cp314-cp314-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:016e9190af8f0a4e3741343777710e3d5717427f175adfdc3e72508f59e2a7f3", size = 194343, upload-time = "2025-10-02T14:35:40.69Z" }, + { url = "https://files.pythonhosted.org/packages/e7/3b/52badfb2aecec2c377ddf1ae75f55db3ba2d321c5e164f14461c90837ef3/xxhash-3.6.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4f6f72232f849eb9d0141e2ebe2677ece15adfd0fa599bc058aad83c714bb2c6", size = 213074, upload-time = "2025-10-02T14:35:42.29Z" }, + { url = "https://files.pythonhosted.org/packages/a2/2b/ae46b4e9b92e537fa30d03dbc19cdae57ed407e9c26d163895e968e3de85/xxhash-3.6.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:63275a8aba7865e44b1813d2177e0f5ea7eadad3dd063a21f7cf9afdc7054063", size = 212388, upload-time = "2025-10-02T14:35:43.929Z" }, + { url = "https://files.pythonhosted.org/packages/f5/80/49f88d3afc724b4ac7fbd664c8452d6db51b49915be48c6982659e0e7942/xxhash-3.6.0-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:3cd01fa2aa00d8b017c97eb46b9a794fbdca53fc14f845f5a328c71254b0abb7", size = 445614, upload-time = "2025-10-02T14:35:45.216Z" }, + { url = "https://files.pythonhosted.org/packages/ed/ba/603ce3961e339413543d8cd44f21f2c80e2a7c5cfe692a7b1f2cccf58f3c/xxhash-3.6.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0226aa89035b62b6a86d3c68df4d7c1f47a342b8683da2b60cedcddb46c4d95b", size = 194024, upload-time = "2025-10-02T14:35:46.959Z" }, + { url = "https://files.pythonhosted.org/packages/78/d1/8e225ff7113bf81545cfdcd79eef124a7b7064a0bba53605ff39590b95c2/xxhash-3.6.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:c6e193e9f56e4ca4923c61238cdaced324f0feac782544eb4c6d55ad5cc99ddd", size = 210541, upload-time = "2025-10-02T14:35:48.301Z" }, + { url = "https://files.pythonhosted.org/packages/6f/58/0f89d149f0bad89def1a8dd38feb50ccdeb643d9797ec84707091d4cb494/xxhash-3.6.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:9176dcaddf4ca963d4deb93866d739a343c01c969231dbe21680e13a5d1a5bf0", size = 198305, upload-time = "2025-10-02T14:35:49.584Z" }, + { url = "https://files.pythonhosted.org/packages/11/38/5eab81580703c4df93feb5f32ff8fa7fe1e2c51c1f183ee4e48d4bb9d3d7/xxhash-3.6.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:c1ce4009c97a752e682b897aa99aef84191077a9433eb237774689f14f8ec152", size = 210848, upload-time = "2025-10-02T14:35:50.877Z" }, + { url = "https://files.pythonhosted.org/packages/5e/6b/953dc4b05c3ce678abca756416e4c130d2382f877a9c30a20d08ee6a77c0/xxhash-3.6.0-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:8cb2f4f679b01513b7adbb9b1b2f0f9cdc31b70007eaf9d59d0878809f385b11", size = 414142, upload-time = "2025-10-02T14:35:52.15Z" }, + { url = "https://files.pythonhosted.org/packages/08/a9/238ec0d4e81a10eb5026d4a6972677cbc898ba6c8b9dbaec12ae001b1b35/xxhash-3.6.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:653a91d7c2ab54a92c19ccf43508b6a555440b9be1bc8be553376778be7f20b5", size = 191547, upload-time = "2025-10-02T14:35:53.547Z" }, + { url = "https://files.pythonhosted.org/packages/f1/ee/3cf8589e06c2164ac77c3bf0aa127012801128f1feebf2a079272da5737c/xxhash-3.6.0-cp314-cp314-win32.whl", hash = "sha256:a756fe893389483ee8c394d06b5ab765d96e68fbbfe6fde7aa17e11f5720559f", size = 31214, upload-time = "2025-10-02T14:35:54.746Z" }, + { url = "https://files.pythonhosted.org/packages/02/5d/a19552fbc6ad4cb54ff953c3908bbc095f4a921bc569433d791f755186f1/xxhash-3.6.0-cp314-cp314-win_amd64.whl", hash = "sha256:39be8e4e142550ef69629c9cd71b88c90e9a5db703fecbcf265546d9536ca4ad", size = 32290, upload-time = "2025-10-02T14:35:55.791Z" }, + { url = "https://files.pythonhosted.org/packages/b1/11/dafa0643bc30442c887b55baf8e73353a344ee89c1901b5a5c54a6c17d39/xxhash-3.6.0-cp314-cp314-win_arm64.whl", hash = "sha256:25915e6000338999236f1eb68a02a32c3275ac338628a7eaa5a269c401995679", size = 28795, upload-time = "2025-10-02T14:35:57.162Z" }, + { url = "https://files.pythonhosted.org/packages/2c/db/0e99732ed7f64182aef4a6fb145e1a295558deec2a746265dcdec12d191e/xxhash-3.6.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:c5294f596a9017ca5a3e3f8884c00b91ab2ad2933cf288f4923c3fd4346cf3d4", size = 32955, upload-time = "2025-10-02T14:35:58.267Z" }, + { url = "https://files.pythonhosted.org/packages/55/f4/2a7c3c68e564a099becfa44bb3d398810cc0ff6749b0d3cb8ccb93f23c14/xxhash-3.6.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:1cf9dcc4ab9cff01dfbba78544297a3a01dafd60f3bde4e2bfd016cf7e4ddc67", size = 31072, upload-time = "2025-10-02T14:35:59.382Z" }, + { url = "https://files.pythonhosted.org/packages/c6/d9/72a29cddc7250e8a5819dad5d466facb5dc4c802ce120645630149127e73/xxhash-3.6.0-cp314-cp314t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:01262da8798422d0685f7cef03b2bd3f4f46511b02830861df548d7def4402ad", size = 196579, upload-time = "2025-10-02T14:36:00.838Z" }, + { url = "https://files.pythonhosted.org/packages/63/93/b21590e1e381040e2ca305a884d89e1c345b347404f7780f07f2cdd47ef4/xxhash-3.6.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:51a73fb7cb3a3ead9f7a8b583ffd9b8038e277cdb8cb87cf890e88b3456afa0b", size = 215854, upload-time = "2025-10-02T14:36:02.207Z" }, + { url = "https://files.pythonhosted.org/packages/ce/b8/edab8a7d4fa14e924b29be877d54155dcbd8b80be85ea00d2be3413a9ed4/xxhash-3.6.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b9c6df83594f7df8f7f708ce5ebeacfc69f72c9fbaaababf6cf4758eaada0c9b", size = 214965, upload-time = "2025-10-02T14:36:03.507Z" }, + { url = "https://files.pythonhosted.org/packages/27/67/dfa980ac7f0d509d54ea0d5a486d2bb4b80c3f1bb22b66e6a05d3efaf6c0/xxhash-3.6.0-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:627f0af069b0ea56f312fd5189001c24578868643203bca1abbc2c52d3a6f3ca", size = 448484, upload-time = "2025-10-02T14:36:04.828Z" }, + { url = "https://files.pythonhosted.org/packages/8c/63/8ffc2cc97e811c0ca5d00ab36604b3ea6f4254f20b7bc658ca825ce6c954/xxhash-3.6.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:aa912c62f842dfd013c5f21a642c9c10cd9f4c4e943e0af83618b4a404d9091a", size = 196162, upload-time = "2025-10-02T14:36:06.182Z" }, + { url = "https://files.pythonhosted.org/packages/4b/77/07f0e7a3edd11a6097e990f6e5b815b6592459cb16dae990d967693e6ea9/xxhash-3.6.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:b465afd7909db30168ab62afe40b2fcf79eedc0b89a6c0ab3123515dc0df8b99", size = 213007, upload-time = "2025-10-02T14:36:07.733Z" }, + { url = "https://files.pythonhosted.org/packages/ae/d8/bc5fa0d152837117eb0bef6f83f956c509332ce133c91c63ce07ee7c4873/xxhash-3.6.0-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:a881851cf38b0a70e7c4d3ce81fc7afd86fbc2a024f4cfb2a97cf49ce04b75d3", size = 200956, upload-time = "2025-10-02T14:36:09.106Z" }, + { url = "https://files.pythonhosted.org/packages/26/a5/d749334130de9411783873e9b98ecc46688dad5db64ca6e04b02acc8b473/xxhash-3.6.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:9b3222c686a919a0f3253cfc12bb118b8b103506612253b5baeaac10d8027cf6", size = 213401, upload-time = "2025-10-02T14:36:10.585Z" }, + { url = "https://files.pythonhosted.org/packages/89/72/abed959c956a4bfc72b58c0384bb7940663c678127538634d896b1195c10/xxhash-3.6.0-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:c5aa639bc113e9286137cec8fadc20e9cd732b2cc385c0b7fa673b84fc1f2a93", size = 417083, upload-time = "2025-10-02T14:36:12.276Z" }, + { url = "https://files.pythonhosted.org/packages/0c/b3/62fd2b586283b7d7d665fb98e266decadf31f058f1cf6c478741f68af0cb/xxhash-3.6.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:5c1343d49ac102799905e115aee590183c3921d475356cb24b4de29a4bc56518", size = 193913, upload-time = "2025-10-02T14:36:14.025Z" }, + { url = "https://files.pythonhosted.org/packages/9a/9a/c19c42c5b3f5a4aad748a6d5b4f23df3bed7ee5445accc65a0fb3ff03953/xxhash-3.6.0-cp314-cp314t-win32.whl", hash = "sha256:5851f033c3030dd95c086b4a36a2683c2ff4a799b23af60977188b057e467119", size = 31586, upload-time = "2025-10-02T14:36:15.603Z" }, + { url = "https://files.pythonhosted.org/packages/03/d6/4cc450345be9924fd5dc8c590ceda1db5b43a0a889587b0ae81a95511360/xxhash-3.6.0-cp314-cp314t-win_amd64.whl", hash = "sha256:0444e7967dac37569052d2409b00a8860c2135cff05502df4da80267d384849f", size = 32526, upload-time = "2025-10-02T14:36:16.708Z" }, + { url = "https://files.pythonhosted.org/packages/0f/c9/7243eb3f9eaabd1a88a5a5acadf06df2d83b100c62684b7425c6a11bcaa8/xxhash-3.6.0-cp314-cp314t-win_arm64.whl", hash = "sha256:bb79b1e63f6fd84ec778a4b1916dfe0a7c3fdb986c06addd5db3a0d413819d95", size = 28898, upload-time = "2025-10-02T14:36:17.843Z" }, + { url = "https://files.pythonhosted.org/packages/93/1e/8aec23647a34a249f62e2398c42955acd9b4c6ed5cf08cbea94dc46f78d2/xxhash-3.6.0-pp311-pypy311_pp73-macosx_10_15_x86_64.whl", hash = "sha256:0f7b7e2ec26c1666ad5fc9dbfa426a6a3367ceaf79db5dd76264659d509d73b0", size = 30662, upload-time = "2025-10-02T14:37:01.743Z" }, + { url = "https://files.pythonhosted.org/packages/b8/0b/b14510b38ba91caf43006209db846a696ceea6a847a0c9ba0a5b1adc53d6/xxhash-3.6.0-pp311-pypy311_pp73-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:5dc1e14d14fa0f5789ec29a7062004b5933964bb9b02aae6622b8f530dc40296", size = 41056, upload-time = "2025-10-02T14:37:02.879Z" }, + { url = "https://files.pythonhosted.org/packages/50/55/15a7b8a56590e66ccd374bbfa3f9ffc45b810886c8c3b614e3f90bd2367c/xxhash-3.6.0-pp311-pypy311_pp73-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:881b47fc47e051b37d94d13e7455131054b56749b91b508b0907eb07900d1c13", size = 36251, upload-time = "2025-10-02T14:37:04.44Z" }, + { url = "https://files.pythonhosted.org/packages/62/b2/5ac99a041a29e58e95f907876b04f7067a0242cb85b5f39e726153981503/xxhash-3.6.0-pp311-pypy311_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c6dc31591899f5e5666f04cc2e529e69b4072827085c1ef15294d91a004bc1bd", size = 32481, upload-time = "2025-10-02T14:37:05.869Z" }, + { url = "https://files.pythonhosted.org/packages/7b/d9/8d95e906764a386a3d3b596f3c68bb63687dfca806373509f51ce8eea81f/xxhash-3.6.0-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:15e0dac10eb9309508bfc41f7f9deaa7755c69e35af835db9cb10751adebc35d", size = 31565, upload-time = "2025-10-02T14:37:06.966Z" }, +] + +[[package]] +name = "yarl" +version = "1.22.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "idna" }, + { name = "multidict" }, + { name = "propcache" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/57/63/0c6ebca57330cd313f6102b16dd57ffaf3ec4c83403dcb45dbd15c6f3ea1/yarl-1.22.0.tar.gz", hash = "sha256:bebf8557577d4401ba8bd9ff33906f1376c877aa78d1fe216ad01b4d6745af71", size = 187169, upload-time = "2025-10-06T14:12:55.963Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/d1/43/a2204825342f37c337f5edb6637040fa14e365b2fcc2346960201d457579/yarl-1.22.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:c7bd6683587567e5a49ee6e336e0612bec8329be1b7d4c8af5687dcdeb67ee1e", size = 140517, upload-time = "2025-10-06T14:08:42.494Z" }, + { url = "https://files.pythonhosted.org/packages/44/6f/674f3e6f02266428c56f704cd2501c22f78e8b2eeb23f153117cc86fb28a/yarl-1.22.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:5cdac20da754f3a723cceea5b3448e1a2074866406adeb4ef35b469d089adb8f", size = 93495, upload-time = "2025-10-06T14:08:46.2Z" }, + { url = "https://files.pythonhosted.org/packages/b8/12/5b274d8a0f30c07b91b2f02cba69152600b47830fcfb465c108880fcee9c/yarl-1.22.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:07a524d84df0c10f41e3ee918846e1974aba4ec017f990dc735aad487a0bdfdf", size = 94400, upload-time = "2025-10-06T14:08:47.855Z" }, + { url = "https://files.pythonhosted.org/packages/e2/7f/df1b6949b1fa1aa9ff6de6e2631876ad4b73c4437822026e85d8acb56bb1/yarl-1.22.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e1b329cb8146d7b736677a2440e422eadd775d1806a81db2d4cded80a48efc1a", size = 347545, upload-time = "2025-10-06T14:08:49.683Z" }, + { url = "https://files.pythonhosted.org/packages/84/09/f92ed93bd6cd77872ab6c3462df45ca45cd058d8f1d0c9b4f54c1704429f/yarl-1.22.0-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:75976c6945d85dbb9ee6308cd7ff7b1fb9409380c82d6119bd778d8fcfe2931c", size = 319598, upload-time = "2025-10-06T14:08:51.215Z" }, + { url = "https://files.pythonhosted.org/packages/c3/97/ac3f3feae7d522cf7ccec3d340bb0b2b61c56cb9767923df62a135092c6b/yarl-1.22.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:80ddf7a5f8c86cb3eb4bc9028b07bbbf1f08a96c5c0bc1244be5e8fefcb94147", size = 363893, upload-time = "2025-10-06T14:08:53.144Z" }, + { url = "https://files.pythonhosted.org/packages/06/49/f3219097403b9c84a4d079b1d7bda62dd9b86d0d6e4428c02d46ab2c77fc/yarl-1.22.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d332fc2e3c94dad927f2112395772a4e4fedbcf8f80efc21ed7cdfae4d574fdb", size = 371240, upload-time = "2025-10-06T14:08:55.036Z" }, + { url = "https://files.pythonhosted.org/packages/35/9f/06b765d45c0e44e8ecf0fe15c9eacbbde342bb5b7561c46944f107bfb6c3/yarl-1.22.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0cf71bf877efeac18b38d3930594c0948c82b64547c1cf420ba48722fe5509f6", size = 346965, upload-time = "2025-10-06T14:08:56.722Z" }, + { url = "https://files.pythonhosted.org/packages/c5/69/599e7cea8d0fcb1694323b0db0dda317fa3162f7b90166faddecf532166f/yarl-1.22.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:663e1cadaddae26be034a6ab6072449a8426ddb03d500f43daf952b74553bba0", size = 342026, upload-time = "2025-10-06T14:08:58.563Z" }, + { url = "https://files.pythonhosted.org/packages/95/6f/9dfd12c8bc90fea9eab39832ee32ea48f8e53d1256252a77b710c065c89f/yarl-1.22.0-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:6dcbb0829c671f305be48a7227918cfcd11276c2d637a8033a99a02b67bf9eda", size = 335637, upload-time = "2025-10-06T14:09:00.506Z" }, + { url = "https://files.pythonhosted.org/packages/57/2e/34c5b4eb9b07e16e873db5b182c71e5f06f9b5af388cdaa97736d79dd9a6/yarl-1.22.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:f0d97c18dfd9a9af4490631905a3f131a8e4c9e80a39353919e2cfed8f00aedc", size = 359082, upload-time = "2025-10-06T14:09:01.936Z" }, + { url = "https://files.pythonhosted.org/packages/31/71/fa7e10fb772d273aa1f096ecb8ab8594117822f683bab7d2c5a89914c92a/yarl-1.22.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:437840083abe022c978470b942ff832c3940b2ad3734d424b7eaffcd07f76737", size = 357811, upload-time = "2025-10-06T14:09:03.445Z" }, + { url = "https://files.pythonhosted.org/packages/26/da/11374c04e8e1184a6a03cf9c8f5688d3e5cec83ed6f31ad3481b3207f709/yarl-1.22.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:a899cbd98dce6f5d8de1aad31cb712ec0a530abc0a86bd6edaa47c1090138467", size = 351223, upload-time = "2025-10-06T14:09:05.401Z" }, + { url = "https://files.pythonhosted.org/packages/82/8f/e2d01f161b0c034a30410e375e191a5d27608c1f8693bab1a08b089ca096/yarl-1.22.0-cp310-cp310-win32.whl", hash = "sha256:595697f68bd1f0c1c159fcb97b661fc9c3f5db46498043555d04805430e79bea", size = 82118, upload-time = "2025-10-06T14:09:11.148Z" }, + { url = "https://files.pythonhosted.org/packages/62/46/94c76196642dbeae634c7a61ba3da88cd77bed875bf6e4a8bed037505aa6/yarl-1.22.0-cp310-cp310-win_amd64.whl", hash = "sha256:cb95a9b1adaa48e41815a55ae740cfda005758104049a640a398120bf02515ca", size = 86852, upload-time = "2025-10-06T14:09:12.958Z" }, + { url = "https://files.pythonhosted.org/packages/af/af/7df4f179d3b1a6dcb9a4bd2ffbc67642746fcafdb62580e66876ce83fff4/yarl-1.22.0-cp310-cp310-win_arm64.whl", hash = "sha256:b85b982afde6df99ecc996990d4ad7ccbdbb70e2a4ba4de0aecde5922ba98a0b", size = 82012, upload-time = "2025-10-06T14:09:14.664Z" }, + { url = "https://files.pythonhosted.org/packages/4d/27/5ab13fc84c76a0250afd3d26d5936349a35be56ce5785447d6c423b26d92/yarl-1.22.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:1ab72135b1f2db3fed3997d7e7dc1b80573c67138023852b6efb336a5eae6511", size = 141607, upload-time = "2025-10-06T14:09:16.298Z" }, + { url = "https://files.pythonhosted.org/packages/6a/a1/d065d51d02dc02ce81501d476b9ed2229d9a990818332242a882d5d60340/yarl-1.22.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:669930400e375570189492dc8d8341301578e8493aec04aebc20d4717f899dd6", size = 94027, upload-time = "2025-10-06T14:09:17.786Z" }, + { url = "https://files.pythonhosted.org/packages/c1/da/8da9f6a53f67b5106ffe902c6fa0164e10398d4e150d85838b82f424072a/yarl-1.22.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:792a2af6d58177ef7c19cbf0097aba92ca1b9cb3ffdd9c7470e156c8f9b5e028", size = 94963, upload-time = "2025-10-06T14:09:19.662Z" }, + { url = "https://files.pythonhosted.org/packages/68/fe/2c1f674960c376e29cb0bec1249b117d11738db92a6ccc4a530b972648db/yarl-1.22.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3ea66b1c11c9150f1372f69afb6b8116f2dd7286f38e14ea71a44eee9ec51b9d", size = 368406, upload-time = "2025-10-06T14:09:21.402Z" }, + { url = "https://files.pythonhosted.org/packages/95/26/812a540e1c3c6418fec60e9bbd38e871eaba9545e94fa5eff8f4a8e28e1e/yarl-1.22.0-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3e2daa88dc91870215961e96a039ec73e4937da13cf77ce17f9cad0c18df3503", size = 336581, upload-time = "2025-10-06T14:09:22.98Z" }, + { url = "https://files.pythonhosted.org/packages/0b/f5/5777b19e26fdf98563985e481f8be3d8a39f8734147a6ebf459d0dab5a6b/yarl-1.22.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ba440ae430c00eee41509353628600212112cd5018d5def7e9b05ea7ac34eb65", size = 388924, upload-time = "2025-10-06T14:09:24.655Z" }, + { url = "https://files.pythonhosted.org/packages/86/08/24bd2477bd59c0bbd994fe1d93b126e0472e4e3df5a96a277b0a55309e89/yarl-1.22.0-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:e6438cc8f23a9c1478633d216b16104a586b9761db62bfacb6425bac0a36679e", size = 392890, upload-time = "2025-10-06T14:09:26.617Z" }, + { url = "https://files.pythonhosted.org/packages/46/00/71b90ed48e895667ecfb1eaab27c1523ee2fa217433ed77a73b13205ca4b/yarl-1.22.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4c52a6e78aef5cf47a98ef8e934755abf53953379b7d53e68b15ff4420e6683d", size = 365819, upload-time = "2025-10-06T14:09:28.544Z" }, + { url = "https://files.pythonhosted.org/packages/30/2d/f715501cae832651d3282387c6a9236cd26bd00d0ff1e404b3dc52447884/yarl-1.22.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:3b06bcadaac49c70f4c88af4ffcfbe3dc155aab3163e75777818092478bcbbe7", size = 363601, upload-time = "2025-10-06T14:09:30.568Z" }, + { url = "https://files.pythonhosted.org/packages/f8/f9/a678c992d78e394e7126ee0b0e4e71bd2775e4334d00a9278c06a6cce96a/yarl-1.22.0-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:6944b2dc72c4d7f7052683487e3677456050ff77fcf5e6204e98caf785ad1967", size = 358072, upload-time = "2025-10-06T14:09:32.528Z" }, + { url = "https://files.pythonhosted.org/packages/2c/d1/b49454411a60edb6fefdcad4f8e6dbba7d8019e3a508a1c5836cba6d0781/yarl-1.22.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:d5372ca1df0f91a86b047d1277c2aaf1edb32d78bbcefffc81b40ffd18f027ed", size = 385311, upload-time = "2025-10-06T14:09:34.634Z" }, + { url = "https://files.pythonhosted.org/packages/87/e5/40d7a94debb8448c7771a916d1861d6609dddf7958dc381117e7ba36d9e8/yarl-1.22.0-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:51af598701f5299012b8416486b40fceef8c26fc87dc6d7d1f6fc30609ea0aa6", size = 381094, upload-time = "2025-10-06T14:09:36.268Z" }, + { url = "https://files.pythonhosted.org/packages/35/d8/611cc282502381ad855448643e1ad0538957fc82ae83dfe7762c14069e14/yarl-1.22.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:b266bd01fedeffeeac01a79ae181719ff848a5a13ce10075adbefc8f1daee70e", size = 370944, upload-time = "2025-10-06T14:09:37.872Z" }, + { url = "https://files.pythonhosted.org/packages/2d/df/fadd00fb1c90e1a5a8bd731fa3d3de2e165e5a3666a095b04e31b04d9cb6/yarl-1.22.0-cp311-cp311-win32.whl", hash = "sha256:a9b1ba5610a4e20f655258d5a1fdc7ebe3d837bb0e45b581398b99eb98b1f5ca", size = 81804, upload-time = "2025-10-06T14:09:39.359Z" }, + { url = "https://files.pythonhosted.org/packages/b5/f7/149bb6f45f267cb5c074ac40c01c6b3ea6d8a620d34b337f6321928a1b4d/yarl-1.22.0-cp311-cp311-win_amd64.whl", hash = "sha256:078278b9b0b11568937d9509b589ee83ef98ed6d561dfe2020e24a9fd08eaa2b", size = 86858, upload-time = "2025-10-06T14:09:41.068Z" }, + { url = "https://files.pythonhosted.org/packages/2b/13/88b78b93ad3f2f0b78e13bfaaa24d11cbc746e93fe76d8c06bf139615646/yarl-1.22.0-cp311-cp311-win_arm64.whl", hash = "sha256:b6a6f620cfe13ccec221fa312139135166e47ae169f8253f72a0abc0dae94376", size = 81637, upload-time = "2025-10-06T14:09:42.712Z" }, + { url = "https://files.pythonhosted.org/packages/75/ff/46736024fee3429b80a165a732e38e5d5a238721e634ab41b040d49f8738/yarl-1.22.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:e340382d1afa5d32b892b3ff062436d592ec3d692aeea3bef3a5cfe11bbf8c6f", size = 142000, upload-time = "2025-10-06T14:09:44.631Z" }, + { url = "https://files.pythonhosted.org/packages/5a/9a/b312ed670df903145598914770eb12de1bac44599549b3360acc96878df8/yarl-1.22.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:f1e09112a2c31ffe8d80be1b0988fa6a18c5d5cad92a9ffbb1c04c91bfe52ad2", size = 94338, upload-time = "2025-10-06T14:09:46.372Z" }, + { url = "https://files.pythonhosted.org/packages/ba/f5/0601483296f09c3c65e303d60c070a5c19fcdbc72daa061e96170785bc7d/yarl-1.22.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:939fe60db294c786f6b7c2d2e121576628468f65453d86b0fe36cb52f987bd74", size = 94909, upload-time = "2025-10-06T14:09:48.648Z" }, + { url = "https://files.pythonhosted.org/packages/60/41/9a1fe0b73dbcefce72e46cf149b0e0a67612d60bfc90fb59c2b2efdfbd86/yarl-1.22.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e1651bf8e0398574646744c1885a41198eba53dc8a9312b954073f845c90a8df", size = 372940, upload-time = "2025-10-06T14:09:50.089Z" }, + { url = "https://files.pythonhosted.org/packages/17/7a/795cb6dfee561961c30b800f0ed616b923a2ec6258b5def2a00bf8231334/yarl-1.22.0-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b8a0588521a26bf92a57a1705b77b8b59044cdceccac7151bd8d229e66b8dedb", size = 345825, upload-time = "2025-10-06T14:09:52.142Z" }, + { url = "https://files.pythonhosted.org/packages/d7/93/a58f4d596d2be2ae7bab1a5846c4d270b894958845753b2c606d666744d3/yarl-1.22.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:42188e6a615c1a75bcaa6e150c3fe8f3e8680471a6b10150c5f7e83f47cc34d2", size = 386705, upload-time = "2025-10-06T14:09:54.128Z" }, + { url = "https://files.pythonhosted.org/packages/61/92/682279d0e099d0e14d7fd2e176bd04f48de1484f56546a3e1313cd6c8e7c/yarl-1.22.0-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:f6d2cb59377d99718913ad9a151030d6f83ef420a2b8f521d94609ecc106ee82", size = 396518, upload-time = "2025-10-06T14:09:55.762Z" }, + { url = "https://files.pythonhosted.org/packages/db/0f/0d52c98b8a885aeda831224b78f3be7ec2e1aa4a62091f9f9188c3c65b56/yarl-1.22.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:50678a3b71c751d58d7908edc96d332af328839eea883bb554a43f539101277a", size = 377267, upload-time = "2025-10-06T14:09:57.958Z" }, + { url = "https://files.pythonhosted.org/packages/22/42/d2685e35908cbeaa6532c1fc73e89e7f2efb5d8a7df3959ea8e37177c5a3/yarl-1.22.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:1e8fbaa7cec507aa24ea27a01456e8dd4b6fab829059b69844bd348f2d467124", size = 365797, upload-time = "2025-10-06T14:09:59.527Z" }, + { url = "https://files.pythonhosted.org/packages/a2/83/cf8c7bcc6355631762f7d8bdab920ad09b82efa6b722999dfb05afa6cfac/yarl-1.22.0-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:433885ab5431bc3d3d4f2f9bd15bfa1614c522b0f1405d62c4f926ccd69d04fa", size = 365535, upload-time = "2025-10-06T14:10:01.139Z" }, + { url = "https://files.pythonhosted.org/packages/25/e1/5302ff9b28f0c59cac913b91fe3f16c59a033887e57ce9ca5d41a3a94737/yarl-1.22.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:b790b39c7e9a4192dc2e201a282109ed2985a1ddbd5ac08dc56d0e121400a8f7", size = 382324, upload-time = "2025-10-06T14:10:02.756Z" }, + { url = "https://files.pythonhosted.org/packages/bf/cd/4617eb60f032f19ae3a688dc990d8f0d89ee0ea378b61cac81ede3e52fae/yarl-1.22.0-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:31f0b53913220599446872d757257be5898019c85e7971599065bc55065dc99d", size = 383803, upload-time = "2025-10-06T14:10:04.552Z" }, + { url = "https://files.pythonhosted.org/packages/59/65/afc6e62bb506a319ea67b694551dab4a7e6fb7bf604e9bd9f3e11d575fec/yarl-1.22.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:a49370e8f711daec68d09b821a34e1167792ee2d24d405cbc2387be4f158b520", size = 374220, upload-time = "2025-10-06T14:10:06.489Z" }, + { url = "https://files.pythonhosted.org/packages/e7/3d/68bf18d50dc674b942daec86a9ba922d3113d8399b0e52b9897530442da2/yarl-1.22.0-cp312-cp312-win32.whl", hash = "sha256:70dfd4f241c04bd9239d53b17f11e6ab672b9f1420364af63e8531198e3f5fe8", size = 81589, upload-time = "2025-10-06T14:10:09.254Z" }, + { url = "https://files.pythonhosted.org/packages/c8/9a/6ad1a9b37c2f72874f93e691b2e7ecb6137fb2b899983125db4204e47575/yarl-1.22.0-cp312-cp312-win_amd64.whl", hash = "sha256:8884d8b332a5e9b88e23f60bb166890009429391864c685e17bd73a9eda9105c", size = 87213, upload-time = "2025-10-06T14:10:11.369Z" }, + { url = "https://files.pythonhosted.org/packages/44/c5/c21b562d1680a77634d748e30c653c3ca918beb35555cff24986fff54598/yarl-1.22.0-cp312-cp312-win_arm64.whl", hash = "sha256:ea70f61a47f3cc93bdf8b2f368ed359ef02a01ca6393916bc8ff877427181e74", size = 81330, upload-time = "2025-10-06T14:10:13.112Z" }, + { url = "https://files.pythonhosted.org/packages/ea/f3/d67de7260456ee105dc1d162d43a019ecad6b91e2f51809d6cddaa56690e/yarl-1.22.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:8dee9c25c74997f6a750cd317b8ca63545169c098faee42c84aa5e506c819b53", size = 139980, upload-time = "2025-10-06T14:10:14.601Z" }, + { url = "https://files.pythonhosted.org/packages/01/88/04d98af0b47e0ef42597b9b28863b9060bb515524da0a65d5f4db160b2d5/yarl-1.22.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:01e73b85a5434f89fc4fe27dcda2aff08ddf35e4d47bbbea3bdcd25321af538a", size = 93424, upload-time = "2025-10-06T14:10:16.115Z" }, + { url = "https://files.pythonhosted.org/packages/18/91/3274b215fd8442a03975ce6bee5fe6aa57a8326b29b9d3d56234a1dca244/yarl-1.22.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:22965c2af250d20c873cdbee8ff958fb809940aeb2e74ba5f20aaf6b7ac8c70c", size = 93821, upload-time = "2025-10-06T14:10:17.993Z" }, + { url = "https://files.pythonhosted.org/packages/61/3a/caf4e25036db0f2da4ca22a353dfeb3c9d3c95d2761ebe9b14df8fc16eb0/yarl-1.22.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b4f15793aa49793ec8d1c708ab7f9eded1aa72edc5174cae703651555ed1b601", size = 373243, upload-time = "2025-10-06T14:10:19.44Z" }, + { url = "https://files.pythonhosted.org/packages/6e/9e/51a77ac7516e8e7803b06e01f74e78649c24ee1021eca3d6a739cb6ea49c/yarl-1.22.0-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e5542339dcf2747135c5c85f68680353d5cb9ffd741c0f2e8d832d054d41f35a", size = 342361, upload-time = "2025-10-06T14:10:21.124Z" }, + { url = "https://files.pythonhosted.org/packages/d4/f8/33b92454789dde8407f156c00303e9a891f1f51a0330b0fad7c909f87692/yarl-1.22.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5c401e05ad47a75869c3ab3e35137f8468b846770587e70d71e11de797d113df", size = 387036, upload-time = "2025-10-06T14:10:22.902Z" }, + { url = "https://files.pythonhosted.org/packages/d9/9a/c5db84ea024f76838220280f732970aa4ee154015d7f5c1bfb60a267af6f/yarl-1.22.0-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:243dda95d901c733f5b59214d28b0120893d91777cb8aa043e6ef059d3cddfe2", size = 397671, upload-time = "2025-10-06T14:10:24.523Z" }, + { url = "https://files.pythonhosted.org/packages/11/c9/cd8538dc2e7727095e0c1d867bad1e40c98f37763e6d995c1939f5fdc7b1/yarl-1.22.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bec03d0d388060058f5d291a813f21c011041938a441c593374da6077fe21b1b", size = 377059, upload-time = "2025-10-06T14:10:26.406Z" }, + { url = "https://files.pythonhosted.org/packages/a1/b9/ab437b261702ced75122ed78a876a6dec0a1b0f5e17a4ac7a9a2482d8abe/yarl-1.22.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:b0748275abb8c1e1e09301ee3cf90c8a99678a4e92e4373705f2a2570d581273", size = 365356, upload-time = "2025-10-06T14:10:28.461Z" }, + { url = "https://files.pythonhosted.org/packages/b2/9d/8e1ae6d1d008a9567877b08f0ce4077a29974c04c062dabdb923ed98e6fe/yarl-1.22.0-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:47fdb18187e2a4e18fda2c25c05d8251a9e4a521edaed757fef033e7d8498d9a", size = 361331, upload-time = "2025-10-06T14:10:30.541Z" }, + { url = "https://files.pythonhosted.org/packages/ca/5a/09b7be3905962f145b73beb468cdd53db8aa171cf18c80400a54c5b82846/yarl-1.22.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:c7044802eec4524fde550afc28edda0dd5784c4c45f0be151a2d3ba017daca7d", size = 382590, upload-time = "2025-10-06T14:10:33.352Z" }, + { url = "https://files.pythonhosted.org/packages/aa/7f/59ec509abf90eda5048b0bc3e2d7b5099dffdb3e6b127019895ab9d5ef44/yarl-1.22.0-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:139718f35149ff544caba20fce6e8a2f71f1e39b92c700d8438a0b1d2a631a02", size = 385316, upload-time = "2025-10-06T14:10:35.034Z" }, + { url = "https://files.pythonhosted.org/packages/e5/84/891158426bc8036bfdfd862fabd0e0fa25df4176ec793e447f4b85cf1be4/yarl-1.22.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e1b51bebd221006d3d2f95fbe124b22b247136647ae5dcc8c7acafba66e5ee67", size = 374431, upload-time = "2025-10-06T14:10:37.76Z" }, + { url = "https://files.pythonhosted.org/packages/bb/49/03da1580665baa8bef5e8ed34c6df2c2aca0a2f28bf397ed238cc1bbc6f2/yarl-1.22.0-cp313-cp313-win32.whl", hash = "sha256:d3e32536234a95f513bd374e93d717cf6b2231a791758de6c509e3653f234c95", size = 81555, upload-time = "2025-10-06T14:10:39.649Z" }, + { url = "https://files.pythonhosted.org/packages/9a/ee/450914ae11b419eadd067c6183ae08381cfdfcb9798b90b2b713bbebddda/yarl-1.22.0-cp313-cp313-win_amd64.whl", hash = "sha256:47743b82b76d89a1d20b83e60d5c20314cbd5ba2befc9cda8f28300c4a08ed4d", size = 86965, upload-time = "2025-10-06T14:10:41.313Z" }, + { url = "https://files.pythonhosted.org/packages/98/4d/264a01eae03b6cf629ad69bae94e3b0e5344741e929073678e84bf7a3e3b/yarl-1.22.0-cp313-cp313-win_arm64.whl", hash = "sha256:5d0fcda9608875f7d052eff120c7a5da474a6796fe4d83e152e0e4d42f6d1a9b", size = 81205, upload-time = "2025-10-06T14:10:43.167Z" }, + { url = "https://files.pythonhosted.org/packages/88/fc/6908f062a2f77b5f9f6d69cecb1747260831ff206adcbc5b510aff88df91/yarl-1.22.0-cp313-cp313t-macosx_10_13_universal2.whl", hash = "sha256:719ae08b6972befcba4310e49edb1161a88cdd331e3a694b84466bd938a6ab10", size = 146209, upload-time = "2025-10-06T14:10:44.643Z" }, + { url = "https://files.pythonhosted.org/packages/65/47/76594ae8eab26210b4867be6f49129861ad33da1f1ebdf7051e98492bf62/yarl-1.22.0-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:47d8a5c446df1c4db9d21b49619ffdba90e77c89ec6e283f453856c74b50b9e3", size = 95966, upload-time = "2025-10-06T14:10:46.554Z" }, + { url = "https://files.pythonhosted.org/packages/ab/ce/05e9828a49271ba6b5b038b15b3934e996980dd78abdfeb52a04cfb9467e/yarl-1.22.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:cfebc0ac8333520d2d0423cbbe43ae43c8838862ddb898f5ca68565e395516e9", size = 97312, upload-time = "2025-10-06T14:10:48.007Z" }, + { url = "https://files.pythonhosted.org/packages/d1/c5/7dffad5e4f2265b29c9d7ec869c369e4223166e4f9206fc2243ee9eea727/yarl-1.22.0-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4398557cbf484207df000309235979c79c4356518fd5c99158c7d38203c4da4f", size = 361967, upload-time = "2025-10-06T14:10:49.997Z" }, + { url = "https://files.pythonhosted.org/packages/50/b2/375b933c93a54bff7fc041e1a6ad2c0f6f733ffb0c6e642ce56ee3b39970/yarl-1.22.0-cp313-cp313t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:2ca6fd72a8cd803be290d42f2dec5cdcd5299eeb93c2d929bf060ad9efaf5de0", size = 323949, upload-time = "2025-10-06T14:10:52.004Z" }, + { url = "https://files.pythonhosted.org/packages/66/50/bfc2a29a1d78644c5a7220ce2f304f38248dc94124a326794e677634b6cf/yarl-1.22.0-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ca1f59c4e1ab6e72f0a23c13fca5430f889634166be85dbf1013683e49e3278e", size = 361818, upload-time = "2025-10-06T14:10:54.078Z" }, + { url = "https://files.pythonhosted.org/packages/46/96/f3941a46af7d5d0f0498f86d71275696800ddcdd20426298e572b19b91ff/yarl-1.22.0-cp313-cp313t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:6c5010a52015e7c70f86eb967db0f37f3c8bd503a695a49f8d45700144667708", size = 372626, upload-time = "2025-10-06T14:10:55.767Z" }, + { url = "https://files.pythonhosted.org/packages/c1/42/8b27c83bb875cd89448e42cd627e0fb971fa1675c9ec546393d18826cb50/yarl-1.22.0-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9d7672ecf7557476642c88497c2f8d8542f8e36596e928e9bcba0e42e1e7d71f", size = 341129, upload-time = "2025-10-06T14:10:57.985Z" }, + { url = "https://files.pythonhosted.org/packages/49/36/99ca3122201b382a3cf7cc937b95235b0ac944f7e9f2d5331d50821ed352/yarl-1.22.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:3b7c88eeef021579d600e50363e0b6ee4f7f6f728cd3486b9d0f3ee7b946398d", size = 346776, upload-time = "2025-10-06T14:10:59.633Z" }, + { url = "https://files.pythonhosted.org/packages/85/b4/47328bf996acd01a4c16ef9dcd2f59c969f495073616586f78cd5f2efb99/yarl-1.22.0-cp313-cp313t-musllinux_1_2_armv7l.whl", hash = "sha256:f4afb5c34f2c6fecdcc182dfcfc6af6cccf1aa923eed4d6a12e9d96904e1a0d8", size = 334879, upload-time = "2025-10-06T14:11:01.454Z" }, + { url = "https://files.pythonhosted.org/packages/c2/ad/b77d7b3f14a4283bffb8e92c6026496f6de49751c2f97d4352242bba3990/yarl-1.22.0-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:59c189e3e99a59cf8d83cbb31d4db02d66cda5a1a4374e8a012b51255341abf5", size = 350996, upload-time = "2025-10-06T14:11:03.452Z" }, + { url = "https://files.pythonhosted.org/packages/81/c8/06e1d69295792ba54d556f06686cbd6a7ce39c22307100e3fb4a2c0b0a1d/yarl-1.22.0-cp313-cp313t-musllinux_1_2_s390x.whl", hash = "sha256:5a3bf7f62a289fa90f1990422dc8dff5a458469ea71d1624585ec3a4c8d6960f", size = 356047, upload-time = "2025-10-06T14:11:05.115Z" }, + { url = "https://files.pythonhosted.org/packages/4b/b8/4c0e9e9f597074b208d18cef227d83aac36184bfbc6eab204ea55783dbc5/yarl-1.22.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:de6b9a04c606978fdfe72666fa216ffcf2d1a9f6a381058d4378f8d7b1e5de62", size = 342947, upload-time = "2025-10-06T14:11:08.137Z" }, + { url = "https://files.pythonhosted.org/packages/e0/e5/11f140a58bf4c6ad7aca69a892bff0ee638c31bea4206748fc0df4ebcb3a/yarl-1.22.0-cp313-cp313t-win32.whl", hash = "sha256:1834bb90991cc2999f10f97f5f01317f99b143284766d197e43cd5b45eb18d03", size = 86943, upload-time = "2025-10-06T14:11:10.284Z" }, + { url = "https://files.pythonhosted.org/packages/31/74/8b74bae38ed7fe6793d0c15a0c8207bbb819cf287788459e5ed230996cdd/yarl-1.22.0-cp313-cp313t-win_amd64.whl", hash = "sha256:ff86011bd159a9d2dfc89c34cfd8aff12875980e3bd6a39ff097887520e60249", size = 93715, upload-time = "2025-10-06T14:11:11.739Z" }, + { url = "https://files.pythonhosted.org/packages/69/66/991858aa4b5892d57aef7ee1ba6b4d01ec3b7eb3060795d34090a3ca3278/yarl-1.22.0-cp313-cp313t-win_arm64.whl", hash = "sha256:7861058d0582b847bc4e3a4a4c46828a410bca738673f35a29ba3ca5db0b473b", size = 83857, upload-time = "2025-10-06T14:11:13.586Z" }, + { url = "https://files.pythonhosted.org/packages/46/b3/e20ef504049f1a1c54a814b4b9bed96d1ac0e0610c3b4da178f87209db05/yarl-1.22.0-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:34b36c2c57124530884d89d50ed2c1478697ad7473efd59cfd479945c95650e4", size = 140520, upload-time = "2025-10-06T14:11:15.465Z" }, + { url = "https://files.pythonhosted.org/packages/e4/04/3532d990fdbab02e5ede063676b5c4260e7f3abea2151099c2aa745acc4c/yarl-1.22.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:0dd9a702591ca2e543631c2a017e4a547e38a5c0f29eece37d9097e04a7ac683", size = 93504, upload-time = "2025-10-06T14:11:17.106Z" }, + { url = "https://files.pythonhosted.org/packages/11/63/ff458113c5c2dac9a9719ac68ee7c947cb621432bcf28c9972b1c0e83938/yarl-1.22.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:594fcab1032e2d2cc3321bb2e51271e7cd2b516c7d9aee780ece81b07ff8244b", size = 94282, upload-time = "2025-10-06T14:11:19.064Z" }, + { url = "https://files.pythonhosted.org/packages/a7/bc/315a56aca762d44a6aaaf7ad253f04d996cb6b27bad34410f82d76ea8038/yarl-1.22.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f3d7a87a78d46a2e3d5b72587ac14b4c16952dd0887dbb051451eceac774411e", size = 372080, upload-time = "2025-10-06T14:11:20.996Z" }, + { url = "https://files.pythonhosted.org/packages/3f/3f/08e9b826ec2e099ea6e7c69a61272f4f6da62cb5b1b63590bb80ca2e4a40/yarl-1.22.0-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:852863707010316c973162e703bddabec35e8757e67fcb8ad58829de1ebc8590", size = 338696, upload-time = "2025-10-06T14:11:22.847Z" }, + { url = "https://files.pythonhosted.org/packages/e3/9f/90360108e3b32bd76789088e99538febfea24a102380ae73827f62073543/yarl-1.22.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:131a085a53bfe839a477c0845acf21efc77457ba2bcf5899618136d64f3303a2", size = 387121, upload-time = "2025-10-06T14:11:24.889Z" }, + { url = "https://files.pythonhosted.org/packages/98/92/ab8d4657bd5b46a38094cfaea498f18bb70ce6b63508fd7e909bd1f93066/yarl-1.22.0-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:078a8aefd263f4d4f923a9677b942b445a2be970ca24548a8102689a3a8ab8da", size = 394080, upload-time = "2025-10-06T14:11:27.307Z" }, + { url = "https://files.pythonhosted.org/packages/f5/e7/d8c5a7752fef68205296201f8ec2bf718f5c805a7a7e9880576c67600658/yarl-1.22.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bca03b91c323036913993ff5c738d0842fc9c60c4648e5c8d98331526df89784", size = 372661, upload-time = "2025-10-06T14:11:29.387Z" }, + { url = "https://files.pythonhosted.org/packages/b6/2e/f4d26183c8db0bb82d491b072f3127fb8c381a6206a3a56332714b79b751/yarl-1.22.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:68986a61557d37bb90d3051a45b91fa3d5c516d177dfc6dd6f2f436a07ff2b6b", size = 364645, upload-time = "2025-10-06T14:11:31.423Z" }, + { url = "https://files.pythonhosted.org/packages/80/7c/428e5812e6b87cd00ee8e898328a62c95825bf37c7fa87f0b6bb2ad31304/yarl-1.22.0-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:4792b262d585ff0dff6bcb787f8492e40698443ec982a3568c2096433660c694", size = 355361, upload-time = "2025-10-06T14:11:33.055Z" }, + { url = "https://files.pythonhosted.org/packages/ec/2a/249405fd26776f8b13c067378ef4d7dd49c9098d1b6457cdd152a99e96a9/yarl-1.22.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:ebd4549b108d732dba1d4ace67614b9545b21ece30937a63a65dd34efa19732d", size = 381451, upload-time = "2025-10-06T14:11:35.136Z" }, + { url = "https://files.pythonhosted.org/packages/67/a8/fb6b1adbe98cf1e2dd9fad71003d3a63a1bc22459c6e15f5714eb9323b93/yarl-1.22.0-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:f87ac53513d22240c7d59203f25cc3beac1e574c6cd681bbfd321987b69f95fd", size = 383814, upload-time = "2025-10-06T14:11:37.094Z" }, + { url = "https://files.pythonhosted.org/packages/d9/f9/3aa2c0e480fb73e872ae2814c43bc1e734740bb0d54e8cb2a95925f98131/yarl-1.22.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:22b029f2881599e2f1b06f8f1db2ee63bd309e2293ba2d566e008ba12778b8da", size = 370799, upload-time = "2025-10-06T14:11:38.83Z" }, + { url = "https://files.pythonhosted.org/packages/50/3c/af9dba3b8b5eeb302f36f16f92791f3ea62e3f47763406abf6d5a4a3333b/yarl-1.22.0-cp314-cp314-win32.whl", hash = "sha256:6a635ea45ba4ea8238463b4f7d0e721bad669f80878b7bfd1f89266e2ae63da2", size = 82990, upload-time = "2025-10-06T14:11:40.624Z" }, + { url = "https://files.pythonhosted.org/packages/ac/30/ac3a0c5bdc1d6efd1b41fa24d4897a4329b3b1e98de9449679dd327af4f0/yarl-1.22.0-cp314-cp314-win_amd64.whl", hash = "sha256:0d6e6885777af0f110b0e5d7e5dda8b704efed3894da26220b7f3d887b839a79", size = 88292, upload-time = "2025-10-06T14:11:42.578Z" }, + { url = "https://files.pythonhosted.org/packages/df/0a/227ab4ff5b998a1b7410abc7b46c9b7a26b0ca9e86c34ba4b8d8bc7c63d5/yarl-1.22.0-cp314-cp314-win_arm64.whl", hash = "sha256:8218f4e98d3c10d683584cb40f0424f4b9fd6e95610232dd75e13743b070ee33", size = 82888, upload-time = "2025-10-06T14:11:44.863Z" }, + { url = "https://files.pythonhosted.org/packages/06/5e/a15eb13db90abd87dfbefb9760c0f3f257ac42a5cac7e75dbc23bed97a9f/yarl-1.22.0-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:45c2842ff0e0d1b35a6bf1cd6c690939dacb617a70827f715232b2e0494d55d1", size = 146223, upload-time = "2025-10-06T14:11:46.796Z" }, + { url = "https://files.pythonhosted.org/packages/18/82/9665c61910d4d84f41a5bf6837597c89e665fa88aa4941080704645932a9/yarl-1.22.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:d947071e6ebcf2e2bee8fce76e10faca8f7a14808ca36a910263acaacef08eca", size = 95981, upload-time = "2025-10-06T14:11:48.845Z" }, + { url = "https://files.pythonhosted.org/packages/5d/9a/2f65743589809af4d0a6d3aa749343c4b5f4c380cc24a8e94a3c6625a808/yarl-1.22.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:334b8721303e61b00019474cc103bdac3d7b1f65e91f0bfedeec2d56dfe74b53", size = 97303, upload-time = "2025-10-06T14:11:50.897Z" }, + { url = "https://files.pythonhosted.org/packages/b0/ab/5b13d3e157505c43c3b43b5a776cbf7b24a02bc4cccc40314771197e3508/yarl-1.22.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1e7ce67c34138a058fd092f67d07a72b8e31ff0c9236e751957465a24b28910c", size = 361820, upload-time = "2025-10-06T14:11:52.549Z" }, + { url = "https://files.pythonhosted.org/packages/fb/76/242a5ef4677615cf95330cfc1b4610e78184400699bdda0acb897ef5e49a/yarl-1.22.0-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:d77e1b2c6d04711478cb1c4ab90db07f1609ccf06a287d5607fcd90dc9863acf", size = 323203, upload-time = "2025-10-06T14:11:54.225Z" }, + { url = "https://files.pythonhosted.org/packages/8c/96/475509110d3f0153b43d06164cf4195c64d16999e0c7e2d8a099adcd6907/yarl-1.22.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c4647674b6150d2cae088fc07de2738a84b8bcedebef29802cf0b0a82ab6face", size = 363173, upload-time = "2025-10-06T14:11:56.069Z" }, + { url = "https://files.pythonhosted.org/packages/c9/66/59db471aecfbd559a1fd48aedd954435558cd98c7d0da8b03cc6c140a32c/yarl-1.22.0-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:efb07073be061c8f79d03d04139a80ba33cbd390ca8f0297aae9cce6411e4c6b", size = 373562, upload-time = "2025-10-06T14:11:58.783Z" }, + { url = "https://files.pythonhosted.org/packages/03/1f/c5d94abc91557384719da10ff166b916107c1b45e4d0423a88457071dd88/yarl-1.22.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e51ac5435758ba97ad69617e13233da53908beccc6cfcd6c34bbed8dcbede486", size = 339828, upload-time = "2025-10-06T14:12:00.686Z" }, + { url = "https://files.pythonhosted.org/packages/5f/97/aa6a143d3afba17b6465733681c70cf175af89f76ec8d9286e08437a7454/yarl-1.22.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:33e32a0dd0c8205efa8e83d04fc9f19313772b78522d1bdc7d9aed706bfd6138", size = 347551, upload-time = "2025-10-06T14:12:02.628Z" }, + { url = "https://files.pythonhosted.org/packages/43/3c/45a2b6d80195959239a7b2a8810506d4eea5487dce61c2a3393e7fc3c52e/yarl-1.22.0-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:bf4a21e58b9cde0e401e683ebd00f6ed30a06d14e93f7c8fd059f8b6e8f87b6a", size = 334512, upload-time = "2025-10-06T14:12:04.871Z" }, + { url = "https://files.pythonhosted.org/packages/86/a0/c2ab48d74599c7c84cb104ebd799c5813de252bea0f360ffc29d270c2caa/yarl-1.22.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:e4b582bab49ac33c8deb97e058cd67c2c50dac0dd134874106d9c774fd272529", size = 352400, upload-time = "2025-10-06T14:12:06.624Z" }, + { url = "https://files.pythonhosted.org/packages/32/75/f8919b2eafc929567d3d8411f72bdb1a2109c01caaab4ebfa5f8ffadc15b/yarl-1.22.0-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:0b5bcc1a9c4839e7e30b7b30dd47fe5e7e44fb7054ec29b5bb8d526aa1041093", size = 357140, upload-time = "2025-10-06T14:12:08.362Z" }, + { url = "https://files.pythonhosted.org/packages/cf/72/6a85bba382f22cf78add705d8c3731748397d986e197e53ecc7835e76de7/yarl-1.22.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:c0232bce2170103ec23c454e54a57008a9a72b5d1c3105dc2496750da8cfa47c", size = 341473, upload-time = "2025-10-06T14:12:10.994Z" }, + { url = "https://files.pythonhosted.org/packages/35/18/55e6011f7c044dc80b98893060773cefcfdbf60dfefb8cb2f58b9bacbd83/yarl-1.22.0-cp314-cp314t-win32.whl", hash = "sha256:8009b3173bcd637be650922ac455946197d858b3630b6d8787aa9e5c4564533e", size = 89056, upload-time = "2025-10-06T14:12:13.317Z" }, + { url = "https://files.pythonhosted.org/packages/f9/86/0f0dccb6e59a9e7f122c5afd43568b1d31b8ab7dda5f1b01fb5c7025c9a9/yarl-1.22.0-cp314-cp314t-win_amd64.whl", hash = "sha256:9fb17ea16e972c63d25d4a97f016d235c78dd2344820eb35bc034bc32012ee27", size = 96292, upload-time = "2025-10-06T14:12:15.398Z" }, + { url = "https://files.pythonhosted.org/packages/48/b7/503c98092fb3b344a179579f55814b613c1fbb1c23b3ec14a7b008a66a6e/yarl-1.22.0-cp314-cp314t-win_arm64.whl", hash = "sha256:9f6d73c1436b934e3f01df1e1b21ff765cd1d28c77dfb9ace207f746d4610ee1", size = 85171, upload-time = "2025-10-06T14:12:16.935Z" }, + { url = "https://files.pythonhosted.org/packages/73/ae/b48f95715333080afb75a4504487cbe142cae1268afc482d06692d605ae6/yarl-1.22.0-py3-none-any.whl", hash = "sha256:1380560bdba02b6b6c90de54133c81c9f2a453dee9912fe58c1dcced1edb7cff", size = 46814, upload-time = "2025-10-06T14:12:53.872Z" }, +] + +[[package]] +name = "zipp" +version = "3.23.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/e3/02/0f2892c661036d50ede074e376733dca2ae7c6eb617489437771209d4180/zipp-3.23.0.tar.gz", hash = "sha256:a07157588a12518c9d4034df3fbbee09c814741a33ff63c05fa29d26a2404166", size = 25547, upload-time = "2025-06-08T17:06:39.4Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/2e/54/647ade08bf0db230bfea292f893923872fd20be6ac6f53b2b936ba839d75/zipp-3.23.0-py3-none-any.whl", hash = "sha256:071652d6115ed432f5ce1d34c336c0adfd6a884660d1e9712a256d3d3bd4b14e", size = 10276, upload-time = "2025-06-08T17:06:38.034Z" }, +]