update docker

Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com>

.buildkite/pipeline.yml

@@ -222,3 +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"

.buildkite/scripts/pr_test.sh

@@ -31,9 +31,9 @@ log "Setting up Modal authentication from Buildkite secrets..."
 MODAL_TOKEN_ID=$(buildkite-agent secret get modal_token_id)
 MODAL_TOKEN_SECRET=$(buildkite-agent secret get modal_token_secret)
 
+# Retrieve other secrets
 WANDB_API_KEY=$(buildkite-agent secret get wandb_api_key)
-
-WANDB_API_KEY=$(buildkite-agent secret get wandb_api_key)
+HF_API_KEY=$(buildkite-agent secret get hf_api_key)
 
 if [ -n "$MODAL_TOKEN_ID" ] && [ -n "$MODAL_TOKEN_SECRET" ]; then
     log "Retrieved Modal credentials from Buildkite secrets"
@@ -63,19 +63,19 @@ MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUI
 case "$TEST_TYPE" in
     "encoder")
         log "Running encoder tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_encoder_tests"
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_encoder_tests"
         ;;
     "vae")
         log "Running VAE tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_vae_tests"
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_vae_tests"
         ;;
     "transformer")
         log "Running transformer tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_transformer_tests"
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_transformer_tests"
         ;;
     "ssim")
         log "Running SSIM tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_ssim_tests"
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_ssim_tests"
         ;;
     "training")
         log "Running training tests..."
@@ -126,6 +126,10 @@ case "$TEST_TYPE" in
         log "Running unit tests..."
         MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_unit_test"
         ;;
+    "lora_extraction")
+        log "Running LoRA extraction tests..."
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_lora_extraction_tests"
+        ;;
     *)
         log "Error: Unknown test type: $TEST_TYPE"
         exit 1

.github/workflows/docs.yml

@@ -1,82 +1,65 @@
-# Sample workflow for building and deploying a Hugo site to GitHub Pages
-name: Deploy FastVideo Docs to Pages
+name: Deploy Documentation
 
 on:
-  # Runs on pushes targeting the default branch
   push:
-    branches:
-      - main
+    branches: [ main ]
     paths:
-      - "docs/**/*.md"
-      - "fastvideo/examples/**/*.py"
+      - 'docs/**'
+      - 'mkdocs.yml'
+      - 'requirements-mkdocs.txt'
+      - '.github/workflows/docs.yml'
   pull_request:
-    branches: 
-      - main
-    types: [opened, ready_for_review, synchronize, reopened]
+    branches: [ main ]
     paths:
-      - "docs/**/*.md"
-      - "fastvideo/examples/**/*.py"
+      - 'docs/**'
+      - 'mkdocs.yml'
+      - 'requirements-mkdocs.txt'
+      - '.github/workflows/docs.yml'
 
-  # Allows you to run this workflow manually from the Actions tab
-  workflow_dispatch:
-
-# Sets permissions of the GITHUB_TOKEN to allow deployment to GitHub Pages
 permissions:
   contents: read
   pages: write
   id-token: write
 
-# Allow only one concurrent deployment, skipping runs queued between the run in-progress and latest queued.
-# However, do NOT cancel in-progress runs as we want to allow these production deployments to complete.
 concurrency:
   group: "pages"
   cancel-in-progress: false
 
-# Default to bash
-defaults:
-  run:
-    shell: bash
-
 jobs:
-  pre-commit:
-    uses: ./.github/workflows/pre-commit.yml
-
-  # Build job
   build:
     runs-on: ubuntu-latest
-    needs: pre-commit
     steps:
       - name: Checkout
         uses: actions/checkout@v4
-      - name: Setup Pages
-        id: pages
-        uses: actions/configure-pages@v5
-      - name: Set up Python
+
+      - name: Setup Python
         uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: '3.12'
+
       - name: Install dependencies
         run: |
-          cd docs
-          pip install -r requirements-docs.txt
-      - name: Build docs
-        run: |
-          cd docs
-          make clean
-          make html
+          python -m pip install --upgrade pip
+          pip install -r requirements-mkdocs.txt
+
+      - name: Setup Pages
+        uses: actions/configure-pages@v4
+
+      - name: Build documentation
+        run: mkdocs build
+
       - name: Upload artifact
         uses: actions/upload-pages-artifact@v3
         with:
-          path: ./docs/build/html
+          path: ./site
 
-  # Deployment job
   deploy:
     environment:
       name: github-pages
       url: ${{ steps.deployment.outputs.page_url }}
-    if: ${{ github.event_name == 'push' }}
     runs-on: ubuntu-latest
     needs: build
+    if: github.ref == 'refs/heads/main'
     steps:
       - name: Deploy to GitHub Pages
         id: deployment

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

@@ -0,0 +1,236 @@
+name: Publish FastVideo Kernel to PyPI on Version Change
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - "csrc/fastvideo_kernel/pyproject.toml"
+  workflow_dispatch:
+
+jobs:
+  check-version-change:
+    runs-on: ubuntu-latest
+    outputs:
+      version-changed: ${{ steps.check-version.outputs.changed }}
+      new-version: ${{ steps.check-version.outputs.new-version }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 2
+
+      - name: Check if version changed
+        id: check-version
+        run: |
+          cd csrc/fastvideo_kernel
+          # Get current commit's version from pyproject.toml
+          NEW_VERSION=$(grep -oP 'version\s*=\s*"\K[^"]+' pyproject.toml)
+          echo "New version: $NEW_VERSION"
+          
+          # Get previous version from git history
+          OLD_VERSION=$(git show HEAD~1:./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', '3.13']
+          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'
+
+    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
+          
+          cd csrc/fastvideo_kernel
+          git submodule update --init --recursive  # Ensure ThunderKittens submodule is initialized
+          python setup.py bdist_wheel --dist-dir=dist
+
+      - name: Rename wheel file
+        run: |
+          cd csrc/fastvideo_kernel
+
+          CUDA_SHORT_VERSION=$(echo ${{ matrix.torch-cuda.cuda-version }} | cut -d. -f1,2 | sed 's/\.//g')
+          TORCH_SHORT_VERSION=$(echo ${{ matrix.torch-cuda.torch-version }} | cut -d. -f1,2)
+          # Get the correct version format
+          tmpname=cu${CUDA_SHORT_VERSION}torch${TORCH_SHORT_VERSION}
+          wheel_name=$(ls dist/*whl | xargs -n 1 basename | sed "s/-/+$tmpname-/2")
+          # Rename with version information
+          ls dist/*whl |xargs -I {} mv {} dist/${wheel_name}
+          echo "wheel_name=${wheel_name}" >> $GITHUB_ENV
+
+      - name: Upload wheel artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: ${{ env.wheel_name }}-py${{ matrix.python-version }}
+          path: csrc/fastvideo_kernel/dist/*.whl
+          retention-days: 90
+
+  publish_package:
+    name: Publish package
+    needs: [build_wheels, check-version-change]
+    if: ${{ needs.check-version-change.outputs.version-changed == 'true' || github.event_name == 'workflow_dispatch' }}
+    runs-on: ubuntu-22.04
+    permissions:
+      id-token: write  # Needed for OIDC Trusted Publishing
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+
+      - name: Install CUDA 12.4.1
+        uses: Jimver/cuda-toolkit@v0.2.21
+        id: cuda-toolkit
+        with:
+          cuda: 12.4.1
+          linux-local-args: '["--toolkit"]'
+          method: 'network'
+          sub-packages: '["nvcc"]'
+
+      - name: Install dependencies (GCC, Clang, CUDA Paths, Git)
+        run: |
+          sudo apt update
+          sudo apt install -y git patchelf gcc-11 g++-11 clang-11
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+
+          # Allow Git to Access Safe Directory
+          git config --global --add safe.directory /__w/FastVideo/FastVideo
+
+          # Set CUDA environment variables
+          export CUDA_HOME=/usr/local/cuda-12.4.1
+          export PATH=${CUDA_HOME}/bin:${PATH}
+          export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+          
+          # Verify installation
+          gcc --version
+          g++ --version
+          clang-11 --version
+          nvcc --version
+
+      - name: Install PyTorch 2.5.1+cu12.4.1
+        run: |
+          pip install --upgrade pip
+          pip install typing-extensions==4.12.2
+          export TORCH_CUDA_VERSION=124
+          pip install --no-cache-dir torch==2.5.1 --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
+          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 source distribution
+        run: |
+          export PYTHONPATH=$GITHUB_WORKSPACE:$PYTHONPATH
+          
+          pip install setuptools ninja packaging wheel triton
+          
+          cd csrc/fastvideo_kernel
+          git submodule update --init --recursive
+          python setup.py sdist --dist-dir=dist
+
+      - name: Publish release distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          packages-dir: csrc/fastvideo_kernel/dist/

.gitignore

@@ -14,6 +14,8 @@ wandb/
 *.pt
 cache_dir/
 wandb/
+venv/
+.venv/
 runs/
 samples/
 *validation/
@@ -28,6 +30,8 @@ env
 **/build/
 **.pyc
 **.txt
+*.log
+weights/
 
 # Distribution / packaging
 build/
@@ -37,12 +41,13 @@ dist/
 eggs/
 .eggs/
 
-# Sphinx documentation
-docs/_build/
-docs/source/getting_started/examples/
-docs/source/inference/examples/
-docs/source/training/examples/
-docs/source/distillation/examples/
+# MkDocs documentation
+site/
+docs/getting_started/examples/
+docs/inference/examples/
+docs/training/examples/
+docs/distillation/examples/
+!requirements-mkdocs.txt
 
 # VSCode
 .vscode/
@@ -61,7 +66,7 @@ docs/source/distillation/examples/
 !fastvideo/tests/ssim/reference_videos/**/*.mp4
 
 # Static images
-!docs/source/_static/images/**/*.png
+!docs/assets/images/**/*.png
 !comfyui/assets/**/*.png
 !comfyui/assets/**/*.gif
 

.pre-commit-config.yaml

@@ -10,6 +10,7 @@ exclude: |
     demo/.*|
     predict\.py|
     scripts/.*|
+    prompts/.*|
     fastvideo/data_preprocess/.*|
     fastvideo/dataset/.*|
     fastvideo/models/.*|

README.md

@@ -1,13 +1,12 @@
 <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://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.html"><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-3csdw1isz-Euq8_Q8~baewG8hxjXs2gQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/tMwknPLY" 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>
 
 <div align="center">
@@ -15,7 +14,8 @@ FastVideo features an end-to-end unified pipeline for accelerating diffusion mod
 </div>
 
 ## NEWS
-- ```2025/08/04```: Release [FastWan](https://hao-ai-lab.github.io/FastVideo/distillation/dmd.html) models and [Sparse-Distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).
+- ```2025/11/19```: Release [CausalWan2.2 I2V A14B Preview](https://huggingface.co/FastVideo/CausalWan2.2-I2V-A14B-Preview-Diffusers) models, [Blog](https://hao-ai-lab.github.io/blogs/fastvideo_causalwan_preview/) and [Inference Code!](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py)
+- ```2025/08/04```: Release [FastWan](https://hao-ai-lab.github.io/FastVideo/distillation/dmd) models and [Sparse-Distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).
 - ```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/).
@@ -49,10 +49,10 @@ conda activate fastvideo
 pip install fastvideo
 ```
 
-Please see our [docs](https://hao-ai-lab.github.io/FastVideo/getting_started/installation.html) for more detailed installation instructions.
+Please see our [docs](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/) for more detailed installation instructions.
 
 ## Sparse Distillation
-For our sparse distillation techniques, please see our [distillation docs](https://hao-ai-lab.github.io/FastVideo/distillation/dmd.html) and check out our [blog](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).
+For our sparse distillation techniques, please see our [distillation docs](https://hao-ai-lab.github.io/FastVideo/distillation/dmd/) and check out our [blog](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).
 
 See below for recipes and datasets:
 
@@ -64,7 +64,7 @@ See below for recipes and datasets:
 
 ## Inference
 ### Generating Your First Video
-Here's a minimal example to generate a video using the default settings. Make sure VSA kernels are [installed](https://hao-ai-lab.github.io/FastVideo/video_sparse_attention/installation.html). Create a file called `example.py` with the following code:
+Here's a minimal example to generate a video using the default settings. Make sure VSA kernels are [installed](https://hao-ai-lab.github.io/FastVideo/video_sparse_attention/installation/). Create a file called `example.py` with the following code:
 
 ```python
 import os
@@ -100,35 +100,32 @@ Run the script with:
 python example.py
 ```
 
-For a more detailed guide, please see our [inference quick start](https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html).
+For a more detailed guide, please see our [inference quick start](https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start/).
 
 ### Other docs:
 
-- [Design Overview](https://hao-ai-lab.github.io/FastVideo/design/overview.html)
-- [Contribution Guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation.html)
+- [Design Overview](https://hao-ai-lab.github.io/FastVideo/design/overview/)
+- [Contribution Guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/)
 
 ## Distillation and Finetuning
-- [Distillation Guide](https://hao-ai-lab.github.io/FastVideo/distillation/dmd.html)
+- [Distillation Guide](https://hao-ai-lab.github.io/FastVideo/distillation/dmd/)
 <!-- - [Finetuning Guide](https://hao-ai-lab.github.io/FastVideo/training/finetune.html) -->
 
-## 📑 Development Plan
-<!-- - More distillation methods -->
-  <!-- - [ ] Add Distribution Matching Distillation -->
-More FastWan Models Coming Soon!
-- [ ] Add FastWan2.1-T2V-14B
-- [ ] Add FastWan2.2-T2V-14B
-- [ ] Add FastWan2.2-I2V-14B
-<!-- - Optimization features
-- Code updates -->
-  <!-- - [ ] fp8 support -->
-  <!-- - [ ] faster load model and save model support -->
+## Awesome work using FastVideo or our research projects
 
-See details in [development roadmap](https://github.com/hao-ai-lab/FastVideo/issues/468).
+- [SGLang](https://github.com/sgl-project/sglang/tree/main/python/sglang/multimodal_gen): SGLang's diffusion inference functionality is based  on a fork of FastVideo on Sept. 24, 2025. [![Star](https://img.shields.io/github/stars/sgl-project/sglang.svg?style=social&label=Star)](https://github.com/sgl-project/sglang)
+
+- [DanceGRPO](https://github.com/XueZeyue/DanceGRPO): A  unified framework to adapt Group Relative Policy Optimization (GRPO) to visual generation paradigms. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/XueZeyue/DanceGRPO.svg?style=social&label=Star)](https://github.com/XueZeyue/DanceGRPO)
+- [SRPO](https://github.com/Tencent-Hunyuan/SRPO): A method to directly align the full diffusion trajectory with fine-grained human preference. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/Tencent-Hunyuan/SRPO.svg?style=social&label=Star)](https://github.com/Tencent-Hunyuan/SRPO)
+- [DCM](https://github.com/Vchitect/DCM): Dual-expert consistency model for efficient and high-quality video generation. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/Vchitect/DCM.svg?style=social&label=Star)](https://github.com/Vchitect/DCM)
+- [Hunyuan Video 1.5](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5): A leading lightweight video generation model, where they proposed SSTA based on Sliding Tile Attention. [![Star](https://img.shields.io/github/stars/Tencent-Hunyuan/HunyuanVideo-1.5.svg?style=social&label=Star)](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5)
+- [Kandinsky-5.0](https://github.com/kandinskylab/kandinsky-5): A family of diffusion models for video & image generation, where their NABLA attention includes a Sliding Tile Attention branch. [![Star](https://img.shields.io/github/stars/kandinskylab/kandinsky-5.svg?style=social&label=Star)](https://github.com/kandinskylab/kandinsky-5)
+- [LongCat Video](https://github.com/meituan-longcat/LongCat-Video): A foundational video generation model with 13.6B parameters with block-sparse attention similar to Video Sparse Attention. [![Star](https://img.shields.io/github/stars/meituan-longcat/LongCat-Video.svg?style=social&label=Star)](https://github.com/meituan-longcat/LongCat-Video)
 
 ## 🤝 Contributing
 
-We welcome all contributions. Please check out our guide [here](https://hao-ai-lab.github.io/FastVideo/contributing/overview.html)
-
+We welcome all contributions. Please check out our guide [here](https://hao-ai-lab.github.io/FastVideo/contributing/overview/).
+See details in [development roadmap](https://github.com/hao-ai-lab/FastVideo/issues/899).
 ## Acknowledgement
 We learned and reused code from the following projects:
 - [Wan-Video](https://github.com/Wan-Video)

benchmarks/fvd/README.md

@@ -0,0 +1,103 @@
+# FVD (Fréchet Video Distance) Benchmark
+
+Evaluate generated video quality using FVD with the I3D feature extractor.
+
+## Quick Start
+
+**Run the benchmark:**
+
+```bash
+bash benchmarks/scripts/run.sh
+```
+
+That's it! The script auto-installs dependencies and runs the benchmark.
+
+**To customize:** Edit `benchmarks/fvd/run_fvd.py` to change:
+- Video paths (`real_dir`, `gen_dir`)
+- Number of videos, frames, sampling strategy
+- Device, batch size, caching, etc.
+
+## Advanced Usage (CLI)
+
+For more control without editing Python files, use the CLI.
+
+**First-time setup** (one-time per pod/environment):
+
+```bash
+bash benchmarks/scripts/setup_fvd.sh
+```
+
+Then run any configuration you want:
+
+```bash
+# Custom configuration
+python -m benchmarks.fvd.cli \
+    --real-path data/real/ \
+    --gen-path outputs/gen/ \
+    --num-videos 1024 \
+    --num-frames 32 \
+    --clip-strategy random \
+    --batch-size 32 \
+    --seed 42
+```
+
+**Standard protocols:**
+
+```bash
+# Use predefined protocols
+python -m benchmarks.fvd.cli \
+    --real-path data/real/ \
+    --gen-path outputs/gen/ \
+    --protocol fvd2048_16f  # or fvd2048_128f, quick_test, etc.
+```
+
+**Feature caching** (speed up repeated evaluations):
+
+```bash
+python -m benchmarks.fvd.cli \
+    --real-path data/real/ \
+    --gen-path outputs/gen/ \
+    --protocol fvd2048_16f \
+    --cache-real-features cache/real  # Directory path (will save/load cache/real/real_features.pkl)
+```
+
+Run `python -m benchmarks.fvd.cli --help` for all options.
+
+## Available Protocols
+
+- `fvd2048_16f` - Standard (2048 videos, 16 frames)
+- `fvd2048_128f` - Long videos (128 frames)
+- `fvd2048_128f_subsample8` - Subsampled long videos
+- `quick_test` - Fast testing (10 videos)
+
+## Configuration Options
+
+Key options in `FVDConfig`:
+
+```python
+num_videos=2048,              # Videos to evaluate
+num_frames_per_clip=16,       # Frames per clip
+clip_strategy='beginning',    # beginning|random|uniform|middle|sliding
+frame_stride=1,               # Frame subsampling
+batch_size=32,                # GPU batch size
+device='cuda',                # cuda|cpu
+cache_real_features=None,     # Cache path for speed
+seed=42,                      # Reproducibility
+```
+
+## Programmatic Usage
+
+```python
+from benchmarks.fvd import compute_fvd_with_config, FVDConfig
+
+config = FVDConfig.fvd2048_16f()  # or custom config
+results = compute_fvd_with_config('data/real/', 'outputs/gen/', config)
+print(f"FVD: {results['fvd']:.2f}")
+```
+
+## Notes
+
+- 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

@@ -0,0 +1,35 @@
+"""
+FastVideo Frechet Video Distance (FVD) Benchmark Module.
+    >>> from fastvideo.benchmarks.fvd import compute_fvd_with_config, FVDConfig
+    >>> config = FVDConfig.fvd2048_16f()  # Standard protocol
+    >>> results = compute_fvd_with_config('data/real/', 'outputs/gen/', config)
+    >>> print(f"FVD: {results['fvd']:.2f}")
+"""
+
+from .fvd import (
+    compute_fvd,
+    compute_fvd_with_config,
+    compute_frechet_distance,
+    compute_statistics,
+    FVDConfig,
+)
+from .i3d_model import I3DFeatureExtractor
+from .video_utils import (
+    load_video_auto,
+    sample_clips_from_video,
+    load_video_clips_streaming,
+    ClipSamplingStrategy,
+)
+
+__all__ = [
+    'compute_fvd',
+    'compute_fvd_with_config',
+    'compute_frechet_distance',
+    'compute_statistics',
+    'FVDConfig',
+    'I3DFeatureExtractor',
+    'load_video_auto',
+    'sample_clips_from_video',
+    'load_video_clips_streaming',
+    'ClipSamplingStrategy',
+]

benchmarks/fvd/cli.py

@@ -0,0 +1,185 @@
+import argparse
+import json
+import sys
+from pathlib import Path
+
+from .fvd import compute_fvd_with_config, FVDConfig
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        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 directory')
+    parser.add_argument('--gen-path',
+                        type=str,
+                        required=True,
+                        help='Path to generated videos directory')
+
+    # Reproducibility
+    parser.add_argument(
+        '--seed',
+        type=int,
+        default=None,
+        help='Random seed for reproducibility (np.random, random, torch)')
+
+    # Protocol presets
+    parser.add_argument('--protocol',
+                        type=str,
+                        default=None,
+                        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 (default: 2048)')
+
+    # Clip sampling
+    parser.add_argument('--num-frames',
+                        type=int,
+                        default=16,
+                        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 (default: 32)')
+    parser.add_argument('--device',
+                        type=str,
+                        default='cuda',
+                        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')
+
+    args = parser.parse_args()
+
+    # Create config
+    if args.protocol:
+        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]()
+
+        # Override device and caching from args
+        config.device = args.device
+        config.cache_real_features = args.cache_real_features
+        config.i3d_model_path = args.i3d_model_path
+        config.batch_size = args.batch_size
+        config.seed = args.seed
+    else:
+        # 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:
+        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)
+        import traceback
+        traceback.print_exc()
+        return 1
+
+
+if __name__ == '__main__':
+    sys.exit(main())

benchmarks/fvd/fvd.py

@@ -0,0 +1,447 @@
+import numpy as np
+import scipy.linalg
+import torch
+from pathlib import Path
+from collections.abc import Iterator
+import pickle
+from dataclasses import dataclass, field
+
+from .i3d_model import I3DFeatureExtractor
+from .video_utils import ClipSamplingStrategy, load_video_clips_streaming
+
+
+def compute_statistics(features: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+    """Compute mean and covariance."""
+    mu = np.mean(features, axis=0)
+    sigma = np.cov(features, rowvar=False)
+    return mu, sigma
+
+
+def compute_frechet_distance(mu1: np.ndarray,
+                             sigma1: np.ndarray,
+                             mu2: np.ndarray,
+                             sigma2: np.ndarray,
+                             eps: float = 1e-6) -> float:
+    """
+    Compute Fréchet distance between two Gaussians.
+    """
+    sigma1 = sigma1 + eps * np.eye(sigma1.shape[0])
+    sigma2 = sigma2 + eps * np.eye(sigma2.shape[0])
+
+    diff = mu1 - mu2
+    mean_distance = np.sum(diff**2)
+
+    trace_sum = np.trace(sigma1 + sigma2)
+
+    covmean = scipy.linalg.sqrtm(sigma1 @ sigma2)
+
+    if np.iscomplexobj(covmean):
+        if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3):
+            print(
+                f"Warning: Imaginary component: {np.max(np.abs(covmean.imag))}")
+        covmean = covmean.real
+
+    trace_product = np.trace(covmean)
+
+    fvd = mean_distance + trace_sum - 2 * trace_product
+
+    return float(fvd)
+
+
+@dataclass
+class FVDConfig:
+    # default configuration for FVD computation:
+
+    # Video selection
+    num_videos: int = 2048
+
+    # Clip sampling
+    num_frames_per_clip: int = 16
+    num_clips_per_video: int = 1
+    clip_strategy: str | ClipSamplingStrategy = 'beginning'
+
+    # Temporal subsampling
+    frame_stride: int = 1  # 1=no subsampling, 2=every 2nd, 8=every 8th
+    temporal_stride: int = 1  # For sliding window clips
+
+    # Data processing
+    video_extensions: list[str] = field(
+        default_factory=lambda: ['.mp4', '.avi', '.mov', '.mkv'])
+    support_frame_dirs: bool = True
+
+    # Computation
+    batch_size: int = 32
+    device: str = 'cuda'
+
+    use_streaming: bool = True
+    resize_before_extraction: bool = True
+
+    # Caching
+    cache_real_features: str | None = None
+    i3d_model_path: str | None = None
+
+    # Reproducibility
+    seed: int | None = None
+
+    @classmethod
+    def fvd2048_16f(cls) -> 'FVDConfig':
+        """
+        Standard FVD protocol: 2048 videos, 16 frames, beginning clip.
+        
+        most common FVD configuration used in papers
+        """
+        return cls(num_videos=2048,
+                   num_frames_per_clip=16,
+                   clip_strategy='beginning',
+                   use_streaming=True)
+
+    @classmethod
+    def fvd2048_128f(cls) -> 'FVDConfig':
+        """Long video protocol: 2048 videos, 128 frames."""
+        return cls(num_videos=2048,
+                   num_frames_per_clip=128,
+                   clip_strategy='beginning',
+                   use_streaming=True)
+
+    @classmethod
+    def 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."""
+        return cls(num_videos=100,
+                   num_frames_per_clip=16,
+                   clip_strategy='beginning')
+
+    def to_dict(self) -> dict:
+        """Export config to dict for logging"""
+        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.num_videos}_{self.num_frames_per_clip}f"
+        if self.frame_stride > 1:
+            desc += f"_subsample{self.frame_stride}"
+        if self.num_clips_per_video > 1:
+            desc += f"_{self.num_clips_per_video}clips"
+        if self.clip_strategy != 'beginning':
+            desc += f"_{self.clip_strategy}"
+        return desc
+
+
+def extract_features_streaming(video_generator: Iterator[torch.Tensor],
+                               extractor: 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}...")
+
+    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_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")
+
+    features = np.concatenate(all_features, axis=0)
+
+    if verbose:
+        print(f"Extracted {len(features)} feature vectors")
+
+    return features
+
+
+def load_or_compute_features(videos: str | Path | torch.Tensor,
+                             extractor: 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:
+        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:
+                features = pickle.load(f)
+
+            # Validate and limit based on config
+            max_features = config.num_videos * config.num_clips_per_video
+            if len(features) < max_features:
+                print(
+                    f"WARNING: Cache has {len(features)} features but need {max_features}"
+                )
+                print("Recomputing features...")
+            elif len(features) > max_features:
+                features = features[:max_features]
+                return features
+            else:
+                return features
+
+    # Compute features
+    if isinstance(videos, str | Path):
+        target_size = (224, 224) if config.resize_before_extraction else None
+
+        video_generator = load_video_clips_streaming(
+            videos,
+            num_frames=config.num_frames_per_clip,
+            max_videos=config.num_videos,
+            clip_strategy=config.clip_strategy,
+            frame_stride=config.frame_stride,
+            num_clips_per_video=config.num_clips_per_video,
+            video_extensions=config.video_extensions,
+            support_frame_dirs=config.support_frame_dirs,
+            target_size=target_size,
+            verbose=True)
+
+        max_clips = config.num_videos * config.num_clips_per_video
+        features = extract_features_streaming(video_generator,
+                                              extractor,
+                                              batch_size=config.batch_size,
+                                              max_clips=max_clips,
+                                              verbose=True)
+
+    else:
+        # Already a tensor
+        print(f"Extracting features from {len(videos)} video tensors...")
+        features = extractor.extract_features(videos,
+                                              batch_size=config.batch_size,
+                                              verbose=True)
+        features = features.numpy()
+
+    # Validate feature count
+    expected_count = config.num_videos * config.num_clips_per_video
+    if len(features) < expected_count:
+        raise ValueError(
+            f"ERROR: Only extracted {len(features)} features, but need {expected_count}!\n"
+            f"Found fewer videos than expected. Check your video directory.")
+    elif len(features) > expected_count:
+        print(f"Truncating {len(features)} features to {expected_count}")
+        features = features[:expected_count]
+
+    # Cache features if requested
+    if cache_path is not None:
+        cache_dir = Path(cache_path)
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        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)
+
+    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)
+            - '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
+        _rnd.seed(config.seed)
+        np.random.seed(config.seed)
+        torch.manual_seed(config.seed)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(config.seed)
+
+    if verbose:
+        print("=" * 70)
+        print(f"Computing FVD with protocol: {config}")
+        print("=" * 70)
+        print("\nConfiguration:")
+        for key, value in config.to_dict().items():
+            print(f"  {key}: {value}")
+        print()
+
+    # Initialize I3D
+    if verbose:
+        print(f"\nInitializing I3D model on {config.device}...")
+
+    extractor = I3DFeatureExtractor(device=config.device,
+                                    cache_dir=config.i3d_model_path)
+
+    # Extract features
+    if verbose:
+        print(f"\n{'='*70}")
+        print("Extracting REAL video features...")
+        print(f"{'='*70}")
+
+    real_features = load_or_compute_features(
+        videos=real_videos,
+        extractor=extractor,
+        config=config,
+        cache_path=config.cache_real_features,
+        cache_name="real_features")
+
+    if verbose:
+        print(f"\n{'='*70}")
+        print("Extracting GENERATED video features...")
+        print(f"{'='*70}")
+
+    gen_features = load_or_compute_features(videos=gen_videos,
+                                            extractor=extractor,
+                                            config=config,
+                                            cache_path=None,
+                                            cache_name="gen_features")
+
+    if verbose:
+        print(f"\nReal videos/clips: {len(real_features)}")
+        print(f"Generated videos/clips: {len(gen_features)}")
+        print(f"\n{'='*70}")
+        print("Computing statistics...")
+        print(f"{'='*70}")
+
+    mu_real, sigma_real = compute_statistics(real_features)
+    mu_gen, sigma_gen = compute_statistics(gen_features)
+
+    if verbose:
+        print(f"\n{'='*70}")
+        print("Computing Fréchet distance...")
+        print(f"{'='*70}")
+
+    fvd = compute_frechet_distance(mu_real, sigma_real, mu_gen, sigma_gen)
+
+    if verbose:
+        print(f"\n{'='*70}")
+        print(f"FVD Score: {fvd:.4f}")
+        print(f"Protocol: {config}")
+        print(f"{'='*70}\n")
+
+    results = {
+        'fvd': fvd,
+        'protocol': str(config),
+        'config': config.to_dict(),
+    }
+
+    return results

benchmarks/fvd/i3d_model.py

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

benchmarks/fvd/run_fvd.py

@@ -0,0 +1,34 @@
+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))
+
+
+def main() -> None:
+    # Get script directory
+    script_dir = Path(__file__).parent.resolve()
+
+    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"
+
+    results = compute_fvd_with_config(real_dir, gen_dir, cfg, verbose=True)
+    print(f"FVD = {results['fvd']:.2f}")
+
+
+if __name__ == '__main__':
+    main()

benchmarks/fvd/validate_fvd.py

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

benchmarks/fvd/video_utils.py

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

benchmarks/scripts/run.sh

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

benchmarks/scripts/setup_fvd.sh

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

csrc/attn/sliding_tile_attn/README.md

@@ -2,12 +2,12 @@
 # Attention Kernel Used in FastVideo
 
 ## Sliding Tile Attention (STA)
-We only support H100 for STA.
+We support H100 (via TK) and any other GPU (via triton) for STA.
 
 ### Installation
 ```bash
 pip install st_attn
-``` 
+```
 
 Install from source:
 
@@ -16,6 +16,14 @@ git submodule update --init --recursive
 python setup.py install
 ```
 
+If you want to skip the compilation of the TK kernel and only use the Triton version, try below:
+
+```bash
+SKIP_SM90_EXT=1 python setup.py install
+or
+SKIP_SM90_EXT=1 pip install --no-build-isolation .
+```
+
 If you encounter error during installation, try below:
 Install C++20 for ThunderKittens:
 ```bash
@@ -30,7 +38,7 @@ 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 PATH=${CUDA_HOME}/bin:${PATH}
 export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
 ```
 
@@ -43,7 +51,7 @@ 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. 
+# 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.
@@ -58,7 +66,6 @@ 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
@@ -67,13 +74,22 @@ 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. 
+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. 
+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.
 

csrc/attn/sliding_tile_attn/setup.py

@@ -51,21 +51,28 @@ for k in kernels:
         source_files.append(sources[k]['source_files'][target])
     cpp_flags.append(f'-DTK_COMPILE_{k.replace(" ", "_").upper()}')
 
+ext_modules = []
+
+if os.environ.get("SKIP_SM90_EXT", "0") != "1":
+    ext_modules.append(
+        CUDAExtension('st_attn_cuda',
+                  sources=source_files,
+                  extra_compile_args={
+                      'cxx': cpp_flags,
+                      'nvcc': cuda_flags
+                  },
+                  libraries=['cuda'])
+    )
+else:
+    print("ENV SKIP_SM90_EXT=1, skip st_attn_cuda compile")
+
 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'])
-      ],
+      ext_modules=ext_modules,
       cmdclass={'build_ext': BuildExtension},
       classifiers=[
           "Programming Language :: Python :: 3",

csrc/attn/sliding_tile_attn/st_attn/__init__.py

@@ -7,12 +7,17 @@ try:
 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'):
+try:
+    from st_attn.st_attn_triton import sliding_tile_attention_triton
+except ImportError:
+    sliding_tile_attention_triton = None
+
+def sliding_tile_attention_SM90(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, 
+        '18x48x80':3,
     }
     if has_text:
         assert q_all.shape[
@@ -46,4 +51,13 @@ def sliding_tile_attention(q_all, k_all, v_all, window_size, text_length, has_te
             _ = 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]
+    return hidden_states[:, :, :seq_length]
+
+def sliding_tile_attention(q_all, k_all, v_all, window_size, text_length, has_text=True, dit_seq_shape='30x48x80'):
+    major, minor = torch.cuda.get_device_capability(q_all.device)
+    if major == 9 and minor == 0 and sta_fwd is not None:
+        return sliding_tile_attention_SM90(q_all, k_all, v_all, window_size, text_length, has_text, dit_seq_shape)
+    elif sliding_tile_attention_triton is not None:
+        return sliding_tile_attention_triton(q_all, k_all, v_all, window_size, text_length, has_text, dit_seq_shape)
+    else:
+        raise ImportError("No suitable sliding tile attention implementation found.")

csrc/attn/sliding_tile_attn/st_attn/st_attn_triton.py

@@ -0,0 +1,327 @@
+import math
+import torch
+import triton
+import triton.language as tl
+
+def is_cuda():
+    return triton.runtime.driver.active.get_current_target().backend == "cuda"
+
+
+def is_hip():
+    target = triton.runtime.driver.active.get_current_target()
+    return target.backend == 'hip'
+
+def get_common_autotune_config():
+    configs = [
+        triton.Config({'BLOCK_Q': BLOCK_Q, 'BLOCK_KV': BLOCK_KV}, num_stages=s, num_warps=w) \
+        for BLOCK_Q in [32, 64, 128]\
+        for BLOCK_KV in [32, 64, 128]\
+        for s in [1, 2, 3, 4]\
+        for w in [4, 8]\
+    ]
+    return configs
+
+
+def get_cuda_autotune_config():
+    # cuda and hip can use differnt autotune configs
+    return get_common_autotune_config()
+
+
+def get_hip_autotune_config():
+    # cuda and hip can use differnt autotune configs
+    return get_common_autotune_config()
+
+
+def get_autotune_config():
+    if is_cuda():
+        return get_cuda_autotune_config()
+    else:
+        return get_hip_autotune_config()
+
+
+@triton.jit
+def clamp_int(value, min_val, max_val):
+    ret = tl.where(value > max_val, max_val, value)
+    ret = tl.where(ret < min_val, min_val, ret)
+    return ret
+
+
+@triton.jit
+def _attn_fwd_loop(
+    q, k, v, kv_mask, m, l, acc, sm_scale,
+    MASK_KV: tl.constexpr,
+):
+    scores = tl.dot(q, k.T) #[BLOCK_Q, BLOCK_KV]
+    scores = scores * sm_scale
+    if MASK_KV:
+        scores = tl.where(kv_mask[None, :], scores, -float('inf'))
+
+    current_m = tl.max(scores, axis=1)
+    new_m = tl.maximum(m, current_m)
+    exp_scores = tl.math.exp2(scores - new_m[:, None])
+    current_l = tl.sum(exp_scores, axis=1)
+
+    # Update L <- L * exp(M - M') + L1, M <- M'
+    alpha = tl.math.exp2(m - new_m)
+    l = l * alpha + current_l
+    m = new_m
+
+    # Update O <- O * exp(M - M') + P @ V
+    acc = (acc * alpha[:, None] + tl.dot(exp_scores.to(v.type.element_ty), v))
+
+    return m, l, acc
+
+
+@triton.autotune(
+    configs=get_autotune_config(),
+    key=['head_dim'],
+)
+@triton.jit
+def triton_sta_kernel(
+    Q, K, V, output,
+    batch_size: int, num_heads: int, seq_len: int,  head_dim: int,
+    img_seq_len: int,
+    text_length: int,
+    canvas_t: int, canvas_h: int, canvas_w: int,
+    kernel_t: int, kernel_h: int, kernel_w: int,
+    tile_t: int, tile_h: int, tile_w: int,
+    scale: float,
+    has_text: tl.constexpr,
+    text_q: tl.constexpr,
+    BLOCK_Q: tl.constexpr,
+    BLOCK_KV: tl.constexpr,
+    BLOCK_DIM: tl.constexpr,
+):
+    total_tile_size = tile_t * tile_h * tile_w
+    q_block_per_tile = (total_tile_size + BLOCK_Q - 1) // BLOCK_Q
+
+    batch_idx = tl.program_id(0)
+    head_idx = tl.program_id(1)
+    if text_q:
+        q_block_idx = tl.program_id(2)
+    else:
+        q_tile_flat = tl.program_id(2) // q_block_per_tile
+        q_block_idx = tl.program_id(2) % q_block_per_tile
+
+    m = tl.full((BLOCK_Q,), -float('inf'), dtype=tl.float32)
+    l = tl.zeros((BLOCK_Q,), dtype=tl.float32)
+    acc = tl.zeros((BLOCK_Q, BLOCK_DIM), dtype=tl.float32)
+
+    q_offset = (batch_idx * num_heads + head_idx) * seq_len * head_dim
+    if text_q:
+        q_base_idx = img_seq_len + q_block_idx * BLOCK_Q
+    else:
+        q_base_idx = q_tile_flat * total_tile_size + q_block_idx * BLOCK_Q
+
+    q_offset_in_tile = tl.arange(0, BLOCK_Q)
+    q_idx = q_base_idx + q_offset_in_tile
+    q_mask = (q_block_idx * BLOCK_Q + tl.arange(0, BLOCK_Q)) < total_tile_size
+
+    q = tl.load(
+        Q + q_offset + q_idx[:, None] * head_dim + tl.arange(0, BLOCK_DIM)[None, :],
+        mask=q_mask[:, None],
+        other=0.0
+    )  # [BLOCK_Q, BLOCK_DIM]
+
+    # Scale sm_scale by log_2(e) and use 2^x instead of exp
+    sm_scale = scale * 1.4426950408889634
+
+    num_tiles_t = canvas_t // tile_t
+    num_tiles_h = canvas_h // tile_h
+    num_tiles_w = canvas_w // tile_w
+    tiles_per_hw = num_tiles_h * num_tiles_w
+
+    if text_q:
+        kv_tile_start_t = 0
+        kv_tile_end_t = num_tiles_t
+
+        kv_tile_start_h = 0
+        kv_tile_end_h = num_tiles_h
+
+        kv_tile_start_w = 0
+        kv_tile_end_w = num_tiles_w
+
+    else:
+        q_tile_t = q_tile_flat // tiles_per_hw
+        remaining = q_tile_flat % tiles_per_hw
+        q_tile_h = remaining // num_tiles_w
+        q_tile_w = remaining % num_tiles_w
+
+        kernel_center_t = clamp_int(q_tile_t, kernel_t // 2, (num_tiles_t - 1) - kernel_t // 2)
+        kernel_center_h = clamp_int(q_tile_h, kernel_h // 2, (num_tiles_h - 1) - kernel_h // 2)
+        kernel_center_w = clamp_int(q_tile_w, kernel_w // 2, (num_tiles_w - 1) - kernel_w // 2)
+
+        kv_tile_start_t = kernel_center_t - kernel_t // 2
+        kv_tile_end_t = kernel_center_t + kernel_t // 2 + 1
+        kv_tile_end_t = tl.where(kv_tile_end_t > num_tiles_t, num_tiles_t, kv_tile_end_t)
+
+        kv_tile_start_h = kernel_center_h - kernel_h // 2
+        kv_tile_end_h = kernel_center_h + kernel_h // 2 + 1
+        kv_tile_end_h = tl.where(kv_tile_end_h > num_tiles_h, num_tiles_h, kv_tile_end_h)
+
+        kv_tile_start_w = kernel_center_w - kernel_w // 2
+        kv_tile_end_w = kernel_center_w + kernel_w // 2 + 1
+        kv_tile_end_w = tl.where(kv_tile_end_w > num_tiles_w, num_tiles_w, kv_tile_end_w)
+
+    # for kv_img
+    for kv_tile_t in tl.range(kv_tile_start_t, kv_tile_end_t):
+        for kv_tile_h in tl.range(kv_tile_start_h, kv_tile_end_h):
+            for kv_tile_w in tl.range(kv_tile_start_w, kv_tile_end_w):
+                kv_base_idx = (kv_tile_t * num_tiles_h * num_tiles_w + kv_tile_h * num_tiles_w + kv_tile_w) * total_tile_size
+
+                for kv_block_idx in tl.range(0, total_tile_size, BLOCK_KV):
+                    kv_offset_in_block = tl.arange(0, BLOCK_KV)
+                    kv_idx = kv_base_idx + kv_block_idx + kv_offset_in_block
+                    kv_mask = (kv_block_idx + tl.arange(0, BLOCK_KV)) < total_tile_size
+
+                    kv_offset = (batch_idx * num_heads + head_idx) * seq_len * head_dim
+
+                    k = tl.load(
+                        K + kv_offset + kv_idx[:, None] * head_dim + tl.arange(0, BLOCK_DIM)[None, :],
+                        mask=kv_mask[:, None],
+                        other=0.0
+                    )  # [BLOCK_KV, BLOCK_DIM]
+                    v = tl.load(
+                        V + kv_offset + kv_idx[:, None] * head_dim + tl.arange(0, BLOCK_DIM)[None, :],
+                        mask=kv_mask[:, None],
+                        other=0.0
+                    )  # [BLOCK_KV, BLOCK_DIM]
+
+                    m, l, acc = _attn_fwd_loop(q, k, v, kv_mask, m, l, acc, sm_scale, False)
+
+
+    # for kv_text
+    if has_text:
+        kv_base_idx = img_seq_len
+        for kv_block_idx in tl.range(0, total_tile_size, BLOCK_KV):
+            kv_offset_in_block = tl.arange(0, BLOCK_KV)
+            kv_idx = kv_base_idx + kv_block_idx + kv_offset_in_block
+            kv_mask = (kv_block_idx + tl.arange(0, BLOCK_KV)) < text_length
+
+            kv_offset = (batch_idx * num_heads + head_idx) * seq_len * head_dim
+
+            k = tl.load(
+                K + kv_offset + kv_idx[:, None] * head_dim + tl.arange(0, BLOCK_DIM)[None, :],
+                mask=kv_mask[:, None],
+                other=0.0
+            )  # [BLOCK_KV, BLOCK_DIM]
+            v = tl.load(
+                V + kv_offset + kv_idx[:, None] * head_dim + tl.arange(0, BLOCK_DIM)[None, :],
+                mask=kv_mask[:, None],
+                other=0.0
+            )  # [BLOCK_KV, BLOCK_DIM]
+
+            m, l, acc = _attn_fwd_loop(q, k, v, kv_mask, m, l, acc, sm_scale, True)
+
+
+    output_acc = acc / l[:, None]
+    tl.store(
+        output + q_offset + q_idx[:, None] * head_dim + tl.arange(0, BLOCK_DIM)[None, :],
+        output_acc,
+        mask=q_mask[:, None]
+    ) # [BLOCK_Q, BLOCK_DIM]
+
+
+def sliding_tile_attention_triton(
+    q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+    window_size, text_length: int,
+    has_text=True, dit_seq_shape='30x48x80') -> torch.Tensor:
+    seq_length = q.shape[2]
+    if has_text:
+        assert q.shape[2] >= 115200 and q.shape[2] <= 115456, f"Unsupported {dit_seq_shape}, current shape is {q.shape}, only support '30x48x80' for HunyuanVideo"
+        target_size = math.ceil(seq_length / 384) * 384
+        pad_size = target_size - seq_length
+        if pad_size > 0:
+            q = torch.cat([q, q[:, :, -pad_size:]], dim=2)
+            k = torch.cat([k, k[:, :, -pad_size:]], dim=2)
+            v = torch.cat([v, v[:, :, -pad_size:]], dim=2)
+    else:
+        if dit_seq_shape == '36x48x48': # Stepvideo
+            assert q.shape[2] == 82944
+        elif dit_seq_shape == '18x48x80': # Wan
+            assert q.shape[2] == 69120
+        else:
+            raise ValueError(f"Unsupported {dit_seq_shape}, current shape is {q.shape}, only support '36x48x48' for Stepvideo and '18x48x80' for Wan")
+    assert q.shape[1] == len(window_size), "Number of heads must match the number of window sizes"
+
+    batch_size, num_heads, seq_len, head_dim = q.shape
+    if dit_seq_shape == '30x48x80': # Hunyuan
+        canvas_t, canvas_h, canvas_w = 30, 48, 80
+        tile_t, tile_h, tile_w = 6, 8, 8
+    elif dit_seq_shape == '36x48x48': # Stepvideo
+        canvas_t, canvas_h, canvas_w = 36, 48, 48
+        tile_t, tile_h, tile_w = 6, 8, 8
+    elif dit_seq_shape == '18x48x80': # Wan
+        canvas_t, canvas_h, canvas_w = 18, 48, 80
+        tile_t, tile_h, tile_w = 6, 8, 8
+
+    img_seq_len = canvas_t * canvas_h * canvas_w
+
+    num_tiles_t = canvas_t // tile_t
+    num_tiles_h = canvas_h // tile_h
+    num_tiles_w = canvas_w // tile_w
+    num_tiles = num_tiles_t * num_tiles_h * num_tiles_w
+
+    total_tile_size = tile_t * tile_h * tile_w
+
+    # BLOCK_Q=128
+    # BLOCK_KV=128
+    BLOCK_DIM = head_dim
+
+    output = torch.empty_like(q)
+
+    # for q_img
+    # kernel_size maybe different for different head
+    # This for loop is ugly. but it is actually quite efficient. The sequence dimension alone can already oversubscribe SMs
+    for head_index, (kernel_t, kernel_h, kernel_w) in enumerate(window_size):
+        for batch in range(batch_size):
+            q_head, k_head, v_head, o_head = (q[batch:batch + 1, head_index:head_index + 1],
+                                              k[batch:batch + 1, head_index:head_index + 1],
+                                              v[batch:batch + 1, head_index:head_index + 1],
+                                              output[batch:batch + 1, head_index:head_index + 1])
+
+            # triton_sta_kernel[(1, 1, num_tiles * triton.cdiv(total_tile_size, BLOCK_Q))](
+            grid = lambda META: (1, 1, num_tiles * triton.cdiv(total_tile_size, META['BLOCK_Q']))
+            triton_sta_kernel[grid](
+                q_head, k_head, v_head, o_head,
+                1, 1, seq_len, head_dim,
+                img_seq_len,
+                text_length,
+                canvas_t, canvas_h, canvas_w,
+                kernel_t, kernel_h, kernel_w,
+                tile_t, tile_h, tile_w,
+                scale=1.0 / (head_dim ** 0.5),
+                has_text=has_text,
+                text_q=False,
+                # BLOCK_Q=BLOCK_Q,
+                # BLOCK_KV=BLOCK_KV,
+                BLOCK_DIM=BLOCK_DIM,
+            )
+
+    # for q_text
+    # kernel_t, kernel_h, kernel_w is not used, set to (3, 3, 3)
+    if has_text:
+        # triton_sta_kernel[(batch_size, num_heads, triton.cdiv(total_tile_size, BLOCK_Q))](
+        grid = lambda META: (batch_size, num_heads, triton.cdiv(total_tile_size, META['BLOCK_Q']))
+        triton_sta_kernel[grid](
+            q, k, v, output,
+            batch_size, num_heads, seq_len, head_dim,
+            img_seq_len,
+            text_length,
+            canvas_t, canvas_h, canvas_w,
+            3, 3, 3,
+            #kernel_t, kernel_h, kernel_w,
+            tile_t, tile_h, tile_w,
+            scale=1.0 / (head_dim ** 0.5),
+            has_text=has_text,
+            text_q=True,
+            # BLOCK_Q=BLOCK_Q,
+            # BLOCK_KV=BLOCK_KV,
+            BLOCK_DIM=BLOCK_DIM,
+        )
+
+    if has_text:
+        if pad_size > 0:
+            output = output[:, :, :seq_length]
+
+    return output

csrc/attn/tests/test_vsa.py

@@ -34,16 +34,16 @@ def pytorch_test(Q, K, V, block_sparse_mask, dO):
     )
 
 
-def block_sparse_kernel_test(Q, K, V, block_sparse_mask, variable_block_sizes, non_pad_index, dO):
+def block_sparse_kernel_test(Q, K, V, block_sparse_mask, variable_block_sizes, q_non_pad_index, kv_non_pad_index, q_num_blocks, kv_num_blocks, 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)
+    q_padded = vsa_pad(Q, q_non_pad_index, q_num_blocks, BLOCK_M)
+    k_padded = vsa_pad(K, kv_non_pad_index, kv_num_blocks, BLOCK_M)
+    v_padded = vsa_pad(V, kv_non_pad_index, kv_num_blocks, BLOCK_M)
     output, _= block_sparse_attn(q_padded, k_padded, v_padded, block_sparse_mask, variable_block_sizes)
-    output = output[:, :, non_pad_index, :]
+    output = output[:, :, q_non_pad_index, :]
     output.backward(dO)
     return output, Q.grad, K.grad, V.grad
 
@@ -64,7 +64,7 @@ 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"):
+def generate_variable_block_sizes(num_blocks, min_size=16, max_size=64, device="cuda"):
     return torch.randint(min_size, max_size + 1, (num_blocks,), device=device, dtype=torch.int32)
 
 
@@ -86,19 +86,21 @@ def check_correctness(h, d, num_blocks, k,  num_iterations=20, error_mode='all')
     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)
+    block_mask = generate_block_sparse_mask_for_function(h, num_blocks, num_blocks, k, device)
+    full_mask = create_full_mask_from_block_mask(block_mask, variable_block_sizes, 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)
+        
+        # print(Q.shape, K.shape, V.shape, dO.shape)
 
         # 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)
+        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, non_pad_index, num_blocks, num_blocks, 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
@@ -118,6 +120,60 @@ def check_correctness(h, d, num_blocks, k,  num_iterations=20, error_mode='all')
 
     return results
 
+def check_correctness_qkdiff(h, d, num_q_blocks, num_kv_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"
+    q_variable_block_sizes = generate_variable_block_sizes(num_q_blocks, device=device)
+    kv_variable_block_sizes = generate_variable_block_sizes(num_kv_blocks, device=device)
+    
+    S_q = int(q_variable_block_sizes.sum().item())
+    S_kv = int(kv_variable_block_sizes.sum().item())
+    
+    q_non_pad_index = get_non_pad_index(q_variable_block_sizes, num_q_blocks, BLOCK_M)
+    kv_non_pad_index = get_non_pad_index(kv_variable_block_sizes, num_kv_blocks, BLOCK_M)
+    
+    block_mask = generate_block_sparse_mask_for_function(h, num_q_blocks, num_kv_blocks, k, device)
+    full_mask = create_full_mask_from_block_mask(block_mask, q_variable_block_sizes, kv_variable_block_sizes, device)
+    
+    for _ in range(num_iterations):
+        Q = generate_tensor((1, h, S_q, d),  torch.bfloat16, device)
+        K = generate_tensor((1, h, S_kv, d), torch.bfloat16, device)
+        V = generate_tensor((1, h, S_kv, d),  torch.bfloat16, device)
+        dO = generate_tensor((1, h, S_q, d), torch.bfloat16, device)
+        
+        # print(Q.shape, K.shape, V.shape, dO.shape)
+
+        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), kv_variable_block_sizes, q_non_pad_index, kv_non_pad_index, num_q_blocks, num_kv_blocks, 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_q = h * S_q * d * num_iterations
+    total_elements_kv = h * S_kv * d * num_iterations
+    
+    for name, data in results.items():
+        total_elements = total_elements_q if name in ['gQ', 'gO'] else total_elements_kv
+        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"},
@@ -147,10 +203,43 @@ def generate_error_graphs(h, d, error_mode='all'):
 
     print("-" * 150)
 
+def generate_error_graphs_qkdiff(h, d, error_mode='all'):
+    test_configs = [
+        {"num_q_blocks": 16, "num_kv_blocks": 32, "k": 2, "description": "Small Q, Med KV"},
+        {"num_q_blocks": 32, "num_kv_blocks": 16, "k": 4, "description": "Med Q, Small KV"},
+        {"num_q_blocks": 53, "num_kv_blocks": 32, "k": 6, "description": "Large Q, Med KV"},
+        {"num_q_blocks": 16, "num_kv_blocks": 48, "k": 2, "description": "Small Q, Large KV"},
+        {"num_q_blocks": 48, "num_kv_blocks": 16, "k": 2, "description": "Large Q, Small KV"},
+    ]
+    
+    print(f"\nError Analysis (QK Diff) for h={h}, d={d}, mode={error_mode}")
+    print("=" * 150)
+    print(f"{'Config':<20} {'Q Blks':<8} {'KV Blks':<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_q_blocks = config["num_q_blocks"]
+        num_kv_blocks = config["num_kv_blocks"]
+        k = config["k"]
+        description = config["description"]
+        results = check_correctness_qkdiff(h, d, num_q_blocks, num_kv_blocks, k, error_mode=error_mode)
+        print(f"{description:<20} {num_q_blocks:<8} {num_kv_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.")
+        generate_error_graphs_qkdiff(h, d, error_mode=mode)
+    print("\nAnalysis completed for all modes.")

csrc/attn/tests/test_vsa_forward.py

@@ -0,0 +1,236 @@
+import os
+import sys
+from typing import Tuple
+
+import torch
+
+# Make sure we can import from the project root (`vsa`, `tests.utils`, etc.)
+CURRENT_DIR = os.path.dirname(os.path.abspath(__file__))
+PROJECT_ROOT = os.path.dirname(CURRENT_DIR)
+if PROJECT_ROOT not in sys.path:
+    sys.path.append(PROJECT_ROOT)
+if CURRENT_DIR not in sys.path:
+    sys.path.append(CURRENT_DIR)
+
+from tests.utils import (
+    generate_block_sparse_mask_for_function,
+    create_full_mask_from_block_mask,
+)
+from vsa import block_sparse_attn, BLOCK_M
+import test_vsa as ref  # reuse helper functions from backward test
+
+
+def pytorch_forward(
+    Q: torch.Tensor,
+    K: torch.Tensor,
+    V: torch.Tensor,
+    block_sparse_mask: torch.Tensor,
+) -> torch.Tensor:
+    """
+    Dense PyTorch reference forward:
+    - Q: [1, h, S_q, d]
+    - K,V: [1, h, S_kv, d]
+    - block_sparse_mask: [h, S_q, S_kv] bool
+    """
+    q = Q.clone().float()
+    k = K.clone().float()
+    v = V.clone().float()
+
+    attn = torch.matmul(q, k.transpose(-2, -1))  # [1, h, S_q, S_kv]
+    attn = attn / (q.size(-1) ** 0.5)
+    attn = attn.masked_fill(~block_sparse_mask.unsqueeze(0), float("-inf"))
+    attn = torch.nn.functional.softmax(attn, dim=-1)
+    out = torch.matmul(attn, v)  # [1, h, S_q, d]
+    return out.to(torch.bfloat16)
+
+
+def block_sparse_forward_test(
+    Q: torch.Tensor,
+    K: torch.Tensor,
+    V: torch.Tensor,
+    block_sparse_mask: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+    q_non_pad_index: torch.Tensor,
+    kv_non_pad_index: torch.Tensor,
+    q_num_blocks: int,
+    kv_num_blocks: int,
+) -> torch.Tensor:
+    """
+    Forward-only wrapper around `block_sparse_attn`, mirroring `block_sparse_kernel_test`
+    but without any backward / grad logic.
+    """
+    Q = Q.detach()
+    K = K.detach()
+    V = V.detach()
+
+    q_padded = ref.vsa_pad(Q, q_non_pad_index, q_num_blocks, BLOCK_M)
+    k_padded = ref.vsa_pad(K, kv_non_pad_index, kv_num_blocks, BLOCK_M)
+    v_padded = ref.vsa_pad(V, kv_non_pad_index, kv_num_blocks, BLOCK_M)
+
+    out_padded, _ = block_sparse_attn(
+        q_padded, k_padded, v_padded, block_sparse_mask, variable_block_sizes
+    )
+    # Remove padding on the query side
+    out = out_padded[:, :, q_non_pad_index, :]
+    return out
+
+
+def run_forward_equal_qk(
+    h: int = 16,
+    d: int = 128,
+    num_blocks: int = 16,
+    k: int = 2,
+    num_iterations: int = 5,
+) -> Tuple[float, float]:
+    """
+    Forward-only correctness test for the case S_q == S_kv.
+    Mirrors `check_correctness` but only compares forward outputs.
+    """
+    assert torch.cuda.is_available(), "VSA kernels require CUDA"
+    device = "cuda"
+
+    variable_block_sizes = ref.generate_variable_block_sizes(
+        num_blocks, device=device
+    )
+    S = int(variable_block_sizes.sum().item())
+    non_pad_index = ref.get_non_pad_index(
+        variable_block_sizes, num_blocks, BLOCK_M
+    )
+
+    block_mask = generate_block_sparse_mask_for_function(
+        h, num_blocks, num_blocks, k, device
+    )
+    full_mask = create_full_mask_from_block_mask(
+        block_mask, variable_block_sizes, variable_block_sizes, device
+    )
+    print(f"[qkequal] h: {h}, d: {d}, num_blocks: {num_blocks}, k: {k}")
+    print(f"[qkequal] variable_block_sizes: {variable_block_sizes}, non_pad_index: {non_pad_index.shape}, block_mask: {block_mask.shape}, full_mask: {full_mask.shape}")
+    sum_diff = 0.0
+    sum_abs = 0.0
+    max_rel_diff = 0.0
+
+    for i in range(num_iterations):
+        Q = ref.generate_tensor((1, h, S, d), torch.bfloat16, device)
+        K = ref.generate_tensor((1, h, S, d), torch.bfloat16, device)
+        V = ref.generate_tensor((1, h, S, d), torch.bfloat16, device)
+
+        if i == 0: print(f"[qkequal] Q: {Q.shape}, K: {K.shape}, V: {V.shape}, full_mask: {full_mask.shape}")
+        if i == 0: print(f"[qkequal] block_mask: {block_mask.shape}")
+
+        pt_o = pytorch_forward(Q, K, V, full_mask)
+        bs_o = block_sparse_forward_test(
+            Q,
+            K,
+            V,
+            block_mask.unsqueeze(0),
+            variable_block_sizes,
+            non_pad_index,
+            non_pad_index,
+            num_blocks,
+            num_blocks,
+        )
+
+        diff = (pt_o - bs_o).abs()
+        sum_diff += diff.sum().item()
+        sum_abs += pt_o.abs().sum().item()
+        rel_max = diff.max() / (pt_o.abs().mean() + 1e-6)
+        max_rel_diff = max(max_rel_diff, rel_max.item())
+
+    total_elems = h * S * d * num_iterations
+    avg_abs_err = sum_diff / total_elems
+    return avg_abs_err, max_rel_diff
+
+
+def run_forward_qk_diff(
+    h: int = 16,
+    d: int = 128,
+    num_q_blocks: int = 16,
+    num_kv_blocks: int = 32,
+    k: int = 2,
+    num_iterations: int = 5,
+) -> Tuple[float, float]:
+    """
+    Forward-only correctness test for the case S_q != S_kv.
+
+    NOTE:
+      - The Triton backend supports different Q/KV logical lengths via padding.
+      - The SM90 (H100) CUDA backend currently assumes the same number of blocks
+        for Q and KV, so we skip this test there.
+    """
+    assert torch.cuda.is_available(), "VSA kernels require CUDA"
+
+    device = "cuda"
+
+    q_variable_block_sizes = ref.generate_variable_block_sizes(
+        num_q_blocks, device=device
+    )
+    kv_variable_block_sizes = ref.generate_variable_block_sizes(
+        num_kv_blocks, device=device
+    )
+
+    S_q = int(q_variable_block_sizes.sum().item())
+    S_kv = int(kv_variable_block_sizes.sum().item())
+
+    q_non_pad_index = ref.get_non_pad_index(
+        q_variable_block_sizes, num_q_blocks, BLOCK_M
+    )
+    kv_non_pad_index = ref.get_non_pad_index(
+        kv_variable_block_sizes, num_kv_blocks, BLOCK_M
+    )
+
+    block_mask = generate_block_sparse_mask_for_function(
+        h, num_q_blocks, num_kv_blocks, k, device
+    )
+    full_mask = create_full_mask_from_block_mask(
+        block_mask, q_variable_block_sizes, kv_variable_block_sizes, device
+    )
+
+    sum_diff = 0.0
+    sum_abs = 0.0
+    max_rel_diff = 0.0
+
+    for _ in range(num_iterations):
+        Q = ref.generate_tensor((1, h, S_q, d), torch.bfloat16, device)
+        K = ref.generate_tensor((1, h, S_kv, d), torch.bfloat16, device)
+        V = ref.generate_tensor((1, h, S_kv, d), torch.bfloat16, device)
+
+        pt_o = pytorch_forward(Q, K, V, full_mask)
+        bs_o = block_sparse_forward_test(
+            Q,
+            K,
+            V,
+            block_mask.unsqueeze(0),
+            kv_variable_block_sizes,
+            q_non_pad_index,
+            kv_non_pad_index,
+            num_q_blocks,
+            num_kv_blocks,
+        )
+
+        diff = (pt_o - bs_o).abs()
+        sum_diff += diff.sum().item()
+        sum_abs += pt_o.abs().sum().item()
+        rel_max = diff.max() / (pt_o.abs().mean() + 1e-6)
+        max_rel_diff = max(max_rel_diff, rel_max.item())
+
+    total_elems = h * S_q * d * num_iterations
+    avg_abs_err = sum_diff / total_elems
+    return avg_abs_err, max_rel_diff
+
+
+if __name__ == "__main__":
+    h, d = 16, 128
+    print("Forward Block Sparse Attention Check (QK Equal)")
+    print("=" * 80)
+    avg_err_eq, max_rel_eq = run_forward_equal_qk(h, d, num_blocks=32, k=2)
+    print(f"QK equal: avg |ΔO| = {avg_err_eq:.6e}, max rel ΔO = {max_rel_eq:.6e}")
+
+    print("\nForward Block Sparse Attention Check (QK Different)")
+    print("=" * 80)
+    avg_err_diff, max_rel_diff = run_forward_qk_diff(
+        h, d, num_q_blocks=32, num_kv_blocks=48, k=2
+    )
+    print(
+        f"QK diff:  avg |ΔO| = {avg_err_diff:.6e}, max rel ΔO = {max_rel_diff:.6e}"
+    )
+

csrc/attn/tests/utils.py

@@ -1,54 +1,60 @@
 import torch
 
-def generate_block_sparse_mask_for_function(h, num_blocks, k, device="cuda"):
+def generate_block_sparse_mask_for_function(h, num_q_blocks, num_kv_blocks, k, device="cuda"):
     """
-    Generate block sparse mask of shape [h, num_blocks, num_blocks].
+    Generate block sparse mask of shape [h, num_q_blocks, num_kv_blocks].
     
     Args:
         h: number of heads
-        num_blocks: number of blocks
+        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:
-        block_sparse_mask: [h, num_blocks, num_blocks] bool tensor
+        block_sparse_mask: [h, num_q_blocks, num_kv_blocks] bool tensor
     """
-    k = min(k, num_blocks)
-    scores = torch.rand(h, num_blocks, num_blocks, device=device)
+    k = min(k, num_kv_blocks)
+    scores = torch.rand(h, num_q_blocks, num_kv_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 = torch.zeros(h, num_q_blocks, num_kv_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"):
+def create_full_mask_from_block_mask(block_sparse_mask, q_variable_block_sizes, 
+                                     kv_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
+        block_sparse_mask: [h, num_q_blocks, num_kv_blocks] bool tensor
+        q_variable_block_sizes: [num_q_blocks] tensor
+        kv_variable_block_sizes: [num_kv_blocks] tensor
         device: device to create tensors on
 
     Returns:
-        full_mask: [h, S, S] bool tensor where S = total sequence length
+        full_mask: [h, S_q, S_kv] 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]])
+    h, num_q_blocks, num_kv_blocks = block_sparse_mask.shape
+    total_q_seq_len = q_variable_block_sizes.sum().item()
+    total_kv_seq_len = kv_variable_block_sizes.sum().item()
+    
+    q_cumsum = torch.cat([torch.tensor([0], device=device), q_variable_block_sizes.cumsum(dim=0)[:-1]])
+    kv_cumsum = torch.cat([torch.tensor([0], device=device), kv_variable_block_sizes.cumsum(dim=0)[:-1]])
 
-    full_mask = torch.zeros(h, total_seq_len, total_seq_len, dtype=torch.bool, device=device)
+    full_mask = torch.zeros(h, total_q_seq_len, total_kv_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 q_block in range(num_q_blocks):
+            q_start = q_cumsum[q_block]
+            q_end = q_start + q_variable_block_sizes[q_block]
 
-            for kv_block in range(num_blocks):
+            for kv_block in range(num_kv_blocks):
                 if block_sparse_mask[head, q_block, kv_block]:
-                    kv_start = cumsum[kv_block]
-                    kv_end = kv_start + variable_block_sizes[kv_block]
+                    kv_start = kv_cumsum[kv_block]
+                    kv_end = kv_start + kv_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/vsa/block_sparse_attn_triton.py

@@ -250,8 +250,9 @@ def _attn_bwd_dq(dq, q, K, V,  #
     
     
     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
-        block_size = tl.load(variable_block_sizes + blk_idx//2) - (blk_idx%2) * step_n
+        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)

csrc/attn/video_sparse_attn/vsa/block_sparse_h100.cu

@@ -672,23 +672,32 @@ block_sparse_attention_forward(
     torch::Tensor v, 
     torch::Tensor q2k_block_sparse_index, 
     torch::Tensor q2k_block_sparse_num, 
-    torch::Tensor block_size
+    torch::Tensor kv_block_size
 )
 {
     CHECK_INPUT(q);
     CHECK_INPUT(k);
     CHECK_INPUT(v);
 
+    // q shape: (batch, qo_heads, q_seq_len, head_dim)
+    // k shape: (batch, kv_heads, kv_seq_len, head_dim)
+    // v shape: (batch, kv_heads, kv_seq_len, head_dim)
+    // q2k_block_sparse_index shape: (batch, qo_heads, num_q_blocks, max_kv_blocks_per_q)
+    // q2k_block_sparse_num shape: (batch, qo_heads, num_q_blocks)
+    // kv_block_size shape: (num_kv_blocks) This does not need other dimensions because across all batch/heads the padding is the same.
+    
     auto batch    = q.size(0);
-    auto seq_len  = q.size(2); 
+    auto q_seq_len  = q.size(2);
+    auto kv_seq_len = k.size(2);
     auto head_dim = q.size(3); 
     auto qo_heads = q.size(1);
     auto kv_heads = k.size(1);
     auto max_kv_blocks_per_q = q2k_block_sparse_index.size(3);
-    auto num_q_blocks = block_size.size(0);
+    auto num_q_blocks = q2k_block_sparse_index.size(2);
+    auto num_kv_blocks = kv_block_size.size(0);
     TORCH_CHECK(batch==1, "Batch size dim will be removed in the future, please set batch to 1");
-    TORCH_CHECK(num_q_blocks * 64 == seq_len, "This kernel supports variable block size, but it assumes the input sequence is properly padded.");
-    TORCH_CHECK(num_q_blocks == q2k_block_sparse_index.size(2), "Number of Q blocks does not match between q2k_block_sparse_index and block_size");
+    TORCH_CHECK(num_q_blocks * BLOCK_M == q_seq_len, "This kernel supports variable q block size, but it assumes the input sequence is properly padded.");
+    TORCH_CHECK(num_kv_blocks * BLOCK_M == kv_seq_len, "This kernel supports variable kv block size, but it assumes the input sequence is properly padded.");
     // check to see that these dimensions match for all inputs
     TORCH_CHECK(q.size(0) == batch, "Q batch dimension - idx 0 - must match for all inputs");
     TORCH_CHECK(k.size(0) == batch, "K batch dimension - idx 0 - must match for all inputs");
@@ -696,11 +705,8 @@ block_sparse_attention_forward(
     TORCH_CHECK(q2k_block_sparse_index.size(0) == batch, "q2k_block_sparse_index batch dimension - idx 0 - must match for all inputs");
     TORCH_CHECK(q2k_block_sparse_num.size(0) == batch, "q2k_block_sparse_num batch dimension - idx 0 - must match for all inputs");
 
-    TORCH_CHECK(q.size(2) == seq_len, "Q sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(k.size(2) == seq_len, "K sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(v.size(2) == seq_len, "V sequence length dimension - idx 2 - must match for all inputs");
-    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(v.size(2) == kv_seq_len, "V sequence length dimension - idx 2 - must match K inputs");
+    TORCH_CHECK(q2k_block_sparse_num.size(2) == num_q_blocks, "q2k_block_sparse_num idx 2 - must match num_q_blocks");
 
 
     TORCH_CHECK(q.size(3) == head_dim, "Q head dimension - idx 3 - must match for all non-vector inputs");
@@ -727,12 +733,12 @@ block_sparse_attention_forward(
     // for the returned outputs
     torch::Tensor o     = torch::empty({static_cast<const uint>(batch), 
                                         static_cast<const uint>(qo_heads), 
-                                        static_cast<const uint>(seq_len), 
+                                        static_cast<const uint>(q_seq_len), 
                                         static_cast<const uint>(head_dim)}, v.options());
     
     torch::Tensor l_vec = torch::empty({static_cast<const uint>(batch), 
                                         static_cast<const uint>(qo_heads), 
-                                        static_cast<const uint>(seq_len), 
+                                        static_cast<const uint>(q_seq_len), 
                                         static_cast<const uint>(1)}, 
                                         torch::TensorOptions().dtype(torch::kFloat).device(q.device()).memory_format(at::MemoryFormat::Contiguous));
         
@@ -762,11 +768,11 @@ block_sparse_attention_forward(
 
         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};
+        q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_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>(q_seq_len)};
+        o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
 
         globals g{
             qg_arg, 
@@ -774,17 +780,17 @@ block_sparse_attention_forward(
             vg_arg, 
             lg_arg, 
             og_arg, 
-            static_cast<int>(seq_len), 
+            static_cast<int>(q_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())
+            reinterpret_cast<int32_t*>(kv_block_size.data_ptr())
         };
 
         constexpr int mem_size = 54000;
 
-        dim3 grid(seq_len/(64), qo_heads, batch);
+        dim3 grid(q_seq_len/(BLOCK_M), qo_heads, batch);
 
         cudaFuncSetAttribute(
             fwd_attend_ker<64>,
@@ -813,11 +819,11 @@ block_sparse_attention_forward(
 
         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};
+        q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_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>(q_seq_len)};
+        o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
 
         globals g{
             qg_arg, 
@@ -825,17 +831,17 @@ block_sparse_attention_forward(
             vg_arg, 
             lg_arg, 
             og_arg, 
-            static_cast<int>(seq_len), 
+            static_cast<int>(q_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())
+            reinterpret_cast<int32_t*>(kv_block_size.data_ptr())
         };
 
         constexpr int mem_size = 54000;
 
-        dim3 grid(seq_len/(64), qo_heads, batch);
+        dim3 grid(q_seq_len/(BLOCK_M), qo_heads, batch);
 
         cudaFuncSetAttribute(
             fwd_attend_ker<128>,
@@ -862,7 +868,7 @@ block_sparse_attention_backward(torch::Tensor q,
                    torch::Tensor og,
                    torch::Tensor k2q_block_sparse_index,
                    torch::Tensor k2q_block_sparse_num,
-                   torch::Tensor block_size)
+                   torch::Tensor kv_block_size)
 {
     CHECK_INPUT(q);
     CHECK_INPUT(k);
@@ -871,11 +877,23 @@ block_sparse_attention_backward(torch::Tensor q,
     CHECK_INPUT(o);
     CHECK_INPUT(og);
 
+    // q: [batch, qo_heads, q_seq_len, head_dim]
+    // k: [batch, kv_heads, kv_seq_len, head_dim]
+    // v: [batch, kv_heads, kv_seq_len, head_dim]
+    // o: [batch, qo_heads, q_seq_len, head_dim]
+    // l_vec: [batch, qo_heads, q_seq_len, 1]
+    // og: [batch, qo_heads, q_seq_len, head_dim]
+    // k2q_block_sparse_index: [batch, kv_heads, num_kv_blocks, max_num_q_blocks]
+    // k2q_block_sparse_num: [batch, kv_heads, num_kv_blocks]
+    // kv_block_size: [num_kv_blocks]
+
     auto batch    = q.size(0);
-    auto seq_len  = q.size(2);
+    auto q_seq_len  = q.size(2);
+    auto kv_seq_len = k.size(2);
     auto head_dim = q.size(3);
     auto max_q_blocks_per_kv = k2q_block_sparse_index.size(3);
-    TORCH_CHECK(k2q_block_sparse_index.size(2) == block_size.size(0), "k2q_block_sparse_index.size(2) must match block_size.size(0)");
+    auto num_kv_blocks = kv_block_size.size(0);
+    TORCH_CHECK(k2q_block_sparse_index.size(2) == num_kv_blocks, "k2q_block_sparse_index.size(2) must match num_kv_blocks (kv_block_size.size(0))");
     // check to see that these dimensions match for all inputs
     TORCH_CHECK(q.size(0)     == batch, "Q  batch dimension - idx 0 - must match for all inputs");
     TORCH_CHECK(k.size(0)     == batch, "K  batch dimension - idx 0 - must match for all inputs");
@@ -886,23 +904,18 @@ block_sparse_attention_backward(torch::Tensor q,
     TORCH_CHECK(k2q_block_sparse_index.size(0) == batch, "k2q_block_sparse_index batch dimension - idx 0 - must match for all inputs");
     TORCH_CHECK(k2q_block_sparse_num.size(0) == batch, "k2q_block_sparse_num batch dimension - idx 0 - must match for all inputs");
 
-
-    TORCH_CHECK(q.size(2)     == seq_len, "Q  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(k.size(2)     == seq_len, "K  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(v.size(2)     == seq_len, "V  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(l_vec.size(2) == seq_len, "L  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(o.size(2)     == seq_len, "O  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(og.size(2)    == seq_len, "OG sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(k2q_block_sparse_index.size(2) == seq_len / BLOCK_N, "k2q_block_sparse_index idx 2 - must match seq_len / BLOCK_N");
-    TORCH_CHECK(k2q_block_sparse_num.size(2) == seq_len / BLOCK_N, "k2q_block_sparse_num idx 2 - must match seq_len / BLOCK_N");
+    TORCH_CHECK(v.size(2)     == kv_seq_len, "V  sequence length dimension - idx 2 - must match K sequence length");
+    TORCH_CHECK(l_vec.size(2) == q_seq_len, "L  sequence length dimension - idx 2 - must match Q sequence length");
+    TORCH_CHECK(o.size(2)     == q_seq_len, "O  sequence length dimension - idx 2 - must match Q sequence length");
+    TORCH_CHECK(og.size(2)    == q_seq_len, "OG sequence length dimension - idx 2 - must match Q sequence length");
+    TORCH_CHECK(k2q_block_sparse_index.size(2) == num_kv_blocks, "k2q_block_sparse_index idx 2 - must match num_kv_blocks (kv_block_size.size(0))");
+    TORCH_CHECK(k2q_block_sparse_num.size(2) == num_kv_blocks, "k2q_block_sparse_num idx 2 - must match num_kv_blocks (kv_block_size.size(0))");
 
     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");
     TORCH_CHECK(o.size(3)  == head_dim, "O  head dimension - idx 3 - must match for all non-vector inputs");
     TORCH_CHECK(og.size(3) == head_dim, "OG head dimension - idx 3 - must match for all non-vector inputs");
-    
-
 
     auto qo_heads = q.size(1);
     auto kv_heads = k.size(1);
@@ -929,20 +942,20 @@ block_sparse_attention_backward(torch::Tensor q,
 
     torch::Tensor qg = torch::zeros({static_cast<const uint>(batch), 
                                      static_cast<const uint>(qo_heads), 
-                                     static_cast<const uint>(seq_len), 
+                                     static_cast<const uint>(q_seq_len), 
                                      static_cast<const uint>(head_dim)},   l_vec.options());
     torch::Tensor kg = torch::zeros({static_cast<const uint>(batch), 
                                      static_cast<const uint>(kv_heads), 
-                                     static_cast<const uint>(seq_len), 
+                                     static_cast<const uint>(kv_seq_len), 
                                      static_cast<const uint>(head_dim)},   l_vec.options());
     torch::Tensor vg = torch::zeros({static_cast<const uint>(batch), 
                                      static_cast<const uint>(kv_heads), 
-                                     static_cast<const uint>(seq_len), 
+                                     static_cast<const uint>(kv_seq_len), 
                                      static_cast<const uint>(head_dim)},   l_vec.options());
     
     torch::Tensor d_vec = torch::empty({static_cast<const uint>(batch), 
                                         static_cast<const uint>(qo_heads), 
-                                        static_cast<const uint>(seq_len), 
+                                        static_cast<const uint>(q_seq_len), 
                                         static_cast<const uint>(1)},       l_vec.options());
 
     float*         qg_ptr = qg.data_ptr<float>();
@@ -971,7 +984,7 @@ block_sparse_attention_backward(torch::Tensor q,
     //  cudaStreamSynchronize(stream);
 
     // TORCH_CHECK(seq_len % (4*kittens::TILE_DIM*4) == 0, "sequence length must be divisible by 256");
-    dim3 grid_bwd(seq_len/(PREP_NUM_WARPS*kittens::TILE_ROW_DIM<bf16>*4), qo_heads, batch);
+    dim3 grid_bwd(q_seq_len/(PREP_NUM_WARPS*kittens::TILE_ROW_DIM<bf16>*4), qo_heads, batch);
 
     if (head_dim == 64)  {
         using og_tile = st_bf<4*16, 64>;
@@ -984,9 +997,9 @@ block_sparse_attention_backward(torch::Tensor q,
 
         using bwd_prep_globals = bwd_prep_globals<64>;
 
-        og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
-        o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
-        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)};
+        og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        d_global  prep_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_seq_len)};
 
         bwd_prep_globals bwd_g{prep_og_arg, prep_o_arg, prep_d_arg};
 
@@ -1023,15 +1036,15 @@ block_sparse_attention_backward(torch::Tensor q,
 
         using bwd_global_args = bwd_globals<64>;
 
-        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), 64U};
-        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), 64U};
-        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), 64U};
-        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), 64U};
-        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), 64U};
-        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), 64U};
-        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), 64U};
-        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_q_global  bwd_q_arg {d_q,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        bwd_k_global  bwd_k_arg {d_k,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        bwd_v_global  bwd_v_arg {d_v,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        bwd_og_global bwd_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        bwd_qg_global bwd_qg_arg{d_qg, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        bwd_kg_global bwd_kg_arg{d_kg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        bwd_vg_global bwd_vg_arg{d_vg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        bwd_l_global  bwd_l_arg {d_l,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_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>(q_seq_len)};
 
         bwd_global_args bwd_global{bwd_q_arg, 
                         bwd_k_arg, 
@@ -1042,14 +1055,14 @@ block_sparse_attention_backward(torch::Tensor q,
                         bwd_vg_arg, 
                         bwd_l_arg, 
                         bwd_d_arg, 
-                        static_cast<int>(seq_len), 
+                        static_cast<int>(kv_seq_len), // N is not used in the kernel
                         static_cast<int>(hr),
                         static_cast<int>(max_q_blocks_per_kv),
                         reinterpret_cast<int32_t*>(k2q_block_sparse_index.data_ptr()),
                         reinterpret_cast<int32_t*>(k2q_block_sparse_num.data_ptr()),
-                        reinterpret_cast<int32_t*>(block_size.data_ptr())};
+                        reinterpret_cast<int32_t*>(kv_block_size.data_ptr())};
 
-        dim3 grid_bwd_2(seq_len/64, qo_heads, batch);
+        dim3 grid_bwd_2(kv_seq_len/BLOCK_N, qo_heads, batch);
         threads = 128;
 
         //cudadevicesynchronize();
@@ -1088,9 +1101,9 @@ block_sparse_attention_backward(torch::Tensor q,
 
         using bwd_prep_globals = bwd_prep_globals<128>;
 
-        og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
-        o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
-        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)};
+        og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        d_global  prep_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_seq_len)};
 
         bwd_prep_globals bwd_g{prep_og_arg, prep_o_arg, prep_d_arg};
 
@@ -1127,15 +1140,15 @@ block_sparse_attention_backward(torch::Tensor q,
 
         using bwd_global_args = bwd_globals<128>;
 
-        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), 128U};
-        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), 128U};
-        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), 128U};
-        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), 128U};
-        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), 128U};
-        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), 128U};
-        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), 128U};
-        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_q_global  bwd_q_arg {d_q,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        bwd_k_global  bwd_k_arg {d_k,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        bwd_v_global  bwd_v_arg {d_v,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        bwd_og_global bwd_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        bwd_qg_global bwd_qg_arg{d_qg, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        bwd_kg_global bwd_kg_arg{d_kg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        bwd_vg_global bwd_vg_arg{d_vg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        bwd_l_global  bwd_l_arg {d_l,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_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>(q_seq_len)};
 
         bwd_global_args bwd_global{bwd_q_arg, 
                         bwd_k_arg, 
@@ -1146,14 +1159,14 @@ block_sparse_attention_backward(torch::Tensor q,
                         bwd_vg_arg, 
                         bwd_l_arg, 
                         bwd_d_arg, 
-                        static_cast<int>(seq_len), 
+                        static_cast<int>(kv_seq_len), // N is not used in the kernel
                         static_cast<int>(hr),
                         static_cast<int>(max_q_blocks_per_kv),
                         reinterpret_cast<int32_t*>(k2q_block_sparse_index.data_ptr()),
                         reinterpret_cast<int32_t*>(k2q_block_sparse_num.data_ptr()),
-                        reinterpret_cast<int32_t*>(block_size.data_ptr())};
+                        reinterpret_cast<int32_t*>(kv_block_size.data_ptr())};
 
-        dim3 grid_bwd_2(seq_len/64, qo_heads, batch);
+        dim3 grid_bwd_2(kv_seq_len/BLOCK_N, qo_heads, batch);
         threads = 128;
 
         //cudadevicesynchronize();

csrc/fastvideo_kernel/.gitignore

@@ -0,0 +1,7 @@
+build/
+dist/
+*.egg-info/
+__pycache__/
+*.so
+*.pyc
+.ipynb_checkpoints/

csrc/fastvideo_kernel/LICENSE

@@ -0,0 +1,187 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.

csrc/fastvideo_kernel/MANIFEST.in

@@ -0,0 +1,6 @@
+include LICENSE
+include README.md
+include pyproject.toml
+recursive-include src/fastvideo_kernel *.cu *.cuh *.cpp *.h
+recursive-include csrc *.cu *.cuh *.cpp *.h
+recursive-include tk *.cu *.cuh *.cpp *.h

csrc/fastvideo_kernel/README.md

@@ -0,0 +1,31 @@
+# FastVideo Kernel
+
+CUDA kernels for FastVideo video generation.
+
+## Installation
+
+```bash
+git submodule update --init --recursive
+cd csrc/fastvideo_kernel
+pip install .
+```
+
+## Usage
+
+```python
+from fastvideo_kernel import sliding_tile_attention, video_sparse_attn, moba_attn_varlen
+
+# Example: Sliding Tile Attention
+out = sliding_tile_attention(q, k, v, window_sizes, text_len)
+
+# Example: Video Sparse Attention (with Triton fallback)
+out = video_sparse_attn(q, k, v, block_sizes, topk=5)
+
+# Example: VMoBA
+out = moba_attn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, ...)
+```
+
+## Requirements
+
+- H100 GPU (sm_90a) for CUDA kernels
+- Triton for non-H100 fallback

csrc/fastvideo_kernel/csrc/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/fastvideo_kernel/csrc/st_attn_h100.cu

@@ -0,0 +1,573 @@
+// # Define TORCH_COMPILE macro
+
+#include "kittens.cuh"
+#include <cooperative_groups.h>
+#include <iostream>
+#include <stdio.h>
+#include <c10/cuda/CUDAGuard.h>
+
+// #define CLAMP(value, min, max) ((value) < (min) ? (min) : ((value) > (max) ? (max) : (value)))
+__device__ __forceinline__ int clamp_int(int value, int min, int max) {
+    return (value < min) ? min : ((value > max) ? max : value);
+}
+// #define ABS(x) ((x) < 0 ? -(x) : (x))
+__device__ __forceinline__ int abs_int(int value) {
+    return (value < 0) ? -value : value;
+}
+
+
+constexpr int CONSUMER_WARPGROUPS = (3); 
+constexpr int PRODUCER_WARPGROUPS = (1); 
+constexpr int NUM_WARPGROUPS      = (CONSUMER_WARPGROUPS+PRODUCER_WARPGROUPS); 
+constexpr int NUM_WORKERS         = (NUM_WARPGROUPS*kittens::WARPGROUP_WARPS); 
+
+using namespace kittens;
+namespace cg = cooperative_groups;
+
+template<int D> struct fwd_attend_ker_tile_dims {};
+template<> struct fwd_attend_ker_tile_dims<64> {
+    constexpr static int tile_width = (64);
+    constexpr static int qo_height  = (4*16);
+    constexpr static int kv_height  = (8*16);
+    constexpr static int stages     = (4); 
+};
+template<> struct fwd_attend_ker_tile_dims<128> {
+    constexpr static int tile_width = (128);
+    constexpr static int qo_height  = (4*16);
+    constexpr static int kv_height  = (8*16);
+    constexpr static int stages     = (2); 
+};
+
+template<int D> struct fwd_globals {
+    using q_tile    =         st_bf<fwd_attend_ker_tile_dims<D>::qo_height, fwd_attend_ker_tile_dims<D>::tile_width>;
+    using k_tile    =         st_bf<fwd_attend_ker_tile_dims<D>::kv_height, fwd_attend_ker_tile_dims<D>::tile_width>;
+    using v_tile    =         st_bf<fwd_attend_ker_tile_dims<D>::kv_height, fwd_attend_ker_tile_dims<D>::tile_width>;
+    using l_col_vec = col_vec<st_fl<fwd_attend_ker_tile_dims<D>::qo_height, fwd_attend_ker_tile_dims<D>::tile_width>>;
+    using o_tile    =         st_bf<fwd_attend_ker_tile_dims<D>::qo_height, fwd_attend_ker_tile_dims<D>::tile_width>;
+
+    using q_gl = gl<bf16,  -1, -1, -1, -1, q_tile>;
+    using k_gl = gl<bf16,  -1, -1, -1, -1, k_tile>;
+    using v_gl = gl<bf16,  -1, -1, -1, -1, v_tile>;
+    using l_gl = gl<float, -1, -1, -1, -1, l_col_vec>;
+    using o_gl = gl<bf16,  -1, -1, -1, -1, o_tile>;
+
+    q_gl q;
+    k_gl k;
+    v_gl v;
+    l_gl l;
+    o_gl o;
+
+    const int N; 
+    const int text_L;
+    const int hr;
+};
+
+
+template<int D, bool is_causal, bool text_q, bool text_kv, int DT, int DH, int DW, int CT, int CH, int CW>
+__global__  __launch_bounds__((NUM_WORKERS)*kittens::WARP_THREADS, 1)
+void fwd_attend_ker(const __grid_constant__ fwd_globals<D> g) {
+    extern __shared__ int __shm[]; 
+    tma_swizzle_allocator al((int*)&__shm[0]);
+    int warpid = kittens::warpid(), warpgroupid = warpid/kittens::WARPGROUP_WARPS;
+
+    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>;
+    
+    q_tile    (&q_smem)[CONSUMER_WARPGROUPS] = al.allocate<q_tile, CONSUMER_WARPGROUPS>();
+    k_tile    (&k_smem)[K::stages]           = al.allocate<k_tile, K::stages          >();
+    v_tile    (&v_smem)[K::stages]           = al.allocate<v_tile, K::stages          >();
+    l_col_vec (&l_smem)[CONSUMER_WARPGROUPS] = al.allocate<l_col_vec, CONSUMER_WARPGROUPS>();
+    auto      (*o_smem)                      = reinterpret_cast<o_tile(*)>(q_smem);
+    int img_kv_blocks;
+    int kv_blocks   = g.N / (K::kv_height);
+    if constexpr (text_kv) {
+        img_kv_blocks = kv_blocks - 3;
+    } else {
+        img_kv_blocks = kv_blocks;
+    }
+    int kv_head_idx = blockIdx.y / g.hr;
+    int seq_idx;
+    if constexpr (text_q) {
+        seq_idx = CT * CH * CW * 6.0 + blockIdx.x * CONSUMER_WARPGROUPS;
+    } else {
+        seq_idx = blockIdx.x * CONSUMER_WARPGROUPS; 
+    }
+    __shared__ kittens::semaphore qsmem_semaphore, k_smem_arrived[K::stages], v_smem_arrived[K::stages], compute_done[K::stages];
+    if (threadIdx.x == 0) { 
+        init_semaphore(qsmem_semaphore, 0, 1); 
+        for(int j = 0; j < K::stages; j++) {
+            init_semaphore(k_smem_arrived[j], 0, 1); 
+            init_semaphore(v_smem_arrived[j], 0, 1); 
+            init_semaphore(compute_done[j], CONSUMER_WARPGROUPS, 0); 
+        }
+
+        tma::expect_bytes(qsmem_semaphore, sizeof(q_smem));
+
+        for (int wg = 0; wg < CONSUMER_WARPGROUPS; wg++) {
+            coord<q_tile> q_tile_idx = {blockIdx.z, blockIdx.y, (seq_idx) + wg, 0};
+            tma::load_async(q_smem[wg], g.q, q_tile_idx, qsmem_semaphore);
+        }
+
+        if constexpr (text_q){
+            for (int j = 0; j < K::stages - 1; j++) {
+                coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, j, 0};
+                tma::expect_bytes(k_smem_arrived[j], sizeof(k_tile));
+                tma::load_async(k_smem[j], g.k, kv_tile_idx, k_smem_arrived[j]);
+                tma::expect_bytes(v_smem_arrived[j], sizeof(v_tile));
+                tma::load_async(v_smem[j], g.v, kv_tile_idx, v_smem_arrived[j]);
+            }
+        } else {
+            int qt = seq_idx / 6 / (CH * CW);
+            int qh = (seq_idx / 6) % (CH * CW) / CW;
+            int qw = (seq_idx / 6) % CW;
+            qt = clamp_int(qt, DT, CT-DT-1);
+            qh = clamp_int(qh, DH, CH-DH-1);
+            qw = clamp_int(qw, DW, CW-DW-1);
+            int count = 0;
+            int j = 0;
+            while (count < K::stages - 1) {
+                int kt = j / 3 / (CH * CW);
+                int kh = (j / 3) % (CH * CW) / CW;
+                int kw = (j / 3) % CW;
+                bool mask = (abs_int(qt - kt) <= DT) && (abs_int(qh - kh) <= DH) && (abs_int(qw - kw) <= DW);
+                if (mask){
+                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, j, 0};
+                    tma::expect_bytes(k_smem_arrived[count], sizeof(k_tile));
+                    tma::load_async(k_smem[count], g.k, kv_tile_idx, k_smem_arrived[count]);
+                    tma::expect_bytes(v_smem_arrived[count], sizeof(v_tile));
+                    tma::load_async(v_smem[count], g.v, kv_tile_idx, v_smem_arrived[count]);
+                    count += 1;
+                }
+                j += 1;
+            }
+        }
+    }
+    __syncthreads(); 
+
+    int pipe_idx = K::stages - 1; 
+    
+    if(warpgroupid == NUM_WARPGROUPS-1) {
+        warpgroup::decrease_registers<32>();      
+        
+        int kv_iters; 
+        if constexpr (is_causal) {
+            kv_iters = (seq_idx * (K::qo_height/kittens::TILE_ROW_DIM<bf16>)) - 1 + (CONSUMER_WARPGROUPS * (K::qo_height/kittens::TILE_ROW_DIM<bf16>)); 
+            kv_iters = ((kv_iters / (K::kv_height/kittens::TILE_ROW_DIM<bf16>)) == 0) ? (0) : ((kv_iters / (K::kv_height/kittens::TILE_ROW_DIM<bf16>)) - 1);
+        }
+        else { kv_iters = kv_blocks-2;}
+
+        if(warpid == NUM_WORKERS-4) {
+            if constexpr (text_q){
+                for (auto kv_idx = pipe_idx - 1; kv_idx <= kv_iters; kv_idx++) {
+                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, kv_idx + 1, 0};
+                    tma::expect_bytes(k_smem_arrived[(kv_idx+1)%K::stages], sizeof(k_tile));
+                    tma::load_async(k_smem[(kv_idx+1)%K::stages], g.k, kv_tile_idx, k_smem_arrived[(kv_idx+1)%K::stages]);
+                    tma::expect_bytes(v_smem_arrived[(kv_idx+1)%K::stages], sizeof(v_tile));
+                    tma::load_async(v_smem[(kv_idx+1)%K::stages], g.v, kv_tile_idx, v_smem_arrived[(kv_idx+1)%K::stages]);
+                    kittens::wait(compute_done[(kv_idx)%K::stages], (kv_idx/K::stages)%2);
+                }
+            } else {
+                int qt = seq_idx / 6 / (CH * CW);
+                int qh = (seq_idx / 6) % (CH * CW) / CW;
+                int qw = (seq_idx / 6) % CW;
+                qt = clamp_int(qt, DT, CT-DT-1);
+                qh = clamp_int(qh, DH, CH-DH-1);
+                qw = clamp_int(qw, DW, CW-DW-1);
+                int k_t_min = clamp_int(qt-DT, 0, CT-1);
+                int k_t_max = clamp_int(qt+DT, 0, CT-1);
+                int k_h_min = clamp_int(qh-DH, 0, CH-1);
+                int k_h_max = clamp_int(qh+DH, 0, CH-1);
+                int k_w_min = clamp_int(qw-DW, 0, CW-1);
+                int k_w_max = clamp_int(qw+DW, 0, CW-1);
+                int count = 0;
+                for (int kt = k_t_min; kt <= k_t_max; kt++) {
+                    for (int kh = k_h_min; kh <= k_h_max; kh++) {
+                        for (int kw = k_w_min; kw <= k_w_max; kw++) {
+                            for (int j = 0; j <= 2; j++){
+                                if (count >= K::stages - 1) {
+                                    int index = ((kt * (CH * CW)) + (kh * CW) + kw) * 3 + j;
+                                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, index, 0};
+                                    tma::expect_bytes(k_smem_arrived[count%K::stages], sizeof(k_tile));
+                                    tma::load_async(k_smem[count%K::stages], g.k, kv_tile_idx, k_smem_arrived[count%K::stages]);
+                                    tma::expect_bytes(v_smem_arrived[count%K::stages], sizeof(v_tile));
+                                    tma::load_async(v_smem[count%K::stages], g.v, kv_tile_idx, v_smem_arrived[count%K::stages]);
+                                    kittens::wait(compute_done[(count - 1)%K::stages], ((count - 1)/K::stages)%2);
+                                    count += 1;
+                                } else {
+                                    count += 1;
+                                }
+                            }
+                        }
+                    }
+                }
+                // for text 
+                for (int index = img_kv_blocks; index < kv_blocks; index++) {
+                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, index, 0};
+                    tma::expect_bytes(k_smem_arrived[count%K::stages], sizeof(k_tile));
+                    tma::load_async(k_smem[count%K::stages], g.k, kv_tile_idx, k_smem_arrived[count%K::stages]);
+                    tma::expect_bytes(v_smem_arrived[count%K::stages], sizeof(v_tile));
+                    tma::load_async(v_smem[count%K::stages], g.v, kv_tile_idx, v_smem_arrived[count%K::stages]);
+                    kittens::wait(compute_done[(count - 1)%K::stages], ((count - 1)/K::stages)%2);
+                    count += 1;
+                }
+            }
+
+
+        }
+    }
+    else {
+        warpgroup::increase_registers<160>();
+
+        rt_fl<16, K::kv_height>  att_block;
+        rt_bf<16, K::kv_height>  att_block_mma;
+        rt_fl<16, K::tile_width> o_reg;
+        
+        col_vec<rt_fl<16, K::kv_height>> max_vec, norm_vec, max_vec_last_scaled, max_vec_scaled;
+        
+        neg_infty(max_vec);
+        zero(norm_vec);
+        zero(o_reg);
+
+        int kv_iters; 
+        if constexpr (is_causal) {
+            kv_iters = (seq_idx * 4) - 1 + (CONSUMER_WARPGROUPS * 4);
+            kv_iters = (kv_iters/8);
+        }
+        else if constexpr (text_q){ 
+            // the last three kv blocks are for text, we process them separately
+            kv_iters = img_kv_blocks - 1;
+        } else {
+            kv_iters = clamp_int(DT*2+1, 1, CT) * clamp_int(DH*2+1, 1, CH) * clamp_int(DW*2+1, 1, CW) * 3 - 1 ; 
+        }
+
+        kittens::wait(qsmem_semaphore, 0);
+        for (auto kv_idx = 0; kv_idx <= kv_iters; kv_idx++) {
+
+            kittens::wait(k_smem_arrived[(kv_idx)%K::stages], (kv_idx/K::stages)%2);
+            warpgroup::mm_ABt(att_block, q_smem[warpgroupid], k_smem[(kv_idx)%K::stages]);
+            
+            copy(max_vec_last_scaled, max_vec);
+            if constexpr (D == 64) { mul(max_vec_last_scaled, max_vec_last_scaled, 1.44269504089f*0.125f); }
+            else                   { mul(max_vec_last_scaled, max_vec_last_scaled, 1.44269504089f*0.08838834764f); }
+            
+            warpgroup::mma_async_wait();
+
+            row_max(max_vec, att_block, max_vec);
+            
+            if constexpr (D == 64) { 
+                mul(att_block, att_block,    1.44269504089f*0.125f); 
+                mul(max_vec_scaled, max_vec, 1.44269504089f*0.125f);
+            }
+            else                   { 
+                mul(att_block, att_block,    1.44269504089f*0.08838834764f); 
+                mul(max_vec_scaled, max_vec, 1.44269504089f*0.08838834764f);
+            }
+
+            sub_row(att_block, att_block, max_vec_scaled);
+            exp2(att_block, att_block);
+            sub(max_vec_last_scaled, max_vec_last_scaled, max_vec_scaled);
+            exp2(max_vec_last_scaled,       max_vec_last_scaled);
+            mul(norm_vec,            norm_vec,     max_vec_last_scaled);
+            row_sum(norm_vec,  att_block, norm_vec);
+            add(att_block, att_block, 0.f);
+            copy(att_block_mma, att_block); 
+            mul_row(o_reg, o_reg, max_vec_last_scaled); 
+
+            kittens::wait(v_smem_arrived[(kv_idx)%K::stages], (kv_idx/K::stages)%2); 
+
+            warpgroup::mma_AB(o_reg, att_block_mma, v_smem[(kv_idx)%K::stages]);
+            warpgroup::mma_async_wait();
+
+            if(warpgroup::laneid() == 0) arrive(compute_done[(kv_idx)%K::stages], 1);
+        }
+        // the last three kv blocks are for text, we process them separately
+        if constexpr(text_kv) {
+            for (auto kv_idx = kv_iters + 1; kv_idx <= kv_iters + 3; kv_idx++) {
+
+                kittens::wait(k_smem_arrived[(kv_idx)%K::stages], (kv_idx/K::stages)%2);
+                warpgroup::mm_ABt(att_block, q_smem[warpgroupid], k_smem[(kv_idx)%K::stages]);
+                
+                copy(max_vec_last_scaled, max_vec);
+                if constexpr (D == 64) { mul(max_vec_last_scaled, max_vec_last_scaled, 1.44269504089f*0.125f); }
+                else                   { mul(max_vec_last_scaled, max_vec_last_scaled, 1.44269504089f*0.08838834764f); }
+                
+                warpgroup::mma_async_wait();
+                // apply non-pad mask
+                int offset = g.text_L - (kv_idx - (kv_iters + 1)) * K::kv_height;
+                // printf("k_idx_start: %d, k_idx_end: %d, text_end: %d, offset: %d\n", k_idx_start, k_idx_end, text_end, offset);
+                right_fill(att_block, att_block, offset, base_types::constants<float>::neg_infty());
+
+
+                row_max(max_vec, att_block, max_vec);
+                
+                if constexpr (D == 64) { 
+                    mul(att_block, att_block,    1.44269504089f*0.125f); 
+                    mul(max_vec_scaled, max_vec, 1.44269504089f*0.125f);
+                }
+                else                   { 
+                    mul(att_block, att_block,    1.44269504089f*0.08838834764f); 
+                    mul(max_vec_scaled, max_vec, 1.44269504089f*0.08838834764f);
+                }
+
+                sub_row(att_block, att_block, max_vec_scaled);
+                exp2(att_block, att_block);
+                sub(max_vec_last_scaled, max_vec_last_scaled, max_vec_scaled);
+                exp2(max_vec_last_scaled,       max_vec_last_scaled);
+                mul(norm_vec,            norm_vec,     max_vec_last_scaled);
+                row_sum(norm_vec,  att_block, norm_vec);
+                add(att_block, att_block, 0.f);
+                copy(att_block_mma, att_block); 
+                mul_row(o_reg, o_reg, max_vec_last_scaled); 
+
+                kittens::wait(v_smem_arrived[(kv_idx)%K::stages], (kv_idx/K::stages)%2); 
+
+                warpgroup::mma_AB(o_reg, att_block_mma, v_smem[(kv_idx)%K::stages]);
+                warpgroup::mma_async_wait();
+
+                if(warpgroup::laneid() == 0) arrive(compute_done[(kv_idx)%K::stages], 1);
+            }
+        }
+
+        div_row(o_reg, o_reg, norm_vec);
+        warpgroup::store(o_smem[warpgroupid], o_reg); 
+        warpgroup::sync(warpgroupid+4);
+
+        if (warpid % 4 == 0) {
+            coord<o_tile> o_tile_idx = {blockIdx.z, blockIdx.y, (seq_idx) + warpgroupid, 0};
+            tma::store_async(g.o, o_smem[warpgroupid], o_tile_idx);
+        }
+
+        mul(max_vec_scaled,   max_vec_scaled, 0.69314718056f);
+        log(norm_vec, norm_vec);
+        add(norm_vec, norm_vec, max_vec_scaled);
+
+        if constexpr (D == 64) { mul(norm_vec, norm_vec, -8.0f); }
+        else                   { mul(norm_vec, norm_vec, -11.313708499f); }
+    
+        warpgroup::store(l_smem[warpgroupid], norm_vec);
+        warpgroup::sync(warpgroupid+4);
+
+        if (warpid % 4 == 0) {
+            coord<l_col_vec> tile_idx = {blockIdx.z, blockIdx.y, 0, (seq_idx) + warpgroupid};
+            tma::store_async(g.l, l_smem[warpgroupid], tile_idx);
+        }
+        tma::store_async_wait();
+    }
+}
+
+
+
+#include "pyutils/torch_helpers.cuh"
+#include <ATen/cuda/CUDAContext.h>
+#include <iostream>
+
+torch::Tensor 
+sta_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o, int kernel_t_size, int kernel_h_size, int kernel_w_size, int text_length, bool process_text, bool has_text, int kernel_aspect_ratio_flag)
+{
+    CHECK_INPUT(q);
+    CHECK_INPUT(k);
+    CHECK_INPUT(v);
+
+    auto batch    = q.size(0);
+    auto seq_len  = q.size(2); 
+    auto head_dim = q.size(3);  
+    auto qo_heads = q.size(1);
+    auto kv_heads = k.size(1);
+
+    // check to see that these dimensions match for all inputs
+    TORCH_CHECK(q.size(0) == batch, "Q batch dimension - idx 0 - must match for all inputs");
+    TORCH_CHECK(k.size(0) == batch, "K batch dimension - idx 0 - must match for all inputs");
+    TORCH_CHECK(v.size(0) == batch, "V batch dimension - idx 0 - must match for all inputs");
+
+    TORCH_CHECK(q.size(2) == seq_len, "Q sequence length dimension - idx 2 - must match for all inputs");
+    TORCH_CHECK(k.size(2) == seq_len, "K sequence length dimension - idx 2 - must match for all inputs");
+    TORCH_CHECK(v.size(2) == seq_len, "V sequence length dimension - idx 2 - must match for all inputs");
+
+    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");
+
+    TORCH_CHECK(qo_heads >= kv_heads, "QO heads must be greater than or equal to KV heads");
+    TORCH_CHECK(qo_heads % kv_heads == 0, "QO heads must be divisible by KV heads");
+    TORCH_CHECK(q.size(1) == qo_heads, "QO head dimension - idx 1 - must match for all inputs");
+    TORCH_CHECK(k.size(1) == kv_heads, "KV head dimension - idx 1 - must match for all inputs");
+    TORCH_CHECK(v.size(1) == kv_heads, "KV head dimension - idx 1 - must match for all inputs");  
+
+    auto hr = qo_heads / kv_heads;
+
+    c10::BFloat16* q_ptr = q.data_ptr<c10::BFloat16>();
+    c10::BFloat16* k_ptr = k.data_ptr<c10::BFloat16>();
+    c10::BFloat16* v_ptr = v.data_ptr<c10::BFloat16>();
+
+    bf16*  d_q = reinterpret_cast<bf16*>(q_ptr);
+    bf16*  d_k = reinterpret_cast<bf16*>(k_ptr);
+    bf16*  d_v = reinterpret_cast<bf16*>(v_ptr);
+    
+
+    
+    torch::Tensor l_vec = torch::empty({static_cast<const uint>(batch), 
+                                        static_cast<const uint>(qo_heads), 
+                                        static_cast<const uint>(seq_len), 
+                                        static_cast<const uint>(1)}, 
+                                        torch::TensorOptions().dtype(torch::kFloat).device(q.device()).memory_format(at::MemoryFormat::Contiguous));
+        
+
+    bf16*  o_ptr = reinterpret_cast<bf16*>(o.data_ptr<c10::BFloat16>());
+    bf16*  d_o   = reinterpret_cast<bf16*>(o_ptr);
+
+    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(); 
+
+
+    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>(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) {
+            // TORCH_CHECK(seq_len % (CONSUMER_WARPGROUPS*kittens::TILE_DIM*4) == 0, "sequence length must be divisible by 192");
+            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)      { 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);
+                }
+#undef LAUNCH_IMAGE_KER
+            } else {
+                cudaFuncSetAttribute(
+                    fwd_attend_ker<128, false, true, true, 1, 1, 1, 5, 6, 10>,
+                    cudaFuncAttributeMaxDynamicSharedMemorySize,
+                    mem_size
+                );
+                fwd_attend_ker<128, false, true, true, 1, 1, 1, 5, 6, 10><<<grid_text, (32*NUM_WORKERS), mem_size, stream>>>(g);
+            }
+
+        } 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)      { 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);
+                }
+#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)      { 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);
+                }
+#undef LAUNCH_IMAGE_KER
+            }
+
+            else {
+                TORCH_CHECK(false, "Unsupported kernel_aspect_ratio_flag: ", kernel_aspect_ratio_flag);
+            }
+
+        }
+        CHECK_CUDA_ERROR(cudaGetLastError());
+        // cudaStreamSynchronize(stream);
+    }
+
+    return o;
+    //cudadevicesynchronize();
+}
+

csrc/fastvideo_kernel/csrc/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/fastvideo_kernel/pyproject.toml

@@ -0,0 +1,27 @@
+[build-system]
+requires = ["setuptools>=61.0", "torch>=2.5.0", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "fastvideo-kernel"
+version = "0.1.0"
+description = "CUDA kernels for FastVideo"
+readme = "README.md"
+requires-python = ">=3.10"
+license = {text = "Apache-2.0"}
+authors = [{name = "Hao AI Lab"}]
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "Environment :: GPU :: NVIDIA CUDA :: 12",
+    "License :: OSI Approved :: Apache Software License",
+]
+dependencies = [
+    "torch>=2.5.0",
+    "triton>=2.0.0"
+]
+
+[project.urls]
+Repository = "https://github.com/hao-ai-lab/FastVideo"
+
+[tool.setuptools.packages.find]
+where = ["src"]

csrc/fastvideo_kernel/setup.py

@@ -0,0 +1,132 @@
+import os
+import subprocess
+import sys
+from pathlib import Path
+from setuptools import find_packages, setup
+from torch.utils.cpp_extension import BuildExtension, CUDAExtension
+
+ROOT = Path(__file__).parent.absolute()
+CSRC_DIR = ROOT / "csrc"
+
+# Path to ThunderKittens (TK)
+def get_tk_dir():
+    tk_env = os.getenv("THUNDERKITTENS_ROOT")
+    if tk_env:
+        return tk_env
+    
+    # Check common locations
+    possible_paths = [
+        ROOT / "tk",
+        ROOT / "csrc" / "tk",
+        ROOT.parent / "attn" / "sliding_tile_attn" / "tk",
+        ROOT.parent / "attn" / "video_sparse_attn" / "tk",
+    ]
+    for p in possible_paths:
+        if (p / "include" / "kittens.cuh").exists():
+            return str(p)
+    
+    # Default fallback
+    return str(ROOT.parent / "attn" / "sliding_tile_attn" / "tk")
+
+TK_DIR = get_tk_dir()
+
+def get_cuda_flags(tk_root: str) -> list:
+    python_include = subprocess.check_output(
+        ["python", "-c", "import sysconfig; print(sysconfig.get_path('include'))"]
+    ).decode().strip()
+    
+    torch_includes = 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().split()
+    
+    return [
+        "-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",
+        "-DKITTENS_HOPPER",
+        "-arch=sm_90a",
+    ] + torch_includes
+
+def get_extensions():
+    if not torch.cuda.is_available():
+        return []
+
+    extensions = []
+    cpp_flags = ["-std=c++20", "-O3"]
+    
+    # Check if TK is available
+    if not os.path.exists(os.path.join(TK_DIR, "include", "kittens.cuh")):
+        print(f"Warning: ThunderKittens not found at {TK_DIR}. CUDA kernels will not be built.")
+        return []
+
+    cuda_flags = get_cuda_flags(TK_DIR)
+    
+    # STA Extension
+    extensions.append(CUDAExtension(
+        "fastvideo_kernel._C.st_attn",
+        sources=[
+            "csrc/st_attn.cpp",
+            "csrc/st_attn_h100.cu",
+        ],
+        extra_compile_args={
+            "cxx": cpp_flags + ["-DTK_COMPILE_ST_ATTN"], 
+            "nvcc": cuda_flags + ["-DTK_COMPILE_ST_ATTN"]
+        },
+        libraries=["cuda"],
+    ))
+    
+    # VSA Extension
+    extensions.append(CUDAExtension(
+        "fastvideo_kernel._C.vsa",
+        sources=[
+            "csrc/vsa.cpp",
+            "csrc/block_sparse_h100.cu",
+        ],
+        extra_compile_args={
+            "cxx": cpp_flags + ["-DTK_COMPILE_BLOCK_SPARSE"], 
+            "nvcc": cuda_flags + ["-DTK_COMPILE_BLOCK_SPARSE"]
+        },
+        libraries=["cuda"],
+    ))
+    
+    return extensions
+
+ext_modules = []
+if not any(arg in sys.argv for arg in ["clean", "egg_info", "--version"]):
+    try:
+        import torch
+        ext_modules = get_extensions()
+    except Exception as e:
+        print(f"Warning: Failed to configure CUDA extensions: {e}")
+
+setup(
+    name="fastvideo-kernel",
+    version="0.1.0",
+    description="Unified CUDA kernels for FastVideo",
+    long_description=open("README.md").read(),
+    long_description_content_type="text/markdown",
+    license="Apache-2.0",
+    author="Hao AI Lab",
+    url="https://github.com/hao-ai-lab/FastVideo",
+    package_dir={"": "src"},
+    packages=find_packages(where="src"),
+    ext_modules=ext_modules,
+    cmdclass={"build_ext": BuildExtension} if ext_modules else {},
+    python_requires=">=3.10",
+    install_requires=["torch>=2.5.0", "triton>=2.0.0"],
+)

csrc/fastvideo_kernel/src/fastvideo_kernel/__init__.py

@@ -0,0 +1,21 @@
+__version__ = "0.1.0"
+
+from fastvideo_kernel.ops import (
+    sliding_tile_attention,
+    video_sparse_attn,
+)
+
+from fastvideo_kernel.vmoba import (
+    moba_attn_varlen,
+    process_moba_input,
+    process_moba_output,
+)
+
+__all__ = [
+    "sliding_tile_attention",
+    "video_sparse_attn",
+    "moba_attn_varlen",
+    "process_moba_input",
+    "process_moba_output",
+    "__version__",
+]

csrc/fastvideo_kernel/src/fastvideo_kernel/ops.py

@@ -0,0 +1,103 @@
+import math
+import torch
+from .triton_kernels.block_sparse_attn_triton import triton_block_sparse_attn_forward
+from .triton_kernels.index import map_to_index
+
+try:
+    from fastvideo_kernel._C.st_attn import sta_fwd
+except ImportError:
+    sta_fwd = None
+
+try:
+    from fastvideo_kernel._C.vsa import block_sparse_fwd, block_sparse_bwd
+except ImportError:
+    block_sparse_fwd = None
+    block_sparse_bwd = None
+
+
+def sliding_tile_attention(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    window_size: list,
+    text_length: int,
+    has_text: bool = True,
+    seq_shape: str = "30x48x80",
+) -> torch.Tensor:
+    if sta_fwd is None:
+        raise RuntimeError("STA kernel not compiled. Requires H100 and ThunderKittens at build time.")
+    
+    seq_length = q.shape[2]
+    shape_map = {"30x48x80": 1, "36x48x48": 2, "18x48x80": 3}
+    
+    if has_text:
+        target_size = math.ceil(seq_length / 384) * 384
+        pad_size = target_size - seq_length
+        if pad_size > 0:
+            q = torch.cat([q, q[:, :, -pad_size:]], dim=2)
+            k = torch.cat([k, k[:, :, -pad_size:]], dim=2)
+            v = torch.cat([v, v[:, :, -pad_size:]], dim=2)
+    
+    output = torch.empty_like(q)
+    flag = shape_map[seq_shape]
+    
+    for head_idx, (t, h, w) in enumerate(window_size):
+        sta_fwd(
+            q[:, head_idx:head_idx+1],
+            k[:, head_idx:head_idx+1],
+            v[:, head_idx:head_idx+1],
+            output[:, head_idx:head_idx+1],
+            t, h, w, text_length, False, has_text, flag
+        )
+    
+    if has_text:
+        sta_fwd(q, k, v, output, 3, 3, 3, text_length, True, True, flag)
+    
+    return output[:, :, :seq_length]
+
+
+def video_sparse_attn(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+    topk: int,
+    block_size: int | tuple = 64,
+    compress_attn_weight: torch.Tensor = None,
+) -> torch.Tensor:
+    if isinstance(block_size, int):
+        block_size = (block_size, block_size, block_size)
+    
+    block_elements = block_size[0] * block_size[1] * block_size[2]
+    batch, heads, seq_len, dim = q.shape
+    
+    # Compression branch
+    q_c = q.view(batch, heads, seq_len // block_elements, block_elements, dim)
+    k_c = k.view(batch, heads, seq_len // block_elements, block_elements, dim)
+    v_c = v.view(batch, heads, seq_len // block_elements, block_elements, dim)
+    
+    q_c = (q_c.float().sum(dim=3) / variable_block_sizes.view(1, 1, -1, 1)).to(q.dtype)
+    k_c = (k_c.float().sum(dim=3) / variable_block_sizes.view(1, 1, -1, 1)).to(k.dtype)
+    v_c = (v_c.float().sum(dim=3) / variable_block_sizes.view(1, 1, -1, 1)).to(v.dtype)
+    
+    scores = torch.matmul(q_c, k_c.transpose(-2, -1)) / (dim ** 0.5)
+    attn = torch.softmax(scores, dim=-1)
+    out_c = torch.matmul(attn, v_c)
+    
+    out_c = out_c.view(batch, heads, seq_len // block_elements, 1, dim)
+    out_c = out_c.repeat(1, 1, 1, block_elements, 1).view(batch, heads, seq_len, dim)
+    
+    # Sparse branch
+    topk_idx = torch.topk(scores, topk, dim=-1).indices
+    mask = torch.zeros_like(scores, dtype=torch.bool).scatter_(-1, topk_idx, True)
+    
+    if block_sparse_fwd is not None:
+        idx, num = map_to_index(mask)
+        out_s, _ = block_sparse_fwd(q, k, v, idx, num, variable_block_sizes.int())
+    else:
+        idx, num = map_to_index(mask)
+        out_s, _ = triton_block_sparse_attn_forward(q, k, v, idx, num, variable_block_sizes)
+    
+    if compress_attn_weight is not None:
+        return out_c * compress_attn_weight + out_s
+    return out_c + out_s

csrc/fastvideo_kernel/src/fastvideo_kernel/triton_kernels/block_sparse_attn_triton.py

@@ -0,0 +1,449 @@
+"""
+Fused Attention
+===============
+
+This is a Triton implementation of the Flash Attention v2 algorithm from Tri Dao
+(https://tridao.me/publications/flash2/flash2.pdf)
+
+Credits: OpenAI kernel team
+"""
+
+
+import torch
+import triton
+import triton.language as tl
+
+# ──────────────────────────── SPARSE ADDITION BEGIN ───────────────────────────
+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.
+configs = [
+    triton.Config({'BLOCK_M': BM, 'BLOCK_N': BN}, num_stages=s, num_warps=w) \
+    for BM in [64]\
+    for BN in [64]\
+    for s in [3, 4, 7]\
+    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):
+    """
+    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_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
+
+    # ----- base pointers -----
+    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))
+
+    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)
+
+    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)
+    m_i = tl.full([BLOCK_M], -float("inf"), tl.float32)
+    l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0
+    acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32)
+    qk_scale = sm_scale * 1.44269504  # 1/ln2
+    q = tl.load(Q_ptr)
+
+    # ----- sparse loop over valid K/V tiles -----
+    for i in range(0, kv_blocks):
+        kv_idx = tl.load(kv_ptr + i).to(tl.int32)
+        block_size = tl.load(variable_block_sizes + kv_idx)
+        K_ptr = tl.advance(K_base, (0, kv_idx * BLOCK_N))
+        V_ptr = tl.advance(V_base, (kv_idx * BLOCK_N, 0))
+
+        k = tl.load(K_ptr)
+        qk = tl.dot(q, k)
+        # mask out invalid columns
+        mask = tl.arange(0, BLOCK_N) < block_size
+        qk = tl.where(mask[None, :], qk, -float("inf"))
+
+        m_ij = tl.maximum(m_i, tl.max(qk, 1) * qk_scale)
+        p = tl.math.exp2(qk * qk_scale - m_ij[:, None])
+        l_ij = tl.sum(p, 1)
+
+        alpha = tl.math.exp2(m_i - m_ij)
+        l_i = l_i * alpha + l_ij
+        acc = acc * alpha[:, None]
+
+        v = tl.load(V_ptr)
+        acc = tl.dot(p.to(tl.bfloat16), v, acc)
+        m_i = m_ij
+
+    # ----- epilogue -----
+    m_i += tl.math.log2(l_i)
+    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  #
+                         ):
+    off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
+    off_hz = tl.program_id(1)
+    off_n = tl.arange(0, HEAD_DIM)
+    # load
+    o = tl.load(O + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :])
+    do = tl.load(DO + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :]).to(tl.float32)
+    delta = tl.sum(o * do, axis=1)
+    # write-back
+    tl.store(Delta + off_hz * N_CTX + off_m, delta)
+
+
+# The main inner-loop logic for computing dK and dV.
+@triton.jit
+def _attn_bwd_dkdv(dk, dv,  #
+                   Q, k, v, sm_scale,  #
+                   DO,  #
+                   M, D,  #
+                   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)
+    qT_ptrs = Q + offs_m[None, :] * stride_tok + offs_k[:, None] * stride_d
+    do_ptrs = DO + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d
+    # BLOCK_N1 must be a multiple of BLOCK_M1, otherwise the code wouldn't work.
+    tl.static_assert(BLOCK_N1 % BLOCK_M1 == 0)
+    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
+    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
+    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
+        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)
+        m = tl.load(M + offs_m)
+        qkT = tl.dot(k, qT)
+        pT = tl.math.exp2(qkT - m[None, :])
+        mask = tl.arange(0, BLOCK_N1) < block_size
+        pT = tl.where(mask[:, None], pT, 0.0)
+
+        do = tl.load(do_ptrs + block_sparse_offset * stride_tok)
+        # Compute dV.
+        ppT = pT
+        ppT = ppT.to(tl.bfloat16)
+        dv += tl.dot(ppT, do)
+        # D (= delta) is pre-divided by ds_scale.
+        Di = tl.load(D + offs_m)
+        # Compute dP and dS.
+        dpT = tl.dot(v, tl.trans(do)).to(tl.float32)
+        dsT = pT * (dpT - Di[None, :])
+        dsT = dsT.to(tl.bfloat16)
+        dk += tl.dot(dsT, tl.trans(qT))
+        # Increment pointers.
+    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):
+    offs_m = start_m + tl.arange(0, BLOCK_M2)
+    offs_n = start_n + tl.arange(0, BLOCK_N2)
+    offs_k = tl.arange(0, HEAD_DIM)
+    kT_ptrs = K + offs_n[None, :] * stride_tok + offs_k[:, None] * stride_d
+    vT_ptrs = V + offs_n[None, :] * stride_tok + offs_k[:, None] * stride_d
+    # D (= delta) is pre-divided by ds_scale.
+    Di = tl.load(D + offs_m)
+    # BLOCK_M2 must be a multiple of BLOCK_N2, otherwise the code wouldn't work.
+    tl.static_assert(BLOCK_M2 % BLOCK_N2 == 0)
+    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_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
+        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)
+        # Compute dP and dS.
+        dp = tl.dot(do, vT).to(tl.float32)
+        ds = p * (dp - Di[:, None])
+        ds = ds.to(tl.bfloat16)
+        # Compute dQ.
+        # NOTE: We need to de-scale dq in the end, because kT was pre-scaled.
+        dq += tl.dot(ds, tl.trans(kT))
+        # Increment pointers.
+    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):
+    LN2 = 0.6931471824645996  # = ln(2)
+
+    bhid = tl.program_id(2)
+    off_chz = (bhid * N_CTX).to(tl.int64)
+    adj = (stride_h * (bhid % H) + stride_z * (bhid // H)).to(tl.int64)
+    pid = tl.program_id(0)
+
+    # offset pointers for batch/head
+    Q += adj
+    K += adj
+    V += adj
+    DO += adj
+    DQ += adj
+    DK += adj
+    DV += adj
+    M += off_chz
+    D += off_chz
+
+    # load scales
+    offs_k = tl.arange(0, HEAD_DIM)
+
+    start_n = pid * BLOCK_N1
+    start_m = 0
+
+    offs_n = start_n + tl.arange(0, BLOCK_N1)
+
+    dv = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32)
+    dk = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32)
+
+    # load K and V: they stay in SRAM throughout the inner loop.
+    k = tl.load(K + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d)
+    v = tl.load(V + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d)
+
+
+    num_steps = N_CTX // BLOCK_M1
+
+    dk, dv = _attn_bwd_dkdv(  #
+        dk, dv,  #
+        Q, k, v, sm_scale,  #
+        DO,  #
+        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  #
+    )
+
+    dv_ptrs = DV + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d
+    tl.store(dv_ptrs, dv)
+
+    # Write back dK.
+    dk *= sm_scale
+    dk_ptrs = DK + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d
+    tl.store(dk_ptrs, dk)
+
+    # THIS BLOCK DOES DQ:
+    start_m = pid * BLOCK_M2
+    end_n = 0
+
+    offs_m = start_m + tl.arange(0, BLOCK_M2)
+
+    q = tl.load(Q + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d)
+    dq = tl.zeros([BLOCK_M2, HEAD_DIM], dtype=tl.float32)
+    do = tl.load(DO + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d)
+
+    m = tl.load(M + offs_m)
+    m = m[:, None]
+
+    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  #
+                      )
+    # 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):
+    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 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
+    )
+
+    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):
+    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)
+    dk = torch.empty_like(k)
+    dv = torch.empty_like(v)
+    BATCH, N_HEAD, N_CTX = q.shape[:3]
+    BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 32, 64, 64, 32
+    RCP_LN2 = 1.4426950408889634  # = 1.0 / ln(2)
+    arg_k = k
+    arg_k = arg_k * (sm_scale * RCP_LN2)
+    PRE_BLOCK = 64
+    assert N_CTX % PRE_BLOCK == 0
+    pre_grid = (N_CTX // PRE_BLOCK, BATCH * N_HEAD)
+    delta = torch.empty_like(M)
+    _attn_bwd_preprocess[pre_grid](
+        o, do,  #
+        delta,  #
+        BATCH, N_HEAD, N_CTX,  #
+        BLOCK_M=PRE_BLOCK, HEAD_DIM=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,
+        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,  #
+        HEAD_DIM=D  #
+    )
+
+    return dq, dk, dv
+
+

csrc/fastvideo_kernel/src/fastvideo_kernel/triton_kernels/index.py

@@ -0,0 +1,152 @@
+
+## 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,
+    index_ptr,
+    map_bs_stride,
+    map_h_stride,
+    map_q_stride,
+    map_kv_stride,
+    index_bs_stride,
+    index_h_stride,
+    index_q_stride,
+    index_kv_stride,
+    topk,
+):
+    b, h, q = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    index_ptr_base = index_ptr + b * index_bs_stride + h * index_h_stride + q * index_q_stride
+    map_ptr_base = map_ptr + b * map_bs_stride + h * map_h_stride + q * map_q_stride
+
+    for i in tl.static_range(topk):
+        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,
+    index_ptr,
+    index_num_ptr,
+    map_bs_stride,
+    map_h_stride,
+    map_q_stride,
+    map_kv_stride,
+    index_bs_stride,
+    index_h_stride,
+    index_q_stride,
+    index_kv_stride,
+    index_num_bs_stride,
+    index_num_h_stride,
+    index_num_q_stride,
+    num_kv_blocks,
+):
+    b, h, q = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    index_ptr_base = index_ptr + b * index_bs_stride + h * index_h_stride + q * index_q_stride
+    map_ptr_base = map_ptr + b * map_bs_stride + h * map_h_stride + q * map_q_stride
+
+    num = 0
+    for i in tl.range(num_kv_blocks):
+        map_entry = tl.load(map_ptr_base + i * map_kv_stride)
+        if map_entry:
+            tl.store(index_ptr_base + num * index_kv_stride, i)
+            num += 1
+
+    tl.store(
+        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):
+    """
+    Convert topk indices to a map.
+    
+    Args:
+        index: [bs, h, num_q_blocks, topk]
+            The topk indices tensor.
+        num_kv_blocks: int
+            The number of key-value blocks in the block_map returned
+        transpose_map: bool
+            If True, the block_map will be transposed on the final two dimensions.
+    
+    Returns:
+        block_map: [bs, h, num_q_blocks, num_kv_blocks]
+            A binary map where 1 indicates that the q block attends to the kv block.
+    """
+    bs, h, num_q_blocks, topk = index.shape
+
+    if transpose_map is False:
+        block_map = torch.zeros((bs, h, num_q_blocks, num_kv_blocks),
+                                dtype=torch.bool,
+                                device=index.device)
+    else:
+        block_map = torch.zeros((bs, h, num_kv_blocks, num_q_blocks),
+                                dtype=torch.bool,
+                                device=index.device)
+        block_map = block_map.transpose(2, 3)
+
+    grid = (bs, h, num_q_blocks)
+    topk_index_to_map_kernel[grid](
+        block_map,
+        index,
+        block_map.stride(0),
+        block_map.stride(1),
+        block_map.stride(2),
+        block_map.stride(3),
+        index.stride(0),
+        index.stride(1),
+        index.stride(2),
+        index.stride(3),
+        topk=topk,
+    )
+
+    return block_map
+
+def map_to_index(block_map: torch.Tensor):
+    """
+    Convert a block map to indices and counts.
+    
+    Args:
+        block_map: [bs, h, num_q_blocks, num_kv_blocks]
+            The block map tensor.
+    
+    Returns:
+        index: [bs, h, num_q_blocks, num_kv_blocks]
+            The indices of the blocks.
+        index_num: [bs, h, num_q_blocks]
+            The number of blocks for each q block.
+    """
+    bs, h, num_q_blocks, num_kv_blocks = block_map.shape
+
+    index = torch.full((block_map.shape),
+                       -1,
+                       dtype=torch.int32,
+                       device=block_map.device)
+    index_num = torch.empty((bs, h, num_q_blocks),
+                            dtype=torch.int32,
+                            device=block_map.device)
+
+    grid = (bs, h, num_q_blocks)
+    map_to_index_kernel[grid](
+        block_map,
+        index,
+        index_num,
+        block_map.stride(0),
+        block_map.stride(1),
+        block_map.stride(2),
+        block_map.stride(3),
+        index.stride(0),
+        index.stride(1),
+        index.stride(2),
+        index.stride(3),
+        index_num.stride(0),
+        index_num.stride(1),
+        index_num.stride(2),
+        num_kv_blocks=num_kv_blocks,
+    )
+
+    return index, index_num

csrc/fastvideo_kernel/src/fastvideo_kernel/vmoba.py

@@ -0,0 +1,868 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapt from https://github.com/KwaiVGI/VMoBA/blob/main/src/vmoba.py
+
+import random
+import time
+import os
+import torch
+from typing import Tuple
+try:
+    from flash_attn import flash_attn_varlen_func  # Use the new flash attention function
+    from flash_attn.flash_attn_interface import _flash_attn_varlen_forward, _flash_attn_varlen_backward
+except ImportError:
+    def _unsupported(*args, **kwargs):
+        raise ImportError("flash-attn is not installed. Please install it, e.g., `pip install flash-attn`.")
+    _flash_attn_varlen_forward = _unsupported
+    _flash_attn_varlen_backward = _unsupported
+    flash_attn_varlen_func = _unsupported
+    
+from functools import lru_cache
+from einops import rearrange
+
+
+@lru_cache(maxsize=16)
+def calc_chunks(cu_seqlen, moba_chunk_size):
+    """
+    Calculate chunk boundaries.
+
+    For vision tasks we include all chunks (even the last one which might be shorter)
+    so that every chunk can be selected.
+    """
+    batch_sizes = cu_seqlen[1:] - cu_seqlen[:-1]
+    batch_num_chunk = (batch_sizes + (moba_chunk_size - 1)) // moba_chunk_size
+    cu_num_chunk = torch.ones(
+        batch_num_chunk.numel() + 1,
+        device=cu_seqlen.device,
+        dtype=batch_num_chunk.dtype,
+    )
+    cu_num_chunk[1:] = batch_num_chunk.cumsum(dim=0)
+    num_chunk = cu_num_chunk[-1]
+    chunk_sizes = torch.full(
+        (num_chunk + 1,), moba_chunk_size, dtype=torch.int32, device=cu_seqlen.device
+    )
+    chunk_sizes[0] = 0
+    batch_last_chunk_size = batch_sizes - (batch_num_chunk - 1) * moba_chunk_size
+    chunk_sizes[cu_num_chunk[1:]] = batch_last_chunk_size
+    cu_chunk = chunk_sizes.cumsum(dim=-1, dtype=torch.int32)
+    chunk_to_batch = torch.zeros(
+        (num_chunk,), dtype=torch.int32, device=cu_seqlen.device
+    )
+    chunk_to_batch[cu_num_chunk[1:-1]] = 1
+    chunk_to_batch = chunk_to_batch.cumsum(dim=0, dtype=torch.int32)
+
+    # Do not filter out any chunk
+    filtered_chunk_indices = torch.arange(
+        num_chunk, device=cu_seqlen.device, dtype=torch.int32
+    )
+    num_filtered_chunk = num_chunk
+
+    return cu_chunk, filtered_chunk_indices, num_filtered_chunk, chunk_to_batch
+
+
+# --- Threshold Selection Helper Functions ---
+
+def _select_threshold_query_head(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects chunks for each <query, head> pair based on threshold.
+    Normalization and sorting happen along the chunk dimension (dim=0).
+    """
+    C, H, S = gate.shape
+    eps = 1e-6
+
+    # LSE‐style normalization per <head, query> (across chunks)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    row_min = gate_min_val.amin(dim=0)        # (H, S)
+    row_max = gate_masked.amax(dim=0)         # (H, S)
+    denom = row_max - row_min
+    denom = torch.where(denom <= eps, torch.ones_like(denom), denom) # avoid divide‑by‑zero
+
+    gate_norm = (gate - row_min.unsqueeze(0)) / denom.unsqueeze(0)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) pull out the self‐chunk’s normalized weight for each <head,seq>
+    self_norm = (gate_norm * gate_self_chunk_mask).sum(dim=0)  # (H, S)
+
+    # 2) compute how much more normalized weight we need beyond self
+    total_norm_sum = gate_norm.sum(dim=0) # (H, S)
+    remain_ratio = simsum_threshold - self_norm / (total_norm_sum + eps)  # (H, S)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # if already ≥ thresh, no extra needed
+
+    # 3) zero out the self‐chunk in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0
+
+    # 4) sort the other chunks by descending norm, per <head,seq>
+    sorted_norm, sorted_idx = torch.sort(others_norm, descending=True, dim=0) # (C, H, S)
+
+    # 5) cumulative‑sum the sorted norms per <head,seq>
+    cumsum_others = sorted_norm.cumsum(dim=0) # (C, H, S)
+
+    # 6) for each <head,seq>, find the smallest k where cumsum_ratio ≥ remain_ratio
+    ratio = cumsum_others / (total_norm_sum.unsqueeze(0) + eps) # (C, H, S)
+    cond = ratio >= remain_ratio.unsqueeze(0) # (C, H, S) boolean mask
+    any_cond = cond.any(dim=0) # (H, S)
+    # Find the index of the first True value along dim 0. If none, use C-1.
+    cutoff = torch.where(any_cond, cond.float().argmax(dim=0), torch.full_like(any_cond, fill_value=C - 1)) # (H, S)
+
+    # 7) build a mask in sorted order up to that cutoff
+    idx_range = torch.arange(C, device=gate.device).view(-1, 1, 1) # (C, 1, 1)
+    sorted_mask = idx_range <= cutoff.unsqueeze(0) # (C, H, S)
+
+    # 8) scatter it back to original chunk order
+    others_mask = torch.zeros_like(gate, dtype=torch.bool)
+    others_mask.scatter_(0, sorted_idx, sorted_mask)
+
+    # 9) finally, include every self‐chunk plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_block(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <query, head> pairs for each block based on threshold.
+    Normalization and sorting happen across the head and sequence dimensions (dim=1, 2).
+    """
+    C, H, S = gate.shape
+    HS = H * S
+    eps = 1e-6
+
+    # LSE‐style normalization per block (across heads and queries)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    block_max = gate_masked.amax(dim=(1, 2), keepdim=True) # (C, 1, 1)
+    block_min = gate_min_val.amin(dim=(1, 2), keepdim=True) # (C, 1, 1)
+    block_denom = block_max - block_min
+    block_denom = torch.where(block_denom <= eps, torch.ones_like(block_denom), block_denom) # (C, 1, 1)
+
+    gate_norm = (gate - block_min) / block_denom # (C, H, S)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) identify normalized weights of entries that *are* self-chunks (from query perspective)
+    self_norm_entries = gate_norm * gate_self_chunk_mask # (C, H, S)
+    # Sum these weights *per block*
+    self_norm_sum_per_block = self_norm_entries.sum(dim=(1, 2)) # (C,)
+
+    # 2) compute how much more normalized weight each block needs beyond its self-chunk contributions
+    total_norm_sum_per_block = gate_norm.sum(dim=(1, 2)) # (C,)
+    remain_ratio = simsum_threshold - self_norm_sum_per_block / (total_norm_sum_per_block + eps) # (C,)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # (C,)
+
+    # 3) zero out the self‐chunk entries in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0 # Zero out self entries
+
+    # 4) sort the other <head, seq> pairs by descending norm, per block
+    others_flat = others_norm.contiguous().view(C, HS) # (C, H*S)
+    sorted_others_flat, sorted_indices_flat = torch.sort(others_flat, dim=1, descending=True) # (C, H*S)
+
+    # 5) cumulative‑sum the sorted norms per block
+    cumsum_others_flat = sorted_others_flat.cumsum(dim=1) # (C, H*S)
+
+    # 6) for each block, find the smallest k where cumsum_ratio ≥ remain_ratio
+    ratio_flat = cumsum_others_flat / (total_norm_sum_per_block.unsqueeze(1) + eps) # (C, H*S)
+    cond_flat = ratio_flat >= remain_ratio.unsqueeze(1) # (C, H*S) boolean mask
+    any_cond = cond_flat.any(dim=1) # (C,)
+    # Find the index of the first True value along dim 1. If none, use HS-1.
+    cutoff_flat = torch.where(any_cond, cond_flat.float().argmax(dim=1), torch.full_like(any_cond, fill_value=HS - 1)) # (C,)
+
+    # 7) build a mask in sorted order up to that cutoff per block
+    idx_range_flat = torch.arange(HS, device=gate.device).unsqueeze(0) # (1, H*S)
+    sorted_mask_flat = idx_range_flat <= cutoff_flat.unsqueeze(1) # (C, H*S)
+
+    # 8) scatter it back to original <head, seq> order per block
+    others_mask_flat = torch.zeros_like(others_flat, dtype=torch.bool) # (C, H*S)
+    others_mask_flat.scatter_(1, sorted_indices_flat, sorted_mask_flat)
+    others_mask = others_mask_flat.view(C, H, S) # (C, H, S)
+
+    # 9) finally, include every self‐chunk entry plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_overall(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <chunk, query, head> triplets globally based on threshold.
+    Normalization and sorting happen across all valid entries.
+    """
+    C, H, S = gate.shape
+    CHS = C * H * S
+    eps = 1e-6
+
+    # LSE‐style normalization globally across all valid entries
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    overall_max = gate_masked.max() # scalar
+    overall_min = gate_min_val.min() # scalar
+    overall_denom = overall_max - overall_min
+    overall_denom = torch.where(overall_denom <= eps, torch.tensor(1.0, device=gate.device, dtype=gate.dtype), overall_denom)
+
+    gate_norm = (gate - overall_min) / overall_denom # (C, H, S)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) identify normalized weights of entries that *are* self-chunks
+    self_norm_entries = gate_norm * gate_self_chunk_mask # (C, H, S)
+    # Sum these weights globally
+    self_norm_sum_overall = self_norm_entries.sum() # scalar
+
+    # 2) compute how much more normalized weight is needed globally beyond self-chunk contributions
+    total_norm_sum_overall = gate_norm.sum() # scalar
+    remain_ratio = simsum_threshold - self_norm_sum_overall / (total_norm_sum_overall + eps) # scalar
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # scalar
+
+    # 3) zero out the self‐chunk entries in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0 # Zero out self entries
+
+    # 4) sort all other entries by descending norm, globally
+    others_flat = others_norm.flatten() # (C*H*S,)
+    valid_others_mask_flat = valid_gate_mask.flatten() & ~gate_self_chunk_mask.flatten() # Mask for valid, non-self entries
+    
+    # Only sort the valid 'other' entries
+    valid_others_indices = torch.where(valid_others_mask_flat)[0]
+    valid_others_values = others_flat[valid_others_indices]
+
+    sorted_others_values, sort_perm = torch.sort(valid_others_values, descending=True) # (N_valid_others,)
+    sorted_original_indices = valid_others_indices[sort_perm] # Original indices in C*H*S space, sorted by value
+
+    # 5) cumulative‑sum the sorted valid 'other' norms globally
+    cumsum_others_values = sorted_others_values.cumsum(dim=0) # (N_valid_others,)
+
+    # 6) find the smallest k where cumsum_ratio ≥ remain_ratio globally
+    ratio_values = cumsum_others_values / (total_norm_sum_overall + eps) # (N_valid_others,)
+    cond_values = ratio_values >= remain_ratio # (N_valid_others,) boolean mask
+    any_cond = cond_values.any() # scalar
+    
+    # Find the index of the first True value in the *sorted* list. If none, use all valid others.
+    cutoff_idx_in_sorted = torch.where(
+        any_cond, 
+        cond_values.float().argmax(dim=0), 
+        torch.tensor(len(sorted_others_values) - 1, device=gate.device, dtype=torch.long)
+    )
+
+    # 7) build a mask selecting the top-k others based on the cutoff
+    # Select the original indices corresponding to the top entries in the sorted list
+    selected_other_indices = sorted_original_indices[:cutoff_idx_in_sorted + 1]
+
+    # 8) create the mask in the original flat shape
+    others_mask_flat = torch.zeros_like(others_flat, dtype=torch.bool) # (C*H*S,)
+    if selected_other_indices.numel() > 0: # Check if any 'other' indices were selected
+        others_mask_flat[selected_other_indices] = True
+    others_mask = others_mask_flat.view(C, H, S) # (C, H, S)
+
+    # 9) finally, include every self‐chunk entry plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_head_global(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <chunk, query> globally for each head based on threshold.
+    """
+    C, H, S = gate.shape
+    eps = 1e-6
+
+    # 1) LSE‐style normalization per head (across chunks and sequence dims)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf)
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf)
+
+    max_per_head = gate_masked.amax(dim=(0, 2), keepdim=True)  # (1, H, 1)
+    min_per_head = gate_min_val.amin(dim=(0, 2), keepdim=True)  # (1, H, 1)
+    denom = max_per_head - min_per_head
+    denom = torch.where(denom <= eps, torch.ones_like(denom), denom)
+
+    gate_norm = (gate - min_per_head) / denom
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0)  # (C, H, S)
+
+    # 2) sum normalized self‐chunk contributions per head
+    self_norm_sum = (gate_norm * gate_self_chunk_mask).sum(dim=(0, 2))  # (H,)
+
+    # 3) total normalized sum per head
+    total_norm_sum = gate_norm.sum(dim=(0, 2))  # (H,)
+
+    # 4) how much more normalized weight needed per head
+    remain_ratio = simsum_threshold - self_norm_sum / (total_norm_sum + eps)  # (H,)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0)
+
+    # 5) zero out self‐chunk entries to focus on "others"
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0  # (C, H, S)
+
+    # 6) flatten chunk and sequence dims, per head
+    CS = C * S
+    others_flat = others_norm.permute(1, 0, 2).reshape(H, CS)  # (H, C*S)
+    valid_flat  = (valid_gate_mask & ~gate_self_chunk_mask) \
+                   .permute(1, 0, 2).reshape(H, CS)           # (H, C*S)
+
+    # 7) vectorized selection of “others” per head
+    masked_flat = torch.where(valid_flat, others_flat, torch.zeros_like(others_flat))
+    sorted_vals, sorted_idx = torch.sort(masked_flat, dim=1, descending=True)  # (H, C*S)
+
+    cumsum_vals = sorted_vals.cumsum(dim=1)                                    # (H, C*S)
+    ratio_vals  = cumsum_vals / (total_norm_sum.unsqueeze(1) + eps)            # (H, C*S)
+    cond        = ratio_vals >= remain_ratio.unsqueeze(1)                      # (H, C*S)
+
+    has_cutoff = cond.any(dim=1)                                               # (H,)
+    default    = torch.full((H,), CS - 1, device=gate.device, dtype=torch.long)
+    cutoff     = torch.where(has_cutoff, cond.float().argmax(dim=1), default)  # (H,)
+
+    idx_range   = torch.arange(CS, device=gate.device).unsqueeze(0)            # (1, C*S)
+    sorted_mask = idx_range <= cutoff.unsqueeze(1)                             # (H, C*S)
+
+    selected_flat = torch.zeros_like(valid_flat)                               # (H, C*S)
+    selected_flat.scatter_(1, sorted_idx, sorted_mask)                         # (H, C*S)
+
+    # 8) reshape selection mask back to (C, H, S)
+    others_mask = selected_flat.reshape(H, C, S).permute(1, 0, 2)               # (C, H, S)
+
+    # 9) include self‐chunks plus selected others, and obey valid mask
+    final_gate_mask = valid_gate_mask & (gate_self_chunk_mask | others_mask)
+
+    return final_gate_mask
+
+
+class MixedAttention(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        q,
+        k,
+        v,
+        self_attn_cu_seqlen,
+        moba_q,
+        moba_kv,
+        moba_cu_seqlen_q,
+        moba_cu_seqlen_kv,
+        max_seqlen,
+        moba_chunk_size,
+        moba_q_sh_indices,
+    ):
+        ctx.max_seqlen = max_seqlen
+        ctx.moba_chunk_size = moba_chunk_size
+        ctx.softmax_scale = softmax_scale = q.shape[-1] ** (-0.5)
+
+        # Non-causal self-attention branch
+        # return out, softmax_lse, S_dmask, rng_state
+        self_attn_out_sh, self_attn_lse_hs, _, _ = _flash_attn_varlen_forward(
+            q=q,
+            k=k,
+            v=v,
+            cu_seqlens_q=self_attn_cu_seqlen,
+            cu_seqlens_k=self_attn_cu_seqlen,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+        )
+        # MOBA attention branch (non-causal)
+        moba_attn_out, moba_attn_lse_hs, _, _ = _flash_attn_varlen_forward(
+            q=moba_q,
+            k=moba_kv[:, 0],
+            v=moba_kv[:, 1],
+            cu_seqlens_q=moba_cu_seqlen_q,
+            cu_seqlens_k=moba_cu_seqlen_kv,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=moba_chunk_size,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+        )
+
+        self_attn_lse_sh = self_attn_lse_hs.t().contiguous()
+        moba_attn_lse = moba_attn_lse_hs.t().contiguous()
+
+        output = torch.zeros((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
+        output_2d = output.view(-1, q.shape[2])
+
+        max_lse_1d = self_attn_lse_sh.view(-1)
+        max_lse_1d = max_lse_1d.index_reduce(
+            0, moba_q_sh_indices, moba_attn_lse.view(-1), "amax"
+        )
+        self_attn_lse_sh = self_attn_lse_sh - max_lse_1d.view_as(self_attn_lse_sh)
+        moba_attn_lse = (
+            moba_attn_lse.view(-1)
+            .sub(max_lse_1d.index_select(0, moba_q_sh_indices))
+            .reshape_as(moba_attn_lse)
+        )
+
+        mixed_attn_se_sh = self_attn_lse_sh.exp()
+        moba_attn_se = moba_attn_lse.exp()
+
+        mixed_attn_se_sh.view(-1).index_add_(
+            0, moba_q_sh_indices, moba_attn_se.view(-1)
+        )
+        mixed_attn_lse_sh = mixed_attn_se_sh.log()
+
+        # Combine self-attention output
+        factor = (self_attn_lse_sh - mixed_attn_lse_sh).exp()  # [S, H]
+        self_attn_out_sh = self_attn_out_sh * factor.unsqueeze(-1)
+        output_2d += self_attn_out_sh.reshape_as(output_2d)
+
+        # Combine MOBA attention output
+        mixed_attn_lse = (
+            mixed_attn_lse_sh.view(-1)
+            .index_select(0, moba_q_sh_indices)
+            .view_as(moba_attn_lse)
+        )
+        factor = (moba_attn_lse - mixed_attn_lse).exp()  # [S, H]
+        moba_attn_out = moba_attn_out * factor.unsqueeze(-1)
+        raw_attn_out = moba_attn_out.view(-1, moba_attn_out.shape[-1])
+        output_2d.index_add_(0, moba_q_sh_indices, raw_attn_out)
+        output = output.to(q.dtype)
+        mixed_attn_lse_sh = mixed_attn_lse_sh + max_lse_1d.view_as(mixed_attn_se_sh)
+        ctx.save_for_backward(
+            output,
+            mixed_attn_lse_sh,
+            q,
+            k,
+            v,
+            self_attn_cu_seqlen,
+            moba_q,
+            moba_kv,
+            moba_cu_seqlen_q,
+            moba_cu_seqlen_kv,
+            moba_q_sh_indices,
+        )
+
+        return output
+
+    @staticmethod
+    def backward(ctx, d_output):
+
+        max_seqlen = ctx.max_seqlen
+        moba_chunk_size = ctx.moba_chunk_size
+        softmax_scale = ctx.softmax_scale
+
+        (
+            output,
+            mixed_attn_vlse_sh,
+            q,
+            k,
+            v,
+            self_attn_cu_seqlen,
+            moba_q,
+            moba_kv,
+            moba_cu_seqlen_q,
+            moba_cu_seqlen_kv,
+            moba_q_sh_indices,
+        ) = ctx.saved_tensors
+
+        d_output = d_output.contiguous()
+
+        dq = torch.empty_like(q)
+        dk = torch.empty_like(k)
+        dv = torch.empty_like(v)
+        _ = _flash_attn_varlen_backward(
+            dout=d_output,
+            q=q,
+            k=k,
+            v=v,
+            out=output,
+            softmax_lse=mixed_attn_vlse_sh.t().contiguous(),
+            dq=dq,
+            dk=dk,
+            dv=dv,
+            cu_seqlens_q=self_attn_cu_seqlen,
+            cu_seqlens_k=self_attn_cu_seqlen,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=True,
+            window_size_left=-1,
+            window_size_right=-1
+        )
+
+        headdim = q.shape[-1]
+        d_moba_output = (
+            d_output.view(-1, headdim).index_select(0, moba_q_sh_indices).unsqueeze(1)
+        )
+        moba_output = (
+            output.view(-1, headdim).index_select(0, moba_q_sh_indices).unsqueeze(1)
+        )
+
+        mixed_attn_vlse = (
+            mixed_attn_vlse_sh.view(-1).index_select(0, moba_q_sh_indices).view(1, -1)
+        )
+
+        dmq = torch.empty_like(moba_q)
+        dmkv = torch.empty_like(moba_kv)
+        _ = _flash_attn_varlen_backward(
+            dout=d_moba_output,
+            q=moba_q,
+            k=moba_kv[:, 0],
+            v=moba_kv[:, 1],
+            out=moba_output,
+            softmax_lse=mixed_attn_vlse,
+            dq=dmq,
+            dk=dmkv[:,0],
+            dv=dmkv[:,1],
+            cu_seqlens_q=moba_cu_seqlen_q,
+            cu_seqlens_k=moba_cu_seqlen_kv,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=moba_chunk_size,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=True,
+            window_size_left=-1,
+            window_size_right=-1
+        )
+
+        return dq, dk, dv, None, dmq, dmkv, None, None, None, None, None
+
+
+def moba_attn_varlen(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    cu_seqlens: torch.Tensor,
+    max_seqlen: int,
+    moba_chunk_size: int,
+    moba_topk: int,
+    select_mode: str = 'threshold', # "topk" or "threshold"
+    simsum_threshold: float = 0.25,
+    threshold_type: str = 'query_head',
+) -> torch.Tensor:
+    """
+    Accelerated MOBA attention for vision tasks with proper LSE normalization.
+
+    This version:
+      - Splits KV into chunks.
+      - For each query head, selects the top-k relevant KV chunks (including the self chunk)
+        by amplifying the diagonal (self-chunk) logits.
+      - Aggregates the attention outputs from the selected chunks using a log-sum-exp
+        reduction so that attending to each query over the selected chunks is equivalent
+        to the original algorithm.
+    """
+    # Stack keys and values.
+    kv = torch.stack((k, v), dim=1)
+    seqlen, num_head, head_dim = q.shape
+
+    # Compute chunk boundaries.
+    cu_chunk, filtered_chunk_indices, num_filtered_chunk, chunk_to_batch = calc_chunks(
+        cu_seqlens, moba_chunk_size
+    )
+
+    self_attn_cu_seqlen = cu_chunk
+
+    # Update top-k selection to include the self chunk.
+    moba_topk = min(moba_topk, num_filtered_chunk)
+
+    # --- Build filtered KV from chunks ---
+    chunk_starts = cu_chunk[filtered_chunk_indices]          # [num_filtered_chunk]
+    chunk_ends = cu_chunk[filtered_chunk_indices + 1]          # [num_filtered_chunk]
+    chunk_lengths = chunk_ends - chunk_starts                  # [num_filtered_chunk]
+    max_chunk_len = int(chunk_lengths.max().item())
+
+    range_tensor = torch.arange(max_chunk_len, device=kv.device, dtype=chunk_starts.dtype).unsqueeze(0)
+    indices = chunk_starts.unsqueeze(1) + range_tensor
+    indices = torch.clamp(indices, max=kv.shape[0] - 1)
+    valid_mask = range_tensor < chunk_lengths.unsqueeze(1)
+    gathered = kv[indices.view(-1)].view(num_filtered_chunk, max_chunk_len, *kv.shape[1:])
+    gathered = gathered * valid_mask.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).type_as(gathered)
+
+    # Compute key_gate_weight over valid tokens.
+    key_values = gathered[:, :, 0].float()  # [num_filtered_chunk, max_chunk_len, num_head, head_dim]
+    valid_mask_exp = valid_mask.unsqueeze(-1).unsqueeze(-1)
+    key_sum = (key_values * valid_mask_exp).sum(dim=1)
+    divisor = valid_mask.sum(dim=1).unsqueeze(-1).unsqueeze(-1)
+    key_gate_weight = key_sum / divisor  # [num_filtered_chunk, num_head, head_dim]
+
+    # Compute gate logits between key_gate_weight and queries.
+    q_float = q.float()
+    # gate = torch.einsum("nhd,shd->nhs", key_gate_weight, q_float)  # [num_filtered_chunk, num_head, seqlen]
+    gate = torch.bmm(key_gate_weight.permute(1, 0, 2), q_float.permute(1, 0, 2).transpose(1, 2)).permute(1, 0, 2)
+
+    # Amplify the diagonal (self chunk) contributions.
+    gate_seq_idx = torch.arange(seqlen, device=q.device, dtype=torch.int32).unsqueeze(0).expand(num_filtered_chunk, seqlen)
+    chunk_start = cu_chunk[filtered_chunk_indices]  # [num_filtered_chunk]
+    chunk_end = cu_chunk[filtered_chunk_indices + 1]  # [num_filtered_chunk]
+    gate_self_chunk_mask = ((gate_seq_idx >= chunk_start.unsqueeze(1)) &
+                            (gate_seq_idx < chunk_end.unsqueeze(1))).unsqueeze(1).expand(-1, num_head, -1)
+    amplification_factor = 1e9  # Example factor; adjust as needed.
+    origin_gate = gate.clone()
+    gate = gate.clone()
+    if select_mode == "topk":
+        gate[gate_self_chunk_mask] += amplification_factor
+
+    # Exclude positions that are outside the valid batch boundaries.
+    batch_starts = cu_seqlens[chunk_to_batch[filtered_chunk_indices]]
+    batch_ends = cu_seqlens[chunk_to_batch[filtered_chunk_indices] + 1]
+    gate_batch_start_mask = gate_seq_idx < batch_starts.unsqueeze(1)
+    gate_batch_end_mask = gate_seq_idx >= batch_ends.unsqueeze(1)
+    gate_inf_mask = gate_batch_start_mask | gate_batch_end_mask
+    gate.masked_fill_(gate_inf_mask.unsqueeze(1), -float("inf"))
+
+    if select_mode == 'topk':
+        # We amplify self‐chunk in gate already, so self entries will rank highest.
+        valid_gate_mask = gate != -float("inf")
+        if threshold_type == 'query_head':
+            # === per‐<head,seq> top-k across chunks (original behavior) ===
+            # gate: (C, H, S)
+            _, gate_topk_idx = torch.topk(gate, k=moba_topk, dim=0, largest=True, sorted=False)
+            gate_idx_mask = torch.zeros_like(gate, dtype=torch.bool)
+            gate_idx_mask.scatter_(0, gate_topk_idx, True)
+            gate_mask = valid_gate_mask & gate_idx_mask
+        elif threshold_type == 'overall':
+            # === global top-k across all (chunk, head, seq) entries ===
+            C, H, S = gate.shape
+            flat_gate = gate.flatten()
+            flat_mask = valid_gate_mask.flatten()
+            flat_gate_masked = torch.where(flat_mask, flat_gate, -float("inf"))
+            # pick topk global entries
+            vals, idx = torch.topk(flat_gate_masked, k=moba_topk * H * S, largest=True, sorted=False)
+            others_mask_flat = torch.zeros_like(flat_mask, dtype=torch.bool)
+            others_mask_flat[idx] = True
+            gate_mask = (valid_gate_mask.flatten() & others_mask_flat).view(gate.shape)
+        elif threshold_type == 'head_global':
+            # per-head top-k across all chunks and sequence positions
+            C, H, S = gate.shape
+            CS = C * S
+            flat_gate  = gate.permute(1, 0, 2).reshape(H, CS)
+            flat_valid = valid_gate_mask.permute(1, 0, 2).reshape(H, CS)
+            flat_gate_masked = torch.where(flat_valid, flat_gate, torch.full_like(flat_gate, -float('inf')))
+            # pick top-k indices per head
+            _, topk_idx = torch.topk(flat_gate_masked, k=moba_topk * S, dim=1, largest=True, sorted=False)
+            gate_idx_flat = torch.zeros_like(flat_valid, dtype=torch.bool)
+            gate_idx_flat.scatter_(1, topk_idx, True)
+            gate_mask = gate_idx_flat.reshape(H, C, S).permute(1, 0, 2)
+        else:
+            raise ValueError(
+                f"Invalid threshold_type for topk: {threshold_type}. "
+                "Choose 'query_head', 'block', or 'overall'."
+            )
+    elif select_mode == 'threshold':
+        # Delegate to the specific thresholding function
+        valid_gate_mask = gate != -float("inf")  # (num_chunk, num_head, seqlen)
+        if threshold_type == 'query_head':
+            gate_mask = _select_threshold_query_head(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'block':
+            gate_mask = _select_threshold_block(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'overall':
+            gate_mask = _select_threshold_overall(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'head_global':
+            gate_mask = _select_threshold_head_global(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        else:
+            raise ValueError(f"Invalid threshold_type: {threshold_type}. Choose 'query_head', 'block', or 'overall'.")
+    else:
+        raise ValueError(f"Invalid select_mode: {select_mode}. Choose 'topk' or 'threshold'.")
+    
+    # eliminate self_chunk in MoBA branch
+    gate_mask = gate_mask & ~gate_self_chunk_mask
+    # if gate_mask is all false, perform flash_attn instead
+    if gate_mask.sum() == 0:
+        return flash_attn_varlen_func(
+            q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen, causal=False
+        )
+
+    # Determine which query positions are selected.
+    # nonzero_indices has shape [N, 3] where each row is [chunk_index, head_index, seq_index].
+    moba_q_indices = gate_mask.reshape(gate_mask.shape[0], -1).nonzero(as_tuple=True)[-1] # [(h s k)]
+    moba_q_sh_indices = (moba_q_indices % seqlen) * num_head + (moba_q_indices // seqlen)
+    moba_q = rearrange(q, "s h d -> (h s) d").index_select(0, moba_q_indices).unsqueeze(1)
+
+    # Build cumulative sequence lengths for the selected queries.
+    moba_seqlen_q = gate_mask.sum(dim=-1).flatten()
+    q_zero_mask = moba_seqlen_q == 0
+    valid_expert_mask = ~q_zero_mask
+    if q_zero_mask.sum() > 0:
+        moba_seqlen_q = moba_seqlen_q[valid_expert_mask]
+    moba_cu_seqlen_q = torch.cat(
+        (
+            torch.tensor([0], device=q.device, dtype=moba_seqlen_q.dtype),
+            moba_seqlen_q.cumsum(dim=0),
+        ),
+        dim=0,
+    ).to(torch.int32)
+
+    # Rearrange gathered KV for the MOBA branch.
+    experts_tensor = rearrange(gathered, "nc cl two h d -> (nc h) cl two d")
+    valid_expert_lengths = chunk_lengths.unsqueeze(1).expand(num_filtered_chunk, num_head).reshape(-1).to(torch.int32)
+    if q_zero_mask.sum() > 0:
+        experts_tensor = experts_tensor[valid_expert_mask]
+        valid_expert_lengths = valid_expert_lengths[valid_expert_mask]
+
+    seq_range = torch.arange(experts_tensor.shape[1], device=experts_tensor.device).unsqueeze(0)
+    mask = seq_range < valid_expert_lengths.unsqueeze(1)
+    moba_kv = experts_tensor[mask]  # Shape: ((nc h cl_valid) two d)
+    moba_kv = moba_kv.unsqueeze(2)   # Shape: ((nc h cl_valid) two 1 d)
+
+    moba_cu_seqlen_kv = torch.cat(
+        [torch.zeros(1, device=experts_tensor.device, dtype=torch.int32),
+         valid_expert_lengths.cumsum(dim=0)],
+        dim=0,
+    ).to(torch.int32)
+
+    assert (
+        moba_cu_seqlen_kv.shape == moba_cu_seqlen_q.shape
+    ), f"Mismatch between moba_cu_seqlen_kv.shape and moba_cu_seqlen_q.shape: {moba_cu_seqlen_kv.shape} vs {moba_cu_seqlen_q.shape}"
+
+    return MixedAttention.apply(
+        q,
+        k,
+        v,
+        self_attn_cu_seqlen,
+        moba_q,
+        moba_kv,
+        moba_cu_seqlen_q,
+        moba_cu_seqlen_kv,
+        max_seqlen,
+        moba_chunk_size,
+        moba_q_sh_indices,
+    )
+
+
+def process_moba_input(
+    x,
+    patch_resolution,
+    chunk_size,
+):
+    """
+    Process inputs for the attention function.
+
+    Args:
+        x (torch.Tensor): Input tensor with shape [batch_size, num_patches, num_heads, head_dim].
+        patch_resolution (tuple): Tuple containing the patch resolution (t, h, w).
+        chunk_size (int): Size of the chunk. (maybe tuple or int, according to chunk type)
+
+    Returns:
+        torch.Tensor: Processed input tensor.
+    """
+    if isinstance(chunk_size, float) or isinstance(chunk_size, int):
+        moba_chunk_size = int(chunk_size * patch_resolution[1] * patch_resolution[2])
+    else:
+        assert isinstance(chunk_size, (Tuple, list)), f"chunk_size should be a tuple, list, or int, now it is: {type(chunk_size)}"
+        if len(chunk_size) == 2:
+            assert patch_resolution[1] % chunk_size[0] == 0 and patch_resolution[2] % chunk_size[1] == 0, f"spatial patch_resolution {patch_resolution[1:]} should be divisible by 2d chunk_size {chunk_size}"
+            nch, ncw = patch_resolution[1] // chunk_size[0], patch_resolution[2] // chunk_size[1]
+            x = rearrange(x, "b (t nch ch ncw cw) n d -> b (nch ncw t ch cw) n d", t=patch_resolution[0], nch=nch, ncw=ncw, ch=chunk_size[0], cw=chunk_size[1])
+            moba_chunk_size = patch_resolution[0] * chunk_size[0] * chunk_size[1]
+        elif len(chunk_size) == 3:
+            assert patch_resolution[0] % chunk_size[0] == 0 and patch_resolution[1] % chunk_size[1] == 0 and patch_resolution[2] % chunk_size[2] == 0, f"patch_resolution {patch_resolution} should be divisible by 3d chunk_size {chunk_size}"
+            nct, nch, ncw = patch_resolution[0] // chunk_size[0], patch_resolution[1] // chunk_size[1], patch_resolution[2] // chunk_size[2]
+            x = rearrange(x, "b (nct ct nch ch ncw cw) n d -> b (nct nch ncw ct ch cw) n d", nct=nct, nch=nch, ncw=ncw, ct=chunk_size[0], ch=chunk_size[1], cw=chunk_size[2])
+            moba_chunk_size = chunk_size[0] * chunk_size[1] * chunk_size[2]
+        else:
+            raise ValueError(f"chunk_size should be a int, or a tuple of length 2 or 3, now it is: {len(chunk_size)}")
+        
+    return x, moba_chunk_size
+
+
+def process_moba_output(
+    x,
+    patch_resolution,
+    chunk_size,
+):
+    if isinstance(chunk_size, float) or isinstance(chunk_size, int):
+        pass
+    elif len(chunk_size) == 2:
+        x = rearrange(x, "b (nch ncw t ch cw) n d -> b (t nch ch ncw cw) n d", nch=patch_resolution[1] // chunk_size[0], ncw=patch_resolution[2] // chunk_size[1], t=patch_resolution[0], ch=chunk_size[0], cw=chunk_size[1])
+    elif len(chunk_size) == 3:
+        x = rearrange(x, "b (nct nch ncw ct ch cw) n d -> b (nct ct nch ch ncw cw) n d", nct=patch_resolution[0] // chunk_size[0], nch=patch_resolution[1] // chunk_size[1], ncw=patch_resolution[2] // chunk_size[2], ct=chunk_size[0], ch=chunk_size[1], cw=chunk_size[2])
+
+    return x
+
+
+# TEST
+def generate_data(batch_size, seqlen, num_head, head_dim, dtype):
+    random.seed(0)
+    torch.manual_seed(0)
+    torch.cuda.manual_seed(0)
+    device = torch.cuda.current_device()
+
+    q = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    k = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    v = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    print(f"q.shape: {q.shape}, k.shape: {k.shape}, v.shape: {v.shape}")
+    cu_seqlens = torch.arange(0, q.shape[0] * q.shape[1] + 1, q.shape[1], dtype=torch.int32, device='cuda')
+    max_seqlen = q.shape[1]
+    q = rearrange(q, "b s ... -> (b s) ...")
+    k = rearrange(k, "b s ... -> (b s) ...")
+    v = rearrange(v, "b s ... -> (b s) ...")
+
+    return q, k, v, cu_seqlens, max_seqlen
+
+
+def test_attn_varlen_moba_speed(batch, head, seqlen, head_dim, moba_chunk_size, moba_topk, dtype=torch.bfloat16, select_mode='threshold', simsum_threshold=0.25, threshold_type='query_head'):
+    """Speed test comparing flash_attn vs moba_attention"""
+    # Get data
+    q, k, v, cu_seqlen, max_seqlen = generate_data(batch, seqlen, head, head_dim, dtype)
+    print(f"batch:{batch} head:{head} seqlen:{seqlen} chunk:{moba_chunk_size} topk:{moba_topk} select_mode: {select_mode} simsum_threshold:{simsum_threshold}")
+    vo_grad = torch.randn_like(q)
+    
+    # Warmup
+    warmup_iters = 3
+    perf_test_iters = 10
+
+    # Warmup
+    for _ in range(warmup_iters):
+        o = flash_attn_varlen_func(q, k, v, cu_seqlen, cu_seqlen, max_seqlen, max_seqlen, causal=False)
+        torch.autograd.backward(o, vo_grad)
+    
+    torch.cuda.synchronize()
+    start_flash = time.perf_counter()
+    for _ in range(perf_test_iters):
+        o = flash_attn_varlen_func(q, k, v, cu_seqlen, cu_seqlen, max_seqlen, max_seqlen, causal=False)
+        torch.autograd.backward(o, vo_grad)
+        
+    torch.cuda.synchronize()
+    time_flash = (time.perf_counter() - start_flash) / perf_test_iters * 1000
+
+    # Warmup
+    for _ in range(warmup_iters):
+        om = moba_attn_varlen(q, k, v, cu_seqlen, max_seqlen, moba_chunk_size=moba_chunk_size, moba_topk=moba_topk, select_mode=select_mode, simsum_threshold=simsum_threshold, threshold_type=threshold_type)
+        torch.autograd.backward(om, vo_grad)
+
+    torch.cuda.synchronize()
+    start_moba = time.perf_counter()
+    for _ in range(perf_test_iters):
+        om = moba_attn_varlen(q, k, v, cu_seqlen, max_seqlen, moba_chunk_size=moba_chunk_size, moba_topk=moba_topk, select_mode=select_mode, simsum_threshold=simsum_threshold, threshold_type=threshold_type)
+        torch.autograd.backward(om, vo_grad)
+    
+    torch.cuda.synchronize()
+    time_moba = (time.perf_counter() - start_moba) / perf_test_iters * 1000
+    
+    print(f"Flash: {time_flash:.2f}ms, MoBA: {time_moba:.2f}ms")
+    print(f"Speedup:  {time_flash / time_moba:.2f}x")
+
+
+
+if __name__ == "__main__":
+    """
+    CUDA_VISIBLE_DEVICES=1 \
+    python -u csrc/attn/vmoba_attn/vmoba/vmoba.py
+    """
+    test_attn_varlen_moba_speed(batch=1, head=12, seqlen=32760, head_dim=128, moba_chunk_size=32760 // 3 // 6 // 4, moba_topk=3, select_mode='threshold', simsum_threshold=0.3, threshold_type='query_head')

csrc/fastvideo_kernel/tests/support_flex_sta.py

@@ -0,0 +1,71 @@
+from typing import Tuple
+
+import torch
+from torch import BoolTensor, IntTensor
+from torch.nn.attention.flex_attention import create_block_mask
+
+# Peiyuan: This is neccesay. Dont know why. see https://github.com/pytorch/pytorch/issues/135028
+torch._inductor.config.realize_opcount_threshold = 100
+
+
+def generate_sta_mask(canvas_twh, kernel_twh, tile_twh, text_length):
+    """Generates a 3D NATTEN attention mask with a given kernel size.
+    
+    Args:
+        canvas_t: The time dimension of the canvas.
+        canvas_h: The height of the canvas.
+        canvas_w: The width of the canvas.
+        kernel_t: The time dimension of the kernel.
+        kernel_h: The height of the kernel.
+        kernel_w: The width of the kernel.
+    """
+    canvas_t, canvas_h, canvas_w = canvas_twh
+    kernel_t, kernel_h, kernel_w = kernel_twh
+    tile_t_size, tile_h_size, tile_w_size = tile_twh
+    total_tile_size = tile_t_size * tile_h_size * tile_w_size
+    canvas_tile_t, canvas_tile_h, canvas_tile_w = canvas_t // tile_t_size, canvas_h // tile_h_size, canvas_w // tile_w_size
+    img_seq_len = canvas_t * canvas_h * canvas_w
+
+    def get_tile_t_x_y(idx: IntTensor) -> Tuple[IntTensor, IntTensor, IntTensor]:
+        tile_id = idx // total_tile_size
+        tile_t = tile_id // (canvas_tile_h * canvas_tile_w)
+        tile_h = (tile_id % (canvas_tile_h * canvas_tile_w)) // canvas_tile_w
+        tile_w = tile_id % canvas_tile_w
+        return tile_t, tile_h, tile_w
+
+    def sta_mask_3d(
+        b: IntTensor,
+        h: IntTensor,
+        q_idx: IntTensor,
+        kv_idx: IntTensor,
+    ) -> BoolTensor:
+        q_t_tile, q_x_tile, q_y_tile = get_tile_t_x_y(q_idx)
+        kv_t_tile, kv_x_tile, kv_y_tile = get_tile_t_x_y(kv_idx)
+        # kernel nominally attempts to center itself on the query, but kernel center
+        # is clamped to a fixed distance (kernel half-length) from the canvas edge
+        kernel_center_t = q_t_tile.clamp(kernel_t // 2, (canvas_tile_t - 1) - kernel_t // 2)
+        kernel_center_x = q_x_tile.clamp(kernel_h // 2, (canvas_tile_h - 1) - kernel_h // 2)
+        kernel_center_y = q_y_tile.clamp(kernel_w // 2, (canvas_tile_w - 1) - kernel_w // 2)
+        time_mask = (kernel_center_t - kv_t_tile).abs() <= kernel_t // 2
+        hori_mask = (kernel_center_x - kv_x_tile).abs() <= kernel_h // 2
+        vert_mask = (kernel_center_y - kv_y_tile).abs() <= kernel_w // 2
+        image_mask = (q_idx < img_seq_len) & (kv_idx < img_seq_len)
+        image_to_text_mask = (q_idx < img_seq_len) & (kv_idx >= img_seq_len) & (kv_idx < img_seq_len + text_length)
+        text_to_all_mask = (q_idx >= img_seq_len) & (kv_idx < img_seq_len + text_length)
+        return (image_mask & time_mask & hori_mask & vert_mask) | image_to_text_mask | text_to_all_mask
+
+    sta_mask_3d.__name__ = f"natten_3d_c{canvas_t}x{canvas_w}x{canvas_h}_k{kernel_t}x{kernel_w}x{kernel_h}"
+    return sta_mask_3d
+
+
+def get_sliding_tile_attention_mask(kernel_size, tile_size, img_size, text_length, device, text_max_len=256):
+    img_seq_len = img_size[0] * img_size[1] * img_size[2]
+    image_mask = generate_sta_mask(img_size, kernel_size, tile_size, text_length)
+    mask = create_block_mask(image_mask,
+                             B=None,
+                             H=None,
+                             Q_LEN=img_seq_len + text_max_len,
+                             KV_LEN=img_seq_len + text_max_len,
+                             device=device,
+                             _compile=True)
+    return mask

csrc/fastvideo_kernel/tests/test_correctness.py

@@ -0,0 +1,63 @@
+import torch
+import sys
+import os
+from tqdm import tqdm
+
+# Local support import
+from .support_flex_sta import get_sliding_tile_attention_mask
+
+# USE OUR NEW PACKAGE!
+from fastvideo_kernel import sliding_tile_attention
+from torch.nn.attention.flex_attention import flex_attention
+
+flex_attention = torch.compile(flex_attention, dynamic=False)
+
+def flex_test(Q, K, V, kernel_size):
+    mask = get_sliding_tile_attention_mask(kernel_size, (6, 8, 8), (18, 48, 80), 0, 'cuda', 0)
+    output = flex_attention(Q, K, V, block_mask=mask)
+    return output
+
+def h100_fwd_kernel_test(Q, K, V, kernel_size):
+    # Using the same parameters as the original test
+    o = sliding_tile_attention(Q, K, V, [kernel_size] * 24, 0, False, '18x48x80')
+    return o
+
+def generate_tensor(shape, mean, std, dtype, device):
+    tensor = torch.randn(shape, dtype=dtype, device=device)
+    magnitude = torch.norm(tensor, dim=-1, keepdim=True)
+    scaled_tensor = tensor * (torch.randn(magnitude.shape, dtype=dtype, device=device) * std + mean) / magnitude
+    return scaled_tensor.contiguous()
+
+def check_correctness(b, h, n, d, causal, mean, std, num_iterations=2):
+    print(f"Running correctness check: batch={b}, heads={h}, seq_len={n}, dim={d}")
+    kernel_size_ls = [(3, 3, 5), (3, 1, 10)]
+    
+    for kernel_size in kernel_size_ls:
+        print(f"Testing kernel_size: {kernel_size}")
+        for xi in tqdm(range(num_iterations)):
+            torch.manual_seed(xi)
+            Q = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+            K = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+            V = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+            
+            tk_o = h100_fwd_kernel_test(Q, K, V, kernel_size)
+            pt_o = flex_test(Q, K, V, kernel_size)
+
+            diff = pt_o - tk_o
+            abs_diff = torch.abs(diff)
+            max_d = torch.max(abs_diff).item()
+            avg_d = torch.sum(abs_diff).item() / (b * h * n * d)
+            
+            if max_d > 0.1:
+                print(f"Warning: Large diff detected! max={max_d}, avg={avg_d}")
+
+    print("\n✅ TEST COMPLETE: New package matches FlexAttention behavior.")
+
+if __name__ == "__main__":
+    b, h, d = 2, 24, 128
+    n = 69120 
+    causal = False
+    mean = 1e-1
+    std = 10
+    
+    check_correctness(b, h, n, d, causal, mean, std, num_iterations=2)

csrc/fastvideo_kernel/tests/test_vmoba_correctness.py

@@ -0,0 +1,97 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import pytest
+import random
+from fastvideo_kernel.vmoba import moba_attn_varlen
+
+def generate_test_data(batch_size, total_seqlen, num_heads, head_dim, dtype, device="cuda"):
+    """
+    Generates random data for testing the variable-length attention function.
+    """
+    torch.manual_seed(42)
+    random.seed(42)
+    torch.cuda.manual_seed_all(42)
+
+    # Generate sequence lengths for each item in the batch
+    if batch_size > 1:
+        # Ensure sequence lengths are reasonably distributed
+        avg_seqlen = total_seqlen // batch_size
+        seqlens = [random.randint(avg_seqlen // 2, avg_seqlen + avg_seqlen // 2) for _ in range(batch_size - 1)]
+        remaining_len = total_seqlen - sum(seqlens)
+        if remaining_len > 0:
+            seqlens.append(remaining_len)
+        else: # Adjust if sum exceeds total_seqlen
+            seqlens.append(avg_seqlen)
+            current_sum = sum(seqlens)
+            seqlens[-1] -= (current_sum - total_seqlen)
+        # Ensure all lengths are positive
+        seqlens = [max(1, s) for s in seqlens]
+        # Final adjustment to match total_seqlen
+        seqlens[-1] += total_seqlen - sum(seqlens)
+
+    else:
+        seqlens = [total_seqlen]
+
+    cu_seqlens = torch.tensor([0] + list(torch.cumsum(torch.tensor(seqlens), 0)), device=device, dtype=torch.int32)
+    max_seqlen = max(seqlens) if seqlens else 0
+
+    q = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+    k = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+    v = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+
+    return q, k, v, cu_seqlens, max_seqlen
+
+
+@pytest.mark.parametrize("batch_size", [1, 2])
+@pytest.mark.parametrize("total_seqlen", [512, 1024])
+@pytest.mark.parametrize("num_heads", [8])
+@pytest.mark.parametrize("head_dim", [64])
+@pytest.mark.parametrize("moba_chunk_size", [64])
+@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])
+def test_moba_attn_varlen_forward(
+    batch_size, total_seqlen, num_heads, head_dim, moba_chunk_size, moba_topk, select_mode, threshold_type, dtype
+):
+    """
+    Tests the forward pass of moba_attn_varlen for basic correctness.
+    It checks output shape, dtype, and for the presence of NaNs/Infs.
+    """
+    if dtype == torch.float32:
+        pytest.skip("float32 is not supported in flash attention")
+
+    q, k, v, cu_seqlens, max_seqlen = generate_test_data(
+        batch_size, total_seqlen, num_heads, head_dim, dtype
+    )
+
+    # Ensure chunk size is not larger than the smallest sequence length
+    min_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).min().item()
+    if moba_chunk_size > min_seqlen:
+        pytest.skip("moba_chunk_size is larger than the minimum sequence length in the batch")
+
+    try:
+        output = moba_attn_varlen(
+            q=q,
+            k=k,
+            v=v,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            moba_chunk_size=moba_chunk_size,
+            moba_topk=moba_topk,
+            select_mode=select_mode,
+            threshold_type=threshold_type,
+            simsum_threshold=0.5, # A reasonable default for threshold mode
+        )
+    except Exception as e:
+        pytest.fail(f"moba_attn_varlen forward pass failed with exception: {e}")
+
+    # 1. Check output shape
+    assert output.shape == q.shape, f"Expected output shape {q.shape}, but got {output.shape}"
+
+    # 2. Check output dtype
+    assert output.dtype == q.dtype, f"Expected output dtype {q.dtype}, but got {output.dtype}"
+
+    # 3. Check for NaNs or Infs in the output
+    assert torch.all(torch.isfinite(output)), "Output contains NaN or Inf values"

docker/Dockerfile.python3.10

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:12.8.0-devel-ubuntu22.04
+FROM nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04
 
 ENV DEBIAN_FRONTEND=noninteractive
 
@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
     uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp310-cp310-linux_x86_64.whl
 
 COPY . .
 

docker/Dockerfile.python3.11

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:12.8.0-devel-ubuntu22.04
+FROM nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04
 
 ENV DEBIAN_FRONTEND=noninteractive
 
@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
     uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp311-cp311-linux_x86_64.whl
 
 COPY . .
 
@@ -55,17 +55,10 @@ RUN source $HOME/.local/bin/env && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
 
-# Install STA (Sliding Tile Attention)
+# Install FastVideo Kernels
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
-
-# Install VSA 
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
+    cd csrc/fastvideo_kernel && \
     git submodule update --init --recursive && \
     python setup.py install
 

docker/Dockerfile.python3.12

@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
     uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp312-cp312-linux_x86_64.whl
 
 COPY . .
 
@@ -55,17 +55,10 @@ RUN source $HOME/.local/bin/env && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
 
-# Install STA (Sliding Tile Attention)
+# Install FastVideo Kernels
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
-
-# Install VSA 
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
+    cd csrc/fastvideo_kernel && \
     git submodule update --init --recursive && \
     python setup.py install
 

docker/Dockerfile.python3.12.cuda12.9.1

@@ -55,17 +55,10 @@ RUN source $HOME/.local/bin/env && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
 
-# Install STA (Sliding Tile Attention)
+# Install FastVideo Kernels
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
-
-# Install VSA 
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
+    cd csrc/fastvideo_kernel && \
     git submodule update --init --recursive && \
     python setup.py install
 

docker/rocm7_1.Dockerfile

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

docs/Makefile

@@ -1,26 +0,0 @@
-# Minimal makefile for Sphinx documentation
-#
-
-# You can set these variables from the command line, and also
-# from the environment for the first two.
-SPHINXOPTS    ?=
-SPHINXBUILD   ?= sphinx-build
-SOURCEDIR     = source
-BUILDDIR      = build
-
-# Put it first so that "make" without argument is like "make help".
-help:
-	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
-
-.PHONY: help Makefile
-
-# Catch-all target: route all unknown targets to Sphinx using the new
-# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
-%: Makefile
-	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
-
-clean:
-	@$(SPHINXBUILD) -M clean "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
-	rm -rf "$(SOURCEDIR)/getting_started/examples"
-	rm -rf "$(SOURCEDIR)/inference/examples"
-	rm -rf "$(SOURCEDIR)/training/examples"

docs/README.md

@@ -1,20 +1,39 @@
-# FastVideo documents
+# FastVideo Documentation
 
-## Build the docs
+This directory contains the FastVideo documentation built with MkDocs.
+
+## Build the docs locally
 
 ```bash
-# Install dependencies.
-pip install -r requirements-docs.txt
+# Install dependencies
+pip install -r requirements-mkdocs.txt
 
-# Build the docs.
-make clean
-make html
+# Serve docs with live reload (recommended for development)
+mkdocs serve
+
+# Or build static site
+mkdocs build
 ```
 
-## Open the docs with your browser
+## View the docs
+
+### Development server (with live reload)
 
 ```bash
-python -m http.server -d build/html/
+mkdocs serve
 ```
 
-Launch your browser and open localhost:8000.
+Then open your browser to: http://127.0.0.1:8000
+
+### Static build
+
+```bash
+mkdocs build
+python -m http.server -d site/
+```
+
+Then open your browser to: http://localhost:8000
+
+## Automatic Deployment
+
+Documentation is automatically built and deployed to GitHub Pages when changes are pushed to the `main` branch via the `.github/workflows/docs.yml` workflow.

docs/api/fastvideo.md

@@ -0,0 +1,248 @@
+# FastVideo API Reference
+
+This page contains the complete API reference for the FastVideo library.
+
+## fastvideo
+
+### Modules
+
+| Name | Description |
+|------|-------------|
+| [attention](#fastvideoattention) | Attention mechanisms and backends for video generation |
+| [configs](#fastvideoconfigs) | Configuration classes for pipelines, models, and sampling |
+| [distributed](#fastvideodistributed) | Distributed execution and communication utilities |
+| [entrypoints](#fastvideoentrypoints) | Main API entry points for video generation |
+| [models](#fastvideomodels) | Model implementations (transformers, VAEs, schedulers) |
+| [pipelines](#fastvideopipelines) | Core pipeline classes for video diffusion |
+| [training](#fastvideotraining) | Training utilities and helpers |
+| [workflow](#fastvideoworkflow) | Workflow management and orchestration |
+| [dataset](#fastvideodataset) | Dataset handling and preprocessing |
+| [layers](#fastvideolayers) | Custom neural network layers |
+| [platforms](#fastvideoplatforms) | Platform-specific implementations |
+| [utils](#fastvideoutils) | Utility functions and helpers |
+| [worker](#fastvideoworker) | Execution workers for video generation |
+
+## fastvideo.attention
+
+::: fastvideo.attention
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+## fastvideo.configs
+
+::: fastvideo.configs
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+### Submodules
+
+#### fastvideo.configs.pipelines
+
+::: fastvideo.configs.pipelines
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+#### fastvideo.configs.models
+
+::: fastvideo.configs.models
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+#### fastvideo.configs.sample
+
+::: fastvideo.configs.sample
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+## fastvideo.distributed
+
+::: fastvideo.distributed
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+## fastvideo.entrypoints
+
+::: fastvideo.entrypoints
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+## fastvideo.models
+
+::: fastvideo.models
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+### Submodules
+
+#### fastvideo.models.registry
+
+::: fastvideo.models.registry
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+#### fastvideo.models.loader
+
+::: fastvideo.models.loader
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+## fastvideo.pipelines
+
+::: fastvideo.pipelines
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+### Submodules
+
+#### fastvideo.pipelines.composed_pipeline_base
+
+::: fastvideo.pipelines.composed_pipeline_base
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+#### fastvideo.pipelines.lora_pipeline
+
+::: fastvideo.pipelines.lora_pipeline
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+#### fastvideo.pipelines.pipeline_batch_info
+
+::: fastvideo.pipelines.pipeline_batch_info
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+#### fastvideo.pipelines.pipeline_registry
+
+::: fastvideo.pipelines.pipeline_registry
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+#### fastvideo.pipelines.stages
+
+::: fastvideo.pipelines.stages
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 4
+
+## fastvideo.training
+
+::: fastvideo.training
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+## fastvideo.workflow
+
+::: fastvideo.workflow
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+## fastvideo.dataset
+
+::: fastvideo.dataset
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+## fastvideo.layers
+
+::: fastvideo.layers
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+## fastvideo.platforms
+
+::: fastvideo.platforms
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3
+
+## fastvideo.utils
+
+::: fastvideo.utils
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      heading_level: 3
+
+## fastvideo.worker
+
+::: fastvideo.worker
+    options:
+      show_source: true
+      show_root_heading: true
+      show_root_toc_entry: true
+      show_submodules: true
+      heading_level: 3

docs/api/summary.md

@@ -0,0 +1,27 @@
+# API Summary
+
+This page provides a quick overview of the main FastVideo API components.
+
+## Video Generator
+
+::: fastvideo.VideoGenerator
+    options:
+      show_root_heading: false
+      show_source: false
+      heading_level: 3
+
+## Initialization Configuration
+
+::: fastvideo.PipelineConfig
+    options:
+      show_root_heading: false
+      show_source: false
+      heading_level: 3
+
+## Sampling Configuration
+
+::: fastvideo.SamplingParam
+    options:
+      show_root_heading: false
+      show_source: false
+      heading_level: 3

docs/assets/custom.css

@@ -0,0 +1,41 @@
+.vertical-table-header th.head:not(.stub) {
+    writing-mode: sideways-lr;
+    white-space: nowrap;
+    max-width: 0;
+    p {
+       margin: 0;
+    }
+}
+
+/* Image sizing classes */
+.image-small {
+    max-width: 200px;
+    height: auto;
+}
+
+.image-medium {
+    max-width: 400px;
+    height: auto;
+}
+
+.image-large {
+    max-width: 600px;
+    height: auto;
+}
+
+.image-full {
+    max-width: 100%;
+    height: auto;
+}
+
+/* Responsive images */
+img {
+    max-width: 100%;
+    height: auto;
+}
+
+/* Center images */
+.image-center {
+    display: block;
+    margin: 0 auto;
+}

docs/assets/logos/icon_simple.svg

@@ -0,0 +1,6 @@
+<svg width="160" height="93" viewBox="0 0 160 93" fill="none" xmlns="http://www.w3.org/2000/svg">
+<path d="M28.8511 91.66L57.6319 1.86368H64.5394L35.7585 91.66H28.8511Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
+<path d="M15.0376 91.66L43.8185 1.86368H46.1209L17.3401 91.66H15.0376Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
+<path d="M1.22217 91.66L30.003 1.86366H31.1543L2.3734 91.66H1.22217Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.15122"/>
+<path d="M71.4465 1.86483L42.666 91.6599H69.144L78.3538 58.2746H123.251L129.007 39.855H84.1099L89.866 22.5868H152.032L157.788 1.86483H71.4465Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
+</svg>

docs/assets/logos/logo.svg

@@ -0,0 +1,18 @@
+<svg width="252" height="105" viewBox="0 0 252 105" fill="none" xmlns="http://www.w3.org/2000/svg">
+<path d="M89.4843 55.5457H101.361L87.7028 101H74.638L89.4843 55.5457Z" fill="#356CFF"/>
+<path fill-rule="evenodd" clip-rule="evenodd" d="M96.0167 1.00057H112.645L118.583 48.273H104.924L103.737 39.7882H79.9827L67.5117 55.5457H85.3273L43.1638 101H28.3174L22.3789 55.5457H33.6621L38.4129 91.3031L58.604 68.2729H44.3515L96.0167 1.00057ZM100.768 13.1217L87.7028 29.4852H103.143L100.768 13.1217Z" fill="#356CFF"/>
+<path d="M37.2252 1.00057L22.3789 48.273H36.0375L40.7884 30.6974L62.6727 30.6974L69.6727 21.0005L43.7576 21.0004L46.7269 11.9096L77.6727 11.9096L86 1.00057L37.2252 1.00057Z" fill="#356CFF"/>
+<path fill-rule="evenodd" clip-rule="evenodd" d="M108.488 55.5457L94.2351 101C94.2351 101 105.518 101 120.959 101C136.399 101 144.078 93.0133 148.276 79.788C152.432 68.0157 153.027 55.5457 136.399 55.5457C119.771 55.5457 108.488 55.5457 108.488 55.5457ZM109.081 90.697L116.802 65.8487C116.802 65.8487 120.959 65.8487 132.242 65.8487C143.525 65.8487 137.586 78.5759 135.211 84.0304C133.307 88.4021 127.491 90.697 122.74 90.697C117.989 90.697 109.081 90.697 109.081 90.697Z" fill="#356CFF"/>
+<path d="M173.188 1.00056L168.625 11.9096C168.625 11.9096 149.386 11.9092 142.525 11.9095C135.664 11.9098 136.586 20.3944 141.337 20.3944H159.747C168.654 20.3944 166.961 33.6899 163.904 38.5761C160.467 44.0675 157.371 48.273 148.463 48.273L125.188 48.273L124 37.97L147.87 37.97C153.808 37.97 156.184 29.4852 151.433 29.4852H131.836C120.142 29.4852 125.897 1.00043 141.337 1.00043L173.188 1.00056Z" fill="#356CFF"/>
+<path d="M179.938 1.00056L175.688 11.9096L191.221 11.9096L179.938 48.273H192.409L203.692 11.9096L219.132 11.9095L223.289 1.00043L179.938 1.00056Z" fill="#356CFF"/>
+<path d="M161.341 55.5457H202.845L198.5 65.8487H169.654L167.279 73.7268H188.5L184.749 82.8177H164.31L161.934 90.697H190.251L186.624 101H146.494L161.341 55.5457Z" fill="#356CFF"/>
+<path fill-rule="evenodd" clip-rule="evenodd" d="M230.821 54.9391C255.169 54.9391 251.776 67.0602 249.231 77.9692C246.686 88.8783 240.917 101 217.757 101C194.596 101 195.606 88.8783 199.347 77.9692C203.089 67.0602 206.473 54.9391 230.821 54.9391ZM237.948 77.9692C239.984 70.6965 240.917 65.242 228.446 65.242C215.975 65.242 211.818 71.9087 210.037 77.9692C208.255 84.0298 208.255 91.3025 219.538 91.3025C230.821 91.3025 235.911 85.2419 237.948 77.9692Z" fill="#356CFF"/>
+<path d="M173.188 1.00056L168.625 11.9096C168.625 11.9096 149.386 11.9092 142.525 11.9095C135.664 11.9098 136.586 20.3944 141.337 20.3944M173.188 1.00056C173.188 1.00056 156.777 1.00043 141.337 1.00043M173.188 1.00056L141.337 1.00043M141.337 20.3944C146.088 20.3944 150.839 20.3944 159.747 20.3944M141.337 20.3944H159.747M159.747 20.3944C168.654 20.3944 166.961 33.6899 163.904 38.5761C160.467 44.0675 157.371 48.273 148.463 48.273M148.463 48.273C139.556 48.273 125.188 48.273 125.188 48.273M148.463 48.273L125.188 48.273M125.188 48.273L124 37.97M124 37.97C124 37.97 141.931 37.97 147.87 37.97M124 37.97L147.87 37.97M147.87 37.97C153.808 37.97 156.184 29.4852 151.433 29.4852M151.433 29.4852C146.682 29.4852 138.962 29.4852 131.836 29.4852M151.433 29.4852H131.836M131.836 29.4852C120.142 29.4852 125.897 1.00043 141.337 1.00043M37.2252 1.00057L22.3789 48.273H36.0375L40.7884 30.6974L62.6727 30.6974L69.6727 21.0005L43.7576 21.0004L46.7269 11.9096L77.6727 11.9096L86 1.00057L37.2252 1.00057ZM96.0167 1.00057H112.645L118.583 48.273H104.924L103.737 39.7882H79.9827L67.5117 55.5457H85.3273L43.1638 101H28.3174L22.3789 55.5457H33.6621L38.4129 91.3031L58.604 68.2729H44.3515L96.0167 1.00057ZM87.7028 29.4852L100.768 13.1217L103.143 29.4852H87.7028ZM89.4843 55.5457H101.361L87.7028 101H74.638L89.4843 55.5457ZM108.488 55.5457L94.2351 101C94.2351 101 105.518 101 120.959 101C136.399 101 144.078 93.0133 148.276 79.788C152.432 68.0157 153.027 55.5457 136.399 55.5457C119.771 55.5457 108.488 55.5457 108.488 55.5457ZM116.802 65.8487L109.081 90.697C109.081 90.697 117.989 90.697 122.74 90.697C127.491 90.697 133.307 88.4021 135.211 84.0304C137.586 78.5759 143.525 65.8487 132.242 65.8487C120.959 65.8487 116.802 65.8487 116.802 65.8487ZM179.938 1.00056L175.688 11.9096L191.221 11.9096L179.938 48.273H192.409L203.692 11.9096L219.132 11.9095L223.289 1.00043L179.938 1.00056ZM161.341 55.5457H202.845L198.5 65.8487H169.654L167.279 73.7268H188.5L184.749 82.8177H164.31L161.934 90.697H190.251L186.624 101H146.494L161.341 55.5457ZM230.821 54.9391C255.169 54.9391 251.776 67.0602 249.231 77.9692C246.686 88.8783 240.917 101 217.757 101C194.596 101 195.606 88.8783 199.347 77.9692C203.089 67.0602 206.473 54.9391 230.821 54.9391ZM228.446 65.242C240.917 65.242 239.984 70.6965 237.948 77.9692C235.911 85.2419 230.821 91.3025 219.538 91.3025C208.255 91.3025 208.255 84.0298 210.037 77.9692C211.818 71.9087 215.975 65.242 228.446 65.242Z" stroke="#356CFF" stroke-width="1.18771"/>
+<path d="M15.2524 55.5451L21.191 100.999L24.7541 100.999L18.8156 55.5451L15.2524 55.5451Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.18771"/>
+<path d="M8.12646 55.5451L14.065 100.999L15.2527 100.999L9.31417 55.5451L8.12646 55.5451Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.18771"/>
+<path d="M1 55.5451L6.93853 100.999L7.53239 100.999L1.59385 55.5451L1 55.5451Z" fill="#356CFF" stroke="#356CFF" stroke-width="0.593853"/>
+<path d="M15.2524 48.2724L30.0988 1H33.6619L18.8156 48.2724H15.2524Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.18771"/>
+<path d="M8.12646 48.2724L22.9728 1H24.1605L9.31417 48.2724H8.12646Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.18771"/>
+<path d="M1 48.2724L15.8463 1H16.4402L1.59385 48.2724H1Z" fill="#356CFF" stroke="#356CFF" stroke-width="0.593853"/>
+<path d="M85.3271 55.5457H67.5116L87 12.7363L44.3513 68.2729H58.6038L43.1636 101L85.3271 55.5457Z" fill="#FDC717" stroke="#FDC717" stroke-width="1.18771" stroke-miterlimit="16"/>
+</svg>

docs/contributing/developer_env/docker.md

@@ -1,4 +1,4 @@
-(docker)=
+
 # 🐳 Using the FastVideo Docker Image
 
 If you prefer a containerized development environment or want to avoid managing dependencies manually, you can use our prebuilt Docker image:

docs/contributing/developer_env/index.md

@@ -3,11 +3,3 @@
 # 🧰 Developer Environment
 
 Accelerate your FastVideo development workflow by leveraging Docker images and cloud GPUs for efficient experimentation and reproducible environments.
-
-:::{toctree}
-:caption: Contents
-:maxdepth: 1
-
-docker
-runpod
-:::

docs/contributing/developer_env/runpod.md

@@ -1,4 +1,3 @@
-(runpod)=
 
 # 📦 Developing FastVideo on RunPod
 
@@ -10,7 +9,7 @@ Choose a GPU that supports CUDA 12.8
 
 Pick 1 or 2 L40S GPU(s)
 
-![RunPod CUDA selection](../../_static/images/runpod_cuda.png)
+![RunPod CUDA selection](../../assets/images/runpod_cuda.png)
 
 When creating your pod template, use this image:
 
@@ -24,11 +23,11 @@ Paste Container Start Command to support SSH ([RunPod Docs](https://docs.runpod.
 bash -c "apt update;DEBIAN_FRONTEND=noninteractive apt-get install openssh-server -y;mkdir -p ~/.ssh;cd $_;chmod 700 ~/.ssh;echo \"$PUBLIC_KEY\" >> authorized_keys;chmod 700 authorized_keys;service ssh start;sleep infinity"
 ```
 
-![RunPod template configuration](../../_static/images/runpod_template.png)
+![RunPod template configuration](../../assets/images/runpod_template.png)
 
 After deploying, the pod will take a few minutes to pull the image and start the SSH service.
 
-![RunPod ssh](../../_static/images/runpod_ssh.png)
+![RunPod ssh](../../assets/images/runpod_ssh.png)
 
 ## Working with the pod
 

docs/contributing/overview.md

@@ -1,4 +1,3 @@
-(developer-overview)=
 
 # 🛠️ Contributing to FastVideo
 
@@ -7,7 +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 support 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 supports 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:
 
@@ -71,3 +70,7 @@ uv pip install ninja
 
 python setup.py install
 ```
+
+## Testing
+
+Please refer to the [Testing Guide](testing.md) for more information on how to add and run tests in 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 the inference.
+    Profiling is only intended for FastVideo developers and maintainers to understand the proportion of time spent in different parts of the codebase. **FastVideo end-users should never turn on profiling** as it will significantly slow down inference.
 
 ## Profiling with PyTorch
 
@@ -49,5 +49,5 @@ Traces can be visualized using <https://ui.perfetto.dev/>.
 ### Best Practices
 
 - Keep the profiled step count small; traces can be large and slow down job shutdown while the profiler flushes data.
-- After profiling, clean up trace directories to avoid filling disks.
+- After profiling, clean up trace directories to avoid filling disk storage.
 - When adding new regions, register them in `fastvideo.profiler` and wrap the corresponding code block with `with self.profiler_controller.region("your_region"):` or the `@profile_region` decorator.

docs/contributing/testing.md

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

docs/design/overview.md

@@ -29,7 +29,6 @@ FastVideo separates model components from execution logic with these principles:
 - **Custom Attention Backends**: Components can support and use different Attention implementations
 - **Pipeline Abstraction**: Consistent interface across diffusion models
 
-(design-fastvideo-args)=
 ## FastVideoArgs
 
 The `FastVideoArgs` class in `fastvideo/fastvideo_args.py` serves as the central configuration system for FastVideo. It contains all parameters needed to control model loading, inference configuration, performance optimization settings, and more.
@@ -61,7 +60,6 @@ with set_current_fastvideo_args(fastvideo_args):
     result = generate_video()
 ```
 
-(design-pipeline-system)=
 ## Pipeline System
 
 ### `ComposedPipelineBase`
@@ -108,7 +106,8 @@ def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> Forward
     return batch
 ```
 
-(design-forwardbatch)=
+![Pipeline execution and data flow](../assets/images/pipeline.png)
+
 ### ForwardBatch
 
 Defined in `fastvideo/pipelines/pipeline_batch_info.py`, `ForwardBatch` encapsulates the data payload passed between pipeline stages. It typically holds:
@@ -120,12 +119,10 @@ Defined in `fastvideo/pipelines/pipeline_batch_info.py`, `ForwardBatch` encapsul
 
 This structure facilitates clear state transitions between stages.
 
-(design-model-components)=
 ## Model Components
 
 The `fastvideo/models/` directory contains implementations of the core neural network models used in video diffusion:
 
-(design-transformer-models)=
 ### Transformer Models
 
 Transformer networks perform the actual denoising during diffusion:
@@ -152,7 +149,6 @@ def forward(
     return noise_pred  # Predicted noise residual
 ```
 
-(design-vae-variational-auto-encoder)=
 ### VAE (Variational Auto-Encoder)
 
 VAEs handle conversion between pixel space and latent space:
@@ -170,7 +166,6 @@ FastVideo's VAE implementations include:
 - Optional tiling for large frames
 - Distributed weight support
 
-(design-text-and-image-encoders)=
 ### Text and Image Encoders
 
 Encoders process conditioning inputs into embeddings:
@@ -188,7 +183,6 @@ FastVideo implements optimizations such as:
 - Caching for common prompts
 - Precision-tuned computation
 
-(design-schedulers)=
 ### Schedulers
 
 Schedulers manage the diffusion sampling process:
@@ -216,7 +210,10 @@ def step(
     return prev_sample
 ```
 
-(design-optimized-attention)=
+This diagram shows how models are discovered, validated, and loaded across entrypoints, executors, pipelines, and model loaders.
+
+![Model loading flow](../assets/images/load_models.png)
+
 ## Optimized Attention
 
 The `fastvideo/attention/` directory contains optimized attention implementations crucial for efficient video diffusion:
@@ -240,17 +237,17 @@ self.attn = LocalAttention(
 # export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
 ```
 
+![Attention backend selector design](../assets/images/attention_backend.png)
+
 ### Attention Patterns
 Supports various patterns with memory optimization techniques:
 - **Cross/Self/Temporal/Global-Local Attention**
 - Chunking, progressive computation, optimized masking
 
-(design-distributed-processing)=
 ## Distributed Processing
 
 The `fastvideo/distributed/` directory contains implementations for distributed model execution:
 
-(design-tensor-parallelism)=
 ### Tensor Parallelism
 
 Tensor parallelism splits model weights across devices:
@@ -307,7 +304,6 @@ Efficient communication primitives minimize distributed overhead:
 - **Tensor-Parallel AllReduce**: Combines partial results
 - **Distributed Synchronization**: Coordinates execution
 
-(design-forwardcontext)=
 ## Forward Context Management
 
 ### ForwardContext
@@ -330,7 +326,6 @@ with set_forward_context(current_timestep, attn_metadata, fastvideo_args):
     output = model(inputs)
 ```
 
-(design-executor-and-worker-abstractions)=
 ## Executor and Worker System
 
 The `fastvideo/worker/` directory contains the distributed execution framework:
@@ -357,7 +352,6 @@ Each GPU worker:
 
 This design allows FastVideo to efficiently utilize multiple GPUs while providing a simple, unified interface for model execution.
 
-(design-platforms)=
 ## Platforms
 
 The `fastvideo/platforms/` directory provides hardware platform abstractions that enable FastVideo to run efficiently on different hardware configurations:
@@ -388,7 +382,6 @@ else:
 
 The platform system is designed to be extensible for future hardware targets.
 
-(design-logger)=
 ## Logger
 See [PR](https://github.com/hao-ai-lab/FastVideo/pull/356)
 

docs/distillation/data_preprocess.md

@@ -1,4 +1,3 @@
-(v0-data-preprocess)=
 
 # 🧱 Data Preprocess for Distillation
 

docs/distillation/dmd.md

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

docs/examples/examples_index.md

@@ -0,0 +1,12 @@
+# 💡 Examples
+
+A collection of examples demonstrating usage of FastVideo.
+
+All documented examples are autogenerated using [generate_examples.py](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/generate_examples.py) from examples found in the [examples](https://github.com/hao-ai-lab/FastVideo/tree/main/examples) directory.
+
+## Examples
+
+- [Examples Distillation Index](distillation/examples/examples_distillation_index.md)
+- [Examples Training Index](training/examples/examples_training_index.md)
+- [Examples Inference Index](inference/examples/examples_inference_index.md)
+

docs/generate_examples.py

@@ -6,10 +6,11 @@ import re
 from dataclasses import dataclass, field
 from pathlib import Path
 
-ROOT_DIR = Path(__file__).parent.parent.parent.resolve()
-ROOT_DIR_RELATIVE = '../../../..'
+ROOT_DIR = Path(__file__).parent.parent.resolve()
+ROOT_DIR_RELATIVE = '../..'
 EXAMPLE_DIR = ROOT_DIR / "examples"
-EXAMPLE_DOC_DIR = ROOT_DIR / "docs/source/getting_started/examples"
+EXAMPLE_DOC_DIR = ROOT_DIR / "docs/getting_started/examples"
+GITHUB_REPO = "hao-ai-lab/FastVideo"  # Update this to your repo
 
 
 def fix_case(text: str) -> str:
@@ -71,9 +72,16 @@ class Index:
 
     def generate(self) -> str:
         content = f"# {self.title}\n\n{self.description}\n\n"
-        content += ":::{toctree}\n"
-        content += f":caption: {self.caption}\n:maxdepth: {self.maxdepth}\n"
-        content += "\n".join(self.documents) + "\n:::\n"
+        if self.caption:
+            content += f"## {self.caption}\n\n"
+        # Generate a simple list of links for MkDocs
+        for doc in self.documents:
+            # Convert document path to proper link
+            doc_link = doc.replace("\\", "/")
+            # Get just the filename for the link text
+            doc_title = fix_case(Path(doc).stem.replace("_", " ").title())
+            content += f"- [{doc_title}]({doc_link}.md)\n"
+        content += "\n"
         return content
 
 
@@ -142,30 +150,66 @@ class Example:
         return fix_case(self.path.stem.replace("_", " ").title())
 
     def generate(self) -> str:
-        # Convert the path to a relative path from __file__
-        make_relative = lambda path: ROOT_DIR_RELATIVE / path.relative_to(
-            ROOT_DIR)
+        # Create GitHub link to source
+        github_path = str(self.path.relative_to(ROOT_DIR)).replace("\\", "/")
+        github_url = f"https://github.com/{GITHUB_REPO}/blob/main/{github_path}"
+        content = f"**Source:** [{github_path}]({github_url})\n\n"
 
-        content = f"Source <gh-file:{self.path.relative_to(ROOT_DIR)}>.\n\n"
-        include = "include" if self.main_file.suffix == ".md" else \
-            "literalinclude"
-        if include == "literalinclude":
+        # Add title for code files
+        if self.main_file.suffix != ".md":
             content += f"# {self.title}\n\n"
-        content += f":::{{{include}}} {make_relative(self.main_file)}\n"  # type: ignore[no-untyped-call]
-        if include == "literalinclude":
-            content += f":language: {self.main_file.suffix[1:]}\n"
-        content += ":::\n\n"
+
+        # Include main file content
+        if self.main_file.suffix == ".md":
+            # For markdown files, include the content directly
+            with open(self.main_file, encoding='utf-8') as f:
+                content += f.read() + "\n\n"
+        else:
+            # For code files, use code blocks
+            language = self.main_file.suffix[1:] if self.main_file.suffix else ""
+            with open(self.main_file, encoding='utf-8') as f:
+                file_content = f.read()
+            content += f"```{language}\n{file_content}\n```\n\n"
 
         if not self.other_files:
             return content
 
-        content += "## Example materials\n\n"
+        content += "## Additional Files\n\n"
+        # Define binary/non-text file extensions to skip
+        binary_extensions = {
+            '.mp4', '.avi', '.mov', '.mkv', '.gif', '.jpg', '.jpeg', '.png',
+            '.webp', '.bmp', '.pdf', '.zip', '.tar', '.gz', '.mp3', '.wav'
+        }
+
         for file in sorted(self.other_files):
-            include = "include" if file.suffix == ".md" else "literalinclude"
-            content += f":::{{admonition}} {file.relative_to(self.path)}\n"
-            content += ":class: dropdown\n\n"
-            content += f":::{{{include}}} {make_relative(file)}\n:::\n"  # type: ignore[no-untyped-call]
-            content += ":::\n\n"
+            # Skip binary files
+            if file.suffix.lower() in binary_extensions:
+                continue
+
+            file_rel_path = file.relative_to(self.path)
+            # Use collapsible admonition syntax for MkDocs
+            content += f"??? note \"{file_rel_path}\"\n\n"
+
+            try:
+                if file.suffix == ".md":
+                    # Include markdown content with indentation
+                    with open(file, encoding='utf-8') as f:
+                        for line in f:
+                            content += f"    {line}"
+                else:
+                    # Include code with proper formatting
+                    language = file.suffix[1:] if file.suffix else ""
+                    with open(file, encoding='utf-8') as f:
+                        file_content = f.read()
+                    # Indent the code block for the admonition
+                    content += f"    ```{language}\n"
+                    for line in file_content.split('\n'):
+                        content += f"    {line}\n"
+                    content += "    ```\n"
+                content += "\n"
+            except UnicodeDecodeError:
+                # Skip files that can't be decoded as UTF-8
+                continue
 
         return content
 
@@ -195,7 +239,7 @@ class NestedStructure:
 
 def create_category_indices() -> dict[str, Index]:
     """Create category indices with their respective configurations."""
-    main_index_dir = ROOT_DIR / "docs/source/examples"
+    main_index_dir = ROOT_DIR / "docs/examples"
     if not main_index_dir.exists():
         main_index_dir.mkdir(parents=True)
 
@@ -203,17 +247,16 @@ def create_category_indices() -> dict[str, Index]:
         "inference":
         Index(
             path=ROOT_DIR /
-            "docs/source/inference/examples/examples_inference_index.md",
+            "docs/inference/examples/examples_inference_index.md",
             title="🚀 Examples",
             description=
-            "Inference examples demonstrate how to use FastVideo inference. We recommend starting with <project:basic.md>.",
+            "Inference examples demonstrate how to use FastVideo inference. We recommend starting with [basic.md](basic.md).",
             caption="Examples",
             maxdepth=1,
         ),
         "training":
         Index(
-            path=ROOT_DIR /
-            "docs/source/training/examples/examples_training_index.md",
+            path=ROOT_DIR / "docs/training/examples/examples_training_index.md",
             title="🚀 Examples",
             description=
             "Training examples demonstrate how to use FastVideo training.",
@@ -223,7 +266,7 @@ def create_category_indices() -> dict[str, Index]:
         "distillation":
         Index(
             path=ROOT_DIR /
-            "docs/source/distillation/examples/examples_distillation_index.md",
+            "docs/distillation/examples/examples_distillation_index.md",
             title="🚀 Examples",
             description=
             "Distillation examples demonstrate how to use FastVideo distillation.",
@@ -246,9 +289,21 @@ def find_examples(category_indices: dict[str, Index],
     examples = []
     glob_patterns = ["*.py", "*.md", "*.sh"]
 
+    # Map category names to actual directory names
+    category_dir_mapping = {
+        "distillation": "distill",  # examples/distill/ -> distillation category
+    }
+
     # Find categorised examples
     for category in category_indices:
-        category_dir = EXAMPLE_DIR / category
+        # Use mapped directory name if available, otherwise use category name
+        dir_name = category_dir_mapping.get(category, category)
+        category_dir = EXAMPLE_DIR / dir_name
+
+        # Skip if directory doesn't exist
+        if not category_dir.exists():
+            continue
+
         globs = [category_dir.glob(pattern) for pattern in glob_patterns]
         for path in itertools.chain(*globs):
             examples.append(Example(path, category))
@@ -279,11 +334,18 @@ def create_nested_structures(
                                                 dict[str,
                                                      NestedStructure]]]] = {}
 
+    # Map category names to actual directory names
+    category_dir_mapping = {
+        "distillation": "distill",
+    }
+
     for example in examples:
         if example.category not in ["training", "distillation"]:
             continue
 
-        category_dir = EXAMPLE_DIR / example.category
+        # Use mapped directory name if available
+        dir_name = category_dir_mapping.get(example.category, example.category)
+        category_dir = EXAMPLE_DIR / dir_name
         relative_path = example.path.relative_to(category_dir)
         path_parts = relative_path.parts
 
@@ -415,7 +477,7 @@ def generate_nested_examples(nested_structures: dict[str, dict[str, dict[
             category_index.documents.append(method)
 
 
-def generate_examples(generate_main_index=False):
+def generate_examples(generate_main_index: bool = False) -> None:
     """
     Generate example documentation.
     
@@ -429,12 +491,14 @@ def generate_examples(generate_main_index=False):
     # Create the main examples index only if requested
     examples_index = None
     if generate_main_index:
-        main_index_dir = ROOT_DIR / "docs/source/examples"
+        main_index_dir = ROOT_DIR / "docs/examples"
         examples_index = Index(
             path=main_index_dir / "examples_index.md",
             title="💡 Examples",
             description=
-            "A collection of examples demonstrating usage of FastVideo.\nAll documented examples are autogenerated using <gh-file:docs/source/generate_examples.py> from examples found in <gh-file:examples>.",
+            "A collection of examples demonstrating usage of FastVideo.\n\n"
+            f"All documented examples are autogenerated using [generate_examples.py](https://github.com/{GITHUB_REPO}/blob/main/docs/generate_examples.py) "
+            f"from examples found in the [examples](https://github.com/{GITHUB_REPO}/tree/main/examples) directory.",
             caption="Examples",
             maxdepth=2)
 
@@ -471,3 +535,19 @@ def generate_examples(generate_main_index=False):
     if generate_main_index and examples_index:
         with open(examples_index.path, "w+") as f:
             f.write(examples_index.generate())
+
+
+def on_pre_build_hook(config, **kwargs):
+    """
+    MkDocs hook to generate examples before building the documentation.
+    This function is called automatically by the mkdocs-simple-hooks plugin.
+    """
+    print("Generating example documentation...")
+    generate_examples(generate_main_index=True)
+    print("Example documentation generated successfully!")
+
+
+if __name__ == "__main__":
+    print("Generating example documentation...")
+    generate_examples(generate_main_index=True)
+    print("Example documentation generated successfully!")

docs/getting_started/installation.md

@@ -0,0 +1,41 @@
+
+# 🔧 Installation
+
+FastVideo supports the following hardware platforms:
+
+- [NVIDIA CUDA](installation/gpu.md)
+- [Apple silicon](installation/mps.md)
+
+## Quick Installation
+
+### Using pip
+
+```bash
+pip install fastvideo
+```
+
+### Using conda
+
+```bash
+conda install -c conda-forge fastvideo
+```
+
+### From source
+
+```bash
+git clone https://github.com/hao-ai-lab/FastVideo.git
+cd FastVideo
+pip install -e .
+```
+
+## Hardware Requirements
+
+- **NVIDIA GPUs**: CUDA 11.8+ with compute capability 7.0+
+- **Apple Silicon**: macOS 12.0+ with M1/M2/M3 chips
+- **CPU**: x86_64 architecture (for CPU-only inference)
+
+## Next Steps
+
+- [Quick Start Guide](quick_start.md) - Get started with your first video generation
+- [Configuration](../inference/configuration.md) - Learn about configuration options
+- [Examples](../inference/examples/) - Explore example scripts and notebooks

docs/getting_started/installation/gpu.md

@@ -30,17 +30,8 @@ conda create -n fastvideo python=3.12 -y
 conda activate fastvideo
 ```
 
-:::{note}
-[PyTorch has deprecated the conda release channel](https://github.com/pytorch/pytorch/issues/138506). If you use `conda`, please only use it to create Python environment rather than installing packages.
-:::
-
 #### uv
 
-:::{tip}
-We highly recommend using `uv` to install FastVideo. In our experience, `uv` speeds up installation by at least 3x.
-Note that you can also use `uv` to install FastVideo in a Conda environment.
-:::
-
 Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
 
 ```console

docs/getting_started/installation/mps.md

@@ -31,17 +31,8 @@ conda create -n fastvideo python=3.12.4 -y
 conda activate fastvideo
 ```
 
-:::{note}
-[PyTorch has deprecated the conda release channel](https://github.com/pytorch/pytorch/issues/138506). If you use `conda`, please only use it to create Python environment rather than installing packages.
-:::
-
 #### uv
 
-:::{tip}
-We highly recommend using `uv` to install FastVideo. In our experience, `uv` speeds up installation by at least 3x.
-Note that you can also use `uv` to install FastVideo in a Conda environment.
-:::
-
 Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
 
 ```console

docs/getting_started/quick_start.md

@@ -0,0 +1,77 @@
+# 🚀 Quick Start
+
+Get up and running with FastVideo in minutes!
+
+## Installation
+
+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
+```
+
+## Basic Usage
+
+### Text-to-Video Generation
+
+```python
+from fastvideo import VideoGenerator
+
+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
+    )
+
+    # Define a prompt for your video
+    prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest."
+
+    # Generate the video
+    video = generator.generate_video(
+        prompt,
+        return_frames=True,  # Also return frames from this call (defaults to False)
+        output_path="my_videos/",  # Controls where videos are saved
+        save_video=True
+    )
+
+if __name__ == '__main__':
+    main()
+```
+
+### Image-to-Video Generation
+
+```python
+from fastvideo import VideoGenerator, SamplingParam
+
+def main():
+    # Create the generator
+    model_name = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+    generator = VideoGenerator.from_pretrained(model_name, num_gpus=1)
+
+    # Set up parameters with an initial image
+    sampling_param = SamplingParam.from_pretrained(model_name)
+    sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+    sampling_param.num_frames = 107
+
+    # Generate 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)
+
+if __name__ == '__main__':
+    main()
+```
+
+## Next Steps
+
+- [Installation Guide](installation.md) - Detailed installation instructions
+- [Configuration](../inference/configuration.md) - Learn about configuration options
+- [Examples](../inference/examples/) - Explore more examples
+- [Optimizations](../inference/optimizations.md) - Performance optimization tips

docs/getting_started/v1_api.md

@@ -0,0 +1,42 @@
+# V1 API
+
+FastVideo's V1 API provides a streamlined interface for video generation tasks with powerful customization options. This page documents the primary components of the API.
+
+## Video Generator
+
+This class will be the primary Python API for generating videos and images.
+
+::: fastvideo.entrypoints.video_generator.VideoGenerator
+    options:
+      show_root_heading: true
+      show_source: false
+      members:
+        - from_pretrained
+      heading_level: 3
+
+`VideoGenerator.from_pretrained()` should be the primary way of creating a new video generator.
+
+## Configuring FastVideo
+
+The following two classes `PipelineConfig` and `SamplingParam` are used to configure initialization and sampling parameters, respectively.
+
+### PipelineConfig
+
+::: fastvideo.configs.pipelines.base.PipelineConfig
+    options:
+      show_root_heading: true
+      show_source: false
+      members:
+        - from_pretrained
+        - dump_to_json
+      heading_level: 4
+
+### SamplingParam
+
+::: fastvideo.configs.sample.base.SamplingParam
+    options:
+      show_root_heading: true
+      show_source: false
+      members:
+        - from_pretrained
+      heading_level: 4

docs/index.md

@@ -1,132 +1,52 @@
 # Welcome to FastVideo
 
-:::{figure} ../../assets/logos/logo.svg
-:align: center
-:alt: FastVideo
-:class: no-scaled-link
-:width: 60%
-:::
+<div style="text-align: center;">
+  <img src="assets/logos/logo.svg" alt="FastVideo" style="width: 60%;" />
+</div>
 
-:::{raw} html
-<p style="text-align:center">
-<strong>FastVideo is a unified inference and post-training framework for accelerated video generation.
-</strong>
-</p>
+<div style="text-align: center;">
+<strong>FastVideo is a unified inference and post-training framework for accelerated video generation.</strong>
+</div>
 
-<p style="text-align:center">
+<div style="text-align: center;">
 <script async defer src="https://buttons.github.io/buttons.js"></script>
 <a class="github-button" href="https://github.com/hao-ai-lab/FastVideo/" data-show-count="true" data-size="large" aria-label="Star">Star</a>
 <a class="github-button" href="https://github.com/hao-ai-lab/FastVideo/subscription" data-icon="octicon-eye" data-size="large" aria-label="Watch">Watch</a>
 <a class="github-button" href="https://github.com/hao-ai-lab/FastVideo/fork" data-icon="octicon-repo-forked" data-size="large" aria-label="Fork">Fork</a>
-</p>
-:::
+</div>
 
 FastVideo is an inference and post-training framework for diffusion models. It 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.
 
 <div style="text-align: center;">
-  <img src=_static/images/fastwan.png width="100%"/>
+  <img src="assets/images/fastwan.png" style="width: 100%;"/>
 </div>
 
 ## Key Features
 
 FastVideo has the following features:
+
 - State-of-the-art performance optimizations for inference
   - [Sliding Tile Attention](https://arxiv.org/pdf/2502.04507)
   - [TeaCache](https://arxiv.org/pdf/2411.19108)
   - [Sage Attention](https://arxiv.org/abs/2410.02367)
 - E2E post-training support
-  - Data preprocessing pipeline for video data.
+  - Data preprocessing pipeline for video data
   - [Sparse distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/) for Wan2.1 and Wan2.2 using [Video Sparse Attention](https://arxiv.org/pdf/2505.13389) and [Distribution Matching Distillation](https://tianweiy.github.io/dmd2/)
   - Support full finetuning and LoRA finetuning for state-of-the-art open video DiTs.
   - Scalable training with FSDP2, sequence parallelism, and selective activation checkpointing, with near linear scaling to 64 GPUs.
 
 ## Documentation
 
-% How to start using FastVideo?
+Welcome to FastVideo! This documentation will help you get started with our unified inference and post-training framework for accelerated video generation.
 
-:::{toctree}
-:caption: Getting Started
-:maxdepth: 1
+Use the navigation menu on the left to explore different sections:
 
-getting_started/installation
-<!-- getting_started/v1_api -->
-:::
-
-:::{toctree}
-:caption: Inference
-:maxdepth: 1
-
-inference/inference_quick_start
-inference/examples/examples_inference_index
-inference/configuration
-inference/optimizations
-inference/comfyui
-inference/support_matrix
-inference/cli
-inference/add_pipeline
-:::
-
-:::{toctree}
-:caption: Training
-:maxdepth: 1
-
-training/examples/examples_training_index
-training/data_preprocess
-<!-- training/finetune -->
-:::
-
-:::{toctree}
-:caption: Distillation
-:maxdepth: 1
-
-distillation/examples/examples_distillation_index
-distillation/data_preprocess
-distillation/dmd
-:::
-
-% What is STA Kernel?
-
-:::{toctree}
-:caption: Sliding Tile Attention
-:maxdepth: 1
-
-sliding_tile_attention/installation
-sliding_tile_attention/demo
-:::
-
-% What is VSA Kernel?
-
-:::{toctree}
-:caption: Video Sparse Attention
-:maxdepth: 1
-
-video_sparse_attention/installation
-:::
-
-:::{toctree}
-:caption: Design
-:maxdepth: 1
-design/overview
-:::
-
-:::{toctree}
-:caption: Developer Guide
-:maxdepth: 2
-
-contributing/overview
-contributing/developer_env/index
-contributing/profiling
-:::
-
-:::{toctree}
-:caption: API Reference
-:maxdepth: 2
-
-<!-- api/summary -->
-api/fastvideo/fastvideo
-:::
-
-## Indices and tables
-
-- {ref}`genindex`
-- {ref}`modindex`
+- **Getting Started**: Installation and quick start guides
+- **Inference**: Learn how to use FastVideo for video generation
+- **Training**: Data preprocessing and fine-tuning workflows
+- **Distillation**: Post-training optimization techniques
+- **Sliding Tile Attention**: Advanced attention mechanisms
+- **Video Sparse Attention**: Efficient attention for video models
+- **Design**: Framework architecture and design principles
+- **Developer Guide**: Contributing and development setup
+- **API Reference**: Complete API documentation

docs/inference/add_pipeline.md

@@ -1,4 +1,3 @@
-(add-pipeline)=
 
 # 🏗️ Adding a New Pipeline
 

docs/inference/configuration.md

@@ -1,4 +1,4 @@
-(inference-configuration)=
+
 # Configuration
 
 ## Multi-GPU Setup

docs/inference/optimizations.md

@@ -1,4 +1,3 @@
-(inference-optimizations)=
 
 # Optimizations
 
@@ -16,8 +15,6 @@ This page describes the various options for speeding up generation times in Fast
 - Caching Techniques
   - [TeaCache](#optimizations-teacache)
 
-(optimizations-backends)=
-
 ## Attention Backends
 
 ### Available Backends
@@ -49,8 +46,6 @@ You can also set the environment variable on the command line:
 FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN python example.py
 ```
 
-(optimizations-flash)=
-
 ### Flash Attention
 
 **`FLASH_ATTN`**
@@ -71,12 +66,6 @@ pip install ninja
 python setup.py install
 ```
 
-:::{note}
-FastVideo will automatically detect and use `FA3` if it is installed when using `FLASH_ATTN` backend.
-:::
-
-(optimizations-sta)=
-
 ### Sliding Tile Attention
 
 **`SLIDING_TILE_ATTN`**
@@ -87,8 +76,6 @@ pip install st_attn==0.0.4
 
 Please see [this page](#sta-installation) for more installation instructions.
 
-(optimizations-vsa)=
-
 ### Video Sparse Attention
 
 **`VIDEO_SPARSE_ATTN`**
@@ -100,8 +87,6 @@ python setup_vsa.py install
 
 Please see [this page](#vsa-installation) for more installation instructions.
 
-(optimizations-sage)=
-
 ### Sage Attention
 
 **`SAGE_ATTN`**
@@ -114,8 +99,6 @@ cd sageattention
 python setup.py install  # or pip install -e .
 ```
 
-(optimizations-sage3)=
-
 ### Sage Attention 3
 
 **`SAGE_ATTN_THREE`**
@@ -136,8 +119,6 @@ To use Sage Attention 3 in FastVideo, first get access to the SageAttention3 cod
 python setup.py install
 ```
 
-(optimizations-teacache)=
-
 ## Teacache
 
 TeaCache is an optimization technique supported in FastVideo that can significantly speed up video generation by skipping redundant calculations across diffusion steps. This guide explains how to enable and configure TeaCache for optimal performance in FastVideo.

docs/inference/support_matrix.md

@@ -0,0 +1,66 @@
+# Compatibility Matrix
+
+The table below shows every supported model and optimizations supported for them.
+
+The symbols used have the following meanings:
+
+- ✅ = Full compatibility
+- ❌ = No compatibility
+- ⭕ = Does not apply to this model
+
+## Models x Optimization
+
+The `HuggingFace Model ID` can be directly pass to `from_pretrained()` methods and FastVideo will use the optimal default parameters when initializing and generating videos.
+
+<style>
+  /* Target tables in this section */
+  #models-x-optimization + p + table {
+    display: block;
+    overflow-x: auto;
+    width: 100%;
+    font-size: 0.85rem;
+  }
+  
+  #models-x-optimization + p + table td,
+  #models-x-optimization + p + table th {
+    text-align: center;
+    white-space: nowrap;
+    padding: 0.5em;
+  }
+  
+  /* First two columns can wrap */
+  #models-x-optimization + p + table td:nth-child(1),
+  #models-x-optimization + p + table td:nth-child(2) {
+    white-space: normal;
+    min-width: 120px;
+  }
+  
+  #models-x-optimization + p + table td:nth-child(2) code {
+    font-size: 0.75rem;
+  }
+</style>
+
+| Model Name | HuggingFace Model ID | Resolutions | TeaCache | Sliding Tile Attn | Sage Attn | VSA |
+|------------|---------------------|-------------|----------|-------------------|-----------|-----|
+| FastWan2.1 T2V 1.3B | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ✅ |
+| FastWan2.2 TI2V 5B Full Attn* | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ |
+| Wan2.2 TI2V 5B | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 720P | ⭕ | ⭕ | ✅ | ⭕ |
+| Wan2.2 T2V A14B | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ |
+| Wan2.2 I2V A14B | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ |
+| HunyuanVideo | `hunyuanvideo-community/HunyuanVideo` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ |
+| FastHunyuan | `FastVideo/FastHunyuan-diffusers` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ |
+| Wan2.1 T2V 1.3B | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | 480P | ✅ | ✅* | ✅ | ⭕ |
+| Wan2.1 T2V 14B | `Wan-AI/Wan2.1-T2V-14B-Diffusers` | 480P, 720P | ✅ | ✅* | ✅ | ⭕ |
+| Wan2.1 I2V 480P | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | 480P | ✅ | ✅* | ✅ | ⭕ |
+| Wan2.1 I2V 720P | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` | 720P | ✅ | ✅ | ✅ | ⭕ |
+| StepVideo T2V | `FastVideo/stepvideo-t2v-diffusers` | 768px768px204f<br>544px992px204f<br>544px992px136f | ❌ | ❌ | ✅ | ⭕ |
+
+**Note**: Wan2.2 TI2V 5B has some quality issues when performing I2V generation. We are working on fixing this issue.
+
+## Special requirements
+
+### StepVideo T2V
+- The self-attention in text-encoder (step_llm) only supports CUDA capabilities sm_80 sm_86 and sm_90
+
+### Sliding Tile Attention
+- Currently only Hopper GPUs (H100s) are supported.

docs/make.bat

@@ -1,35 +0,0 @@
-@ECHO OFF
-
-pushd %~dp0
-
-REM Command file for Sphinx documentation
-
-if "%SPHINXBUILD%" == "" (
-	set SPHINXBUILD=sphinx-build
-)
-set SOURCEDIR=source
-set BUILDDIR=build
-
-%SPHINXBUILD% >NUL 2>NUL
-if errorlevel 9009 (
-	echo.
-	echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
-	echo.installed, then set the SPHINXBUILD environment variable to point
-	echo.to the full path of the 'sphinx-build' executable. Alternatively you
-	echo.may add the Sphinx directory to PATH.
-	echo.
-	echo.If you don't have Sphinx installed, grab it from
-	echo.https://www.sphinx-doc.org/
-	exit /b 1
-)
-
-if "%1" == "" goto help
-
-%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
-goto end
-
-:help
-%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
-
-:end
-popd

docs/requirements-docs.txt

@@ -1,15 +0,0 @@
-sphinx==7.4.7
-sphinx-argparse==0.5.2
-sphinx-autodoc2==0.5.0
-sphinx-book-theme==1.1.4
-sphinx-copybutton==0.5.2
-sphinx-design==0.6.1
-sphinx-togglebutton==0.3.2
-myst-parser==3.0.1
-msgspec
-commonmark # Required by sphinx-argparse when using :markdownhelp:
-
-# packages to install to build the documentation
-cachetools
-# -f https://download.pytorch.org/whl/cpu
-torch

docs/seed_parameter_behavior.md

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

docs/sliding_tile_attention/demo.md

@@ -1,4 +1,3 @@
-(sta-demo)=
 
 # 🔍 Demo
 This is is a demo for 2D STA with window size (6,6) operating on a (10, 10) image.

docs/sliding_tile_attention/installation.md

@@ -1,4 +1,3 @@
-(sta-installation)=
 
 # 🔧 Installation
 You can install the Sliding Tile Attention package using
@@ -8,7 +7,7 @@ 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.
+We test our code on Pytorch 2.5.0 and CUDA>=12.4. Currently, we only have an implementation for H100s.
 First, install C++20 for ThunderKittens:
 
 ```bash

docs/source/_static/custom.css

@@ -1,8 +0,0 @@
-.vertical-table-header th.head:not(.stub) {
-    writing-mode: sideways-lr;
-    white-space: nowrap;
-    max-width: 0;
-    p {
-       margin: 0;
-    }
-}

docs/source/_templates/sections/header.html

@@ -1,39 +0,0 @@
-<style>
-  .notification-bar {
-    width: 100vw;
-    display: flex;
-    justify-content: center;
-    align-items: center;
-    font-size: 16px;
-    padding: 0 6px 0 6px;
-  }
-  .notification-bar p {
-    margin: 0;
-  }
-  .notification-bar a {
-    font-weight: bold;
-    text-decoration: none;
-  }
-
-  /* Light mode styles (default) */
-  .notification-bar {
-    background-color: #fff3cd;
-    color: #856404;
-  }
-  .notification-bar a {
-    color: #d97706;
-  }
-
-  /* Dark mode styles */
-  html[data-theme=dark] .notification-bar {
-    background-color: #333;
-    color: #ddd;
-  }
-  html[data-theme=dark] .notification-bar a {
-    color: #ffa500; /* Brighter color for visibility */
-  }
-</style>
-
-<!-- <div class="notification-bar">
-  <p>You are viewing the latest developer preview docs. <a href="https://docs.vllm.ai/en/stable/">Click here</a> to view docs for the latest stable release.</p>
-</div> -->

docs/source/api/summary.md

@@ -1,19 +0,0 @@
-# Summary
-
-## Video Generator
-
-```{autodoc2-summary}
-    fastvideo.VideoGenerator
-```
-
-## Initialization Configuration
-
-```{autodoc2-summary}
-    fastvideo.configs.pipelines.PipelineConfig
-```
-
-## Sampling Configuration
-
-```{autodoc2-summary}
-    fastvideo.configs.sample.SamplingParam
-```

docs/source/autodoc2_docstring_parser.py

@@ -1,22 +0,0 @@
-# type: ignore
-# SPDX-License-Identifier: Apache-2.0
-from docutils import nodes
-from myst_parser.parsers.sphinx_ import MystParser
-from sphinx.ext.napoleon import docstring
-
-
-class NapoleonParser(MystParser):
-
-    def parse(self, input_string: str, document: nodes.document) -> None:
-        # Get the Sphinx configuration
-        config = document.settings.env.config
-
-        parsed_content = str(
-            docstring.GoogleDocstring(
-                str(docstring.NumpyDocstring(input_string, config)),
-                config,
-            ))
-        return super().parse(parsed_content, document)
-
-
-Parser = NapoleonParser