fix

.buildkite/pipeline.yml

@@ -22,7 +22,7 @@ steps:
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
-                command: "timeout 20m .buildkite/scripts/pr_test.sh"
+                command: "timeout 15m .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 20m .buildkite/scripts/pr_test.sh"
+                command: "timeout 15m .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 60m .buildkite/scripts/pr_test.sh"
+                command: "timeout 45m .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 20m .buildkite/scripts/pr_test.sh"
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
                 label: "LoRA Inference Tests"
                 env:
                   - TEST_TYPE=inference_lora
@@ -129,7 +129,11 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/**"
-                - "fastvideo-kernel/**"
+                - "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"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -141,7 +145,10 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/**"
-                - "fastvideo-kernel/**"
+                - "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"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -152,16 +159,48 @@ steps:
                 agents:
                   queue: "default"
             - path:
-                - "fastvideo-kernel/**"
+                - "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"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
                 command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "Kernel Tests"
+                label: "Precision Tests STA"
                 env:
-                  - TEST_TYPE=kernel_tests
+                  - TEST_TYPE=precision_sta
+                agents:
+                  queue: "default"
             - path:
-                - "fastvideo-kernel/**"
+                - "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/attention/backends/vmoba.py"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
@@ -183,14 +222,14 @@ 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"
+            - 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

@@ -93,9 +93,13 @@ case "$TEST_TYPE" in
         log "Running inference STA tests..."
         MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_inference_tests_STA"
         ;;
-    "kernel_tests")
-        log "Running kernel tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_kernel_tests"
+    "precision_sta")
+        log "Running precision STA tests..."
+        MODAL_COMMAND="$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 $MODAL_TEST_FILE::run_precision_tests_VSA"
         ;;
     "inference_lora")
         log "Running LoRA tests..."
@@ -114,6 +118,10 @@ case "$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"

.github/workflows/fastvideo-kernel-publish.yml

@@ -1,207 +0,0 @@
-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="${CMAKE_ARGS:-} -DFASTVIDEO_KERNEL_BUILD_TK=ON -DCMAKE_CUDA_ARCHITECTURES=90a"
-          
-          # Build standard wheel (no local version suffix) for PyPI
-          python -m build --wheel --outdir dist
-          
-          # Fix the wheel to be manylinux compliant
-          pip install auditwheel
-          # Target manylinux_2_35 (Ubuntu 22.04 native)
-          auditwheel repair dist/*.whl --plat manylinux_2_35_x86_64 -w fixed_dist
-          # Move fixed wheels back to dist for upload consistency
-          rm dist/*.whl
-          mv fixed_dist/*.whl dist/
-
-      - name: Upload wheel artifact
-        # Only upload if it's the "main" CUDA version we want on PyPI
-        # We upload all to artifacts for inspection/GH releases, but give them distinct artifact names
-        uses: actions/upload-artifact@v4
-        with:
-          name: fastvideo_kernel-py${{ 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

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

.gitmodules

@@ -1,6 +1,7 @@
-[submodule "fastvideo-kernel/include/tk"]
-	path = fastvideo-kernel/include/tk
+[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
 	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

@@ -4,7 +4,7 @@ default_stages:
 exclude: |
   (?x)(
     fastvideo/third_party/.*|
-    fastvideo-kernel/.*|
+    csrc/.*|
     assets/.*|
     tests/.*|
     demo/.*|

README.md

@@ -1,47 +1,41 @@
 <div align="center">
 <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://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> |
+    | 🕹️ <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/c7g1qdD" target="_blank"> <b> WeChat </b> </a> |
 </p>
 
-**FastVideo is a unified post-training and inference framework for accelerated video generation.**
+<div align="center">
+<img src=assets/fastwan.png width="90%"/>
+</div>
 
 ## NEWS
 - ```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>
-
+- ```2025/08/04```: 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/).
 - ```2025/06/14```: 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 for bidirectional and autoregressive models:
+- 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
   - Support full finetuning and LoRA finetuning for state-of-the-art open video DiTs
-  - 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.
+  - Scalable training with FSDP2, sequence parallelism, and selective activation checkpointing, with near linear scaling to 64 GPUs
 - State-of-the-art performance optimizations for inference
-  - 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.
+  - [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)
 - 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:
@@ -131,7 +125,7 @@ For a more detailed guide, please see our [inference quick start](https://hao-ai
 ## 🤝 Contributing
 
 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).
+See details in [development roadmap](https://github.com/hao-ai-lab/FastVideo/issues/468).
 ## Acknowledgement
 We learned and reused code from the following projects:
 - [Wan-Video](https://github.com/Wan-Video)

benchmarks/fvd/README.md

@@ -38,8 +38,7 @@ python -m benchmarks.fvd.cli \
     --num-frames 32 \
     --clip-strategy random \
     --batch-size 32 \
-    --seed 42 \
-    --extractor clip
+    --seed 42
 ```
 
 **Standard protocols:**
@@ -52,8 +51,6 @@ python -m benchmarks.fvd.cli \
     --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
@@ -61,7 +58,7 @@ 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)
+    --cache-real-features cache/real  # Directory path (will save/load cache/real/real_features.pkl)
 ```
 
 Run `python -m benchmarks.fvd.cli --help` for all options.
@@ -86,7 +83,6 @@ 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
@@ -101,6 +97,7 @@ print(f"FVD: {results['fvd']:.2f}")
 
 ## Notes
 
+- I3D model auto-downloads from Hugging Face on first run
 - 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

@@ -13,8 +13,7 @@ from .fvd import (
     compute_statistics,
     FVDConfig,
 )
-from .feature_extractors import (BaseFeatureExtractor, I3DFeatureExtractor,
-                                 load_extractor)
+from .i3d_model import I3DFeatureExtractor
 from .video_utils import (
     load_video_auto,
     sample_clips_from_video,
@@ -28,9 +27,7 @@ __all__ = [
     'compute_frechet_distance',
     'compute_statistics',
     'FVDConfig',
-    'BaseFeatureExtractor',
     'I3DFeatureExtractor',
-    'load_extractor',
     'load_video_auto',
     'sample_clips_from_video',
     'load_video_clips_streaming',

benchmarks/fvd/cli.py

@@ -1,64 +1,122 @@
 import argparse
+import json
 import sys
-import traceback
+from pathlib import Path
+
 from .fvd import compute_fvd_with_config, FVDConfig
 
 
 def main() -> int:
     parser = argparse.ArgumentParser(
-        description='Compute Fréchet Video Distance (FVD)')
+        description='Compute Fréchet Video Distance (FVD)',
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Examples:
+  # Standard FVD2048_16f protocol
+  python -m fastvideo.benchmarks.fvd.cli \\
+      --real-path data/real/ \\
+      --gen-path outputs/gen/ \\
+      --protocol fvd2048_16f
+
+  # Custom configuration
+  python -m fastvideo.benchmarks.fvd.cli \\
+      --real-path data/real/ \\
+      --gen-path outputs/gen/ \\
+      --num-videos 1024 \\
+      --num-frames 32 \\
+      --clip-strategy random \\
+      --frame-stride 2
+        """)
 
     # Required arguments
     parser.add_argument('--real-path',
                         type=str,
                         required=True,
-                        help='Path to real videos')
+                        help='Path to real videos directory')
     parser.add_argument('--gen-path',
                         type=str,
                         required=True,
-                        help='Path to generated videos')
+                        help='Path to generated videos directory')
 
-    # Extractor selection
-    parser.add_argument('--extractor',
-                        type=str,
-                        default='i3d',
-                        choices=['i3d', 'clip', 'videomae'],
-                        help='Feature extractor model to use (default: i3d)')
+    # Reproducibility
+    parser.add_argument(
+        '--seed',
+        type=int,
+        default=None,
+        help='Random seed for reproducibility (np.random, random, torch)')
 
-    # Standard args
-    parser.add_argument('--seed',
-                        type=int,
-                        default=None,
-                        help='Random seed for reproducibility')
+    # Protocol presets
     parser.add_argument('--protocol',
                         type=str,
                         default=None,
-                        choices=['fvd2048_16f', 'fvd2048_128f', 'quick_test'],
+                        choices=[
+                            'fvd2048_16f', 'fvd2048_128f',
+                            'fvd2048_128f_subsample8', 'quick_test'
+                        ],
                         help='Use standard protocol (overrides other settings)')
+
+    # Video selection
     parser.add_argument('--num-videos',
                         type=int,
                         default=2048,
-                        help='Number of videos to use')
+                        help='Number of videos to use (default: 2048)')
+
+    # Clip sampling
     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')
+                        help='Number of frames per clip (default: 16)')
+    parser.add_argument('--num-clips',
+                        type=int,
+                        default=1,
+                        help='Number of clips per video (default: 1)')
+    parser.add_argument(
+        '--clip-strategy',
+        type=str,
+        default='beginning',
+        choices=['beginning', 'random', 'uniform', 'middle', 'sliding', 'all'],
+        help='Clip sampling strategy (default: beginning)')
+    parser.add_argument(
+        '--frame-stride',
+        type=int,
+        default=1,
+        help='Frame stride for FPS subsampling (default: 1, no subsampling)')
+    parser.add_argument('--temporal-stride',
+                        type=int,
+                        default=1,
+                        help='Temporal stride for sliding window (default: 1)')
+
+    # Data processing
+    parser.add_argument('--no-frame-dirs',
+                        action='store_true',
+                        help='Disable frame directory support')
+
+    # Computation
     parser.add_argument('--batch-size',
                         type=int,
                         default=32,
-                        help='Batch size for feature extraction')
+                        help='Batch size for feature extraction (default: 32)')
     parser.add_argument('--device',
                         type=str,
                         default='cuda',
-                        help='Device to use (cuda or cpu)')
+                        choices=['cuda', 'cpu'],
+                        help='Device to use (default: cuda)')
+
+    # Caching
     parser.add_argument('--cache-real-features',
                         type=str,
                         default=None,
                         help='Path to cache real video features')
+    parser.add_argument('--i3d-model-path',
+                        type=str,
+                        default=None,
+                        help='Custom cache path for I3D model')
+
+    # Output
+    parser.add_argument('--output',
+                        type=str,
+                        default='fvd_results.json',
+                        help='Output JSON file (default: fvd_results.json)')
     parser.add_argument('--quiet',
                         action='store_true',
                         help='Suppress progress output')
@@ -70,36 +128,56 @@ def main() -> int:
         protocol_map = {
             'fvd2048_16f': FVDConfig.fvd2048_16f,
             'fvd2048_128f': FVDConfig.fvd2048_128f,
+            'fvd2048_128f_subsample8': FVDConfig.fvd2048_128f_subsample8,
             'quick_test': FVDConfig.quick_test,
         }
         config = protocol_map[args.protocol]()
-        # Apply overrides
+
+        # Override device and caching from args
         config.device = args.device
         config.cache_real_features = args.cache_real_features
-        config.extractor_model = args.extractor  # Apply extractor arg
+        config.i3d_model_path = args.i3d_model_path
+        config.batch_size = args.batch_size
+        config.seed = args.seed
     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)
+        # Custom config from args
+        config = FVDConfig(num_videos=args.num_videos,
+                           num_frames_per_clip=args.num_frames,
+                           num_clips_per_video=args.num_clips,
+                           clip_strategy=args.clip_strategy,
+                           frame_stride=args.frame_stride,
+                           temporal_stride=args.temporal_stride,
+                           support_frame_dirs=not args.no_frame_dirs,
+                           batch_size=args.batch_size,
+                           device=args.device,
+                           cache_real_features=args.cache_real_features,
+                           i3d_model_path=args.i3d_model_path,
+                           seed=args.seed)
 
+    # Compute FVD
     try:
-        _ = compute_fvd_with_config(
-            args.real_path,  # noqa: F841
-            args.gen_path,
-            config,
-            verbose=not args.quiet)
+        results = compute_fvd_with_config(real_videos=args.real_path,
+                                          gen_videos=args.gen_path,
+                                          config=config,
+                                          verbose=not args.quiet)
+
+        # Save results
+        output_path = Path(args.output)
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+
+        with open(output_path, 'w') as f:
+            json.dump(results, f, indent=2)
+
+        print(f"\nResults saved to {output_path}")
+        print(f"FVD: {results['fvd']:.2f}")
+        print(f"Protocol: {results['protocol']}")
 
         return 0
 
     except Exception as e:
         print(f"Error: {e}", file=sys.stderr)
-        traceback.print_exc(file=sys.stderr)
+        import traceback
+        traceback.print_exc()
         return 1
 
 

benchmarks/fvd/feature_extractors.py

@@ -1,264 +0,0 @@
-"""
-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

@@ -5,7 +5,8 @@ from pathlib import Path
 from collections.abc import Iterator
 import pickle
 from dataclasses import dataclass, field
-from .feature_extractors import BaseFeatureExtractor, load_extractor
+
+from .i3d_model import I3DFeatureExtractor
 from .video_utils import ClipSamplingStrategy, load_video_clips_streaming
 
 
@@ -54,9 +55,6 @@ class FVDConfig:
     # 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
@@ -87,7 +85,11 @@ class FVDConfig:
 
     @classmethod
     def fvd2048_16f(cls) -> 'FVDConfig':
-        """Standard FVD protocol: 2048 videos, 16 frames, beginning clip."""
+        """
+        Standard FVD protocol: 2048 videos, 16 frames, beginning clip.
+        
+        most common FVD configuration used in papers
+        """
         return cls(num_videos=2048,
                    num_frames_per_clip=16,
                    clip_strategy='beginning',
@@ -101,6 +103,18 @@ class FVDConfig:
                    clip_strategy='beginning',
                    use_streaming=True)
 
+    @classmethod
+    def fvd2048_128f_subsample8(cls) -> 'FVDConfig':
+        """
+        Long video with FPS subsampling: 2048 videos, 128 frames (every 8th).
+        Used for very long videos - samples every 8th frame
+        """
+        return cls(num_videos=2048,
+                   num_frames_per_clip=16,
+                   frame_stride=8,
+                   clip_strategy='beginning',
+                   use_streaming=True)
+
     @classmethod
     def quick_test(cls) -> 'FVDConfig':
         """Quick test config: 100 videos, 16 frames."""
@@ -110,13 +124,22 @@ class FVDConfig:
 
     def to_dict(self) -> dict:
         """Export config to dict for logging"""
-        d = self.__dict__.copy()
-        d['clip_strategy'] = str(self.clip_strategy)
-        return d
+        return {
+            'num_videos': self.num_videos,
+            'num_frames_per_clip': self.num_frames_per_clip,
+            'num_clips_per_video': self.num_clips_per_video,
+            'clip_strategy': str(self.clip_strategy),
+            'frame_stride': self.frame_stride,
+            'temporal_stride': self.temporal_stride,
+            'batch_size': self.batch_size,
+            'device': self.device,
+            'seed': self.seed,
+            'use_streaming': self.use_streaming,
+        }
 
     def __str__(self) -> str:
         """Human-readable protocol name"""
-        desc = f"FVD_{self.extractor_model.upper()}_{self.num_videos}_{self.num_frames_per_clip}f"
+        desc = f"FVD{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:
@@ -127,42 +150,57 @@ class FVDConfig:
 
 
 def extract_features_streaming(video_generator: Iterator[torch.Tensor],
-                               extractor: BaseFeatureExtractor,
+                               extractor: I3DFeatureExtractor,
                                batch_size: int = 32,
                                max_clips: int | None = None,
                                verbose: bool = True) -> np.ndarray:
     """
     Extract features from a video clip generator using streaming.
+
+    Args:
+        video_generator: Iterator yielding clips [T, C, H, W]
+        extractor: I3D feature extractor
+        batch_size: Batch size for processing
+        max_clips: Maximum clips to process (for validation)
+        verbose: Show progress
+    
+    Returns:
+        features: [N, 400] numpy array
     """
     all_features = []
     batch = []
+    clip_count = 0
 
     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)
+    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:
+        # 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())
+            features = extractor.extract_features(batch_tensor,
+                                                  batch_size=batch_size,
+                                                  verbose=False)
+            all_features.append(features.cpu().numpy())
+
+            batch = []  # Clear batch
+
+            if verbose and clip_count % (batch_size * 10) == 0:
+                print(f"Processed {clip_count} clips...")
+
+        # Stop if we've reached max_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_tensor,
+                                              batch_size=len(batch),
+                                              verbose=False)
+        all_features.append(features.cpu().numpy())
 
     if len(all_features) == 0:
         raise RuntimeError("No features extracted - check video loading")
@@ -176,17 +214,14 @@ def extract_features_streaming(video_generator: Iterator[torch.Tensor],
 
 
 def load_or_compute_features(videos: str | Path | torch.Tensor,
-                             extractor: BaseFeatureExtractor,
+                             extractor: I3DFeatureExtractor,
                              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"
-
+        cache_file = Path(cache_path) / f"{cache_name}.pkl"
         if cache_file.exists():
             print(f"Loading cached features from {cache_file}")
             with open(cache_file, 'rb') as f:
@@ -194,25 +229,19 @@ def load_or_compute_features(videos: str | Path | torch.Tensor,
 
             # 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...")
+                print("Recomputing features...")
             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)):
+    # Compute features
+    if isinstance(videos, str | Path):
         target_size = (224, 224) if config.resize_before_extraction else None
 
         video_generator = load_video_clips_streaming(
@@ -233,7 +262,9 @@ def load_or_compute_features(videos: str | Path | torch.Tensor,
                                               batch_size=config.batch_size,
                                               max_clips=max_clips,
                                               verbose=True)
+
     else:
+        # Already a tensor
         print(f"Extracting features from {len(videos)} video tensors...")
         features = extractor.extract_features(videos,
                                               batch_size=config.batch_size,
@@ -252,10 +283,9 @@ def load_or_compute_features(videos: str | Path | torch.Tensor,
 
     # Cache features if requested
     if cache_path is not None:
-        script_dir = Path(__file__).parent
-        cache_dir = script_dir / cache_path
+        cache_dir = Path(cache_path)
         cache_dir.mkdir(parents=True, exist_ok=True)
-        cache_file = cache_dir / f"{config.extractor_model}_{cache_name}.pkl"
+        cache_file = cache_dir / f"{cache_name}.pkl"
         print(f"Caching features to {cache_file}")
         with open(cache_file, 'wb') as f:
             pickle.dump(features, f)
@@ -263,34 +293,79 @@ def load_or_compute_features(videos: str | Path | torch.Tensor,
     return features
 
 
+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:
+    """
+    Compute Fréchet Video Distance (FVD)
+    
+    For advanced control, use compute_fvd_with_config() instead.
+    
+    Args:
+        real_videos: Path to real videos or tensor [N, T, C, H, W]
+        gen_videos: Path to generated videos or tensor [N, T, C, H, W]
+        num_frames: Frames per video (default: 16)
+        batch_size: Batch size (default: 32)
+        device: 'cuda' or 'cpu' (default: 'cuda')
+        num_videos: Max videos (default: 2048)
+        cache_real_features: Cache path for real features
+        i3d_model_path: Custom I3D model cache path
+        seed: Random seed for reproducibility
+        verbose: Print progress
+    
+    Returns:
+        FVD score (float). Lower is better.
+    """
+    num_videos = num_videos if num_videos is not None else 2048
+
+    config = FVDConfig(
+        num_videos=num_videos,
+        num_frames_per_clip=num_frames,
+        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']
+
+
 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}")
+    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)
+            - 'config': Configuration dict
+    
+    Example:
+        >>> config = FVDConfig.fvd2048_16f()
+        >>> results = compute_fvd_with_config('data/real/', 'outputs/gen/', config)
+        >>> print(f"FVD: {results['fvd']:.2f}")
+        >>> print(f"Protocol: {results['protocol']}")  # "FVD2048_16f"
     """
-
     # Seed for reproducibility
     if config.seed is not None:
         import random as _rnd
@@ -303,20 +378,18 @@ def compute_fvd_with_config(real_videos: str | Path | torch.Tensor,
     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
+    # Initialize I3D
     if verbose:
-        print(
-            f"\nInitializing {config.extractor_model.upper()} model on {config.device}..."
-        )
+        print(f"\nInitializing I3D model on {config.device}...")
 
-    extractor = load_extractor(config.extractor_model, device=config.device)
+    extractor = I3DFeatureExtractor(device=config.device,
+                                    cache_dir=config.i3d_model_path)
 
     # Extract features
     if verbose:
@@ -361,45 +434,14 @@ def compute_fvd_with_config(real_videos: str | Path | torch.Tensor,
 
     if verbose:
         print(f"\n{'='*70}")
-        print(f"FVD Score ({config.extractor_model.upper()}): {fvd:.4f}")
+        print(f"FVD Score: {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/run_fvd.py

@@ -1,53 +1,33 @@
 import sys
 from pathlib import Path
+from benchmarks.fvd.fvd import FVDConfig, compute_fvd_with_config
 
 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:
+    # Get script directory
     script_dir = Path(__file__).parent.resolve()
 
-    # Define directories
+    clip_strategy = 'beginning'  # Options: 'uniform', 'random', 'beginning', 'end', 'all'
+    cfg = FVDConfig(
+        num_videos=650,
+        num_frames_per_clip=16,
+        num_clips_per_video=1,
+        clip_strategy=clip_strategy,
+        frame_stride=1,
+        batch_size=32,
+        device='cuda',
+        seed=42,
+        cache_real_features=str(script_dir / f'fvd-cache/{clip_strategy}'),
+    )
+
     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}")
+    results = compute_fvd_with_config(real_dir, gen_dir, cfg, verbose=True)
+    print(f"FVD = {results['fvd']:.2f}")
 
 
 if __name__ == '__main__':

benchmarks/scripts/run.sh

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

csrc/attn/README.md

@@ -0,0 +1,113 @@
+
+
+# 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

@@ -0,0 +1,145 @@
+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

@@ -0,0 +1,224 @@
+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

@@ -0,0 +1,217 @@
+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

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

csrc/attn/sliding_tile_attn/README.md

@@ -0,0 +1,96 @@
+
+# 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

@@ -0,0 +1,15 @@
+### 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

@@ -0,0 +1,76 @@
+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

@@ -0,0 +1,23 @@
+#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

@@ -0,0 +1,49 @@
+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/sliding_tile_attn/st_attn/st_attn_h100.cu

@@ -451,10 +451,6 @@ sta_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o,
 
         globals g{qg_arg, kg_arg, vg_arg, lg_arg, og_arg, static_cast<int>(seq_len),  static_cast<int>(text_length), static_cast<int>(hr)};
 
-        // Shared memory size for the kernel.
-        // We use the maximum available shared memory (kittens::MAX_SHARED_MEMORY) 
-        // which is approximately 227KB on H100, necessary for the high-performance 
-        // TMA-based attention tiles with multiple stages.
         constexpr int mem_size = kittens::MAX_SHARED_MEMORY;
         int threads = NUM_WORKERS * kittens::WARP_THREADS;
         if (has_text) {
@@ -462,31 +458,104 @@ sta_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o,
             dim3 grid_image(seq_len/(CONSUMER_WARPGROUPS*kittens::TILE_ROW_DIM<bf16>*4)-2, qo_heads, batch);
             dim3 grid_text(2, qo_heads, batch);
             if (!process_text) {
-#define LAUNCH_IMAGE_KER(DT_VAL, DH_VAL, DW_VAL) \
-                    cudaFuncSetAttribute( \
-                        fwd_attend_ker<128, false, false, true, DT_VAL, DH_VAL, DW_VAL, 5, 6, 10>, \
-                        cudaFuncAttributeMaxDynamicSharedMemorySize, \
-                        mem_size \
-                    ); \
-                    fwd_attend_ker<128, false, false, true, DT_VAL, DH_VAL, DW_VAL, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3) {
 
-                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3)      { LAUNCH_IMAGE_KER(1, 1, 1); }
-                else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 5) { LAUNCH_IMAGE_KER(1, 1, 2); }
-                else if (kernel_t_size == 5 && kernel_h_size == 3 && kernel_w_size == 3) { LAUNCH_IMAGE_KER(2, 1, 1); }
-                else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 5) { LAUNCH_IMAGE_KER(1, 2, 2); }
-                else if (kernel_t_size == 5 && kernel_h_size == 6 && kernel_w_size == 1) { LAUNCH_IMAGE_KER(2, 3, 0); }
-                else if (kernel_t_size == 5 && kernel_h_size == 3 && kernel_w_size == 5) { LAUNCH_IMAGE_KER(2, 1, 2); }
-                else if (kernel_t_size == 5 && kernel_h_size == 5 && kernel_w_size == 5) { LAUNCH_IMAGE_KER(2, 2, 2); }
-                else if (kernel_t_size == 5 && kernel_h_size == 5 && kernel_w_size == 7) { LAUNCH_IMAGE_KER(2, 2, 3); }
-                else if (kernel_t_size == 5 && kernel_h_size == 6 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(2, 3, 5); }
-                else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(1, 3, 5); }
-                else if (kernel_t_size == 5 && kernel_h_size == 1 && kernel_w_size == 1) { LAUNCH_IMAGE_KER(2, 0, 0); }
-                else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(0, 3, 5); }
-                else if (kernel_t_size == 5 && kernel_h_size == 1 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(2, 0, 5); }
-                else {
-                    TORCH_CHECK(false, "Invalid kernel size: ", kernel_t_size, "x", kernel_h_size, "x", kernel_w_size);
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 1, 1, 1, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 1, 1, 1, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                }  else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 5) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 1, 1, 2, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true,1, 1, 2, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 5 && kernel_h_size == 3 && kernel_w_size == 3) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 1, 1, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 1, 1, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                }else if (kernel_t_size ==3 && kernel_h_size == 5 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 1, 2, 2, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 1, 2, 2, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size ==5 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 3, 0, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 3, 0, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size ==5 && kernel_h_size == 3 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 1, 2, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 1, 2, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 5 && kernel_h_size == 5 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 2, 2, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 2, 2, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 5 && kernel_h_size == 5 && kernel_w_size == 7){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 2, 3, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 2, 3, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 5 && kernel_h_size == 6 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 3, 5, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 3, 5, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 5 && kernel_h_size == 1 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 0, 0, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 0, 0, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 0, 3, 5, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 0, 3, 5, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 5 && kernel_h_size == 1 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 0, 5, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true,2, 0, 5, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else {
+                    // print error
+                    std::cout << "Invalid kernel size" << std::endl;
+                    //print kernel size
+                    std::cout << "Kernel size: " << kernel_t_size << " " << kernel_h_size << " " << kernel_w_size << std::endl;
                 }
-#undef LAUNCH_IMAGE_KER
             } else {
                 cudaFuncSetAttribute(
                     fwd_attend_ker<128, false, true, true, 1, 1, 1, 5, 6, 10>,
@@ -499,67 +568,266 @@ sta_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o,
         } else {
             dim3 grid_image(seq_len/(CONSUMER_WARPGROUPS*kittens::TILE_ROW_DIM<bf16>*4), qo_heads, batch);
             if (kernel_aspect_ratio_flag == 2){
-#define LAUNCH_IMAGE_KER(DT_VAL, DH_VAL, DW_VAL) \
-                    cudaFuncSetAttribute( \
-                        fwd_attend_ker<128, false, false, false, DT_VAL, DH_VAL, DW_VAL, 6, 6, 6>, \
-                        cudaFuncAttributeMaxDynamicSharedMemorySize, \
-                        mem_size \
-                    ); \
-                    fwd_attend_ker<128, false, false, false, DT_VAL, DH_VAL, DW_VAL, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 1, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 1, 1, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
 
-                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3)      { LAUNCH_IMAGE_KER(1, 1, 1); }
-                else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 6) { LAUNCH_IMAGE_KER(1, 1, 3); }
-                else if (kernel_t_size == 6 && kernel_h_size == 3 && kernel_w_size == 3) { LAUNCH_IMAGE_KER(3, 1, 1); }
-                else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 6) { LAUNCH_IMAGE_KER(1, 3, 3); }
-                else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 3) { LAUNCH_IMAGE_KER(1, 3, 1); }
-                else if (kernel_t_size == 6 && kernel_h_size == 3 && kernel_w_size == 6) { LAUNCH_IMAGE_KER(3, 1, 3); }
-                else if (kernel_t_size == 6 && kernel_h_size == 6 && kernel_w_size == 6) { LAUNCH_IMAGE_KER(3, 3, 3); }
-                else if (kernel_t_size == 6 && kernel_h_size == 1 && kernel_w_size == 1) { LAUNCH_IMAGE_KER(3, 0, 0); }
-                else if (kernel_t_size == 6 && kernel_h_size == 1 && kernel_w_size == 6) { LAUNCH_IMAGE_KER(3, 0, 3); }
-                else if (kernel_t_size == 6 && kernel_h_size == 6 && kernel_w_size == 1) { LAUNCH_IMAGE_KER(3, 3, 0); }
-                else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 6) { LAUNCH_IMAGE_KER(0, 3, 3); }
-                else if (kernel_t_size == 1 && kernel_h_size == 1 && kernel_w_size == 6) { LAUNCH_IMAGE_KER(0, 0, 3); }
-                else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 1) { LAUNCH_IMAGE_KER(0, 3, 0); }
-                else {
-                    TORCH_CHECK(false, "Invalid kernel size: ", kernel_t_size, "x", kernel_h_size, "x", kernel_w_size);
+                }  else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 6) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,1, 1, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 6 && kernel_h_size == 3 && kernel_w_size == 3) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 1, 1, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 1, 1, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size ==3 && kernel_h_size == 6 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                }else if (kernel_t_size ==3 && kernel_h_size == 6 && kernel_w_size == 3){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 1, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 1, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size ==6 && kernel_h_size == 3 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 1, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 1, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 6 && kernel_h_size == 6 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 3, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 3, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 6 && kernel_h_size == 1 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 0, 0, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 0, 0, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 6 && kernel_h_size == 1 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 0, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 0, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                }  else if (kernel_t_size == 6 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 3, 0, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 3, 0, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 3, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                }  else if (kernel_t_size == 1 && kernel_h_size == 1 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 0, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 0, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 0, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 3, 0, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 6 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 3, 0, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 3, 0, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 6 && kernel_h_size == 1 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 0, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 0, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else {
+                    // print error
+                    std::cout << "Invalid kernel size" << std::endl;
+                    //print kernel size
+                    std::cout << "Kernel size: " << kernel_t_size << " " << kernel_h_size << " " << kernel_w_size << std::endl;
                 }
-#undef LAUNCH_IMAGE_KER
             }
             else if (kernel_aspect_ratio_flag == 3) {
-#define LAUNCH_IMAGE_KER(DT_VAL, DH_VAL, DW_VAL) \
-                    cudaFuncSetAttribute( \
-                        fwd_attend_ker<128, false, false, false, DT_VAL, DH_VAL, DW_VAL, 3, 6, 10>, \
-                        cudaFuncAttributeMaxDynamicSharedMemorySize, \
-                        mem_size \
-                    ); \
-                    fwd_attend_ker<128, false, false, false, DT_VAL, DH_VAL, DW_VAL, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3) {
 
-                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3)      { LAUNCH_IMAGE_KER(1, 1, 1); }
-                else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 5) { LAUNCH_IMAGE_KER(1, 1, 2); }
-                else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 5) { LAUNCH_IMAGE_KER(1, 2, 2); }
-                else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 1) { LAUNCH_IMAGE_KER(1, 3, 0); }
-                else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 7) { LAUNCH_IMAGE_KER(1, 2, 3); }
-                else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 9) { LAUNCH_IMAGE_KER(1, 2, 4); }
-                else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(1, 3, 5); }
-                else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 3) { LAUNCH_IMAGE_KER(1, 3, 1); }
-                else if (kernel_t_size == 3 && kernel_h_size == 1 && kernel_w_size == 1) { LAUNCH_IMAGE_KER(1, 0, 0); }
-                else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(0, 3, 5); }
-                else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(0, 2, 5); }
-                else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 7) { LAUNCH_IMAGE_KER(0, 3, 3); }
-                else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 7) { LAUNCH_IMAGE_KER(0, 2, 3); }
-                else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 9) { LAUNCH_IMAGE_KER(0, 2, 4); }
-                else if (kernel_t_size == 3 && kernel_h_size == 1 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(1, 0, 5); }
-                else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(1, 1, 5); }
-                else if (kernel_t_size == 1 && kernel_h_size == 3 && kernel_w_size == 10){ LAUNCH_IMAGE_KER(0, 1, 5); }
-                else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 5) { LAUNCH_IMAGE_KER(0, 3, 2); }
-                else {
-                    TORCH_CHECK(false, "Invalid kernel size: ", kernel_t_size, "x", kernel_h_size, "x", kernel_w_size);
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 1, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 1, 1, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 5) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 2, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,1, 1, 2, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 2, 2, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 2, 2, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 0, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 0, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 7){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 2, 3, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 2, 3, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 9){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 2, 4, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 2, 4, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 3){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 1, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 1, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 1 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 0, 0, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 0, 0, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 3, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 2, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 2, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 7){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 3, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 3, 3, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 7){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 2, 3, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 2, 3, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 9){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 2, 4, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 2, 4, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 1 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 0, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,1, 0, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,1, 1, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 3 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 1, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,0, 1, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 2, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,0, 3, 2, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else {
+                    // print error
+                    std::cout << "Invalid kernel size" << std::endl;
+                    //print kernel size
+                    std::cout << "Kernel size: " << kernel_t_size << " " << kernel_h_size << " " << kernel_w_size << std::endl;
                 }
-#undef LAUNCH_IMAGE_KER
             }
 
             else {
-                TORCH_CHECK(false, "Unsupported kernel_aspect_ratio_flag: ", kernel_aspect_ratio_flag);
+                std::cout << "Unsupported kernel_aspect_ratio_flag: " << kernel_aspect_ratio_flag << std::endl;
             }
 
         }

csrc/attn/tests/test_sta.py

@@ -1,6 +1,6 @@
 import torch
-from .support_flex_sta import get_sliding_tile_attention_mask
-from fastvideo_kernel import sliding_tile_attention
+from flex_sta_ref import get_sliding_tile_attention_mask
+from st_attn import sliding_tile_attention
 from torch.nn.attention.flex_attention import flex_attention
 # from flash_attn_interface import flash_attn_func
 from tqdm import tqdm
@@ -73,23 +73,15 @@ def check_correctness(b, h, n, d, causal, mean, std, num_iterations=50, error_mo
 
 
 # Example usage
-def test_sliding_tile_attention():
-    if not torch.cuda.is_available():
-        return
-
-    b, h, d = 2, 24, 128
-    n = 69120  # Sequence length
-    causal = False
-    mean = 1e-1
-    std = 10
-
-    # Run correctness check directly
-    results = check_correctness(b, h, n, d, causal, mean, std, error_mode='output')
-    assert results['TK vs FLEX']['avg_diff'] < 3e-6, f"Average difference: {results['TK vs FLEX']['avg_diff']} is too large"
-    assert results['TK vs FLEX']['max_diff'] < 4e-2, f"Maximum difference: {results['TK vs FLEX']['max_diff']} is too large"
-    print(f"Average difference: {results['TK vs FLEX']['avg_diff']}")
-    print(f"Maximum difference: {results['TK vs FLEX']['max_diff']}")
-
-if __name__ == "__main__":
-    test_sliding_tile_attention()
+b, h, d = 2, 24, 128
+n = 69120  # Sequence length
+causal = False
+mean = 1e-1
+std = 10
 
+# Run correctness check directly
+results = check_correctness(b, h, n, d, causal, mean, std, error_mode='output')
+assert results['TK vs FLEX']['avg_diff'] < 3e-6, f"Average difference: {results['TK vs FLEX']['avg_diff']} is too large"
+assert results['TK vs FLEX']['max_diff'] < 4e-2, f"Maximum difference: {results['TK vs FLEX']['max_diff']} is too large"
+print(f"Average difference: {results['TK vs FLEX']['avg_diff']}")
+print(f"Maximum difference: {results['TK vs FLEX']['max_diff']}")

csrc/attn/tests/test_vsa.py

@@ -0,0 +1,156 @@
+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

@@ -0,0 +1,54 @@
+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

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

csrc/attn/video_sparse_attn/README.md

@@ -0,0 +1,61 @@
+
+
+# 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

@@ -0,0 +1,15 @@
+### 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

@@ -0,0 +1,81 @@
+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

@@ -0,0 +1,27 @@
+#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

@@ -0,0 +1,80 @@
+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_attn_triton.py

@@ -8,6 +8,7 @@ This is a Triton implementation of the Flash Attention v2 algorithm from Tri Dao
 Credits: OpenAI kernel team
 """
 
+import pytest
 import torch
 import triton
 import triton.language as tl
@@ -16,6 +17,7 @@ import triton.language as tl
 import math  # small utility needed by the sparse wrapper
 # ──────────────────────────── SPARSE ADDITION END ─────────────────────────────
 
+
 # We don't run auto-tuning every time to keep the tutorial fast. Keeping
 # the code below and commenting out the equivalent parameters is convenient for
 # re-tuning.
@@ -27,92 +29,65 @@ configs = [
     for w in [4, 8]\
 ]
 
-
 # ──────────────────────────── SPARSE ADDITION BEGIN ───────────────────────────
 @triton.autotune(configs, key=["N_CTX", "HEAD_DIM"])
 @triton.jit
-def _attn_fwd_sparse(
-        Q,
-        K,
-        V,
-        sm_scale,  #
-        q2k_index,
-        q2k_num,
-        max_kv_blks,  #
-        variable_block_sizes,
-        M,
-        Out,  #
-        stride_qz,
-        stride_qh,
-        stride_qm,
-        stride_qk,
-        stride_kz,
-        stride_kh,
-        stride_kn,
-        stride_kk,
-        stride_vz,
-        stride_vh,
-        stride_vk,
-        stride_vn,
-        stride_oz,
-        stride_oh,
-        stride_om,
-        stride_on,
-        Z,
-        H,
-        N_CTX,  #
-        HEAD_DIM: tl.constexpr,  #
-        BLOCK_M: tl.constexpr,
-        BLOCK_N: tl.constexpr,
-        STAGE: tl.constexpr):
+def _attn_fwd_sparse(Q, K, V, sm_scale,                         #
+                     q2k_index, q2k_num, max_kv_blks,           #
+                     variable_block_sizes,
+                     M, Out,                                    #
+                     stride_qz, stride_qh, stride_qm, stride_qk,
+                     stride_kz, stride_kh, stride_kn, stride_kk,
+                     stride_vz, stride_vh, stride_vk, stride_vn,
+                     stride_oz, stride_oh, stride_om, stride_on,
+                     Z, H, N_CTX,                               #
+                     HEAD_DIM: tl.constexpr,                    #
+                     BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr,
+                     STAGE: tl.constexpr):
     """
     64×64 **block-sparse** forward kernel. Back-prop kernels remain dense
     (32×64 and 64×32) – memory footprint unchanged.
     """
 
     # ----- program-id mapping -----
-    q_blk = tl.program_id(0)  # Q-tile index
-    off_hz = tl.program_id(1)  # fused (batch, head)
-    b = off_hz // H
-    h = off_hz % H
+    q_blk   = tl.program_id(0)          # Q-tile index
+    off_hz  = tl.program_id(1)          # fused (batch, head)
+    b       = off_hz // H
+    h       = off_hz %  H
     q_tiles = N_CTX // BLOCK_M
     meta_base = ((b * H + h) * q_tiles + q_blk)
 
-    kv_blocks = tl.load(q2k_num + meta_base)  # int32
-    kv_ptr = q2k_index + meta_base * max_kv_blks  # ptr to list
+    kv_blocks = tl.load(q2k_num  + meta_base)                 # int32
+    kv_ptr    = q2k_index + meta_base * max_kv_blks           # ptr to list
 
     # ----- base pointers -----
-    qvk_off = (b.to(tl.int64) * stride_qz + h.to(tl.int64) * stride_qh)
+    qvk_off = (b.to(tl.int64) * stride_qz +
+               h.to(tl.int64) * stride_qh)
 
-    Q_ptr = tl.make_block_ptr(base=Q + qvk_off,
-                              shape=(N_CTX, HEAD_DIM),
-                              strides=(stride_qm, stride_qk),
-                              offsets=(q_blk * BLOCK_M, 0),
-                              block_shape=(BLOCK_M, HEAD_DIM),
-                              order=(1, 0))
+    Q_ptr = tl.make_block_ptr(
+        base=Q + qvk_off, shape=(N_CTX, HEAD_DIM),
+        strides=(stride_qm, stride_qk),
+        offsets=(q_blk * BLOCK_M, 0),
+        block_shape=(BLOCK_M, HEAD_DIM), order=(1, 0))
 
-    K_base = tl.make_block_ptr(base=K + qvk_off,
-                               shape=(HEAD_DIM, N_CTX),
-                               strides=(stride_kk, stride_kn),
-                               offsets=(0, 0),
-                               block_shape=(HEAD_DIM, BLOCK_N),
-                               order=(0, 1))
+    K_base = tl.make_block_ptr(
+        base=K + qvk_off, shape=(HEAD_DIM, N_CTX),
+        strides=(stride_kk, stride_kn),
+        offsets=(0, 0),
+        block_shape=(HEAD_DIM, BLOCK_N), order=(0, 1))
 
-    v_order: tl.constexpr = (0, 1) if V.dtype.element_ty == tl.float8e5 else (1,
-                                                                              0)
-    V_base = tl.make_block_ptr(base=V + qvk_off,
-                               shape=(N_CTX, HEAD_DIM),
-                               strides=(stride_vk, stride_vn),
-                               offsets=(0, 0),
-                               block_shape=(BLOCK_N, HEAD_DIM),
-                               order=v_order)
+    v_order: tl.constexpr = (0, 1) if V.dtype.element_ty == tl.float8e5 else (1, 0)
+    V_base = tl.make_block_ptr(
+        base=V + qvk_off, shape=(N_CTX, HEAD_DIM),
+        strides=(stride_vk, stride_vn),
+        offsets=(0, 0),
+        block_shape=(BLOCK_N, HEAD_DIM), order=v_order)
 
-    O_ptr = tl.make_block_ptr(base=Out + qvk_off,
-                              shape=(N_CTX, HEAD_DIM),
-                              strides=(stride_om, stride_on),
-                              offsets=(q_blk * BLOCK_M, 0),
-                              block_shape=(BLOCK_M, HEAD_DIM),
-                              order=(1, 0))
+    O_ptr = tl.make_block_ptr(
+        base=Out + qvk_off, shape=(N_CTX, HEAD_DIM),
+        strides=(stride_om, stride_on),
+        offsets=(q_blk * BLOCK_M, 0),
+        block_shape=(BLOCK_M, HEAD_DIM), order=(1, 0))
 
     # ----- accumulators -----
     offs_m = q_blk * BLOCK_M + tl.arange(0, BLOCK_M)
@@ -152,30 +127,23 @@ def _attn_fwd_sparse(
     acc = acc / l_i[:, None]
     tl.store(M + off_hz * N_CTX + offs_m, m_i)
     tl.store(O_ptr, acc.to(Out.type.element_ty))
-
-
 # ──────────────────────────── SPARSE ADDITION END ─────────────────────────────
 
 
+
+
 @triton.jit
-def _attn_bwd_preprocess(
-        O,
-        DO,  #
-        Delta,  #
-        Z,
-        H,
-        N_CTX,  #
-        BLOCK_M: tl.constexpr,
-        HEAD_DIM: tl.constexpr  #
-):
+def _attn_bwd_preprocess(O, DO,  #
+                         Delta,  #
+                         Z, H, N_CTX,  #
+                         BLOCK_M: tl.constexpr, HEAD_DIM: tl.constexpr  #
+                         ):
     off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
     off_hz = tl.program_id(1)
     off_n = tl.arange(0, HEAD_DIM)
     # load
-    o = tl.load(O + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM +
-                off_n[None, :])
-    do = tl.load(DO + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM +
-                 off_n[None, :]).to(tl.float32)
+    o = tl.load(O + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :])
+    do = tl.load(DO + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :]).to(tl.float32)
     delta = tl.sum(o * do, axis=1)
     # write-back
     tl.store(Delta + off_hz * N_CTX + off_m, delta)
@@ -183,32 +151,19 @@ def _attn_bwd_preprocess(
 
 # The main inner-loop logic for computing dK and dV.
 @triton.jit
-def _attn_bwd_dkdv(
-        dk,
-        dv,  #
-        Q,
-        k,
-        v,
-        sm_scale,  #
-        DO,  #
-        M,
-        D,  #
-        k2q_index,
-        k2q_num,
-        max_q_blks,
-        variable_block_sizes,
-        # shared by Q/K/V/DO.
-        stride_tok,
-        stride_d,  #
-        H,
-        N_CTX,
-        BLOCK_M1: tl.constexpr,  #
-        BLOCK_N1: tl.constexpr,  #
-        HEAD_DIM: tl.constexpr,  #
-        # Filled in by the wrapper.
-    start_n,
-        start_m,
-        num_steps):
+def _attn_bwd_dkdv(dk, dv,  #
+                   Q, k, v, sm_scale,  #
+                   DO,  #
+                   M, D,  #
+                   k2q_index, k2q_num, max_q_blks,
+                   variable_block_sizes,
+                   # shared by Q/K/V/DO.
+                   stride_tok, stride_d,  #
+                   H, N_CTX, BLOCK_M1: tl.constexpr,  #
+                   BLOCK_N1: tl.constexpr,  #
+                   HEAD_DIM: tl.constexpr,  #
+                   # Filled in by the wrapper.
+                   start_n, start_m, num_steps):
     offs_m = start_m + tl.arange(0, BLOCK_M1)
     offs_n = start_n + tl.arange(0, BLOCK_N1)
     offs_k = tl.arange(0, HEAD_DIM)
@@ -217,20 +172,21 @@ def _attn_bwd_dkdv(
     # BLOCK_N1 must be a multiple of BLOCK_M1, otherwise the code wouldn't work.
     tl.static_assert(BLOCK_N1 % BLOCK_M1 == 0)
     step_m = BLOCK_M1
-    kv_blk = tl.program_id(0)  # Q-tile index
-    off_hz = tl.program_id(2)  # fused (batch, head)
-    b = off_hz // H
-    h = off_hz % H
+    kv_blk   = tl.program_id(0)          # Q-tile index
+    off_hz  = tl.program_id(2)          # fused (batch, head)
+    b       = off_hz // H
+    h       = off_hz %  H
     q_tiles = N_CTX // BLOCK_N1
     meta_base = ((b * H + h) * q_tiles + kv_blk)
 
-    q_blocks = tl.load(k2q_num + meta_base)  # int32
-    q_ptr = k2q_index + meta_base * max_q_blks  # ptr to list
+    q_blocks = tl.load(k2q_num  + meta_base)                 # int32
+    q_ptr    = k2q_index + meta_base * max_q_blks           # ptr to list
     block_size = tl.load(variable_block_sizes + kv_blk)
-
-    for blk_idx in range(q_blocks * 2):
-        block_sparse_offset = (tl.load(q_ptr + blk_idx // 2).to(tl.int32) * 2 +
-                               blk_idx % 2) * step_m
+    
+    
+        
+    for blk_idx in range(q_blocks*2):
+        block_sparse_offset = (tl.load(q_ptr + blk_idx//2).to(tl.int32)*2 + blk_idx%2) *step_m
         qT = tl.load(qT_ptrs + block_sparse_offset * stride_tok)
         # Load m before computing qk to reduce pipeline stall.
         offs_m = start_m + block_sparse_offset + tl.arange(0, BLOCK_M1)
@@ -256,32 +212,21 @@ def _attn_bwd_dkdv(
     return dk, dv
 
 
+
 # the main inner-loop logic for computing dQ
 @triton.jit
-def _attn_bwd_dq(
-        dq,
-        q,
-        K,
-        V,  #
-        do,
-        m,
-        D,
-        # shared by Q/K/V/DO.
-        q2k_index,
-        q2k_num,
-        max_kv_blks,
-        variable_block_sizes,
-        stride_tok,
-        stride_d,  #
-        H,
-        N_CTX,  #
-        BLOCK_M2: tl.constexpr,  #
-        BLOCK_N2: tl.constexpr,  #
-        HEAD_DIM: tl.constexpr,
-        # Filled in by the wrapper.
-        start_m,
-        start_n,
-        num_steps):
+def _attn_bwd_dq(dq, q, K, V,  #
+                 do, m, D,
+                 # shared by Q/K/V/DO.
+                 q2k_index, q2k_num, max_kv_blks,
+                 variable_block_sizes,
+                 stride_tok, stride_d,  #
+                 H, N_CTX,  #
+                 BLOCK_M2: tl.constexpr,  #
+                 BLOCK_N2: tl.constexpr,  #
+                 HEAD_DIM: tl.constexpr,
+                 # Filled in by the wrapper.
+                 start_m, start_n, num_steps):
     offs_m = start_m + tl.arange(0, BLOCK_M2)
     offs_n = start_n + tl.arange(0, BLOCK_N2)
     offs_k = tl.arange(0, HEAD_DIM)
@@ -292,27 +237,28 @@ def _attn_bwd_dq(
     # BLOCK_M2 must be a multiple of BLOCK_N2, otherwise the code wouldn't work.
     tl.static_assert(BLOCK_M2 % BLOCK_N2 == 0)
     step_n = BLOCK_N2
-
-    q_blk = tl.program_id(0)  # Q-tile index
-    off_hz = tl.program_id(2)  # fused (batch, head)
-    b = off_hz // H
-    h = off_hz % H
+    
+    q_blk   = tl.program_id(0)          # Q-tile index
+    off_hz  = tl.program_id(2)          # fused (batch, head)
+    b       = off_hz // H
+    h       = off_hz %  H
     q_tiles = N_CTX // BLOCK_M2
     meta_base = ((b * H + h) * q_tiles + q_blk)
 
-    kv_blocks = tl.load(q2k_num + meta_base)  # int32
-    kv_ptr = q2k_index + meta_base * max_kv_blks  # ptr to list
-    block_size = tl.load(variable_block_sizes + q_blk)
-
-    for blk_idx in range(kv_blocks * 2):
-        block_sparse_offset = (tl.load(kv_ptr + blk_idx // 2).to(tl.int32) * 2 +
-                               blk_idx % 2) * step_n * stride_tok
+    kv_blocks = tl.load(q2k_num  + meta_base)                 # int32
+    kv_ptr    = q2k_index + meta_base * max_kv_blks           # ptr to list
+    
+    
+    for blk_idx in range(kv_blocks*2):
+        kv_idx = tl.load(kv_ptr + blk_idx//2).to(tl.int32)
+        block_size = tl.load(variable_block_sizes + kv_idx) -  (blk_idx % 2) * step_n
+        block_sparse_offset = (kv_idx*2 + blk_idx%2) * step_n * stride_tok
         kT = tl.load(kT_ptrs + block_sparse_offset)
         vT = tl.load(vT_ptrs + block_sparse_offset)
         qk = tl.dot(q, kT)
         p = tl.math.exp2(qk - m)
         mask = tl.arange(0, BLOCK_N2) < block_size.to(tl.int32)
-        p = tl.where(mask[None, :], p, 0.0)
+        p = tl.where(mask[None, :], p , 0.0)
         # Compute dP and dS.
         dp = tl.dot(do, vT).to(tl.float32)
         ds = p * (dp - Di[:, None])
@@ -324,37 +270,23 @@ def _attn_bwd_dq(
     return dq
 
 
+
 @triton.jit
-def _attn_bwd(
-        Q,
-        K,
-        V,
-        sm_scale,  #
-        DO,  #
-        DQ,
-        DK,
-        DV,  #
-        M,
-        D,
-        q2k_index,
-        q2k_num,
-        max_kv_blks,
-        k2q_index,
-        k2q_num,
-        max_q_blks,
-        variable_block_sizes,
-        # shared by Q/K/V/DO.
-        stride_z,
-        stride_h,
-        stride_tok,
-        stride_d,  #
-        H,
-        N_CTX,  #
-        BLOCK_M1: tl.constexpr,  #
-        BLOCK_N1: tl.constexpr,  #
-        BLOCK_M2: tl.constexpr,  #
-        BLOCK_N2: tl.constexpr,  #
-        HEAD_DIM: tl.constexpr):
+def _attn_bwd(Q, K, V, sm_scale,  #
+              DO,  #
+              DQ, DK, DV,  #
+              M, D,
+              q2k_index, q2k_num, max_kv_blks,
+              k2q_index, k2q_num, max_q_blks,
+              variable_block_sizes,
+              # shared by Q/K/V/DO.
+              stride_z, stride_h, stride_tok, stride_d,  #
+              H, N_CTX,  #
+              BLOCK_M1: tl.constexpr,  #
+              BLOCK_N1: tl.constexpr,  #
+              BLOCK_M2: tl.constexpr,  #
+              BLOCK_N2: tl.constexpr,  #
+              HEAD_DIM: tl.constexpr):
     LN2 = 0.6931471824645996  # = ln(2)
 
     bhid = tl.program_id(2)
@@ -388,32 +320,20 @@ def _attn_bwd(
     k = tl.load(K + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d)
     v = tl.load(V + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d)
 
+
     num_steps = N_CTX // BLOCK_M1
 
     dk, dv = _attn_bwd_dkdv(  #
-        dk,
-        dv,  #
-        Q,
-        k,
-        v,
-        sm_scale,  #
+        dk, dv,  #
+        Q, k, v, sm_scale,  #
         DO,  #
-        M,
-        D,  #
-        k2q_index,
-        k2q_num,
-        max_q_blks,
+        M, D,  #
+        k2q_index, k2q_num, max_q_blks,
         variable_block_sizes,
-        stride_tok,
-        stride_d,  #
-        H,
-        N_CTX,  #
-        BLOCK_M1,
-        BLOCK_N1,
-        HEAD_DIM,  #
-        start_n,
-        start_m,
-        num_steps  #
+        stride_tok, stride_d,  #
+        H, N_CTX,  #
+        BLOCK_M1, BLOCK_N1, HEAD_DIM,  #
+        start_n, start_m, num_steps  #
     )
 
     dv_ptrs = DV + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d
@@ -438,88 +358,54 @@ def _attn_bwd(
     m = m[:, None]
 
     num_steps = N_CTX // BLOCK_N2
-    dq = _attn_bwd_dq(
-        dq,
-        q,
-        K,
-        V,  #
-        do,
-        m,
-        D,  #
-        q2k_index,
-        q2k_num,
-        max_kv_blks,
-        variable_block_sizes,
-        stride_tok,
-        stride_d,  #
-        H,
-        N_CTX,  #
-        BLOCK_M2,
-        BLOCK_N2,
-        HEAD_DIM,  #
-        start_m,
-        end_n,
-        num_steps  #
-    )
+    dq = _attn_bwd_dq(dq, q, K, V,  #
+                      do, m, D,  #
+                      q2k_index, q2k_num, max_kv_blks,
+                      variable_block_sizes,
+                      stride_tok, stride_d,  #
+                      H, N_CTX,  #
+                      BLOCK_M2, BLOCK_N2, HEAD_DIM,  #
+                      start_m, end_n, num_steps  #
+                      )
     # Write back dQ.
     dq_ptrs = DQ + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d
     dq *= LN2
     tl.store(dq_ptrs, dq)
+    
+    
+
 
 
 # ──────────────────────────── SPARSE ADDITION BEGIN ───────────────────────────
-def triton_block_sparse_attn_forward(q, k, v, q2k_index, q2k_num,
-                                     variable_block_sizes):
+def triton_block_sparse_attn_forward(q, k, v, q2k_index, q2k_num, variable_block_sizes):
     B, H, T, D = q.shape
     sm_scale = 1.0 / math.sqrt(D)
     max_kv_blks = q2k_index.shape[-1]
     assert T % 64 == 0, f"T must be a multiple of 64, but got {T}"
-    assert q2k_num.shape[
-        -1] == T // 64, f"shape mismatch, T // 64 = {T // 64}, q2k_num.shape[-2] = {q2k_num.shape[-2]}"
+    assert T // 64 == q2k_num.shape[-1], f"shape mismatch, T // 64 = {T // 64}, q2k_num.shape[-2] = {q2k_num.shape[-2]}"
     o = torch.empty_like(q)
     M = torch.empty((B, H, T), dtype=torch.float32, device=q.device)
 
     grid = lambda _: (triton.cdiv(T, 64), B * H, 1)
-    _attn_fwd_sparse[grid](q,
-                           k,
-                           v,
-                           sm_scale,
-                           q2k_index,
-                           q2k_num,
-                           max_kv_blks,
-                           variable_block_sizes,
-                           M,
-                           o,
-                           q.stride(0),
-                           q.stride(1),
-                           q.stride(2),
-                           q.stride(3),
-                           k.stride(0),
-                           k.stride(1),
-                           k.stride(2),
-                           k.stride(3),
-                           v.stride(0),
-                           v.stride(1),
-                           v.stride(2),
-                           v.stride(3),
-                           o.stride(0),
-                           o.stride(1),
-                           o.stride(2),
-                           o.stride(3),
-                           B,
-                           H,
-                           T,
-                           HEAD_DIM=D,
-                           STAGE=3)
+    _attn_fwd_sparse[grid](
+        q, k, v, sm_scale,
+        q2k_index, q2k_num, max_kv_blks,
+        variable_block_sizes,
+        M, o,
+        q.stride(0), q.stride(1), q.stride(2), q.stride(3),
+        k.stride(0), k.stride(1), k.stride(2), k.stride(3),
+        v.stride(0), v.stride(1), v.stride(2), v.stride(3),
+        o.stride(0), o.stride(1), o.stride(2), o.stride(3),
+        B, H, T,
+        HEAD_DIM=D, STAGE=3
+    )
 
     return o, M
 
-
-def triton_block_sparse_attn_backward(do, q, k, v, o, M, q2k_index, q2k_num,
-                                      k2q_index, k2q_num, variable_block_sizes):
+def triton_block_sparse_attn_backward(do, q, k, v, o, M, q2k_index, q2k_num, k2q_index, k2q_num, variable_block_sizes):
     assert do.is_contiguous()
     assert q.stride() == k.stride() == v.stride() == o.stride() == do.stride()
-
+    
     B, H, T, D = q.shape
     sm_scale = 1.0 / math.sqrt(D)
     dq = torch.empty_like(q)
@@ -535,49 +421,30 @@ def triton_block_sparse_attn_backward(do, q, k, v, o, M, q2k_index, q2k_num,
     pre_grid = (N_CTX // PRE_BLOCK, BATCH * N_HEAD)
     delta = torch.empty_like(M)
     _attn_bwd_preprocess[pre_grid](
-        o,
-        do,  #
+        o, do,  #
         delta,  #
-        BATCH,
-        N_HEAD,
-        N_CTX,  #
-        BLOCK_M=PRE_BLOCK,
-        HEAD_DIM=D  #
+        BATCH, N_HEAD, N_CTX,  #
+        BLOCK_M=PRE_BLOCK, HEAD_DIM=D  #
     )
-
+    
+    
     max_q_blks = k2q_index.shape[-1]
     max_kv_blks = q2k_index.shape[-1]
-
+    
     grid = (N_CTX // BLOCK_N1, 1, BATCH * N_HEAD)
     _attn_bwd[grid](
-        q,
-        arg_k,
-        v,
-        sm_scale,
-        do,
-        dq,
-        dk,
-        dv,  #
-        M,
-        delta,  #
-        q2k_index,
-        q2k_num,
-        max_kv_blks,
-        k2q_index,
-        k2q_num,
-        max_q_blks,
+        q, arg_k, v, sm_scale, do, dq, dk, dv,  #
+        M, delta,  #
+        q2k_index, q2k_num, max_kv_blks,
+        k2q_index, k2q_num, max_q_blks,
         variable_block_sizes,
-        q.stride(0),
-        q.stride(1),
-        q.stride(2),
-        q.stride(3),  #
-        N_HEAD,
-        N_CTX,  #
-        BLOCK_M1=BLOCK_M1,
-        BLOCK_N1=BLOCK_N1,  #
-        BLOCK_M2=BLOCK_M2,
-        BLOCK_N2=BLOCK_N2,  #
+        q.stride(0), q.stride(1), q.stride(2), q.stride(3),  #
+        N_HEAD, N_CTX,  #
+        BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1,  #
+        BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2,  #
         HEAD_DIM=D  #
     )
 
     return dq, dk, dv
+
+

csrc/attn/video_sparse_attn/vsa/block_sparse_h100.cu

@@ -639,7 +639,8 @@ void bwd_attend_ker(const __grid_constant__ bwd_globals<D> g) {
 
     // store kq and vq
 
-    // ensuring all writes are finished
+    // ! the following two line seems unnecessary.
+    // tma::store_async_wait(); // ensure qg is finished
     __syncthreads();
 
     warpgroup::store(kg_smem[0], kg_reg);
@@ -660,145 +661,6 @@ void bwd_attend_ker(const __grid_constant__ bwd_globals<D> g) {
     tma::store_async_wait(); 
 }
 
-
-template<int D>
-void block_sparse_attention_forward_impl(
-    bf16* d_q, bf16* d_k, bf16* d_v, float* d_l, bf16* d_o,
-    int batch, int qo_heads, int kv_heads, int seq_len, int hr,
-    int max_kv_blocks_per_q, 
-    int32_t* q2k_block_sparse_index_ptr,
-    int32_t* q2k_block_sparse_num_ptr,
-    int32_t* block_size_ptr,
-    cudaStream_t stream
-) {
-    using K = fwd_attend_ker_tile_dims<D>;
-    using q_tile    =         st_bf<K::qo_height, K::tile_width>;
-    using k_tile    =         st_bf<K::kv_height, K::tile_width>;
-    using v_tile    =         st_bf<K::kv_height, K::tile_width>;
-    using l_col_vec = col_vec<st_fl<K::qo_height, K::tile_width>>;
-    using o_tile    =         st_bf<K::qo_height, K::tile_width>;
-
-    using q_global = gl<bf16,  -1, -1, -1, -1, q_tile>;
-    using k_global = gl<bf16,  -1, -1, -1, -1, k_tile>;
-    using v_global = gl<bf16,  -1, -1, -1, -1, v_tile>;
-    using l_global = gl<float, -1, -1, -1, -1, l_col_vec>;
-    using o_global = gl<bf16,  -1, -1, -1, -1, o_tile>;
-
-    using globals      = fwd_globals<D>;
-
-    q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    l_global lg_arg{d_l, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,   static_cast<unsigned int>(seq_len)};
-    o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-
-    globals g{
-        qg_arg, kg_arg, vg_arg, lg_arg, og_arg, 
-        static_cast<int>(seq_len), static_cast<int>(hr), static_cast<int>(max_kv_blocks_per_q), 
-        q2k_block_sparse_index_ptr, q2k_block_sparse_num_ptr, block_size_ptr
-    };
-
-    // Shared memory size for the kernel
-    // 54000 bytes is calibrated for H100 shared memory constraints for these tile sizes
-    constexpr int mem_size = 54000;
-
-    dim3 grid(seq_len/(64), qo_heads, batch);
-
-    cudaFuncSetAttribute(
-        fwd_attend_ker<D>,
-        cudaFuncAttributeMaxDynamicSharedMemorySize,
-        mem_size
-    );
-
-    fwd_attend_ker<D><<<grid, (128), mem_size, stream>>>(g);
-}
-
-template<int D>
-void block_sparse_attention_backward_impl(
-    bf16* d_q, bf16* d_k, bf16* d_v, bf16* d_o, bf16* d_og, float* d_l, float* d_d, float* d_qg, float* d_kg, float* d_vg,
-    int batch, int qo_heads, int kv_heads, int seq_len, int hr, int max_q_blocks_per_kv,
-    int32_t* k2q_block_sparse_index_ptr,
-    int32_t* k2q_block_sparse_num_ptr,
-    int32_t* block_size_ptr,
-    cudaStream_t stream
-) {
-    using G = bwd_attend_ker_tile_dims<D>;
-    using og_tile = st_bf<4*16, D>;
-    using o_tile  = st_bf<4*16, D>;
-    using d_tile  = col_vec<st_fl<4*16, D>>;
-
-    using og_global = gl<bf16,  -1, -1, -1, -1, og_tile>;
-    using o_global  = gl<bf16,  -1, -1, -1, -1, o_tile>;
-    using d_global  = gl<float, -1, -1, -1, -1, d_tile>;
-
-    using prep_globals = bwd_prep_globals<D>;
-
-    constexpr int mem_size_prep = kittens::MAX_SHARED_MEMORY;
-    int threads_prep = PREP_NUM_WARPS * kittens::WARP_THREADS;
-    dim3 grid_bwd_prep(seq_len/(PREP_NUM_WARPS*kittens::TILE_ROW_DIM<bf16>*4), qo_heads, batch);
-
-    og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    d_global  prep_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
-
-    prep_globals bwd_g{prep_og_arg, prep_o_arg, prep_d_arg};
-
-    cudaFuncSetAttribute(
-        bwd_attend_prep_ker<D>,
-        cudaFuncAttributeMaxDynamicSharedMemorySize,
-        mem_size_prep
-    );
-    bwd_attend_prep_ker<D><<<grid_bwd_prep, threads_prep, mem_size_prep, stream>>>(bwd_g);
-
-    using bwd_q_tile    =         st_bf<G::tile_h_qo, G::tile_width>;
-    using bwd_k_tile    =         st_bf<G::tile_h,    G::tile_width>;
-    using bwd_v_tile    =         st_bf<G::tile_h,    G::tile_width>;
-    using bwd_og_tile   =         st_bf<G::tile_h_qo, G::tile_width>;
-    using bwd_qg_tile   =         st_fl<G::tile_h_qo, G::tile_width>;
-    using bwd_kg_tile   =         st_fl<G::tile_h,    G::tile_width>;
-    using bwd_vg_tile   =         st_fl<G::tile_h,    G::tile_width>;
-    using bwd_l_tile    = row_vec<st_fl<G::tile_h_qo, G::tile_h>>;
-    using bwd_d_tile    = row_vec<st_fl<G::tile_h_qo, G::tile_h>>;
-
-    using bwd_q_global  = gl<bf16,  -1, -1, -1, -1, bwd_q_tile>;
-    using bwd_k_global  = gl<bf16,  -1, -1, -1, -1, bwd_k_tile>;
-    using bwd_v_global  = gl<bf16,  -1, -1, -1, -1, bwd_v_tile>;
-    using bwd_og_global = gl<bf16,  -1, -1, -1, -1, bwd_og_tile>;
-    using bwd_qg_global = gl<float, -1, -1, -1, -1, bwd_qg_tile>;
-    using bwd_kg_global = gl<float, -1, -1, -1, -1, bwd_kg_tile>;
-    using bwd_vg_global = gl<float, -1, -1, -1, -1, bwd_vg_tile>;
-    using bwd_l_global  = gl<float, -1, -1, -1, -1, bwd_l_tile>;
-    using bwd_d_global  = gl<float, -1, -1, -1, -1, bwd_d_tile>;
-
-    using bwd_global_args = bwd_globals<D>;
-
-    bwd_q_global  bwd_q_arg {d_q,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_k_global  bwd_k_arg {d_k,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_v_global  bwd_v_arg {d_v,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_og_global bwd_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_qg_global bwd_qg_arg{d_qg, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_kg_global bwd_kg_arg{d_kg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_vg_global bwd_vg_arg{d_vg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_l_global  bwd_l_arg {d_l,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
-    bwd_d_global  bwd_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
-
-    bwd_global_args bwd_global{bwd_q_arg, bwd_k_arg, bwd_v_arg, bwd_og_arg, bwd_qg_arg, bwd_kg_arg, bwd_vg_arg, bwd_l_arg, bwd_d_arg, 
-                        static_cast<int>(seq_len), static_cast<int>(hr), static_cast<int>(max_q_blocks_per_kv),
-                        k2q_block_sparse_index_ptr, k2q_block_sparse_num_ptr, block_size_ptr};
-
-    dim3 grid_bwd_main(seq_len/64, qo_heads, batch);
-    int threads_main = 128;
-    // Calibrated shared memory sizes for different head dimensions
-    int bwd_mem_size = (D == 64) ? 72000 : 113000;
-
-    cudaFuncSetAttribute(
-        bwd_attend_ker<D>,
-        cudaFuncAttributeMaxDynamicSharedMemorySize,
-        bwd_mem_size
-    );
-    bwd_attend_ker<D><<<grid_bwd_main, threads_main, bwd_mem_size, stream>>>(bwd_global);
-}
-
 #include "pyutils/torch_helpers.cuh"
 #include <ATen/cuda/CUDAContext.h>
 #include <iostream>
@@ -840,6 +702,7 @@ block_sparse_attention_forward(
     TORCH_CHECK(q2k_block_sparse_index.size(2) == seq_len / BLOCK_M, "q2k_block_sparse_index idx 2 - must match seq_len / BLOCK_M");
     TORCH_CHECK(q2k_block_sparse_num.size(2) == seq_len / BLOCK_M, "q2k_block_sparse_num idx 2 - must match seq_len / BLOCK_M");
 
+
     TORCH_CHECK(q.size(3) == head_dim, "Q head dimension - idx 3 - must match for all non-vector inputs");
     TORCH_CHECK(k.size(3) == head_dim, "K head dimension - idx 3 - must match for all non-vector inputs");
     TORCH_CHECK(v.size(3) == head_dim, "V head dimension - idx 3 - must match for all non-vector inputs");
@@ -880,32 +743,110 @@ block_sparse_attention_forward(
     float* l_ptr = reinterpret_cast<float*>(l_vec.data_ptr<float>());
     float* d_l   = reinterpret_cast<float*>(l_ptr);
 
+    //cudadevicesynchronize();
     const c10::cuda::OptionalCUDAGuard device_guard(q.device());    
     const cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream(); 
 
-    // Temporated implementation to avoid code duplication between head_dim=64 and 128
     if (head_dim == 64) {
-        block_sparse_attention_forward_impl<64>(
-            d_q, d_k, d_v, d_l, d_o,
-            batch, qo_heads, kv_heads, seq_len, hr,
-            max_kv_blocks_per_q,
-            reinterpret_cast<int32_t*>(q2k_block_sparse_index.data_ptr()),
+        using q_tile    =         st_bf<fwd_attend_ker_tile_dims<64>::qo_height, fwd_attend_ker_tile_dims<64>::tile_width>;
+        using k_tile    =         st_bf<fwd_attend_ker_tile_dims<64>::kv_height, fwd_attend_ker_tile_dims<64>::tile_width>;
+        using v_tile    =         st_bf<fwd_attend_ker_tile_dims<64>::kv_height, fwd_attend_ker_tile_dims<64>::tile_width>;
+        using l_col_vec = col_vec<st_fl<fwd_attend_ker_tile_dims<64>::qo_height, fwd_attend_ker_tile_dims<64>::tile_width>>;
+        using o_tile    =         st_bf<fwd_attend_ker_tile_dims<64>::qo_height, fwd_attend_ker_tile_dims<64>::tile_width>;
+
+        using q_global = gl<bf16,  -1, -1, -1, -1, q_tile>;
+        using k_global = gl<bf16,  -1, -1, -1, -1, k_tile>;
+        using v_global = gl<bf16,  -1, -1, -1, -1, v_tile>;
+        using l_global = gl<float, -1, -1, -1, -1, l_col_vec>;
+        using o_global = gl<bf16,  -1, -1, -1, -1, o_tile>;
+
+        using globals      = fwd_globals<64>;
+
+        q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
+        k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 64U};
+        v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 64U};
+        l_global lg_arg{d_l, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,   static_cast<unsigned int>(seq_len)};
+        o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
+
+        globals g{
+            qg_arg, 
+            kg_arg, 
+            vg_arg, 
+            lg_arg, 
+            og_arg, 
+            static_cast<int>(seq_len), 
+            static_cast<int>(hr), 
+            static_cast<int>(max_kv_blocks_per_q), 
+            reinterpret_cast<int32_t*>(q2k_block_sparse_index.data_ptr()), 
             reinterpret_cast<int32_t*>(q2k_block_sparse_num.data_ptr()),
-            reinterpret_cast<int32_t*>(block_size.data_ptr()),
-            stream
+            reinterpret_cast<int32_t*>(block_size.data_ptr())
+        };
+
+        constexpr int mem_size = 54000;
+
+        dim3 grid(seq_len/(64), qo_heads, batch);
+
+        cudaFuncSetAttribute(
+            fwd_attend_ker<64>,
+            cudaFuncAttributeMaxDynamicSharedMemorySize,
+            mem_size
         );
-    } else if (head_dim == 128) {
-        block_sparse_attention_forward_impl<128>(
-            d_q, d_k, d_v, d_l, d_o,
-            batch, qo_heads, kv_heads, seq_len, hr,
-            max_kv_blocks_per_q,
-            reinterpret_cast<int32_t*>(q2k_block_sparse_index.data_ptr()),
+
+        fwd_attend_ker<64><<<grid, (128), mem_size, stream>>>(g);
+
+        CHECK_CUDA_ERROR(cudaGetLastError());
+        // cudaStreamSynchronize(stream);
+    }
+
+    if (head_dim == 128) {
+        using q_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using k_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::kv_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using v_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::kv_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using l_col_vec = col_vec<st_fl<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>>;
+        using o_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+
+        using q_global = gl<bf16,  -1, -1, -1, -1, q_tile>;
+        using k_global = gl<bf16,  -1, -1, -1, -1, k_tile>;
+        using v_global = gl<bf16,  -1, -1, -1, -1, v_tile>;
+        using l_global = gl<float, -1, -1, -1, -1, l_col_vec>;
+        using o_global = gl<bf16,  -1, -1, -1, -1, o_tile>;
+
+        using globals      = fwd_globals<128>;
+
+        q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
+        k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 128U};
+        v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 128U};
+        l_global lg_arg{d_l, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,   static_cast<unsigned int>(seq_len)};
+        o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
+
+        globals g{
+            qg_arg, 
+            kg_arg, 
+            vg_arg, 
+            lg_arg, 
+            og_arg, 
+            static_cast<int>(seq_len), 
+            static_cast<int>(hr), 
+            static_cast<int>(max_kv_blocks_per_q), 
+            reinterpret_cast<int32_t*>(q2k_block_sparse_index.data_ptr()), 
             reinterpret_cast<int32_t*>(q2k_block_sparse_num.data_ptr()),
-            reinterpret_cast<int32_t*>(block_size.data_ptr()),
-            stream
+            reinterpret_cast<int32_t*>(block_size.data_ptr())
+        };
+
+        constexpr int mem_size = 54000;
+
+        dim3 grid(seq_len/(64), qo_heads, batch);
+
+        cudaFuncSetAttribute(
+            fwd_attend_ker<128>,
+            cudaFuncAttributeMaxDynamicSharedMemorySize,
+            mem_size
         );
-    } else {
-        TORCH_CHECK(false, "Unsupported head_dim: ", head_dim, ". Only 64 and 128 are supported.");
+
+        fwd_attend_ker<128><<<grid, (128), mem_size, stream>>>(g);
+
+        CHECK_CUDA_ERROR(cudaGetLastError());
+        // cudaStreamSynchronize(stream);
     }
 
     return {o, l_vec};

csrc/attn/video_sparse_attn/vsa/block_sparse_wrapper.py

@@ -0,0 +1,185 @@
+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/video_sparse_attn/vsa/index.py

@@ -1,9 +1,9 @@
+
 ## pytorch sdpa version of block sparse ##
 import triton
 import triton.language as tl
 import torch
 
-
 @triton.jit
 def topk_index_to_map_kernel(
     map_ptr,
@@ -26,7 +26,6 @@ def topk_index_to_map_kernel(
         index = tl.load(index_ptr_base + i * index_kv_stride)
         tl.store(map_ptr_base + index * map_kv_stride, 1.0)
 
-
 @triton.jit
 def map_to_index_kernel(
     map_ptr,
@@ -60,7 +59,6 @@ def map_to_index_kernel(
         index_num_ptr + b * index_num_bs_stride + h * index_num_h_stride +
         q * index_num_q_stride, num)
 
-
 def topk_index_to_map(index: torch.Tensor,
                       num_kv_blocks: int,
                       transpose_map: bool = False):
@@ -108,7 +106,6 @@ def topk_index_to_map(index: torch.Tensor,
 
     return block_map
 
-
 def map_to_index(block_map: torch.Tensor):
     """
     Convert a block map to indices and counts.

csrc/attn/vmoba_attn/README.md

@@ -0,0 +1,32 @@
+# 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

@@ -0,0 +1,26 @@
+# 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/tests/test_vmoba_attn.py

@@ -3,7 +3,7 @@
 import torch
 import pytest
 import random
-from fastvideo_kernel import moba_attn_varlen
+from csrc.attn.vmoba_attn.vmoba  import moba_attn_varlen
 
 def generate_test_data(batch_size, total_seqlen, num_heads, head_dim, dtype, device="cuda"):
     """
@@ -51,7 +51,7 @@ def generate_test_data(batch_size, total_seqlen, num_heads, head_dim, dtype, dev
 @pytest.mark.parametrize("moba_topk", [2, 4])
 @pytest.mark.parametrize("select_mode", ["topk", "threshold"])
 @pytest.mark.parametrize("threshold_type", ["query_head", "head_global", "overall"])
-@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
+@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
 def test_moba_attn_varlen_forward(
     batch_size, total_seqlen, num_heads, head_dim, moba_chunk_size, moba_topk, select_mode, threshold_type, dtype
 ):

csrc/attn/vmoba_attn/vmoba/__init__.py

@@ -0,0 +1,2 @@
+# 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-cudnn-devel-ubuntu22.04
+FROM nvidia/cuda:12.8.0-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 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
+    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
 
 COPY . .
 
@@ -55,12 +55,18 @@ 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 FastVideo Unified Kernel
+# Install STA (Sliding Tile Attention)
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd fastvideo-kernel && \
+    cd csrc/attn/sliding_tile_attn && \
     git submodule update --init --recursive && \
-    ./build.sh
+    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
 
 EXPOSE 22

docker/Dockerfile.python3.11

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04
+FROM nvidia/cuda:12.8.0-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 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
+    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
 
 COPY . .
 
@@ -55,12 +55,18 @@ 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 FastVideo Unified Kernel
+# Install STA (Sliding Tile Attention)
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd fastvideo-kernel && \
+    cd csrc/attn/sliding_tile_attn && \
     git submodule update --init --recursive && \
-    ./build.sh
+    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
 
 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 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
+    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
 
 COPY . .
 
@@ -55,11 +55,18 @@ 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 FastVideo Unified Kernel
+# Install STA (Sliding Tile Attention)
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd fastvideo-kernel && \
+    cd csrc/attn/sliding_tile_attn && \
     git submodule update --init --recursive && \
-    ./build.sh
+    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
 
 EXPOSE 22

docker/Dockerfile.python3.12.cuda12.9.1

@@ -55,12 +55,18 @@ 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 FastVideo Unified Kernel
+# Install STA (Sliding Tile Attention)
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd fastvideo-kernel && \
+    cd csrc/attn/sliding_tile_attn && \
     git submodule update --init --recursive && \
-    ./build.sh
+    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
 
 EXPOSE 22

docker/rocm7_1.Dockerfile

@@ -1,58 +0,0 @@
-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 requirements-mkdocs.txt
+pip install -r docs/requirements-mkdocs.txt
 
 # Serve docs with live reload (recommended for development)
 mkdocs serve

docs/assets/custom.css

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

docs/attention/developer/index.md

@@ -1,186 +0,0 @@
-# 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

@@ -1,53 +0,0 @@
-# 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.
-
-## 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

@@ -1,36 +0,0 @@
-# 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

@@ -1,36 +0,0 @@
-# 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
-# topk: Number of blocks to attend
-
-output = video_sparse_attn(
-    q, k, v, 
-    variable_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/overview.md

@@ -6,7 +6,7 @@ Thank you for your interest in contributing to FastVideo. We want to make the pr
 Our community is open to everyone and welcomes any contributions no matter how large or small.
 
 # Developer Environment:
-Do make sure you have CUDA 12.4 installed and supported. FastVideo currently only supports Linux and CUDA GPUs, but we hope to support other platforms in the future.
+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.
 
 We recommend using a fresh Python 3.10 Conda environment to develop FastVideo:
 

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 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 the 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 disk storage.
+- After profiling, clean up trace directories to avoid filling disks.
 - 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

@@ -74,11 +74,9 @@ To add a new SSIM test, follow these steps:
        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
-   ```
+       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`.

docs/design/overview.md

@@ -4,26 +4,25 @@ This document outlines FastVideo's architecture for developers interested in fra
 
 ## Table of Contents - Directory Structure and Files
 
-- [`fastvideo/pipelines/`](#pipeline-system) - Core diffusion pipeline components
-- [`fastvideo/models/`](#model-components) - Model implementations
-  - [`dits/`](#transformer-models) - Transformer-based diffusion models
-  - [`vaes/`](#vae-variational-auto-encoder) - Variational autoencoders
-  - [`encoders/`](#text-and-image-encoders) - Text and image encoders
-  - [`schedulers/`](#schedulers) - Diffusion schedulers
-- [`fastvideo/attention/`](#optimized-attention) - Optimized attention implementations
-- [`fastvideo/distributed/`](#distributed-processing) - Distributed computing utilities
-- [`fastvideo/layers/`](#tensor-parallelism) - Custom neural network layers
-- [`fastvideo/platforms/`](#platforms) - Hardware platform abstractions
-- [`fastvideo/worker/`](#executor-and-worker-system) - Multi-GPU process management
-- [`fastvideo/fastvideo_args.py`](#fastvideoargs) - Argument handling
-- [`fastvideo/forward_context.py`](#forward-context-management) - Forward pass context management
+- [`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`](#logger) - Logging infrastructure
+- [`fastvideo/logger.py`](#design-logger) - Logging infrastructure
 
 ## Core Architecture
 
 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)
@@ -35,14 +34,12 @@ FastVideo separates model components from execution logic with these principles:
 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`
@@ -93,9 +90,7 @@ class MyCustomPipeline(ComposedPipelineBase):
 ```
 
 ### 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
@@ -138,7 +133,6 @@ Transformer networks perform the actual denoising during diffusion:
   - `HunyuanVideoTransformer3DModel`
 
 Features include:
-
 - Text/image conditioning
 - Standardized interface for model-specific optimizations
 
@@ -167,7 +161,6 @@ VAEs handle conversion between pixel space and latent space:
 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
@@ -186,7 +179,6 @@ Encoders process conditioning inputs into embeddings:
   - `CLIPVisionModel`
 
 FastVideo implements optimizations such as:
-
 - Vocab parallelism for distributed processing
 - Caching for common prompts
 - Precision-tuned computation
@@ -201,7 +193,6 @@ Schedulers manage the diffusion sampling process:
   - `FlowMatchEulerDiscreteScheduler`
 
 These components control:
-
 - Diffusion timestep sequences
 - Noise prediction to latent update conversions
 - Quality/speed trade-offs
@@ -228,9 +219,7 @@ This diagram shows how models are discovered, validated, and loaded across entry
 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
@@ -251,9 +240,7 @@ self.attn = LocalAttention(
 ![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
 
@@ -309,7 +296,6 @@ self.attn = DistributedAttention(
 ```
 
 ### Communication Primitives
-
 Efficient distributed operations via AllGather, AllReduce, and synchronization mechanisms.
 
 Efficient communication primitives minimize distributed overhead:
@@ -328,7 +314,6 @@ Defined in `fastvideo/forward_context.py`, `ForwardContext` manages execution-sp
 - **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
 
@@ -354,14 +339,12 @@ FastVideo implements a flexible execution model for distributed processing:
 - **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
@@ -376,13 +359,11 @@ The `fastvideo/platforms/` directory provides hardware platform abstractions tha
 ### 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
@@ -402,7 +383,6 @@ else:
 The platform system is designed to be extensible for future hardware targets.
 
 ## Logger
-
 See [PR](https://github.com/hao-ai-lab/FastVideo/pull/356)
 
 *TODO*: (help wanted) Add an environment variable that disables process-aware logging.
@@ -417,7 +397,6 @@ If you're a new contributor, here are some common areas to explore:
 4. **Hardware support**: Extend the `platforms` module for new hardware targets
 
 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

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 computation, 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 computations, 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](../attention/vsa/index.md). Set `MODEL_BASE` to your own model path and run:
+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:
 
 ```bash
 bash scripts/inference/v1_inference_wan_dmd.sh

docs/getting_started/installation.md

@@ -11,13 +11,15 @@ 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
@@ -26,12 +28,6 @@ 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+
@@ -42,4 +38,4 @@ pip install flash-attn --no-build-isolation
 
 - [Quick Start Guide](quick_start.md) - Get started with your first video generation
 - [Configuration](../inference/configuration.md) - Learn about configuration options
-- [Examples](../inference/examples/examples_inference_index.md) - Explore example scripts and notebooks
+- [Examples](../inference/examples/) - 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](../../contributing/developer_env/docker.md)
+[Docker Images](#docker)
 
 ## Development Environment Setup
 
 If you're planning to contribute to FastVideo please see the following page:
-[Contributor Guide](../../contributing/overview.md)
+[Contributor Guide](#developer-overview)
 
 ## 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](../../contributing/overview.md)
+[Contributor Guide](#developer-overview)
 
 ## Hardware Requirements
 

docs/getting_started/quick_start.md

@@ -7,72 +7,44 @@ 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 VideoGenerator
+from fastvideo import FastVideoPipeline
 
-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
-    )
+# Initialize the pipeline
+pipe = FastVideoPipeline.from_pretrained("wan2.1-t2v-1.3B")
 
-    # Define a prompt for your video
-    prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest."
+# Generate a video
+prompt = "A cat playing with a ball of yarn"
+video = pipe(prompt, num_frames=16, height=512, width=512)
 
-    # 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()
+# Save the video
+video.save("output.mp4")
 ```
 
 ### Image-to-Video Generation
 
 ```python
-from fastvideo import VideoGenerator, SamplingParam
+from fastvideo import FastVideoPipeline
+from PIL import Image
 
-def main():
-    # Create the generator
-    model_name = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
-    generator = VideoGenerator.from_pretrained(model_name, num_gpus=1)
+# Load an image
+image = Image.open("input.jpg")
 
-    # 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
+# Initialize the pipeline
+pipe = FastVideoPipeline.from_pretrained("wan2.1-i2v-14B-480p")
 
-    # 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)
+# Generate a video from the image
+video = pipe(image, num_frames=16, height=480, width=480)
 
-if __name__ == '__main__':
-    main()
+# Save the video
+video.save("output.mp4")
 ```
 
 ## Next Steps

docs/inference/add_pipeline.md

@@ -45,7 +45,6 @@ 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/`
@@ -54,15 +53,12 @@ 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
@@ -95,7 +91,6 @@ self.out_proj = RowParallelLinear(
 ```
 
 ### Attention Layers
-
 Replace standard attention with FastVideo's optimized attention:
 
 ```python
@@ -168,10 +163,6 @@ Pipelines are composed of stages, each handling a specific part of the diffusion
 - **DenoisingStage**: Performs denoising diffusion
 - **DecodingStage**: Converts latents to pixels
 
-Note: `DenoisingStage` uses the unified denoising engine under the hood. You
-can inject a custom strategy via `strategy_cls` if a pipeline needs specialized
-denoising behavior.
-
 ### Creating Your Pipeline
 
 ```python
@@ -234,7 +225,7 @@ class MyCustomPipeline(ComposedPipelineBase):
             stage_name="denoising_stage",
             stage=DenoisingStage(
                 transformer=self.get_module("transformer"),
-                scheduler=self.get_module("scheduler"),
+                scheduler=self.get_module("scheduler")
             )
         )
         
@@ -249,24 +240,6 @@ class MyCustomPipeline(ComposedPipelineBase):
 EntryClass = MyCustomPipeline
 ```
 
-### Customizing Denoising Strategies
-
-If your model requires a custom denoising loop, pass a strategy class:
-
-```python
-from fastvideo.pipelines.stages.denoising_cosmos_strategy import (
-    CosmosStrategy)
-
-        self.add_stage(
-            stage_name="denoising_stage",
-            stage=DenoisingStage(
-                transformer=self.get_module("transformer"),
-                scheduler=self.get_module("scheduler"),
-                strategy_cls=CosmosStrategy,
-            )
-        )
-```
-
 ### Creating Custom Stages (Optional)
 
 If existing stages don't meet your needs, create custom ones:
@@ -331,7 +304,6 @@ 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](examples/basic.md).
+see the Python interface [here](https://hao-ai-lab.github.io/FastVideo/inference/examples/basic.html).
 
 ## Basic Usage
 

docs/inference/configuration.md

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

docs/inference/inference_quick_start.md

@@ -3,7 +3,6 @@
 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
@@ -22,10 +21,9 @@ conda activate fastvideo
 pip install fastvideo
 ```
 
-For advanced installation options, see the [Installation Guide](../getting_started/installation.md).
+For advanced installation options, see the [Installation Guide](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
@@ -62,10 +60,9 @@ 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.md) for the list of supported models and their available optimizations.
+Please see the [support matrix](#support-matrix) for the list of supported models and their available optimizations.
 
 ## Image-to-Video Generation
 
@@ -99,28 +96,20 @@ 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
@@ -128,8 +117,8 @@ If the generated video doesn't match your prompt:
 
 ## Next Steps
 
-- Learn about [Advanced Inference Configurations](configuration.md)
-- Learn about using [Optimizations](optimizations.md)
-- See [Examples](examples/examples_inference_index.md) for more usage scenarios
+- Learn about [Advanced Inference Configurations](#inference-configuration)
+- Learn about using [Optimizations](#inference-optimizations)
+- 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/optimizations.md

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

docs/inference/support_matrix.md

@@ -40,26 +40,20 @@ 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 | 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 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| 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 | ❌ | ❌ | ✅ | ⭕ |
 
 **Note**: Wan2.2 TI2V 5B has some quality issues when performing I2V generation. We are working on fixing this issue.
 
@@ -70,13 +64,3 @@ 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/sliding_tile_attention/demo.md

@@ -0,0 +1,16 @@
+
+# 🔍 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

@@ -0,0 +1,65 @@
+
+# 🔧 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

@@ -1,130 +1,45 @@
-# 🧱 Data Preprocessing
 
-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.
+# 🧱 Data Preprocess
 
-## Quick Start
+To save GPU memory, we precompute text embeddings and VAE latents to eliminate the need to load the text encoder and VAE during training.
 
-Download the sample dataset and run preprocessing:
+We provide a sample dataset to help you get started. Download the source media using the following command:
 
 ```bash
-# 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
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/mini_i2v_dataset --local_dir=data/mini_i2v_dataset --repo_type=dataset
 ```
 
-## Preprocessing Pipeline
+The folder `crush-smol_raw/` contains raw videos and captions for testing preprocessing, while `crush-smol_preprocessed/` contains latents prepared for testing training.
 
-The new preprocessing pipeline supports multiple dataset formats and video loaders:
-
-```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
-```
-
-### Key Parameters
-
-| 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:
+To preprocess the dataset for fine-tuning or distillation, run:
 
 ```
-your_dataset/
-├── videos/
-│   ├── video_001.mp4
-│   ├── video_002.mp4
-│   └── ...
-└── videos2caption.json
+bash scripts/preprocess/v1_preprocess_wan_data_t2v # for wan
 ```
 
-The `videos2caption.json` maps video filenames to captions:
+## Process your own dataset
 
-```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:
+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:
 
 ```
-your_raw_data/
+path_to_your_dataset_folder/
 ├── videos/
 │   ├── 0.mp4
 │   ├── 1.mp4
-│   └── ...
-├── videos.txt    # list of video filenames
-└── prompt.txt    # corresponding captions (one per line)
+├── videos.txt
+└── prompt.txt
 ```
 
-## Output Format
+To geranate the `videos2caption.json` and `merge.txt`, run
 
-Preprocessing outputs Parquet files in the `combined_parquet_dataset/` subdirectory containing:
+``` python
+python scripts/dataset_preparation/prepare_json_file.py --data_folder mini_i2v_dataset/crush-smol_raw/ --output your_output_folder
+```
 
-- `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
+Adjust the `DATA_MERGE_PATH` and `OUTPUT_DIR` in `scripts/preprocess/v1_preprocess_****.sh` accordingly and run:
 
-## Examples
+```
+bash scripts/preprocess/v1_preprocess_****.sh
+```
 
-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)**
+The preprocessed data will be put into the `OUTPUT_DIR` and the `videos2caption.json` can be used in finetune and distill scripts.

docs/training/finetune.md

@@ -1,176 +1,78 @@
-# 🧠 Finetuning
 
-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.
+# 🧠 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:
 
 ```bash
-# Example: Wan2.1 T2V 1.3B full finetune (4 GPUs)
-bash examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/Mochi-Black-Myth --local_dir=data/Mochi-Black-Myth --repo_type=dataset
 ```
 
-**Typical settings:**
+Download the original model weights as specified in the [Distillation Section](../distillation/dmd.md):
 
-- 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
+Then you can run the finetune with:
 
-## LoRA Finetuning
+```
+bash scripts/finetune/finetune_mochi.sh # for mochi
+```
 
-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
+**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:
 
 ```bash
-# Example: Wan2.1 T2V 1.3B LoRA finetune (1 GPU)
-bash examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_lora.sh
+bash scripts/finetune/finetune_v1_VSA.sh
 ```
 
-Key differences from full finetune:
+## ⚡ Lora Finetune
 
-- 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:
+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:
 
 ```bash
-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
+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
 ```
 
-| 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) |
+### 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.
 
-### Merge LoRA Adapter
+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 an adapter back into a base model:
+### 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.
 
 ```bash
-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
+bash scripts/finetune/finetune_hunyuan.sh
+bash scripts/finetune/finetune_mochi_lora_mix.sh
 ```
 
-| 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
+For Image-Video Mixture Fine-tuning, make sure to enable the `--group_frame` option in your script.

docs/training/overview.md

@@ -1,67 +0,0 @@
-# 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

@@ -1,71 +0,0 @@
-# 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

@@ -0,0 +1,65 @@
+
+# 🔧 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/SFWan2.1-T2V/README.md

@@ -1,19 +0,0 @@
-# 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/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) first.
+If you have not installed FastVideo, please following these [instructions](https://hao-ai-lab.github.io/FastVideo/getting_started/installation.html) 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

@@ -12,7 +12,7 @@ def main():
         "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory
+        use_fsdp_inference=True,
         dit_cpu_offload=False,
         vae_cpu_offload=False,
         text_encoder_cpu_offload=True,

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=False, # set to True if GPU is out of memory
+        use_fsdp_inference=True,
         # 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

@@ -1,42 +0,0 @@
-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_longcat_i2v.py

@@ -1,210 +0,0 @@
-"""
-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()
-
-

examples/inference/basic/basic_longcat_t2v.py

@@ -1,198 +0,0 @@
-"""
-LongCat Text-to-Video (T2V) Example Script
-
-This script demonstrates LongCat T2V inference using the FastVideo Python API.
-It runs both basic generation (50 steps) and distill+refine generation
-(16 steps distill + 50 steps refinement to 720p).
-
-Usage:
-    python examples/inference/basic/basic_longcat_t2v.py
-"""
-
-import glob
-import os
-
-from fastvideo import VideoGenerator
-
-# Common prompts and settings matching the shell script examples
-PROMPT = (
-    "In a realistic photography style, a white boy around seven or eight years old "
-    "sits on a park bench, wearing a light blue T-shirt, denim shorts, and white sneakers. "
-    "He holds an ice cream cone with vanilla and chocolate flavors, and beside him is a "
-    "medium-sized golden Labrador. Smiling, the boy offers the ice cream to the dog, "
-    "who eagerly licks it with its tongue. The sun is shining brightly, and the background "
-    "features a green lawn and several tall trees, creating a warm and loving scene."
-)
-
-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"
-)
-
-SEED = 42
-
-
-def basic_generation():
-    """
-    Run basic LongCat T2V generation (50 steps at 480p).
-    
-    This uses the full 50-step denoising process for highest quality.
-    """
-    print("=" * 60)
-    print("LongCat T2V: Basic Generation (50 steps, 480p)")
-    print("=" * 60)
-    
-    generator = VideoGenerator.from_pretrained(
-        "FastVideo/LongCat-Video-T2V-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_t2v_basic"
-    
-    generator.generate_video(
-        prompt=PROMPT,
-        negative_prompt=NEGATIVE_PROMPT,
-        output_path=output_path,
-        save_video=True,
-        height=480,
-        width=832,
-        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 T2V with distill+refine pipeline (16 steps + refinement to 720p).
-    
-    This uses the distilled LoRA for fast 480p generation (16 steps),
-    then refines to 720p using the refinement LoRA with BSA enabled.
-    """
-    print("\n" + "=" * 60)
-    print("LongCat T2V: 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-T2V-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_t2v_distill"
-    
-    generator.generate_video(
-        prompt=PROMPT,
-        negative_prompt=NEGATIVE_PROMPT,
-        output_path=distill_output_path,
-        save_video=True,
-        height=480,
-        width=832,
-        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 -> 720p)
-    print("\n[Stage 2] Refinement (480p -> 720p 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
-    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, 8],
-        bsa_chunk_k=[4, 4, 8],
-        lora_path="FastVideo/LongCat-Video-T2V-Refinement-LoRA",
-        lora_nickname="refinement",
-    )
-    
-    refine_output_path = "outputs_video/longcat_t2v_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=1280,
-        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 Text-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()
-
-

examples/inference/basic/basic_longcat_vc.py

@@ -1,228 +0,0 @@
-"""
-LongCat Video Continuation (VC) Example Script
-
-This script demonstrates LongCat VC inference using the FastVideo Python API.
-LongCat VC takes an input video and generates a continuation of it.
-
-It runs both basic generation (50 steps) and distill+refine generation
-(16 steps distill + 50 steps refinement to 720p).
-
-Usage:
-    python examples/inference/basic/basic_longcat_vc.py
-
-Prerequisites:
-    - Ensure the input video exists at assets/motorcycle.mp4
-      (or provide your own video)
-"""
-
-import glob
-import os
-
-from fastvideo import VideoGenerator
-
-# Common prompts and settings matching the shell script examples
-PROMPT = (
-    "A person rides a motorcycle along a long, straight road that stretches between "
-    "a body of water and a forested hillside. The rider steadily accelerates, keeping "
-    "the motorcycle centered between the guardrails, while the scenery passes by on "
-    "both sides. The video captures the journey from the rider's perspective, emphasizing "
-    "the sense of motion and adventure."
-)
-
-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 video path
-VIDEO_PATH = "assets/motorcycle.mp4"
-
-# Number of conditioning frames from the input video
-NUM_COND_FRAMES = 13
-
-SEED = 42
-
-
-def basic_generation():
-    """
-    Run basic LongCat VC generation (50 steps at 480p).
-    
-    This uses the full 50-step denoising process for highest quality.
-    """
-    print("=" * 60)
-    print("LongCat VC: Basic Generation (50 steps, 480p)")
-    print("=" * 60)
-    
-    # Check if video exists
-    if not os.path.exists(VIDEO_PATH):
-        raise FileNotFoundError(
-            f"Video not found at {VIDEO_PATH}. "
-            "Please provide a valid video path."
-        )
-    
-    generator = VideoGenerator.from_pretrained(
-        "FastVideo/LongCat-Video-VC-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_vc_basic"
-    
-    generator.generate_video(
-        prompt=PROMPT,
-        negative_prompt=NEGATIVE_PROMPT,
-        video_path=VIDEO_PATH,
-        num_cond_frames=NUM_COND_FRAMES,
-        output_path=output_path,
-        save_video=True,
-        height=480,
-        width=832,
-        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 VC with distill+refine pipeline (16 steps + refinement to 720p).
-    
-    This uses the distilled LoRA for fast 480p generation (16 steps),
-    then refines to 720p using the refinement LoRA with BSA enabled.
-    """
-    print("\n" + "=" * 60)
-    print("LongCat VC: Distill + Refine Pipeline")
-    print("=" * 60)
-    
-    # Check if video exists
-    if not os.path.exists(VIDEO_PATH):
-        raise FileNotFoundError(
-            f"Video not found at {VIDEO_PATH}. "
-            "Please provide a valid video path."
-        )
-    
-    # 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-VC-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_vc_distill"
-    
-    generator.generate_video(
-        prompt=PROMPT,
-        negative_prompt=NEGATIVE_PROMPT,
-        video_path=VIDEO_PATH,
-        num_cond_frames=NUM_COND_FRAMES,
-        output_path=distill_output_path,
-        save_video=True,
-        height=480,
-        width=832,
-        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 -> 720p)
-    print("\n[Stage 2] Refinement (480p -> 720p 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 VC) since it's upscaling the generated video
-    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, 8],
-        bsa_chunk_k=[4, 4, 8],
-        lora_path="FastVideo/LongCat-Video-T2V-Refinement-LoRA",
-        lora_nickname="refinement",
-    )
-    
-    refine_output_path = "outputs_video/longcat_vc_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,  # For refinement, no conditioning frames
-        height=720,
-        width=1280,
-        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 Video Continuation 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()
-
-

examples/inference/basic/basic_matrixgame.py

@@ -1,75 +0,0 @@
-from fastvideo import VideoGenerator
-from fastvideo.configs.pipelines.wan import MatrixGameI2V480PConfig
-from fastvideo.models.dits.matrix_game.utils import create_action_presets
-
-import torch
-
-# Available variants: "base_distilled_model", "gta_distilled_model", "templerun_distilled_model"
-# Each variant has different keyboard_dim:
-#   - base_distilled_model: keyboard_dim=4
-#   - gta_distilled_model: keyboard_dim=2
-#   - templerun_distilled_model: keyboard_dim=7 (keyboard only, no mouse)
-MODEL_VARIANT = "base_distilled_model"
-
-# Variant-specific settings
-VARIANT_CONFIG = {
-    "base_distilled_model": {
-        "model_path": "FastVideo/Matrix-Game-2.0-Base-Diffusers",
-        "keyboard_dim": 4,
-        "image_url": "https://raw.githubusercontent.com/SkyworkAI/Matrix-Game/main/Matrix-Game-2/demo_images/universal/0000.png",
-    },
-    "gta_distilled_model": {
-        "model_path": "FastVideo/Matrix-Game-2.0-GTA-Diffusers",
-        "keyboard_dim": 2,
-        "image_url": "https://raw.githubusercontent.com/SkyworkAI/Matrix-Game/main/Matrix-Game-2/demo_images/gta_drive/0000.png",
-    },
-    "templerun_distilled_model": {
-        "model_path": "FastVideo/Matrix-Game-2.0-TempleRun-Diffusers",
-        "keyboard_dim": 7,
-        "image_url": "https://raw.githubusercontent.com/SkyworkAI/Matrix-Game/main/Matrix-Game-2/demo_images/temple_run/0000.png",
-    },
-}
-
-
-OUTPUT_PATH = "video_samples_matrixgame2"
-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.
-    config = VARIANT_CONFIG[MODEL_VARIANT]
-
-    generator = VideoGenerator.from_pretrained(
-        config["model_path"],
-        # 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, # DiT need to be offloaded for MoE
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=True,
-        # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer
-        pin_cpu_memory=True,
-        # image_encoder_cpu_offload=False,
-    )
-
-    num_frames = 597
-    actions = create_action_presets(num_frames, keyboard_dim=config["keyboard_dim"])
-    grid_sizes = torch.tensor([150, 44, 80])
-
-    generator.generate_video(
-        prompt="",
-        image_path=config["image_url"],
-        mouse_cond=actions["mouse"].unsqueeze(0),
-        keyboard_cond=actions["keyboard"].unsqueeze(0),
-        grid_sizes=grid_sizes,
-        num_frames=num_frames,
-        height=352,
-        width=640,
-        num_inference_steps=50,
-        output_path=OUTPUT_PATH,
-        save_video=True,
-    )
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_matrixgame_streaming.py

@@ -1,98 +0,0 @@
-from fastvideo.entrypoints.streaming_generator import StreamingVideoGenerator
-from fastvideo.models.dits.matrix_game.utils import get_current_action_async, expand_action_to_frames
-
-import torch
-import asyncio
-
-# Available variants: "base_distilled_model", "gta_distilled_model", "templerun_distilled_model"
-# Each variant has different keyboard_dim:
-#   - base_distilled_model: keyboard_dim=4
-#   - gta_distilled_model: keyboard_dim=2
-#   - templerun_distilled_model: keyboard_dim=7 (keyboard only, no mouse)
-MODEL_VARIANT = "base_distilled_model"
-
-# Variant-specific settings
-VARIANT_CONFIG = {
-    "base_distilled_model": {
-        "model_path": "FastVideo/Matrix-Game-2.0-Base-Diffusers",
-        "keyboard_dim": 4,
-        "mode": "universal",
-        "image_url": "https://raw.githubusercontent.com/SkyworkAI/Matrix-Game/main/Matrix-Game-2/demo_images/universal/0000.png",
-    },
-    "gta_distilled_model": {
-        "model_path": "FastVideo/Matrix-Game-2.0-GTA-Diffusers",
-        "keyboard_dim": 2,
-        "mode": "gta_drive",
-        "image_url": "https://raw.githubusercontent.com/SkyworkAI/Matrix-Game/main/Matrix-Game-2/demo_images/gta_drive/0000.png",
-    },
-    "templerun_distilled_model": {
-        "model_path": "FastVideo/Matrix-Game-2.0-TempleRun-Diffusers",
-        "keyboard_dim": 7,
-        "mode": "templerun",
-        "image_url": "https://raw.githubusercontent.com/SkyworkAI/Matrix-Game/main/Matrix-Game-2/demo_images/temple_run/0000.png",
-    },
-}
-
-
-OUTPUT_PATH = "video_samples_matrixgame2"
-async 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.
-    config = VARIANT_CONFIG[MODEL_VARIANT]
-
-    generator = StreamingVideoGenerator.from_pretrained(
-        config["model_path"],
-        # 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, # DiT need to be offloaded for MoE
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=True,
-        # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer
-        pin_cpu_memory=True,
-        # image_encoder_cpu_offload=False,
-    )
-
-    max_blocks = 50
-    num_frames = 597    
-    actions = {
-        "keyboard": torch.zeros((num_frames, config["keyboard_dim"])),
-        "mouse": torch.zeros((num_frames, 2))
-    }
-    grid_sizes = torch.tensor([150, 44, 80])
-    mode = config["mode"]
-
-    generator.reset(
-        prompt="",
-        image_path=config["image_url"],
-        mouse_cond=actions["mouse"].unsqueeze(0),
-        keyboard_cond=actions["keyboard"].unsqueeze(0),
-        grid_sizes=grid_sizes,
-        num_frames=num_frames,
-        height=352,
-        width=640,
-        num_inference_steps=50,
-        output_path=OUTPUT_PATH,
-        save_video=True,
-    )
-    print("Initialization complete.")
-
-    for block_id in range(max_blocks):
-        print(f"\n=== Block {block_id + 1}/{max_blocks} ===")
-        
-        action = await get_current_action_async(mode)
-        keyboard_cond, mouse_cond = expand_action_to_frames(action, 12)
-        await generator.step_async(keyboard_cond, mouse_cond)
-        
-        if (await asyncio.to_thread(input, "\nContinue? (y/n): ")).lower() == 'n':
-            break
-
-    # Save final video
-    generator.finalize()
-    generator.shutdown()
-
-
-if __name__ == "__main__":
-    asyncio.run(main())

examples/inference/basic/basic_self_forcing_causal.py

@@ -13,7 +13,7 @@ def main():
         model_name,
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory
+        use_fsdp_inference=True,
         text_encoder_cpu_offload=False,
         dit_cpu_offload=False,
     )

examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py

@@ -14,7 +14,7 @@ def main():
         "FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers",
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory
+        use_fsdp_inference=True,
         dit_cpu_offload=True, # DiT need to be offloaded for MoE
         dit_precision="fp32",
         vae_cpu_offload=False,

examples/inference/basic/basic_self_forcing_causal_wan2_2_t2v.py

@@ -14,7 +14,7 @@ def main():
         "rand0nmr/SFWan2.2-T2V-A14B-Diffusers",
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory
+        use_fsdp_inference=True,
         dit_cpu_offload=True, # DiT need to be offloaded for MoE
         vae_cpu_offload=False,
         text_encoder_cpu_offload=True,

examples/inference/basic/basic_turbodiffusion.py

@@ -1,55 +0,0 @@
-import os
-
-# Set SLA attention backend BEFORE fastvideo imports
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLA_ATTN"
-
-from fastvideo import VideoGenerator
-
-OUTPUT_PATH = "video_samples_turbodiffusion"
-
-
-def main() -> None:
-    # TurboDiffusion: 1-4 step video generation using RCM scheduler + SLA attention
-    # FastVideo will automatically use TurboDiffusionPipeline when specified
-    generator = VideoGenerator.from_pretrained(
-        "loayrashid/TurboWan2.1-T2V-1.3B-Diffusers",
-        # FastVideo will automatically handle distributed setup
-        num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory
-
-        # set to false if using RTX 4090 
-        # pin_cpu_memory=False,
-    )
-
-    # Generate videos with the same simple API, regardless of GPU count
-    # TurboDiffusion defaults: guidance_scale=1.0 and num_inference_steps=4 (from config)
-    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,
-        seed=42,
-    )
-
-    # Generate another video with a different prompt, without reloading the model!
-    prompt2 = (
-        "A majestic lion strides across the golden savanna, its powerful frame "
-        "glistening under the warm afternoon sun. The tall grass ripples gently in "
-        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-        "cinematic."
-    )
-    video2 = generator.generate_video(
-        prompt2,
-        output_path=OUTPUT_PATH,
-        save_video=True,
-        seed=42,
-    )
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_turbodiffusion_14b.py

@@ -1,49 +0,0 @@
-import os
-
-# Set SLA attention backend BEFORE fastvideo imports
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLA_ATTN"
-
-from fastvideo import VideoGenerator
-
-OUTPUT_PATH = "video_samples_turbodiffusion_14B"
-
-
-def main() -> None:
-    # TurboDiffusion 14B: 1-4 step video generation using RCM scheduler + SLA attention
-    # FastVideo will automatically use TurboDiffusionPipeline when specified
-    generator = VideoGenerator.from_pretrained(
-        "loayrashid/TurboWan2.1-T2V-14B-Diffusers",
-        # 14B model needs more GPUs
-        num_gpus=2,
-    )
-
-    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,
-        seed=42,
-    )
-
-    # Generate another video with a different prompt, without reloading the model!
-    prompt2 = (
-        "A majestic lion strides across the golden savanna, its powerful frame "
-        "glistening under the warm afternoon sun. The tall grass ripples gently in "
-        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-        "cinematic."
-    )
-    video2 = generator.generate_video(
-        prompt2,
-        output_path=OUTPUT_PATH,
-        save_video=True,
-        seed=42,
-    )
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_turbodiffusion_i2v.py

@@ -1,37 +0,0 @@
-import os
-
-# Set SLA attention backend BEFORE fastvideo imports
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLA_ATTN"
-
-from fastvideo import VideoGenerator
-
-# Use local model path
-MODEL_PATH = "loayrashid/TurboWan2.2-I2V-A14B-Diffusers"
-OUTPUT_PATH = "video_samples_turbodiffusion_i2v"
-
-
-def main() -> None:
-    # TurboDiffusion I2V: 1-4 step image-to-video generation
-    generator = VideoGenerator.from_pretrained(
-        MODEL_PATH,
-        num_gpus=2,
-    )
-
-    # Example prompt and image for I2V
-    prompt = ("Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside.")
-
-    # Use an example image path
-    image_path = "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG"
-
-
-    video = generator.generate_video(
-        prompt,
-        image_path=image_path,
-        output_path=OUTPUT_PATH,
-        save_video=True,
-        seed=42,
-    )
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_wan2_2.py

@@ -12,7 +12,7 @@ def main():
         "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory 
+        use_fsdp_inference=True,
         dit_cpu_offload=True, # DiT need to be offloaded for MoE
         vae_cpu_offload=False,
         text_encoder_cpu_offload=True,

examples/inference/basic/basic_wan2_2_Fun.py

@@ -14,7 +14,7 @@ def main():
         # "alibaba-pai/Wan2.2-Fun-A14B-Control",
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory
+        use_fsdp_inference=True,
         dit_cpu_offload=True, # DiT need to be offloaded for MoE
         vae_cpu_offload=False,
         text_encoder_cpu_offload=True,

examples/inference/basic/basic_wan2_2_i2v.py

@@ -12,7 +12,7 @@ def main():
         "Wan-AI/Wan2.2-I2V-A14B-Diffusers",
         # FastVideo will automatically handle distributed setup
         num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory
+        use_fsdp_inference=True,
         dit_cpu_offload=True, # DiT need to be offloaded for MoE
         vae_cpu_offload=False,
         text_encoder_cpu_offload=True,