ckpt

Fixes the sequence parallel sharding on t, now SP shards on t*h*w

.buildkite/pipeline.yml

@@ -22,7 +22,7 @@ steps:
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                command: "timeout 20m .buildkite/scripts/pr_test.sh"
                 label: "Encoder Tests"
                 env:
                   - TEST_TYPE=encoder
@@ -35,7 +35,7 @@ steps:
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                command: "timeout 20m .buildkite/scripts/pr_test.sh"
                 label: "VAE Tests"
                 env:
                   - TEST_TYPE=vae
@@ -61,7 +61,7 @@ steps:
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
-                command: "timeout 45m .buildkite/scripts/pr_test.sh"
+                command: "timeout 90m .buildkite/scripts/pr_test.sh"
                 label: "SSIM Tests"
                 env:
                   - TEST_TYPE=ssim
@@ -76,7 +76,7 @@ steps:
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                command: "timeout 20m .buildkite/scripts/pr_test.sh"
                 label: "LoRA Inference Tests"
                 env:
                   - TEST_TYPE=inference_lora
@@ -129,11 +129,7 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/**"
-                - "csrc/attn/video_sparse_attn/**"
-                - "csrc/attn/video_sparse_attn/tk/**"
-                - "csrc/attn/video_sparse_attn/setup.py"
-                - "csrc/attn/video_sparse_attn/config_vsa.py"
-                - "csrc/attn/video_sparse_attn/vsa.cpp"
+                - "fastvideo-kernel/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -145,10 +141,7 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/**"
-                - "csrc/attn/sliding_tile_attn/**"
-                - "csrc/attn/sliding_tile_attn/setup.py"
-                - "csrc/attn/sliding_tile_attn/config_sta.py"
-                - "csrc/attn/sliding_tile_attn/st_attn.cpp"
+                - "fastvideo-kernel/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -159,48 +152,16 @@ steps:
                 agents:
                   queue: "default"
             - path:
-                - "csrc/attn/sliding_tile_attn/**"
-                - "csrc/attn/sliding_tile_attn/setup.py"
-                - "csrc/attn/sliding_tile_attn/config_sta.py"
-                - "csrc/attn/sliding_tile_attn/st_attn.cpp"
+                - "fastvideo-kernel/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
                 command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "Precision Tests STA"
+                label: "Kernel Tests"
                 env:
-                  - TEST_TYPE=precision_sta
-                agents:
-                  queue: "default"
+                  - TEST_TYPE=kernel_tests
             - path:
-                - "csrc/attn/video_sparse_attn/**"
-                - "csrc/attn/video_sparse_attn/tk/**"
-                - "csrc/attn/tests/test_vsa.py"
-                - "csrc/attn/video_sparse_attn/setup.py"
-                - "csrc/attn/video_sparse_attn/config_vsa.py"
-                - "csrc/attn/video_sparse_attn/vsa.cpp"
-                - "pyproject.toml"
-                - "docker/Dockerfile.python3.12"
-              config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "Precision Tests VSA"
-                env:
-                  - TEST_TYPE=precision_vsa
-                agents:
-                  queue: "default"
-            - path:
-                - "csrc/attn/vmoba_attn/**"
-                - "pyproject.toml"
-                - "docker/Dockerfile.python3.12"
-              config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "Precision Tests VMoBA"
-                env:
-                  - TEST_TYPE=precision_vmoba
-                agents:
-                  queue: "default"
-            - path:
-                - "csrc/attn/vmoba_attn/vmoba/**"
+                - "fastvideo-kernel/**"
                 - "fastvideo/attention/backends/vmoba.py"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
@@ -222,3 +183,14 @@ steps:
                   - TEST_TYPE=unit_test
                 agents:
                   queue: "default"
+            # - path:
+            #     - "scripts/lora_extraction/**"
+            #     - "pyproject.toml"
+            #     - "docker/Dockerfile.python3.12"
+            #   config:
+            #     command: "timeout 90m .buildkite/scripts/pr_test.sh"
+            #     label: "LoRA Extraction Tests"
+            #     env:
+            #       - TEST_TYPE=lora_extraction
+            #     agents:
+            #       queue: "default"

.buildkite/scripts/pr_test.sh

@@ -75,7 +75,7 @@ case "$TEST_TYPE" in
         ;;
     "ssim")
         log "Running SSIM tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_ssim_tests"
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_ssim_tests"
         ;;
     "training")
         log "Running training tests..."
@@ -93,13 +93,9 @@ case "$TEST_TYPE" in
         log "Running inference STA tests..."
         MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_inference_tests_STA"
         ;;
-    "precision_sta")
-        log "Running precision STA tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_precision_tests_STA"
-        ;;
-    "precision_vsa")
-        log "Running precision VSA tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_precision_tests_VSA"
+    "kernel_tests")
+        log "Running kernel tests..."
+        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_kernel_tests"
         ;;
     "inference_lora")
         log "Running LoRA tests..."
@@ -118,14 +114,14 @@ case "$TEST_TYPE" in
         log "Running V-MoBA inference tests..."
         MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_inference_tests_vmoba"
         ;;
-    "precision_vmoba")
-        log "Running V-MoBA precision tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_precision_tests_vmoba"
-        ;;
     "unit_test")
         log "Running unit tests..."
         MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_unit_test"
         ;;
+    "lora_extraction")
+        log "Running LoRA extraction tests..."
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_lora_extraction_tests"
+        ;;
     *)
         log "Error: Unknown test type: $TEST_TYPE"
         exit 1

.github/workflows/docs.yml

@@ -3,8 +3,18 @@ name: Deploy Documentation
 on:
   push:
     branches: [ main ]
+    paths:
+      - 'docs/**'
+      - 'mkdocs.yml'
+      - 'requirements-mkdocs.txt'
+      - '.github/workflows/docs.yml'
   pull_request:
     branches: [ main ]
+    paths:
+      - 'docs/**'
+      - 'mkdocs.yml'
+      - 'requirements-mkdocs.txt'
+      - '.github/workflows/docs.yml'
 
 permissions:
   contents: read

.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/

.gitignore

@@ -30,6 +30,8 @@ env
 **/build/
 **.pyc
 **.txt
+*.log
+weights/
 
 # Distribution / packaging
 build/

.gitmodules

@@ -1,7 +1,6 @@
-[submodule "csrc/attn/video_sparse_attn/tk"]
-	path = csrc/attn/video_sparse_attn/tk
-	url = https://github.com/HazyResearch/ThunderKittens.git
-
-[submodule "csrc/attn/sliding_tile_attn/tk"]
-	path = csrc/attn/sliding_tile_attn/tk
+[submodule "fastvideo-kernel/include/tk"]
+	path = fastvideo-kernel/include/tk
 	url = https://github.com/HazyResearch/ThunderKittens.git
+[submodule "fastvideo-kernel/include/cutlass"]
+	path = fastvideo-kernel/include/cutlass
+	url = https://github.com/NVIDIA/cutlass.git

.pre-commit-config.yaml

@@ -4,12 +4,13 @@ default_stages:
 exclude: |
   (?x)(
     fastvideo/third_party/.*|
-    csrc/.*|
+    fastvideo-kernel/.*|
     assets/.*|
     tests/.*|
     demo/.*|
     predict\.py|
     scripts/.*|
+    prompts/.*|
     fastvideo/data_preprocess/.*|
     fastvideo/dataset/.*|
     fastvideo/models/.*|
@@ -67,7 +68,7 @@ repos:
     entry: bash
     args:
       - -c
-      - 'git ls-files | grep -v "^fastvideo/tests/ssim/" | grep -v "^fastvideo/tests/inference/lora/L40S_reference_videos/" | grep " " && echo "Filenames should not contain spaces!" && exit 1 || exit 0'
+      - 'git ls-files | grep -v "^\"*fastvideo/tests/ssim/" | grep -v "^\"*fastvideo/tests/inference/lora/L40S_reference_videos/" | grep " " && echo "Filenames should not contain spaces!" && exit 1 || exit 0'
     language: system
     always_run: true
     pass_filenames: false

README.md

@@ -2,40 +2,46 @@
 <img src=assets/logos/logo.svg width="30%"/>
 </div>
 
-**FastVideo is a unified post-training and inference framework for accelerated video generation.**
-
-FastVideo features an end-to-end unified pipeline for accelerating diffusion models, starting from data preprocessing to model training, finetuning, distillation, and inference. FastVideo is designed to be modular and extensible, allowing users to easily add new optimizations and techniques. Whether it is training-free optimizations or post-training optimizations, FastVideo has you covered.
-
 <p align="center">
-    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start/"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/XcY0Cpv" target="_blank"> <b> WeChat </b> </a> |
+     | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start/"><b> Quick Start</b></a> | <a href="https://github.com/hao-ai-lab/FastVideo/discussions/982"  target="_blank"><b>Weekly Dev Meeting</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/sv3MMKyv" target="_blank"> <b> WeChat </b> </a> |
 </p>
 
-<div align="center">
-<img src=assets/fastwan.png width="90%"/>
-</div>
+**FastVideo is a unified post-training and inference framework for accelerated video generation.**
 
 ## NEWS
-- ```2025/08/04```: Release [FastWan](https://hao-ai-lab.github.io/FastVideo/distillation/dmd.html) models and [Sparse-Distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).
+- ```2025/11/19```: Release [CausalWan2.2 I2V A14B Preview](https://huggingface.co/FastVideo/CausalWan2.2-I2V-A14B-Preview-Diffusers) models, [Blog](https://hao-ai-lab.github.io/blogs/fastvideo_causalwan_preview/) and [Inference Code!](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py)
+- ```2025/08/04```: Release [FastWan](https://hao-ai-lab.github.io/FastVideo/distillation/dmd) models and [Sparse-Distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).
+
+<details>
+<summary>More</summary>
+
 - ```2025/06/14```: Release finetuning and inference code for [VSA](https://arxiv.org/pdf/2505.13389)
 - ```2025/04/24```: [FastVideo V1](https://hao-ai-lab.github.io/blogs/fastvideo/) is released!
 - ```2025/02/18```: Release the inference code for [Sliding Tile Attention](https://hao-ai-lab.github.io/blogs/sta/).
 
+</details>
+
 ## Key Features
 
 FastVideo has the following features:
-- End-to-end post-training support:
-  - [Sparse distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/) for Wan2.1 and Wan2.2 to achineve >50x denoising speedup
-  - Data preprocessing pipeline for video data
+- End-to-end post-training support for bidirectional and autoregressive models:
   - Support full finetuning and LoRA finetuning for state-of-the-art open video DiTs
-  - Scalable training with FSDP2, sequence parallelism, and selective activation checkpointing, with near linear scaling to 64 GPUs
+  - Data preprocessing pipeline for video, image, and text data
+  - Distribution Matching Distillation (DMD2) stepwise distillation.
+  - Sparse attention with [Video Sparse Attention](https://arxiv.org/pdf/2505.13389)
+  - [Sparse distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/) to achineve >50x denoising speedup
+  - Scalable training with FSDP2, sequence parallelism, and selective activation checkpointing.
+  - Causal distillation through Self-Forcing
+  - See this [page](https://hao-ai-lab.github.io/FastVideo/training/overview/) for full list of supported models and recipes.
 - State-of-the-art performance optimizations for inference
-  - [Video Sparse Attention](https://arxiv.org/pdf/2505.13389)
-  - [Sliding Tile Attention](https://arxiv.org/pdf/2502.04507)
-  - [TeaCache](https://arxiv.org/pdf/2411.19108)
-  - [Sage Attention](https://arxiv.org/abs/2410.02367)
+  - Sequence Parallelism for distributed inference
+  - Multiple state-of-the-art attention backends
+  - User-friendly CLI and Python API
+  - See this [page](https://hao-ai-lab.github.io/FastVideo/inference/optimizations/) for full list of supported optimizations.
 - Diverse hardware and OS support
   - Support H100, A100, 4090
   - Support Linux, Windows, MacOS
+  - See this [page](https://hao-ai-lab.github.io/FastVideo/inference/hardware_support/) for full list of supported hardware and OS.
 
 ## Getting Started
 We recommend using an environment manager such as `Conda` to create a clean environment:
@@ -111,24 +117,21 @@ For a more detailed guide, please see our [inference quick start](https://hao-ai
 - [Distillation Guide](https://hao-ai-lab.github.io/FastVideo/distillation/dmd/)
 <!-- - [Finetuning Guide](https://hao-ai-lab.github.io/FastVideo/training/finetune.html) -->
 
-## 📑 Development Plan
-<!-- - More distillation methods -->
-  <!-- - [ ] Add Distribution Matching Distillation -->
-More FastWan Models Coming Soon!
-- [ ] Add FastWan2.1-T2V-14B
-- [ ] Add FastWan2.2-T2V-14B
-- [ ] Add FastWan2.2-I2V-14B
-<!-- - Optimization features
-- Code updates -->
-  <!-- - [ ] fp8 support -->
-  <!-- - [ ] faster load model and save model support -->
+## Awesome work using FastVideo or our research projects
 
-See details in [development roadmap](https://github.com/hao-ai-lab/FastVideo/issues/468).
+- [SGLang](https://github.com/sgl-project/sglang/tree/main/python/sglang/multimodal_gen): SGLang's diffusion inference functionality is based  on a fork of FastVideo on Sept. 24, 2025. [![Star](https://img.shields.io/github/stars/sgl-project/sglang.svg?style=social&label=Star)](https://github.com/sgl-project/sglang)
+
+- [DanceGRPO](https://github.com/XueZeyue/DanceGRPO): A  unified framework to adapt Group Relative Policy Optimization (GRPO) to visual generation paradigms. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/XueZeyue/DanceGRPO.svg?style=social&label=Star)](https://github.com/XueZeyue/DanceGRPO)
+- [SRPO](https://github.com/Tencent-Hunyuan/SRPO): A method to directly align the full diffusion trajectory with fine-grained human preference. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/Tencent-Hunyuan/SRPO.svg?style=social&label=Star)](https://github.com/Tencent-Hunyuan/SRPO)
+- [DCM](https://github.com/Vchitect/DCM): Dual-expert consistency model for efficient and high-quality video generation. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/Vchitect/DCM.svg?style=social&label=Star)](https://github.com/Vchitect/DCM)
+- [Hunyuan Video 1.5](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5): A leading lightweight video generation model, where they proposed SSTA based on Sliding Tile Attention. [![Star](https://img.shields.io/github/stars/Tencent-Hunyuan/HunyuanVideo-1.5.svg?style=social&label=Star)](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5)
+- [Kandinsky-5.0](https://github.com/kandinskylab/kandinsky-5): A family of diffusion models for video & image generation, where their NABLA attention includes a Sliding Tile Attention branch. [![Star](https://img.shields.io/github/stars/kandinskylab/kandinsky-5.svg?style=social&label=Star)](https://github.com/kandinskylab/kandinsky-5)
+- [LongCat Video](https://github.com/meituan-longcat/LongCat-Video): A foundational video generation model with 13.6B parameters with block-sparse attention similar to Video Sparse Attention. [![Star](https://img.shields.io/github/stars/meituan-longcat/LongCat-Video.svg?style=social&label=Star)](https://github.com/meituan-longcat/LongCat-Video)
 
 ## 🤝 Contributing
 
-We welcome all contributions. Please check out our guide [here](https://hao-ai-lab.github.io/FastVideo/contributing/overview/)
-
+We welcome all contributions. Please check out our guide [here](https://hao-ai-lab.github.io/FastVideo/contributing/overview/).
+See details in [development roadmap](https://github.com/hao-ai-lab/FastVideo/issues/899).
 ## Acknowledgement
 We learned and reused code from the following projects:
 - [Wan-Video](https://github.com/Wan-Video)

benchmarks/fvd/README.md

@@ -0,0 +1,106 @@
+# FVD (Fréchet Video Distance) Benchmark
+
+Evaluate generated video quality using FVD with the I3D feature extractor.
+
+## Quick Start
+
+**Run the benchmark:**
+
+```bash
+bash benchmarks/scripts/run.sh
+```
+
+That's it! The script auto-installs dependencies and runs the benchmark.
+
+**To customize:** Edit `benchmarks/fvd/run_fvd.py` to change:
+- Video paths (`real_dir`, `gen_dir`)
+- Number of videos, frames, sampling strategy
+- Device, batch size, caching, etc.
+
+## Advanced Usage (CLI)
+
+For more control without editing Python files, use the CLI.
+
+**First-time setup** (one-time per pod/environment):
+
+```bash
+bash benchmarks/scripts/setup_fvd.sh
+```
+
+Then run any configuration you want:
+
+```bash
+# Custom configuration
+python -m benchmarks.fvd.cli \
+    --real-path data/real/ \
+    --gen-path outputs/gen/ \
+    --num-videos 1024 \
+    --num-frames 32 \
+    --clip-strategy random \
+    --batch-size 32 \
+    --seed 42 \
+    --extractor clip
+```
+
+**Standard protocols:**
+
+```bash
+# Use predefined protocols
+python -m benchmarks.fvd.cli \
+    --real-path data/real/ \
+    --gen-path outputs/gen/ \
+    --protocol fvd2048_16f  # or fvd2048_128f, quick_test, etc.
+```
+
+This would use i3d model by default as the feature extractor
+
+**Feature caching** (speed up repeated evaluations):
+
+```bash
+python -m benchmarks.fvd.cli \
+    --real-path data/real/ \
+    --gen-path outputs/gen/ \
+    --protocol fvd2048_16f \
+    --cache-real-features fvd-cache/extractor_name  # Directory path (will save/load fvd-cache/extractor_name/extractor-name_real_features.pkl)
+```
+
+Run `python -m benchmarks.fvd.cli --help` for all options.
+
+## Available Protocols
+
+- `fvd2048_16f` - Standard (2048 videos, 16 frames)
+- `fvd2048_128f` - Long videos (128 frames)
+- `fvd2048_128f_subsample8` - Subsampled long videos
+- `quick_test` - Fast testing (10 videos)
+
+## Configuration Options
+
+Key options in `FVDConfig`:
+
+```python
+num_videos=2048,              # Videos to evaluate
+num_frames_per_clip=16,       # Frames per clip
+clip_strategy='beginning',    # beginning|random|uniform|middle|sliding
+frame_stride=1,               # Frame subsampling
+batch_size=32,                # GPU batch size
+device='cuda',                # cuda|cpu
+cache_real_features=None,     # Cache path for speed
+seed=42,                      # Reproducibility
+extractor='i3d',              # i3d|clip|videomae
+```
+
+## Programmatic Usage
+
+```python
+from benchmarks.fvd import compute_fvd_with_config, FVDConfig
+
+config = FVDConfig.fvd2048_16f()  # or custom config
+results = compute_fvd_with_config('data/real/', 'outputs/gen/', config)
+print(f"FVD: {results['fvd']:.2f}")
+```
+
+## Notes
+
+- Requires minimum 10 frames per clip
+- Supports both video files (.mp4, .avi, etc.) and frame directories
+- `--cache-real-features` expects a **directory path** (e.g., `cache/real`), it will automatically create/load `real_features.pkl` inside that directory

benchmarks/fvd/__init__.py

@@ -0,0 +1,38 @@
+"""
+FastVideo Frechet Video Distance (FVD) Benchmark Module.
+    >>> from fastvideo.benchmarks.fvd import compute_fvd_with_config, FVDConfig
+    >>> config = FVDConfig.fvd2048_16f()  # Standard protocol
+    >>> results = compute_fvd_with_config('data/real/', 'outputs/gen/', config)
+    >>> print(f"FVD: {results['fvd']:.2f}")
+"""
+
+from .fvd import (
+    compute_fvd,
+    compute_fvd_with_config,
+    compute_frechet_distance,
+    compute_statistics,
+    FVDConfig,
+)
+from .feature_extractors import (BaseFeatureExtractor, I3DFeatureExtractor,
+                                 load_extractor)
+from .video_utils import (
+    load_video_auto,
+    sample_clips_from_video,
+    load_video_clips_streaming,
+    ClipSamplingStrategy,
+)
+
+__all__ = [
+    'compute_fvd',
+    'compute_fvd_with_config',
+    'compute_frechet_distance',
+    'compute_statistics',
+    'FVDConfig',
+    'BaseFeatureExtractor',
+    'I3DFeatureExtractor',
+    'load_extractor',
+    'load_video_auto',
+    'sample_clips_from_video',
+    'load_video_clips_streaming',
+    'ClipSamplingStrategy',
+]

benchmarks/fvd/cli.py

@@ -0,0 +1,107 @@
+import argparse
+import sys
+import traceback
+from .fvd import compute_fvd_with_config, FVDConfig
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description='Compute Fréchet Video Distance (FVD)')
+
+    # Required arguments
+    parser.add_argument('--real-path',
+                        type=str,
+                        required=True,
+                        help='Path to real videos')
+    parser.add_argument('--gen-path',
+                        type=str,
+                        required=True,
+                        help='Path to generated videos')
+
+    # Extractor selection
+    parser.add_argument('--extractor',
+                        type=str,
+                        default='i3d',
+                        choices=['i3d', 'clip', 'videomae'],
+                        help='Feature extractor model to use (default: i3d)')
+
+    # Standard args
+    parser.add_argument('--seed',
+                        type=int,
+                        default=None,
+                        help='Random seed for reproducibility')
+    parser.add_argument('--protocol',
+                        type=str,
+                        default=None,
+                        choices=['fvd2048_16f', 'fvd2048_128f', 'quick_test'],
+                        help='Use standard protocol (overrides other settings)')
+    parser.add_argument('--num-videos',
+                        type=int,
+                        default=2048,
+                        help='Number of videos to use')
+    parser.add_argument('--num-frames',
+                        type=int,
+                        default=16,
+                        help='Number of frames per clip')
+    parser.add_argument('--clip-strategy',
+                        type=str,
+                        default='beginning',
+                        help='Clip sampling strategy')
+    parser.add_argument('--batch-size',
+                        type=int,
+                        default=32,
+                        help='Batch size for feature extraction')
+    parser.add_argument('--device',
+                        type=str,
+                        default='cuda',
+                        help='Device to use (cuda or cpu)')
+    parser.add_argument('--cache-real-features',
+                        type=str,
+                        default=None,
+                        help='Path to cache real video features')
+    parser.add_argument('--quiet',
+                        action='store_true',
+                        help='Suppress progress output')
+
+    args = parser.parse_args()
+
+    # Create config
+    if args.protocol:
+        protocol_map = {
+            'fvd2048_16f': FVDConfig.fvd2048_16f,
+            'fvd2048_128f': FVDConfig.fvd2048_128f,
+            'quick_test': FVDConfig.quick_test,
+        }
+        config = protocol_map[args.protocol]()
+        # Apply overrides
+        config.device = args.device
+        config.cache_real_features = args.cache_real_features
+        config.extractor_model = args.extractor  # Apply extractor arg
+    else:
+        config = FVDConfig(
+            num_videos=args.num_videos,
+            num_frames_per_clip=args.num_frames,
+            extractor_model=args.extractor,  # Apply extractor arg
+            clip_strategy=args.clip_strategy,
+            batch_size=args.batch_size,
+            device=args.device,
+            cache_real_features=args.cache_real_features,
+            seed=args.seed)
+
+    try:
+        _ = compute_fvd_with_config(
+            args.real_path,  # noqa: F841
+            args.gen_path,
+            config,
+            verbose=not args.quiet)
+
+        return 0
+
+    except Exception as e:
+        print(f"Error: {e}", file=sys.stderr)
+        traceback.print_exc(file=sys.stderr)
+        return 1
+
+
+if __name__ == '__main__':
+    sys.exit(main())

benchmarks/fvd/feature_extractors.py

@@ -0,0 +1,264 @@
+"""
+Pluggable Feature Extractors for FVD Computation.
+Supports I3D (standard), CLIP, and VideoMAE via a common interface.
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from abc import ABC, abstractmethod
+from huggingface_hub import hf_hub_download
+from tqdm import tqdm
+
+try:
+    from transformers import CLIPModel, CLIPProcessor, VideoMAEModel
+    TRANSFORMERS_AVAILABLE = True
+except ImportError:
+    TRANSFORMERS_AVAILABLE = False
+
+
+class BaseFeatureExtractor(ABC, nn.Module):
+    """Abstract base class for all video feature extractors."""
+
+    def __init__(self, device: str = 'cuda'):
+        super().__init__()
+        self.device = torch.device(
+            device if torch.cuda.is_available() else 'cpu')
+
+    @property
+    @abstractmethod
+    def feature_dim(self) -> int:
+        """Dimension of the output feature vector."""
+        pass
+
+    @abstractmethod
+    def preprocess(self, videos: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            videos: [B, T, C, H, W] in [0, 255] range.
+        Returns:
+            Preprocessed tensor ready for the model.
+        """
+        pass
+
+    @abstractmethod
+    def extract_features_batch(self, videos: torch.Tensor) -> torch.Tensor:
+        """
+        Extract features for a single batch.
+        Args:
+            videos: [B, T, C, H, W] (raw input)
+        Returns:
+            Features: [B, feature_dim]
+        """
+        pass
+
+    @torch.no_grad()
+    def extract_features(self,
+                         videos: torch.Tensor,
+                         batch_size: int = 32,
+                         verbose: bool = True) -> torch.Tensor:
+        """
+        Extract features for a large tensor of videos by batching.
+        """
+        N = len(videos)
+        all_features = []
+
+        iterator = range(0, N, batch_size)
+        if verbose:
+            iterator = tqdm(
+                iterator,
+                desc=f"Extracting features ({self.__class__.__name__})")
+
+        for i in iterator:
+            batch = videos[i:i + batch_size].to(self.device)
+            features = self.extract_features_batch(batch)
+            all_features.append(features.cpu())
+
+        return torch.cat(all_features, dim=0)
+
+
+# 1. I3D Extractor (The Standard FVD Metric)
+class I3DFeatureExtractor(BaseFeatureExtractor):
+    REPO_ID = 'flateon/FVD-I3D-torchscript'
+    MODEL_FILENAME = 'i3d_torchscript.pt'
+
+    def __init__(self, device: str = 'cuda', cache_dir: str | None = None):
+        super().__init__(device)
+        self.cache_dir = cache_dir
+        self.model = self._load_model()
+        self.model.eval()
+        self.model.to(self.device)
+
+    @property
+    def feature_dim(self) -> int:
+        return 400
+
+    def _load_model(self) -> torch.nn.Module:
+        try:
+            model_path = hf_hub_download(repo_id=self.REPO_ID,
+                                         filename=self.MODEL_FILENAME,
+                                         cache_dir=self.cache_dir)
+            return torch.jit.load(model_path, map_location=self.device)
+        except Exception as e:
+            raise RuntimeError(f"Failed to load I3D model: {e}") from e
+
+    def preprocess(self, videos: torch.Tensor) -> torch.Tensor:
+        """Standard I3D preprocessing: Resize to 224, Norm to [-1, 1]."""
+        B, T, C, H, W = videos.shape
+
+        if T < 10:
+            raise ValueError(f"I3D requires at least 10 frames, got {T}")
+
+        # Normalize to [0, 1]
+        if videos.max() > 1.0:
+            videos = videos / 255.0
+
+        # Scale to [-1, 1]
+        videos = videos * 2.0 - 1.0
+
+        # Resize to 224x224
+        if H != 224 or W != 224:
+            videos = videos.reshape(B * T, C, H, W)
+            videos = F.interpolate(videos,
+                                   size=(224, 224),
+                                   mode='bilinear',
+                                   align_corners=False)
+            videos = videos.reshape(B, T, C, 224, 224)
+
+        # [B, T, C, H, W] -> [B, C, T, H, W]
+        return videos.permute(0, 2, 1, 3, 4).contiguous()
+
+    def extract_features_batch(self, videos: torch.Tensor) -> torch.Tensor:
+        batch = self.preprocess(videos)
+        # TorchScript I3D returns raw logits when return_features=True
+        return self.model(batch,
+                          rescale=False,
+                          resize=False,
+                          return_features=True)
+
+
+# 2. CLIP Extractor (Semantic/Content Quality)
+class CLIPFeatureExtractor(BaseFeatureExtractor):
+
+    def __init__(self,
+                 device: str = 'cuda',
+                 model_name: str = "openai/clip-vit-base-patch32"):
+        if not TRANSFORMERS_AVAILABLE:
+            raise ImportError(
+                "Please install transformers: pip install transformers")
+        super().__init__(device)
+        self.processor = CLIPProcessor.from_pretrained(model_name)
+        self.model = CLIPModel.from_pretrained(model_name).to(self.device)
+        self.model.eval()
+        self._feature_dim = self.model.config.projection_dim
+
+    @property
+    def feature_dim(self) -> int:
+        return self._feature_dim
+
+    def preprocess(self, videos: torch.Tensor) -> torch.Tensor:
+        # Ensure values are [0, 255]
+        if videos.max() <= 1.0:
+            videos = videos * 255.0
+
+        return videos.to(torch.uint8)
+
+    def extract_features_batch(self, videos: torch.Tensor) -> torch.Tensor:
+        # Input: [B, T, C, H, W]
+        B, T, C, H, W = videos.shape
+        videos = self.preprocess(videos)
+
+        # Flatten B*T to treat frames as images
+        images = videos.view(B * T, C, H, W)
+
+        # HF Processor
+        inputs = self.processor(images=images,
+                                return_tensors="pt",
+                                padding=True)
+        inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+        # Extract features [B*T, Dim]
+        outputs = self.model.get_image_features(**inputs)
+
+        # Reshape [B, T, Dim] and Average Pooling over time
+        outputs = outputs.view(B, T, -1)
+        return outputs.mean(dim=1)
+
+
+# 3. VideoMAE Extractor (Structure/Motion Quality)
+class VideoMAEFeatureExtractor(BaseFeatureExtractor):
+
+    def __init__(self,
+                 device: str = 'cuda',
+                 model_name: str = "MCG-NJU/videomae-base"):
+        if not TRANSFORMERS_AVAILABLE:
+            raise ImportError(
+                "Please install transformers: pip install transformers")
+        super().__init__(device)
+        self.model = VideoMAEModel.from_pretrained(model_name).to(self.device)
+        self.model.eval()
+
+        self.register_buffer(
+            'mean',
+            torch.tensor([0.485, 0.456, 0.406],
+                         device=self.device).view(1, 1, 3, 1, 1))
+        self.register_buffer(
+            'std',
+            torch.tensor([0.229, 0.224, 0.225],
+                         device=self.device).view(1, 1, 3, 1, 1))
+
+    @property
+    def feature_dim(self) -> int:
+        return self.model.config.hidden_size
+
+    def preprocess(self, videos: torch.Tensor) -> torch.Tensor:
+        """
+        Efficient GPU-based preprocessing.
+        Input: [B, T, C, H, W] in range [0, 255]
+        """
+        B, T, C, H, W = videos.shape
+
+        # 1. Resize to 224x224
+        if H != 224 or W != 224:
+            videos = videos.view(B * T, C, H, W)
+            videos = F.interpolate(videos,
+                                   size=(224, 224),
+                                   mode='bilinear',
+                                   align_corners=False)
+            videos = videos.view(B, T, C, 224, 224)
+
+        # 2. Normalize to [0, 1]
+        if videos.dtype != torch.float32:
+            videos = videos.float()
+
+        if videos.max() > 1.0:
+            videos = videos / 255.0
+
+        # 3. Apply ImageNet Mean/Std
+        return (videos - self.mean) / self.std
+
+    def extract_features_batch(self, videos: torch.Tensor) -> torch.Tensor:
+        # Input: [B, T, C, H, W]
+
+        # Fast GPU Preprocessing
+        pixel_values = self.preprocess(videos)
+
+        # Forward pass
+        outputs = self.model(pixel_values)
+
+        # Global Average Pooling of last hidden state [B, T_patches, 768] -> [B, 768]
+        return outputs.last_hidden_state.mean(dim=1)
+
+
+# Factory
+def load_extractor(name: str, device: str = 'cuda') -> BaseFeatureExtractor:
+    name = name.lower()
+    if name == 'i3d':
+        return I3DFeatureExtractor(device)
+    elif name == 'clip':
+        return CLIPFeatureExtractor(device)
+    elif name == 'videomae':
+        return VideoMAEFeatureExtractor(device)
+    else:
+        raise ValueError(
+            f"Unknown extractor: {name}. Options: i3d, clip, videomae")

benchmarks/fvd/fvd.py

@@ -0,0 +1,405 @@
+import numpy as np
+import scipy.linalg
+import torch
+from pathlib import Path
+from collections.abc import Iterator
+import pickle
+from dataclasses import dataclass, field
+from .feature_extractors import BaseFeatureExtractor, load_extractor
+from .video_utils import ClipSamplingStrategy, load_video_clips_streaming
+
+
+def compute_statistics(features: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+    """Compute mean and covariance."""
+    mu = np.mean(features, axis=0)
+    sigma = np.cov(features, rowvar=False)
+    return mu, sigma
+
+
+def compute_frechet_distance(mu1: np.ndarray,
+                             sigma1: np.ndarray,
+                             mu2: np.ndarray,
+                             sigma2: np.ndarray,
+                             eps: float = 1e-6) -> float:
+    """
+    Compute Fréchet distance between two Gaussians.
+    """
+    sigma1 = sigma1 + eps * np.eye(sigma1.shape[0])
+    sigma2 = sigma2 + eps * np.eye(sigma2.shape[0])
+
+    diff = mu1 - mu2
+    mean_distance = np.sum(diff**2)
+
+    trace_sum = np.trace(sigma1 + sigma2)
+
+    covmean = scipy.linalg.sqrtm(sigma1 @ sigma2)
+
+    if np.iscomplexobj(covmean):
+        if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3):
+            print(
+                f"Warning: Imaginary component: {np.max(np.abs(covmean.imag))}")
+        covmean = covmean.real
+
+    trace_product = np.trace(covmean)
+
+    fvd = mean_distance + trace_sum - 2 * trace_product
+
+    return float(fvd)
+
+
+@dataclass
+class FVDConfig:
+    # default configuration for FVD computation:
+
+    # Video selection
+    num_videos: int = 2048
+
+    # Feature Extractor Selection
+    extractor_model: str = 'i3d'  # Options: 'i3d', 'clip', 'videomae'
+
+    # Clip sampling
+    num_frames_per_clip: int = 16
+    num_clips_per_video: int = 1
+    clip_strategy: str | ClipSamplingStrategy = 'beginning'
+
+    # Temporal subsampling
+    frame_stride: int = 1  # 1=no subsampling, 2=every 2nd, 8=every 8th
+    temporal_stride: int = 1  # For sliding window clips
+
+    # Data processing
+    video_extensions: list[str] = field(
+        default_factory=lambda: ['.mp4', '.avi', '.mov', '.mkv'])
+    support_frame_dirs: bool = True
+
+    # Computation
+    batch_size: int = 32
+    device: str = 'cuda'
+
+    use_streaming: bool = True
+    resize_before_extraction: bool = True
+
+    # Caching
+    cache_real_features: str | None = None
+    i3d_model_path: str | None = None
+
+    # Reproducibility
+    seed: int | None = None
+
+    @classmethod
+    def fvd2048_16f(cls) -> 'FVDConfig':
+        """Standard FVD protocol: 2048 videos, 16 frames, beginning clip."""
+        return cls(num_videos=2048,
+                   num_frames_per_clip=16,
+                   clip_strategy='beginning',
+                   use_streaming=True)
+
+    @classmethod
+    def fvd2048_128f(cls) -> 'FVDConfig':
+        """Long video protocol: 2048 videos, 128 frames."""
+        return cls(num_videos=2048,
+                   num_frames_per_clip=128,
+                   clip_strategy='beginning',
+                   use_streaming=True)
+
+    @classmethod
+    def quick_test(cls) -> 'FVDConfig':
+        """Quick test config: 100 videos, 16 frames."""
+        return cls(num_videos=100,
+                   num_frames_per_clip=16,
+                   clip_strategy='beginning')
+
+    def to_dict(self) -> dict:
+        """Export config to dict for logging"""
+        d = self.__dict__.copy()
+        d['clip_strategy'] = str(self.clip_strategy)
+        return d
+
+    def __str__(self) -> str:
+        """Human-readable protocol name"""
+        desc = f"FVD_{self.extractor_model.upper()}_{self.num_videos}_{self.num_frames_per_clip}f"
+        if self.frame_stride > 1:
+            desc += f"_subsample{self.frame_stride}"
+        if self.num_clips_per_video > 1:
+            desc += f"_{self.num_clips_per_video}clips"
+        if self.clip_strategy != 'beginning':
+            desc += f"_{self.clip_strategy}"
+        return desc
+
+
+def extract_features_streaming(video_generator: Iterator[torch.Tensor],
+                               extractor: BaseFeatureExtractor,
+                               batch_size: int = 32,
+                               max_clips: int | None = None,
+                               verbose: bool = True) -> np.ndarray:
+    """
+    Extract features from a video clip generator using streaming.
+    """
+    all_features = []
+    batch = []
+
+    if verbose:
+        print(f"Extracting features with batch_size={batch_size}...")
+
+    with torch.no_grad():
+        for clip_count, clip in enumerate(video_generator):
+            batch.append(clip)
+
+            # Process batch when full
+            if len(batch) == batch_size:
+                batch_tensor = torch.stack(batch).to(extractor.device)
+                features = extractor.extract_features_batch(batch_tensor)
+
+                all_features.append(features.detach().cpu().numpy())
+                batch = []
+
+                if verbose and clip_count % (batch_size * 10) == 0:
+                    print(f"Processed {clip_count} clips...")
+
+            if max_clips is not None and clip_count >= max_clips:
+                break
+
+        # Process remaining clips
+        if len(batch) > 0:
+            batch_tensor = torch.stack(batch).to(extractor.device)
+            features = extractor.extract_features_batch(batch_tensor)
+            all_features.append(features.detach().cpu().numpy())
+
+    if len(all_features) == 0:
+        raise RuntimeError("No features extracted - check video loading")
+
+    features = np.concatenate(all_features, axis=0)
+
+    if verbose:
+        print(f"Extracted {len(features)} feature vectors")
+
+    return features
+
+
+def load_or_compute_features(videos: str | Path | torch.Tensor,
+                             extractor: BaseFeatureExtractor,
+                             config: FVDConfig,
+                             cache_path: str | None = None,
+                             cache_name: str = "real_features") -> np.ndarray:
+    """Load features from cache or compute (with streaming support)"""
+
+    if cache_path is not None:
+        script_dir = Path(__file__).parent
+        cache_dir = script_dir / cache_path
+        cache_file = cache_dir / f"{config.extractor_model}_{cache_name}.pkl"
+
+        if cache_file.exists():
+            print(f"Loading cached features from {cache_file}")
+            with open(cache_file, 'rb') as f:
+                features = pickle.load(f)
+
+            # Validate and limit based on config
+            max_features = config.num_videos * config.num_clips_per_video
+
+            if len(features) < max_features:
+                print(
+                    f"WARNING: Cache has {len(features)} features but need {max_features}"
+                )
+                print("Cached features insufficient - will recompute...")
+            elif len(features) > max_features:
+                print(
+                    f"Using {max_features} features from cache (truncated from {len(features)})"
+                )
+                features = features[:max_features]
+                return features
+            else:
+                print(f"Using all {len(features)} cached features")
+                return features
+
+    print("Computing features from scratch...")
+
+    if isinstance(videos, (str | Path)):
+        target_size = (224, 224) if config.resize_before_extraction else None
+
+        video_generator = load_video_clips_streaming(
+            videos,
+            num_frames=config.num_frames_per_clip,
+            max_videos=config.num_videos,
+            clip_strategy=config.clip_strategy,
+            frame_stride=config.frame_stride,
+            num_clips_per_video=config.num_clips_per_video,
+            video_extensions=config.video_extensions,
+            support_frame_dirs=config.support_frame_dirs,
+            target_size=target_size,
+            verbose=True)
+
+        max_clips = config.num_videos * config.num_clips_per_video
+        features = extract_features_streaming(video_generator,
+                                              extractor,
+                                              batch_size=config.batch_size,
+                                              max_clips=max_clips,
+                                              verbose=True)
+    else:
+        print(f"Extracting features from {len(videos)} video tensors...")
+        features = extractor.extract_features(videos,
+                                              batch_size=config.batch_size,
+                                              verbose=True)
+        features = features.numpy()
+
+    # Validate feature count
+    expected_count = config.num_videos * config.num_clips_per_video
+    if len(features) < expected_count:
+        raise ValueError(
+            f"ERROR: Only extracted {len(features)} features, but need {expected_count}!\n"
+            f"Found fewer videos than expected. Check your video directory.")
+    elif len(features) > expected_count:
+        print(f"Truncating {len(features)} features to {expected_count}")
+        features = features[:expected_count]
+
+    # Cache features if requested
+    if cache_path is not None:
+        script_dir = Path(__file__).parent
+        cache_dir = script_dir / cache_path
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        cache_file = cache_dir / f"{config.extractor_model}_{cache_name}.pkl"
+        print(f"Caching features to {cache_file}")
+        with open(cache_file, 'wb') as f:
+            pickle.dump(features, f)
+
+    return features
+
+
+def compute_fvd_with_config(real_videos: str | Path | torch.Tensor,
+                            gen_videos: str | Path | torch.Tensor,
+                            config: FVDConfig,
+                            verbose: bool = True) -> dict:
+    """
+        Compute FVD using a standardized configuration.
+
+        This is the recommended way to compute FVD for reproducibility.
+
+        Args:
+            real_videos: Path or tensors
+            gen_videos: Path or tensors
+            config: FVDConfig specifying protocol
+            verbose: Print progress
+
+        Returns:
+            results: Dictionary with:
+                - 'fvd': FVD score (float)
+                - 'protocol': Protocol name (str)
+                - 'model': Feature extractor model name (str)
+                - 'config': Configuration dict
+
+        Example:
+            >>> config = FVDConfig.fvd2048_16f()
+            >>> results = compute_fvd_with_config('data/real/', 'outputs/gen/', config)
+            >>> print(f"FVD: {results['fvd']:.2f}")
+    """
+
+    # Seed for reproducibility
+    if config.seed is not None:
+        import random as _rnd
+        _rnd.seed(config.seed)
+        np.random.seed(config.seed)
+        torch.manual_seed(config.seed)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(config.seed)
+
+    if verbose:
+        print("=" * 70)
+        print(f"Computing FVD with protocol: {config}")
+        print(f"Model: {config.extractor_model.upper()}")
+        print("=" * 70)
+        print("\nConfiguration:")
+        for key, value in config.to_dict().items():
+            print(f"  {key}: {value}")
+        print()
+
+    # Initialize Extractor using Factory
+    if verbose:
+        print(
+            f"\nInitializing {config.extractor_model.upper()} model on {config.device}..."
+        )
+
+    extractor = load_extractor(config.extractor_model, device=config.device)
+
+    # Extract features
+    if verbose:
+        print(f"\n{'='*70}")
+        print("Extracting REAL video features...")
+        print(f"{'='*70}")
+
+    real_features = load_or_compute_features(
+        videos=real_videos,
+        extractor=extractor,
+        config=config,
+        cache_path=config.cache_real_features,
+        cache_name="real_features")
+
+    if verbose:
+        print(f"\n{'='*70}")
+        print("Extracting GENERATED video features...")
+        print(f"{'='*70}")
+
+    gen_features = load_or_compute_features(videos=gen_videos,
+                                            extractor=extractor,
+                                            config=config,
+                                            cache_path=None,
+                                            cache_name="gen_features")
+
+    if verbose:
+        print(f"\nReal videos/clips: {len(real_features)}")
+        print(f"Generated videos/clips: {len(gen_features)}")
+        print(f"\n{'='*70}")
+        print("Computing statistics...")
+        print(f"{'='*70}")
+
+    mu_real, sigma_real = compute_statistics(real_features)
+    mu_gen, sigma_gen = compute_statistics(gen_features)
+
+    if verbose:
+        print(f"\n{'='*70}")
+        print("Computing Fréchet distance...")
+        print(f"{'='*70}")
+
+    fvd = compute_frechet_distance(mu_real, sigma_real, mu_gen, sigma_gen)
+
+    if verbose:
+        print(f"\n{'='*70}")
+        print(f"FVD Score ({config.extractor_model.upper()}): {fvd:.4f}")
+        print(f"Protocol: {config}")
+        print(f"{'='*70}\n")
+
+    results = {
+        'fvd': fvd,
+        'protocol': str(config),
+        'model': config.extractor_model,
+        'config': config.to_dict(),
+    }
+
+    return results
+
+
+def compute_fvd(real_videos: str | Path | torch.Tensor,
+                gen_videos: str | Path | torch.Tensor,
+                num_frames: int = 16,
+                batch_size: int = 32,
+                device: str = 'cuda',
+                num_videos: int | None = 2048,
+                cache_real_features: str | None = None,
+                i3d_model_path: str | None = None,
+                seed: int | None = None,
+                verbose: bool = True) -> float:
+    """
+    Backward compatibility wrapper for computing FVD (defaults to I3D).
+    """
+    num_videos = num_videos if num_videos is not None else 2048
+
+    config = FVDConfig(
+        num_videos=num_videos,
+        num_frames_per_clip=num_frames,
+        extractor_model='i3d',  # Default to I3D
+        batch_size=batch_size,
+        device=device,
+        cache_real_features=cache_real_features,
+        i3d_model_path=i3d_model_path,
+        seed=seed,
+    )
+
+    result = compute_fvd_with_config(real_videos, gen_videos, config, verbose)
+    return result['fvd']

benchmarks/fvd/i3d_model.py

@@ -0,0 +1,142 @@
+"""I3D Feature Extractor for FVD Computation"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from pathlib import Path
+from huggingface_hub import hf_hub_download
+from tqdm import tqdm
+from contextlib import suppress
+
+
+class I3DFeatureExtractor(nn.Module):
+    """
+    I3D feature extractor for FVD computation.
+    Extracts 400-dimensional features from videos using I3D model
+    trained on Kinetics-400.
+    """
+
+    REPO_ID = 'flateon/FVD-I3D-torchscript'
+    MODEL_FILENAME = 'i3d_torchscript.pt'
+
+    def __init__(self,
+                 device: str = 'cuda',
+                 cache_dir: str | Path | None = None):
+        super().__init__()
+
+        self.device_str = device
+        if device == 'cuda' and not torch.cuda.is_available():
+            print(
+                "Warning: CUDA requested but not available – falling back to CPU"
+            )
+            self.device = torch.device('cpu')
+        else:
+            self.device = torch.device(device)
+
+        self.cache_dir: str | None
+        if cache_dir is not None:
+            self.cache_dir = str(Path(cache_dir).resolve())
+        else:
+            self.cache_dir = None  # Use HF default cache
+
+        self.model = self._load_model()
+        self.model.eval()
+
+        with suppress(Exception):
+            self.model.to(self.device)
+
+    def _load_model(self) -> torch.nn.Module:
+        """Download and load I3D TorchScript model from Hugging Face Hub."""
+        print(f"Loading I3D model from Hugging Face Hub ({self.REPO_ID})...")
+
+        try:
+            # Download model from Hugging Face Hub
+            model_path = hf_hub_download(repo_id=self.REPO_ID,
+                                         filename=self.MODEL_FILENAME,
+                                         cache_dir=self.cache_dir)
+
+            # Load directly to chosen device
+            model = torch.jit.load(model_path, map_location=self.device)
+            print("I3D model loaded successfully")
+            return model
+
+        except Exception as e:
+            raise RuntimeError(
+                f"Failed to load I3D model from Hugging Face Hub. Error: {e}\n"
+                f"Ensure you have internet connection and huggingface_hub installed:\n"
+                f"pip install huggingface_hub") from e
+
+    def preprocess(self, videos: torch.Tensor) -> torch.Tensor:
+        """
+        Preprocess videos for I3D.
+        
+        Args:
+            videos: [B, T, C, H, W], values in [0, 255]
+        
+        Returns:
+            Preprocessed videos [B, C, T, 224, 224] (normalized and resized)
+        """
+        B, T, C, H, W = videos.shape
+
+        if T < 10:
+            raise ValueError(f"I3D requires at least 10 frames, got {T}")
+
+        # Normalize to [0, 1] if needed
+        if videos.max() > 1.0:
+            videos = videos / 255.0
+
+        # Resize to 224x224 if needed
+        if H != 224 or W != 224:
+            videos = videos.reshape(B * T, C, H, W)
+            videos = F.interpolate(videos,
+                                   size=(224, 224),
+                                   mode='bilinear',
+                                   align_corners=False)
+            videos = videos.reshape(B, T, C, 224, 224)
+
+        # Convert to [B, C, T, H, W] format
+        videos = videos.permute(0, 2, 1, 3, 4).contiguous()
+
+        return videos
+
+    @torch.no_grad()
+    def extract_features(self,
+                         videos: torch.Tensor,
+                         batch_size: int = 32,
+                         verbose: bool = True) -> torch.Tensor:
+        """
+        Extract I3D features
+        
+        Args:
+            videos: [N, T, C, H, W], values in [0, 255]
+            batch_size: Batch size for processing
+            verbose: Show progress bar
+        
+        Returns:
+            Features [N, 400]
+        """
+        N = len(videos)
+        all_features = []
+
+        iterator = range(0, N, batch_size)
+        if verbose:
+            iterator = tqdm(iterator, desc="Extracting I3D features")
+
+        for i in iterator:
+            batch = videos[i:i + batch_size].to(self.device)
+            batch = self.preprocess(batch)  # Now returns [B, C, T, H, W]
+
+            # Use the HF model without rescale/resize (we handle it in preprocess)
+            features = self.model(batch,
+                                  rescale=False,
+                                  resize=False,
+                                  return_features=True)
+
+            all_features.append(features.cpu())
+
+        return torch.cat(all_features, dim=0)
+
+    def __call__(self,
+                 videos: torch.Tensor,
+                 batch_size: int = 32) -> torch.Tensor:
+        return self.extract_features(videos, batch_size=batch_size)

benchmarks/fvd/run_fvd.py

@@ -0,0 +1,54 @@
+import sys
+from pathlib import Path
+
+root_dir = Path(__file__).parent.parent.parent
+sys.path.insert(0, str(root_dir))
+
+from benchmarks.fvd.fvd import FVDConfig, compute_fvd_with_config  # noqa: E402
+
+
+def main() -> None:
+    script_dir = Path(__file__).parent.resolve()
+
+    # Define directories
+    real_dir = "benchmarks/data/real_videos"
+    gen_dir = "benchmarks/data/generated_videos"
+
+    # Compare all 3 models
+    models_to_test = ['i3d', 'clip', 'videomae']
+
+    print(f"\n{'='*60}")
+    print("STARTING COMPARISON BENCHMARK")
+    print(f"{'='*60}")
+
+    for model_name in models_to_test:
+        print(f"\n>>> Running evaluation with {model_name.upper()}...")
+
+        try:
+            cfg = FVDConfig(
+                num_videos=650,
+                num_frames_per_clip=16,
+                extractor_model=model_name,
+                clip_strategy='beginning',
+                device='cuda',
+                seed=42,
+                # Use separate cache folders for each model to avoid conflicts
+                cache_real_features=str(script_dir / f'fvd-cache/{model_name}'),
+            )
+
+            results = compute_fvd_with_config(real_dir,
+                                              gen_dir,
+                                              cfg,
+                                              verbose=False)
+            print(f"FVD: {results['fvd']}\nModel: {results['model']}")
+
+        except Exception as e:
+            print(f"{model_name.upper()} Failed: {e}")
+
+    print(f"\n{'='*60}")
+    print("BENCHMARK COMPLETE")
+    print(f"{'='*60}")
+
+
+if __name__ == '__main__':
+    main()

benchmarks/fvd/validate_fvd.py

@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+import sys
+from pathlib import Path
+import shutil
+import random
+from fvd import compute_fvd_with_config, FVDConfig
+
+script_path = Path(__file__).resolve()
+fastvideo_root = script_path.parent.parent.parent
+sys.path.insert(0, str(fastvideo_root))
+
+
+def split_videos(video_dir: Path, n_per_subset: int = 128, seed: int = 42):
+    subset_a = video_dir.parent / 'bair_full_subset_A'
+    subset_b = video_dir.parent / 'bair_full_subset_B'
+
+    if subset_a.exists():
+        shutil.rmtree(subset_a)
+    if subset_b.exists():
+        shutil.rmtree(subset_b)
+
+    subset_a.mkdir(parents=True)
+    subset_b.mkdir(parents=True)
+
+    videos = sorted(video_dir.glob('*.mp4'))
+
+    random.seed(seed)
+    shuffled = list(videos)
+    random.shuffle(shuffled)
+
+    needed = n_per_subset * 2
+    if len(shuffled) > needed:
+        shuffled = shuffled[:needed]
+
+    mid = len(shuffled) // 2
+
+    print(f"\nSplitting {len(shuffled)} BAIR FULL videos:")
+    print(f"  Subset A: {mid} videos")
+    print(f"  Subset B: {len(shuffled) - mid} videos")
+
+    for v in shuffled[:mid]:
+        shutil.copy2(v, subset_a / v.name)
+
+    for v in shuffled[mid:]:
+        shutil.copy2(v, subset_b / v.name)
+
+    return subset_a, subset_b, mid
+
+
+def validate_fvd(subset_a: Path, subset_b: Path, num_videos: int):
+    config = FVDConfig(num_videos=num_videos,
+                       num_frames_per_clip=16,
+                       clip_strategy='beginning',
+                       batch_size=8,
+                       device='cuda',
+                       seed=42)
+
+    print("\n" + "=" * 70)
+    print("TEST 1: Identity Test")
+    print("=" * 70)
+
+    result1 = compute_fvd_with_config(real_videos=str(subset_a),
+                                      gen_videos=str(subset_a),
+                                      config=config,
+                                      verbose=False)
+    fvd_identity = result1['fvd']
+    print(f"\nIdentity FVD: {fvd_identity:.2f}")
+
+    print("\n" + "=" * 70)
+    print("TEST 2: Real vs Real")
+    print("=" * 70)
+
+    result2 = compute_fvd_with_config(real_videos=str(subset_a),
+                                      gen_videos=str(subset_b),
+                                      config=config,
+                                      verbose=False)
+    fvd_real = result2['fvd']
+    print(f"\nReal vs Real FVD: {fvd_real:.2f}")
+
+    print("\n" + "=" * 70)
+    print("RESULTS")
+    print("=" * 70)
+    print(f"Identity:        {fvd_identity:.2f}")
+    print(f"Real vs Real:    {fvd_real:.2f}")
+
+
+def main() -> None:
+    bair_dir = Path('benchmarks/data/bair_full_videos')
+
+    subset_a, subset_b, count = split_videos(bair_dir,
+                                             n_per_subset=128,
+                                             seed=42)
+    validate_fvd(subset_a, subset_b, count)
+
+
+if __name__ == '__main__':
+    main()

benchmarks/fvd/video_utils.py

@@ -0,0 +1,490 @@
+import torch
+import cv2
+import numpy as np
+from pathlib import Path
+from collections.abc import Iterator
+from tqdm import tqdm
+from enum import Enum
+
+
+class ClipSamplingStrategy(Enum):
+    """Clip sampling strategies for FVD evaluation."""
+    BEGINNING = 'beginning'  # Take first N frames (most common)
+    RANDOM = 'random'  # Random N consecutive frames
+    UNIFORM = 'uniform'  # Uniformly spaced frames across video
+    MIDDLE = 'middle'  # Middle N frames
+    SLIDING = 'sliding'  # Multiple sliding windows
+    ALL = 'all'  # All possible clips
+
+
+def _load_video_cv2(video_path: str | Path,
+                    num_frames: int | None = 16,
+                    sample_strategy: str = 'uniform') -> torch.Tensor:
+    """
+    Load video from video file using OpenCV.
+    
+    Args:
+        video_path: Path to video file (MP4, AVI, MOV, MKV)
+        num_frames: Number of frames to extract
+        sample_strategy: 'uniform' or 'random'
+    
+    Returns:
+        video: [T, C, H, W]
+    """
+    video_path = str(video_path)
+    cap = cv2.VideoCapture(video_path)
+
+    if not cap.isOpened():
+        raise RuntimeError(f"Cannot open video: {video_path}")
+
+    frames = []
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+
+    if num_frames is None:
+        # Read all available frames
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frames.append(frame)
+
+        cap.release()
+        if len(frames) == 0:
+            raise RuntimeError(f"Video has 0 frames: {video_path}")
+
+        frames = np.stack(frames)  # [T, H, W, C]
+        frames = torch.from_numpy(frames).permute(0, 3, 1,
+                                                  2).float()  # [T, C, H, W]
+        return frames
+
+    if total_frames == 0:
+        raise RuntimeError(f"Video has 0 frames: {video_path}")
+
+    # Determine frame indices for sampling
+    if total_frames < num_frames:
+        frame_indices = list(range(
+            total_frames)) + [total_frames - 1] * (num_frames - total_frames)
+    elif sample_strategy == 'uniform':
+        frame_indices = np.linspace(0, total_frames - 1, num_frames,
+                                    dtype=int).tolist()
+    elif sample_strategy == 'random':
+        frame_indices = sorted(
+            np.random.choice(total_frames, num_frames, replace=False))
+    else:
+        raise ValueError(f"Unknown sample_strategy: {sample_strategy}")
+
+    # Extract frames
+    for idx in frame_indices:
+        cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
+        ret, frame = cap.read()
+
+        if not ret:
+            if len(frames) > 0:
+                frames.append(frames[-1].copy())
+            else:
+                h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+                frames.append(np.zeros((h, w, 3), dtype=np.uint8))
+            continue
+
+        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        frames.append(frame)
+
+    cap.release()
+
+    frames = np.stack(frames)  # [T, H, W, C]
+    frames = torch.from_numpy(frames).permute(0, 3, 1,
+                                              2).float()  # [T, C, H, W]
+
+    return frames
+
+
+def _load_video_from_frames(
+        frame_dir: str | Path,
+        num_frames: int | None = 16,
+        sample_strategy: str = 'uniform',
+        frame_extensions: list[str] | None = None) -> torch.Tensor:
+    """
+    Load video from directory of frame images.
+    
+    Args:
+        frame_dir: Directory containing frames
+        num_frames: Number of frames to sample
+        sample_strategy: 'uniform' or 'random'
+        frame_extensions: Image file extensions to look for
+    
+    Returns:
+        video: [T, C, H, W]
+    """
+    if frame_extensions is None:
+        frame_extensions = ['.jpg', '.png', '.jpeg', '.bmp']
+
+    frame_dir = Path(frame_dir)
+
+    if not frame_dir.exists():
+        raise FileNotFoundError(f"Frame directory not found: {frame_dir}")
+
+    # Find all frames
+    frame_files: list[Path] = []
+    for ext in frame_extensions:
+        frame_files.extend(frame_dir.glob(f"*{ext}"))
+
+    if len(frame_files) == 0:
+        raise ValueError(
+            f"No frames found in {frame_dir} with extensions {frame_extensions}"
+        )
+
+    frame_files = sorted(frame_files, key=lambda x: x.name)
+    total_frames = len(frame_files)
+
+    # Determine frame indices
+    if num_frames is None:
+        frame_indices = list(range(total_frames))
+    else:
+        if total_frames < num_frames:
+            frame_indices = list(range(total_frames)) + [total_frames - 1] * (
+                num_frames - total_frames)
+        elif sample_strategy == 'uniform':
+            frame_indices = np.linspace(0,
+                                        total_frames - 1,
+                                        num_frames,
+                                        dtype=int).tolist()
+        elif sample_strategy == 'random':
+            frame_indices = sorted(
+                np.random.choice(total_frames, num_frames, replace=False))
+        else:
+            raise ValueError(f"Unknown sample_strategy: {sample_strategy}")
+
+    # Load frames
+    frames = []
+    for idx in frame_indices:
+        frame_path = frame_files[idx]
+        frame = cv2.imread(str(frame_path))
+
+        if frame is None:
+            raise RuntimeError(f"Failed to load frame: {frame_path}")
+
+        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        frames.append(frame)
+
+    # Stack and convert to tensor
+    frames = np.stack(frames)  # [T, H, W, C]
+    frames = torch.from_numpy(frames).permute(0, 3, 1,
+                                              2).float()  # [T, C, H, W]
+
+    return frames
+
+
+def _detect_video_format(path: str | Path) -> str:
+    """
+    Detect if path is a video file or frame directory.
+    
+    Returns:
+        'video_file', 'frame_directory', or 'unknown'
+    """
+    path = Path(path)
+
+    if path.is_file():
+        return 'video_file'
+    elif path.is_dir():
+        # Check if contains image files
+        image_extensions = ['.jpg', '.jpeg', '.png', '.bmp']
+        for ext in image_extensions:
+            if list(path.glob(f"*{ext}")):
+                return 'frame_directory'
+        return 'unknown'
+    else:
+        raise ValueError(f"Path does not exist: {path}")
+
+
+def load_video_auto(video_path: str | Path,
+                    num_frames: int | None = 16,
+                    sample_strategy: str = 'uniform') -> torch.Tensor:
+    """
+    Automatically detect format and load video.
+    
+    Supports:
+    - Video files (MP4, AVI, MOV, MKV)
+    - Frame directories (JPG, PNG)
+    
+    Args:
+        video_path: Path to video file or frame directory
+        num_frames: Number of frames to extract
+        sample_strategy: 'uniform' or 'random'
+    
+    Returns:
+        video: [T, C, H, W]
+    """
+    format_type = _detect_video_format(video_path)
+
+    if format_type == 'video_file':
+        return _load_video_cv2(video_path, num_frames, sample_strategy)
+    elif format_type == 'frame_directory':
+        return _load_video_from_frames(video_path, num_frames, sample_strategy)
+    else:
+        raise ValueError(f"Unknown video format at {video_path}")
+
+
+def sample_clips_from_video(
+        video: torch.Tensor,
+        num_frames_per_clip: int = 16,
+        num_clips: int = 1,
+        strategy: str | ClipSamplingStrategy = ClipSamplingStrategy.BEGINNING,
+        frame_stride: int = 1,
+        temporal_stride: int = 1) -> list[torch.Tensor]:
+    """
+    Sample clips from a video with various strategies.
+        
+    Args:
+        video: [T, C, H, W] full video
+        num_frames_per_clip: Frames per clip
+        num_clips: Number of clips to extract
+        strategy: ClipSamplingStrategy or string ('beginning', 'random', etc.)
+        frame_stride: Skip frames (FPS control: 1=all, 2=every 2nd, 8=every 8th)
+        temporal_stride: Stride between clips for sliding window
+    
+    Returns:
+        List of clips, each [num_frames_per_clip, C, H, W]
+    
+    Examples:
+        >>> # Beginning clip (most common for FVD)
+        >>> clips = sample_clips_from_video(video, 16, strategy='beginning')
+        
+        >>> # Multiple random clips
+        >>> clips = sample_clips_from_video(video, 16, num_clips=4, strategy='random')
+        
+        >>> # Subsample FPS by 2x (every 2nd frame)
+        >>> clips = sample_clips_from_video(video, 16, frame_stride=2)
+        
+        >>> # Sliding window with overlap
+        >>> clips = sample_clips_from_video(video, 16, strategy='sliding', temporal_stride=8)
+    """
+    # Convert string to enum if needed
+    if isinstance(strategy, str):
+        strategy = ClipSamplingStrategy(strategy)
+
+    T, C, H, W = video.shape
+
+    # Apply frame stride (FPS subsampling)
+    if frame_stride > 1:
+        video = video[::frame_stride]
+        T = len(video)
+
+    effective_clip_length = num_frames_per_clip
+
+    # Handle videos shorter than clip length
+    if effective_clip_length > T:
+        pad_length = effective_clip_length - T
+        last_frame = video[-1:].repeat(pad_length, 1, 1, 1)
+        video = torch.cat([video, last_frame], dim=0)
+        T = len(video)
+
+    clips = []
+
+    if strategy == ClipSamplingStrategy.BEGINNING:
+        # Take first clip (most common for FVD evaluation)
+        clip = video[:effective_clip_length]
+        clips.append(clip)
+
+    elif strategy == ClipSamplingStrategy.MIDDLE:
+        # Take middle clip
+        start = (T - effective_clip_length) // 2
+        clip = video[start:start + effective_clip_length]
+        clips.append(clip)
+
+    elif strategy == ClipSamplingStrategy.RANDOM:
+        # Sample N random clips
+        for _ in range(num_clips):
+            if effective_clip_length == T:
+                start = 0
+            else:
+                start = np.random.randint(0, T - effective_clip_length + 1)
+            clip = video[start:start + effective_clip_length]
+            clips.append(clip)
+
+    elif strategy == ClipSamplingStrategy.UNIFORM:
+        # Uniformly spaced clips
+        if num_clips == 1:
+            # Single clip from middle
+            start = (T - effective_clip_length) // 2
+            clip = video[start:start + effective_clip_length]
+            clips.append(clip)
+        else:
+            # Multiple uniformly spaced clips
+            step = (T - effective_clip_length) / (num_clips -
+                                                  1) if num_clips > 1 else 0
+            for i in range(num_clips):
+                start = int(i * step)
+                start = min(start, T - effective_clip_length)
+                clip = video[start:start + effective_clip_length]
+                clips.append(clip)
+
+    elif strategy == ClipSamplingStrategy.SLIDING:
+        # Sliding window with stride
+        for start in range(0, T - effective_clip_length + 1, temporal_stride):
+            clip = video[start:start + effective_clip_length]
+            clips.append(clip)
+            if len(clips) >= num_clips:
+                break
+
+    elif strategy == ClipSamplingStrategy.ALL:
+        # All possible clips (overlapping)
+        for start in range(T - effective_clip_length + 1):
+            clip = video[start:start + effective_clip_length]
+            clips.append(clip)
+
+    else:
+        raise ValueError(f"Unknown strategy: {strategy}")
+
+    return clips
+
+
+def load_video_clips_streaming(directory: str | Path,
+                               num_frames: int = 16,
+                               max_videos: int | None = None,
+                               clip_strategy: str
+                               | ClipSamplingStrategy = 'beginning',
+                               frame_stride: int = 1,
+                               num_clips_per_video: int = 1,
+                               video_extensions: list[str] | None = None,
+                               support_frame_dirs: bool = True,
+                               target_size: tuple[int, int] | None = (224, 224),
+                               verbose: bool = True) -> Iterator[torch.Tensor]:
+    """
+    This generator yields clips one-by-one instead of loading all videos into RAM.
+    Perfect for large datasets where memory is limited.
+    
+    Args:
+        directory: Path to directory with videos
+        num_frames: Frames per clip
+        max_videos: Max videos to load
+        clip_strategy: 'beginning', 'random', 'uniform', etc.
+        frame_stride: Frame skip (1=all, 2=every 2nd, 8=every 8th)
+        num_clips_per_video: Number of clips per video
+        video_extensions: Video file extensions
+        support_frame_dirs: Also load frame directories
+        target_size: Resize clips to (H, W). If None, keep original size.
+        verbose: Show progress
+    
+    Yields:
+        clip: [T, C, H, W] individual clips
+    
+    Example:
+        >>> for clip in load_video_clips_streaming('data/videos/', num_frames=16):
+        >>>     features = model.extract_features(clip.unsqueeze(0))
+        >>>     # Process one clip at a time - low memory usage!
+    """
+    if video_extensions is None:
+        video_extensions = ['.mp4', '.avi', '.mov', '.mkv']
+
+    directory = Path(directory)
+
+    if not directory.exists():
+        raise FileNotFoundError(f"Directory not found: {directory}")
+
+    # Find video paths
+    video_paths: list[Path] = []
+
+    # Find video files
+    for ext in video_extensions:
+        video_paths.extend(directory.glob(f"**/*{ext}"))
+
+    # Find frame directories if enabled
+    if support_frame_dirs:
+        for subdir in directory.iterdir():
+            if subdir.is_dir():
+                # Check if it contains frames
+                image_extensions = ['.jpg', '.jpeg', '.png', '.bmp']
+                for ext in image_extensions:
+                    if list(subdir.glob(f"*{ext}")):
+                        video_paths.append(subdir)
+                        break
+
+    if len(video_paths) == 0:
+        raise ValueError(f"No videos found in {directory}")
+
+    video_paths = sorted(video_paths)
+
+    if max_videos is not None:
+        video_paths = video_paths[:max_videos]
+
+    if verbose:
+        print(f"Found {len(video_paths)} videos in {directory}")
+        if num_clips_per_video > 1:
+            print(f"Extracting {num_clips_per_video} clips per video...")
+        if frame_stride > 1:
+            print(f"Subsampling frames with stride {frame_stride}...")
+        if target_size:
+            print(f"Resizing clips to {target_size}...")
+
+    # Track statistics
+    failed_count = 0
+    total_clips = 0
+
+    iterator = tqdm(video_paths,
+                    desc="Loading videos") if verbose else video_paths
+
+    for video_path in iterator:
+        try:
+            # Load full video
+            video = load_video_auto(video_path,
+                                    num_frames=None,
+                                    sample_strategy='uniform')
+
+            # Sample clips from video
+            clips = sample_clips_from_video(video,
+                                            num_frames_per_clip=num_frames,
+                                            num_clips=num_clips_per_video,
+                                            strategy=clip_strategy,
+                                            frame_stride=frame_stride)
+
+            if target_size is not None:
+                resized_clips = []
+                for clip in clips:
+                    T, C, H, W = clip.shape
+                    if target_size != (H, W):
+                        # Resize to target size
+                        clip = clip.contiguous(
+                        )  # Fix non-contiguous tensors first
+                        clip_flat = clip.view(T * C, H,
+                                              W).unsqueeze(0)  # [1, T*C, H, W]
+                        clip_resized = torch.nn.functional.interpolate(
+                            clip_flat,
+                            size=target_size,
+                            mode='bilinear',
+                            align_corners=False)
+                        clip = clip_resized.squeeze(0).view(
+                            T, C, target_size[0],
+                            target_size[1])  # Back to [T, C, H, W]
+                    resized_clips.append(clip)
+                clips = resized_clips
+
+            # Yield clips one by one
+            for clip in clips:
+                yield clip
+                total_clips += 1
+
+            # Free memory
+            del video, clips
+
+        except Exception as e:
+            failed_count += 1
+            if verbose:
+                print(f"\nWarning: Failed to load {video_path}: {e}")
+            continue
+
+    # Validate
+    if total_clips == 0:
+        raise RuntimeError(f"Failed to load any videos from {directory}")
+
+    failure_rate = failed_count / len(video_paths)
+    if failure_rate > 0.1:  # More than 10% failed
+        print(
+            f"\nWARNING: {failure_rate:.1%} of videos failed to load ({failed_count}/{len(video_paths)})"
+        )
+
+    if verbose:
+        print(
+            f"\nSuccessfully loaded {total_clips} clips from {len(video_paths) - failed_count} videos"
+        )

benchmarks/scripts/run.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+# 1. Install missing dependency
+pip install -q opencv-python-headless transformers huggingface_hub
+
+# 2. Run FVD script
+python benchmarks/fvd/run_fvd.py

benchmarks/scripts/setup_fvd.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+# 1. Install missing dependency
+pip install -q opencv-python-headless

csrc/attn/README.md

@@ -1,113 +0,0 @@
-
-
-# Attention Kernel Used in FastVideo
-
-
-
-## Video Sparse Attention (VSA)
-
-### Installation
-We support H100 (via TK) and any other GPU (via triton) for VSA.
-```bash
-git submodule update --init --recursive
-python setup_vsa.py install
-```
-
-
-If you encounter error during installation, try below:
-Install C++20 for ThunderKittens:
-```bash
-sudo apt update
-sudo apt install gcc-11 g++-11
-
-sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
-
-sudo apt update
-sudo apt install clang-11
-```
-(If you use CUDA12.8)
-```bash
-export CUDA_HOME=/usr/local/cuda-12.8
-export PATH=${CUDA_HOME}/bin:${PATH} 
-export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
-```
-
-### Verify if you have successfully installed
-
-```bash
-# test numerical
-python tests/test_vsa.py
-# (For H100) test speed
-python benchmarks/bench_vsa_hopper.py
-```
-bench_vsa_hopper.py should print something like this:
-```bash
-Using topk=76 kv blocks per q block (out of 768 total kv blocks)
-
-=== BLOCK SPARSE ATTENTION BENCHMARK ===
-Block Sparse Forward  - TFLOPS: 5622.26
-Block Sparse Backward - TFLOPS: 3865.68
-```
-
-
-## Sliding Tile Attention (STA)
-We only support H100 for STA.
-```bash
-git submodule update --init --recursive
-python setup_sta.py install
-```
-
-
-
-
-###  Usage
-End-2-end inference with FastVideo:
-```bash
-bash scripts/inference/v1_inference_wan_STA.sh
-```
-
-If you want to use sliding tile attention in your custom model:
-```python
-from st_attn import sliding_tile_attention
-# assuming video size (T, H, W) = (30, 48, 80), text tokens = 256 with padding. 
-# q, k, v: [batch_size, num_heads, seq_length, head_dim], seq_length = T*H*W + 256
-# a tile is a cube of size (6, 8, 8)
-# window_size in tiles: [(window_t, window_h, window_w), (..)...]. For example, window size (3, 3, 3) means a query can attend to (3x6, 3x8, 3x8) = (18, 24, 24) tokens out of the total 30x48x80 video.
-# text_length: int ranging from 0 to 256
-# If your attention contains text token (Hunyuan)
-out = sliding_tile_attention(q, k, v, window_size, text_length)
-# If your attention does not contain text token (StepVideo)
-out = sliding_tile_attention(q, k, v, window_size, 0, False)
-```
-
-
-### Test
-```bash
-python tests/test_sta.py # test STA
-python tests/test_vsa.py # test VSA
-```
-### Benchmark
-```bash
-python benchmarks/bench_sta.py
-```
-
-
-### How Does STA Work?
-We give a demo for 2D STA with window size (6,6) operating on a (10, 10) image. 
-
-
-https://github.com/user-attachments/assets/f3b6dd79-7b43-4b60-a0fa-3d6495ec5747
-
-## Why is STA Fast?
-2D/3D Sliding Window Attention (SWA) creates many mixed blocks in the attention map. Even though mixed blocks have less output value,a mixed block is significantly slower than a dense block due to the GPU-unfriendly masking operation. 
-
-STA removes mixed blocks.
-
-
-<div align="center">
-<img src=../../assets/sliding_tile_attn_map.png width="80%"/>
-</div>
-
-## Acknowledgement
-
-We learned or reuse code from FlexAtteniton, NATEN, and ThunderKittens.

csrc/attn/benchmarks/bench_sta.py

@@ -1,145 +0,0 @@
-import os
-from collections import defaultdict
-
-import matplotlib.pyplot as plt
-import numpy as np
-import torch
-from st_attn import sliding_tile_attention
-from triton.testing import do_bench
-
-
-def flops(batch, seqlen, nheads, headdim, causal, mode="fwd"):
-    assert mode in ["fwd", "bwd", "fwd_bwd"]
-    f = 4 * batch * seqlen**2 * nheads * headdim // (2 if causal else 1)
-    return f if mode == "fwd" else (2.5 * f if mode == "bwd" else 3.5 * f)
-
-
-def compute_TFLOPS(flops, ms):
-    flops = flops / 1e12
-    ms = ms / 1e3
-    return flops / ms
-
-
-def benchmark_attention(configurations):
-    results = {'fwd': defaultdict(list), 'bwd': defaultdict(list)}
-
-    for B, H, N, D, causal, dit_seq_shape, window_size in configurations:
-        print("=" * 60)
-        print(f"Timing forward and backward pass for B={B}, H={H}, N={N}, D={D}, causal={causal}")
-
-        q = torch.randn(B, H, N, D, dtype=torch.bfloat16, device='cuda', requires_grad=False).contiguous()
-        k = torch.randn(B, H, N, D, dtype=torch.bfloat16, device='cuda', requires_grad=False).contiguous()
-        v = torch.randn(B, H, N, D, dtype=torch.bfloat16, device='cuda', requires_grad=False).contiguous()
-
-        # grad_output = torch.randn_like(q, requires_grad=False).contiguous()
-        # qg = torch.zeros_like(q, requires_grad=False, dtype=torch.float).contiguous()
-        # kg = torch.zeros_like(k, requires_grad=False, dtype=torch.float).contiguous()
-        # vg = torch.zeros_like(v, requires_grad=False, dtype=torch.float).contiguous()
-
-
-        # # Warmup for forward pass
-        # for _ in range(10):
-        #     o = sliding_tile_attention(q, k, v, [[3, 6, 10]] * 24, 0, False, dit_seq_shape)
-
-        # # Time the forward pass
-        # for i in range(10):
-        #     start_events_fwd[i].record()
-        #     o = sliding_tile_attention(q, k, v, [[3, 6, 10]] * 24, 0, False, dit_seq_shape)
-        #     end_events_fwd[i].record()
-        ms = do_bench(lambda: sliding_tile_attention(q, k, v, [window_size] * 24, 0, False, dit_seq_shape))
-
-        # times_fwd = [s.elapsed_time(e) for s, e in zip(start_events_fwd, end_events_fwd)]
-        # time_us_fwd = np.mean(times_fwd) * 1000
-
-        tflops_fwd = compute_TFLOPS(flops(B, N, H, D, causal, 'fwd'), ms)
-        results['fwd'][(D, causal)].append((N, tflops_fwd))
-
-        print(f"Average time for forward pass (ms): {ms:.2f}")
-        print(f"Average TFLOPS: {tflops_fwd}")
-        print("-" * 60)
-
-        # torch.cuda.empty_cache()
-        # torch.cuda.synchronize()
-
-        # # Prepare for timing backward pass
-        # start_events_bwd = [torch.cuda.Event(enable_timing=True) for _ in range(10)]
-        # end_events_bwd = [torch.cuda.Event(enable_timing=True) for _ in range(10)]
-
-        # # Warmup for backward pass
-        # for _ in range(10):
-        #     qg, kg, vg = tk.mha_backward(q, k, v, o, l_vec, grad_output, causal)
-
-        # # Time the backward pass
-        # for i in range(10):
-        #     start_events_bwd[i].record()
-        #     qg, kg, vg = tk.mha_backward(q, k, v, o, l_vec, grad_output, causal)
-        #     end_events_bwd[i].record()
-
-        # torch.cuda.synchronize()
-        # times_bwd = [s.elapsed_time(e) for s, e in zip(start_events_bwd, end_events_bwd)]
-        # time_us_bwd = np.mean(times_bwd) * 1000
-
-        # tflops_bwd = compute_TFLOPS(flops(B, N, H, D, causal, 'bwd'), ms)
-        # results['bwd'][(D, causal)].append((N, tflops_bwd))
-
-        # print(f"Average time for backward pass(ms): {ms:.2f}")
-        # print(f"Average TFLOPS: {tflops_bwd}")
-        # print("=" * 60)
-
-    return results
-
-
-def plot_results(results):
-    os.makedirs('benchmark_results', exist_ok=True)
-    for mode in ['fwd', 'bwd']:
-        for (D, causal), values in results[mode].items():
-            seq_lens = [x[0] for x in values]
-            tflops = [x[1] for x in values]
-
-            plt.figure(figsize=(10, 6))
-            bars = plt.bar(range(len(seq_lens)), tflops, tick_label=seq_lens)
-            plt.xlabel('Sequence Length')
-            plt.ylabel('TFLOPS')
-            plt.title(f'{mode.upper()} Pass - Head Dim: {D}, Causal: {causal}')
-            plt.grid(True)
-
-            # Adding the numerical y value on top of each bar
-            for bar in bars:
-                yval = bar.get_height()
-                plt.text(bar.get_x() + bar.get_width() / 2, yval, round(yval, 2), ha='center', va='bottom')
-
-            filename = f'benchmark_results/{mode}_D{D}_causal{causal}.png'
-            plt.savefig(filename)
-            plt.close()
-
-
-# Example list of configurations to test
-configurations = [
-    (2, 24, 69120, 128, False, '18x48x80', [3, 6, 10]),
-    (2, 24, 69120, 128, True, '18x48x80', [3, 6, 10]),
-    (2, 24, 82944, 128, False, '36x48x48', [3, 3, 6]), # Stepvideo
-    (2, 24, 82944, 128, True, '36x48x48', [3, 3, 6]),
-    # (16, 16, 768*16,  128, False),
-    # (16, 16, 768*2,  128, False),
-    # (16, 16, 768*4,  128, False),
-    # (16, 16, 768*8,  128, False),
-    # (16, 16, 768*16, 128, False),
-    # (16, 16, 768,    128, True),
-    # (16, 16, 768*2,  128, True),
-    # (16, 16, 768*4,  128, True),
-    # (16, 16, 768*8,  128, True),
-    # (16, 16, 768*16, 128, True),
-    # (16, 32, 768,    64,  False),
-    # (16, 32, 768*2,  64,  False),
-    # (16, 32, 768*4,  64,  False),
-    # (16, 32, 768*8,  64,  False),
-    # (16, 32, 768*16, 64,  False),
-    # (16, 32, 768,    64,  True),
-    # (16, 32, 768*2,  64,  True),
-    # (16, 32, 768*4,  64,  True),
-    # (16, 32, 768*8,  64,  True),
-    # (16, 32, 768*16, 64,  True),
-]
-
-results = benchmark_attention(configurations)
-# plot_results(results)

csrc/attn/benchmarks/bench_vsa_hopper.py

@@ -1,224 +0,0 @@
-import torch
-import argparse
-from triton.testing import do_bench
-from vsa import block_sparse_fwd, block_sparse_bwd
-from vsa import BLOCK_M, BLOCK_N
-import  triton
-import numpy as np
-import random
-
-def set_seed(seed: int = 42):
-    # Python random module
-    random.seed(seed)
-
-    # NumPy
-    np.random.seed(seed)
-
-    # PyTorch
-    torch.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)  # if using multi-GPU
-
-def parse_arguments():
-    parser = argparse.ArgumentParser(description='Benchmark Block Sparse Attention')
-    parser.add_argument('--batch_size', type=int, default=1, help='Batch size')
-    parser.add_argument('--num_heads', type=int, default=12, help='Number of heads')
-    parser.add_argument('--head_dim', type=int, default=128, help='Head dimension')
-    parser.add_argument('--topk', type=int, default=None, help='Number of kv blocks each q block attends to')
-    parser.add_argument('--seq_lengths', type=int, nargs='+', default=[49152], help='Sequence lengths to benchmark')
-    return parser.parse_args()
-
-def create_input_tensors(batch, head, seq_len, headdim):
-    """Create random input tensors for attention."""
-    q = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    k = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    v = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    return q, k, v
-
-def generate_block_sparse_pattern(bs, h, num_q_blocks, num_kv_blocks, k, device="cuda"):
-    """
-    Generate a block sparse pattern where each q block attends to exactly k kv blocks.
-    
-    Args:
-        bs: batch size
-        h: number of heads
-        num_q_blocks: number of query blocks
-        num_kv_blocks: number of key-value blocks
-        k: number of kv blocks each q block attends to
-        device: device to create tensors on
-        
-    Returns:
-        q2k_block_sparse_index: [bs, h, num_q_blocks, k]
-            Contains the indices of kv blocks that each q block attends to.
-        q2k_block_sparse_num: [bs, h, num_q_blocks]
-            Contains the number of kv blocks that each q block attends to (all equal to k).
-        k2q_block_sparse_index: [bs, h, num_kv_blocks, num_q_blocks]
-            Contains the indices of q blocks that attend to each kv block.
-        k2q_block_sparse_num: [bs, h, num_kv_blocks]
-            Contains the number of q blocks that attend to each kv block.
-        block_sparse_mask: [bs, h, num_q_blocks, num_kv_blocks]
-            Binary mask where 1 indicates attention connection.
-    """
-    # Ensure k is not larger than num_kv_blocks
-    k = min(k, num_kv_blocks)
-    
-    # Create random scores for sampling
-    scores = torch.rand(bs, h, num_q_blocks, num_kv_blocks, device=device)
-    
-    # Get top-k indices for each q block
-    _, q2k_block_sparse_index = torch.topk(scores, k, dim=-1)
-    q2k_block_sparse_index = q2k_block_sparse_index.to(torch.int32)
-    
-    # sort q2k_block_sparse_index
-    q2k_block_sparse_index, _ = torch.sort(q2k_block_sparse_index, dim=-1)
-
-    # All q blocks attend to exactly k kv blocks
-    q2k_block_sparse_num = torch.full((bs, h, num_q_blocks), k, dtype=torch.int32, device=device)
-    
-    # Create the corresponding mask
-    block_sparse_mask = torch.zeros(bs, h, num_q_blocks, num_kv_blocks, dtype=torch.bool, device=device)
-    
-    # Fill in the mask based on the indices
-    for b in range(bs):
-        for head in range(h):
-            for q_idx in range(num_q_blocks):
-                kv_indices = q2k_block_sparse_index[b, head, q_idx]
-                block_sparse_mask[b, head, q_idx, kv_indices] = True
-    
-    # Create the reverse mapping (k2q)
-    # First, initialize lists to collect q indices for each kv block
-    k2q_indices_list = [[[] for _ in range(num_kv_blocks)] for _ in range(bs * h)]
-    
-    # Populate the lists based on q2k mapping
-    for b in range(bs):
-        for head in range(h):
-            flat_idx = b * h + head
-            for q_idx in range(num_q_blocks):
-                kv_indices = q2k_block_sparse_index[b, head, q_idx].tolist()
-                for kv_idx in kv_indices:
-                    k2q_indices_list[flat_idx][kv_idx].append(q_idx)
-    
-    # Find the maximum number of q blocks that attend to any kv block
-    max_q_per_kv = 0
-    for flat_idx in range(bs * h):
-        for kv_idx in range(num_kv_blocks):
-            max_q_per_kv = max(max_q_per_kv, len(k2q_indices_list[flat_idx][kv_idx]))
-    
-    # Create tensors for k2q mapping
-    k2q_block_sparse_index = torch.full((bs, h, num_kv_blocks, max_q_per_kv), -1, 
-                                        dtype=torch.int32, device=device)
-    k2q_block_sparse_num = torch.zeros((bs, h, num_kv_blocks), 
-                                       dtype=torch.int32, device=device)
-    
-    # Fill the tensors
-    for b in range(bs):
-        for head in range(h):
-            flat_idx = b * h + head
-            for kv_idx in range(num_kv_blocks):
-                q_indices = k2q_indices_list[flat_idx][kv_idx]
-                num_q = len(q_indices)
-                k2q_block_sparse_num[b, head, kv_idx] = num_q
-                if num_q > 0:
-                    k2q_block_sparse_index[b, head, kv_idx, :num_q] = torch.tensor(
-                        q_indices, dtype=torch.int32, device=device)
-                
-    return q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask
-
-def benchmark_block_sparse_attention(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, flops):
-    """Benchmark block sparse attention forward and backward passes."""
-    print("\n=== BLOCK SPARSE ATTENTION BENCHMARK ===")
-    
-    # Forward pass
-    # Warm-up run
-    variable_block_sizes = torch.ones(q2k_block_sparse_index.shape[2], device=q.device).int() * BLOCK_M
-    o, l_vec = block_sparse_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, variable_block_sizes)
-    torch.cuda.synchronize()
-    
-    # Benchmark forward
-    fwd_time = do_bench(
-        lambda: block_sparse_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, variable_block_sizes),
-        warmup=5,
-        rep=20,
-        quantiles=None
-    )
-    
-    sparse_tflops = flops / fwd_time * 1e-12 * 1e3
-    print(f"Block Sparse Forward - TFLOPS: {sparse_tflops:.2f}")
-    
-    # Backward pass
-    grad_output = torch.randn_like(o)
-    
-    # Warm-up runs
-    for _ in range(5):
-        block_sparse_bwd(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num, variable_block_sizes)
-    torch.cuda.synchronize()
-    
-    # Benchmark backward
-    bwd_time = do_bench(
-        lambda: block_sparse_bwd(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num, variable_block_sizes),
-        warmup=5,
-        rep=20,
-        quantiles=None
-    )
-    bwd_flops = 2.5 * flops  # Approximation
-
-    sparse_bwd_tflops = bwd_flops / bwd_time * 1e-12 * 1e3
-    print(f"Block Sparse Backward - TFLOPS: {sparse_bwd_tflops:.2f}")
-    
-    return sparse_tflops, sparse_bwd_tflops
-
-def main():
-    args = parse_arguments()
-
-    set_seed(42)
-    
-    # Extract parameters
-    batch = args.batch_size
-    head = args.num_heads
-    headdim = args.head_dim
-    
-    print(f"Block Sparse Attention Benchmark")
-    print(f"batch: {batch}, head: {head}, headdim: {headdim}")
-    
-    # Test with different sequence lengths
-    for seq_len in args.seq_lengths:
-        # Skip very long sequences if they might cause OOM
-        if seq_len > 16384 and batch > 1:
-            continue
-            
-        print("="*100)
-        print(f"\nSequence length: {seq_len}")
-        
-        # Calculate theoretical FLOPs for attention
-        flops = 4 * batch * head * headdim * seq_len * seq_len
-        
-        # Create input tensors
-        q, k, v = create_input_tensors(batch, head, seq_len, headdim)
-        
-        # Setup block sparse parameters
-        num_q_blocks = seq_len // BLOCK_M
-        num_kv_blocks = seq_len // BLOCK_N
-        
-        # Determine k value (number of kv blocks per q block)
-        topk = args.topk
-        if topk is None:
-            topk = num_kv_blocks // 10  # Default to ~90% sparsity if k is not specified
-        topk = max(1, topk)       
-        print(f"Using topk={topk} kv blocks per q block (out of {num_kv_blocks} total kv blocks)")
-        
-        # Generate block sparse pattern
-        q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, _ = generate_block_sparse_pattern(
-            batch, head, num_q_blocks, num_kv_blocks, topk, device="cuda")
-        
-        # Benchmark block sparse attention
-        sparse_fwd, sparse_bwd = benchmark_block_sparse_attention(
-            q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, flops
-        )
-        
-        # Print results
-        print("\n=== PERFORMANCE RESULTS ===")
-        print(f"Block Sparse Forward - TFLOPS: {sparse_fwd:.2f}")
-        print(f"Block Sparse Backward - TFLOPS: {sparse_bwd:.2f}")
-
-if __name__ == "__main__":
-    main()

csrc/attn/benchmarks/bench_vsa_triton.py

@@ -1,217 +0,0 @@
-import torch
-import argparse
-import triton.testing
-from vsa import block_sparse_attn
-from vsa import BLOCK_M, BLOCK_N
-
-import numpy as np
-import random
-
-def set_seed(seed: int = 42):
-    # Python random module
-    random.seed(seed)
-
-    # NumPy
-    np.random.seed(seed)
-
-    # PyTorch
-    torch.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)  # if using multi-GPU
-
-def parse_arguments():
-    parser = argparse.ArgumentParser(description='Benchmark Block Sparse Attention')
-    parser.add_argument('--batch_size', type=int, default=1, help='Batch size')
-    parser.add_argument('--num_heads', type=int, default=12, help='Number of heads')
-    parser.add_argument('--head_dim', type=int, default=64, help='Head dimension')
-    parser.add_argument('--topk', type=int, default=None, help='Number of kv blocks each q block attends to')
-    parser.add_argument('--seq_lengths', type=int, nargs='+', default=[49152], help='Sequence lengths to benchmark')
-    return parser.parse_args()
-
-def create_input_tensors(batch, head, seq_len, headdim):
-    """Create random input tensors for attention."""
-    q = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    k = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    v = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    return q, k, v
-
-def generate_block_sparse_pattern(bs, h, num_q_blocks, num_kv_blocks, k, device="cuda"):
-    """
-    Generate a block sparse pattern where each q block attends to exactly k kv blocks.
-    
-    Args:
-        bs: batch size
-        h: number of heads
-        num_q_blocks: number of query blocks
-        num_kv_blocks: number of key-value blocks
-        k: number of kv blocks each q block attends to
-        device: device to create tensors on
-        
-    Returns:
-        q2k_block_sparse_index: [bs, h, num_q_blocks, k]
-            Contains the indices of kv blocks that each q block attends to.
-        q2k_block_sparse_num: [bs, h, num_q_blocks]
-            Contains the number of kv blocks that each q block attends to (all equal to k).
-        k2q_block_sparse_index: [bs, h, num_kv_blocks, num_q_blocks]
-            Contains the indices of q blocks that attend to each kv block.
-        k2q_block_sparse_num: [bs, h, num_kv_blocks]
-            Contains the number of q blocks that attend to each kv block.
-        block_sparse_mask: [bs, h, num_q_blocks, num_kv_blocks]
-            Binary mask where 1 indicates attention connection.
-    """
-    # Ensure k is not larger than num_kv_blocks
-    k = min(k, num_kv_blocks)
-    
-    # Create random scores for sampling
-    scores = torch.rand(bs, h, num_q_blocks, num_kv_blocks, device=device)
-    
-    # Get top-k indices for each q block
-    _, q2k_block_sparse_index = torch.topk(scores, k, dim=-1)
-    q2k_block_sparse_index = q2k_block_sparse_index.to(torch.int32)
-    
-    # sort q2k_block_sparse_index
-    q2k_block_sparse_index, _ = torch.sort(q2k_block_sparse_index, dim=-1)
-
-    # All q blocks attend to exactly k kv blocks
-    q2k_block_sparse_num = torch.full((bs, h, num_q_blocks), k, dtype=torch.int32, device=device)
-    
-    # Create the corresponding mask
-    block_sparse_mask = torch.zeros(bs, h, num_q_blocks, num_kv_blocks, dtype=torch.bool, device=device)
-    
-    # Fill in the mask based on the indices
-    for b in range(bs):
-        for head in range(h):
-            for q_idx in range(num_q_blocks):
-                kv_indices = q2k_block_sparse_index[b, head, q_idx]
-                block_sparse_mask[b, head, q_idx, kv_indices] = True
-    
-    # Create the reverse mapping (k2q)
-    # First, initialize lists to collect q indices for each kv block
-    k2q_indices_list = [[[] for _ in range(num_kv_blocks)] for _ in range(bs * h)]
-    
-    # Populate the lists based on q2k mapping
-    for b in range(bs):
-        for head in range(h):
-            flat_idx = b * h + head
-            for q_idx in range(num_q_blocks):
-                kv_indices = q2k_block_sparse_index[b, head, q_idx].tolist()
-                for kv_idx in kv_indices:
-                    k2q_indices_list[flat_idx][kv_idx].append(q_idx)
-    
-    # Find the maximum number of q blocks that attend to any kv block
-    max_q_per_kv = 0
-    for flat_idx in range(bs * h):
-        for kv_idx in range(num_kv_blocks):
-            max_q_per_kv = max(max_q_per_kv, len(k2q_indices_list[flat_idx][kv_idx]))
-    
-    # Create tensors for k2q mapping
-    k2q_block_sparse_index = torch.full((bs, h, num_kv_blocks, max_q_per_kv), -1, 
-                                        dtype=torch.int32, device=device)
-    k2q_block_sparse_num = torch.zeros((bs, h, num_kv_blocks), 
-                                       dtype=torch.int32, device=device)
-    
-    # Fill the tensors
-    for b in range(bs):
-        for head in range(h):
-            flat_idx = b * h + head
-            for kv_idx in range(num_kv_blocks):
-                q_indices = k2q_indices_list[flat_idx][kv_idx]
-                num_q = len(q_indices)
-                k2q_block_sparse_num[b, head, kv_idx] = num_q
-                if num_q > 0:
-                    k2q_block_sparse_index[b, head, kv_idx, :num_q] = torch.tensor(
-                        q_indices, dtype=torch.int32, device=device)
-                
-    return q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask
-
-def benchmark_block_sparse_attention(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, flops):
-    """Benchmark block sparse attention forward+backward pass."""
-    print("\n=== BLOCK SPARSE ATTENTION FORWARD+BACKWARD BENCHMARK ===")
-    
-    # Combined forward+backward pass
-    # Warm-up run
-    q_fwd = q.clone().requires_grad_(True)
-    k_fwd = k.clone().requires_grad_(True)
-    v_fwd = v.clone().requires_grad_(True)
-    o = block_sparse_attn(q_fwd, k_fwd, v_fwd, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num)
-    grad_output = torch.randn_like(o)
-    o.backward(grad_output)
-    torch.cuda.synchronize()
-    
-    # Benchmark forward+backward
-    def forward_backward_fn():
-        q_fwd = q.clone().requires_grad_(True)
-        k_fwd = k.clone().requires_grad_(True)
-        v_fwd = v.clone().requires_grad_(True)
-        o = block_sparse_attn(q_fwd, k_fwd, v_fwd, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num)
-        grad_output = torch.randn_like(o)
-        o.backward(grad_output)
-    
-    total_time = triton.testing.do_bench(
-        forward_backward_fn,
-        warmup=25,
-        rep=100,
-        return_mode='mean'
-    )
-    
-    # Total flops for forward + backward (forward + 2.5x backward approximation)
-    total_flops = flops + 2.5 * flops  # 3.5x the forward flops
-    sparse_tflops = total_flops / total_time * 1e-12 * 1e3
-    print(f"Block Sparse Forward+Backward - TFLOPS: {sparse_tflops:.2f}")
-    
-    return sparse_tflops
-
-def main():
-    args = parse_arguments()
-
-    set_seed(42)
-    
-    # Extract parameters
-    batch = args.batch_size
-    head = args.num_heads
-    headdim = args.head_dim
-    
-    print(f"Block Sparse Attention Benchmark")
-    print(f"batch: {batch}, head: {head}, headdim: {headdim}")
-    
-    # Test with different sequence lengths
-    for seq_len in args.seq_lengths:
-        # Skip very long sequences if they might cause OOM
-        if seq_len > 16384 and batch > 1:
-            continue
-            
-        print("="*100)
-        print(f"\nSequence length: {seq_len}")
-        
-        # Calculate theoretical FLOPs for attention
-        flops = 4 * batch * head * headdim * seq_len * seq_len
-        
-        # Create input tensors
-        q, k, v = create_input_tensors(batch, head, seq_len, headdim)
-        
-        # Setup block sparse parameters
-        num_q_blocks = seq_len // BLOCK_M
-        num_kv_blocks = seq_len // BLOCK_N
-        
-        # Determine k value (number of kv blocks per q block)
-        topk = args.topk
-        if topk is None:
-            topk = num_kv_blocks // 10  # Default to ~90% sparsity if k is not specified
-        topk = max(1, topk)       
-        print(f"Using topk={topk} kv blocks per q block (out of {num_kv_blocks} total kv blocks)")
-        
-        # Generate block sparse pattern
-        q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, _ = generate_block_sparse_pattern(
-            batch, head, num_q_blocks, num_kv_blocks, topk, device="cuda")
-        
-        # Benchmark block sparse attention
-        sparse_fwd = benchmark_block_sparse_attention(
-            q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, flops
-        )
-        
-        # Print results
-        print("\n=== PERFORMANCE RESULTS ===")
-        print(f"Block Sparse Forward+Backward - TFLOPS: {sparse_fwd:.2f}")
-
-if __name__ == "__main__":
-    main()

csrc/attn/sliding_tile_attn/MANIFEST.in

@@ -1,2 +0,0 @@
-recursive-include tk *
-include config_sta.py

csrc/attn/sliding_tile_attn/README.md

@@ -1,96 +0,0 @@
-
-# Attention Kernel Used in FastVideo
-
-## Sliding Tile Attention (STA)
-We only support H100 for STA.
-
-### Installation
-```bash
-pip install st_attn
-``` 
-
-Install from source:
-
-```bash
-git submodule update --init --recursive
-python setup.py install
-```
-
-If you encounter error during installation, try below:
-Install C++20 for ThunderKittens:
-```bash
-sudo apt update
-sudo apt install gcc-11 g++-11
-
-sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
-
-sudo apt update
-sudo apt install clang-11
-```
-(If you use CUDA12.8)
-```bash
-export CUDA_HOME=/usr/local/cuda-12.8
-export PATH=${CUDA_HOME}/bin:${PATH} 
-export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
-```
-
-###  Usage
-End-2-end inference with FastVideo:
-```bash
-bash scripts/inference/v1_inference_wan_STA.sh
-```
-
-If you want to use sliding tile attention in your custom model:
-```python
-from st_attn import sliding_tile_attention
-# assuming video size (T, H, W) = (30, 48, 80), text tokens = 256 with padding. 
-# q, k, v: [batch_size, num_heads, seq_length, head_dim], seq_length = T*H*W + 256
-# a tile is a cube of size (6, 8, 8)
-# window_size in tiles: [(window_t, window_h, window_w), (..)...]. For example, window size (3, 3, 3) means a query can attend to (3x6, 3x8, 3x8) = (18, 24, 24) tokens out of the total 30x48x80 video.
-# text_length: int ranging from 0 to 256
-# If your attention contains text token (Hunyuan)
-out = sliding_tile_attention(q, k, v, window_size, text_length)
-# If your attention does not contain text token (StepVideo)
-out = sliding_tile_attention(q, k, v, window_size, 0, False)
-```
-
-
-### Test
-```bash
-python ../tests/test_sta.py # test STA
-python ../tests/test_vsa.py # test VSA
-```
-### Benchmark
-```bash
-python ../benchmarks/bench_sta.py
-```
-
-
-### How Does STA Work?
-We give a demo for 2D STA with window size (6,6) operating on a (10, 10) image. 
-
-
-https://github.com/user-attachments/assets/f3b6dd79-7b43-4b60-a0fa-3d6495ec5747
-
-
-## STA Configuration Logic
-Here is a diagram of how the window is configured and passed through the FastVideo pipeline:
-
-<div align="center">
-<img src="../../../docs/assets/images/STA_configuration.png" width="80%"/>
-</div>
-
-
-## Why is STA Fast?
-2D/3D Sliding Window Attention (SWA) creates many mixed blocks in the attention map. Even though mixed blocks have less output value,a mixed block is significantly slower than a dense block due to the GPU-unfriendly masking operation. 
-
-STA removes mixed blocks.
-
-
-<div align="center">
-<img src=../../../assets/sliding_tile_attn_map.png width="80%"/>
-</div>
-
-## Acknowledgement
-
-We learned or reuse code from FlexAtteniton, NATEN, and ThunderKittens.

csrc/attn/sliding_tile_attn/config_sta.py

@@ -1,15 +0,0 @@
-### ADD TO THIS TO REGISTER NEW KERNELS
-sources = {
-    'st_attn': {
-        'source_files': {
-            'h100': 'st_attn/st_attn_h100.cu'  # define these source files for each GPU target desired.
-        }
-    }
-}
-
-### WHICH KERNELS DO WE WANT TO BUILD?
-# (oftentimes during development work you don't need to redefine them all.)
-kernels = ['st_attn']
-
-### WHICH GPU TARGET DO WE WANT TO BUILD FOR?
-target = 'h100'

csrc/attn/sliding_tile_attn/setup.py

@@ -1,76 +0,0 @@
-import os
-import subprocess
-
-from config_sta import kernels, sources, target
-from setuptools import find_packages, setup
-from torch.utils.cpp_extension import BuildExtension, CUDAExtension
-
-target = target.lower()
-
-# Package metadata
-PACKAGE_NAME = "st_attn"
-VERSION = "0.0.6"
-AUTHOR = "Hao AI Lab"
-DESCRIPTION = "Sliding Tile Atteniton Kernel Used in FastVideo"
-URL = "https://github.com/hao-ai-lab/FastVideo/tree/main/csrc/sliding_tile_attention"
-
-# Set environment variables
-tk_root = os.getenv('THUNDERKITTENS_ROOT', os.path.abspath(os.path.join(os.getcwd(), 'tk/')))
-python_include = subprocess.check_output(['python', '-c',
-                                          "import sysconfig; print(sysconfig.get_path('include'))"]).decode().strip()
-torch_include = subprocess.check_output([
-    'python', '-c',
-    "import torch; from torch.utils.cpp_extension import include_paths; print(' '.join(['-I' + p for p in include_paths()]))"
-]).decode().strip()
-print('st_attn root:', tk_root)
-print('Python include:', python_include)
-print('Torch include directories:', torch_include)
-
-# CUDA flags
-cuda_flags = [
-    '-DNDEBUG', '-Xcompiler=-Wno-psabi', '-Xcompiler=-fno-strict-aliasing', '--expt-extended-lambda',
-    '--expt-relaxed-constexpr', '-forward-unknown-to-host-compiler', '--use_fast_math', '-std=c++20', '-O3',
-    '-Xnvlink=--verbose', '-Xptxas=--verbose', '-Xptxas=--warn-on-spills', f'-I{tk_root}/include',
-    f'-I{tk_root}/prototype', f'-I{python_include}', '-DTORCH_COMPILE'
-] + torch_include.split()
-cpp_flags = ['-std=c++20', '-O3']
-
-if target == 'h100':
-    cuda_flags.append('-DKITTENS_HOPPER')
-    cuda_flags.append('-arch=sm_90a')
-else:
-    raise ValueError(f'Target {target} not supported')
-
-source_files = ['st_attn.cpp']
-for k in kernels:
-    if target not in sources[k]['source_files']:
-        raise KeyError(f'Target {target} not found in source files for kernel {k}')
-    if isinstance(sources[k]['source_files'][target], list):
-        source_files.extend(sources[k]['source_files'][target])
-    else:
-        source_files.append(sources[k]['source_files'][target])
-    cpp_flags.append(f'-DTK_COMPILE_{k.replace(" ", "_").upper()}')
-
-setup(name=PACKAGE_NAME,
-      version=VERSION,
-      author=AUTHOR,
-      description=DESCRIPTION,
-      url=URL,
-      packages=find_packages(),
-      ext_modules=[
-          CUDAExtension('st_attn_cuda',
-                        sources=source_files,
-                        extra_compile_args={
-                            'cxx': cpp_flags,
-                            'nvcc': cuda_flags
-                        },
-                        libraries=['cuda'])
-      ],
-      cmdclass={'build_ext': BuildExtension},
-      classifiers=[
-          "Programming Language :: Python :: 3",
-          "Environment :: GPU :: NVIDIA CUDA :: 12",
-          "License :: OSI Approved :: Apache Software License",
-      ],
-      python_requires='>=3.10',
-      install_requires=["torch>=2.5.0"])

csrc/attn/sliding_tile_attn/st_attn.cpp

@@ -1,23 +0,0 @@
-#include <torch/extension.h>
-#include <ATen/ATen.h>
-
-#include <vector>
-#include <cuda_fp16.h>
-#include <cuda_bf16.h>
-
-#include <cuda_runtime.h>
-
-#ifdef TK_COMPILE_ST_ATTN
-extern torch::Tensor sta_forward(
-    torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o, int kernel_t_size, int kernel_w_size, int kernel_h_size, int text_length, bool process_text, bool has_text, int kernel_aspect_ratio_flag
-); 
-#endif
-
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-    m.doc() = "Sliding Block Attention Kernels"; // optional module docstring
-    
-#ifdef TK_COMPILE_ST_ATTN
-    m.def("sta_fwd",  torch::wrap_pybind_function(sta_forward), "sliding tile attention, assuming tile size is (6,8,8)");
-#endif
-}

csrc/attn/sliding_tile_attn/st_attn/__init__.py

@@ -1,49 +0,0 @@
-import math
-
-import torch
-from torch.utils.checkpoint import detach_variable
-try:
-    from st_attn_cuda import sta_fwd
-except ImportError:
-    sta_fwd = None
-
-def sliding_tile_attention(q_all, k_all, v_all, window_size, text_length, has_text=True, dit_seq_shape='30x48x80'):
-    seq_length = q_all.shape[2]
-    dit_seq_shape_mapping = {
-        '30x48x80':1,
-        '36x48x48':2,
-        '18x48x80':3, 
-    }
-    if has_text:
-        assert q_all.shape[
-            2] >= 115200 and q_all.shape[2] <= 115456, f"Unsupported {dit_seq_shape}, current shape is {q_all.shape}, only support '30x48x80' for HunyuanVideo"
-        assert q_all.shape[1] == len(window_size), "Number of heads must match the number of window sizes"
-        target_size = math.ceil(seq_length / 384) * 384
-        pad_size = target_size - seq_length
-        if pad_size > 0:
-            q_all = torch.cat([q_all, q_all[:, :, -pad_size:]], dim=2)
-            k_all = torch.cat([k_all, k_all[:, :, -pad_size:]], dim=2)
-            v_all = torch.cat([v_all, v_all[:, :, -pad_size:]], dim=2)
-    else:
-        if dit_seq_shape == '36x48x48': # Stepvideo 204x768x68
-            assert q_all.shape[2] == 82944
-        elif dit_seq_shape == '18x48x80': # Wan 69x768x1280
-            assert q_all.shape[2] == 69120
-        else:
-            raise ValueError(f"Unsupported {dit_seq_shape}, current shape is {q_all.shape}, only support '36x48x48' for Stepvideo and '18x48x80' for Wan")
-
-    kernel_aspect_ratio_flag = dit_seq_shape_mapping[dit_seq_shape]
-    hidden_states = torch.empty_like(q_all)
-    # This for loop is ugly. but it is actually quite efficient. The sequence dimension alone can already oversubscribe SMs
-    for head_index, (t_kernel, h_kernel, w_kernel) in enumerate(window_size):
-        for batch in range(q_all.shape[0]):
-            q_head, k_head, v_head, o_head = (q_all[batch:batch + 1, head_index:head_index + 1],
-                                              k_all[batch:batch + 1,
-                                                    head_index:head_index + 1], v_all[batch:batch + 1,
-                                                                                      head_index:head_index + 1],
-                                              hidden_states[batch:batch + 1, head_index:head_index + 1])
-
-            _ = sta_fwd(q_head, k_head, v_head, o_head, t_kernel, h_kernel, w_kernel, text_length, False, has_text, kernel_aspect_ratio_flag)
-    if has_text:
-        _ = sta_fwd(q_all, k_all, v_all, hidden_states, 3, 3, 3, text_length, True, True, kernel_aspect_ratio_flag)
-    return hidden_states[:, :, :seq_length]

csrc/attn/tests/test_vsa.py

@@ -1,156 +0,0 @@
-import torch
-import sys
-import os
-import numpy as np
-from tqdm import tqdm
-
-# Add the parent directory to the path to import block_sparse_attn
-sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-from tests.utils import generate_block_sparse_mask_for_function, create_full_mask_from_block_mask
-from vsa import block_sparse_attn
-
-BLOCK_M = 64
-BLOCK_N = 64
-
-def pytorch_test(Q, K, V, block_sparse_mask, dO):
-    q_ = Q.clone().float().requires_grad_()
-    k_ = K.clone().float().requires_grad_()
-    v_ = V.clone().float().requires_grad_()
-
-    QK = torch.matmul(q_, k_.transpose(-2, -1))
-    QK /= (q_.size(-1) ** 0.5)
-    QK = QK.masked_fill(~block_sparse_mask.unsqueeze(0), float('-inf'))
-
-    QK = torch.nn.functional.softmax(QK, dim=-1)
-    output = torch.matmul(QK, v_)
-
-    dO_ = dO
-    output.backward(dO_)
-    return (
-        output.to(torch.bfloat16),
-        q_.grad.to(torch.bfloat16),
-        k_.grad.to(torch.bfloat16),
-        v_.grad.to(torch.bfloat16),
-    )
-
-
-def block_sparse_kernel_test(Q, K, V, block_sparse_mask, variable_block_sizes, non_pad_index, dO):
-    Q = Q.detach().requires_grad_()
-    K = K.detach().requires_grad_()
-    V = V.detach().requires_grad_()
-    
-    q_padded = vsa_pad(Q, non_pad_index, variable_block_sizes.shape[0], BLOCK_M)
-    k_padded = vsa_pad(K, non_pad_index, variable_block_sizes.shape[0], BLOCK_M)
-    v_padded = vsa_pad(V, non_pad_index, variable_block_sizes.shape[0], BLOCK_M)
-    output, _= block_sparse_attn(q_padded, k_padded, v_padded, block_sparse_mask, variable_block_sizes)
-    output = output[:, :, non_pad_index, :]
-    output.backward(dO)
-    return output, Q.grad, K.grad, V.grad
-
-
-def get_non_pad_index(
-    vid_len: torch.LongTensor,
-    n_win: int,
-    win_size: int,
-):
-    device = vid_len.device
-    starts_pad = torch.arange(n_win, device=device) * win_size
-    index_pad = starts_pad[:, None] + torch.arange(win_size, device=device)[None, :]
-    index_mask = torch.arange(win_size, device=device)[None, :] < vid_len[:, None]
-
-    return index_pad[index_mask]
-
-def generate_tensor(shape, dtype, device):
-    tensor = torch.randn(shape, dtype=dtype, device=device)
-    return tensor
-
-def generate_variable_block_sizes(num_blocks, min_size=32, max_size=64, device="cuda"):
-    return torch.randint(min_size, max_size + 1, (num_blocks,), device=device, dtype=torch.int32)
-
-
-def vsa_pad(x, non_pad_index, num_blocks, block_size):
-    padded_x = torch.zeros((1, x.shape[1], num_blocks * BLOCK_M, x.shape[3]), device=x.device, dtype=x.dtype)
-    padded_x[:, :, non_pad_index, :] = x
-    return padded_x
-
-def check_correctness(h, d, num_blocks, k,  num_iterations=20, error_mode='all'):
-    results = {
-        'gO': {'sum_diff': 0.0, 'sum_abs': 0.0, 'max_diff': 0.0},
-        'gQ': {'sum_diff': 0.0, 'sum_abs': 0.0, 'max_diff': 0.0},
-        'gK': {'sum_diff': 0.0, 'sum_abs': 0.0, 'max_diff': 0.0},
-        'gV': {'sum_diff': 0.0, 'sum_abs': 0.0, 'max_diff': 0.0},
-    }
-
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-    variable_block_sizes = generate_variable_block_sizes(num_blocks, device=device)
-    S = int(variable_block_sizes.sum().item())
-    padded_S = num_blocks * BLOCK_M
-    non_pad_index = get_non_pad_index(variable_block_sizes, num_blocks, BLOCK_M)
-    block_mask = generate_block_sparse_mask_for_function(h, num_blocks, k, device)
-    full_mask = create_full_mask_from_block_mask(block_mask, variable_block_sizes, device)
-    for _ in range(num_iterations):
-        Q = generate_tensor((1, h, S, d),  torch.bfloat16, device)
-        K = generate_tensor((1, h, S, d), torch.bfloat16, device)
-        V = generate_tensor((1, h, S, d),  torch.bfloat16, device)
-        dO = generate_tensor((1, h, S, d), torch.bfloat16, device)
-
-        # dO_padded = torch.zeros_like(dO_padded)
-        # dO_padded[:, :, non_pad_index, :] = dO
-
-        pt_o, pt_qg, pt_kg, pt_vg = pytorch_test(Q, K, V, full_mask, dO)
-        bs_o, bs_qg, bs_kg, bs_vg = block_sparse_kernel_test(Q, K, V, block_mask.unsqueeze(0), variable_block_sizes,non_pad_index, dO)
-        for name, (pt, bs) in zip(['gQ', 'gK', 'gV', 'gO'], [(pt_qg, bs_qg), (pt_kg, bs_kg), (pt_vg, bs_vg), (pt_o, bs_o)]):
-            if bs is not None:
-                diff = pt - bs
-                abs_diff = torch.abs(diff)
-                results[name]['sum_diff'] += torch.sum(abs_diff).item()
-                results[name]['sum_abs'] += torch.sum(torch.abs(pt)).item()
-                rel_max_diff = torch.max(abs_diff) / torch.mean(torch.abs(pt))
-                results[name]['max_diff'] = max(results[name]['max_diff'], rel_max_diff.item())
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-
-    total_elements = h * S * d * num_iterations
-    for name, data in results.items():
-        avg_diff = data['sum_diff'] / total_elements
-        max_diff = data['max_diff']
-        results[name] = {'avg_diff': avg_diff, 'max_diff': max_diff}
-
-    return results
-
-def generate_error_graphs(h, d, error_mode='all'):
-    test_configs = [
-        {"num_blocks": 16, "k": 2, "description": "Small sequence"},
-        {"num_blocks": 32, "k": 4, "description": "Medium sequence"},
-        {"num_blocks": 53, "k": 6, "description": "Large sequence"},
-    ]
-    
-    print(f"\nError Analysis for h={h}, d={d}, mode={error_mode}")
-    print("=" * 150)
-    print(f"{'Config':<20} {'Blocks':<8} {'K':<4} "
-          f"{'gQ Avg':<12} {'Rel gQ Max':<12} "
-          f"{'gK Avg':<12} {'Rel gK Max':<12} "
-          f"{'gV Avg':<12} {'Rel gV Max':<12} "
-          f"{'gO Avg':<12} {'Rel gO Max':<12}")
-    print("-" * 150)
-
-    for config in test_configs:
-        num_blocks = config["num_blocks"]
-        k = config["k"]
-        description = config["description"]
-        results = check_correctness(h, d, num_blocks, k, error_mode=error_mode)
-        print(f"{description:<20} {num_blocks:<8} {k:<4} "
-              f"{results['gQ']['avg_diff']:<12.6e} {results['gQ']['max_diff']:<12.6e} "
-              f"{results['gK']['avg_diff']:<12.6e} {results['gK']['max_diff']:<12.6e} "
-              f"{results['gV']['avg_diff']:<12.6e} {results['gV']['max_diff']:<12.6e} "
-              f"{results['gO']['avg_diff']:<12.6e} {results['gO']['max_diff']:<12.6e}")
-
-    print("-" * 150)
-
-if __name__ == "__main__":
-    h, d = 16, 128
-    print("Block Sparse Attention with Variable Block Sizes Analysis")
-    print("=" * 60)
-    for mode in ['backward']:
-        generate_error_graphs(h, d, error_mode=mode)
-    print("\nAnalysis completed for all modes.")

csrc/attn/tests/utils.py

@@ -1,54 +0,0 @@
-import torch
-
-def generate_block_sparse_mask_for_function(h, num_blocks, k, device="cuda"):
-    """
-    Generate block sparse mask of shape [h, num_blocks, num_blocks].
-    
-    Args:
-        h: number of heads
-        num_blocks: number of blocks
-        k: number of kv blocks each q block attends to
-        device: device to create tensors on
-
-    Returns:
-        block_sparse_mask: [h, num_blocks, num_blocks] bool tensor
-    """
-    k = min(k, num_blocks)
-    scores = torch.rand(h, num_blocks, num_blocks, device=device)
-    _, indices = torch.topk(scores, k, dim=-1)
-    block_sparse_mask = torch.zeros(h, num_blocks, num_blocks, dtype=torch.bool, device=device)
-
-    block_sparse_mask = block_sparse_mask.scatter_(2, indices, 1).bool()
-    return block_sparse_mask
-
-
-def create_full_mask_from_block_mask(block_sparse_mask, variable_block_sizes, device="cuda"):
-    """
-    Convert block-level sparse mask to full attention mask.
-    
-    Args:
-        block_sparse_mask: [h, num_blocks, num_blocks] bool tensor
-        variable_block_sizes: [num_blocks] tensor
-        device: device to create tensors on
-
-    Returns:
-        full_mask: [h, S, S] bool tensor where S = total sequence length
-    """
-    h, num_blocks, _ = block_sparse_mask.shape
-    total_seq_len = variable_block_sizes.sum().item()
-    cumsum = torch.cat([torch.tensor([0], device=device), variable_block_sizes.cumsum(dim=0)[:-1]])
-
-    full_mask = torch.zeros(h, total_seq_len, total_seq_len, dtype=torch.bool, device=device)
-
-    for head in range(h):
-        for q_block in range(num_blocks):
-            q_start = cumsum[q_block]
-            q_end = q_start + variable_block_sizes[q_block]
-
-            for kv_block in range(num_blocks):
-                if block_sparse_mask[head, q_block, kv_block]:
-                    kv_start = cumsum[kv_block]
-                    kv_end = kv_start + variable_block_sizes[kv_block]
-                    full_mask[head, q_start:q_end, kv_start:kv_end] = True
-
-    return full_mask

csrc/attn/video_sparse_attn/MANIFEST.in

@@ -1,2 +0,0 @@
-recursive-include tk *
-include config_vsa.py

csrc/attn/video_sparse_attn/README.md

@@ -1,61 +0,0 @@
-
-
-# Attention Kernel Used in FastVideo
-
-## Video Sparse Attention (VSA)
-
-### Installation
-We support H100 (via TK) and any other GPU (via triton) for VSA.
-
-```bash
-pip install vsa
-``` 
-
-Install from source:
-
-```bash
-git submodule update --init --recursive
-python setup.py install
-```
-
-
-If you encounter error during installation, try below:
-Install C++20 for ThunderKittens:
-```bash
-sudo apt update
-sudo apt install gcc-11 g++-11
-
-sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
-
-sudo apt update
-sudo apt install clang-11
-```
-(If you use CUDA12.8)
-```bash
-export CUDA_HOME=/usr/local/cuda-12.8
-export PATH=${CUDA_HOME}/bin:${PATH} 
-export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
-```
-
-### Verify if you have successfully installed
-
-```bash
-# test numerical
-python ../tests/test_vsa.py
-# (For H100) test speed
-python ../benchmarks/bench_vsa_hopper.py
-```
-
-bench_vsa_hopper.py should print something like this:
-
-```bash
-Using topk=76 kv blocks per q block (out of 768 total kv blocks)
-
-=== BLOCK SPARSE ATTENTION BENCHMARK ===
-Block Sparse Forward  - TFLOPS: 5622.26
-Block Sparse Backward - TFLOPS: 3865.68
-```
-
-## Acknowledgement
-
-We learned or reuse code from FlexAtteniton, NATEN, and ThunderKittens.

csrc/attn/video_sparse_attn/config_vsa.py

@@ -1,15 +0,0 @@
-### ADD TO THIS TO REGISTER NEW KERNELS
-sources = {
-    'block_sparse': {
-        'source_files': {
-            'h100': 'vsa/block_sparse_h100.cu'
-        }
-    }
-}
-
-### WHICH KERNELS DO WE WANT TO BUILD?
-# (oftentimes during development work you don't need to redefine them all.)
-kernels = ['block_sparse']
-
-### WHICH GPU TARGET DO WE WANT TO BUILD FOR?
-target = 'h100'

csrc/attn/video_sparse_attn/setup.py

@@ -1,81 +0,0 @@
-import os
-import subprocess
-
-from config_vsa import kernels, sources, target
-from setuptools import find_packages, setup
-from torch.utils.cpp_extension import BuildExtension, CUDAExtension
-
-target = target.lower()
-
-# Package metadata
-PACKAGE_NAME = "vsa"
-VERSION = "0.0.3"
-AUTHOR = "Hao AI Lab"
-DESCRIPTION = "Video Sparse Attention Kernel Used in FastVideo"
-URL = "https://github.com/hao-ai-lab/FastVideo/tree/main/csrc/attn/video_sparse_attn"
-
-# Set environment variables
-tk_root = os.getenv('THUNDERKITTENS_ROOT', os.path.abspath(os.path.join(os.getcwd(), 'tk/')))
-python_include = subprocess.check_output(['python', '-c',
-                                          "import sysconfig; print(sysconfig.get_path('include'))"]).decode().strip()
-torch_include = subprocess.check_output([
-    'python', '-c',
-    "import torch; from torch.utils.cpp_extension import include_paths; print(' '.join(['-I' + p for p in include_paths()]))"
-]).decode().strip()
-print('vsa root:', tk_root)
-print('Python include:', python_include)
-print('Torch include directories:', torch_include)
-
-# CUDA flags
-cuda_flags = [
-    '-DNDEBUG', '-Xcompiler=-Wno-psabi', '-Xcompiler=-fno-strict-aliasing', '--expt-extended-lambda',
-    '--expt-relaxed-constexpr', '-forward-unknown-to-host-compiler', '--use_fast_math', '-std=c++20', '-O3',
-    '-Xnvlink=--verbose', '-Xptxas=--verbose', '-Xptxas=--warn-on-spills', f'-I{tk_root}/include',
-    f'-I{tk_root}/prototype', f'-I{python_include}', '-DTORCH_COMPILE'
-] + torch_include.split()
-cpp_flags = ['-std=c++20', '-O3']
-
-if target == 'h100':
-    cuda_flags.append('-DKITTENS_HOPPER')
-    cuda_flags.append('-arch=sm_90a')
-else:
-    raise ValueError(f'Target {target} not supported')
-
-source_files = ['vsa.cpp']
-for k in kernels:
-    if target not in sources[k]['source_files']:
-        raise KeyError(f'Target {target} not found in source files for kernel {k}')
-    if isinstance(sources[k]['source_files'][target], list):
-        source_files.extend(sources[k]['source_files'][target])
-    else:
-        source_files.append(sources[k]['source_files'][target])
-    cpp_flags.append(f'-DTK_COMPILE_{k.replace(" ", "_").upper()}')
-
-
-ext_modules = [
-    CUDAExtension('vsa_cuda',
-        sources=source_files,
-        extra_compile_args={
-            'cxx': cpp_flags,
-            'nvcc': cuda_flags
-        },
-        libraries=['cuda'])
-]
-
-
-
-setup(name=PACKAGE_NAME,
-      version=VERSION,
-      author=AUTHOR,
-      description=DESCRIPTION,
-      url=URL,
-      packages=find_packages(),
-      ext_modules=ext_modules,
-      cmdclass={'build_ext': BuildExtension},
-      classifiers=[
-          "Programming Language :: Python :: 3",
-          "Environment :: GPU :: NVIDIA CUDA :: 12",
-          "License :: OSI Approved :: Apache Software License",
-      ],
-      python_requires='>=3.10',
-      install_requires=["torch>=2.5.0"])

csrc/attn/video_sparse_attn/vsa.cpp

@@ -1,27 +0,0 @@
-#include <torch/extension.h>
-#include <ATen/ATen.h>
-
-#include <vector>
-#include <cuda_fp16.h>
-#include <cuda_bf16.h>
-
-#include <cuda_runtime.h>
-
-#ifdef TK_COMPILE_BLOCK_SPARSE
-extern std::vector<torch::Tensor> block_sparse_attention_forward(
-    torch::Tensor q, torch::Tensor k, torch::Tensor v,  torch::Tensor q2k_block_sparse_index, torch::Tensor q2k_block_sparse_num, torch::Tensor block_size
-); 
-extern std::vector<torch::Tensor> block_sparse_attention_backward(
-    torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o, torch::Tensor l_vec, torch::Tensor og, torch::Tensor k2q_block_sparse_index, torch::Tensor k2q_block_sparse_num, torch::Tensor block_size
-);
-#endif
-
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-    m.doc() = "Video Sparse Attention Kernels"; // optional module docstring
-
-#ifdef TK_COMPILE_BLOCK_SPARSE
-    m.def("block_sparse_fwd",  torch::wrap_pybind_function(block_sparse_attention_forward), "block sparse attention");
-    m.def("block_sparse_bwd",  torch::wrap_pybind_function(block_sparse_attention_backward), "block sparse attention backward");
-#endif
-}

csrc/attn/video_sparse_attn/vsa/__init__.py

@@ -1,80 +0,0 @@
-import torch
-from typing import Tuple
-block_sparse_attn=None
-import torch
-major, minor = torch.cuda.get_device_capability(0)
-if major == 9 and minor == 0:# check if H100
-    from vsa_cuda import block_sparse_fwd, block_sparse_bwd
-    from vsa.block_sparse_wrapper import block_sparse_attn_SM90
-    block_sparse_attn = block_sparse_attn_SM90
-else:
-    from vsa.block_sparse_wrapper import block_sparse_attn_triton
-    block_sparse_fwd = None
-    block_sparse_bwd = None
-    block_sparse_attn = block_sparse_attn_triton
-
-BLOCK_M = 64
-BLOCK_N = 64
-
-
-def torch_attention(q, k, v) -> Tuple[torch.Tensor, torch.Tensor]:
-    QK = torch.matmul(q, k.transpose(-2, -1))
-    QK /= (q.size(-1)**0.5)
-
-    # Causal mask removed since causal is always false
-
-    QK = torch.nn.functional.softmax(QK, dim=-1)
-    output = torch.matmul(QK, v)
-    return output, QK
-
-
-def video_sparse_attn(q, k, v, variable_block_sizes, topk, block_size, compress_attn_weight=None):
-    """
-    q: [batch_size, num_heads, seq_len, head_dim]
-    k: [batch_size, num_heads, seq_len, head_dim]
-    v: [batch_size, num_heads, seq_len, head_dim]
-    topk: int
-    block_size: int or tuple of 3 ints
-    video_shape: tuple of (T, H, W)
-    compress_attn_weight: [batch_size, num_heads, seq_len, head_dim]
-    select_attn_weight: [batch_size, num_heads, seq_len, head_dim]
-    NOTE: We assume q, k, v is zero padded!!
-    V1 of sparse attention. Include compress attn and sparse attn branch, use average pooling to compress. 
-    Assume q, k, v is flattened in this way: [batch_size, num_heads, T//block_size[0], H//block_size[1], W//block_size[2], block_size[0], block_size[1], block_size[2]]
-    """
-
-    if isinstance(block_size, int):
-        block_size = (block_size, block_size, block_size)
-
-    block_elements = block_size[0] * block_size[1] * block_size[2]
-    assert block_elements == 64
-    assert q.shape[2] % block_elements == 0
-    batch_size, num_heads, seq_len, head_dim = q.shape
-    # compress attn
-    q_compress = (q.view(batch_size, num_heads, seq_len // block_elements,
-                        block_elements, head_dim).float().sum(dim=3) / variable_block_sizes.view(1, 1, -1, 1)).to(q.dtype)
-    k_compress = (k.view(batch_size, num_heads, seq_len // block_elements,
-                        block_elements, head_dim).float().sum(dim=3) / variable_block_sizes.view(1, 1, -1, 1)).to(k.dtype)
-    v_compress = (v.view(batch_size, num_heads, seq_len // block_elements,
-                        block_elements, head_dim).float().sum(dim=3) / variable_block_sizes.view(1, 1, -1, 1)).to(v.dtype)
-
-    output_compress, block_attn_score = torch_attention(q_compress, k_compress,
-                                                        v_compress)
-
-    output_compress = output_compress.view(batch_size, num_heads,
-                                           seq_len // block_elements, 1,
-                                           head_dim)
-    output_compress = output_compress.repeat(1, 1, 1, block_elements,
-                                             1).view(batch_size, num_heads,
-                                                     seq_len, head_dim)
-
-    topK_indices = torch.topk(block_attn_score, topk, dim=-1).indices
-    block_mask = torch.zeros_like(block_attn_score, dtype=torch.bool).scatter_(-1, topK_indices, True)
-    output_select, _ = block_sparse_attn(q, k, v, block_mask, variable_block_sizes)
-
-    if compress_attn_weight is not None:
-        final_output = output_compress * compress_attn_weight + output_select
-    else:
-        final_output = output_compress + output_select
-    return final_output
-

csrc/attn/video_sparse_attn/vsa/block_sparse_wrapper.py

@@ -1,185 +0,0 @@
-import torch
-try:
-    from vsa_cuda import block_sparse_fwd, block_sparse_bwd
-except ImportError:
-    block_sparse_fwd = None
-    block_sparse_bwd = None
-from vsa.block_sparse_attn_triton import triton_block_sparse_attn_forward, triton_block_sparse_attn_backward
-assert torch.__version__ >= "2.4.0", "VSA requires PyTorch 2.4.0 or higher"
-from vsa.index import map_to_index
-from typing import Tuple, Optional
-
-
-
-@torch.library.custom_op("vsa::block_sparse_attn_triton", mutates_args=(), device_types="cuda")
-def block_sparse_attn_triton(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    v: torch.Tensor,
-    block_map: torch.Tensor,
-    variable_block_sizes: torch.Tensor,
-) -> Tuple[torch.Tensor, torch.Tensor]:
-    q = q.contiguous()
-    k = k.contiguous()
-    v = v.contiguous()
-    block_map = block_map.int()
-    q2k_block_sparse_index, q2k_block_sparse_num = map_to_index(block_map)
-    o, M = triton_block_sparse_attn_forward(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, variable_block_sizes)
-    return o, M
-
-
-@torch.library.register_fake("vsa::block_sparse_attn_triton")
-def _block_sparse_attn_triton_fake(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    v: torch.Tensor,
-    block_map: torch.Tensor,
-    variable_block_sizes: torch.Tensor,
-) -> Tuple[torch.Tensor, torch.Tensor]:
-    q = q.contiguous()
-    k = k.contiguous()
-    v = v.contiguous()
-    o = torch.empty_like(q)
-    M = torch.empty((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
-    return o, M
-
-
-@torch.library.custom_op("vsa::block_sparse_attn_backward_triton", mutates_args=(), device_types="cuda")
-def block_sparse_attn_backward_triton(
-    grad_output_padded: torch.Tensor, 
-    q_padded: torch.Tensor, 
-    k_padded: torch.Tensor, 
-    v_padded: torch.Tensor, 
-    o_padded: torch.Tensor, 
-    M: torch.Tensor, 
-    block_map: torch.Tensor,
-    variable_block_sizes: torch.Tensor,
-) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    grad_output_padded = grad_output_padded.contiguous()
-    q2k_block_sparse_index, q2k_block_sparse_num = map_to_index(block_map)
-    k2q_block_sparse_index, k2q_block_sparse_num = map_to_index(block_map.transpose(-1, -2))
-    dq, dk, dv = triton_block_sparse_attn_backward(grad_output_padded, q_padded, k_padded, v_padded, o_padded, M, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, variable_block_sizes)
-    return dq, dk, dv
-
-@torch.library.register_fake("vsa::block_sparse_attn_backward_triton")
-def _block_sparse_attn_backward_triton_fake(
-    grad_output_padded: torch.Tensor, 
-    q_padded: torch.Tensor, 
-    k_padded: torch.Tensor, 
-    v_padded: torch.Tensor, 
-    o_padded: torch.Tensor, 
-    M: torch.Tensor,
-    block_map: torch.Tensor,
-    variable_block_sizes: torch.Tensor,
-) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    grad_output_padded = grad_output_padded.contiguous()
-    dq = torch.empty_like(grad_output_padded)
-    dk = torch.empty_like(grad_output_padded)
-    dv = torch.empty_like(grad_output_padded)
-    return dq, dk, dv
-
-
-def backward_triton(ctx, grad_output1, grad_output2):
-    q_padded, k_padded, v_padded, o_padded, M, block_map, variable_block_sizes = ctx.saved_tensors
-    dq, dk, dv = block_sparse_attn_backward_triton(grad_output1, q_padded, k_padded, v_padded, o_padded, M, block_map, variable_block_sizes)
-    return dq, dk, dv, None, None
-
-
-def setup_context_triton(ctx, inputs, output):
-    q_padded, k_padded, v_padded, block_map, variable_block_sizes = inputs
-    o_padded, M = output
-    ctx.save_for_backward(q_padded, k_padded, v_padded, o_padded, M, block_map, variable_block_sizes)
-    
-block_sparse_attn_triton.register_autograd(backward_triton, setup_context=setup_context_triton)
-
-
-major, minor = torch.cuda.get_device_capability(0)
-
-if major == 9 and minor == 0:# check if H100
-    @torch.library.custom_op("vsa::block_sparse_attn_SM90", mutates_args=(), device_types="cuda")
-    def block_sparse_attn_SM90(
-        q_padded: torch.Tensor,
-        k_padded: torch.Tensor, 
-        v_padded: torch.Tensor, 
-        block_map: torch.Tensor,
-        variable_block_sizes: torch.Tensor, 
-        )-> Tuple[torch.Tensor, torch.Tensor]:
-        q_padded = q_padded.contiguous()
-        k_padded = k_padded.contiguous()
-        v_padded = v_padded.contiguous()
-        q2k_block_sparse_index, q2k_block_sparse_num = map_to_index(block_map)
-        variable_block_sizes = variable_block_sizes.int()
-        o_padded, lse_padded = block_sparse_fwd(q_padded, k_padded, v_padded, q2k_block_sparse_index, q2k_block_sparse_num, variable_block_sizes)
-        return o_padded, lse_padded
-
-
-
-
-    @torch.library.register_fake("vsa::block_sparse_attn_SM90")
-    def _block_sparse_attn_SM90_fake(
-        q_padded: torch.Tensor,
-        k_padded: torch.Tensor, 
-        v_padded: torch.Tensor, 
-        block_map: torch.Tensor,
-        variable_block_sizes: torch.Tensor, 
-    ) -> Tuple[torch.Tensor,  torch.Tensor]:
-        q_padded, k_padded, v_padded = [x.contiguous() for x in (q_padded, k_padded, v_padded)]
-        B, H, S, D = q_padded.shape
-        o_padded = torch.empty_like(q_padded)
-        lse_padded = torch.empty((B, H, S, 1), device=q_padded.device, dtype=torch.float32)
-        return o_padded, lse_padded
-
-
-    @torch.library.custom_op("vsa::block_sparse_attn_backward_SM90", mutates_args=(), device_types="cuda")
-    def block_sparse_attn_backward_SM90(
-        grad_output_padded: torch.Tensor, 
-        q_padded: torch.Tensor, 
-        k_padded: torch.Tensor, 
-        v_padded: torch.Tensor, 
-        o_padded: torch.Tensor, 
-        lse_padded: torch.Tensor, 
-        block_map: torch.Tensor,
-        variable_block_sizes: torch.Tensor, 
-        )-> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        grad_output_padded = grad_output_padded.contiguous()
-        k2q_block_sparse_index, k2q_block_sparse_num = map_to_index(block_map.transpose(-1, -2))
-        grad_q_padded, grad_k_padded, grad_v_padded = block_sparse_bwd(
-            q_padded, k_padded, v_padded, o_padded, lse_padded, grad_output_padded, k2q_block_sparse_index, k2q_block_sparse_num, variable_block_sizes
-        )
-        grad_q_padded = grad_q_padded.to(grad_output_padded.dtype)
-        grad_k_padded = grad_k_padded.to(grad_output_padded.dtype)
-        grad_v_padded = grad_v_padded.to(grad_output_padded.dtype)
-        return grad_q_padded, grad_k_padded, grad_v_padded
-
-    @torch.library.register_fake("vsa::block_sparse_attn_backward_SM90")
-    def _block_sparse_attn_backward_SM90_fake(
-        grad_output_padded: torch.Tensor, 
-        q_padded: torch.Tensor, 
-        k_padded: torch.Tensor, 
-        v_padded: torch.Tensor, 
-        o_padded: torch.Tensor, 
-        lse_padded: torch.Tensor, 
-        block_map: torch.Tensor,
-        variable_block_sizes: torch.Tensor, 
-    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
-        torch._check(grad_output_padded.dtype == torch.bfloat16)
-        torch._check(lse_padded.dtype == torch.float32)
-        grad_output_padded = grad_output_padded.contiguous()
-        dq = torch.empty_like(grad_output_padded)
-        dk = torch.empty_like(grad_output_padded)
-        dv = torch.empty_like(grad_output_padded)
-        return dq, dk, dv
-
-
-    def backward_SM90(ctx, grad_output1, grad_output2):
-        q_padded, k_padded, v_padded, o_padded, lse_padded, block_map, variable_block_sizes= ctx.saved_tensors
-        dq, dk, dv = block_sparse_attn_backward_SM90(grad_output1, q_padded, k_padded, v_padded, o_padded, lse_padded, block_map, variable_block_sizes)
-        return dq, dk, dv, None, None
-
-    def setup_context_SM90(ctx, inputs, output):
-        q_padded, k_padded, v_padded, block_map, variable_block_sizes = inputs
-        o_padded, lse_padded = output
-        ctx.save_for_backward(q_padded, k_padded, v_padded, o_padded, lse_padded, block_map, variable_block_sizes)
-
-
-    block_sparse_attn_SM90.register_autograd(backward_SM90, setup_context=setup_context_SM90)

csrc/attn/vmoba_attn/README.md

@@ -1,32 +0,0 @@
-# Attention Kernel Used in FastVideo
-
-## VMoBA: Mixture-of-Block Attention for Video Diffusion Models (VMoBA)
-
-### Installation
-Please ensure that you have installed FlashAttention version **2.7.1 or higher**, as some interfaces have changed in recent releases.
-
-### Usage
-
-You can use `moba_attn_varlen` in the following ways:
-
-**Install from source:**
-```bash
-python setup.py install
-```
-
-**Import after installation:**
-```python
-from vmoba import moba_attn_varlen
-```
-
-**Or import directly from the project root:**
-```python
-from csrc.attn.vmoba_attn.vmoba import moba_attn_varlen
-```
-
-### Verify if you have successfully installed
-
-```bash
-python csrc/attn/vmoba_attn/vmoba/vmoba.py
-```
-

csrc/attn/vmoba_attn/setup.py

@@ -1,26 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-from setuptools import find_packages, setup
-
-PACKAGE_NAME = "vmoba"
-VERSION = "0.0.0"
-AUTHOR = "JianzongWu"
-DESCRIPTION = "VMoBA: Mixture-of-Block Attention for Video Diffusion Models"
-URL = "https://github.com/KwaiVGI/VMoBA"
-
-setup(
-    name=PACKAGE_NAME,
-    version=VERSION,
-    author=AUTHOR,
-    description=DESCRIPTION,
-    url=URL,
-    packages=find_packages(),
-    classifiers=[
-        "Programming Language :: Python :: 3",
-        "License :: OSI Approved :: Apache Software License",
-    ],
-    python_requires='>=3.12',
-    install_requires=[
-        "flash-attn >= 2.7.1",
-    ]
-)

csrc/attn/vmoba_attn/vmoba/__init__.py

@@ -1,2 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from .vmoba import moba_attn_varlen, process_moba_input, process_moba_output

docker/Dockerfile.python3.10

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:12.8.0-devel-ubuntu22.04
+FROM nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04
 
 ENV DEBIAN_FRONTEND=noninteractive
 
@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
     uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp310-cp310-linux_x86_64.whl
 
 COPY . .
 
@@ -55,18 +55,12 @@ RUN source $HOME/.local/bin/env && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
 
-# Install STA (Sliding Tile Attention)
+# Install FastVideo Unified Kernel
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
+    cd fastvideo-kernel && \
     git submodule update --init --recursive && \
-    python setup.py install
+    ./build.sh
 
-# Install VSA 
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
 
 EXPOSE 22

docker/Dockerfile.python3.11

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:12.8.0-devel-ubuntu22.04
+FROM nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04
 
 ENV DEBIAN_FRONTEND=noninteractive
 
@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
     uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp311-cp311-linux_x86_64.whl
 
 COPY . .
 
@@ -55,18 +55,12 @@ RUN source $HOME/.local/bin/env && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
 
-# Install STA (Sliding Tile Attention)
+# Install FastVideo Unified Kernel
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
+    cd fastvideo-kernel && \
     git submodule update --init --recursive && \
-    python setup.py install
+    ./build.sh
 
-# Install VSA 
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
 
 EXPOSE 22

docker/Dockerfile.python3.12

@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
     uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp312-cp312-linux_x86_64.whl
 
 COPY . .
 
@@ -55,18 +55,11 @@ RUN source $HOME/.local/bin/env && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
 
-# Install STA (Sliding Tile Attention)
+# Install FastVideo Unified Kernel
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
+    cd fastvideo-kernel && \
     git submodule update --init --recursive && \
-    python setup.py install
-
-# Install VSA 
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
+    ./build.sh
 
 EXPOSE 22

docker/Dockerfile.python3.12.cuda12.9.1

@@ -55,18 +55,12 @@ RUN source $HOME/.local/bin/env && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
 
-# Install STA (Sliding Tile Attention)
+# Install FastVideo Unified Kernel
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
+    cd fastvideo-kernel && \
     git submodule update --init --recursive && \
-    python setup.py install
+    ./build.sh
 
-# Install VSA 
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
 
 EXPOSE 22

docker/rocm7_1.Dockerfile

@@ -0,0 +1,58 @@
+FROM rocm/pytorch:rocm7.1_ubuntu22.04_py3.10_pytorch_release_2.9.1
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+SHELL ["/bin/bash", "-c"]
+
+WORKDIR /FastVideo
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    wget \
+    git \
+    ca-certificates \
+    openssh-server \
+    zsh \
+    vim \
+    curl \
+    gcc-11 \
+    g++-11 \
+    clang-11 \
+    && rm -rf /var/lib/apt/lists/*
+
+# Set up C++20 compilers for ThunderKittens
+RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+
+# Install uv and source its environment
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh && \
+    echo 'source $HOME/.local/bin/env' >> /root/.bashrc
+
+# Copy just the pyproject.toml first to leverage Docker cache
+COPY pyproject_other.toml ./pyproject.toml
+
+# Create a dummy README to satisfy the installation
+RUN echo "# Placeholder" > README.md
+
+# Create and activate virtual environment with specific Python version and seed
+RUN source $HOME/.local/bin/env && \
+    uv venv --python 3.10 --seed /opt/venv && \
+    source /opt/venv/bin/activate && \
+    uv pip install --no-cache-dir --upgrade pip
+
+COPY . .
+
+# Install dependencies using uv and set up shell configuration
+RUN source $HOME/.local/bin/env && \
+    source /opt/venv/bin/activate && \
+    uv pip install --no-cache-dir -e .[rocm] && \
+    git config --unset-all http.https://github.com/.extraheader || true && \
+    echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
+    echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
+
+# Install FastVideo Unified Kernel
+RUN source $HOME/.local/bin/env && \
+    source /opt/venv/bin/activate && \
+    cd fastvideo-kernel && \
+    git submodule update --init --recursive && \
+    ./build.sh --rocm
+
+EXPOSE 22

docs/README.md

@@ -6,7 +6,7 @@ This directory contains the FastVideo documentation built with MkDocs.
 
 ```bash
 # Install dependencies
-pip install -r docs/requirements-mkdocs.txt
+pip install -r requirements-mkdocs.txt
 
 # Serve docs with live reload (recommended for development)
 mkdocs serve

docs/assets/custom.css

@@ -2,9 +2,11 @@
     writing-mode: sideways-lr;
     white-space: nowrap;
     max-width: 0;
-    p {
-       margin: 0;
-    }
+}
+
+/* Keep header cell paragraph content tight (avoid CSS nesting for compatibility) */
+.vertical-table-header th.head:not(.stub) p {
+    margin: 0;
 }
 
 /* Image sizing classes */

docs/attention/developer/index.md

@@ -0,0 +1,186 @@
+# Adding a New Attention Backend
+
+FastVideo allows integrating new attention mechanisms easily. This guide walks you through adding a new backend (e.g., `MyNewAttn`).
+
+## 1. Implement the Backend (Python)
+
+Create a new file in `fastvideo/attention/backends/` (e.g., `mynew_attn.py`).
+
+Your implementation should inherit from `AttentionBackend` defined in `abstract.py`.
+
+```python
+# fastvideo/attention/backends/mynew_attn.py
+import torch
+from .abstract import AttentionBackend
+# Import the context manager to access metadata (optional)
+from fastvideo.forward_context import get_forward_context
+
+# Import compiled kernel if applicable (see Section 2)
+try:
+    # Import from the top-level package
+    from fastvideo_kernel import my_compiled_attn_func
+except ImportError:
+    my_compiled_attn_func = None
+
+class MyNewAttnBackend(AttentionBackend):
+    def process_inputs(self, q, k, v, **kwargs):
+        # Pre-process inputs if necessary
+        return q, k, v
+
+    def forward(self, q, k, v, **kwargs):
+        # Optional: Access extra metadata passed via ForwardContext
+        # Only needed if your backend requires global state (e.g. window_size)
+        try:
+            context = get_forward_context()
+            metadata = context.attn_metadata
+            # Example: window_size = metadata.window_size
+        except (AssertionError, AttributeError):
+            # Handle case where context is not set (e.g. standard inference)
+            pass
+
+        if my_compiled_attn_func is not None:
+            return my_compiled_attn_func(q, k, v)
+        else:
+            # Fallback implementation (e.g., Triton or pure PyTorch)
+            return self.fallback_impl(q, k, v)
+```
+
+## 2. Passing Extra Information via ForwardContext (Optional)
+
+FastVideo uses a `ForwardContext` to pass global metadata (like current timestep, batch info, or custom attention configurations) to attention backends without changing the `forward` signature of every layer. **This is optional and only required if your backend needs dynamic per-step information.**
+
+To use this:
+
+1. **Set Context**: In your pipeline or generation loop, use the `set_forward_context` context manager.
+2. **Access Context**: Inside your attention backend, use `get_forward_context()`.
+
+See [`docs/attention/sta/index.md`](../sta/index.md) (Sliding Tile Attention) for an example of how complex configuration (window sizes) is passed this way.
+
+## 3. Adding Compiled Kernels (C++/CUDA)
+
+If your backend requires custom CUDA kernels, you need to add them to the `fastvideo-kernel` package.
+
+### A. Add Source Files
+
+Place your kernel implementation files in `fastvideo-kernel/csrc/attention/`.
+
+* `mynew_attn.cu` (CUDA implementation)
+* `mynew_attn.h` (Optional headers)
+
+### B. Register in Extension
+
+Update `fastvideo-kernel/csrc/common_extension.cpp` to expose your function to Python.
+
+```cpp
+// 1. Declare external function
+#ifdef COMPILE_MYNEW_ATTN
+extern torch::Tensor mynew_attn_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v);
+#endif
+
+// 2. Register in module
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    // ... other kernels ...
+
+#ifdef COMPILE_MYNEW_ATTN
+    m.def("mynew_attn_fwd", torch::wrap_pybind_function(mynew_attn_forward), "My New Attention Forward");
+#endif
+}
+```
+
+### C. Update CMakeLists.txt
+
+Update `fastvideo-kernel/CMakeLists.txt` to compile your new files.
+
+**Case 1: General CUDA Kernel (Runs on all GPUs)**
+Add your source file directly to `EXTENSION_SOURCES` and define the compilation flag.
+
+```cmake
+# Add to EXTENSION_SOURCES
+list(APPEND EXTENSION_SOURCES csrc/attention/mynew_attn.cu)
+
+# Add compilation definition for common_extension.cpp
+list(APPEND COMPILE_DEFS COMPILE_MYNEW_ATTN)
+```
+
+**Case 2: ThunderKittens Kernel (Hopper H100 Only)**
+If your kernel uses ThunderKittens (TK), it requires specific architecture flags (`sm_90a`). Add it inside the `ENABLE_TK_KERNELS` block.
+
+```cmake
+if(ENABLE_TK_KERNELS)
+    # Add source only if TK is enabled
+    list(APPEND EXTENSION_SOURCES csrc/attention/mynew_attn_tk.cu)
+    
+    # Add definition to guard registration
+    list(APPEND COMPILE_DEFS TK_COMPILE_MYNEW_ATTN)
+endif()
+```
+
+### D. Expose in Python Ops
+
+Update `fastvideo-kernel/python/fastvideo_kernel/ops.py` to make the function importable and handle fallbacks gracefully.
+
+```python
+# fastvideo-kernel/python/fastvideo_kernel/ops.py
+
+# Try to load C++ extension symbols
+try:
+    from fastvideo_kernel._C import fastvideo_kernel_ops
+    mynew_attn_fwd = getattr(fastvideo_kernel_ops, "mynew_attn_fwd", None)
+except ImportError:
+    mynew_attn_fwd = None
+
+def my_compiled_attn_func(q, k, v):
+    # Runtime check: use C++ kernel if available, else fallback
+    if mynew_attn_fwd is not None:
+        return mynew_attn_fwd(q, k, v)
+    else:
+        # Call Triton/Python fallback
+        return mynew_attn_triton(q, k, v)
+```
+
+### E. Expose in Package Init
+
+Update `fastvideo-kernel/python/fastvideo_kernel/__init__.py` to export the function.
+
+```python
+from fastvideo_kernel.ops import (
+    my_compiled_attn_func,
+    # ...
+)
+
+__all__ = [
+    "my_compiled_attn_func",
+    # ...
+]
+```
+
+## 4. Register the Backend
+
+Update `fastvideo/attention/backends/__init__.py` to export your new class.
+
+```python
+from .mynew_attn import MyNewAttnBackend
+```
+
+## 5. Platform Integration
+
+If your backend requires specific platform checks (e.g., checking for H100 support), handle that in `fastvideo/platforms/cuda.py` or within your backend's `__init__`.
+
+## 6. Add Documentation
+
+Create a new documentation page for your backend to explain its usage, installation (if custom kernels are needed), and features.
+
+1. **Create Directory**: `docs/attention/mynew_attn/`
+2. **Create Index**: `docs/attention/mynew_attn/index.md`
+3. **Update Navigation**: Add an entry to `mkdocs.yml` under the "Attention" tab.
+
+## Checklist
+
+* [ ] Created `fastvideo/attention/backends/mynew_attn.py`.
+* [ ] (Optional) Added CUDA kernels in `fastvideo-kernel/csrc/attention/`.
+* [ ] (Optional) Updated `common_extension.cpp` and `CMakeLists.txt`.
+* [ ] (Optional) Exposed kernel in `fastvideo-kernel/python/fastvideo_kernel/ops.py`.
+* [ ] (Optional) Exported kernel in `fastvideo-kernel/python/fastvideo_kernel/__init__.py`.
+* [ ] Implemented `forward` method respecting the standard signature.
+* [ ] Added unit tests in `tests/`.
+* [ ] Added documentation in `docs/attention/` and updated `mkdocs.yml`.

docs/attention/index.md

@@ -0,0 +1,55 @@
+# FastVideo Attention Kernels
+
+FastVideo provides highly optimized custom attention kernels to accelerate video generation.
+
+## Supported Kernels
+
+* **[Video Sparse Attention (VSA)](vsa/index.md)**: Sparse attention mechanism selecting top-k blocks.
+* **[Sliding Tile Attention (STA)](sta/index.md)**: Optimized attention for window-based video generation.
+* **Backend development guide**: See the developer guide at
+  [Attention Backend Development](../contributing/attention_backend.md).
+
+## General Build Instructions
+
+These instructions apply to building the `fastvideo-kernel` package from source, which includes both STA and VSA kernels.
+
+### Prerequisites
+
+* **PyTorch**: 2.5.0+
+* **CUDA**: 12.4+ (12.8 recommended for best performance)
+* **C++ Compiler**: GCC 11+ (C++20 support required for ThunderKittens)
+
+Install system dependencies:
+
+```bash
+sudo apt update
+sudo apt install -y gcc-11 g++-11 clang-11 ninja-build
+
+# Set gcc-11 as default
+sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+```
+
+Set up your CUDA environment variables (adjust version as needed):
+
+```bash
+export CUDA_HOME=/usr/local/cuda-12.8
+export PATH=${CUDA_HOME}/bin:${PATH}
+export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+```
+
+### Compile and Install
+
+Clone the repository and build the kernel:
+
+```bash
+# Clone recursively to get ThunderKittens submodule
+git clone https://github.com/hao-ai-lab/FastVideo.git
+cd FastVideo/fastvideo-kernel
+
+# Build and install
+./build.sh
+```
+
+The build script automatically detects your GPU architecture:
+* **H100 (sm_90a)**: Compiles optimized C++ ThunderKittens kernels.
+* **Other (A100, etc.)**: Skips C++ compilation; installs Python package with Triton kernels.

docs/attention/sta/index.md

@@ -0,0 +1,36 @@
+# Sliding Tile Attention (STA)
+
+Optimized attention for window-based video generation (e.g., HunyuanVideo).
+
+## Installation
+
+STA is included in the `fastvideo-kernel` package. See the [main Attention page](../index.md) for build instructions.
+
+## Usage
+
+```python
+from fastvideo_kernel import sliding_tile_attention
+
+# q, k, v: [batch_size, num_heads, seq_length, head_dim]
+# window_size: List of (t, h, w) tiles. Tile size is (6, 8, 8).
+# text_length: Number of text tokens (0-256)
+
+out = sliding_tile_attention(
+    q, k, v, 
+    window_size=[(3, 3, 3)],  # Example window
+    text_length=256
+)
+```
+
+## Citation
+
+If you use Sliding Tile Attention in your research, please cite:
+
+```bibtex
+@article{zhang2025fast,
+  title={Fast video generation with sliding tile attention},
+  author={Zhang, Peiyuan and Chen, Yongqi and Su, Runlong and Ding, Hangliang and Stoica, Ion and Liu, Zhengzhong and Zhang, Hao},
+  journal={arXiv preprint arXiv:2502.04507},
+  year={2025}
+}
+```

docs/attention/vsa/index.md

@@ -0,0 +1,38 @@
+# Video Sparse Attention (VSA)
+
+Sparse attention mechanism selecting top-k blocks.
+
+## Installation
+
+VSA is included in the `fastvideo-kernel` package. See the [main Attention page](../index.md) for build instructions.
+
+## Usage
+
+```python
+from fastvideo_kernel import video_sparse_attn
+
+# q, k, v: [batch_size, num_heads, seq_len, head_dim]
+# variable_block_sizes: Number of valid tokens per block
+# q_variable_block_sizes: Number of valid tokens per q block (can differ from KV for q/k of different lengths)
+# topk: Number of blocks to attend
+
+output = video_sparse_attn(
+    q, k, v, 
+    block_sizes,
+    block_sizes,
+    topk=32
+)
+```
+
+## Citation
+
+If you use Video Sparse Attention in your research, please cite:
+
+```bibtex
+@article{zhang2025vsa,
+  title={Vsa: Faster video diffusion with trainable sparse attention},
+  author={Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
+  journal={arXiv preprint arXiv:2505.13389},
+  year={2025}
+}
+```

docs/contributing/attention_backend.md

@@ -0,0 +1,176 @@
+# Attention Backend Development
+
+This guide is for contributors adding a new attention backend (kernel or
+implementation) to FastVideo. If you just want to use existing kernels or build
+fastvideo-kernel, see [Attention overview](../attention/index.md).
+
+## When you need this guide
+
+Use this guide when you are:
+
+- Adding a new attention kernel or algorithm.
+- Wiring an existing kernel into FastVideo's attention selection.
+- Extending attention support to a new platform.
+
+## 0) Choose a backend name and scope
+
+Pick a backend name in `UPPER_SNAKE_CASE` and decide where it should run.
+Example: `MY_NEW_ATTN`.
+
+You will use this name in:
+
+- `AttentionBackendEnum` (global list of backends).
+- `get_name()` in your backend class (match the enum name).
+- Platform selectors (CUDA/ROCm/MPS/NPU) to return your backend class.
+
+## 1) Add enum + platform selection
+
+1) Add your backend to `fastvideo/platforms/interface.py`:
+
+```python
+class AttentionBackendEnum(enum.Enum):
+    ...
+    MY_NEW_ATTN = enum.auto()
+```
+
+1) Register it in platform selection (example: CUDA). Update
+`fastvideo/platforms/cuda.py` inside `get_attn_backend_cls`:
+
+```python
+elif selected_backend == AttentionBackendEnum.MY_NEW_ATTN:
+    try:
+        from fastvideo.attention.backends.my_new_attn import MyNewAttnBackend
+        return "fastvideo.attention.backends.my_new_attn.MyNewAttnBackend"
+    except ImportError as e:
+        logger.error("Failed to import MY_NEW_ATTN backend: %s", str(e))
+        raise
+```
+
+If you want support on other platforms, add a similar branch in
+`fastvideo/platforms/rocm.py`, `fastvideo/platforms/mps.py`, or `fastvideo/platforms/npu.py`.
+
+## 2) Implement the backend
+
+Create `fastvideo/attention/backends/my_new_attn.py` and implement the required
+classes.
+
+Minimal skeleton (no custom metadata):
+
+```python
+import torch
+from dataclasses import dataclass
+from fastvideo.attention.backends.abstract import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+)
+
+class MyNewAttnBackend(AttentionBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "MY_NEW_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["MyNewAttnImpl"]:
+        return MyNewAttnImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["AttentionMetadata"]:
+        return AttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["AttentionMetadataBuilder"]:
+        return MyNewAttnMetadataBuilder
+
+
+@dataclass
+class MyNewAttnMetadata(AttentionMetadata):
+    current_timestep: int
+
+
+class MyNewAttnMetadataBuilder(AttentionMetadataBuilder):
+    def __init__(self) -> None:
+        pass
+
+    def prepare(self) -> None:
+        pass
+
+    def build(self, current_timestep: int, **kwargs):
+        return MyNewAttnMetadata(current_timestep=current_timestep)
+
+
+class MyNewAttnImpl(AttentionImpl):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        softmax_scale: float,
+        causal: bool = False,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.softmax_scale = softmax_scale
+        self.causal = causal
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: MyNewAttnMetadata,
+    ) -> torch.Tensor:
+        # Implement attention
+        return torch.nn.functional.scaled_dot_product_attention(
+            query.transpose(1, 2),
+            key.transpose(1, 2),
+            value.transpose(1, 2),
+            is_causal=self.causal,
+            scale=self.softmax_scale,
+        ).transpose(1, 2)
+```
+
+Optional:
+
+- Implement `preprocess_qkv` / `postprocess_output` if your kernel needs tiling
+  or reshaping.
+- Use `fastvideo.forward_context.get_forward_context()` if you need dynamic
+  per-step data (e.g., window sizes).
+- Set `accept_output_buffer = True` if your backend writes into a provided
+  output buffer.
+
+## 3) Wire into attention layers
+
+Backends are used by `LocalAttention` and `DistributedAttention`. These layers
+accept a `supported_attention_backends` tuple. If your backend should be
+eligible, update the call sites that construct these layers (search for
+`supported_attention_backends=`).
+
+## 4) Add compiled kernels (optional)
+
+If you have a custom CUDA kernel:
+
+1) Add sources in `fastvideo-kernel/csrc/attention/`.
+2) Register bindings in `fastvideo-kernel/csrc/common_extension.cpp`.
+3) Add to `fastvideo-kernel/CMakeLists.txt` (and any feature flags).
+4) Expose in `fastvideo-kernel/python/fastvideo_kernel/ops.py`.
+5) Export in `fastvideo-kernel/python/fastvideo_kernel/__init__.py`.
+
+Keep a Python/Triton fallback so the backend runs even when the kernel is not
+available.
+
+## 5) Testing and debugging
+
+- Add a small parity test or microbenchmark comparing to SDPA.
+- Force your backend with the env var:
+  `FASTVIDEO_ATTENTION_BACKEND=MY_NEW_ATTN`.
+- Check logs from `fastvideo/attention/selector.py` to confirm selection.
+
+## Checklist
+
+- [ ] Added enum entry in `fastvideo/platforms/interface.py`.
+- [ ] Implemented backend in `fastvideo/attention/backends/`.
+- [ ] Registered selection in platform(s).
+- [ ] Updated layer call sites to include the backend where appropriate.
+- [ ] Added tests and documentation.

docs/contributing/coding_agents.md

@@ -0,0 +1,480 @@
+# FastVideo + Coding Agents
+
+Coding agents are now strong at navigating large codebases and iterating fast
+with parity tests and examples. This guide shows how to use them to add new
+model pipelines and ship PRs in a production-grade video diffusion framework.
+
+FastVideo is a great project to contribute to, with production-grade
+infrastructure, active collaborations (including NVIDIA), and a pipeline design
+and inference architecture that has been forked by [SGLang’s
+multimodal generation stack](https://github.com/sgl-project/sglang/tree/main/python/sglang/multimodal_gen).
+
+Goal: run the new pipeline with a minimal script like
+`examples/inference/basic/basic.py`. In production, FastVideo can download
+models automatically via `HF_HOME`; for development, use local directories so
+agents can run scripts and tests deterministically. We standardize local paths
+as:
+
+- `official_weights/<model_name>/` for official checkpoints
+- `converted_weights/<model_name>/` if conversion is required
+
+## Tips when prompting the agent
+
+When prompting the agent, include:
+
+- This guide and the [FastVideo design overview](../design/overview.md).
+- Exact file paths to edit.
+- A closest reference example file in FastVideo.
+- Expected behavior and acceptance criteria.
+- Repro steps (command, inputs, logs).
+- Constraints (performance, memory, compatibility).
+- Local paths (e.g., `official_weights/<model_name>/` or
+  `converted_weights/<model_name>/`) for parity tests.
+
+## FastVideo structure at a glance
+
+Before diving in, scan these references:
+
+- [Contributing overview](overview.md) for environment/setup context.
+- [FastVideo design overview](../design/overview.md) for pipeline architecture, configs, and HF layout.
+
+FastVideo maps a Diffusers-style repo into a pipeline like:
+
+- `fastvideo/models/*`: model implementations (DiT, VAE, encoders, upsamplers).
+- `fastvideo/configs/models/*`: arch configs and `param_names_mapping` for
+  weight name translation.
+- `fastvideo/configs/pipelines/*`: pipeline wiring (component classes + names).
+- `fastvideo/configs/sample/*`: default runtime sampling parameters.
+- `fastvideo/pipelines/basic/*`: end-to-end pipeline logic built from stages.
+- `model_index.json`: the HF repo entrypoint that maps component names to
+  classes and weight files.
+- Component loading happens in `VideoGenerator.from_pretrained`, which reads
+  `model_index.json`, resolves configs, and loads weights.
+
+Minimal usage example (based on `examples/inference/basic/basic.py`):
+
+```python
+from fastvideo import VideoGenerator
+from fastvideo.configs.sample import SamplingParam
+
+model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # or official_weights/<model_name>/
+generator = VideoGenerator.from_pretrained(model_id, num_gpus=1)
+
+sampling = SamplingParam.from_pretrained(model_id)
+sampling.num_frames = 45
+video = generator.generate_video(
+    "A vibrant city street at sunset.",
+    sampling_param=sampling,
+    output_path="video_samples",
+    save_video=True,
+)
+```
+
+## Some questions to ask yourself before starting
+
+Answering these upfront clarifies the work and speeds up implementation.
+
+### Is the model already supported by SGLang's multimodal generation stack?
+If yes, you can port many components from SGLang. It is a FastVideo fork, so
+interfaces line up, but you still need to swap layers/modules to match
+FastVideo's architecture and attention stack.
+
+If not, implement the model directly in FastVideo.
+
+### Is there an official implementation of the model you are adding?
+
+If yes, use it as the numerical reference. For example, LTX‑2 has an official
+implementation here: https://github.com/Lightricks/LTX-2. Prefer official code
+even if Diffusers also has one.
+
+### Is there a HuggingFace repo for the model you are adding? Is it in Diffusers format?
+
+If yes, load it directly in FastVideo after setting tensor mapping rules in the
+config. Otherwise, convert the weights to Diffusers format. See [Weights and
+Diffusers format](../design/overview.md#weights-and-diffusers-format) for details.
+
+### Do I have official weights + local paths ready?
+
+Standardize local paths as:
+
+- `official_weights/<model_name>/` for official checkpoints
+- `converted_weights/<model_name>/` if conversion is required (can be created later)
+
+### What pipeline components are required for the model you are adding?
+
+Usually you need a transformer (DiT), VAE, text encoder, and tokenizer. Some
+models add extra components.
+
+### What tasks does the model support?
+
+Usually a video diffusion model supports text‑to‑video (T2V),
+image‑to‑video (I2V), and video‑to‑video (V2V). Some add extra tasks (two‑stage
+generation, keyframe interpolation), which require extra components.
+
+It's usually easiest to start with a T2V pipeline and add the other tasks later.
+
+You can refer to the [Pipeline system](../design/overview.md#pipeline-system)
+section for more details.
+
+### Am I able to generate videos with the official implementation?
+
+These videos and prompts are your reference. Once the FastVideo pipeline works,
+compare outputs to the official implementation. Due to seeding and other
+factors, outputs may not match exactly, but they should be comparable.
+
+## Workflow: adding a full pipeline
+
+This is an example workflow for adding a full model pipeline (model + configs +
+examples + tests). This guide is in active development; feedback is welcome.
+
+!!! note
+    If you get stuck, refer to existing models/pipelines in FastVideo or ask in Slack.
+
+### 0) Fetch official model's code and weights
+
+Purpose:
+
+- Keep official checkpoints and source code local so conversion, parity tests,
+  and reference runs are reproducible.
+- Clone the official repo so you can use it as a numerical reference.
+
+Action:
+
+- Download official weights into `official_weights/<model_name>/`
+  (Diffusers format or not).
+- Clone the official repo under the project root (e.g., `FastVideo/LTX-2/`).
+- If a Diffusers-format HF repo already exists, you can skip manual weight
+  handling and download it directly with
+  `scripts/huggingface/download_hf.py`.
+
+!!! note
+    This step is best done manually because large downloads can time out.
+    Example:
+    ```bash
+    python scripts/huggingface/download_hf.py \
+      --repo_id Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
+      --local_dir official_weights/Wan2.1-T2V-1.3B-Diffusers \
+      --repo_type model
+    ```
+  
+### 1) Implement the model + config mapping
+
+Purpose:
+
+- Model weights are a dictionary of named tensors (`state_dict`). If the names
+  don’t line up with FastVideo’s module names, weights won’t load correctly (or
+  will silently load into the wrong layer).
+- Official checkpoints often use different prefixes or module layouts than
+  FastVideo, so we translate names via the mapping (during load or conversion).
+- Mapping aligns three things:
+  1. the official implementation’s module names,
+  2. the checkpoint `state_dict` keys,
+  3. FastVideo’s model classes and layer naming conventions.
+- If names don’t align, weights won’t load; implement the FastVideo model and
+  define mapping rules first.
+
+Action:
+
+- Implement the FastVideo model + config mapping.
+  - Add/extend the model in `fastvideo/models/...` and config in
+    `fastvideo/configs/models/...` (including `param_names_mapping`).
+  - Reuse existing FastVideo layers/modules where possible.
+  - Use FastVideo’s attention layers:
+    - `DistributedAttention` only for full‑sequence self‑attention in the DiT.
+    - `LocalAttention` for cross‑attention and other attention layers.
+  - See the “Configuration System” and “Weights and Diffusers format” sections
+    in `docs/design/overview.md` for how these pieces connect.
+  - If you are using an agent, ask it to implement the model, config mapping,
+    and a parity test together so you can validate numerics immediately.
+
+!!! note
+    After the first component is aligned and parity‑tested, open a **DRAFT PR**
+    on FastVideo so the rest of the pipeline work can build on top of it.
+
+!!! note
+    If a Diffusers-format HF repo already exists and loads correctly, you can
+    skip conversion entirely (no conversion script needed) and just download it
+    with `scripts/huggingface/download_hf.py`. Otherwise, you may need a
+    conversion script + a `converted_weights/<model>/` staging directory.
+
+Example (key renaming via arch config mapping, Wan2.1‑style):
+
+```python
+# Official model (simplified) in the upstream repo.
+class OfficialWanTransformer(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.patch_embedding = torch.nn.Conv3d(16, 1536, kernel_size=2, padding=0)
+
+    def forward(self, x):
+        return self.patch_embedding(x)
+
+# FastVideo model (simplified) in fastvideo/models/dits/wanvideo.py
+class PatchEmbed(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.proj = torch.nn.Conv3d(16, 1536, kernel_size=2, padding=0)
+
+    def forward(self, x):
+        return self.proj(x)
+
+class WanTransformer3DModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.patch_embedding = PatchEmbed()
+
+    def forward(self, x):
+        return self.patch_embedding(x)
+
+# Mapping defined in a config (simplified; see the real mapping in
+# fastvideo/configs/models/dits/wanvideo.py)
+param_names_mapping = {
+    r"^patch_embedding\.(.*)$": r"patch_embedding.proj.\1",
+    r"^blocks\.(\d+)\.attn1\.to_q\.(.*)$": r"blocks.\1.to_q.\2",
+}
+
+def apply_regex_map(state_dict, mapping):
+    # Pseudocode: apply regex substitutions in order
+    ...
+
+# Official checkpoint keys (example)
+official = {
+    "patch_embedding.weight": ...,
+    "blocks.0.attn1.to_q.weight": ...,
+}
+
+# Apply mapping so keys match FastVideo modules
+converted = apply_regex_map(official, param_names_mapping)
+
+```
+
+Optional helper (print a few checkpoint keys quickly):
+
+```bash
+python - <<'PY'
+import safetensors.torch as st
+keys = list(st.load_file("official_weights/<model>/transformer/diffusion_pytorch_model.safetensors").keys())
+print(keys[:20])
+PY
+```
+
+Example agent prompt (task request):
+
+```
+Please add the Wan2.1 T2V 1.3B Diffusers pipeline to FastVideo:
+- Add a FastVideo native Wan2.1 DiT implementation + config mapping.
+- Make sure to use the existing FastVideo layers and attention modules where possible.
+- Add a parity test that loads the official model alongside the FastVideo model and compares outputs numerically with fixed seeds and inputs.
+
+Paths:
+  - Official repo: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+  - Local download: official_weights/Wan2.1-T2V-1.3B-Diffusers
+Mapping steps:
+  - Load the official DiT weights from
+    official_weights/Wan2.1-T2V-1.3B-Diffusers/transformer/diffusion_pytorch_model.safetensors.
+  - Instantiate the FastVideo DiT (`WanTransformer3DModel`) and compare
+    its `state_dict().keys()` to the official keys.
+  - Update `param_names_mapping` in
+    fastvideo/configs/models/dits/wanvideo.py to resolve missing/unexpected keys.
+  - Use `load_state_dict(strict=False)` during iteration to surface mismatches.
+```
+
+External examples of the same pattern:
+- SGLang uses prefix-based routing in its weight loader to map checkpoint keys
+  into internal submodules (e.g., stripping a top-level prefix before delegating).
+- vLLM includes model-specific renamers for certain checkpoints that adjust
+  key prefixes so weights match its internal naming.
+
+### 2) Test numerical alignment with the official implementation
+
+Purpose:
+
+- Verify that the FastVideo component is numerically aligned with the official
+  implementation.
+
+Action:
+
+- Add or reuse a numerical parity test that loads the official model and the
+  FastVideo model and compares outputs.
+- See examples in `tests/local_tests/` (e.g., `tests/local_tests/upsamplers/`)
+  and the commands in `tests/local_tests/README.md`.
+- If there are discrepancies, add opt‑in logging to both models and compare
+  activation summaries (layer output sums, per‑stage logs).
+- First align the loaded weights (validate `param_names_mapping`).
+- Then align forward outputs using fixed seeds and inputs.
+  - Start with `atol=1e-4, rtol=1e-4` in `assert_close`.
+  - Keep dtype consistent (bf16 if available; otherwise fp32).
+  - If attention parity is unstable, align backends (e.g.,
+    `FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA`).
+
+### 3) Repeat the process for each component
+
+If the model requires additional components, repeat Steps 1–2 for each one.
+For example, implement the VAE in `fastvideo/models/vaes/` and its config in
+`fastvideo/configs/models/vaes/`, then add parity coverage for it.
+
+### 4) Add a pipeline config + sample defaults
+
+Purpose:
+
+- `fastvideo/configs/pipelines/` describes pipeline wiring and model module
+  names.
+- `fastvideo/configs/sample/` defines default runtime parameters.
+
+Action:
+
+- Add a new pipeline config + sampling params.
+- Register them in `fastvideo/configs/pipelines/registry.py` and
+  `fastvideo/configs/sample/registry.py`.
+
+### 5) Wire pipeline stages
+
+Purpose:
+
+- `fastvideo/pipelines/basic/<pipeline>/` contains the actual pipeline logic.
+- `fastvideo/pipelines/stages/` holds reusable, testable stages.
+
+Action:
+
+- Build the pipeline using stages; keep new stages isolated and documented.
+- Prefer opt‑in flags for expensive or optional steps.
+
+### 6) Add pipeline‑level tests
+
+Purpose:
+
+- Ensure the end‑to‑end pipeline works and stays aligned as pieces evolve.
+
+Action:
+
+- Add a pipeline parity test under `tests/local_tests/pipelines/`.
+- See the [Testing Guide](testing.md) for test conventions.
+
+### 7) Add user‑facing examples
+
+Purpose:
+
+- `examples/inference/basic/` is the entry point for simple, runnable scripts.
+
+Action:
+
+- Provide a minimal “hello world” example plus advanced variations.
+- Use fixed seeds and stable prompts.
+- Run the example locally to confirm end‑to‑end behavior.
+
+### 8) Add SSIM tests for CI checks
+
+Purpose:
+
+- Ensure visual similarity stays within expected bounds for regression testing.
+- SSIM tests act as a higher‑level guardrail beyond unit/parity tests.
+
+Action:
+
+- Add SSIM tests under `fastvideo/tests/ssim/` and include reference videos
+  (see the structure in the Testing Guide).
+- Use stable prompts/seeds and document any GPU‑specific requirements.
+- Follow the [Testing Guide](testing.md) for reference video placement and
+  execution details.
+
+### 9) Document it
+
+Purpose:
+
+- `docs/` is where users find the new pipeline usage and limitations.
+
+Action:
+
+- Add a short doc page or update an existing one.
+- Mention any caveats (memory, speed, constraints).
+
+## Common pitfalls when porting models
+
+- **Attention backend mismatch**: parity can fail if the official model uses a
+  different attention backend (e.g., SDPA vs custom). Align backends before
+  debugging deeper issues.
+- **Patchifier shape mistakes**: wrong patchification or reshape lengths can
+  silently corrupt outputs. Validate patch shapes early.
+- **Mask handling**: attention masks must match the official behavior (padding,
+  causal masks, and broadcast shapes).
+- **Scheduler / sigma schedule mismatch**: even small differences in schedules
+  or timestep shapes can cause noticeable drift.
+
+## Diffusers vs manual conversion
+
+If a model already ships in Diffusers format (with a proper `model_index.json`),
+prefer downloading it directly and loading it via FastVideo. In that case:
+
+- You usually **do not need** a conversion script.
+- You still need a correct `param_names_mapping` if the internal module names
+  differ from FastVideo’s implementation.
+
+If the model does **not** have a Diffusers-format repo:
+
+- You will need a conversion script to rewrite `state_dict` keys into FastVideo
+  naming and stage the result (e.g., under `converted_weights/<model>/`).
+- You may still use the official repo for reference parity and debugging.
+
+In both cases, parity testing is required to validate correctness.
+
+If you want to publish a Diffusers‑style repo after conversion, use
+`scripts/checkpoint_conversion/create_hf_repo.py` to assemble a HuggingFace‑ready
+directory before uploading.
+
+## FAQ
+
+**Q: Why do we implement the FastVideo model before conversion?**  
+A: You can’t define the key‑mapping rules until the FastVideo module names are
+known. The implementation determines the target `state_dict` schema.
+
+**Q: Do we always need a conversion script?**  
+A: No. If a Diffusers‑format repo exists and loads correctly, download it and
+skip conversion.
+
+**Q: How do I figure out `param_names_mapping` quickly?**  
+A: Load the official weights, instantiate the FastVideo model, and diff
+`state_dict().keys()` on both sides. Add regex rules until missing/unexpected
+keys are resolved. Agents can help you with this.
+
+**Q: What if parity fails even after mapping?**  
+A: Align attention backends, sigma schedules, and timestep shapes first. Then
+add opt‑in activation logging to locate the first divergent layer.
+
+## Case study: LTX‑2 port (from PLAN.md)
+
+The LTX‑2 port in `PLAN.md` shows the real sequence of steps and backtracking
+that happened during integration. Use it as a reference for how parity work
+actually unfolds:
+
+- Ported components first (transformer, VAE, audio, text encoder).
+- Added parity tests per component; used SDPA for reference parity.
+- Added debug logging to compare per‑block activations and isolate divergence.
+- Fixed cross‑attention reshape and patch grid bounds issues after logging.
+- Aligned sigma schedule and masking behavior to match the official pipeline.
+
+Recommendation:
+
+- Keep raw step‑by‑step logs in your own local `PLAN.md` for large ports.
+
+## Worked example: Wan2.1 T2V 1.3B pipeline
+
+The Wan2.1 T2V 1.3B Diffusers pipeline is a good “standard” example for
+FastVideo integration.
+
+1. Verify model config + mapping.
+   - DiT mapping: `fastvideo/configs/models/dits/wanvideo.py`
+   - VAE: `fastvideo/models/vaes/wanvae.py`
+   - Text encoder: `fastvideo/models/encoders/t5.py`
+
+2. Parity test the core components.
+   - Example tests: `fastvideo/tests/transformers/test_wanvideo.py`,
+     `fastvideo/tests/vaes/test_wan_vae.py`,
+     `fastvideo/tests/encoders/test_t5_encoder.py`
+
+3. Pipeline wiring.
+   - Pipeline: `fastvideo/pipelines/basic/wan/wan_pipeline.py`
+   - Pipeline config: `fastvideo/configs/pipelines/wan.py`
+   - Sampling defaults: `fastvideo/configs/sample/wan.py`
+
+4. Minimal example.
+   - Script: `examples/inference/basic/basic.py`

docs/contributing/developer_env/runpod.md

@@ -1,16 +1,55 @@
 
 # 📦 Developing FastVideo on RunPod
 
-You can easily use the FastVideo Docker image as a custom container on [RunPod](https://www.runpod.io) for development or experimentation.
+You can easily use the FastVideo Pod Template on [RunPod](https://www.runpod.io) for development or experimentation.
 
 ## Creating a new pod
 
-Choose a GPU that supports CUDA 12.8
-
-Pick 1 or 2 L40S GPU(s)
+- Make sure you are using the correct RunPod account.
+![RunPod Account Selection](../../assets/images/runpod_account.png)
 
+- Use "Additional Filters" to select CUDA 12.8.
 ![RunPod CUDA selection](../../assets/images/runpod_cuda.png)
 
+- Click "Deploy" and Pick a single A40 or RTX 4090 GPU.
+![RunPod GPU Selection](../../assets/images/runpod_deploy.png)
+
+- Select the "FastVideo" or "fastvideo-dev" Pod Template.
+![RunPod Pod Template Selection](../../assets/images/runpod_create.png)
+
+- Set the Pod name to "<name>-<FastVideo>-<date>".
+
+- Finally, once the pod is deployed (will take a few minutes as the image is being pulled), you can SSH into it using "SSH exposed over TCP". You'll need to use the matching private ssh key you provided.
+![RunPod SSH](../../assets/images/runpod_ssh.png)
+
+## Working with the pod
+
+After SSH'ing into your pod, you'll find the correct `uv` environment already activated and you should be in /FastVideo/ directory. Make sure to use /FastVideo/ for all your work.
+
+To pull in the latest changes from the GitHub repo:
+
+```bash
+cd /FastVideo
+git pull
+```
+
+Run your development workflows as usual:
+
+```bash
+# Run linters
+pre-commit run --all-files
+
+# Run tests
+pytest tests/
+```
+
+Make sure to push your changes back to the GitHub repo as nothing will be saved to the pod when it is terminated.
+
+After you are done with your work, you can terminate the pod by clicking the "Terminate" and "Delete" buttons. Remember if the pod is not completely deleted, Runpod will keep charging you for it.
+
+## Extra Information:
+If you need to customize the pod template this section has some useful information. For the most part you can leave the defaults of the FastVideo Pod Template.
+
 When creating your pod template, use this image:
 
 ```
@@ -24,30 +63,3 @@ bash -c "apt update;DEBIAN_FRONTEND=noninteractive apt-get install openssh-serve
 ```
 
 ![RunPod template configuration](../../assets/images/runpod_template.png)
-
-After deploying, the pod will take a few minutes to pull the image and start the SSH service.
-
-![RunPod ssh](../../assets/images/runpod_ssh.png)
-
-## Working with the pod
-
-After SSH'ing into your pod, you'll find the `fastvideo-dev` Conda environment already activated.
-
-To pull in the latest changes from the GitHub repo:
-
-```bash
-cd /FastVideo
-git pull
-```
-
-`If you have a persistent volume and want to keep your code changes, you can move /FastVideo to /workspace/FastVideo, or simply clone the repository there.`
-
-Run your development workflows as usual:
-
-```bash
-# Run linters
-pre-commit run --all-files
-
-# Run tests
-pytest tests/
-```

docs/contributing/overview.md

@@ -1,72 +1,84 @@
 
 # 🛠️ Contributing to FastVideo
 
-Thank you for your interest in contributing to FastVideo. We want to make the process as smooth for you as possible and this is a guide to help get you started!
+Thank you for your interest in contributing to FastVideo. We want the process
+to be smooth and beginner‑friendly, whether you are adding a new pipeline,
+improving performance, or fixing a bug.
 
-Our community is open to everyone and welcomes any contributions no matter how large or small.
+## Quick prerequisites
 
-# Developer Environment:
-Do make sure you have CUDA 12.4 installed and supported. FastVideo currently only support Linux and CUDA GPUs, but we hope to support other platforms in the future.
+- **OS**: Linux is the primary development target (WSL can work).
+- **GPU**: NVIDIA GPU recommended for inference and training workflows.
+- **CUDA**: Use a recent CUDA 12.x toolchain (see the installation guide for
+  the current recommendation).
 
-We recommend using a fresh Python 3.10 Conda environment to develop FastVideo:
+For a full install checklist, see `docs/getting_started/installation/gpu.md`.
+
+## Local development (Conda + editable install)
 
 Install Miniconda:
 
-```
+```bash
 wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
 bash Miniconda3-latest-Linux-x86_64.sh
 source ~/.bashrc
 ```
 
-Create and activate a Conda environment for FastVideo:
+Create and activate a Conda environment:
 
-```
+```bash
 conda create -n fastvideo python=3.12 -y
 conda activate fastvideo
 ```
 
 Install `uv` (optional, but recommended):
 
-From instructions on [uv](https://astral.sh/uv/):
-
-```
+```bash
 curl -LsSf https://astral.sh/uv/install.sh | sh
-# or 
+# or
 wget -qO- https://astral.sh/uv/install.sh | sh
 ```
 
-Clone the FastVideo repository and go to the FastVideo directory:
+Clone the repo:
 
-```
+```bash
 git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
-
 ```
 
-Now you can install FastVideo and setup git hooks for running linting. By using `pre-commit`, the linters will run and have to pass before you'll be able to make a commit.
+Install FastVideo in editable mode and set up hooks:
 
 ```bash
 uv pip install -e .[dev]
 
-# Can also install flash-attn (optional)
-uv pip install flash-attn --no-build-isolation 
+# Optional: FlashAttention (builds native kernels)
+uv pip install flash-attn --no-build-isolation
 
-# Linting, formatting and static type checking
+# Linting, formatting, static typing
 pre-commit install --hook-type pre-commit --hook-type commit-msg
-
-# You can manually run pre-commit with
 pre-commit run --all-files
 
 # Unit tests
 pytest tests/
 ```
 
-If you are on a Hopper GPU, you should also install [FA3](https://github.com/Dao-AILab/flash-attention) for much better performance:
+If you are on a Hopper GPU, installing FlashAttention 3 can improve
+performance (see `docs/inference/optimizations.md`).
 
-```
-git clone https://github.com/Dao-AILab/flash-attention.git && cd flash-attention/hopper
+## Docker development (optional)
 
-# make sure you have ninja installed
-uv pip install ninja
+If you prefer a containerized environment, use the dev image documented in
+`docs/contributing/developer_env/docker.md`.
 
-python setup.py install
-```
+## Testing
+
+See the [Testing Guide](testing.md) for how to add and run tests in FastVideo.
+
+## Attention backend development
+
+If you are adding a new attention kernel or backend, follow
+[Attention Backend Development](attention_backend.md).
+
+## Contributing with coding agents
+
+For a step‑by‑step workflow on adding pipelines or components with coding
+agents, see `docs/contributing/coding_agents.md`.

docs/contributing/profiling.md

@@ -1,7 +1,7 @@
 # Profiling FastVideo
 
 !!! warning
-    Profiling is only intended for FastVideo developers and maintainers to understand the proportion of time spent in different parts of the codebase. **FastVideo end-users should never turn on profiling** as it will significantly slow down the inference.
+    Profiling is only intended for FastVideo developers and maintainers to understand the proportion of time spent in different parts of the codebase. **FastVideo end-users should never turn on profiling** as it will significantly slow down inference.
 
 ## Profiling with PyTorch
 
@@ -49,5 +49,5 @@ Traces can be visualized using <https://ui.perfetto.dev/>.
 ### Best Practices
 
 - Keep the profiled step count small; traces can be large and slow down job shutdown while the profiler flushes data.
-- After profiling, clean up trace directories to avoid filling disks.
+- After profiling, clean up trace directories to avoid filling disk storage.
 - When adding new regions, register them in `fastvideo.profiler` and wrap the corresponding code block with `with self.profiler_controller.region("your_region"):` or the `@profile_region` decorator.

docs/contributing/testing.md

@@ -0,0 +1,131 @@
+# Testing in FastVideo
+
+This guide explains how to add and run tests in FastVideo. The testing suite is divided into several categories to ensure correctness across components, training workflows, and inference quality.
+
+## Test Types
+
+* **Unit Tests**: Located in `fastvideo/tests/dataset`, `fastvideo/tests/entrypoints`, and `fastvideo/tests/workflow`. These test individual functions and classes.
+* **Component Tests**: Located in `fastvideo/tests/encoders`, `fastvideo/tests/transformers`, and `fastvideo/tests/vaes`. These verify the loading and basic functionality of model components.
+* **SSIM Tests**: Located in `fastvideo/tests/ssim`. These are regression tests that compare generated videos against reference videos using the Structural Similarity Index Measure (SSIM) to detect quality degradation.
+* **Training Tests**: Located in `fastvideo/tests/training`. These validate training loops, loss calculations, and specific training techniques like LoRA, Distillation, and VSA.
+* **Inference Tests**: Located in `fastvideo/tests/inference`. These test specialized inference pipelines and optimizations (e.g., STA, V-MoBA).
+
+For now, we will focus on **SSIM Tests**.
+
+## SSIM Tests
+
+SSIM tests are located in `fastvideo/tests/ssim`. These tests generate videos using specific models and parameters, and compare them against reference videos to ensure that changes in the codebase do not degrade generation quality or alter the output unexpectedly.
+
+!!! note
+    If you are adding an SSIM test, this serves as a safeguard. Any future code changes that break or cause errors with the specific arguments and configurations you defined will trigger a failure. Therefore, it is important to include multiple settings and arguments that cover the core features of your new pipeline to ensure robust regression testing.
+
+### Directory Structure
+
+```
+fastvideo/tests/ssim/
+├── <GPU>_reference_videos/   # Reference videos organized by GPU type (e.g., L40S_reference_videos)
+│   ├── <Model_Name>/
+│   │   ├── <Backend>/        # e.g., FLASH_ATTN, TORCH_SDPA
+│   │   │   └── <Video_File>
+├── test_causal_similarity.py
+├── test_inference_similarity.py
+├── update_reference_videos.sh
+└── ...
+```
+
+### Adding a New SSIM Test
+
+To add a new SSIM test, follow these steps:
+
+1. **Create or Update a Test File**: You can add a new test function to an existing file (like `test_inference_similarity.py`) or create a new one if testing a distinct category of models.
+
+2. **Define Model Parameters**: Define the configuration for the model you want to test. This includes model path, dimensions, inference steps, and other generation parameters. **Note:** Consider using lower `num_inference_steps` or reduced resolution (e.g., 480p instead of 720p) to keep test execution time reasonable, provided it doesn't compromise the test's ability to detect regression.
+
+   ```python
+   MY_MODEL_PARAMS = {
+       "num_gpus": 1,
+       "model_path": "organization/model-name",
+       "height": 480,
+       "width": 832,
+       "num_frames": 45,
+       "num_inference_steps": 20,
+       # ... other parameters
+   }
+   ```
+
+3. **Implement the Test Function**:
+   * Use `pytest.mark.parametrize` to run the test with different prompts, backends, and models.
+   * Set the attention backend environment variable.
+   * Initialize the `VideoGenerator`.
+   * Generate the video.
+   * Compare the generated video with the reference video using `compute_video_ssim_torchvision`.
+
+   Example structure:
+
+   ```python
+   @pytest.mark.parametrize("prompt", TEST_PROMPTS)
+   @pytest.mark.parametrize("ATTENTION_BACKEND", ["FLASH_ATTN"])
+   def test_my_model_similarity(prompt, ATTENTION_BACKEND):
+       # Setup output directories
+       # ...
+
+       # Initialize Generator
+       generator = VideoGenerator.from_pretrained(...)
+       generator.generate_video(prompt, ...)
+
+       # Compare with Reference
+       ssim_values = compute_video_ssim_torchvision(
+           reference_path, generated_path, use_ms_ssim=True
+       )
+       assert ssim_values[0] >= 0.98  # Threshold
+   ```
+
+4. **Reference Videos**:
+   * When running the test for the first time (or when updating the reference), the test will fail because the reference video is missing. The generated video will be saved in `fastvideo/tests/ssim/generated_videos`.
+   * Inspect the generated video to ensure it meets quality expectations.
+   * Move the generated video to the appropriate reference folder: `fastvideo/tests/ssim/<GPU>_reference_videos/<Model>/<Backend>/`.
+   * You can use the helper script `update_reference_videos.sh` to automate copying videos from `generated_videos` to `L40S_reference_videos`. Note: Check the script to ensure paths match your environment (it defaults to `L40S_reference_videos`).
+
+### Running Tests Locally
+
+To run the SSIM tests locally:
+
+```bash
+pytest fastvideo/tests/ssim/ -vs
+```
+
+Ensure you have the necessary GPUs available as defined in your test parameters.
+
+## Modal Workflow
+
+FastVideo uses [Modal](https://modal.com/) for running tests in a CI environment. The workflow scripts are located in `fastvideo/tests/modal/`.
+
+### `pr_test.py`
+
+The main entry point for CI tests is `fastvideo/tests/modal/pr_test.py`. This script defines Modal functions that execute the pytest suites on specific hardware (e.g., L40S, H100).
+
+### Updating Modal Configuration
+
+If you add a new test that requires:
+* **Different GPU Hardware**: You may need to change the `@app.function(gpu=...)` decorator.
+* **Longer Execution Time**: Increase the `timeout` parameter.
+* **New Environment Variables/Secrets**: Add them to `secrets=[...]` or the image environment. For example, if your model is gated on Hugging Face, ensure `HF_API_KEY` is passed.
+
+For SSIM tests, the `run_ssim_tests` function in `pr_test.py` currently runs:
+
+```python
+@app.function(gpu="L40S:2", image=image, timeout=2700, secrets=[modal.Secret.from_dict({"HF_API_KEY": os.environ.get("HF_API_KEY", "")})])
+def run_ssim_tests():
+    run_test("hf auth login --token $HF_API_KEY && pytest ./fastvideo/tests/ssim -vs")
+```
+
+If your new test file is inside `fastvideo/tests/ssim`, it will automatically be picked up by this command. However, ensure that the `gpu="L40S:2"` configuration is sufficient for your model. If your model requires more GPUs (e.g., 4 or 8), you might need to create a separate Modal function or update the existing one.
+
+### Workflow Scripts
+
+The shell script that triggers these tests in the CI pipeline is located at `.buildkite/scripts/pr_test.sh`. If you add a new test category (e.g., a new folder outside of `ssim`), you will need to:
+1. Add a new function in `fastvideo/tests/modal/pr_test.py`.
+2. Add a new case in `.buildkite/scripts/pr_test.sh` to handle the new test type.
+
+!!! note
+    If you are a maintainer, you'll need to finally manually update the workflow script in Buildkite. Otherwise, a maintainer will help you update.

docs/design/overview.md

@@ -1,403 +1,178 @@
-# 🔍 FastVideo Overview
+# FastVideo Architecture Overview
 
-This document outlines FastVideo's architecture for developers interested in framework internals or contributions. It serves as an onboarding guide for new contributors by providing an overview of the most important directories and files within the `fastvideo/` codebase.
+This document summarizes how FastVideo is structured and how a Diffusers-style
+model repo maps into a runnable pipeline. It is intended for contributors who
+need the high-level layout and key entrypoints, not every internal detail.
 
-## Table of Contents - Directory Structure and Files
+## FastVideo structure at a glance
 
-- [`fastvideo/pipelines/`](#design-pipeline-system) - Core diffusion pipeline components
-- [`fastvideo/models/`](#design-model-components) - Model implementations
-  - [`dits/`](#design-transformer-models) - Transformer-based diffusion models
-  - [`vaes/`](#design-vae-variational-auto-encoder) - Variational autoencoders
-  - [`encoders/`](#design-text-and-image-encoders) - Text and image encoders
-  - [`schedulers/`](#design-schedulers) - Diffusion schedulers
-- [`fastvideo/attention/`](#design-optimized-attention) - Optimized attention implementations
-- [`fastvideo/distributed/`](#design-distributed-processing) - Distributed computing utilities
-- [`fastvideo/layers/`](#design-tensor-parallelism) - Custom neural network layers
-- [`fastvideo/platforms/`](#design-platforms) - Hardware platform abstractions
-- [`fastvideo/worker/`](#design-executor-and-worker-abstractions) - Multi-GPU process management
-- [`fastvideo/fastvideo_args.py`](#design-fastvideo-args) - Argument handling
-- [`fastvideo/forward_context.py`](#design-forwardcontext) - Forward pass context management
-- `fastvideo/utils.py` - Utility functions
-- [`fastvideo/logger.py`](#design-logger) - Logging infrastructure
+FastVideo maps a Diffusers-style repo into a pipeline like this:
 
-## Core Architecture
+- `fastvideo/models/*`: model implementations (DiT, VAE, encoders, upsamplers).
+- `fastvideo/configs/models/*`: arch configs and `param_names_mapping` for
+  weight name translation.
+- `fastvideo/configs/pipelines/*`: pipeline wiring (component classes + names).
+- `fastvideo/configs/sample/*`: default runtime sampling parameters.
+- `fastvideo/pipelines/basic/*`: end-to-end pipelines.
+- `fastvideo/pipelines/stages/*`: reusable pipeline stages.
+- `fastvideo/models/loader/*`: component loaders for Diffusers-style repos.
+- `model_index.json`: HF repo entrypoint mapping component names to classes.
 
-FastVideo separates model components from execution logic with these principles:
-- **Component Isolation**: Models (encoders, VAEs, transformers) are isolated from execution (pipelines, stages, distributed processing)
-- **Modular Design**: Components can be independently replaced
-- **Distributed Execution**: Supports various parallelism strategies (Tensor, Sequence)
-- **Custom Attention Backends**: Components can support and use different Attention implementations
-- **Pipeline Abstraction**: Consistent interface across diffusion models
+Flow:
+`model_index.json` -> component loaders -> model modules -> pipeline stages ->
+sampling params.
 
-## FastVideoArgs
-
-The `FastVideoArgs` class in `fastvideo/fastvideo_args.py` serves as the central configuration system for FastVideo. It contains all parameters needed to control model loading, inference configuration, performance optimization settings, and more.
-
-Key features include:
-- **Command-line Interface**: Automatic conversion between CLI arguments and dataclass fields
-- **Configuration Groups**: Organized by functional areas (model loading, video params, optimization settings)
-- **Context Management**: Global access to current settings via `get_current_fastvideo_args()`
-- **Parameter Validation**: Ensures valid combinations of settings
-
-Common configuration areas:
-- **Model paths and loading options**: `model_path`, `trust_remote_code`, `revision`
-- **Distributed execution settings**: `num_gpus`, `tp_size`, `sp_size`
-- **Video generation parameters**: `height`, `width`, `num_frames`, `num_inference_steps`
-- **Precision settings**: Control computation precision for different components
-
-Example usage:
+Minimal usage (from `examples/inference/basic/basic.py`):
 
 ```python
-# Load arguments from command line
-fastvideo_args = prepare_fastvideo_args(sys.argv[1:])
+from fastvideo import VideoGenerator
+from fastvideo.configs.sample import SamplingParam
 
-# Access parameters
-model = load_model(fastvideo_args.model_path)
+model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # or official_weights/<model_name>/
+generator = VideoGenerator.from_pretrained(model_id, num_gpus=1)
 
-# Set as global context
-with set_current_fastvideo_args(fastvideo_args):
-    # Code that requires access to these arguments
-    result = generate_video()
-```
-
-## Pipeline System
-
-### `ComposedPipelineBase`
-
-This foundational class provides:
-
-- **Model Loading**: Automatically loads components from HuggingFace-Diffusers-compatible model directories
-- **Stage Management**: Creates and orchestrates processing stages
-- **Data Flow Coordination**: Ensures proper state flow between stages
-
-```python
-class MyCustomPipeline(ComposedPipelineBase):
-    _required_config_modules = [
-        "text_encoder", "tokenizer", "vae", "transformer", "scheduler"
-    ]
-    
-    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
-        # Pipeline-specific initialization
-        pass
-        
-    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
-        self.add_stage("input_validation_stage", InputValidationStage())
-        self.add_stage("text_encoding_stage", CLIPTextEncodingStage(
-            text_encoder=self.get_module("text_encoder"),
-            tokenizer=self.get_module("tokenizer")
-        ))
-        # Additional stages...
-```
-
-### Pipeline Stages
-Each stage handles a specific diffusion process component:
-- **Input Validation**: Parameter verification
-- **Text Encoding**: CLIP, LLaMA, or T5-based encoding
-- **Image Encoding**: Image input processing
-- **Timestep & Latent Preparation**: Setup for diffusion
-- **Denoising**: Core diffusion loop
-- **Decoding**: Latent-to-pixel conversion
-
-Each stage implements a standard interface:
-
-```python
-def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
-    # Process batch and update state
-    return batch
-```
-
-![Pipeline execution and data flow](../assets/images/pipeline.png)
-
-### ForwardBatch
-
-Defined in `fastvideo/pipelines/pipeline_batch_info.py`, `ForwardBatch` encapsulates the data payload passed between pipeline stages. It typically holds:
-
-- **Input Data**: Prompts, images, generation parameters
-- **Intermediate State**: Embeddings, latents, timesteps, accumulated during stage execution
-- **Output Storage**: Generated results and metadata
-- **Configuration**: Sampling parameters, precision settings
-
-This structure facilitates clear state transitions between stages.
-
-## Model Components
-
-The `fastvideo/models/` directory contains implementations of the core neural network models used in video diffusion:
-
-### Transformer Models
-
-Transformer networks perform the actual denoising during diffusion:
-
-- **Location**: `fastvideo/models/dits/`
-- **Examples**:
-  - `WanTransformer3DModel`
-  - `HunyuanVideoTransformer3DModel`
-
-Features include:
-- Text/image conditioning
-- Standardized interface for model-specific optimizations
-
-```python
-def forward(
-    self, 
-    latents,                    # [B, T, C, H, W]
-    encoder_hidden_states,      # Text embeddings
-    timestep,                   # Current diffusion timestep
-    encoder_hidden_states_image=None,  # Optional image embeddings
-    **kwargs
-):
-    # Perform denoising computation
-    return noise_pred  # Predicted noise residual
-```
-
-### VAE (Variational Auto-Encoder)
-
-VAEs handle conversion between pixel space and latent space:
-
-- **Location**: `fastvideo/models/vaes/`
-- **Examples**:
-  - `AutoencoderKLWan`
-  - `AutoencoderKLHunyuanVideo`
-
-These models compress image/video data to a more efficient latent representation (typically 4x-8x smaller in each dimension).
-
-FastVideo's VAE implementations include:
-- Efficient video batch processing
-- Memory optimization
-- Optional tiling for large frames
-- Distributed weight support
-
-### Text and Image Encoders
-
-Encoders process conditioning inputs into embeddings:
-
-- **Location**: `fastvideo/models/encoders/`
-- **Text Encoders**:
-  - `CLIPTextModel`
-  - `LlamaModel`
-  - `UMT5EncoderModel`
-- **Image Encoders**:
-  - `CLIPVisionModel`
-
-FastVideo implements optimizations such as:
-- Vocab parallelism for distributed processing
-- Caching for common prompts
-- Precision-tuned computation
-
-### Schedulers
-
-Schedulers manage the diffusion sampling process:
-
-- **Location**: `fastvideo/models/schedulers/`
-- **Examples**:
-  - `UniPCMultistepScheduler`
-  - `FlowMatchEulerDiscreteScheduler`
-
-These components control:
-- Diffusion timestep sequences
-- Noise prediction to latent update conversions
-- Quality/speed trade-offs
-
-```python
-def step(
-    self, 
-    model_output: torch.Tensor,
-    timestep: torch.LongTensor,
-    sample: torch.Tensor,
-    **kwargs
-) -> torch.Tensor:
-    # Process model output and update latents
-    # Return updated latents
-    return prev_sample
-```
-
-This diagram shows how models are discovered, validated, and loaded across entrypoints, executors, pipelines, and model loaders.
-
-![Model loading flow](../assets/images/load_models.png)
-
-## Optimized Attention
-
-The `fastvideo/attention/` directory contains optimized attention implementations crucial for efficient video diffusion:
-
-### Attention Backends
-Multiple implementations with automatic selection:
-- **FLASH_ATTN**: Optimized for supporting hardware
-- **TORCH_SDPA**: Built-in PyTorch scaled dot-product attention
-- **SLIDING_TILE_ATTN**: For very long sequences
-
-```python
-# Configure available attention backends for this layer
-self.attn = LocalAttention(
-    num_heads=num_heads,
-    head_size=head_dim,
-    causal=False,
-    supported_attention_backends=(_Backend.FLASH_ATTN, _Backend.TORCH_SDPA)
-)
-
-# Override via environment variable
-# export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
-```
-
-![Attention backend selector design](../assets/images/attention_backend.png)
-
-### Attention Patterns
-Supports various patterns with memory optimization techniques:
-- **Cross/Self/Temporal/Global-Local Attention**
-- Chunking, progressive computation, optimized masking
-
-## Distributed Processing
-
-The `fastvideo/distributed/` directory contains implementations for distributed model execution:
-
-### Tensor Parallelism
-
-Tensor parallelism splits model weights across devices:
-
-- **Implementation**: Through `RowParallelLinear` and `ColumnParallelLinear` layers
-- **Use cases**: Will be used by encoder models as their sequence lengths are shorter and enables efficient sharding.
-
-```python
-# Tensor-parallel layers in a transformer block
-from fastvideo.layers.linear import ColumnParallelLinear, RowParallelLinear
-
-# Split along output dimension
-self.qkv_proj = ColumnParallelLinear(
-    input_size=hidden_size,
-    output_size=3 * hidden_size,
-    bias=True,
-    gather_output=False
-)
-
-# Split along input dimension
-self.out_proj = RowParallelLinear(
-    input_size=hidden_size,
-    output_size=hidden_size,
-    bias=True,
-    input_is_parallel=True
+sampling = SamplingParam.from_pretrained(model_id)
+sampling.num_frames = 45
+video = generator.generate_video(
+    "A vibrant city street at sunset.",
+    sampling_param=sampling,
+    output_path="video_samples",
+    save_video=True,
 )
 ```
 
-### Sequence Parallelism
+## Configuration system
 
-Sequence parallelism splits sequences across devices:
+FastVideo uses typed configs to keep model definitions, pipeline wiring, and
+runtime parameters consistent:
 
-- **Implementation**: Through `DistributedAttention` and sequence splitting
-- **Use cases**: Long video sequences or high-resolution processing. Used by DiT models.
+- `fastvideo/configs/models/`: architecture definitions, layer shapes, and
+  `param_names_mapping` rules for key renaming.
+- `fastvideo/configs/pipelines/`: pipeline wiring and required components.
+- `fastvideo/configs/sample/`: default sampling parameters (steps, frames,
+  guidance scale, resolution, fps).
+- `fastvideo/configs/registry.py`: pipeline registry.
+- `fastvideo/configs/sample/registry.py`: sampling param registry.
 
-```python
-# Distributed attention for long sequences
-from fastvideo.attention import DistributedAttention
+`FastVideoArgs` (in `fastvideo/fastvideo_args.py`) provides runtime settings and
+is passed into pipeline construction and stages.
 
-self.attn = DistributedAttention(
-    num_heads=num_heads,
-    head_size=head_dim,
-    causal=False,
-    supported_attention_backends=(_Backend.SLIDING_TILE_ATTN, _Backend.FLASH_ATTN)
-)
+## Weights and Diffusers format
+
+FastVideo follows the HuggingFace Diffusers repo layout. This keeps loaders
+compatible with HF repos and makes it easy to add new components.
+
+Typical Diffusers repo:
+
+```
+<model-repo>/
+  model_index.json
+  scheduler/
+    scheduler_config.json
+  transformer/               # or unet/ for image models
+    config.json
+    diffusion_pytorch_model.safetensors
+  vae/
+    config.json
+    diffusion_pytorch_model.safetensors
+  text_encoder/
+    config.json
+    model.safetensors
+  tokenizer/
+    tokenizer_config.json
+    tokenizer.json
 ```
 
-### Communication Primitives
-Efficient distributed operations via AllGather, AllReduce, and synchronization mechanisms.
+Key points:
 
-Efficient communication primitives minimize distributed overhead:
+- `model_index.json` is the root map that tells FastVideo which components to
+  load and which classes implement them.
+- Each component lives in its own folder with a `config.json` and weights.
+- Weights are usually in `diffusion_pytorch_model.safetensors`.
 
-- **Sequence-Parallel AllGather**: Collects sequence chunks
-- **Tensor-Parallel AllReduce**: Combines partial results
-- **Distributed Synchronization**: Coordinates execution
+Note on tensor names:
 
-## Forward Context Management
+Official checkpoints often use different `state_dict` names than FastVideo's
+module layout. We translate tensor names via the DiT arch config mapping
+(`param_names_mapping` under `fastvideo/configs/models/dits/`). This is similar
+in spirit to name-translation layers used in systems like vLLM and SGLang.
 
-### ForwardContext
+Example HF repo (Wan 2.1 T2V 1.3B Diffusers):
 
-Defined in `fastvideo/forward_context.py`, `ForwardContext` manages execution-specific state *within* a forward pass, particularly for low-level optimizations. It is accessed via `get_forward_context()`.
-
-- **Attention Metadata**: Configuration for optimized attention kernels (`attn_metadata`)
-- **Profiling Data**: Potential hooks for performance metrics collection
-
-This context-based approach enables:
-- Dynamic optimization based on execution state (e.g., attention backend selection)
-- Step-specific customizations within model components
-
-Usage example:
-
-```python
-with set_forward_context(current_timestep, attn_metadata, fastvideo_args):
-    # During this forward pass, components can access context
-    # through get_forward_context()
-    output = model(inputs)
+```
+https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers/tree/main
 ```
 
-## Executor and Worker System
+Example `model_index.json` from that repo:
 
-The `fastvideo/worker/` directory contains the distributed execution framework:
-
-### Executor Abstraction
-
-FastVideo implements a flexible execution model for distributed processing:
-
-- **Executor Base Class**: An abstract base class defining the interface for all executors
-- **MultiProcExecutor**: Primary implementation that spawns and manages worker processes
-- **GPU Workers**: Handle actual model execution on individual GPUs
-
-The MultiProcExecutor implementation:
-1. Spawns worker processes for each GPU
-2. Establishes communication channels via pipes
-3. Coordinates distributed operations across workers
-4. Handles graceful startup and shutdown of the process group
-
-Each GPU worker:
-1. Initializes the distributed environment
-2. Builds the pipeline for the specified model
-3. Executes requested operations on its assigned GPU
-4. Manages local resources and communicates results back to the executor
-
-This design allows FastVideo to efficiently utilize multiple GPUs while providing a simple, unified interface for model execution.
-
-## Platforms
-
-The `fastvideo/platforms/` directory provides hardware platform abstractions that enable FastVideo to run efficiently on different hardware configurations:
-
-### Platform Abstraction
-
-FastVideo's platform abstraction layer enables:
-- **Hardware Detection**: Automatic detection of available hardware
-- **Backend Selection**: Appropriate selection of compute kernels
-- **Memory Management**: Efficient utilization of hardware-specific memory features
-
-The primary components include:
-- **Platform Interface**: Defines the common API for all platform implementations
-- **CUDA Platform**: Optimized implementation for NVIDIA GPUs
-- **Backend Enum**: Used throughout the codebase for feature selection
-
-Usage example:
-
-```python
-from fastvideo.platforms import current_platform, _Backend
-
-# Check hardware capabilities
-if current_platform.supports_backend(_Backend.FLASH_ATTN):
-    # Use FlashAttention implementation
-else:
-    # Fall back to standard implementation
+```json
+{
+  "_class_name": "WanPipeline",
+  "_diffusers_version": "0.33.0.dev0",
+  "scheduler": [
+    "diffusers",
+    "UniPCMultistepScheduler"
+  ],
+  "text_encoder": [
+    "transformers",
+    "UMT5EncoderModel"
+  ],
+  "tokenizer": [
+    "transformers",
+    "T5TokenizerFast"
+  ],
+  "transformer": [
+    "diffusers",
+    "WanTransformer3DModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderKLWan"
+  ]
+}
 ```
 
-The platform system is designed to be extensible for future hardware targets.
+How this maps to FastVideo:
 
-## Logger
-See [PR](https://github.com/hao-ai-lab/FastVideo/pull/356)
+- `WanPipeline` -> `fastvideo/pipelines/basic/wan/wan_pipeline.py`
+- `WanTransformer3DModel` -> `fastvideo/models/dits/wanvideo.py`
+- `AutoencoderKLWan` -> `fastvideo/models/vaes/wanvae.py`
+- `UMT5EncoderModel` -> `fastvideo/models/encoders/t5.py`
+- `T5TokenizerFast` -> loaded via HF in `fastvideo/models/loader/`
+- `UniPCMultistepScheduler` -> loaded via Diffusers scheduler utilities
+- Pipeline defaults -> `fastvideo/configs/pipelines/wan.py`
+- Sampling defaults -> `fastvideo/configs/sample/wan.py`
 
-*TODO*: (help wanted) Add an environment variable that disables process-aware logging.
+## Pipeline system
 
-## Contributing to FastVideo
+- `fastvideo/pipelines/basic/*` contains end-to-end pipelines for each model
+  family.
+- `fastvideo/pipelines/stages/*` contains reusable, testable stages.
+- Pipelines subclass `ComposedPipelineBase` and declare required components via
+  `_required_config_modules`.
+- `ForwardBatch` (in `fastvideo/pipelines/pipeline_batch_info.py`) carries
+  prompts, latents, timesteps, and intermediate state across stages.
 
-If you're a new contributor, here are some common areas to explore:
+## Model components
 
-1. **Adding a new model**: Implement new model types in the appropriate subdirectory of `fastvideo/models/`
-2. **Optimizing performance**: Look at attention implementations or memory management
-3. **Adding a new pipeline**: Create a new pipeline subclass in `fastvideo/pipelines/`
-4. **Hardware support**: Extend the `platforms` module for new hardware targets
+- DiT models: `fastvideo/models/dits/`
+- VAEs: `fastvideo/models/vaes/`
+- Text/image encoders: `fastvideo/models/encoders/`
+- Schedulers: `fastvideo/models/schedulers/`
+- Upsamplers: `fastvideo/models/upsamplers/`
+- Optional audio models: `fastvideo/models/audio/`
 
-When adding code, follow these practices:
-- Use type hints for better code readability
-- Add appropriate docstrings
-- Maintain the separation between model components and execution logic
-- Follow existing patterns for distributed processing
+## Attention and distributed execution
+
+- Attention backends live in `fastvideo/attention/` and can be selected via
+  `FASTVIDEO_ATTENTION_BACKEND`.
+- `LocalAttention` is used for cross-attention and most attention layers.
+- `DistributedAttention` is used for full-sequence self-attention in the DiT.
+- Tensor-parallel layers live in `fastvideo/layers/`.
+- Sequence/tensor parallel utilities live in `fastvideo/distributed/`.
+
+## Related docs
+
+- [Contributing overview](../contributing/overview.md)
+- [Coding agents workflow](../contributing/coding_agents.md)
+- [Testing guide](../contributing/testing.md)

docs/distillation/dmd.md

@@ -1,6 +1,6 @@
 # 🎯 Distillation
 
-We introduce a new finetuning strategy - **Sparse-distill**, which jointly integrates **[DMD](https://arxiv.org/abs/2405.14867)** and **[VSA](https://arxiv.org/abs/2505.13389)** in a single training process. This approach combines the benefits of both distillation to shorten diffusion steps and sparse attention to reduce attention computations, enabling much faster video generation.
+We introduce a new finetuning strategy - **Sparse-distill**, which jointly integrates **[DMD](https://arxiv.org/abs/2405.14867)** and **[VSA](https://arxiv.org/abs/2505.13389)** in a single training process. This approach combines the benefits of both distillation to shorten diffusion steps and sparse attention to reduce attention computation, enabling much faster video generation.
 
 ## 📊 Model Overview
 
@@ -13,7 +13,7 @@ We provide two distilled models:
 Both models are trained on **61×448×832** resolution but support generating videos with **any resolution** (1.3B  model mainly support 480P, 14B model support 480P and 720P, quality may degrade for different resolutions).
 
 ## ⚙️ Inference
-First install [VSA](https://hao-ai-lab.github.io/FastVideo/video_sparse_attention/installation.html). Set `MODEL_BASE` to your own model path and run:
+First install [VSA](../attention/vsa/index.md). Set `MODEL_BASE` to your own model path and run:
 
 ```bash
 bash scripts/inference/v1_inference_wan_dmd.sh

docs/getting_started/installation.md

@@ -11,15 +11,13 @@ FastVideo supports the following hardware platforms:
 ### Using pip
 
 ```bash
+# Create and activate a new conda environment
+conda create -n fastvideo python=3.12
+conda activate fastvideo
+
 pip install fastvideo
 ```
 
-### Using conda
-
-```bash
-conda install -c conda-forge fastvideo
-```
-
 ### From source
 
 ```bash
@@ -28,6 +26,12 @@ cd FastVideo
 pip install -e .
 ```
 
+Also optionally install flash-attn:
+
+```bash
+pip install flash-attn --no-build-isolation
+```
+
 ## Hardware Requirements
 
 - **NVIDIA GPUs**: CUDA 11.8+ with compute capability 7.0+
@@ -38,4 +42,4 @@ pip install -e .
 
 - [Quick Start Guide](quick_start.md) - Get started with your first video generation
 - [Configuration](../inference/configuration.md) - Learn about configuration options
-- [Examples](../inference/examples/) - Explore example scripts and notebooks
+- [Examples](../inference/examples/examples_inference_index.md) - Explore example scripts and notebooks

docs/getting_started/installation/gpu.md

@@ -84,12 +84,12 @@ pip install flash-attn --no-build-isolation
 
 ## Set up using Docker
 We also have prebuilt docker images with FastVideo dependencies pre-installed:
-[Docker Images](#docker)
+[Docker Images](../../contributing/developer_env/docker.md)
 
 ## Development Environment Setup
 
 If you're planning to contribute to FastVideo please see the following page:
-[Contributor Guide](#developer-overview)
+[Contributor Guide](../../contributing/overview.md)
 
 ## Hardware Requirements
 

docs/getting_started/installation/mps.md

@@ -78,7 +78,7 @@ uv pip install -e .
 ## Development Environment Setup
 
 If you're planning to contribute to FastVideo please see the following page:
-[Contributor Guide](#developer-overview)
+[Contributor Guide](../../contributing/overview.md)
 
 ## Hardware Requirements
 

docs/getting_started/quick_start.md

@@ -7,44 +7,72 @@ Get up and running with FastVideo in minutes!
 First, install FastVideo:
 
 ```bash
+# Create and activate a new conda environment
+conda create -n fastvideo python=3.12
+conda activate fastvideo
+
+# Install FastVideo
 pip install fastvideo
 ```
 
+Also optionally install flash-attn:
+
+```bash
+pip install flash-attn --no-build-isolation
+```
+
 ## Basic Usage
 
 ### Text-to-Video Generation
 
 ```python
-from fastvideo import FastVideoPipeline
+from fastvideo import VideoGenerator
 
-# Initialize the pipeline
-pipe = FastVideoPipeline.from_pretrained("wan2.1-t2v-1.3B")
+def main():
+    # Create a video generator with a pre-trained model
+    generator = VideoGenerator.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        num_gpus=1,  # Adjust based on your hardware
+    )
 
-# Generate a video
-prompt = "A cat playing with a ball of yarn"
-video = pipe(prompt, num_frames=16, height=512, width=512)
+    # Define a prompt for your video
+    prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest."
 
-# Save the video
-video.save("output.mp4")
+    # Generate the video
+    video = generator.generate_video(
+        prompt,
+        return_frames=True,  # Also return frames from this call (defaults to False)
+        output_path="my_videos/",  # Controls where videos are saved
+        save_video=True
+    )
+
+if __name__ == '__main__':
+    main()
 ```
 
 ### Image-to-Video Generation
 
 ```python
-from fastvideo import FastVideoPipeline
-from PIL import Image
+from fastvideo import VideoGenerator, SamplingParam
 
-# Load an image
-image = Image.open("input.jpg")
+def main():
+    # Create the generator
+    model_name = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+    generator = VideoGenerator.from_pretrained(model_name, num_gpus=1)
 
-# Initialize the pipeline
-pipe = FastVideoPipeline.from_pretrained("wan2.1-i2v-14B-480p")
+    # Set up parameters with an initial image
+    sampling_param = SamplingParam.from_pretrained(model_name)
+    sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+    sampling_param.num_frames = 107
 
-# Generate a video from the image
-video = pipe(image, num_frames=16, height=480, width=480)
+    # Generate video based on the image
+    prompt = "A photograph coming to life with gentle movement"
+    generator.generate_video(prompt, sampling_param=sampling_param,
+                             output_path="my_videos/",
+                             save_video=True)
 
-# Save the video
-video.save("output.mp4")
+if __name__ == '__main__':
+    main()
 ```
 
 ## Next Steps

docs/index.md

@@ -24,12 +24,13 @@ FastVideo is an inference and post-training framework for diffusion models. It f
 ## Key Features
 
 FastVideo has the following features:
+
 - State-of-the-art performance optimizations for inference
   - [Sliding Tile Attention](https://arxiv.org/pdf/2502.04507)
   - [TeaCache](https://arxiv.org/pdf/2411.19108)
   - [Sage Attention](https://arxiv.org/abs/2410.02367)
 - E2E post-training support
-  - Data preprocessing pipeline for video data.
+  - Data preprocessing pipeline for video data
   - [Sparse distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/) for Wan2.1 and Wan2.2 using [Video Sparse Attention](https://arxiv.org/pdf/2505.13389) and [Distribution Matching Distillation](https://tianweiy.github.io/dmd2/)
   - Support full finetuning and LoRA finetuning for state-of-the-art open video DiTs.
   - Scalable training with FSDP2, sequence parallelism, and selective activation checkpointing, with near linear scaling to 64 GPUs.
@@ -42,7 +43,7 @@ Use the navigation menu on the left to explore different sections:
 
 - **Getting Started**: Installation and quick start guides
 - **Inference**: Learn how to use FastVideo for video generation
-- **Training**: Data preprocessing and fine-tuning workflows  
+- **Training**: Data preprocessing and fine-tuning workflows
 - **Distillation**: Post-training optimization techniques
 - **Sliding Tile Attention**: Advanced attention mechanisms
 - **Video Sparse Attention**: Efficient attention for video models

docs/inference/add_pipeline.md

@@ -45,6 +45,7 @@ FastVideo uses the Hugging Face Diffusers format for model organization:
 ### Implementing Modules
 
 Place new modules in the appropriate directories:
+
 - Encoders: `fastvideo/models/encoders/`
 - VAEs: `fastvideo/models/vaes/`
 - Transformer models: `fastvideo/models/dits/`
@@ -53,12 +54,15 @@ Place new modules in the appropriate directories:
 ### Adapting Model Layers
 
 #### Layer Replacements
+
 Replace standard PyTorch layers with FastVideo optimized versions:
+
 - nn.LayerNorm → fastvideo.layers.layernorm.RMSNorm
 - Embedding layers → fastvideo.layers.vocab_parallel_embedding modules
 - Activation functions → versions from fastvideo.layers.activation
 
 #### Distributed Linear Layers
+
 Use appropriate parallel layers for distribution:
 
 ```python
@@ -91,6 +95,7 @@ self.out_proj = RowParallelLinear(
 ```
 
 ### Attention Layers
+
 Replace standard attention with FastVideo's optimized attention:
 
 ```python
@@ -304,6 +309,7 @@ EntryClass = [MyCustomPipeline, MyOtherPipeline]
 ```
 
 The registry will automatically:
+
 1. Scan all packages under `fastvideo/pipelines/`
 2. Look for `EntryClass` variables
 3. Register pipelines using their class names as identifiers

docs/inference/cli.md

@@ -1,7 +1,7 @@
 # FastVideo CLI Inference
 
 The FastVideo CLI provides a quick way to access the FastVideo inference pipeline for video generation. For more advanced usage,
-see the Python interface [here](https://hao-ai-lab.github.io/FastVideo/inference/examples/basic.html).
+see the Python interface [here](examples/basic.md).
 
 ## Basic Usage
 

docs/inference/configuration.md

@@ -74,4 +74,4 @@ if __name__ == '__main__':
 
 ## Performance Optimization
 
-For configuring optimizations, please see our [optimizations guide](#inference-optimizations)
+For configuring optimizations, please see our [optimizations guide](optimizations.md)

docs/inference/inference_quick_start.md

@@ -3,6 +3,7 @@
 This page contains step-by-step instructions to get you quickly started with video generation using FastVideo.
 
 ## Requirements
+
 - **OS**: Linux (Tested on Ubuntu 22.04+)
 - **Python**: 3.10-3.12
 - **CUDA**: 12.8
@@ -21,9 +22,10 @@ conda activate fastvideo
 pip install fastvideo
 ```
 
-For advanced installation options, see the [Installation Guide](installation.md).
+For advanced installation options, see the [Installation Guide](../getting_started/installation.md).
 
 ## Generating Your First Video
+
 Here's a minimal example to generate a video using the default settings. Create a file called `example.py` with the following code:
 
 ```python
@@ -60,9 +62,10 @@ python example.py
 The generated video will be saved in the current directory under `my_videos/`  
 
 More inference example scripts can be found in `scripts/inference/`
+
 ## Available Models
 
-Please see the [support matrix](#support-matrix) for the list of supported models and their available optimizations.
+Please see the [support matrix](support_matrix.md) for the list of supported models and their available optimizations.
 
 ## Image-to-Video Generation
 
@@ -96,20 +99,28 @@ if __name__ == '__main__':
 Common issues and their solutions:
 
 ### Out of Memory Errors
+
 If you encounter CUDA out of memory errors:
+
 - Reduce `num_frames` or video resolution
 - Enable memory optimization with `enable_model_cpu_offload`
 - 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
+
 To speed up generation:
+
 - Reduce `num_inference_steps` (20-30 is usually sufficient)
 - Use half precision (`fp16`) for the VAE
 - Use multiple GPUs if available
 
 ### Unexpected Results
+
 If the generated video doesn't match your prompt:
+
 - Try increasing `guidance_scale` (7.0-9.0 works well)
 - Make your prompt more detailed and specific
 - Experiment with different random seeds
@@ -117,8 +128,8 @@ If the generated video doesn't match your prompt:
 
 ## Next Steps
 
-- Learn about [Advanced Inference Configurations](#inference-configuration)
-- Learn about using [Optimizations](#inference-optimizations)
-- See [Examples](../examples/examples_inference_index.md) for more usage scenarios
+- Learn about [Advanced Inference Configurations](configuration.md)
+- Learn about using [Optimizations](optimizations.md)
+- See [Examples](examples/examples_inference_index.md) for more usage scenarios
 - Join our [Community Discord](https://discord.gg/JA7cksDz86).
 - Join our [Community Slack](https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ).

docs/inference/offloading.md

@@ -0,0 +1,140 @@
+# Offloading
+
+This page describes how to use offloading techniques for inference to reduce GPU memory usage while maintaining acceptable performance.
+
+## Default Behavior
+
+```python
+dit_cpu_offload: bool = True
+use_fsdp_inference: bool = False
+dit_layerwise_offload: bool = True
+text_encoder_cpu_offload: bool = True
+image_encoder_cpu_offload: bool = True
+vae_cpu_offload: bool = True
+pin_cpu_memory: bool = True
+```
+
+## Behavior Explanation
+
+!!! note
+    For CLI usage, replace underscores (`_`) with hyphens (`-`).
+
+### `use_fsdp_inference`
+
+Enables [FSDP](https://docs.pytorch.org/tutorials/intermediate/FSDP_tutorial.html) for inference. The model weights are sharded across multiple GPUs to reduce memory usage per GPU, and weights are broadcast to all GPUs layer by layer during inference.
+
+#### Performance Impact
+
+FSDP inference introduces negligible performance overhead due to weight prefetching. Performance overhead may be visible when GPU interconnect is slow (e.g., multiple consumer-level GPUs connected by slow PCIe without GPU P2P support).
+
+#### Usage Recommendation
+
+We recommend enabling this option when multiple GPUs are available.
+
+### `dit_cpu_offload`
+
+Enables CPU offloading for FSDP inference. When enabled, the model weights are offloaded to CPU memory, and the weight of each layer is moved to GPU memory only when that layer is being computed.
+
+#### Performance Impact
+
+The PyTorch FSDP implementation does not overlap computation and data transfer perfectly for inference, so enabling this option will harm performance.
+
+#### Usage Recommendation
+
+This option only takes effect when FSDP is enabled. For single GPU usage, we recommend using `dit_layerwise_offload` instead.
+
+### `dit_layerwise_offload`
+
+This option is similar to `dit_cpu_offload`, but with two key differences:
+
+1. It overlaps computation and PCIe data transfer
+2. It only works for single GPU inference
+
+#### Performance Impact
+
+This option introduces negligible performance overhead.
+
+#### Usage Recommendation
+
+We recommend enabling this option for single GPU usage. This option is not compatible with FSDP.
+
+### `text_encoder_cpu_offload`
+
+When enabled, the text encoder model weights are offloaded to CPU memory, and text encoding is computed on CPU.
+
+#### Performance Impact
+
+This option significantly slows down text encoding computation, but text encoding is usually not the bottleneck.
+
+#### Usage Recommendation
+
+We recommend enabling this option only when OOM happens.
+
+### `image_encoder_cpu_offload` and `vae_cpu_offload`
+
+When enabled, the weights are stored in CPU memory and moved to GPU memory when the corresponding module is being computed. After computation, the weights are moved back to CPU memory.
+
+#### Performance Impact
+
+These options introduce performance overhead due to PCIe data transfer.
+
+#### Usage Recommendation
+
+We recommend enabling these options when OOM happens.
+
+## General Recommendations
+
+### Single GPU Inference
+
+We recommend enabling `dit_layerwise_offload`. If OOM happens, also enable `image_encoder_cpu_offload` and `vae_cpu_offload`. If OOM still happens, consider enabling `text_encoder_cpu_offload`.
+
+### Multi-GPU Inference
+
+We recommend enabling `use_fsdp_inference` and disabling both `dit_layerwise_offload` and `dit_cpu_offload`. If OOM happens, consider enabling `text_encoder_cpu_offload`, `image_encoder_cpu_offload`, and `vae_cpu_offload`. If OOM still happens, consider enabling `dit_cpu_offload`.
+
+## Examples
+
+### Single GPU with Layerwise Offloading
+
+```python
+from fastvideo import VideoGenerator
+
+generator = VideoGenerator.from_pretrained(
+    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+    num_gpus=1,
+    # Recommended for single GPU
+    dit_layerwise_offload=True,
+    # Enable if OOM happens
+    vae_cpu_offload=True,
+    image_encoder_cpu_offload=True,
+    text_encoder_cpu_offload=True,
+    # Speeds up CPU-GPU transfer
+    pin_cpu_memory=True,
+)
+
+prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers."
+video = generator.generate_video(prompt, output_path="output/", save_video=True)
+```
+
+### Multi-GPU with FSDP
+
+```python
+from fastvideo import VideoGenerator
+
+generator = VideoGenerator.from_pretrained(
+    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+    num_gpus=2,
+    # Recommended for multi-GPU
+    use_fsdp_inference=True,
+    dit_layerwise_offload=False,
+    dit_cpu_offload=False,
+    # Enable if OOM happens
+    vae_cpu_offload=True,
+    image_encoder_cpu_offload=True,
+    text_encoder_cpu_offload=True,
+    pin_cpu_memory=True,
+)
+
+prompt = "A majestic lion strides across the golden savanna."
+video = generator.generate_video(prompt, output_path="output/", save_video=True)
+```

docs/inference/optimizations.md

@@ -7,13 +7,13 @@ This page describes the various options for speeding up generation times in Fast
 
 - Optimized Attention Backends
 
-  - [Flash Attention](#optimizations-flash)
-  - [Sliding Tile Attention](#optimizations-sta)
-  - [Sage Attention](#optimizations-sage)
-  - [Sage Attention 3](#optimizations-sage3)
+  - [Flash Attention](#flash-attention)
+  - [Sliding Tile Attention](#sliding-tile-attention)
+  - [Sage Attention](#sage-attention)
+  - [Sage Attention 3](#sage-attention-3)
 
 - Caching Techniques
-  - [TeaCache](#optimizations-teacache)
+  - [TeaCache](#teacache)
 
 ## Attention Backends
 
@@ -74,7 +74,7 @@ python setup.py install
 pip install st_attn==0.0.4
 ```
 
-Please see [this page](#sta-installation) for more installation instructions.
+Please see [this page](../attention/sta/index.md) for more installation instructions.
 
 ### Video Sparse Attention
 
@@ -85,7 +85,7 @@ git submodule update --init --recursive
 python setup_vsa.py install
 ```
 
-Please see [this page](#vsa-installation) for more installation instructions.
+Please see [this page](../attention/vsa/index.md) for more installation instructions.
 
 ### Sage Attention
 

docs/inference/support_matrix.md

@@ -40,20 +40,26 @@ The `HuggingFace Model ID` can be directly pass to `from_pretrained()` methods a
   }
 </style>
 
-| Model Name | HuggingFace Model ID | Resolutions | TeaCache | Sliding Tile Attn | Sage Attn | VSA |
-|------------|---------------------|-------------|----------|-------------------|-----------|-----|
-| FastWan2.1 T2V 1.3B | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ✅ |
-| FastWan2.2 TI2V 5B Full Attn* | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ |
-| Wan2.2 TI2V 5B | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 720P | ⭕ | ⭕ | ✅ | ⭕ |
-| Wan2.2 T2V A14B | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ |
-| Wan2.2 I2V A14B | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ |
-| HunyuanVideo | `hunyuanvideo-community/HunyuanVideo` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ |
-| FastHunyuan | `FastVideo/FastHunyuan-diffusers` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ |
-| Wan2.1 T2V 1.3B | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | 480P | ✅ | ✅* | ✅ | ⭕ |
-| Wan2.1 T2V 14B | `Wan-AI/Wan2.1-T2V-14B-Diffusers` | 480P, 720P | ✅ | ✅* | ✅ | ⭕ |
-| Wan2.1 I2V 480P | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | 480P | ✅ | ✅* | ✅ | ⭕ |
-| Wan2.1 I2V 720P | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` | 720P | ✅ | ✅ | ✅ | ⭕ |
-| StepVideo T2V | `FastVideo/stepvideo-t2v-diffusers` | 768px768px204f<br>544px992px204f<br>544px992px136f | ❌ | ❌ | ✅ | ⭕ |
+| Model Name | HuggingFace Model ID | Resolutions | TeaCache | Sliding Tile Attn | Sage Attn | VSA | BSA |
+|------------|---------------------|-------------|----------|-------------------|-----------|-----|-----|
+| FastWan2.1 T2V 1.3B | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
+| FastWan2.2 TI2V 5B Full Attn* | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
+| Wan2.2 TI2V 5B | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 720P | ⭕ | ⭕ | ✅ | ⭕ | ⭕ |
+| Wan2.2 T2V A14B | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |
+| Wan2.2 I2V A14B | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |
+| HunyuanVideo | `hunyuanvideo-community/HunyuanVideo` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ | ⭕ |
+| FastHunyuan | `FastVideo/FastHunyuan-diffusers` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ | ⭕ |
+| Wan2.1 T2V 1.3B | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | 480P | ✅ | ✅* | ✅ | ⭕ | ⭕ |
+| Wan2.1 T2V 14B | `Wan-AI/Wan2.1-T2V-14B-Diffusers` | 480P, 720P | ✅ | ✅* | ✅ | ⭕ | ⭕ |
+| Wan2.1 I2V 480P | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | 480P | ✅ | ✅* | ✅ | ⭕ | ⭕ |
+| Wan2.1 I2V 720P | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` | 720P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
+| StepVideo T2V | `FastVideo/stepvideo-t2v-diffusers` | 768px768px204f<br>544px992px204f<br>544px992px136f | ❌ | ❌ | ✅ | ⭕ | ⭕ |
+| TurboWan2.1 T2V 1.3B | `loayrashid/TurboWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| TurboWan2.1 T2V 14B | `loayrashid/TurboWan2.1-T2V-14B-Diffusers` | 480P, 720P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| LongCat T2V 13.6B | See note** | 480P<br>720P | ❌ | ❌ | ❌ | ⭕ | ✅ |
+| Matrix Game 2.0 Base | `FastVideo/Matrix-Game-2.0-Base-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| Matrix Game 2.0 GTA | `FastVideo/Matrix-Game-2.0-GTA-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| Matrix Game 2.0 TempleRun | `FastVideo/Matrix-Game-2.0-TempleRun-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 
 **Note**: Wan2.2 TI2V 5B has some quality issues when performing I2V generation. We are working on fixing this issue.
 
@@ -64,3 +70,13 @@ The `HuggingFace Model ID` can be directly pass to `from_pretrained()` methods a
 
 ### Sliding Tile Attention
 - Currently only Hopper GPUs (H100s) are supported.
+
+### TurboWan2.1 (TurboDiffusion)
+- Uses TurboDiffusionPipeline with RCM scheduler for 1-4 step generation
+- Requires SLA attention backend: `export FASTVIDEO_ATTENTION_BACKEND=SLA_ATTN`
+- Uses `guidance_scale=1.0` (no classifier-free guidance)
+
+### Matrix Game 2.0
+- Image-to-video game world models with keyboard/mouse control input
+- Three variants available: Base (universal), GTA, and TempleRun
+- Each variant has different keyboard dimensions for control inputs

docs/seed_parameter_behavior.md

@@ -1,51 +0,0 @@
-# Seed Parameter Behavior in vLLM
-
-## Overview
-
-The `seed` parameter in vLLM is used to control the random states for various random number generators. This parameter can affect the behavior of random operations in user code, especially when working with models in vLLM.
-
-## Default Behavior
-
-By default, the `seed` parameter is set to `None`. When the `seed` parameter is `None`, the global random states for `random`, `np.random`, and `torch.manual_seed` are not set. This means that the random operations will behave as expected, without any fixed random states.
-
-## Specifying a Seed
-
-If a specific seed value is provided, the global random states for `random`, `np.random`, and `torch.manual_seed` will be set accordingly. This can be useful for reproducibility, as it ensures that the random operations produce the same results across multiple runs.
-
-## Example Usage
-
-### Without Specifying a Seed
-
-```python
-import random
-from vllm import LLM
-
-# Initialize a vLLM model without specifying a seed
-model = LLM(model="Qwen/Qwen2.5-0.5B-Instruct")
-
-# Try generating random numbers
-print(random.randint(0, 100))  # Outputs different numbers across runs
-```
-
-### Specifying a Seed
-
-```python
-import random
-from vllm import LLM
-
-# Initialize a vLLM model with a specific seed
-model = LLM(model="Qwen/Qwen2.5-0.5B-Instruct", seed=42)
-
-# Try generating random numbers
-print(random.randint(0, 100))  # Outputs the same number across runs
-```
-
-## Important Notes
-
-- If the `seed` parameter is not specified, the behavior of global random states remains unaffected.
-- If a specific seed value is provided, the global random states for `random`, `np.random`, and `torch.manual_seed` will be set to that value.
-- This behavior can be useful for reproducibility but may lead to non-intuitive behavior if the user is not explicitly aware of it.
-
-## Conclusion
-
-Understanding the behavior of the `seed` parameter in vLLM is crucial for ensuring the expected behavior of random operations in your code. By default, the `seed` parameter is set to `None`, which means that the global random states are not affected. However, specifying a seed value can help achieve reproducibility in your experiments.

docs/sliding_tile_attention/demo.md

@@ -1,16 +0,0 @@
-
-# 🔍 Demo
-This is is a demo for 2D STA with window size (6,6) operating on a (10, 10) image.
-
-<div style="text-align: center;">
-  <video controls width="800">
-    <source src="https://github.com/user-attachments/assets/f3b6dd79-7b43-4b60-a0fa-3d6495ec5747" type="video/mp4">
-    Your browser does not support the video tag.
-  </video>
-</div>
-
-You can run STA using the following command:
-
-```bash
-bash scripts/inference/v1_inference_wan_STA.sh
-```

docs/sliding_tile_attention/installation.md

@@ -1,65 +0,0 @@
-
-# 🔧 Installation
-You can install the Sliding Tile Attention package using
-
-```
-pip install st_attn
-``` 
-
-# Building from Source
-We test our code on Pytorch 2.5.0 and CUDA>=12.4. Currently we only have implementation on H100.
-First, install C++20 for ThunderKittens:
-
-```bash
-sudo apt update
-sudo apt install gcc-11 g++-11
-
-sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
-
-sudo apt update
-sudo apt install clang-11
-```
-
-Set up CUDA environment (if using CUDA 12.4):
-
-```bash
-export CUDA_HOME=/usr/local/cuda-12.4
-export PATH=${CUDA_HOME}/bin:${PATH} 
-export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
-```
-
-Install STA:
-
-```bash
-cd csrc/attn/sliding_tile_attn/
-git submodule update --init --recursive
-python setup.py install
-```
-
-# 🧪 Test
-
-```bash
-python csrc/attn/tests/test_sta.py
-```
-
-# 📋 Usage
-
-```python
-from st_attn import sliding_tile_attention
-# assuming video size (T, H, W) = (30, 48, 80), text tokens = 256 with padding. 
-# q, k, v: [batch_size, num_heads, seq_length, head_dim], seq_length = T*H*W + 256
-# a tile is a cube of size (6, 8, 8)
-# window_size in tiles: [(window_t, window_h, window_w), (..)...]. For example, window size (3, 3, 3) means a query can attend to (3x6, 3x8, 3x8) = (18, 24, 24) tokens out of the total 30x48x80 video.
-# text_length: int ranging from 0 to 256
-# If your attention contains text token (Hunyuan)
-out = sliding_tile_attention(q, k, v, window_size, text_length)
-# If your attention does not contain text token (StepVideo)
-out = sliding_tile_attention(q, k, v, window_size, 0, False)
-
-```
-
-# 🚀Inference
-
-```bash
-bash scripts/inference/v1_inference_wan_STA.sh
-```

docs/training/data_preprocess.md

@@ -1,45 +1,130 @@
+# 🧱 Data Preprocessing
 
-# 🧱 Data Preprocess
+To save GPU memory during training, FastVideo precomputes text embeddings and VAE latents. This eliminates the need to load the text encoder and VAE during training.
 
-To save GPU memory, we precompute text embeddings and VAE latents to eliminate the need to load the text encoder and VAE during training.
+## Quick Start
 
-We provide a sample dataset to help you get started. Download the source media using the following command:
+Download the sample dataset and run preprocessing:
 
 ```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/mini_i2v_dataset --local_dir=data/mini_i2v_dataset --repo_type=dataset
+# Download the crush-smol dataset
+python scripts/huggingface/download_hf.py \
+    --repo_id "wlsaidhi/crush-smol-merged" \
+    --local_dir "data/crush-smol" \
+    --repo_type "dataset"
+
+# Run preprocessing
+bash examples/training/finetune/wan_t2v_1.3B/crush_smol/preprocess_wan_data_t2v_new.sh
 ```
 
-The folder `crush-smol_raw/` contains raw videos and captions for testing preprocessing, while `crush-smol_preprocessed/` contains latents prepared for testing training.
+## Preprocessing Pipeline
 
-To preprocess the dataset for fine-tuning or distillation, run:
+The new preprocessing pipeline supports multiple dataset formats and video loaders:
 
-```
-bash scripts/preprocess/v1_preprocess_wan_data_t2v # for wan
+```bash
+GPU_NUM=2
+MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+DATASET_PATH="data/crush-smol/"
+OUTPUT_DIR="data/crush-smol_processed_t2v/"
+
+torchrun --nproc_per_node=$GPU_NUM \
+    -m fastvideo.pipelines.preprocess.v1_preprocessing_new \
+    --model_path $MODEL_PATH \
+    --mode preprocess \
+    --workload_type t2v \
+    --preprocess.video_loader_type torchvision \
+    --preprocess.dataset_type merged \
+    --preprocess.dataset_path $DATASET_PATH \
+    --preprocess.dataset_output_dir $OUTPUT_DIR \
+    --preprocess.preprocess_video_batch_size 2 \
+    --preprocess.dataloader_num_workers 0 \
+    --preprocess.max_height 480 \
+    --preprocess.max_width 832 \
+    --preprocess.num_frames 77 \
+    --preprocess.train_fps 16 \
+    --preprocess.samples_per_file 8 \
+    --preprocess.flush_frequency 8 \
+    --preprocess.video_length_tolerance_range 5
 ```
 
-## Process your own dataset
+### Key Parameters
 
-If you wish to create your own dataset for finetuning or distillation, please refer `mini_i2v_dataset/crush-smol_raw/` to structure you video dataset in the following format:
+| Parameter | Description |
+|-----------|-------------|
+| `--workload_type` | Task type: `t2v` (text-to-video) or `i2v` (image-to-video) |
+| `--preprocess.dataset_type` | Input format: `hf` (HuggingFace) or `merged` (local folder) |
+| `--preprocess.dataset_path` | Path to dataset (HF repo ID or local folder) |
+| `--preprocess.dataset_output_dir` | Output directory for Parquet files |
+| `--preprocess.video_loader_type` | Video decoder: `torchcodec` or `torchvision` |
+| `--preprocess.max_height` / `max_width` | Target resolution for videos |
+| `--preprocess.num_frames` | Number of frames to extract per video |
+| `--preprocess.train_fps` | Target FPS for frame extraction |
+
+## Dataset Formats
+
+### Merged Dataset (Local Folder)
+
+Structure your dataset as follows:
 
 ```
-path_to_your_dataset_folder/
+your_dataset/
+├── videos/
+│   ├── video_001.mp4
+│   ├── video_002.mp4
+│   └── ...
+└── videos2caption.json
+```
+
+The `videos2caption.json` maps video filenames to captions:
+
+```json
+[
+  {"path": "video_001.mp4", "cap": "A cat playing with yarn..."},
+  {"path": "video_002.mp4", "cap": "Ocean waves at sunset..."}
+]
+```
+
+### HuggingFace Dataset
+
+Use `--preprocess.dataset_type hf` and point `--preprocess.dataset_path` to a HuggingFace dataset with `video` and `caption` columns.
+
+## Creating Your Own Dataset
+
+If you have raw videos and captions in separate files, generate the `videos2caption.json`:
+
+```bash
+python scripts/dataset_preparation/prepare_json_file.py \
+    --data_folder path/to/your_raw_data/ \
+    --output path/to/output_folder
+```
+
+Your raw data folder should contain:
+
+```
+your_raw_data/
 ├── videos/
 │   ├── 0.mp4
 │   ├── 1.mp4
-├── videos.txt
-└── prompt.txt
+│   └── ...
+├── videos.txt    # list of video filenames
+└── prompt.txt    # corresponding captions (one per line)
 ```
 
-To geranate the `videos2caption.json` and `merge.txt`, run
+## Output Format
 
-``` python
-python scripts/dataset_preparation/prepare_json_file.py --data_folder mini_i2v_dataset/crush-smol_raw/ --output your_output_folder
-```
+Preprocessing outputs Parquet files in the `combined_parquet_dataset/` subdirectory containing:
 
-Adjust the `DATA_MERGE_PATH` and `OUTPUT_DIR` in `scripts/preprocess/v1_preprocess_****.sh` accordingly and run:
+- `vae_latent_bytes` — VAE-encoded video latent
+- `text_embedding_bytes` — text encoder output
+- `clip_feature_bytes` — CLIP image features (I2V only)
+- `first_frame_latent_bytes` — first frame latent (I2V only)
+- Metadata: shapes, dtypes, and sample identifiers
 
-```
-bash scripts/preprocess/v1_preprocess_****.sh
-```
+## Examples
 
-The preprocessed data will be put into the `OUTPUT_DIR` and the `videos2caption.json` can be used in finetune and distill scripts.
+See ready-to-run preprocessing scripts in the training examples:
+
+- **T2V**: `examples/training/finetune/wan_t2v_1.3B/crush_smol/preprocess_wan_data_t2v_new.sh`
+- **I2V**: `examples/training/finetune/wan_i2v_14B_480p/crush_smol/preprocess_wan_data_i2v_new.sh`
+
+**→ [Browse all training examples](examples/examples_training_index.md)**

docs/training/finetune.md

@@ -1,78 +1,176 @@
+# 🧠 Finetuning
 
-# 🧠 Finetune
-## ⚡ Full Finetune
-Ensure your data is prepared and preprocessed in the format specified in [data_preprocess.md](#v0-data-preprocess). For convenience, we also provide a mochi preprocessed Black Myth Wukong data that can be downloaded directly:
+This guide covers finetuning video diffusion models with FastVideo, including full finetuning and LoRA.
+
+## Training Arguments
+
+FastVideo training scripts use several argument groups:
+
+### Training Arguments
+
+| Argument | Description |
+|----------|-------------|
+| `--max_train_steps` | Total training steps |
+| `--train_batch_size` | Batch size per GPU |
+| `--gradient_accumulation_steps` | Steps to accumulate before optimizer update |
+| `--num_latent_t` | Temporal latent dimension (reduce to save memory) |
+| `--num_height` / `--num_width` | Video resolution |
+| `--num_frames` | Number of frames per video |
+| `--output_dir` | Directory for checkpoints |
+
+### Parallelism Arguments
+
+| Argument | Description |
+|----------|-------------|
+| `--num_gpus` | Total number of GPUs |
+| `--sp_size` | Sequence parallel size (increase to reduce memory per GPU) |
+| `--tp_size` | Tensor parallel size |
+| `--hsdp_replicate_dim` | HSDP replication dimension |
+| `--hsdp_shard_dim` | HSDP sharding dimension |
+
+### Optimizer Arguments
+
+| Argument | Description |
+|----------|-------------|
+| `--learning_rate` | Base learning rate |
+| `--mixed_precision` | Precision mode (`bf16` recommended) |
+| `--weight_decay` | Weight decay for regularization |
+| `--max_grad_norm` | Gradient clipping threshold |
+
+### Validation Arguments
+
+| Argument | Description |
+|----------|-------------|
+| `--log_validation` | Enable validation logging |
+| `--validation_dataset_file` | JSON file with validation prompts |
+| `--validation_steps` | Run validation every N steps |
+| `--validation_sampling_steps` | Inference steps for validation |
+| `--validation_guidance_scale` | CFG scale for validation |
+
+## Full Finetuning
+
+Full finetuning updates all model weights. This provides the best quality but requires more GPU memory.
 
 ```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/Mochi-Black-Myth --local_dir=data/Mochi-Black-Myth --repo_type=dataset
+# Example: Wan2.1 T2V 1.3B full finetune (4 GPUs)
+bash examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh
 ```
 
-Download the original model weights as specified in the [Distillation Section](../distillation/dmd.md):
+**Typical settings:**
 
-Then you can run the finetune with:
+- Learning rate: `1e-5` to `5e-5`
+- Gradient checkpointing: `--enable_gradient_checkpointing_type "full"`
+- Memory scaling: Increase `--sp_size` or reduce `--num_latent_t` to fit in memory
 
-```
-bash scripts/finetune/finetune_mochi.sh # for mochi
-```
+## LoRA Finetuning
 
-**Note that for finetuning, we did not tune the hyperparameters in the provided script.**
-## ⚡ Finetune with VSA
-Follow [data_preprocess.md](#v0-data-preprocess) to get parquet files for preproccessed latent, and then run:
+LoRA (Low-Rank Adaptation) trains lightweight adapters while keeping the base model frozen. This significantly reduces memory usage and training time.
+
+### LoRA-Specific Arguments
+
+| Argument | Description |
+|----------|-------------|
+| `--lora_training True` | Enable LoRA mode |
+| `--lora_rank` | Rank of LoRA adapters (16, 32, 64, 128) |
+
+### Learning Rate for LoRA
+
+**Important:** LoRA typically requires a **10–20× higher learning rate** than full finetuning because only the low-rank adapters are being trained while the base model is frozen.
+
+| Training Mode | Recommended Learning Rate |
+|---------------|---------------------------|
+| Full finetune | `1e-5` to `5e-5` |
+| LoRA | `1e-4` to `2e-4` |
+
+### Example LoRA Training
 
 ```bash
-bash scripts/finetune/finetune_v1_VSA.sh
+# Example: Wan2.1 T2V 1.3B LoRA finetune (1 GPU)
+bash examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_lora.sh
 ```
 
-## ⚡ Lora Finetune
+Key differences from full finetune:
 
-Hunyuan supports Lora fine-tuning of videos up to 720p. Demos and prompts of Black-Myth-Wukong can be found in [here](https://huggingface.co/FastVideo/Hunyuan-Black-Myth-Wukong-lora-weight). You can download the Lora weight through:
+- Add `--lora_training True --lora_rank 32`
+- Use higher learning rate (10–20× full finetune)
+- Can run on fewer GPUs (even single GPU)
+- Outputs adapter weights instead of full model
+
+## LoRA Extraction and Merging
+
+FastVideo provides tools to extract LoRA adapters from finetuned models and merge them back.
+
+### Extract LoRA Adapter
+
+Extract a LoRA adapter by comparing a finetuned model to its base:
 
 ```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/Hunyuan-Black-Myth-Wukong-lora-weight --local_dir=data/Hunyuan-Black-Myth-Wukong-lora-weight --repo_type=model
+python scripts/lora_extraction/extract_lora.py \
+  --base Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
+  --ft path/to/your/finetuned_model \
+  --out adapter_r32.safetensors \
+  --rank 32
 ```
 
-### Minimum Hardware Requirement
-- 40 GB GPU memory each for 2 GPUs with lora.
-- 30 GB GPU memory each for 2 GPUs with CPU offload and lora.
+| Argument | Description |
+|----------|-------------|
+| `--base` | Base model (HuggingFace ID or local path) |
+| `--ft` | Finetuned model path |
+| `--out` | Output adapter file (.safetensors) |
+| `--rank` | LoRA rank (16, 32, 64, 128) |
+| `--full-rank` | Extract full-rank adapter (optional) |
 
-Currently, both Mochi and Hunyuan models support Lora finetuning through diffusers. To generate personalized videos from your own dataset, you'll need to follow three main steps: dataset preparation, finetuning, and inference.
+### Merge LoRA Adapter
 
-### Dataset Preparation
-We provide scripts to better help you get started to train on your own characters!
-You can run this to organize your dataset to get the videos2caption.json before preprocess. Specify your video folder and corresponding caption folder (caption files should be .txt files and have the same name with its video):
-
-```
-python scripts/dataset_preparation/prepare_json_file.py --video_dir data/input_videos/ --prompt_dir data/captions/ --output_path data/output_folder/videos2caption.json --verbose
-```
-
-Also, we provide script to resize your videos:
-
-```
-python scripts/data_preprocess/resize_videos.py
-```
-
-### Finetuning
-After basic dataset preparation and preprocess, you can start to finetune your model using Lora:
-
-```
-bash scripts/finetune/finetune_hunyuan_hf_lora.sh
-```
-
-### Inference
-For inference with Lora checkpoint, you can run the following scripts with additional parameter `--lora_checkpoint_dir`:
-
-```
-bash scripts/inference/inference_hunyuan_hf.sh
-```
-
-**We also provide scripts for Mochi in the same directory.**
-
-### Finetune with Both Image and Video
-Our codebase support finetuning with both image and video.
+Merge an adapter back into a base model:
 
 ```bash
-bash scripts/finetune/finetune_hunyuan.sh
-bash scripts/finetune/finetune_mochi_lora_mix.sh
+python scripts/lora_extraction/merge_lora.py \
+  --base Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
+  --adapter adapter_r32.safetensors \
+  --ft path/to/your/finetuned_model \
+  --output merged_model
 ```
 
-For Image-Video Mixture Fine-tuning, make sure to enable the `--group_frame` option in your script.
+| Argument | Description |
+|----------|-------------|
+| `--base` | Base model path |
+| `--adapter` | LoRA adapter file |
+| `--ft` | Finetuned model (for config reference) |
+| `--output` | Output directory for merged model |
+
+### Validate Merged Model
+
+Compare the merged model against the original finetuned model:
+
+```bash
+python scripts/lora_extraction/lora_inference_comparison.py \
+  --base merged_model \
+  --ft path/to/your/finetuned_model \
+  --adapter NONE \
+  --output-dir results \
+  --prompt "A cat sitting on a windowsill" \
+  --compute-ssim \
+  --compute-lpips
+```
+
+## Training Examples
+
+Ready-to-run training scripts are available for multiple models:
+
+**→ [Browse all training examples](examples/examples_training_index.md)**
+
+| Model | Type | Example |
+|-------|------|---------|
+| Wan2.1 T2V 1.3B | T2V | `examples/training/finetune/wan_t2v_1.3B/crush_smol/` |
+| Wan2.1 I2V 14B | I2V | `examples/training/finetune/wan_i2v_14B_480p/crush_smol/` |
+| Wan2.1-Fun 1.3B InP | I2V | `examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/` |
+| Wan2.1 VSA | T2V/I2V | `examples/training/finetune/Wan2.1-VSA/Wan-Syn-Data/` |
+
+Each example includes:
+
+- `download_dataset.sh` — download sample data
+- `preprocess_*.sh` — run preprocessing
+- `finetune_*.sh` — full finetune launcher
+- `finetune_*_lora.sh` — LoRA finetune launcher
+- `validation.json` — validation prompts

docs/training/overview.md

@@ -0,0 +1,67 @@
+# Training Overview
+
+FastVideo supports finetuning video diffusion models on custom datasets. This page explains what data you need and how to get started.
+
+## Data Requirements
+
+To save GPU memory during training, FastVideo precomputes embeddings and latents ahead of time. This eliminates the need to load the text encoder and VAE during training, significantly reducing memory usage.
+
+### Text-to-Video (T2V) Finetuning
+
+For T2V models, you need:
+
+| Component | Description |
+|-----------|-------------|
+| **Text embeddings** | Precomputed embeddings from the model's text encoder (e.g., T5 or LLaMA). Stored as numpy arrays in Parquet files. |
+| **Video latents** | VAE-encoded representations of your training videos. Each video is encoded into a compressed latent tensor. |
+
+### Image-to-Video (I2V) Finetuning
+
+For I2V models, you need everything from T2V plus additional image conditioning. Note that not all I2V architectures require encoded images—this depends on how the model conditions on the input frame. Wan2.1 and Wan2.2 A14B I2V models do require these additional components:
+
+| Component | Description |
+|-----------|-------------|
+| **Text embeddings** | Same as T2V—precomputed from the text encoder. |
+| **Video latents** | Same as T2V—VAE-encoded video representations. |
+| **First frame latent** | VAE-encoded representation of the first frame, used as the conditioning image. |
+| **CLIP features** | Image embeddings from a CLIP vision encoder for the conditioning frame. |
+
+## Preprocessing
+
+Before training, you need to preprocess your raw videos and captions into Parquet files containing precomputed latents and embeddings.
+
+FastVideo supports two input formats:
+
+- **HuggingFace datasets** — load directly from HF Hub or local HF datasets
+- **Merged datasets** — local folder with videos and a `videos2caption.json` metadata file
+
+**→ See [Data Preprocessing](data_preprocess.md) for full details and examples.**
+
+## Training Examples
+
+Ready-to-run examples with preprocessing scripts, training launchers, and validation configs are available for multiple models and datasets:
+
+**→ [Browse all training examples](examples/examples_training_index.md)**
+
+Each example includes:
+
+- `download_dataset.sh` — download sample data
+- `preprocess_*.sh` — run preprocessing
+- `finetune_*.sh` — launch training (full finetune or LoRA)
+- `validation.json` — validation prompts for checkpoints
+
+## Training Methods
+
+FastVideo supports several training approaches:
+
+| Method | Use Case |
+|--------|----------|
+| **Full finetune** | Adapt entire model to a new domain or style |
+| **LoRA finetune** | Lightweight adaptation with frozen base weights |
+| **VSA finetune** | Finetune with Variable Sparse Attention for efficiency |
+
+## Next Steps
+
+1. **Get started**: Pick an example from the [training examples index](examples/examples_training_index.md)
+2. **Prepare data**: Follow [data preprocessing](data_preprocess.md) for your own dataset
+3. **Run inference**: After training, see [inference examples](../inference/examples/examples_inference_index.md)

docs/utilities/lora.md

@@ -0,0 +1,71 @@
+# LoRA Extraction and Merging
+
+Tools for extracting and merging LoRA adapters for FastVideo models.
+
+## Extract LoRA Adapter
+
+```bash
+python scripts/lora_extraction/extract_lora.py \
+  --base Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+  --ft FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers \
+  --out adapter_r32.safetensors \
+  --rank 32
+```
+
+**Options:**
+
+- `--base`: Base model (HuggingFace ID or local path)
+- `--ft`: Fine-tuned model (HuggingFace ID or local path)
+- `--out`: Output adapter file
+- `--rank`: LoRA rank (16, 32, 64, 128)
+- `--full-rank`: Extract full-rank adapter (optional)
+
+## Merge Adapter
+
+```bash
+python scripts/lora_extraction/merge_lora.py \
+  --base Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+  --adapter adapter_r32.safetensors \
+  --ft FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers \
+  --output merged_model
+```
+
+**Options:**
+
+- `--base`: Base model (HuggingFace ID or local path)
+- `--adapter`: LoRA adapter file (.safetensors)
+- `--ft`: Fine-tuned model (for configuration)
+- `--output`: Output directory
+
+## Validate Quality (Optional)
+
+```bash
+python scripts/lora_extraction/lora_inference_comparison.py \
+  --base merged_model \
+  --ft FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers \
+  --adapter NONE \
+  --output-dir results \
+  --prompt "A cat sitting on a windowsill" \
+  --seed 42 \
+  --height 480 \
+  --width 480 \
+  --num-frames 49 \
+  --num-inference-steps 32 \
+  --compute-ssim \
+  --compute-lpips
+```
+
+**Options:**
+
+- `--base`: Merged model or base model path
+- `--ft`: Fine-tuned model (reference)
+- `--adapter`: Path to adapter or NONE
+- `--output-dir`: Output directory
+- `--prompt`: Text prompt (default: "A cat sitting on a windowsill")
+- `--seed`: Random seed (default: 42)
+- `--height`: Video height (default: 480)
+- `--width`: Video width (default: 832)
+- `--num-frames`: Number of frames (default: 49)
+- `--num-inference-steps`: Inference steps (default: 32)
+- `--compute-ssim`: Compute SSIM metric
+- `--compute-lpips`: Compute LPIPS metric

docs/video_sparse_attention/installation.md

@@ -1,65 +0,0 @@
-
-# 🔧 Installation
-You can install the Video Sparse Attention package using
-
-```bash
-pip install vsa
-```
-
-# Building from Source
-We support H100 (via ThunderKittens) and any other GPU (via Triton) for VSA.
-
-First, install C++20 for ThunderKittens (if using H100):
-
-```bash
-sudo apt update
-sudo apt install gcc-11 g++-11
-
-sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
-
-sudo apt update
-sudo apt install clang-11
-```
-
-Set up CUDA environment (if using CUDA 12.8):
-
-```bash
-export CUDA_HOME=/usr/local/cuda-12.8
-export PATH=${CUDA_HOME}/bin:${PATH} 
-export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
-```
-
-Install VSA:
-
-```bash
-cd csrc/attn/video_sparse_attn/
-git submodule update --init --recursive
-python setup.py install
-```
-
-# 🧪 Test
-
-```bash
-python csrc/attn/tests/test_vsa.py
-```
-
-# 📋 Usage
-
-```python
-from vsa import video_sparse_attn
-
-# q, k, v: [batch_size, num_heads, seq_len, head_dim]
-# variable_block_sizes: [num_blocks] - number of valid tokens in each block
-# topk: int - number of top-k blocks to attend to
-# block_size: int or tuple of 3 ints - size of each block (default: 64 tokens)
-# compress_attn_weight: optional weight for compressed attention branch
-
-output = video_sparse_attn(q, k, v, variable_block_sizes, topk, block_size, compress_attn_weight)
-
-``` 
-
-# 🚀Inference
-
-```bash
-bash scripts/inference/v1_inference_wan_VSA.sh
-```

examples/distill/SFHunyuan1.5-T2V/distill_dmd.slurm

@@ -0,0 +1,172 @@
+#!/bin/bash
+#SBATCH --job-name=dmd_3333
+#SBATCH --partition=main
+#SBATCH --nodes=4
+#SBATCH --ntasks=4
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=dmd_3333_output/dmd_3333_%j.out
+#SBATCH --error=dmd_3333_output/dmd_3333_%j.err
+#SBATCH --exclusive
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+export NCCL_DEBUG_SUBSYS=INIT,NET
+# 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 TOKENIZERS_PARALLELISM=false
+# export WANDB_API_KEY="8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde"
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+# export WANDB_API_KEY='your_wandb_api_key_here'
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+source ~/conda/miniconda/bin/activate
+conda activate wei-fv-distill
+export HOME="/mnt/weka/home/hao.zhang/wei"
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=32
+
+# Model paths for Self-Forcing DMD distillation with Hunyuan1.5:
+GENERATOR_MODEL_PATH="weizhou03/SFHunyuanVideo-1.5-Diffusers-480p_t2v"  # Updated to Wan2.2
+REAL_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v" # Teacher model
+FAKE_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v"
+# REAL_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled"
+# FAKE_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled" # Critic model
+
+# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
+# DATA_DIR="/mnt/weka/home/hao.zhang/matthew/FastVideo/data/test-text-preprocessing"
+DATA_DIR="/mnt/weka/home/hao.zhang/wl/sharefs/wl/release/vidprom_16k_text_embed"
+DATA_DIR_2="/mnt/weka/home/hao.zhang/wl/sharefs/wl/release/fv-ode-preprocessing-16k-hy15-121"
+# DATA_DIR=data/crush-smol_processed_t2v/combined_parquet_dataset
+# DATA_DIR="/mnt/sharefs/users/hao.zhang/Vchitect-2M/Wan-Syn-upload/latents_i2v/train/"
+# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
+
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+training_args=(
+  --tracker_project_name SFhy1.5_t2v_distill_self_forcing_dmd  # Updated for Wan2.2
+  --output_dir "/mnt/weka/home/hao.zhang/wei/SFhy1.5_distill_3333_1e-5_1e-5_cfg6_corrected_scheduler"
+  --wandb_run_name "self_forcing_3333_context_forcing"
+  --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 848
+  --enable_gradient_checkpointing_type "full"
+  --simulate_generator_forward
+  --log_visualization
+  --num_frames 81
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
+  # --resume-from-checkpoint "/mnt/weka/home/hao.zhang/wei/SFhy1.5_distill_1333/checkpoint-300"
+  --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wei/hy15_worldplay_tf_init_3333/checkpoint-2800/transformer/diffusion_pytorch_model.safetensors
+  # --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wei/hy15_worldplay_df_init_1333/checkpoint-11600/transformer/diffusion_pytorch_model.safetensors
+  # --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wei/hy15_distilled_ode_init_3333/checkpoint-2400/transformer/diffusion_pytorch_model.safetensors
+)
+
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+dataset_args=(
+  --data_path "$DATA_DIR"
+  # --data_path_2 "$DATA_DIR_2"
+  --dataloader_num_workers 4
+)
+
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "8"
+  --validation_guidance_scale "1.0" # not used for dmd inference
+  --text-encoder-cpu-offload
+  # --vae_cpu_offload True
+)
+
+optimizer_args=(
+  --optimizer-type "muon"
+  --learning_rate 1e-5
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 100
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 1.0
+)
+
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 5
+  --seed 1000
+  --use_ema True
+  --ema_decay 0.99
+  --ema_start_step 200
+)
+
+dmd_args=(
+  --warp_denoising_step
+  # --dmd_denoising_steps '1000,875,750,625,500,375,250,125'
+  # --dmd_denoising_steps '1000,760,520,280'
+  # --dmd_denoising_steps '1000,750,500,250'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --dfake_gen_update_ratio 5
+  --real_score_guidance_scale 3.5
+  --fake_score_learning_rate 5e-6
+  --fake_score_betas '0.0,0.999'
+)
+
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+  # --use-ode-init True
+  # --use-context-forcing True
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/hy15_self_forcing_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"

examples/distill/SFWan2.1-T2V/README.md

@@ -0,0 +1,19 @@
+# Self-Forcing Distillation for SFWan2.1 T2V 1.3B
+
+These scripts demonstrate self-forcing distillation (SFwan) for the causal Wan2.1 T2V 1.3B model. The workflow mirrors DMD2 while injecting self-forcing blocks so the student can autoregressively refine later frames.
+
+## Run the recipe
+1. Download the preprocessed text-video dataset:
+   ```bash
+   bash examples/distill/SFWan2.1-T2V/download_dataset.sh
+   ```
+2. (Optional) Regenerate parquet shards locally:
+   ```bash
+   bash examples/distill/SFWan2.1-T2V/preprocess_data.sh
+   ```
+3. Launch self-forcing distillation with your cluster settings:
+   ```bash
+   sbatch examples/distill/SFWan2.1-T2V/distill_dmd_t2v_1.3B.sh
+   ```
+
+Update the dataset paths and wandb credentials inside the script before running on your environment.

examples/distill/SFWan2.1-T2V/distill_dmd.slurm

@@ -0,0 +1,172 @@
+#!/bin/bash
+#SBATCH --job-name=distill_dmd_t2v_1.3b
+#SBATCH --nodes=8
+#SBATCH --ntasks=8
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=distill_dmd_t2v_1.3b_output/distill_dmd_t2v_1.3b_%j.out
+#SBATCH --error=distill_dmd_t2v_1.3b_output/distill_dmd_t2v_1.3b_%j.err
+#SBATCH --exclusive
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+export NCCL_DEBUG_SUBSYS=INIT,NET
+# 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 TOKENIZERS_PARALLELISM=false
+# export WANDB_API_KEY="8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde"
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+# export WANDB_API_KEY='your_wandb_api_key_here'
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+export HF_HOME=/mnt/data/huggingface
+export HF_DATASETS_CACHE="/tmp/hf_datasets_${SLURM_JOB_ID}_${SLURM_PROCID}"
+mkdir -p "${HF_DATASETS_CACHE}"
+export CUDA_HOME=/usr/local/cuda
+export PATH="$CUDA_HOME/bin:$PATH"
+export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/home/shared-bin/local/usr/lib/aarch64-linux-gnu:$CUDA_HOME/lib64"
+export CUDACXX="$CUDA_HOME/bin/nvcc"
+
+source ~/miniconda3/bin/activate
+conda activate fastvideo
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=64
+
+# Model paths for Self-Forcing DMD distillation:
+GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
+
+DATA_DIR="/mnt/data/vidprom_filtered_extended_umt5_text_embed"
+# DATA_DIR="/mnt/data/mixkit_wan_1.3b_processed_t2v_distributed"
+DATA_DIR_2="/mnt/data/mixkit_wan_1.3b_processed_t2v_distributed"
+VALIDATION_DATASET_FILE="/mnt/home/zhouw.jerry2017/FastVideo/examples/distill/SFWan2.1-T2V/validation_64.json"
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+training_args=(
+  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd
+  --wandb_run_name causal_forcing_vidprom_extended
+  --output_dir /mnt/data/wan_cf_df_distill_1.3b
+  --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
+  --enable_gradient_checkpointing_type "full"
+  --log_visualization
+  --simulate_generator_forward
+  --num_frames 81
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
+  --enable_gradient_masking
+  --gradient_mask_last_n_frames 21
+)
+
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+dataset_args=(
+  --data_path "$DATA_DIR"
+  # --data_path_2 "$DATA_DIR_2"
+  --dataloader_num_workers 4
+)
+
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 100
+  --validation_sampling_steps "4"
+  --validation_guidance_scale "1.0"
+)
+
+optimizer_args=(
+  --optimizer-type "adamw"
+  --learning_rate 1e-5
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 1.0
+)
+
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 5
+  --seed 1000
+  --use_ema True
+  --ema_decay 0.99
+  --ema_start_step 200
+#   --init_weights_from_safetensors "/mnt/data/wan_tf_ode_init_3333/checkpoint-1600/transformer/diffusion_pytorch_model.safetensors"
+#   --init_weights_from_safetensors "/mnt/data/wan_tf_init_3333/checkpoint-800/transformer/diffusion_pytorch_model.safetensors"
+  # --init_weights_from_safetensors "/mnt/data/Causal-Forcing-wan1.3b/causal_ode.safetensors"
+  # --init_weights_from_safetensors "/mnt/data/wan_ar_diffusion_3333_mixkit_shift5/checkpoint-7200/transformer/diffusion_pytorch_model.safetensors"
+  --init_weights_from_safetensors "/mnt/data/wan_ar_diffusion_3333_mixkit_shift5/checkpoint-4800/transformer/diffusion_pytorch_model.safetensors"
+  # --init_weights_from_safetensors "/mnt/data/wan_tf_ode_init_3333_mixkit_shift5/checkpoint-2000/transformer/diffusion_pytorch_model.safetensors"
+  # --init_weights_from_safetensors "/mnt/data/wan_ar_df_diffusion_3333_mixkit_shift5/checkpoint-4800/transformer/diffusion_pytorch_model.safetensors"
+)
+
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --dfake_gen_update_ratio 5
+  --real_score_guidance_scale 3.0
+  --fake_score_learning_rate 2e-6
+  --fake_score_betas '0.0,0.999'
+  --warp_denoising_step
+  # --use-decoupled-dmd True
+  # --use-gan-loss True
+)
+
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"

examples/inference/basic/README.md

@@ -6,7 +6,7 @@ The `VideoGenerator` class provides the primary Python interface for doing offli
 - Python 3.10-3.12
 
 ## Installation
-If you have not installed FastVideo, please following these [instructions](https://hao-ai-lab.github.io/FastVideo/getting_started/installation.html) first.
+If you have not installed FastVideo, please following these [instructions](https://hao-ai-lab.github.io/FastVideo/getting_started/installation) first.
 
 ## Usage
 The first script in this example shows the most basic usage of FastVideo. If you are new to Python and FastVideo, you should start here.

examples/inference/basic/basic.py

@@ -12,10 +12,10 @@ def main():
         "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=True,
+        use_fsdp_inference=False, # set to True if GPU is out of memory
         dit_cpu_offload=False,
         vae_cpu_offload=False,
-        text_encoder_cpu_offload=True,
+        text_encoder_cpu_offload=False,
         pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
         # image_encoder_cpu_offload=False,
     )

examples/inference/basic/basic_cosmos2_5_t2w.py

@@ -0,0 +1,42 @@
+from fastvideo import VideoGenerator
+
+
+def main():
+    # Point this to your local diffusers model dir (or replace with a HF model ID).
+    model_path = "KyleShao/Cosmos-Predict2.5-2B-Diffusers"
+
+    generator = VideoGenerator.from_pretrained(
+        model_path,
+        num_gpus=1,
+        use_fsdp_inference=False,  # set True if GPU is out of memory
+        dit_cpu_offload=False,
+        vae_cpu_offload=False,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+    )
+
+    prompt = (
+       "A high-definition video captures the precision of robotic welding in an industrial setting. The first frame showcases a robotic arm, equipped with a welding torch, positioned over a large metal structure. The welding process is in full swing, with bright sparks and intense light illuminating the scene, creating a vivid display of blue and white hues. A significant amount of smoke billows around the welding area, partially obscuring the view but emphasizing the heat and activity. The background reveals parts of the workshop environment, including a ventilation system and various pieces of machinery, indicating a busy and functional industrial workspace. As the video progresses, the robotic arm maintains its steady position, continuing the welding process and moving to its left. The welding torch consistently emits sparks and light, and the smoke continues to rise, diffusing slightly as it moves upward. The metal surface beneath the torch shows ongoing signs of heating and melting. The scene retains its industrial ambiance, with the welding sparks and smoke dominating the visual field, underscoring the ongoing nature of the welding operation."
+    )
+
+    video = generator.generate_video(
+        prompt,
+        negative_prompt="",
+        height=704,
+        width=1280,
+        num_frames=77,
+        num_inference_steps=35,
+        guidance_scale=7.0,
+        fps=24,
+        output_path="outputs_video/cosmos2_5_t2w.mp4",
+        save_video=True,
+    )
+
+    generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()
+
+
+

examples/inference/basic/basic_dmd.py

@@ -14,7 +14,7 @@ def main():
         model_name,
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=True,
+        use_fsdp_inference=False, # set to True if GPU is out of memory
         # Adjust these offload parameters if you have < 32GB of VRAM
         text_encoder_cpu_offload=True,
         pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"

examples/inference/basic/basic_hy15.py

@@ -0,0 +1,42 @@
+from fastvideo import VideoGenerator
+import json
+# from fastvideo.configs.sample import SamplingParam
+
+OUTPUT_PATH = "video_samples_hy15"
+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=False, # set to True if GPU is out of memory
+        dit_cpu_offload=True,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
+        # image_encoder_cpu_offload=False,
+    )
+
+    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, negative_prompt="", num_frames=81, fps=16)
+
+    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, negative_prompt="", num_frames=81, fps=16)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_hyworld.py

@@ -0,0 +1,224 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Basic example for HYWorld (HY-WorldPlay) video generation using FastVideo.
+
+This example replicates the same functionality as HY-WorldPlay/run.sh,
+demonstrating image-to-video generation with camera trajectory control.
+"""
+
+import time
+import math
+import numpy as np
+import imageio
+import torchvision
+from einops import rearrange
+
+from fastvideo import VideoGenerator
+from fastvideo.pipelines import ForwardBatch
+from fastvideo.utils import shallow_asdict, align_to
+from fastvideo.logger import init_logger
+
+from fastvideo.models.dits.hyworld.pose import pose_to_input, compute_latent_num
+from fastvideo.models.dits.hyworld.resolution_utils import get_resolution_from_image
+
+logger = init_logger(__name__)
+
+class HYWorldVideoGenerator(VideoGenerator):
+    """Extended VideoGenerator that adds HYWorld-specific parameters to batch.extra."""
+
+    def _generate_single_video(self, prompt: str, sampling_param=None, **kwargs):
+        """Override to add viewmats, Ks, and action to batch.extra."""
+        fastvideo_args = self.fastvideo_args
+        pipeline_config = fastvideo_args.pipeline_config
+
+        if sampling_param is None:
+            from fastvideo.configs.sample import SamplingParam
+            sampling_param = SamplingParam.from_pretrained(fastvideo_args.model_path)
+
+        # Update sampling param with kwargs
+        if kwargs:
+            for key, value in kwargs.items():
+                if hasattr(sampling_param, key):
+                    setattr(sampling_param, key, value)
+
+        # Get pose string from sampling_param or kwargs
+        pose = kwargs.get('pose', getattr(sampling_param, 'POSE', 'w-31'))
+        num_frames = kwargs.get('num_frames', getattr(sampling_param, 'num_frames', 125))
+
+        # Calculate number of latents
+        latent_num = compute_latent_num(num_frames)
+
+        # Convert pose to viewmats, Ks, and action
+        viewmats, Ks, action = pose_to_input(pose, latent_num)
+
+        # Convert to tensors and add batch dimension
+        viewmats = viewmats.unsqueeze(0)  # (1, T, 4, 4)
+        Ks = Ks.unsqueeze(0)  # (1, T, 3, 3)
+        action = action.unsqueeze(0)  # (1, T)
+
+        # Validate inputs
+        prompt = prompt.strip()
+        sampling_param = sampling_param.__class__(**shallow_asdict(sampling_param))
+        output_path = kwargs.get("output_path", sampling_param.output_path)
+        sampling_param.prompt = prompt
+
+        if sampling_param.negative_prompt is not None:
+            sampling_param.negative_prompt = sampling_param.negative_prompt.strip()
+
+        # Validate dimensions
+        if (sampling_param.height <= 0 or sampling_param.width <= 0 or
+            sampling_param.num_frames <= 0):
+            raise ValueError(
+                f"Height, width, and num_frames must be positive integers")
+
+        temporal_scale_factor = pipeline_config.vae_config.arch_config.temporal_compression_ratio
+        num_frames = sampling_param.num_frames
+        num_gpus = fastvideo_args.num_gpus
+        use_temporal_scaling_frames = pipeline_config.vae_config.use_temporal_scaling_frames
+
+        # Adjust number of frames based on number of GPUs
+        if use_temporal_scaling_frames:
+            orig_latent_num_frames = (num_frames - 1) // temporal_scale_factor + 1
+        else:
+            orig_latent_num_frames = sampling_param.num_frames // 17 * 3
+
+        if orig_latent_num_frames % fastvideo_args.num_gpus != 0:
+            if use_temporal_scaling_frames:
+                new_num_frames = (orig_latent_num_frames - 1) * temporal_scale_factor + 1
+            else:
+                divisor = math.lcm(3, num_gpus)
+                orig_latent_num_frames = (
+                    (orig_latent_num_frames + divisor - 1) // divisor) * divisor
+                new_num_frames = orig_latent_num_frames // 3 * 17
+
+            logger.info(
+                "Adjusting number of frames from %s to %s based on number of GPUs (%s)",
+                sampling_param.num_frames, new_num_frames, fastvideo_args.num_gpus)
+            sampling_param.num_frames = new_num_frames
+
+        # Calculate sizes
+        target_height = align_to(sampling_param.height, 16)
+        target_width = align_to(sampling_param.width, 16)
+
+        # Calculate latent sizes
+        latents_size = [(sampling_param.num_frames - 1) // 4 + 1,
+                        sampling_param.height // 8, sampling_param.width // 8]
+        n_tokens = latents_size[0] * latents_size[1] * latents_size[2]
+
+        # Prepare batch
+        batch = ForwardBatch(
+            **shallow_asdict(sampling_param),
+            eta=0.0,
+            n_tokens=n_tokens,
+            VSA_sparsity=fastvideo_args.VSA_sparsity,
+        )
+
+        # Add HYWorld-specific parameters to batch.extra
+        batch.extra['viewmats'] = viewmats
+        batch.extra['Ks'] = Ks
+        batch.extra['action'] = action
+        batch.extra['chunk_latent_frames'] = 16  # For bidirectional model
+
+        # Run inference
+        start_time = time.perf_counter()
+        output_batch = self.executor.execute_forward(batch, fastvideo_args)
+        samples = output_batch.output
+        logging_info = output_batch.logging_info
+
+        gen_time = time.perf_counter() - start_time
+        logger.info("Generated successfully in %.2f seconds", gen_time)
+
+        # Process outputs
+        videos = rearrange(samples, "b c t h w -> t b c h w")
+        frames = []
+        for x in videos:
+            x = torchvision.utils.make_grid(x, nrow=6)
+            x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+            frames.append((x * 255).numpy().astype(np.uint8))
+
+        # Save video if requested
+        if batch.save_video:
+            imageio.mimsave(output_path, frames, fps=batch.fps, format="mp4")
+            logger.info("Saved video to %s", output_path)
+
+        if batch.return_frames:
+            return frames
+        else:
+            return {
+                "samples": samples,
+                "frames": frames,
+                "prompts": prompt,
+                "size": (target_height, target_width, batch.num_frames),
+                "generation_time": gen_time,
+                "logging_info": logging_info,
+                "trajectory": output_batch.trajectory_latents,
+                "trajectory_timesteps": output_batch.trajectory_timesteps,
+                "trajectory_decoded": output_batch.trajectory_decoded,
+            }
+
+# Default prompt from HY-WorldPlay run.sh
+DEFAULT_PROMPT = 'A paved pathway leads towards a stone arch bridge spanning a calm body of water.  Lush green trees and foliage line the path and the far bank of the water. A traditional-style pavilion with a tiered, reddish-brown roof sits on the far shore. The water reflects the surrounding greenery and the sky.  The scene is bathed in soft, natural light, creating a tranquil and serene atmosphere. The pathway is composed of large, rectangular stones, and the bridge is constructed of light gray stone.  The overall composition emphasizes the peaceful and harmonious nature of the landscape.'
+DEFAULT_IMAGE = 'https://raw.githubusercontent.com/Tencent-Hunyuan/HY-WorldPlay/main/assets/img/test.png'
+
+def main():
+    import argparse
+
+    # pose: (a, w, s, d) - (15, 31)
+    # num_frames: (61, 125)
+    parser = argparse.ArgumentParser(description="HYWorld video generation with FastVideo")
+    parser.add_argument("--prompt", type=str, default=DEFAULT_PROMPT, help="Text prompt for video generation")
+    parser.add_argument("--image", type=str, default=DEFAULT_IMAGE, help="Path or URL to input image")
+    parser.add_argument("--pose", type=str, default='w-31', help="Pose string (e.g., 'a-31', 'w-31', 's-31', 'd-31')")
+    parser.add_argument("--output_path", type=str, default='video_samples_hyworld', help="Output video path")
+    parser.add_argument("--num-frames", type=int, default=125, help="Number of frames")
+    parser.add_argument("--seed", type=int, default=1, help="Random seed")
+    parser.add_argument("--resolution", type=str, default="480p", help="Only support 480p for now")
+    args = parser.parse_args()
+
+    # Automatically determine resolution from input image
+    HEIGHT, WIDTH = get_resolution_from_image(args.image, args.resolution)
+    print(f"Image: {args.image}")
+    print(f"Pose: {args.pose}")
+    print(f"Resolution: {HEIGHT}x{WIDTH} (from {args.resolution} buckets)")
+    print(f"Num frames: {args.num_frames}")
+    print(f"Output path: {args.output_path}")
+
+    # Initialize generator
+    print("\nInitializing VideoGenerator for HYWorld...")
+    
+    generator = HYWorldVideoGenerator.from_pretrained(
+        "FastVideo/HY-WorldPlay-Bidirectional-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+        image_encoder_cpu_offload=True,
+    )
+
+    # Generate video
+    print("\nGenerating video...")
+    start_time = time.time()
+    video = generator.generate_video(
+        prompt=args.prompt,
+        image_path=args.image,
+        output_path=args.output_path,
+        save_video=True,
+        negative_prompt="",
+        num_frames=args.num_frames,
+        fps=24,
+        height=HEIGHT,
+        width=WIDTH,
+        seed=args.seed,
+        pose=args.pose,
+    )
+    elapsed = time.time() - start_time
+
+    print(f"\nVideo generated successfully!")
+    print(f"Saved to: {args.output_path}")
+    print(f"Time: {elapsed:.2f}s")
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_longcat_i2v.py

@@ -0,0 +1,210 @@
+"""
+LongCat Image-to-Video (I2V) Example Script
+
+This script demonstrates LongCat I2V inference using the FastVideo Python API.
+LongCat I2V takes an input image and generates a video from it.
+
+It runs both basic generation (50 steps) and distill+refine generation
+(16 steps distill + 50 steps refinement to 720p with BSA).
+
+Usage:
+    python examples/inference/basic/basic_longcat_i2v.py
+
+Note:
+    Refinement uses 768x768 dimensions where latent (48x48) is divisible by 8,
+    compatible with BSA chunks [4, 4, 8].
+"""
+
+import glob
+import os
+
+from fastvideo import VideoGenerator
+
+# Common prompts and settings matching the shell script examples
+PROMPT = (
+    "A woman sits at a wooden table by the window in a cozy café. She reaches out "
+    "with her right hand, picks up the white coffee cup from the saucer, and gently "
+    "brings it to her lips to take a sip. After drinking, she places the cup back on "
+    "the table and looks out the window, enjoying the peaceful atmosphere."
+)
+
+NEGATIVE_PROMPT = (
+    "Bright tones, overexposed, static, blurred details, subtitles, style, works, "
+    "paintings, images, static, overall gray, worst quality, low quality, JPEG compression "
+    "residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, "
+    "deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, "
+    "three legs, many people in the background, walking backwards"
+)
+
+# Input image path
+IMAGE_PATH = "assets/girl.png"
+
+SEED = 42
+
+
+def basic_generation():
+    """
+    Run basic LongCat I2V generation (50 steps at 480p).
+    
+    This uses the full 50-step denoising process for highest quality.
+    """
+    print("=" * 60)
+    print("LongCat I2V: Basic Generation (50 steps, 480p)")
+    print("=" * 60)
+    
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-I2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=False, # set to True if GPU is out of memory
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=False,
+    )
+    
+    output_path = "outputs_video/longcat_i2v_basic"
+    
+    generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        image_path=IMAGE_PATH,
+        output_path=output_path,
+        save_video=True,
+        height=480,
+        width=480,  # Square
+        num_frames=93,
+        num_inference_steps=50,
+        fps=15,
+        guidance_scale=4.0,
+        seed=SEED,
+    )
+    
+    print(f"\nBasic generation complete! Video saved to: {output_path}")
+    generator.shutdown()
+
+
+def distill_refine_generation():
+    """
+    Run LongCat I2V with distill+refine pipeline (16 steps + refinement to 768p).
+    
+    This uses the distilled LoRA for fast 480p generation (16 steps),
+    then refines to 768p using the refinement LoRA with BSA enabled.
+    """
+    print("\n" + "=" * 60)
+    print("LongCat I2V: Distill + Refine Pipeline")
+    print("=" * 60)
+    
+    # Stage 1: Distilled generation (16 steps at 480p)
+    print("\n[Stage 1] Distilled generation (16 steps, 480p)")
+    print("-" * 40)
+    
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-I2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=False,
+        lora_path="FastVideo/LongCat-Video-T2V-Distilled-LoRA",
+        lora_nickname="distilled",
+    )
+    
+    distill_output_path = "outputs_video/longcat_i2v_distill"
+    
+    generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        image_path=IMAGE_PATH,
+        output_path=distill_output_path,
+        save_video=True,
+        height=480,
+        width=480,  # Square
+        num_frames=93,
+        num_inference_steps=16,
+        fps=15,
+        guidance_scale=1.0,
+        seed=SEED,
+    )
+    
+    print(f"Distilled generation complete! Video saved to: {distill_output_path}")
+    generator.shutdown()
+    
+    # Stage 2: Refinement (480p -> 768p)
+    print("\n[Stage 2] Refinement (480p -> 768p with BSA)")
+    print("-" * 40)
+    
+    # Find the actual saved video file from stage 1
+    video_files = glob.glob(os.path.join(distill_output_path, "*.mp4"))
+    if not video_files:
+        raise FileNotFoundError(f"No video file found in {distill_output_path}")
+    # Use the most recently created video file
+    distill_video_path = max(video_files, key=os.path.getmtime)
+    print(f"Using stage 1 video: {distill_video_path}")
+    
+    # Create a new generator with refinement LoRA and BSA enabled
+    # Note: Refinement uses the T2V model (not I2V) since it's upscaling the generated video
+    # For BSA [4, 4, 8]: latent must be divisible by 8
+    # 768x768: latent 48x48, 48%8=0 ✓
+    refine_generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-T2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=True,
+        bsa_sparsity=0.875,
+        bsa_chunk_q=[4, 4, 4],
+        bsa_chunk_k=[4, 4, 4],
+        lora_path="FastVideo/LongCat-Video-T2V-Refinement-LoRA",
+        lora_nickname="refinement",
+    )
+    
+    refine_output_path = "outputs_video/longcat_i2v_refine_720p"
+    
+    refine_generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        output_path=refine_output_path,
+        save_video=True,
+        refine_from=distill_video_path,
+        t_thresh=0.5,
+        spatial_refine_only=False,
+        num_cond_frames=0,
+        height=720,
+        width=720,
+        num_inference_steps=50,
+        fps=30,
+        guidance_scale=1.0,
+        seed=SEED,
+    )
+    
+    print(f"Refinement complete! Video saved to: {refine_output_path}")
+    refine_generator.shutdown()
+
+
+def main():
+    """Run both basic and distill+refine generation pipelines."""
+    print("\n" + "=" * 60)
+    print("LongCat Image-to-Video Example")
+    print("=" * 60 + "\n")
+    
+    # Run basic generation
+    basic_generation()
+    
+    # Run distill+refine pipeline
+    distill_refine_generation()
+    
+    print("\n" + "=" * 60)
+    print("All generations complete!")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()
+
+