fp8

Co-authored-by: roywang <roywang@amd.com>

.buildkite/pipeline.yml

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

.buildkite/scripts/pr_test.sh

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

.github/workflows/docs.yml

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

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

@@ -0,0 +1,207 @@
+name: Publish FastVideo Kernel to PyPI on Version Change
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - "fastvideo-kernel/pyproject.toml"
+  workflow_dispatch:
+
+jobs:
+  check-version-change:
+    runs-on: ubuntu-latest
+    outputs:
+      version-changed: ${{ steps.check-version.outputs.changed }}
+      new-version: ${{ steps.check-version.outputs.new-version }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 2
+
+      - name: Check if version changed
+        id: check-version
+        run: |
+          cd fastvideo-kernel
+          # Get current commit's version from pyproject.toml
+          # Use ^ to match start of line to avoid matching minimum-version
+          NEW_VERSION=$(grep -oP '^version\s*=\s*"\K[^"]+' pyproject.toml)
+          echo "New version: $NEW_VERSION"
+          
+          # Get previous version from git history
+          # Note: git show expects path relative to repo root
+          OLD_VERSION=$(git show HEAD~1:fastvideo-kernel/pyproject.toml | grep -oP '^version\s*=\s*"\K[^"]+' || echo "0.0.0")
+          echo "Old version: $OLD_VERSION"
+          
+          if [ "$NEW_VERSION" != "$OLD_VERSION" ]; then
+            echo "Version changed from $OLD_VERSION to $NEW_VERSION"
+            echo "changed=true" >> $GITHUB_OUTPUT
+            echo "new-version=$NEW_VERSION" >> $GITHUB_OUTPUT
+          else
+            echo "Version did not change"
+            echo "changed=false" >> $GITHUB_OUTPUT
+          fi
+
+  build_wheels:
+    name: Build Wheel
+    needs: check-version-change
+    if: ${{ needs.check-version-change.outputs.version-changed == 'true' || github.event_name == 'workflow_dispatch' }}
+    runs-on: ${{ matrix.os }}
+
+    strategy:
+      fail-fast: false
+      matrix:
+          os: [ubuntu-22.04]
+          python-version: ['3.10', '3.11', '3.12']
+          torch-cuda:
+            # - torch-version: '2.5.1'
+            #   cuda-version: '12.4.1'
+            #   torch-cuda-short: 'cu124'
+            # - torch-version: '2.6.0'
+            #   cuda-version: '12.6.3'
+            #   torch-cuda-short: 'cu126'
+            # - torch-version: '2.7.1'
+            #   cuda-version: '12.8.0'
+            #   torch-cuda-short: 'cu128'
+            - torch-version: '2.9.1'
+              cuda-version: '12.8.0'
+              torch-cuda-short: 'cu128'
+
+    steps:
+      - name: Free up disk space
+        run: |
+          echo "Initial disk space:"
+          df -h
+          
+          # Remove large directories
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /usr/local/share/boost
+          sudo rm -rf /usr/share/swift
+          sudo rm -rf /usr/local/lib/node_modules
+          sudo rm -rf /usr/local/share/powershell
+          sudo rm -rf /usr/share/rust
+          sudo rm -rf /usr/local/.ghcup
+          
+          # Remove cached files
+          sudo rm -rf /var/lib/apt/lists/*
+          sudo rm -rf /var/cache/apt/archives/*
+          
+          echo "Disk space after cleanup:"
+          df -h
+
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install CUDA ${{ matrix.torch-cuda.cuda-version }}
+        uses: Jimver/cuda-toolkit@v0.2.21
+        id: cuda-toolkit
+        with:
+          cuda: ${{ matrix.torch-cuda.cuda-version }}
+          linux-local-args: '["--toolkit"]'
+          method: 'network'
+
+      - name: Install dependencies (GCC, Clang, CUDA Paths, Git)
+        run: |
+          sudo apt update
+          sudo apt install -y git patchelf gcc-11 g++-11 clang-11
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+
+          # Allow Git to Access Safe Directory
+          git config --global --add safe.directory /__w/FastVideo/FastVideo
+
+          # Set CUDA environment variables
+          export CUDA_HOME=/usr/local/cuda-${{ matrix.torch-cuda.cuda-version }}
+          export PATH=${CUDA_HOME}/bin:${PATH}
+          export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+          
+          # Verify installation
+          gcc --version
+          g++ --version
+          clang-11 --version
+          nvcc --version
+
+      - name: Install PyTorch ${{ matrix.torch-cuda.torch-version }}+cu${{ matrix.torch-cuda.cuda-version }}
+        run: |
+          pip install --upgrade pip
+          pip install typing-extensions==4.12.2
+          pip install --no-cache-dir torch==${{ matrix.torch-cuda.torch-version }} --index-url https://download.pytorch.org/whl/${{matrix.torch-cuda.torch-cuda-short}}
+          nvcc --version
+          python --version
+          python -c "import torch; print('PyTorch:', torch.__version__)"
+          python -c "import torch; print('CUDA:', torch.version.cuda)"
+          python -c "from torch.utils import cpp_extension; print (cpp_extension.CUDA_HOME)"
+
+      - name: Build wheel
+        run: |
+          export PYTHONPATH=$GITHUB_WORKSPACE:$PYTHONPATH
+          
+          pip install setuptools ninja packaging wheel triton scikit-build-core cmake build
+          
+          cd fastvideo-kernel
+          git submodule update --init --recursive  # Ensure ThunderKittens submodule is initialized
+          # Release builds are produced on GPU-less runners, so force-enable TK and target Hopper.
+          export TORCH_CUDA_ARCH_LIST="9.0a"
+          export CMAKE_ARGS="${CMAKE_ARGS:-} -DFASTVIDEO_KERNEL_BUILD_TK=ON -DCMAKE_CUDA_ARCHITECTURES=90a"
+          
+          # Build standard wheel (no local version suffix) for PyPI
+          python -m build --wheel --outdir dist
+          
+          # Fix the wheel to be manylinux compliant
+          pip install auditwheel
+          # Target manylinux_2_35 (Ubuntu 22.04 native)
+          auditwheel repair dist/*.whl --plat manylinux_2_35_x86_64 -w fixed_dist
+          # Move fixed wheels back to dist for upload consistency
+          rm dist/*.whl
+          mv fixed_dist/*.whl dist/
+
+      - name: Upload wheel artifact
+        # Only upload if it's the "main" CUDA version we want on PyPI
+        # We upload all to artifacts for inspection/GH releases, but give them distinct artifact names
+        uses: actions/upload-artifact@v4
+        with:
+          name: fastvideo_kernel-py${{ matrix.python-version }}-${{ matrix.torch-cuda.torch-cuda-short }}-torch${{ matrix.torch-cuda.torch-version }}
+          path: fastvideo-kernel/dist/*.whl
+          retention-days: 90
+
+  publish_package:
+    name: Publish package
+    needs: [build_wheels, check-version-change]
+    if: ${{ needs.check-version-change.outputs.version-changed == 'true' || github.event_name == 'workflow_dispatch' }}
+    runs-on: ubuntu-22.04
+    permissions:
+      id-token: write  # Needed for OIDC Trusted Publishing
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+
+      - name: Download PyPI wheels
+        uses: actions/download-artifact@v4
+        with:
+          path: fastvideo-kernel/dist/
+          pattern: 'fastvideo_kernel-py*'
+          merge-multiple: true
+
+      - name: Build source distribution
+        run: |
+          pip install build scikit-build-core cmake ninja
+          
+          cd fastvideo-kernel
+          # We don't need full CUDA/Torch to just package the source (sdist)
+          python -m build --sdist --outdir dist
+
+      - name: Publish release distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          packages-dir: fastvideo-kernel/dist/

.gitignore

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

.gitmodules

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

.pre-commit-config.yaml

@@ -4,12 +4,13 @@ default_stages:
 exclude: |
   (?x)(
     fastvideo/third_party/.*|
-    csrc/.*|
+    fastvideo-kernel/.*|
     assets/.*|
     tests/.*|
     demo/.*|
     predict\.py|
     scripts/.*|
+    prompts/.*|
     fastvideo/data_preprocess/.*|
     fastvideo/dataset/.*|
     fastvideo/models/.*|

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

benchmarks/fvd/__init__.py

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

benchmarks/fvd/cli.py

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

benchmarks/fvd/feature_extractors.py

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

benchmarks/fvd/fvd.py

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

benchmarks/fvd/i3d_model.py

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

benchmarks/fvd/run_fvd.py

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

benchmarks/fvd/validate_fvd.py

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

benchmarks/fvd/video_utils.py

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

benchmarks/scripts/run.sh

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

benchmarks/scripts/setup_fvd.sh

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

csrc/attn/README.md

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

csrc/attn/benchmarks/bench_sta.py

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

csrc/attn/benchmarks/bench_vsa_hopper.py

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

csrc/attn/benchmarks/bench_vsa_triton.py

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

csrc/attn/sliding_tile_attn/MANIFEST.in

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

csrc/attn/sliding_tile_attn/README.md

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

csrc/attn/sliding_tile_attn/config_sta.py

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

csrc/attn/sliding_tile_attn/setup.py

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

csrc/attn/sliding_tile_attn/st_attn.cpp

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

csrc/attn/sliding_tile_attn/st_attn/__init__.py

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

csrc/attn/tests/test_vsa.py

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

csrc/attn/tests/utils.py

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

csrc/attn/video_sparse_attn/MANIFEST.in

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

csrc/attn/video_sparse_attn/README.md

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

csrc/attn/video_sparse_attn/config_vsa.py

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

csrc/attn/video_sparse_attn/setup.py

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

csrc/attn/video_sparse_attn/vsa.cpp

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

csrc/attn/video_sparse_attn/vsa/__init__.py

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

csrc/attn/video_sparse_attn/vsa/block_sparse_wrapper.py

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

csrc/attn/vmoba_attn/README.md

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

csrc/attn/vmoba_attn/setup.py

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

csrc/attn/vmoba_attn/vmoba/__init__.py

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

docker/Dockerfile.python3.10

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

docker/Dockerfile.python3.11

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

docker/Dockerfile.python3.12

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

docker/Dockerfile.python3.12.cuda12.9.1

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

docker/Dockerfile.yaml

@@ -1,80 +0,0 @@
-FROM nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04
-
-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
-
-# Set CUDA environment variables
-ENV CUDA_HOME=/usr/local/cuda-12.8
-ENV PATH=${CUDA_HOME}/bin:${PATH}
-ENV LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
-
-# Install Miniconda and create conda environment
-COPY Miniconda3-latest-Linux-x86_64.sh /tmp/miniconda.sh
-RUN bash /tmp/miniconda.sh -b -p /opt/conda && rm /tmp/miniconda.sh
-ENV PATH=/opt/conda/bin:${PATH}
-RUN /opt/conda/bin/conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main
-RUN /opt/conda/bin/conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r
-RUN /opt/conda/bin/conda update -y -n base conda && \
-    /opt/conda/bin/conda create -y -n fastvideo python=3.12 && \
-    /opt/conda/bin/conda clean -afy
-ENV PATH=/opt/conda/envs/fastvideo/bin:/opt/conda/bin:${PATH}
-
-# Copy just the pyproject.toml first to leverage Docker cache
-COPY pyproject.toml ./
-
-# Create a dummy README to satisfy the installation
-RUN echo "# Placeholder" > README.md
-
-# Install project dependencies inside the conda environment
-RUN source /opt/conda/etc/profile.d/conda.sh && \
-    conda activate fastvideo && \
-    pip install --no-cache-dir --upgrade pip && \
-    pip install --no-cache-dir .[dev] && \
-    pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
-
-COPY . .
-
-# Install dependencies using pip inside conda env and set up shell configuration
-RUN source /opt/conda/etc/profile.d/conda.sh && \
-    conda activate fastvideo && \
-    pip install --no-cache-dir -e .[dev] && \
-    git config --unset-all http.https://github.com/.extraheader || true && \
-    echo 'source /opt/conda/etc/profile.d/conda.sh && conda activate fastvideo' >> /root/.bashrc && \
-    echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
-
-# Install STA (Sliding Tile Attention)
-RUN source /opt/conda/etc/profile.d/conda.sh && \
-    conda activate fastvideo && \
-    cd csrc/attn/sliding_tile_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
-
-# Install VSA 
-RUN source /opt/conda/etc/profile.d/conda.sh && \
-    conda activate fastvideo && \
-    cd csrc/attn/video_sparse_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
-
-EXPOSE 22
-
-ENTRYPOINT ["/bin/bash", "-lc", "source /opt/conda/etc/profile.d/conda.sh && conda activate fastvideo && exec /FastVideo/examples/inference/gradio/start.sh"]

docker/rocm7_1.Dockerfile

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

docs/README.md

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

docs/assets/custom.css

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

docs/attention/developer/index.md

@@ -0,0 +1,179 @@
+# Adding a New Attention Backend
+
+FastVideo allows integrating new attention mechanisms easily. This guide walks you through adding a new backend (e.g., `MyNewAttn`).
+
+## 1. Implement the Backend (Python)
+
+Create a new file in `fastvideo/attention/backends/` (e.g., `mynew_attn.py`).
+
+Your implementation should inherit from `AttentionBackend` defined in `abstract.py`.
+
+```python
+# fastvideo/attention/backends/mynew_attn.py
+import torch
+from .abstract import AttentionBackend
+# Import the context manager to access metadata (optional)
+from fastvideo.forward_context import get_forward_context
+
+# Import compiled kernel if applicable (see Section 2)
+try:
+    # Import from the top-level package
+    from fastvideo_kernel import my_compiled_attn_func
+except ImportError:
+    my_compiled_attn_func = None
+
+class MyNewAttnBackend(AttentionBackend):
+    def process_inputs(self, q, k, v, **kwargs):
+        # Pre-process inputs if necessary
+        return q, k, v
+
+    def forward(self, q, k, v, **kwargs):
+        # Optional: Access extra metadata passed via ForwardContext
+        # Only needed if your backend requires global state (e.g. window_size)
+        try:
+            context = get_forward_context()
+            metadata = context.attn_metadata
+            # Example: window_size = metadata.window_size
+        except (AssertionError, AttributeError):
+            # Handle case where context is not set (e.g. standard inference)
+            pass
+
+        if my_compiled_attn_func is not None:
+            return my_compiled_attn_func(q, k, v)
+        else:
+            # Fallback implementation (e.g., Triton or pure PyTorch)
+            return self.fallback_impl(q, k, v)
+```
+
+## 2. Passing Extra Information via ForwardContext (Optional)
+
+FastVideo uses a `ForwardContext` to pass global metadata (like current timestep, batch info, or custom attention configurations) to attention backends without changing the `forward` signature of every layer. **This is optional and only required if your backend needs dynamic per-step information.**
+
+To use this:
+1. **Set Context**: In your pipeline or generation loop, use the `set_forward_context` context manager.
+2. **Access Context**: Inside your attention backend, use `get_forward_context()`.
+
+See `docs/attention/sta/index.md` (Sliding Tile Attention) for an example of how complex configuration (window sizes) is passed this way.
+
+## 3. Adding Compiled Kernels (C++/CUDA)
+
+If your backend requires custom CUDA kernels, you need to add them to the `fastvideo-kernel` package.
+
+### A. Add Source Files
+Place your kernel implementation files in `fastvideo-kernel/csrc/attention/`.
+* `mynew_attn.cu` (CUDA implementation)
+* `mynew_attn.h` (Optional headers)
+
+### B. Register in Extension
+Update `fastvideo-kernel/csrc/common_extension.cpp` to expose your function to Python.
+
+```cpp
+// 1. Declare external function
+#ifdef COMPILE_MYNEW_ATTN
+extern torch::Tensor mynew_attn_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v);
+#endif
+
+// 2. Register in module
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    // ... other kernels ...
+
+#ifdef COMPILE_MYNEW_ATTN
+    m.def("mynew_attn_fwd", torch::wrap_pybind_function(mynew_attn_forward), "My New Attention Forward");
+#endif
+}
+```
+
+### C. Update CMakeLists.txt
+Update `fastvideo-kernel/CMakeLists.txt` to compile your new files.
+
+**Case 1: General CUDA Kernel (Runs on all GPUs)**
+Add your source file directly to `EXTENSION_SOURCES` and define the compilation flag.
+
+```cmake
+# Add to EXTENSION_SOURCES
+list(APPEND EXTENSION_SOURCES csrc/attention/mynew_attn.cu)
+
+# Add compilation definition for common_extension.cpp
+list(APPEND COMPILE_DEFS COMPILE_MYNEW_ATTN)
+```
+
+**Case 2: ThunderKittens Kernel (Hopper H100 Only)**
+If your kernel uses ThunderKittens (TK), it requires specific architecture flags (`sm_90a`). Add it inside the `ENABLE_TK_KERNELS` block.
+
+```cmake
+if(ENABLE_TK_KERNELS)
+    # Add source only if TK is enabled
+    list(APPEND EXTENSION_SOURCES csrc/attention/mynew_attn_tk.cu)
+    
+    # Add definition to guard registration
+    list(APPEND COMPILE_DEFS TK_COMPILE_MYNEW_ATTN)
+endif()
+```
+
+### D. Expose in Python Ops
+Update `fastvideo-kernel/python/fastvideo_kernel/ops.py` to make the function importable and handle fallbacks gracefully.
+
+```python
+# fastvideo-kernel/python/fastvideo_kernel/ops.py
+
+# Try to load C++ extension symbols
+try:
+    from fastvideo_kernel._C import fastvideo_kernel_ops
+    mynew_attn_fwd = getattr(fastvideo_kernel_ops, "mynew_attn_fwd", None)
+except ImportError:
+    mynew_attn_fwd = None
+
+def my_compiled_attn_func(q, k, v):
+    # Runtime check: use C++ kernel if available, else fallback
+    if mynew_attn_fwd is not None:
+        return mynew_attn_fwd(q, k, v)
+    else:
+        # Call Triton/Python fallback
+        return mynew_attn_triton(q, k, v)
+```
+
+### E. Expose in Package Init
+Update `fastvideo-kernel/python/fastvideo_kernel/__init__.py` to export the function.
+
+```python
+from fastvideo_kernel.ops import (
+    my_compiled_attn_func,
+    # ...
+)
+
+__all__ = [
+    "my_compiled_attn_func",
+    # ...
+]
+```
+
+## 4. Register the Backend
+
+Update `fastvideo/attention/backends/__init__.py` to export your new class.
+
+```python
+from .mynew_attn import MyNewAttnBackend
+```
+
+## 5. Platform Integration
+
+If your backend requires specific platform checks (e.g., checking for H100 support), handle that in `fastvideo/platforms/cuda.py` or within your backend's `__init__`.
+
+## 6. Add Documentation
+
+Create a new documentation page for your backend to explain its usage, installation (if custom kernels are needed), and features.
+
+1. **Create Directory**: `docs/attention/mynew_attn/`
+2. **Create Index**: `docs/attention/mynew_attn/index.md`
+3. **Update Navigation**: Add an entry to `mkdocs.yml` under the "Attention" tab.
+
+## Checklist
+
+* [ ] Created `fastvideo/attention/backends/mynew_attn.py`.
+* [ ] (Optional) Added CUDA kernels in `fastvideo-kernel/csrc/attention/`.
+* [ ] (Optional) Updated `common_extension.cpp` and `CMakeLists.txt`.
+* [ ] (Optional) Exposed kernel in `fastvideo-kernel/python/fastvideo_kernel/ops.py`.
+* [ ] (Optional) Exported kernel in `fastvideo-kernel/python/fastvideo_kernel/__init__.py`.
+* [ ] Implemented `forward` method respecting the standard signature.
+* [ ] Added unit tests in `tests/`.
+* [ ] Added documentation in `docs/attention/` and updated `mkdocs.yml`.

docs/attention/index.md

@@ -0,0 +1,53 @@
+# FastVideo Attention Kernels
+
+FastVideo provides highly optimized custom attention kernels to accelerate video generation.
+
+## Supported Kernels
+
+* **[Video Sparse Attention (VSA)](vsa/index.md)**: Sparse attention mechanism selecting top-k blocks.
+* **[Sliding Tile Attention (STA)](sta/index.md)**: Optimized attention for window-based video generation.
+
+## General Build Instructions
+
+These instructions apply to building the `fastvideo-kernel` package from source, which includes both STA and VSA kernels.
+
+### Prerequisites
+
+* **PyTorch**: 2.5.0+
+* **CUDA**: 12.4+ (12.8 recommended for best performance)
+* **C++ Compiler**: GCC 11+ (C++20 support required for ThunderKittens)
+
+Install system dependencies:
+
+```bash
+sudo apt update
+sudo apt install -y gcc-11 g++-11 clang-11 ninja-build
+
+# Set gcc-11 as default
+sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+```
+
+Set up your CUDA environment variables (adjust version as needed):
+
+```bash
+export CUDA_HOME=/usr/local/cuda-12.8
+export PATH=${CUDA_HOME}/bin:${PATH}
+export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+```
+
+### Compile and Install
+
+Clone the repository and build the kernel:
+
+```bash
+# Clone recursively to get ThunderKittens submodule
+git clone --recursive https://github.com/hao-ai-lab/FastVideo.git
+cd FastVideo/fastvideo-kernel
+
+# Build and install
+./build.sh
+```
+
+The build script automatically detects your GPU architecture:
+* **H100 (sm_90a)**: Compiles optimized C++ ThunderKittens kernels.
+* **Other (A100, etc.)**: Skips C++ compilation; installs Python package with Triton kernels.

docs/attention/sta/index.md

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

docs/attention/vsa/index.md

@@ -0,0 +1,36 @@
+# Video Sparse Attention (VSA)
+
+Sparse attention mechanism selecting top-k blocks.
+
+## Installation
+
+VSA is included in the `fastvideo-kernel` package. See the [main Attention page](../index.md) for build instructions.
+
+## Usage
+
+```python
+from fastvideo_kernel import video_sparse_attn
+
+# q, k, v: [batch_size, num_heads, seq_len, head_dim]
+# variable_block_sizes: Number of valid tokens per block
+# topk: Number of blocks to attend
+
+output = video_sparse_attn(
+    q, k, v, 
+    variable_block_sizes=block_sizes, 
+    topk=32
+)
+```
+
+## Citation
+
+If you use Video Sparse Attention in your research, please cite:
+
+```bibtex
+@article{zhang2025vsa,
+  title={Vsa: Faster video diffusion with trainable sparse attention},
+  author={Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
+  journal={arXiv preprint arXiv:2505.13389},
+  year={2025}
+}
+```

docs/contributing/overview.md

@@ -6,7 +6,7 @@ Thank you for your interest in contributing to FastVideo. We want to make the pr
 Our community is open to everyone and welcomes any contributions no matter how large or small.
 
 # Developer Environment:
-Do make sure you have CUDA 12.4 installed and supported. FastVideo currently only 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:
 
@@ -70,3 +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

@@ -106,6 +106,8 @@ def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> Forward
     return batch
 ```
 
+![Pipeline execution and data flow](../assets/images/pipeline.png)
+
 ### ForwardBatch
 
 Defined in `fastvideo/pipelines/pipeline_batch_info.py`, `ForwardBatch` encapsulates the data payload passed between pipeline stages. It typically holds:
@@ -208,6 +210,10 @@ def step(
     return prev_sample
 ```
 
+This diagram shows how models are discovered, validated, and loaded across entrypoints, executors, pipelines, and model loaders.
+
+![Model loading flow](../assets/images/load_models.png)
+
 ## Optimized Attention
 
 The `fastvideo/attention/` directory contains optimized attention implementations crucial for efficient video diffusion:
@@ -231,6 +237,8 @@ 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**

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/getting_started/quick_start.md

@@ -7,6 +7,11 @@ Get up and running with FastVideo in minutes!
 First, install FastVideo:
 
 ```bash
+# Create and activate a new conda environment
+conda create -n fastvideo python=3.12
+conda activate fastvideo
+
+# Install FastVideo
 pip install fastvideo
 ```
 
@@ -15,36 +20,53 @@ pip install fastvideo
 ### Text-to-Video Generation
 
 ```python
-from fastvideo import FastVideoPipeline
+from fastvideo import VideoGenerator
 
-# Initialize the pipeline
-pipe = FastVideoPipeline.from_pretrained("wan2.1-t2v-1.3B")
+def main():
+    # Create a video generator with a pre-trained model
+    generator = VideoGenerator.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        num_gpus=1,  # Adjust based on your hardware
+    )
 
-# Generate a video
-prompt = "A cat playing with a ball of yarn"
-video = pipe(prompt, num_frames=16, height=512, width=512)
+    # Define a prompt for your video
+    prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest."
 
-# Save the video
-video.save("output.mp4")
+    # Generate the video
+    video = generator.generate_video(
+        prompt,
+        return_frames=True,  # Also return frames from this call (defaults to False)
+        output_path="my_videos/",  # Controls where videos are saved
+        save_video=True
+    )
+
+if __name__ == '__main__':
+    main()
 ```
 
 ### Image-to-Video Generation
 
 ```python
-from fastvideo import FastVideoPipeline
-from PIL import Image
+from fastvideo import VideoGenerator, SamplingParam
 
-# Load an image
-image = Image.open("input.jpg")
+def main():
+    # Create the generator
+    model_name = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+    generator = VideoGenerator.from_pretrained(model_name, num_gpus=1)
 
-# Initialize the pipeline
-pipe = FastVideoPipeline.from_pretrained("wan2.1-i2v-14B-480p")
+    # Set up parameters with an initial image
+    sampling_param = SamplingParam.from_pretrained(model_name)
+    sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+    sampling_param.num_frames = 107
 
-# Generate a video from the image
-video = pipe(image, num_frames=16, height=480, width=480)
+    # Generate video based on the image
+    prompt = "A photograph coming to life with gentle movement"
+    generator.generate_video(prompt, sampling_param=sampling_param,
+                             output_path="my_videos/",
+                             save_video=True)
 
-# Save the video
-video.save("output.mp4")
+if __name__ == '__main__':
+    main()
 ```
 
 ## Next Steps
@@ -53,3 +75,4 @@ video.save("output.mp4")
 - [Configuration](../inference/configuration.md) - Learn about configuration options
 - [Examples](../inference/examples/) - Explore more examples
 - [Optimizations](../inference/optimizations.md) - Performance optimization tips
+- [Low VRAM Inference](../inference/low_vram_inference.md) - Memory-saving settings (CPU offload, sharded loading, etc.)

docs/index.md

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

docs/inference/inference_quick_start.md

@@ -98,7 +98,7 @@ Common issues and their solutions:
 ### Out of Memory Errors
 If you encounter CUDA out of memory errors:
 - Reduce `num_frames` or video resolution
-- Enable memory optimization with `enable_model_cpu_offload`
+- Enable memory optimization with CPU-offload and sharded loading flags (see [Low VRAM Inference](low_vram_inference.md))
 - Try a smaller model or use distilled versions
 - Use `num_gpus` > 1 if multiple GPUs are available
 

docs/seed_parameter_behavior.md

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

docs/sliding_tile_attention/demo.md

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

docs/sliding_tile_attention/installation.md

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

docs/training/data_preprocess.md

@@ -30,7 +30,7 @@ path_to_your_dataset_folder/
 └── prompt.txt
 ```
 
-To geranate the `videos2caption.json` and `merge.txt`, run
+To generate the `videos2caption.json` and `merge.txt`, run
 
 ``` python
 python scripts/dataset_preparation/prepare_json_file.py --data_folder mini_i2v_dataset/crush-smol_raw/ --output your_output_folder

docs/video_sparse_attention/installation.md

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

examples/inference/basic/README.md

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

examples/inference/basic/basic_hy15.py

@@ -0,0 +1,42 @@
+from fastvideo import VideoGenerator
+import json
+# from fastvideo.configs.sample import SamplingParam
+
+OUTPUT_PATH = "video_samples_hy15"
+def main():
+    # FastVideo will automatically use the optimal default arguments for the
+    # model.
+    # If a local path is provided, FastVideo will make a best effort
+    # attempt to identify the optimal arguments.
+    generator = VideoGenerator.from_pretrained(
+        "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v",
+        # FastVideo will automatically handle distributed setup
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
+        # image_encoder_cpu_offload=False,
+    )
+
+    prompt = (
+        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+        "wide with interest. The playful yet serene atmosphere is complemented by soft "
+        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
+    )
+
+    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, negative_prompt="", num_frames=81, fps=16)
+
+    prompt2 = (
+        "A majestic lion strides across the golden savanna, its powerful frame "
+        "glistening under the warm afternoon sun. The tall grass ripples gently in "
+        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
+        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
+        "cinematic.")
+
+    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True, negative_prompt="", num_frames=81, fps=16)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_matrixgame.py

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

examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py

@@ -0,0 +1,44 @@
+# NOTE: This is still a work in progress, and the checkpoints are not released yet.
+
+from fastvideo import VideoGenerator, SamplingParam
+import json
+# from fastvideo.configs.sample import SamplingParam
+
+OUTPUT_PATH = "video_samples_self_forcing_causal_wan2_2_14B_i2v"
+def main():
+    # FastVideo will automatically use the optimal default arguments for the
+    # model.
+    # If a local path is provided, FastVideo will make a best effort
+    # attempt to identify the optimal arguments.
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers",
+        # FastVideo will automatically handle distributed setup
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True, # DiT need to be offloaded for MoE
+        dit_precision="fp32",
+        vae_cpu_offload=False,
+        text_encoder_cpu_offload=True,
+        dmd_denoising_steps=[1000, 850, 700, 550, 350, 275, 200, 125],
+        # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer
+        pin_cpu_memory=True,
+        # image_encoder_cpu_offload=False,
+    )
+
+    sampling_param = SamplingParam.from_pretrained("FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers")
+    sampling_param.num_frames = 81
+    sampling_param.width = 832
+    sampling_param.height = 480
+    sampling_param.seed = 1000
+
+    with open("prompts/mixkit_i2v.jsonl", "r") as f:
+        prompt_image_pairs = json.load(f)
+
+    for prompt_image_pair in prompt_image_pairs:
+        prompt = prompt_image_pair["prompt"]
+        image_path = prompt_image_pair["image_path"]
+        _ = generator.generate_video(prompt, image_path=image_path, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_self_forcing_causal_wan2_2_t2v.py

@@ -0,0 +1,43 @@
+# NOTE: This is still a work in progress, and the checkpoints are not released yet.
+
+from fastvideo import VideoGenerator
+
+# from fastvideo.configs.sample import SamplingParam
+
+OUTPUT_PATH = "video_samples_self_forcing_causal_wan2_2_14B_t2v"
+def main():
+    # FastVideo will automatically use the optimal default arguments for the
+    # model.
+    # If a local path is provided, FastVideo will make a best effort
+    # attempt to identify the optimal arguments.
+    generator = VideoGenerator.from_pretrained(
+        "rand0nmr/SFWan2.2-T2V-A14B-Diffusers",
+        # FastVideo will automatically handle distributed setup
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True, # DiT need to be offloaded for MoE
+        vae_cpu_offload=False,
+        text_encoder_cpu_offload=True,
+        dmd_denoising_steps=[1000, 850, 700, 550, 350, 275, 200, 125],
+        # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer
+        pin_cpu_memory=True,
+        init_weights_from_safetensors="/mnt/sharefs/users/hao.zhang/wei/SFwan2.2_distill_self_forcing_release_cfg2/checkpoint-246_weight_only/generator_inference_transformer/",
+        init_weights_from_safetensors_2="/mnt/sharefs/users/hao.zhang/wei/SFwan2.2_distill_self_forcing_release_cfg2/checkpoint-246_weight_only/generator_2_inference_transformer/",
+        num_frame_per_block=7,
+        # image_encoder_cpu_offload=False,
+    )
+
+    # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
+    # sampling_param.num_frames = 45
+    # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+    # Generate videos with the same simple API, regardless of GPU count
+    prompt = (
+        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+        "wide with interest. The playful yet serene atmosphere is complemented by soft "
+        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
+    )
+    _ = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, num_frames=81)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/gradio/serving/gradio_frontend.py

@@ -3,6 +3,9 @@ import os
 import requests
 import base64
 import time
+import json
+from pathlib import Path
+import tempfile
 
 import gradio as gr
 
@@ -12,6 +15,7 @@ from fastvideo.configs.sample.base import SamplingParam
 MODEL_PATH_MAPPING = {
     "FastWan2.1-T2V-1.3B": "FastVideo/FastWan2.1-T2V-1.3B-Diffusers",
     "FastWan2.2-TI2V-5B-FullAttn": "FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers",
+    "CausalWan2.2-I2V-A14B-Preview": "FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers",
 }
 
 
@@ -37,7 +41,7 @@ class RayServeClient:
                 f"{self.backend_url}/generate_video",
                 json=request_data,
                 headers=headers,
-                timeout=300
+                timeout=900  # 15 minutes timeout for longer video generation
             )
             
             round_trip_time = time.time() - start_time
@@ -81,49 +85,78 @@ def save_video_from_base64(video_data: str, output_dir: str, prompt: str) -> str
         return None
 
 
-def create_timing_display(inference_time, encoding_time, network_time, total_time, stage_execution_times, num_frames):
-    dit_denoising_time = f"{stage_execution_times[5]:.2f}s" if len(stage_execution_times) > 5 else "N/A"
+def encode_image_to_base64(image_path: str) -> str:
+    """Encode an image file to base64 string."""
+    if not image_path or not os.path.exists(image_path):
+        return None
     
-    timing_html = f"""
-    <div style="margin: 10px 0;">
-        <h3 style="text-align: center; margin-bottom: 10px;">⏱️ Timing Breakdown</h3>
-        <div style="display: grid; grid-template-columns: repeat(5, 1fr); gap: 10px; margin-bottom: 10px;">
-            <div class="timing-card timing-card-highlight">
-                <div style="font-size: 20px;">🚀</div>
-                <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">DiT Denoising</div>
-                <div style="font-size: 16px; color: #ffa200; font-weight: bold;">{dit_denoising_time}</div>
-            </div>
-            <div class="timing-card">
-                <div style="font-size: 20px;">🧠</div>
-                <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">E2E (w. vae/text encoder)</div>
-                <div style="font-size: 16px; color: #2563eb;">{inference_time:.2f}s</div>
-            </div>
-            <div class="timing-card">
-                <div style="font-size: 20px;">🎬</div>
-                <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">Video Encoding</div>
-                <div style="font-size: 16px; color: #dc2626;">{encoding_time:.2f}s</div>
-            </div>
-            <div class="timing-card">
-                <div style="font-size: 20px;">🌐</div>
-                <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">Network Transfer</div>
-                <div style="font-size: 16px; color: #059669;">{network_time:.2f}s</div>
-            </div>
-            <div class="timing-card">
-                <div style="font-size: 20px;">📊</div>
-                <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">Total Processing</div>
-                <div style="font-size: 18px; color: #0277bd;">{total_time:.2f}s</div>
-            </div>
-        </div>"""
-    
-    if inference_time > 0:
-        fps = num_frames / inference_time
-        timing_html += f"""
-        <div class="performance-card" style="margin-top: 15px;">
-            <span style="font-weight: bold;">Generation Speed: </span>
-            <span style="font-size: 18px; color: #6366f1; font-weight: bold;">{fps:.1f} frames/second</span>
-        </div>"""
-    
-    return timing_html + "</div>"
+    try:
+        with open(image_path, 'rb') as f:
+            image_bytes = f.read()
+        
+        image_base64 = base64.b64encode(image_bytes).decode('utf-8')
+        
+        # Determine image type from extension
+        ext = os.path.splitext(image_path)[1].lower()
+        mime_types = {
+            '.jpg': 'image/jpeg',
+            '.jpeg': 'image/jpeg',
+            '.png': 'image/png',
+            '.gif': 'image/gif',
+            '.webp': 'image/webp',
+        }
+        mime_type = mime_types.get(ext, 'image/jpeg')
+        
+        return f"data:{mime_type};base64,{image_base64}"
+        
+    except Exception as e:
+        print(f"Failed to encode image: {e}")
+        return None
+
+
+# def create_timing_display(inference_time, encoding_time, network_time, total_time, stage_execution_times, num_frames):
+#     dit_denoising_time = f"{stage_execution_times[5]:.2f}s" if len(stage_execution_times) > 5 else "N/A"
+#     
+#     timing_html = f"""
+#     <div style="margin: 10px 0;">
+#         <h3 style="text-align: center; margin-bottom: 10px;">⏱️ Timing Breakdown</h3>
+#         <div style="display: grid; grid-template-columns: repeat(5, 1fr); gap: 10px; margin-bottom: 10px;">
+#             <div class="timing-card timing-card-highlight">
+#                 <div style="font-size: 20px;">🚀</div>
+#                 <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">DiT Denoising</div>
+#                 <div style="font-size: 16px; color: #ffa200; font-weight: bold;">{dit_denoising_time}</div>
+#             </div>
+#             <div class="timing-card">
+#                 <div style="font-size: 20px;">🧠</div>
+#                 <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">E2E (w. vae/text encoder)</div>
+#                 <div style="font-size: 16px; color: #2563eb;">{inference_time:.2f}s</div>
+#             </div>
+#             <div class="timing-card">
+#                 <div style="font-size: 20px;">🎬</div>
+#                 <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">Video Encoding</div>
+#                 <div style="font-size: 16px; color: #dc2626;">{encoding_time:.2f}s</div>
+#             </div>
+#             <div class="timing-card">
+#                 <div style="font-size: 20px;">🌐</div>
+#                 <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">Network Transfer</div>
+#                 <div style="font-size: 16px; color: #059669;">{network_time:.2f}s</div>
+#             </div>
+#             <div class="timing-card">
+#                 <div style="font-size: 20px;">📊</div>
+#                 <div style="font-weight: bold; margin: 3px 0; font-size: 14px;">Total Processing</div>
+#                 <div style="font-size: 18px; color: #0277bd;">{total_time:.2f}s</div>
+#             </div>
+#         </div>"""
+#     
+#     if inference_time > 0:
+#         fps = num_frames / inference_time
+#         timing_html += f"""
+#         <div class="performance-card" style="margin-top: 15px;">
+#             <span style="font-weight: bold;">Generation Speed: </span>
+#             <span style="font-size: 18px; color: #6366f1; font-weight: bold;">{fps:.1f} frames/second</span>
+#         </div>"""
+#     
+#     return timing_html + "</div>"
 
 
 def load_example_prompts():
@@ -144,26 +177,83 @@ def load_example_prompts():
             print(f"Warning: Could not read {filepath}: {e}")
         return prompts, labels
     
-    examples, example_labels = load_from_file("prompts/prompts_final.txt")
+    # Load prompts from prompts.txt
+    examples, example_labels = load_from_file("examples/inference/gradio/serving/prompts.txt")
     
     if not examples:
         examples = ["A crowded rooftop bar buzzes with energy, the city skyline twinkling like a field of stars in the background."]
         example_labels = ["Crowded rooftop bar at night"]
     
-    return examples, example_labels
+    # Load image mappings from JSON file
+    prompt_to_image = {}
+    # Try to find the JSON file relative to project root
+    possible_json_paths = [
+        Path("prompts/mixkit_i2v.jsonl"),
+        Path(__file__).parent.parent.parent.parent / "prompts" / "mixkit_i2v.jsonl",
+    ]
+    json_path = None
+    for path in possible_json_paths:
+        if path.exists():
+            json_path = path
+            break
+    
+    if json_path and json_path.exists():
+        try:
+            with open(json_path, "r", encoding='utf-8') as f:
+                data = json.load(f)
+                # Get the project root directory (parent of prompts directory)
+                project_root = json_path.parent.parent
+                for item in data:
+                    prompt_text = item.get("prompt", "").strip()
+                    image_path = item.get("image_path", "")
+                    if prompt_text and image_path:
+                        # Resolve image path relative to project root
+                        full_image_path = project_root / image_path
+                        if full_image_path.exists():
+                            prompt_to_image[prompt_text] = str(full_image_path.absolute())
+        except Exception as e:
+            print(f"Warning: Could not load image mappings from {json_path}: {e}")
+    
+    # Create image paths list matching the prompts
+    example_images = []
+    for prompt in examples:
+        # Try exact match first
+        image_path = prompt_to_image.get(prompt)
+        if not image_path:
+            # Try fuzzy match (case-insensitive, whitespace normalized)
+            normalized_prompt = " ".join(prompt.split())
+            for json_prompt, img_path in prompt_to_image.items():
+                normalized_json = " ".join(json_prompt.split())
+                if normalized_prompt.lower() == normalized_json.lower():
+                    image_path = img_path
+                    break
+        example_images.append(image_path if image_path and os.path.exists(image_path) else None)
+    
+    return examples, example_labels, example_images
 
 
 def create_gradio_interface(backend_url: str, default_params: dict[str, SamplingParam]):
     
     client = RayServeClient(backend_url)
     
+    def is_i2v_model(model_name: str) -> bool:
+        """Check if the model is an I2V model."""
+        return "I2V" in model_name
+    
     def generate_video(
-        prompt, negative_prompt, use_negative_prompt, seed, guidance_scale,
-        num_frames, height, width, randomize_seed, model_selection, progress
+        prompt, negative_prompt, use_negative_prompt, guidance_scale,
+        num_frames, height, width, model_selection, input_image, progress
     ):
+        # Use default seed value (randomize_seed disabled)
+        seed = 1000
+        randomize_seed = False
         if not client.check_health():
             return None, f"Backend is not available. Please check if Ray Serve is running at {backend_url}", ""
         
+        # Check if I2V model requires an image
+        if is_i2v_model(model_selection) and not input_image:
+            return None, "I2V models require an input image. Please upload an image.", ""
+        
         # Validate dimensions
         max_pixels = 720 * 1280
         if height * width > max_pixels:
@@ -172,6 +262,15 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
         if progress:
             progress(0.1, desc="Checking backend health...")
         
+        # Encode image if provided
+        image_data = None
+        if input_image:
+            if progress:
+                progress(0.2, desc="Encoding input image...")
+            image_data = encode_image_to_base64(input_image)
+            if not image_data:
+                return None, "Failed to encode input image", ""
+        
         request_data = {
             "prompt": prompt,
             "negative_prompt": negative_prompt,
@@ -183,7 +282,7 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
             "width": width,
             "randomize_seed": randomize_seed,
             "return_frames": False,
-            "image_path": None,
+            "image_data": image_data,
             "model_path": MODEL_PATH_MAPPING.get(model_selection, "FastVideo/FastWan2.1-T2V-1.3B-Diffusers")
         }
         
@@ -198,16 +297,16 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
         if response.get("success", False):
             video_data = response.get("video_data", "")
             used_seed = response.get("seed", seed)
-            inference_time = response.get("inference_time", 0.0)
-            encoding_time = response.get("encoding_time", 0.0)
-            total_time = response.get("total_time", 0.0)
-            network_time = response.get("network_time", 0.0)
-            stage_execution_times = response.get("stage_execution_times", [])
+            # inference_time = response.get("inference_time", 0.0)
+            # encoding_time = response.get("encoding_time", 0.0)
+            # total_time = response.get("total_time", 0.0)
+            # network_time = response.get("network_time", 0.0)
+            # stage_execution_times = response.get("stage_execution_times", [])
             
-            timing_details = create_timing_display(
-                inference_time, encoding_time, network_time, total_time, 
-                stage_execution_times, num_frames
-            )
+            # timing_details = create_timing_display(
+            #     inference_time, encoding_time, network_time, total_time, 
+            #     stage_execution_times, num_frames
+            # )
             
             if video_data:
                 if progress:
@@ -219,7 +318,7 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
                     progress(1.0, desc="Generation complete!")
                 
                 if video_path and os.path.exists(video_path):
-                    return video_path, used_seed, timing_details
+                    return video_path, used_seed, ""
                 else:
                     return None, "Failed to save video", ""
             else:
@@ -228,7 +327,7 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
             error_msg = response.get("error_message", "Unknown error occurred")
             return None, f"Generation failed: {error_msg}", ""
     
-    examples, example_labels = load_example_prompts()
+    examples, example_labels, example_images = load_example_prompts()
     
     theme = gr.themes.Base().set(
         button_primary_background_fill="#2563eb",
@@ -239,33 +338,39 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
     )
     
     def get_default_values(model_name):
-        model_path = MODEL_PATH_MAPPING.get(model_name)
-        if model_path and model_path in default_params:
-            params = default_params[model_path]
-            return {
-                'height': params.height,
-                'width': params.width,
-                'num_frames': params.num_frames,
-                'guidance_scale': params.guidance_scale,
-                'seed': params.seed,
-            }
+        # model_path = MODEL_PATH_MAPPING.get(model_name)
+        # if model_path and model_path in default_params:
+        #     params = default_params[model_path]
+        #     return {
+        #         'height': params.height,
+        #         'width': params.width,
+        #         'num_frames': params.num_frames,
+        #         'guidance_scale': params.guidance_scale,
+        #     }
         
         return {
-            'height': 448,
+            'height': 480,
             'width': 832,
-            'num_frames': 61,
-            'guidance_scale': 3.0,
-            'seed': 1024,
+            'num_frames': 73,
         }
     
-    initial_values = get_default_values("FastWan2.1-T2V-1.3B")
+    # Get available models based on what's loaded
+    available_models = []
+    for model_name, model_path in MODEL_PATH_MAPPING.items():
+        if model_path in default_params:
+            available_models.append(model_name)
     
-    with gr.Blocks(title="FastWan", theme=theme) as demo:
+    # Select first available model as default
+    default_model = available_models[0] if available_models else "FastWan2.1-T2V-1.3B"
+    initial_values = get_default_values(default_model)
+    initial_show_image = is_i2v_model(default_model)
+    
+    with gr.Blocks(title="CausalWan", theme=theme) as demo:
         gr.Image("assets/logos/logo.svg", show_label=False, container=False, height=80)
         gr.HTML("""
         <div style="text-align: center; margin-bottom: 10px;">
             <p style="font-size: 18px;"> Make Video Generation Go Blurrrrrrr </p>
-            <p style="font-size: 18px;"> <a href="https://github.com/hao-ai-lab/FastVideo/tree/main" target="_blank">Code</a> | <a href="https://hao-ai-lab.github.io/blogs/fastvideo_post_training/" target="_blank">Blog</a> | <a href="https://hao-ai-lab.github.io/FastVideo/" target="_blank">Docs</a>  </p>
+            <p style="font-size: 18px;"> <a href="https://github.com/hao-ai-lab/FastVideo/tree/main" target="_blank">Code</a> | <a href="https://hao-ai-lab.github.io/blogs/fastvideo_causalwan_preview/" target="_blank">Blog</a> | <a href="https://hao-ai-lab.github.io/FastVideo/" target="_blank">Docs</a>  </p>
         </div>
         """)
         
@@ -280,8 +385,8 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
         
         with gr.Row():
             model_selection = gr.Dropdown(
-                choices=list(MODEL_PATH_MAPPING.keys()),
-                value="FastWan2.1-T2V-1.3B",
+                choices=available_models,
+                value=default_model,
                 label="Select Model",
                 interactive=True
             )
@@ -312,69 +417,70 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
         with gr.Row():
             with gr.Column():
                 error_output = gr.Text(label="Error", visible=False)
-                timing_display = gr.Markdown(label="Timing Breakdown", visible=False)
+                # timing_display = gr.Markdown(label="Timing Breakdown", visible=False)
 
-        with gr.Row(equal_height=True, elem_classes="main-content-row"):
-            with gr.Column(scale=1, elem_classes="advanced-options-column"):
-                with gr.Group():
-                    gr.HTML("<div style='margin: 0 0 15px 0; text-align: center; font-size: 16px;'>Advanced Options</div>")
-                    with gr.Row():
-                        height = gr.Number(
-                            label="Height",
-                            value=initial_values['height'],
-                            interactive=False,
-                            container=True
-                        )
-                        width = gr.Number(
-                            label="Width",
-                            value=initial_values['width'],
-                            interactive=False,
-                            container=True
+        with gr.Row(equal_height=False):
+            with gr.Column(scale=1):
+                with gr.Tabs():
+                    with gr.Tab("Input Image", visible=initial_show_image) as image_tab:
+                        gr.Markdown("**Please make sure you upload a 480x832 image**")
+                        input_image = gr.Image(
+                            label="",
+                            type="filepath",
+                            height=400,
                         )
                     
-                    with gr.Row():
-                        num_frames = gr.Number(
-                            label="Number of Frames",
-                            value=initial_values['num_frames'],
-                            interactive=False,
-                            container=True
-                        )
-                        guidance_scale = gr.Slider(
-                            label="Guidance Scale",
-                            minimum=1,
-                            maximum=12,
-                            value=initial_values['guidance_scale'],
-                        )
-                    
-                    with gr.Row():
-                        use_negative_prompt = gr.Checkbox(
-                            label="Use negative prompt", value=False)
-                        negative_prompt = gr.Text(
-                            label="Negative prompt",
-                            max_lines=3,
-                            lines=3,
-                            placeholder="Enter a negative prompt",
-                            visible=False,
-                        )
+                    with gr.Tab("Advanced Options"):
+                        with gr.Group():
+                            with gr.Row():
+                                height = gr.Number(
+                                    label="Height",
+                                    value=initial_values['height'],
+                                    interactive=False,
+                                    container=True
+                                )
+                                width = gr.Number(
+                                    label="Width",
+                                    value=initial_values['width'],
+                                    interactive=False,
+                                    container=True
+                                )
+                            
+                            with gr.Row():
+                                num_frames = gr.Number(
+                                    label="Number of Frames",
+                                    value=initial_values['num_frames'],
+                                    interactive=False,
+                                    container=True
+                                )
+                                guidance_scale = gr.Number(
+                                    label="Guidance Scale",
+                                    value=1.0,
+                                    interactive=False,
+                                    container=True
+                                )
+                            
+                            with gr.Row():
+                                use_negative_prompt = gr.Checkbox(
+                                    label="Use negative prompt", value=False)
+                                negative_prompt = gr.Text(
+                                    label="Negative prompt",
+                                    max_lines=3,
+                                    lines=3,
+                                    placeholder="Enter a negative prompt",
+                                    visible=False,
+                                )
 
-                    seed = gr.Slider(
-                        label="Seed",
-                        minimum=0,
-                        maximum=1000000,
-                        step=1,
-                        value=initial_values['seed'],
-                    )
-                    randomize_seed = gr.Checkbox(label="Randomize seed", value=False)
-                    seed_output = gr.Number(label="Used Seed")
+                            # randomize_seed = gr.Checkbox(label="Randomize seed", value=False)
+                            seed_output = gr.Number(label="Used Seed", value=1000)
         
-            with gr.Column(scale=1, elem_classes="video-column"):
+            with gr.Column(scale=1):
                 result = gr.Video(
                     label="Generated Video", 
                     show_label=True,
-                    height=466,
-                    width=600,
+                    height=500,
                     container=True,
-                    elem_classes="video-component"
+                    autoplay=True,
                 )
         
         gr.HTML("""
@@ -387,116 +493,10 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
         }
         
         .gradio-container {
-            max-width: 1200px !important;
+            max-width: 1400px !important;
             margin: 0 auto !important;
         }
         
-        .main {
-            max-width: 1200px !important;
-            margin: 0 auto !important;
-        }
-        
-        .gr-form, .gr-box, .gr-group {
-            max-width: 1200px !important;
-        }
-        
-        .gr-video {
-            max-width: 500px !important;
-            margin: 0 auto !important;
-        }
-        
-        .main-content-row {
-            display: flex !important;
-            align-items: flex-start !important;
-            min-height: 500px !important;
-            gap: 20px !important;
-        }
-        
-        .advanced-options-column,
-        .video-column {
-            display: flex !important;
-            flex-direction: column !important;
-            flex: 1 !important;
-            min-height: 400px !important;
-            align-items: stretch !important;
-        }
-        
-        .video-column > * {
-            margin-top: 0 !important;
-        }
-        
-        .video-column .gr-video,
-        .video-component {
-            margin-top: 0 !important;
-            padding-top: 0 !important;
-        }
-        
-        .video-column .gr-video .gr-form {
-            margin-top: 0 !important;
-        }
-        
-        .advanced-options-column .gr-group,
-        .video-column .gr-video {
-            margin-top: 0 !important;
-            vertical-align: top !important;
-        }
-        
-        .advanced-options-column > *:last-child,
-        .video-column > *:last-child {
-            flex-grow: 0 !important;
-        }
-        
-        @media (max-width: 1400px) {
-            .main-content-row {
-                min-height: 600px !important;
-            }
-            
-            .advanced-options-column,
-            .video-column {
-                min-height: 600px !important;
-            }
-        }
-        
-        @media (max-width: 1200px) {
-            .main-content-row {
-                flex-direction: column !important;
-                align-items: stretch !important;
-            }
-            
-            .advanced-options-column,
-            .video-column {
-                min-height: auto !important;
-                width: 100% !important;
-            }
-        }
-        
-        .timing-card {
-            background: var(--background-fill-secondary) !important;
-            border: 1px solid var(--border-color-primary) !important;
-            color: var(--body-text-color) !important;
-            padding: 10px;
-            border-radius: 8px;
-            text-align: center;
-            min-height: 80px;
-            display: flex;
-            flex-direction: column;
-            justify-content: center;
-        }
-        
-        .timing-card-highlight {
-            background: var(--background-fill-primary) !important;
-            border: 2px solid var(--color-accent) !important;
-        }
-        
-        .performance-card {
-            background: var(--background-fill-secondary) !important;
-            border: 1px solid var(--border-color-primary) !important;
-            color: var(--body-text-color) !important;
-            padding: 10px;
-            border-radius: 6px;
-            text-align: center;
-        }
-        
         .gr-number input[readonly] {
             background-color: var(--background-fill-secondary) !important;
             border: 1px solid var(--border-color-primary) !important;
@@ -511,18 +511,20 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
         def on_example_select(example_label):
             if example_label and example_label in example_labels:
                 index = example_labels.index(example_label)
-                return examples[index]
-            return ""
+                selected_prompt = examples[index]
+                selected_image = example_images[index] if index < len(example_images) else None
+                return selected_prompt, selected_image
+            return "", None
         
         example_dropdown.change(
             fn=on_example_select,
             inputs=example_dropdown,
-            outputs=prompt,
+            outputs=[prompt, input_image],
         )
         
         gr.HTML("""
         <div style="text-align: center; margin-top: 10px; margin-bottom: 15px;">
-            <p style="font-size: 16px; margin: 0;">The compute for this demo is generously provided by <a href="https://www.gmicloud.ai/" target="_blank">GMI Cloud</a>. Note that this demo is meant to showcase FastWan's quality and that under a large number of requests, generation speed may be affected. We are also rate-limiting users to 3 requests per minute.</p>
+            <p style="font-size: 16px; margin: 0;">The compute for this demo is generously provided by <a href="https://www.gmicloud.ai/" target="_blank">GMI Cloud</a>.  Note that this demo is meant as a preview of our distilled I2V model. Outside of few-step distillation, we have not yet fully optimized it for speed. Stay tuned for updates!</p>
         </div>
         """)
         
@@ -537,6 +539,7 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
                 selected_model = "FastWan2.1-T2V-1.3B"
             
             model_path = MODEL_PATH_MAPPING.get(selected_model)
+            show_image_input = is_i2v_model(selected_model)
             
             if model_path and model_path in default_params:
                 params = default_params[model_path]
@@ -545,29 +548,29 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
                     gr.update(value=params.width),
                     gr.update(value=params.num_frames),
                     gr.update(value=params.guidance_scale),
-                    gr.update(value=params.seed),
+                    gr.update(visible=show_image_input),
                 )
             
             return (
                 gr.update(value=448),
                 gr.update(value=832),
-                gr.update(value=61),
+                gr.update(value=20),
                 gr.update(value=3.0),
-                gr.update(value=1024),
+                gr.update(visible=show_image_input),
             )
         
         model_selection.change(
             fn=on_model_selection_change,
             inputs=model_selection,
-            outputs=[height, width, num_frames, guidance_scale, seed],
+            outputs=[height, width, num_frames, guidance_scale, image_tab],
         )
         
         def handle_generation(*args, progress=None, request: gr.Request = None):
-            model_selection, prompt, negative_prompt, use_negative_prompt, seed, guidance_scale, num_frames, height, width, randomize_seed = args
+            model_selection, prompt, negative_prompt, use_negative_prompt, guidance_scale, num_frames, height, width, input_image = args
             
-            result_path, seed_or_error, timing_details = generate_video(
-                prompt, negative_prompt, use_negative_prompt, seed, guidance_scale, 
-                num_frames, height, width, randomize_seed, model_selection, progress
+            result_path, seed_or_error, _ = generate_video(
+                prompt, negative_prompt, use_negative_prompt, guidance_scale, 
+                num_frames, height, width, model_selection, input_image, progress
             )
             
             if result_path and os.path.exists(result_path):
@@ -575,14 +578,12 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
                     result_path, 
                     seed_or_error, 
                     gr.update(visible=False),
-                    gr.update(visible=True, value=timing_details),
                 )
             else:
                 return (
                     None, 
                     seed_or_error, 
                     gr.update(visible=True, value=seed_or_error),
-                    gr.update(visible=False),
                 )
         
         run_button.click(
@@ -592,14 +593,14 @@ def create_gradio_interface(backend_url: str, default_params: dict[str, Sampling
                 prompt,
                 negative_prompt,
                 use_negative_prompt,
-                seed,
                 guidance_scale,
                 num_frames,
                 height,
                 width,
-                randomize_seed,
+                # randomize_seed,
+                input_image,
             ],
-            outputs=[result, seed_output, error_output, timing_display],
+            outputs=[result, seed_output, error_output],  # timing_display removed
             concurrency_limit=20,
         )
     
@@ -611,8 +612,11 @@ def main():
     parser.add_argument("--backend_url", type=str, default="http://localhost:8000",
                         help="URL of the Ray Serve backend")
     parser.add_argument("--t2v_model_paths", type=str,
-                        default="FastVideo/FastWan2.1-T2V-1.3B-Diffusers,FastVideo/FastWan2.1-T2V-14B-Diffusers",
+                        default="",
                         help="Comma separated list of paths to the T2V model(s)")
+    parser.add_argument("--i2v_model_paths", type=str,
+                        default="FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers",
+                        help="Comma separated list of paths to the I2V model(s)")
     parser.add_argument("--host", type=str, default="0.0.0.0",
                         help="Host to bind to")
     parser.add_argument("--port", type=int, default=7860,
@@ -621,8 +625,15 @@ def main():
     args = parser.parse_args()
     
     default_params = {}
-    model_paths = args.t2v_model_paths.split(",")
-    for model_path in model_paths:
+    
+    # Load T2V model params
+    t2v_paths = [p.strip() for p in args.t2v_model_paths.split(",") if p.strip()]
+    for model_path in t2v_paths:
+        default_params[model_path] = SamplingParam.from_pretrained(model_path)
+    
+    # Load I2V model params
+    i2v_paths = [p.strip() for p in args.i2v_model_paths.split(",") if p.strip()]
+    for model_path in i2v_paths:
         default_params[model_path] = SamplingParam.from_pretrained(model_path)
     
     demo = create_gradio_interface(args.backend_url, default_params)
@@ -630,6 +641,8 @@ def main():
     print(f"Starting Gradio frontend at http://{args.host}:{args.port}")
     print(f"Backend URL: {args.backend_url}")
     print(f"T2V Models: {args.t2v_model_paths}")
+    if args.i2v_model_paths:
+        print(f"I2V Models: {args.i2v_model_paths}")
     
     from fastapi import FastAPI, Request, HTTPException
     from fastapi.responses import HTMLResponse, FileResponse
@@ -674,23 +687,23 @@ def main():
             <meta charset="UTF-8" />
             <meta name="viewport" content="width=device-width, initial-scale=1.0" />
             
-            <title>FastWan</title>
-            <meta name="title" content="FastWan">
+            <title>CausalWan</title>
+            <meta name="title" content="CausalWan">
             <meta name="description" content="Make video generation go blurrrrrrr">
-            <meta name="keywords" content="FastVideo, video generation, AI, machine learning, FastWan">
+            <meta name="keywords" content="FastVideo, video generation, AI, machine learning, CausalWan">
             
             <meta property="og:type" content="website">
             <meta property="og:url" content="{base_url}/">
-            <meta property="og:title" content="FastWan">
+            <meta property="og:title" content="CausalWan">
             <meta property="og:description" content="Make video generation go blurrrrrrr">
             <meta property="og:image" content="{base_url}/logo.svg">
             <meta property="og:image:width" content="1200">
             <meta property="og:image:height" content="630">
-            <meta property="og:site_name" content="FastWan">
+            <meta property="og:site_name" content="CausalWan">
             
             <meta property="twitter:card" content="summary_large_image">
             <meta property="twitter:url" content="{base_url}/">
-            <meta property="twitter:title" content="FastWan">
+            <meta property="twitter:title" content="CausalWan">
             <meta property="twitter:description" content="Make video generation go blurrrrrrr">
             <meta property="twitter:image" content="{base_url}/logo.svg">
             <link rel="icon" type="image/png" sizes="32x32" href="/favicon.ico">
@@ -720,7 +733,14 @@ def main():
         app, 
         demo, 
         path="/gradio",
-        allowed_paths=[os.path.abspath("outputs"), os.path.abspath("fastvideo-logos")]
+        allowed_paths=[
+            os.path.abspath("outputs"), 
+            os.path.abspath("fastvideo-logos"),
+            os.path.abspath("prompts"),
+            os.path.abspath("images"),
+            os.path.abspath(tempfile.gettempdir()),
+            os.path.abspath(os.path.join(tempfile.gettempdir(), "gradio")),
+        ]
     )
     
     uvicorn.run(app, host=args.host, port=args.port)

examples/inference/gradio/serving/prompts.txt

@@ -0,0 +1,15 @@
+Some friends dancing and having fun together in circles, at a party surrounded by colored lights at a party, in a fancy old place, in a view from below them.
+A man wearing grey shorts jumps rope in a gym, weights and gym equipment in the background.
+Flying over a peninsula covered in bushy trees, while discovering the sea around it, painted a beautiful turquoise blue, on a sunny day.
+Skillful cyclist doing a wheelie on a bike while riding through a forest, on a dirt road, surrounded by many trees, in the morning.
+In the video, a lone rider guides a majestic horse across an expansive, open field as the sun sets in the background. The rider, dressed in a classic blue shirt and wide-brimmed hat, sits confidently in the saddle, silhouetted against the warm glow of the evening sky. The horse moves gracefully, its mane and tail flowing with each step, creating a sense of harmony between horse and rider. Surrounding the pair, towering trees form a natural border, their leaves gently rustling in the breeze. The shadows lengthen on the ground, accentuating the serene and timeless feel of the scene. The distant hills and wooden fences frame the horizon, adding depth to the tranquil landscape. A few horses graze peacefully in the background, blending into the pastoral setting. The overall ambiance evokes a sense of calmness and quietude, capturing a perfect moment in the golden light of dusk.
+A silver SUV drives along a winding, snow-covered mountain road, with dense pine trees blanketed in snow lining both sides. The scene is serene, with the vehicle moving smoothly, possibly on a winter journey or vacation. As the SUV disappears around the bend, another, darker SUV follows, creating a sense of motion and perspective on the snow-dusted asphalt. The towering, snow-laden rock formation to the right contrasts with the dark green of the pines, highlighting the peacefulness of the wintry landscape.
+A man and a woman playing in a field with grass, during a bright afternoon, while cars pass by in the distance.
+A saxophonist wearing a blazer dances while playing a song in a park.
+Romantic couple embracing and looking at each other in the middle of a forest, during a break on a road trip through nature.
+Young woman cleaning her house decorated with plants and decorations, while dancing happily to music in her headphones.
+Man dressed in 80's style dances very happily in his kitchen while listening to music on his radio and drinking wine.
+Pair of jazz musicians performing a song with their saxophone and trombone on an abandoned train.
+A young woman with short hair wearing pink sunglasses chews gum and makes a bubble gum with the city in the background.
+Natural aerial landscape with a relief covered with abundant trees and vegetation and a thick layer of mist.
+Loving couple sitting on a log on the shore of a lake outside, sharing an affectionate hug.

examples/inference/gradio/serving/ray_serve_backend.py

@@ -26,6 +26,7 @@ SEED_RANGE_MAX = 1_000_000
 SUPPORTED_MODELS = [
     "FastVideo/FastWan2.1-T2V-1.3B-Diffusers", 
     "FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers",
+    "FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers",
 ]
 
 MODEL_CONFIGS = {
@@ -42,6 +43,13 @@ MODEL_CONFIGS = {
         "dit_cpu_offload": True,
         "vae_cpu_offload": False,
         "VSA_sparsity": 0.9,
+    },
+    "I2V-A14B": {
+        "num_cpus": 15,
+        "text_encoder_cpu_offload": True,
+        "dit_cpu_offload": True,
+        "vae_cpu_offload": False,
+        "VSA_sparsity": 0.0,
     }
 }
 
@@ -58,6 +66,7 @@ class VideoGenerationRequest(BaseModel):
     randomize_seed: bool = False
     return_frames: bool = False
     model_path: Optional[str] = None
+    image_data: Optional[str] = None  # Base64 encoded image for I2V
 
 
 class VideoGenerationResponse(BaseModel):
@@ -91,11 +100,38 @@ def encode_video_to_base64(frames: List[np.ndarray], fps: int = DEFAULT_FPS) ->
         return ""
 
 
+def save_image_from_base64(image_data: str, output_dir: str) -> Optional[str]:
+    """Save base64 image data to a temporary file and return the path."""
+    if not image_data:
+        return None
+    
+    try:
+        # Remove data URL prefix if present
+        if image_data.startswith('data:image/'):
+            image_data = image_data.split(',')[1]
+        
+        image_bytes = base64.b64decode(image_data)
+        
+        # Save to temporary file
+        os.makedirs(output_dir, exist_ok=True)
+        temp_image_path = os.path.join(output_dir, f"temp_input_{int(time.time() * 1000)}.png")
+        
+        with open(temp_image_path, 'wb') as f:
+            f.write(image_bytes)
+        
+        return temp_image_path
+        
+    except Exception as e:
+        print(f"Warning: Failed to save image: {e}")
+        return None
+
+
 def setup_model_environment(model_path: str) -> None:
-    if "fullattn" in model_path.lower():
-        os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "FLASH_ATTN"
-    else:
-        os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "VIDEO_SPARSE_ATTN"
+    # if "fullattn" in model_path.lower():
+    #     os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "FLASH_ATTN"
+    # else:
+    #     os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "VIDEO_SPARSE_ATTN"
+    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "FLASH_ATTN"
     os.environ["FASTVIDEO_STAGE_LOGGING"] = "1"
 
 
@@ -157,22 +193,41 @@ class BaseModelDeployment:
             num_gpus=1,
             use_fsdp_inference=True,
             text_encoder_cpu_offload=config["text_encoder_cpu_offload"],
+            dmd_denoising_steps=[1000, 850, 700, 550, 350, 275, 200, 125], # TODO: hardocde for I2V
+            dit_precision="fp32",  # TODO: hardocde for I2V
             dit_cpu_offload=config["dit_cpu_offload"],
             vae_cpu_offload=config["vae_cpu_offload"],
             VSA_sparsity=config["VSA_sparsity"],
             enable_stage_verification=False,
         )
         self.default_params = SamplingParam.from_pretrained(self.model_path)
+        self.default_params.seed = 1000
+        self.default_params.num_frames = 73
+        self.default_params.width = 832
+        self.default_params.height = 480
 
     def generate_video(self, video_request: VideoGenerationRequest) -> VideoGenerationResponse:
         total_start_time = time.time()
         
         params = prepare_sampling_params(video_request, self.default_params)
 
+        # Save image if provided (for I2V)
+        image_path = None
+        if video_request.image_data:
+            image_path = save_image_from_base64(video_request.image_data, self.output_path)
+            if image_path is None:
+                return VideoGenerationResponse(
+                    video_data=None,
+                    seed=params.seed,
+                    success=False,
+                    error_message="Failed to save input image",
+                )
+
         inference_start_time = time.time()
         result = self.generator.generate_video(
             prompt=video_request.prompt,
             sampling_param=params,
+            image_path=image_path,
             save_video=False,
             return_frames=False,
         )
@@ -185,6 +240,13 @@ class BaseModelDeployment:
         encoding_time = time.time() - encoding_start_time
         
         total_time = time.time() - total_start_time
+        
+        # Clean up temporary image file
+        if image_path and os.path.exists(image_path):
+            try:
+                os.remove(image_path)
+            except Exception as e:
+                print(f"Warning: Failed to remove temporary image file {image_path}: {e}")
 
         return VideoGenerationResponse(
             video_data=video_data,
@@ -200,7 +262,7 @@ class BaseModelDeployment:
 
 
 @serve.deployment(
-    ray_actor_options={"num_cpus": 15, "num_gpus": 1, "runtime_env": {"conda": "fastvideo"}},
+    ray_actor_options={"num_cpus": 2, "num_gpus": 1, "runtime_env": {"conda": "demo-fv"}},
 )
 class T2VModelDeployment(BaseModelDeployment):
     def __init__(self, t2v_model_path: str, output_path: str = "outputs"):
@@ -210,7 +272,7 @@ class T2VModelDeployment(BaseModelDeployment):
 
 
 @serve.deployment(
-    ray_actor_options={"num_cpus": 16, "num_gpus": 1, "runtime_env": {"conda": "fastvideo"}},
+    ray_actor_options={"num_cpus": 16, "num_gpus": 1, "runtime_env": {"conda": "demo-fv"}},
 )
 class T2V14BModelDeployment(BaseModelDeployment):
     def __init__(self, t2v_14b_model_path: str, output_path: str = "outputs"):
@@ -221,18 +283,32 @@ class T2V14BModelDeployment(BaseModelDeployment):
         print("✅ T2V 14B model initialized successfully")
 
 
+@serve.deployment(
+    ray_actor_options={"num_cpus": 15, "num_gpus": 1, "runtime_env": {"conda": "demo-fv"}},
+)
+class I2VModelDeployment(BaseModelDeployment):
+    def __init__(self, i2v_model_path: str, output_path: str = "outputs"):
+        super().__init__(i2v_model_path, output_path)
+        # Override environment for I2V model
+        os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "FLASH_ATTN"
+        self._initialize_generator(MODEL_CONFIGS["I2V-A14B"])
+        print("✅ I2V model initialized successfully")
+
+
 app = FastAPI()
 limiter = Limiter(key_func=get_remote_address)
 app.state.limiter = limiter
 app.add_exception_handler(RateLimitExceeded, _rate_limit_exceeded_handler)
 
 
-@serve.deployment(num_replicas=1, ray_actor_options={"num_cpus": 2})
+@serve.deployment(num_replicas=1, ray_actor_options={"num_cpus": 1})
 @serve.ingress(app)
 class FastVideoAPI:
 
-    def __init__(self, t2v_deployments: Dict[str, DeploymentHandle]):
+    def __init__(self, t2v_deployments: Dict[str, DeploymentHandle], i2v_deployments: Dict[str, DeploymentHandle] = None):
         self.t2v_deployments = t2v_deployments
+        self.i2v_deployments = i2v_deployments or {}
+        self.all_deployments = {**self.t2v_deployments, **self.i2v_deployments}
         
         # Initialize Prometheus metrics
         self.request_count = Counter('fastvideo_requests_total', 'Total FastVideo requests', ['model_type', 'status'])
@@ -257,10 +333,10 @@ class FastVideoAPI:
         model_name = self._get_model_name(video_request.model_path)
         
         try:
-            if video_request.model_path not in self.t2v_deployments:
+            if video_request.model_path not in self.all_deployments:
                 raise ValueError(f"Model {video_request.model_path} not found")
             
-            response_ref = self.t2v_deployments[video_request.model_path].generate_video.remote(video_request)
+            response_ref = self.all_deployments[video_request.model_path].generate_video.remote(video_request)
             response = await response_ref
             
             self._record_metrics(model_name, "success", time.time() - start_time, response)
@@ -291,18 +367,21 @@ class FastVideoAPI:
 
 
 def validate_configuration(model_paths: List[str], replicas: List[int]) -> None:
+    assert len(model_paths) > 0, "At least one model must be specified"
     assert len(model_paths) == len(replicas), "Number of models and replicas must match"
     assert sum(replicas) <= NUM_GPUS, f"Total replicas ({sum(replicas)}) must be <= {NUM_GPUS}"
     
     for model, replica_count in zip(model_paths, replicas):
-        assert model in SUPPORTED_MODELS, f"Model {model} not supported"
+        assert model in SUPPORTED_MODELS, f"Model {model} not supported. Supported models: {SUPPORTED_MODELS}"
         assert replica_count > 0, f"Replicas must be greater than 0"
 
 
 def start_ray_serve(
     *,
-    t2v_model_paths: str,
-    t2v_model_replicas: str,
+    t2v_model_paths: str = "",
+    t2v_model_replicas: str = "",
+    i2v_model_paths: str = "",
+    i2v_model_replicas: str = "",
     output_path: str = "outputs",
     host: str = "0.0.0.0",
     port: int = 8000,
@@ -310,21 +389,39 @@ def start_ray_serve(
     if not ray.is_initialized():
         ray.init()
 
-    model_paths = t2v_model_paths.split(",")
-    replicas = [int(r) for r in t2v_model_replicas.split(",")]
-    validate_configuration(model_paths, replicas)
+    # Parse T2V models
+    t2v_paths = [p.strip() for p in t2v_model_paths.split(",") if p.strip()]
+    t2v_reps = [int(r.strip()) for r in t2v_model_replicas.split(",") if r.strip()] if t2v_model_replicas else []
+    
+    # Parse I2V models
+    i2v_paths = [p.strip() for p in i2v_model_paths.split(",") if p.strip()]
+    i2v_reps = [int(r.strip()) for r in i2v_model_replicas.split(",") if r.strip()] if i2v_model_replicas else []
+    
+    # Validate configurations
+    all_paths = t2v_paths + i2v_paths
+    all_replicas = t2v_reps + i2v_reps
+    validate_configuration(all_paths, all_replicas)
 
+    # Create T2V deployments
     t2v_deps = {}
-    for model_path, replica_count in zip(model_paths, replicas):
+    for model_path, replica_count in zip(t2v_paths, t2v_reps):
         t2v_dep = T2VModelDeployment.options(num_replicas=replica_count).bind(model_path, output_path)
         t2v_deps[model_path] = t2v_dep
 
-    api = FastVideoAPI.bind(t2v_deps)
+    # Create I2V deployments
+    i2v_deps = {}
+    for model_path, replica_count in zip(i2v_paths, i2v_reps):
+        i2v_dep = I2VModelDeployment.options(num_replicas=replica_count).bind(model_path, output_path)
+        i2v_deps[model_path] = i2v_dep
+
+    api = FastVideoAPI.bind(t2v_deps, i2v_deps)
     serve.run(api, route_prefix="/", name="fast_video")
 
     print(f"Ray Serve backend started at http://{host}:{port}")
-    for model_path, replica_count in zip(model_paths, replicas):
+    for model_path, replica_count in zip(t2v_paths, t2v_reps):
         print(f"T2V Model: {model_path} | Replicas: {replica_count}")
+    for model_path, replica_count in zip(i2v_paths, i2v_reps):
+        print(f"I2V Model: {model_path} | Replicas: {replica_count}")
     print(f"Health check: http://{host}:{port}/health")
     print(f"Video generation endpoint: http://{host}:{port}/generate_video")
 
@@ -340,12 +437,20 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="FastVideo Ray Serve Backend")
     parser.add_argument("--t2v_model_paths",
                         type=str,
-                        default="FastVideo/FastWan2.1-T2V-1.3B-Diffusers,FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers",
+                        default="",
                         help="Comma separated list of paths to the T2V model(s)")
     parser.add_argument("--t2v_model_replicas",
                         type=str,
-                        default="4,4",
+                        default="",
                         help="Comma separated list of number of replicas for the T2V model(s)")
+    parser.add_argument("--i2v_model_paths",
+                        type=str,
+                        default="FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers",
+                        help="Comma separated list of paths to the I2V model(s)")
+    parser.add_argument("--i2v_model_replicas",
+                        type=str,
+                        default="1",
+                        help="Comma separated list of number of replicas for the I2V model(s)")
     parser.add_argument("--output_path",
                         type=str,
                         default="outputs",
@@ -361,13 +466,21 @@ if __name__ == "__main__":
     
     args = parser.parse_args()
 
-    model_paths = args.t2v_model_paths.split(",")
-    replicas = [int(r) for r in args.t2v_model_replicas.split(",")]
-    validate_configuration(model_paths, replicas)
+    # Parse and validate all models
+    t2v_paths = [p.strip() for p in args.t2v_model_paths.split(",") if p.strip()]
+    t2v_reps = [int(r.strip()) for r in args.t2v_model_replicas.split(",") if r.strip()] if args.t2v_model_replicas else []
+    i2v_paths = [p.strip() for p in args.i2v_model_paths.split(",") if p.strip()]
+    i2v_reps = [int(r.strip()) for r in args.i2v_model_replicas.split(",") if r.strip()] if args.i2v_model_replicas else []
+    
+    all_paths = t2v_paths + i2v_paths
+    all_replicas = t2v_reps + i2v_reps
+    validate_configuration(all_paths, all_replicas)
     
     start_ray_serve(
         t2v_model_paths=args.t2v_model_paths,
         t2v_model_replicas=args.t2v_model_replicas,
+        i2v_model_paths=args.i2v_model_paths,
+        i2v_model_replicas=args.i2v_model_replicas,
         output_path=args.output_path,
         host=args.host,
         port=args.port,
@@ -376,4 +489,4 @@ if __name__ == "__main__":
     setup_signal_handlers()
     print("✅ FastVideo backend is running. Press Ctrl-C to stop.")
     while True:
-        time.sleep(3600) 
+        time.sleep(3600) 

examples/inference/gradio/serving/start.sh

@@ -1,3 +1,5 @@
 python examples/inference/gradio/serving/start_ray_serve_app.py \
---t2v_model_paths "FastVideo/FastWan2.1-T2V-1.3B-Diffusers" \
---t2v_model_replicas "1"
+--t2v_model_paths "" \
+--t2v_model_replicas "" \
+--i2v_model_paths "FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers" \
+--i2v_model_replicas "1"

examples/inference/gradio/serving/start_ray_serve_app.py

@@ -20,8 +20,10 @@ DEFAULT_BACKEND_PORT = 8000
 DEFAULT_FRONTEND_HOST = "0.0.0.0"
 DEFAULT_FRONTEND_PORT = 7860
 DEFAULT_OUTPUT_PATH = "outputs"
-DEFAULT_T2V_MODELS = "FastVideo/FastWan2.1-T2V-1.3B-Diffusers,FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers"
-DEFAULT_T2V_REPLICAS = "4,4"
+DEFAULT_T2V_MODELS = ""
+DEFAULT_T2V_REPLICAS = ""
+DEFAULT_I2V_MODELS = "FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers"
+DEFAULT_I2V_REPLICAS = "1"
 
 HEALTH_CHECK_TIMEOUT = 5
 HEALTH_CHECK_MAX_RETRIES = 100
@@ -100,6 +102,12 @@ class ServiceManager:
             "port": self.args.backend_port
         }
         
+        # Add I2V parameters if provided
+        if self.args.i2v_model_paths:
+            backend_args["i2v_model_paths"] = self.args.i2v_model_paths
+        if self.args.i2v_model_replicas:
+            backend_args["i2v_model_replicas"] = self.args.i2v_model_replicas
+        
         self.backend_process = self._start_service("ray_serve_backend.py", backend_args, "backend")
         return self.backend_process
     
@@ -111,6 +119,10 @@ class ServiceManager:
             "port": self.args.frontend_port
         }
         
+        # Add I2V parameters if provided
+        if self.args.i2v_model_paths:
+            frontend_args["i2v_model_paths"] = self.args.i2v_model_paths
+        
         self.frontend_process = self._start_service("gradio_frontend.py", frontend_args, "frontend")
         return self.frontend_process
     
@@ -173,6 +185,9 @@ def print_startup_info(args: argparse.Namespace) -> None:
     print("=" * 50)
     print(f"T2V Models: {args.t2v_model_paths}")
     print(f"T2V Model Replicas: {args.t2v_model_replicas}")
+    if args.i2v_model_paths:
+        print(f"I2V Models: {args.i2v_model_paths}")
+        print(f"I2V Model Replicas: {args.i2v_model_replicas}")
     print(f"Output: {args.output_path}")
     print(f"Backend: http://{args.backend_host}:{args.backend_port}")
     print(f"Frontend: http://{args.frontend_host}:{args.frontend_port}")
@@ -190,6 +205,14 @@ def parse_arguments() -> argparse.Namespace:
                         type=str, 
                         default=DEFAULT_T2V_REPLICAS,
                         help="Comma separated list of number of replicas for the T2V model(s)")
+    parser.add_argument("--i2v_model_paths", 
+                        type=str,
+                        default=DEFAULT_I2V_MODELS,
+                        help="Comma separated list of paths to the I2V model(s)")
+    parser.add_argument("--i2v_model_replicas", 
+                        type=str, 
+                        default=DEFAULT_I2V_REPLICAS,
+                        help="Comma separated list of number of replicas for the I2V model(s)")
     parser.add_argument("--output_path", 
                         type=str, 
                         default=DEFAULT_OUTPUT_PATH,

examples/training/finetune/wan_t2v_1.3B/crush_smol/README.md

@@ -5,12 +5,12 @@ These are e2e example scripts for finetuning Wan2.1 T2V 1.3B on the crush-smol d
 
 ### Download crush-smol dataset:
 
-`bash examples/training/finetune/wan_t2v_1_3b/crush_smol/download_dataset.sh`
+`bash examples/training/finetune/wan_t2v_1.3B/crush_smol/download_dataset.sh`
 
 ### Preprocess the videos and captions into latents:
 
-`bash examples/training/finetune/wan_t2v_1_3b/crush_smol/preprocess_wan_data_t2v.sh`
+`bash examples/training/finetune/wan_t2v_1.3B/crush_smol/preprocess_wan_data_t2v.sh`
 
 ### Edit the following file and run finetuning:
 
-`bash examples/training/finetune/wan_t2v_1_3b/crush_smol/finetune_t2v.sh`
+`bash examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh`

examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh

@@ -54,7 +54,7 @@ validation_args=(
   --validation_dataset_file $VALIDATION_DATASET_FILE
   --validation_steps 200
   --validation_sampling_steps "50" 
-  --validation_guidance_scale "6.0"
+  --validation_guidance_scale "3.0"
 )
 
 # Optimizer arguments

examples/training/finetune/wan_t2v_1.3B/crush_smol/profile_finetune_t2v.sh

@@ -0,0 +1,107 @@
+#!/bin/bash
+
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export TOKENIZERS_PARALLELISM=false
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
+NUM_GPUS=4
+# export CUDA_VISIBLE_DEVICES=4,5
+
+mkdir -p ../profiler_traces/wan_t2v_finetune/
+
+# Torch Profiler Configuration
+export FASTVIDEO_TORCH_PROFILE_REGIONS="profiler_region_training_train_one_step"
+export FASTVIDEO_TORCH_PROFILER_RECORD_SHAPES=1
+export FASTVIDEO_TORCH_PROFILER_WITH_STACK=1  
+export FASTVIDEO_TORCH_PROFILER_WITH_FLOPS=1
+export FASTVIDEO_TORCH_PROFILER_WITH_PROFILE_MEMORY=1
+export FASTVIDEO_TORCH_PROFILER_WAIT_STEPS=2
+export FASTVIDEO_TORCH_PROFILER_WARMUP_STEPS=1
+export FASTVIDEO_TORCH_PROFILER_ACTIVE_STEPS=1
+export FASTVIDEO_TORCH_PROFILER_DIR="../profiler_traces/wan_t2v_finetune/"
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_t2v_finetune"
+  --output_dir "checkpoints/wan_t2v_finetune"
+  --max_train_steps 5000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 8
+  --num_latent_t 20
+  --num_height 480
+  --num_width 832
+  --num_frames 77
+  --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS 
+  --sp_size $NUM_GPUS 
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim $NUM_GPUS
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path $DATA_DIR
+  --dataloader_num_workers 1
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation 
+  --validation_dataset_file $VALIDATION_DATASET_FILE
+  --validation_steps 200
+  --validation_sampling_steps "50" 
+  --validation_guidance_scale "6.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 5e-5
+  --mixed_precision "bf16"
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
+  --weight_decay 1e-4
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.1
+  --multi_phased_distill_schedule "4000-1"
+  --not_apply_cfg_solver
+  --dit_precision "fp32"
+  --num_euler_timesteps 50
+  --ema_start_step 0
+  --enable_gradient_checkpointing_type "full"
+  # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500"
+)
+
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+  --master_port 29502 \
+    fastvideo/training/wan_training_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"

fastvideo-kernel/CMakeLists.txt

@@ -0,0 +1,159 @@
+cmake_minimum_required(VERSION 3.26 FATAL_ERROR)
+project(fastvideo-kernel LANGUAGES CXX)
+
+# Prefer environment variable (used by CI or pip install git+repo_addr) if CMake var is not explicitly set.
+if(NOT DEFINED GPU_BACKEND AND DEFINED ENV{GPU_BACKEND})
+    set(GPU_BACKEND "$ENV{GPU_BACKEND}")
+endif()
+
+if(GPU_BACKEND STREQUAL "ROCM")
+    enable_language(HIP)
+else()
+    enable_language(CUDA)
+endif()
+
+# Import common utils if needed, but we keep it simple for now
+
+# Find Python and Torch
+find_package(Python COMPONENTS Interpreter Development.Module REQUIRED)
+
+# Robustly find Torch include paths using Python
+execute_process(
+    COMMAND "${Python_EXECUTABLE}" -c "import torch; from torch.utils.cpp_extension import include_paths; print(';'.join(include_paths()))"
+    OUTPUT_VARIABLE TORCH_INCLUDE_PATHS
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+list(APPEND TORCH_INCLUDE_DIRS ${TORCH_INCLUDE_PATHS})
+
+# Find Torch package (still useful for libraries)
+find_package(Torch REQUIRED)
+
+# Include directories
+include_directories(
+    ${CMAKE_SOURCE_DIR}/include
+    ${CMAKE_SOURCE_DIR}/include/cutlass/include
+    ${CMAKE_SOURCE_DIR}/include/tk/include
+    ${CMAKE_SOURCE_DIR}/include/tk/prototype
+    ${CMAKE_SOURCE_DIR}/csrc
+    ${CMAKE_SOURCE_DIR}/csrc/turbodiffusion
+    ${TORCH_INCLUDE_DIRS}
+)
+
+# ---------------------------
+# ThunderKittens (TK) toggles
+# ---------------------------
+# AUTO: enable TK only when we can confidently target Hopper (sm_90a).
+# ON:   force-enable TK kernels (intended for release wheels/images; does NOT require a GPU).
+# OFF:  never build TK kernels.
+set(FASTVIDEO_KERNEL_BUILD_TK "AUTO" CACHE STRING "Build ThunderKittens kernels: AUTO/ON/OFF")
+set_property(CACHE FASTVIDEO_KERNEL_BUILD_TK PROPERTY STRINGS AUTO ON OFF)
+
+# Prefer environment variable (used by CI) if CMake var is not explicitly set.
+if(NOT DEFINED TORCH_CUDA_ARCH_LIST AND DEFINED ENV{TORCH_CUDA_ARCH_LIST})
+    set(TORCH_CUDA_ARCH_LIST "$ENV{TORCH_CUDA_ARCH_LIST}")
+endif()
+
+message(STATUS "TORCH_CUDA_ARCH_LIST (cmake/env): ${TORCH_CUDA_ARCH_LIST}")
+message(STATUS "FASTVIDEO_KERNEL_BUILD_TK: ${FASTVIDEO_KERNEL_BUILD_TK}")
+
+set(ENABLE_TK_KERNELS OFF)
+if(FASTVIDEO_KERNEL_BUILD_TK STREQUAL "ON")
+    set(ENABLE_TK_KERNELS ON)
+elseif(FASTVIDEO_KERNEL_BUILD_TK STREQUAL "OFF")
+    set(ENABLE_TK_KERNELS OFF)
+else()
+    # AUTO: detect Hopper if possible.
+    if(TORCH_CUDA_ARCH_LIST)
+        # Accept common spellings: 9.0a, 90a, sm_90a.
+        string(REGEX MATCH "(^|[; ,])((9\\.0a)|(90a)|(sm_90a))([; ,]|$)" _HAS_90A "${TORCH_CUDA_ARCH_LIST}")
+        if(_HAS_90A)
+            set(ENABLE_TK_KERNELS ON)
+        endif()
+    else()
+        # Best-effort local detection (works when a CUDA device is visible).
+        execute_process(
+            COMMAND "${Python_EXECUTABLE}" -c "import torch; import sys; \nprint('1' if (torch.cuda.is_available() and torch.version.cuda and torch.cuda.get_device_capability()[0] >= 9) else '0')"
+            OUTPUT_VARIABLE _LOCAL_HAS_HOPPER
+            OUTPUT_STRIP_TRAILING_WHITESPACE
+            ERROR_QUIET
+        )
+        if(_LOCAL_HAS_HOPPER STREQUAL "1")
+            set(ENABLE_TK_KERNELS ON)
+        endif()
+    endif()
+endif()
+
+if(ENABLE_TK_KERNELS)
+    message(STATUS "ThunderKittens kernels: ENABLED")
+else()
+    message(STATUS "ThunderKittens kernels: DISABLED (will use Triton fallbacks at runtime)")
+endif()
+
+# Always try to build the extension if CUDA is available, but conditionally add sources/flags
+set(BUILD_CXX_KERNELS ON)
+
+# Compiler flags
+set(CUDA_FLAGS
+    "-DNDEBUG"
+    "-O3"
+    "-std=c++20"
+    "--use_fast_math"
+    "--expt-extended-lambda"
+    "--expt-relaxed-constexpr"
+    "-Xcompiler=-fno-strict-aliasing"
+    "-Xcompiler=-fPIC"
+    "-DTORCH_COMPILE"
+    "-Xnvlink=--verbose"
+    "-Xptxas=--verbose"
+    "-Xptxas=--warn-on-spills"
+)
+
+# If TK is enabled, ensure we target Hopper. This is required even on GPU-less builders (CI).
+if(ENABLE_TK_KERNELS)
+    if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES OR CMAKE_CUDA_ARCHITECTURES STREQUAL "")
+        set(CMAKE_CUDA_ARCHITECTURES "90a" CACHE STRING "CUDA architectures" FORCE)
+    endif()
+    list(APPEND CUDA_FLAGS "-DKITTENS_HOPPER")
+    message(STATUS "CMAKE_CUDA_ARCHITECTURES: ${CMAKE_CUDA_ARCHITECTURES}")
+endif()
+
+if(BUILD_CXX_KERNELS)
+    # Source files
+    set(EXTENSION_SOURCES 
+        csrc/common_extension.cpp
+        csrc/turbodiffusion/gemm/gemm.cu
+        csrc/turbodiffusion/norm/rmsnorm.cu
+        csrc/turbodiffusion/norm/layernorm.cu
+        csrc/turbodiffusion/quant/quant.cu
+    )
+    
+    # Conditionally add TK kernels
+    if(ENABLE_TK_KERNELS)
+        list(APPEND EXTENSION_SOURCES
+            csrc/attention/st_attn_h100.cu
+            csrc/attention/block_sparse_h100.cu
+        )
+    endif()
+
+    # Combined FastVideo Extension
+    # Using name 'fastvideo_kernel_ops' to distinguish from the python package namespace
+    Python_add_library(fastvideo_kernel_ops MODULE USE_SABI ${SKBUILD_SABI_VERSION} WITH_SOABI
+        ${EXTENSION_SOURCES}
+    )
+
+    # Build compile definitions list
+    set(COMPILE_DEFS TORCH_EXTENSION_NAME=fastvideo_kernel_ops)
+    if(ENABLE_TK_KERNELS)
+        list(APPEND COMPILE_DEFS TK_COMPILE_ST_ATTN TK_COMPILE_BLOCK_SPARSE)
+    endif()
+
+    target_compile_definitions(fastvideo_kernel_ops PRIVATE ${COMPILE_DEFS})
+
+    target_compile_options(fastvideo_kernel_ops PRIVATE
+        $<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>
+    )
+
+    # We install it to fastvideo_kernel/_C so we can load it to register the ops
+    install(TARGETS fastvideo_kernel_ops LIBRARY DESTINATION fastvideo_kernel/_C)
+endif()
+

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.

fastvideo-kernel/MANIFEST.in

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

fastvideo-kernel/README.md

@@ -0,0 +1,69 @@
+# FastVideo Kernel
+
+CUDA kernels for FastVideo video generation.
+
+## Installation
+
+### Standard Installation (Local Development)
+This will automatically detect your GPU architecture. If an NVIDIA Hopper (H100/sm_90a) GPU is detected, ThunderKittens kernels will be enabled. Otherwise, they will be skipped, and the package will use Triton fallbacks at runtime.
+
+```bash
+git submodule update --init --recursive
+cd fastvideo-kernel
+./build.sh
+```
+
+### Rocm Build
+If you are in a rocm environment without the compilation toolchaine of CUDA.
+
+```bash
+cd fastvideo-kernel
+./build.sh --rocm
+```
+
+## Usage
+
+### Sliding Tile Attention (STA) & Video Sparse Attention (VSA)
+
+For detailed usage, please check the [Attention Documentation](../docs/attention/index.md).
+
+```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, ...)
+```
+
+### TurboDiffusion Kernels
+
+This package also includes kernels from [TurboDiffusion](https://github.com/thu-ml/TurboDiffusion), including INT8 GEMM, Quantization, RMSNorm and LayerNorm.
+
+## Requirements
+
+- **Runtime**:
+  - NVIDIA H100 (sm_90a) for C++ optimized kernels.
+  - Any CUDA GPU for Triton-based fallbacks.
+- **Build**:
+  - CUDA Toolkit 12.3+
+  - C++20 compatible compiler (GCC 10+, Clang 11+)
+
+## Acknowledgement
+
+This package structure and build system are based on [sgl-kernel](https://github.com/sgl-project/sglang/tree/main/sgl-kernel) from the SGLang project.
+
+The implementation of `turbodiffusion` kernels is adapted from [TurboDiffusion](https://github.com/thu-ml/TurboDiffusion). If you use these kernels, please cite:
+
+```bibtex
+@article{zhang2025turbodiffusion,
+  title={TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times},
+  author={Zhang, Jintao and Zheng, Kaiwen and Jiang, Kai and Wang, Haoxu and Stoica, Ion and Gonzalez, Joseph E and Chen, Jianfei and Zhu, Jun},
+  journal={arXiv preprint arXiv:2512.16093},
+  year={2025}
+}
+```

fastvideo-kernel/build.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+set -ex
+
+# Simple build script wrapping uv/pip
+# Usage:
+#   ./build.sh                # local dev build (auto-detect / skip TK kernels when not available)
+#   ./build.sh --release      # force-enable Hopper/TK kernels for release builds (no GPU required)
+
+echo "Building fastvideo-kernel..."
+
+# Ensure submodules are initialized if needed (tk)
+git submodule update --init --recursive
+
+# Install build dependencies
+pip install scikit-build-core cmake ninja
+
+RELEASE=0
+GPU_BACKEND=CUDA
+for arg in "$@"; do
+    case "$arg" in
+        --rocm)
+            GPU_BACKEND=ROCM
+            ;;
+    esac
+done
+
+# Force-enable ThunderKittens kernels and compile for Hopper.
+export TORCH_CUDA_ARCH_LIST="9.0a"
+export CMAKE_ARGS="${CMAKE_ARGS:-} -DFASTVIDEO_KERNEL_BUILD_TK=ON -DCMAKE_CUDA_ARCHITECTURES=90a"
+
+export CMAKE_ARGS="${CMAKE_ARGS:-} -DGPU_BACKEND=${GPU_BACKEND}"
+
+echo "TORCH_CUDA_ARCH_LIST: ${TORCH_CUDA_ARCH_LIST:-<unset>}"
+echo "CMAKE_ARGS: ${CMAKE_ARGS:-<unset>}"
+echo "GPU_BACKEND: ${GPU_BACKEND:-<unset>}"
+# Build and install
+# Use -v for verbose output
+pip install . -v --no-build-isolation

fastvideo-kernel/csrc/attention/block_sparse_h100.cu

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

fastvideo-kernel/csrc/attention/st_attn_h100.cu

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

fastvideo-kernel/csrc/common_extension.cpp

@@ -0,0 +1,48 @@
+#include <torch/extension.h>
+#include <ATen/ATen.h>
+#include <vector>
+
+// Forward declarations
+#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
+
+#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
+
+// TurboDiffusion kernels
+void register_quant(pybind11::module_ &);
+void register_rms_norm(pybind11::module_ &);
+void register_layer_norm(pybind11::module_ &);
+void register_gemm(pybind11::module_ &);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.doc() = "FastVideo CUDA Kernels";
+
+#ifdef TK_COMPILE_ST_ATTN
+    m.def("sta_fwd", torch::wrap_pybind_function(sta_forward), "sliding tile attention (Hopper)");
+#endif
+
+#ifdef TK_COMPILE_BLOCK_SPARSE
+    m.def("block_sparse_fwd", torch::wrap_pybind_function(block_sparse_attention_forward), "block sparse attention forward (Hopper)");
+    m.def("block_sparse_bwd", torch::wrap_pybind_function(block_sparse_attention_backward), "block sparse attention backward (Hopper)");
+#endif
+
+    // TurboDiffusion
+    register_quant(m);
+    register_rms_norm(m);
+    register_layer_norm(m);
+    register_gemm(m);
+}

fastvideo-kernel/csrc/turbodiffusion/common/common.hpp

@@ -0,0 +1,85 @@
+#pragma once
+
+#include <torch/torch.h>
+#include <torch/extension.h>
+
+CUTLASS_HOST_DEVICE int64_t cdiv(int64_t const& a, int64_t const &b) {
+  return (a + b - 1) / b;
+}
+
+template <class T>
+CUTLASS_HOST_DEVICE T max(T a, T b) { return a > b ? a : b; }
+
+template <class T>
+CUTLASS_HOST_DEVICE T min(T a, T b) { return a > b ? b : a; }
+
+#define MIN(a, b) ((a) > (b) ? (b) : (a))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+#define BOOL_SWITCH(COND, CONST_NAME, ...)    \
+[&] {                                         \
+  if (COND) {                                 \
+    static constexpr bool CONST_NAME = true;  \
+    return (__VA_ARGS__)();                   \
+  } else {                                    \
+    static constexpr bool CONST_NAME = false; \
+    return (__VA_ARGS__)();                   \
+  }                                           \
+}() 
+
+#define CUDA_CHECK(call)    \
+{                           \
+  cudaError_t err = call;   \
+  if (err != cudaSuccess) { \
+    fprintf(stderr, "CUDA Error at %s:%d: %s\n", __FILE__, __LINE__, cudaGetErrorString(err)); \
+    exit(err);              \
+  }                         \
+}
+
+#define CONFIG_SWITCH(N, ...)                                   \
+[&] {                                                           \
+  if (N <= 1024) {                                              \
+    constexpr int NUM_THR_PER_CTA = 128;                        \
+    constexpr int MAX_HIDDEN_SIZE = 1024;                       \
+    return (__VA_ARGS__)();                                     \
+  } else if (N <= 2048) {                                       \
+    constexpr int NUM_THR_PER_CTA = 128;                        \
+    constexpr int MAX_HIDDEN_SIZE = 2048;                       \
+    return (__VA_ARGS__)();                                     \
+  } else if (N <= 4096) {                                       \
+    constexpr int NUM_THR_PER_CTA = 128;                        \
+    constexpr int MAX_HIDDEN_SIZE = 4096;                       \
+    return (__VA_ARGS__)();                                     \
+  } else if (N <= 8192) {                                       \
+    constexpr int NUM_THR_PER_CTA = 256;                        \
+    constexpr int MAX_HIDDEN_SIZE = 8192;                       \
+    return (__VA_ARGS__)();                                     \
+  } else {                                                      \
+    constexpr int NUM_THR_PER_CTA = 256;                        \
+    constexpr int MAX_HIDDEN_SIZE = 16384;                      \
+    return (__VA_ARGS__)();                                     \
+  }                                                             \
+}()
+
+
+template <int BlockSize>
+  void create_tensor(
+    torch::Device const &device,
+    std::optional<at::Tensor> &output,
+    std::optional<at::Tensor> &scale,
+    int m, int n
+  ) {
+    int num_block_m = cdiv(m, BlockSize);
+    int num_block_n = cdiv(n, BlockSize);
+    if (!output.has_value()) {
+      output.emplace(torch::empty(
+        {m, n},
+        torch::TensorOptions().device(device).dtype(torch::kInt8)
+      ));
+      scale.emplace(torch::empty(
+        {num_block_m, num_block_n},
+        torch::TensorOptions().device(device).dtype(torch::kFloat32)
+      ));
+    }
+  }
+

fastvideo-kernel/csrc/turbodiffusion/common/launch.hpp

@@ -0,0 +1,44 @@
+#pragma once
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+
+template <class Kernel>
+__global__ void device_kernel(
+  __grid_constant__ typename Kernel::Params const params
+) {
+  extern __shared__ char smem[];
+  Kernel op;
+  op(params, smem);
+}
+
+template <class Kernel>
+__global__ __launch_bounds__(Kernel::MaxThreadsPerBlock, Kernel::MinBlocksPerMultiprocessor)
+void device_kernel_with_launch_bounds(
+  __grid_constant__ typename Kernel::Params const params
+) {
+  extern __shared__ char smem[];
+  Kernel op;
+  op(params, smem);
+}
+
+template <class Kernel>
+void launch_kernel(
+  typename Kernel::Params const &params,
+  dim3 grid_shape,
+  dim3 cta_shape,
+  size_t ShmSize,
+  cudaStream_t stream = nullptr
+) {
+  auto func = device_kernel<Kernel>;
+  if (ShmSize >= 48 * 1024) {
+    CUDA_CHECK(cudaFuncSetAttribute(
+      func,
+      cudaFuncAttributeMaxDynamicSharedMemorySize,
+      ShmSize
+    ));
+  }
+  func<<<grid_shape, cta_shape, ShmSize, stream>>>(params);
+  CUDA_CHECK(cudaGetLastError());
+}

fastvideo-kernel/csrc/turbodiffusion/common/load.hpp

@@ -0,0 +1,75 @@
+#pragma once
+
+template <
+  class InputDtype_,
+  int TileM_,
+  int TileN_,
+  int NumThrPerCta_,
+  bool IsEvenM,
+  bool IsEvenN
+>
+class Loader {
+public: 
+  using InputDtype = InputDtype_;
+  static constexpr int TileM = TileM_;
+  static constexpr int TileN = TileN_;
+  static constexpr int NumThrPerCta = NumThrPerCta_;
+  static constexpr int NumElementPerThread = TileM * TileN / NumThrPerCta;
+  static constexpr int NumThrPerRow = TileN / NumElementPerThread;
+
+  static_assert(NumThrPerCta % TileM == 0);
+  static_assert(TileM * TileN % NumThrPerCta == 0);
+
+
+  CUTLASS_DEVICE void
+  load(void const *input_ptr, void *thr_output_reg, int64_t m, int64_t n, int blk_m, int blk_n, int tid) {
+    int n_alignment = (n & 31) * sizeof(InputDtype);
+    int thr_m_offset = tid / NumThrPerRow;
+    int thr_n_offset = (tid % NumThrPerRow) * NumElementPerThread;
+    void const *cta_input_ptr = (void*)((InputDtype*)input_ptr + blk_m * TileM * n + blk_n * TileN);
+    void const *thr_input_ptr = (void*)((InputDtype*)cta_input_ptr + thr_m_offset * n + thr_n_offset);
+    InputDtype tmp_reg[NumElementPerThread];
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < NumElementPerThread; ++i)
+      tmp_reg[i] = InputDtype(0.f);
+    bool pred = IsEvenM ? true : thr_m_offset + blk_m * TileM < m;
+    int limit = IsEvenN ? NumElementPerThread : MIN(NumElementPerThread, n - (blk_n * TileN + thr_n_offset));
+    if (n_alignment % 128 == 0)
+      _load<int4, IsEvenN>(thr_input_ptr, (void*)tmp_reg, limit, pred);
+    else if (n_alignment % 64 == 0)
+      _load<int2, IsEvenN>(thr_input_ptr, (void*)tmp_reg, limit, pred);
+    else
+      _load<InputDtype, IsEvenN>(thr_input_ptr, (void*)tmp_reg, limit, pred);
+
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < NumElementPerThread; ++i)
+      *((float*)thr_output_reg + i) = static_cast<float>(reinterpret_cast<InputDtype const&>(tmp_reg[i]));
+  }
+
+private:
+  template <class LoadDataType, bool IsEven>
+  CUTLASS_DEVICE void
+  _load(void const *thr_input_ptr, void *thr_output_reg, int limit, bool pred) {
+    static constexpr int NumElementPerLoad = sizeof(LoadDataType) / sizeof(InputDtype);
+    if (pred) {
+      if constexpr (IsEven) {
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < NumElementPerThread; i += NumElementPerLoad) {
+          *(LoadDataType*)((InputDtype*)thr_output_reg + i) = *(LoadDataType*)((InputDtype*)thr_input_ptr + i);
+        }
+      } else {
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < limit; i += NumElementPerLoad) {
+          if (limit - i > NumElementPerLoad)
+            *(LoadDataType*)((InputDtype*)thr_output_reg + i) = *(LoadDataType*)((InputDtype*)thr_input_ptr + i);
+          else {
+            for (int j = 0; j < NumElementPerLoad; ++j) {
+              if (i + j < limit)
+                *((InputDtype*)thr_output_reg + i + j) = *((InputDtype*)thr_input_ptr + i + j);
+            }
+          }
+        }
+      }
+    }
+  }
+};

fastvideo-kernel/csrc/turbodiffusion/common/store.hpp

@@ -0,0 +1,77 @@
+#pragma once
+#include "common/common.hpp"
+
+
+template <
+  class OutputDtype_,
+  int TileM_,
+  int TileN_,
+  int NumThrPerCta_,
+  bool IsEvenM,
+  bool IsEvenN,
+  bool Round = true,
+  bool SaveScale = true
+>
+class Saver {
+public:
+  using OutputDtype = OutputDtype_;
+
+  static constexpr int TileM = TileM_;
+  static constexpr int TileN = TileN_;
+  static constexpr int NumThrPerCta = NumThrPerCta_;
+  static constexpr int NumElementPerThread = TileM * TileN / NumThrPerCta;
+  static constexpr int NumThrPerRow = TileN / NumElementPerThread;
+
+  static_assert(TileM * TileN % NumThrPerCta == 0);
+  static_assert(NumThrPerCta % TileM == 0);
+
+  CUTLASS_DEVICE void
+  store(void *Optr, void *OSptr, void *reg, float scale_inv, int64_t m, int64_t n, int blk_m, int blk_n, int tid) {
+    int n_alignment = (n & 31) * sizeof(OutputDtype);
+    int thr_m_offset = tid / NumThrPerRow;
+    int thr_n_offset = (tid % NumThrPerRow) * NumElementPerThread;
+    void *cta_output_ptr = (void*)((OutputDtype*)Optr + blk_m * TileM * (Round ? cdiv(n, TileN) * TileN : n) + blk_n * TileN);
+    void *thr_output_ptr = (void*)((OutputDtype*)cta_output_ptr + thr_m_offset * (Round ? cdiv(n, TileN) * TileN : n) + thr_n_offset);
+    bool pred = IsEvenM ? true : thr_m_offset + blk_m * TileM < m;
+    int limit = IsEvenN ? NumElementPerThread : MIN(NumElementPerThread, n - (blk_n * TileN + thr_n_offset));
+    if (n_alignment % 128 == 0)
+      _store<int4, IsEvenN>(thr_output_ptr, reg, limit, pred);
+    else if (n_alignment % 64 == 0)
+      _store<int2, IsEvenN>(thr_output_ptr, reg, limit, pred);
+    else
+      _store<OutputDtype, IsEvenN>(thr_output_ptr, reg, limit, pred);
+
+    if constexpr (SaveScale) {
+      if (tid == 0) {
+        *((float*)OSptr + blk_m * cdiv(n, TileN)+ blk_n) = scale_inv;
+      }
+    }
+  }
+
+private:
+  template <class StoreDataType, bool IsEven>
+  CUTLASS_DEVICE void
+  _store(void *thr_output_ptr, void *reg, int limit, bool pred) {
+    static constexpr int NumElementPerStore = sizeof(StoreDataType) / sizeof(OutputDtype);
+    if (pred) {
+      if constexpr (IsEven) {
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < NumElementPerThread; i += NumElementPerStore) {
+          *(StoreDataType*)((OutputDtype*)thr_output_ptr + i) = *(StoreDataType*)((OutputDtype*)reg + i);
+        }
+      } else {
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < limit; i += NumElementPerStore) {
+          if (limit - i > NumElementPerStore)
+            *(StoreDataType*)((OutputDtype*)thr_output_ptr + i) = *(StoreDataType*)((OutputDtype*)reg + i);
+          else {
+            for (int j = 0; j < limit - i; ++j) {
+              *((OutputDtype*)thr_output_ptr + i + j) = *((OutputDtype*)reg + i + j);
+            }
+          }
+        }
+      }
+    }
+  }
+
+};

fastvideo-kernel/csrc/turbodiffusion/gemm/gemm.cu

@@ -0,0 +1,77 @@
+/* 
+ * Copyright (c) 2025 by TurboDiffusion team.
+ * 
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * 
+ * Citation (please cite if you use this code):
+ * 
+ * @article{zhang2025turbodiffusion,
+ *   title={TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times},
+ *   author={Zhang, Jintao and Zheng, Kaiwen and Jiang, Kai and Wang, Haoxu and Stoica, Ion and Gonzalez, Joseph E and Chen, Jianfei and Zhu, Jun},
+ *   journal={arXiv preprint arXiv:2512.16093},
+ *   year={2025}
+ * }
+ */
+
+#include <pybind11/pybind11.h>
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <torch/all.h>
+#include <torch/python.h>
+#include <cutlass/cutlass.h>
+#include <cutlass/numeric_types.h>
+
+#include "common/common.hpp"
+#include "gemm/launch.hpp"
+
+void int8_gemm(
+  at::Tensor const& A, at::Tensor const& A_S, 
+  at::Tensor const& B, at::Tensor const& B_S, 
+  torch::Tensor& C
+) {
+
+  
+  static constexpr int swizzle_dir = 1;
+  static constexpr int swizzle_size_log = 5;
+
+  int k = B.size(1);
+  int m = A.size(0);
+  int n = B.size(0);
+
+  switch (C.scalar_type()) {
+    case torch::kHalf:{
+        int8_gemm_<cutlass::half_t> (
+            (int8_t*)A.data_ptr(), A_S.data_ptr<float>(),
+            (int8_t*)B.data_ptr(), B_S.data_ptr<float>(), 
+            (cutlass::half_t*)C.data_ptr(), 
+            m, n, k, swizzle_dir, swizzle_size_log, at::cuda::getCurrentCUDAStream().stream()
+        );
+        break;
+    }
+
+    case torch::kBFloat16:{
+        int8_gemm_<cutlass::bfloat16_t> (
+                (int8_t*)A.data_ptr(), A_S.data_ptr<float>(),
+                (int8_t*)B.data_ptr(), B_S.data_ptr<float>(), 
+                (cutlass::bfloat16_t*)C.data_ptr(), 
+                m, n, k, swizzle_dir, swizzle_size_log, at::cuda::getCurrentCUDAStream().stream()
+            );
+        break;
+    }
+
+    default: {
+      std::cerr << "Observing: " << C.scalar_type() << " for the output datatype which is invalid";
+      throw std::runtime_error("Unsupported output data type for int8 gemm.");
+    }
+  }
+
+}
+
+void register_gemm(pybind11::module_ &m) {
+    m.def("gemm_cuda", &int8_gemm);
+}
+
+
+  
+
+

fastvideo-kernel/csrc/turbodiffusion/gemm/kernel.hpp

@@ -0,0 +1,522 @@
+/* 
+ * Copyright (c) 2025 by TurboDiffusion team.
+ * 
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * 
+ * Citation (please cite if you use this code):
+ * 
+ * @article{zhang2025turbodiffusion,
+ *   title={TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times},
+ *   author={Zhang, Jintao and Zheng, Kaiwen and Jiang, Kai and Wang, Haoxu and Stoica, Ion and Gonzalez, Joseph E and Chen, Jianfei and Zhu, Jun},
+ *   journal={arXiv preprint arXiv:2512.16093},
+ *   year={2025}
+ * }
+ */
+
+#pragma once
+
+#include <cuda.h>
+#include "cute/tensor.hpp"
+
+#include "common/common.hpp"
+#include "gemm/utils.hpp"
+
+using namespace cute;
+
+template <
+  class OutputDtype_,
+  bool IsEvenM,
+  bool IsEvenN
+>
+struct GemmKernel {
+  using ElementA = int8_t;
+  using ElementB = int8_t;
+  using OutputDtype = OutputDtype_;
+  using AccumulatorDtype = int32_t;
+  static constexpr int BlockSize = 128;
+  static constexpr int TileM = 128;
+  static constexpr int TileN = 128;
+  static constexpr int TileK = 128;
+  static constexpr int Stage = 3;
+  static constexpr int EpiStage = 2;
+  
+  static_assert(
+       BlockSize % TileM == 0
+    && BlockSize % TileN == 0
+    && BlockSize % TileK == 0
+  );
+
+  static constexpr int NumTilePerBlock = BlockSize / TileK;
+
+  using SmemLayoutAtom = decltype(
+    composition(
+      Swizzle<3, 4, 3>{},
+      make_layout(
+        make_shape(Int<8>{}, Int<TileK>{}),
+        make_stride(Int<TileK>{}, Int<1>{})
+      )
+    )
+  );
+
+  using SmemLayoutA = decltype(
+    tile_to_shape(
+      SmemLayoutAtom{},
+      make_shape(Int<TileM>{}, Int<TileK>{}, Int<Stage>{})
+    )
+  );
+
+  using SmemLayoutB = decltype(
+    tile_to_shape(
+      SmemLayoutAtom{},
+      make_shape(Int<TileN>{}, Int<TileK>{}, Int<Stage>{})
+    )
+  );
+
+  using MmaOP = cute::SM80_16x8x32_S32S8S8S32_TN;
+  using TiledMma = decltype(
+    make_tiled_mma(
+      MMA_Atom<MMA_Traits<MmaOP>>{},
+      make_layout(make_shape(
+        _4{}, _2{}, _1{}
+      )),
+      make_tile(Int<64>{}, Int<32>{}, Int<32>{})
+    )
+  );
+
+  using G2SCopyAtomA = Copy_Atom<Copy_Traits<SM80_CP_ASYNC_CACHEGLOBAL<cute::uint128_t>>, ElementA>;
+  using G2SCopyAtomB = Copy_Atom<Copy_Traits<SM80_CP_ASYNC_CACHEGLOBAL<cute::uint128_t>>, ElementB>;
+  using G2STiledCopyA = decltype(
+    make_tiled_copy(
+      G2SCopyAtomA{},
+      make_layout(
+        make_shape(Int<64>{}, Int<4>{}),
+        make_stride(Int<4>{}, Int<1>{})
+      ),
+      make_layout(make_shape(Int<1>{}, Int<16>{}))
+    )
+  );
+  using G2STiledCopyB = decltype(
+    make_tiled_copy(
+      G2SCopyAtomB{},
+      make_layout(
+        make_shape(Int<64>{}, Int<4>{}),
+        make_stride(Int<4>{}, Int<1>{})
+      ),
+      make_layout(make_shape(Int<1>{}, Int<16>{}))
+    )
+  );
+
+  using S2RCopyAtomA = Copy_Atom<Copy_Traits<SM75_U32x4_LDSM_N>, ElementA>;
+  using S2RCopyAtomB = Copy_Atom<Copy_Traits<SM75_U32x4_LDSM_N>, ElementB>;
+  using S2RTiledCopyA = decltype(make_tiled_copy_A(S2RCopyAtomA{}, TiledMma{}));
+  using S2RTiledCopyB = decltype(make_tiled_copy_B(S2RCopyAtomB{}, TiledMma{}));
+
+  // epilogue
+  using SmemLayoutAtomD = decltype(
+    composition(
+      Swizzle<2, 3, 3>{},
+      make_layout(
+        make_shape(Int<32>{}, Int<32>{}),
+        LayoutRight{}
+      )
+    )
+  );
+
+  using SmemLayoutD = decltype(
+    tile_to_shape(
+      SmemLayoutAtomD{},
+      make_shape(Int<64>{}, Int<32>{}, Int<EpiStage>{})
+    )
+  );
+
+  using R2SCopyAtomD = Copy_Atom<UniversalCopy<std::conditional_t<sizeof(OutputDtype) == 4, int32_t, int16_t>>, OutputDtype>;
+  using R2STiledCopyD = decltype(make_tiled_copy_C(R2SCopyAtomD{}, TiledMma{}));
+
+  using S2GCopyAtomD = Copy_Atom<UniversalCopy<uint128_t>, OutputDtype>;
+  using S2GCopyD = decltype(make_tiled_copy(
+    S2GCopyAtomD{},
+    make_layout(Shape<_64, _4>{}),
+    make_layout(Shape<_1, _8>{})
+  ));
+
+  using TileShape = decltype(make_shape(Int<TileM>{}, Int<TileN>{}, Int<TileK>{}));
+
+  struct SharedStorageAB: cute::aligned_struct<128> {
+    array_aligned<typename TiledMma::ValTypeA, cosize_v<SmemLayoutA>, 128> smem_A;
+    array_aligned<typename TiledMma::ValTypeB, cosize_v<SmemLayoutB>, 128> smem_B;
+    array_aligned<float, 1> smem_AS;
+    array_aligned<float, 1> smem_BS;
+    array_aligned<int32_t, 1> smem_AF;
+  };
+
+  struct SharedStorageD: cute::aligned_struct<128> {
+    array_aligned<OutputDtype, cosize_v<SmemLayoutD>> smem_D;
+  };
+
+  union SharedStorage {
+    SharedStorageAB storage_AB;
+    SharedStorageD storage_D;
+  };
+
+
+  struct Params {
+    void const* Aptr;
+    void const* ASptr;
+    void const* Bptr;
+    void const* BSptr;
+    void* Dptr;
+    int64_t const m;
+    int64_t const n;
+    int64_t const k;
+    int const swizzle_dir;
+    int const swizzle_size;
+  };
+
+  using Arguments = Params;
+
+  static constexpr int ThreadNum = size(TiledMma{});
+  static constexpr int ShmSize = sizeof(SharedStorage);
+  static constexpr bool FastInt2Float = false;
+
+  static bool can_implement(int64_t m, int64_t n, int64_t k) {
+    if (k % BlockSize != 0) return false;
+    if ((n * sizeof(OutputDtype)) % 16 != 0)
+      return false;
+    return true;
+  }
+
+  static Params to_underlying_arguments(Arguments const& args) {
+    return args;
+  }
+
+  static dim3 get_grid_size(int64_t m, int64_t n) {
+    return dim3(cdiv(m, TileM) * cdiv(n, TileN));
+  }
+
+  CUTLASS_HOST_DEVICE 
+  static auto get_block_coord(
+    int64_t m_blocks, 
+    int64_t n_blocks, 
+    int const swizzle_dir,
+    int64_t const swizzle_size_log
+  ) {
+    int64_t blk_m;
+    int64_t blk_n;
+
+    if (swizzle_dir == 1)
+      std::swap(m_blocks, n_blocks);
+
+    if (swizzle_size_log == 0) {
+      blk_m = blockIdx.x % m_blocks;
+      blk_n = blockIdx.x / m_blocks;
+    } else {
+      int64_t group_size = n_blocks << swizzle_size_log;
+      int64_t num_groups = m_blocks >> swizzle_size_log;
+      int64_t group_idx = blockIdx.x / group_size;
+      int64_t local_idx = blockIdx.x % group_size;
+      if (group_idx == num_groups) {
+        blk_m = (num_groups << swizzle_size_log) + local_idx % (m_blocks - (num_groups << swizzle_size_log));
+        blk_n = local_idx / (m_blocks - (num_groups << swizzle_size_log));
+      } else {
+        blk_m = (local_idx & ((1LL << swizzle_size_log) - 1)) + (group_idx << swizzle_size_log);
+        blk_n = local_idx >> swizzle_size_log;
+      }
+    }
+
+    if (swizzle_dir == 1)
+      std::swap(blk_m, blk_n);
+
+    return make_coord(blk_m, blk_n);
+  }
+
+  CUTLASS_DEVICE
+  void operator()(
+    Params const& params, char* smem_data
+  ) {
+
+    SharedStorage& shared_storage = *reinterpret_cast<SharedStorage*>(smem_data);
+
+    auto t_idx = threadIdx.x;
+
+    int64_t const m = params.m;
+    int64_t const n = params.n;
+    int64_t const k = params.k;
+    int const swizzle_dir = params.swizzle_dir;
+    int const swizzle_size = params.swizzle_size;
+
+    Tensor A = make_tensor(
+      make_gmem_ptr<ElementA>(params.Aptr),
+      make_shape(m, k),
+      make_stride(k, _1{})
+    );
+    Tensor B = make_tensor(
+      make_gmem_ptr<ElementB>(params.Bptr),
+      make_shape(m, k),
+      make_stride(k, _1{})
+    );
+    Tensor AS = make_tensor(
+      make_gmem_ptr<float>(params.ASptr),
+      make_shape(cdiv(m, BlockSize), cdiv(k, BlockSize)),
+      make_stride(cdiv(k, BlockSize), _1{})
+    );
+    Tensor BS = make_tensor(
+      make_gmem_ptr<float>(params.BSptr),
+      make_shape(cdiv(n, BlockSize), cdiv(k, BlockSize)),
+      make_stride(cdiv(k, BlockSize), _1{})
+    );
+    Tensor D = make_tensor(
+      make_gmem_ptr<OutputDtype>(params.Dptr),
+      make_shape(m, n),
+      LayoutRight{}
+    );
+
+    auto [m_coord, n_coord] = get_block_coord(
+      cdiv(m, size<0>(TileShape{})),
+      cdiv(n, size<1>(TileShape{})),
+      swizzle_dir, swizzle_size
+    );
+
+    int32_t blk_m_coord = m_coord / (BlockSize / TileM);
+    int32_t blk_n_coord = n_coord / (BlockSize / TileN);
+
+    // local tile
+    auto gA = local_tile(A, TileShape{}, make_coord(m_coord, n_coord, _), Step<_1, X, _1>{});
+    auto gB = local_tile(B, TileShape{}, make_coord(m_coord, n_coord, _), Step<X, _1, _1>{});
+    auto gD = local_tile(D, TileShape{}, make_coord(m_coord, n_coord, _), Step<_1, _1, X>{});
+
+    // shared memory
+    Tensor sA = make_tensor(
+      make_smem_ptr<ElementA>(shared_storage.storage_AB.smem_A.data()),
+      SmemLayoutA{}
+    );
+
+    Tensor sB = make_tensor(
+      make_smem_ptr<ElementB>(shared_storage.storage_AB.smem_B.data()),
+      SmemLayoutB{}
+    );
+
+    // register
+    TiledMma tiled_mma;
+    auto thr_mma = tiled_mma.get_slice(t_idx);
+    auto tCrA = thr_mma.partition_fragment_A(gA(_, _, 0));
+    auto tCrB = thr_mma.partition_fragment_B(gB(_, _, 0));
+    auto tDrC = thr_mma.partition_fragment_C(gD); // mma accumulator
+    auto tDrD = make_tensor_like<float>(tDrC); // float accumulator
+
+    if constexpr (FastInt2Float) {
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < size(tDrC); ++i)
+        tDrC(i) = 0x4B400000;
+    } else {
+      clear(tDrC);
+    }
+
+    clear(tDrD);
+
+
+    // global to shared copy
+    G2STiledCopyA g2s_tiled_copy_a;
+    auto g2s_thr_copy_a = g2s_tiled_copy_a.get_slice(t_idx);
+    auto tAgA = g2s_thr_copy_a.partition_S(gA);
+    auto tAsA = g2s_thr_copy_a.partition_D(sA);
+    auto cA = make_identity_tensor(make_shape(size<0>(sA), size<1>(sA)));
+    auto tAcA = g2s_thr_copy_a.partition_S(cA);
+    int const m_limit = m - TileM * m_coord;
+    int const n_limit = n - TileN * n_coord;
+
+    G2STiledCopyB g2s_tiled_copy_b;
+    auto g2s_thr_copy_b = g2s_tiled_copy_b.get_slice(t_idx);
+    auto tBgB = g2s_thr_copy_b.partition_S(gB);
+    auto tBsB = g2s_thr_copy_b.partition_D(sB);
+    auto cB = make_identity_tensor(make_shape(size<0>(sB), size<1>(sB)));
+    auto tBcB = g2s_thr_copy_a.partition_S(cB);
+
+
+    // shared to register copy
+    S2RTiledCopyA s2r_tiled_copy_a;
+    auto s2r_thr_copy_a = s2r_tiled_copy_a.get_slice(t_idx);
+    auto tCsA = s2r_thr_copy_a.partition_S(sA);
+    auto tCrA_view = s2r_thr_copy_a.retile_D(tCrA);
+
+    S2RTiledCopyB s2r_tiled_copy_b;
+    auto s2r_thr_copy_b = s2r_tiled_copy_b.get_slice(t_idx);
+    auto tCsB = s2r_thr_copy_b.partition_S(sB);
+    auto tCrB_view = s2r_thr_copy_b.retile_D(tCrB);
+
+    // pipeline status
+    int64_t g2s_a_tile = 0;
+    int64_t g2s_b_tile = 0;
+    int g2s_a_smem = 0;
+    int g2s_b_smem = 0;
+
+    int g2s_tile_in_block = 0;
+    int g2s_block = 0; // b block idx
+
+    int s2r_a_smem = 0;
+    int s2r_b_smem = 0;
+    int s2r_tile_in_block = 0;
+
+    int mma_block_a = 0;
+    int mma_block_b = 0;
+
+    int ntile = k / TileK;
+    // load scale and fallback
+    // we assume all ptrs are 128bit aligned
+    // auto smem_fallback_A = raw_pointer_cast(make_smem_ptr<int32_t>(shared_storage.storage_AB.smem_AF.data()));
+    // auto smem_scale_A = raw_pointer_cast(make_smem_ptr<float>(shared_storage.storage_AB.smem_AS.data()));
+    // auto smem_scale_B = raw_pointer_cast(make_smem_ptr<float>(shared_storage.storage_AB.smem_BS.data()));
+    __syncthreads();
+
+
+    int32_t fallbackA_load = 0;
+    int32_t fallbackA_mma = 0;
+
+    // copy first Stage - 1 tile
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0, _i = min(Stage - 1, ntile); i < _i; ++i) {
+      if (g2s_b_tile < ntile) {
+        g2s_tile_in_block = (g2s_tile_in_block + 1) % NumTilePerBlock;
+        copy_AB<IsEvenM>(g2s_tiled_copy_a, tAgA, tAsA, tAcA, g2s_a_tile, g2s_a_smem, m_limit);
+        copy_AB<IsEvenN>(g2s_tiled_copy_b, tBgB, tBsB, tBcB, g2s_b_tile, g2s_b_smem, n_limit);
+        ++g2s_b_tile;
+        ++g2s_b_smem;
+        ++g2s_block;
+        g2s_a_tile = g2s_block * NumTilePerBlock;
+        ++g2s_a_smem;
+      }
+      cp_async_fence();
+    }
+
+    constexpr int nk = size<2>(tCrA);
+    float scale_a = AS(blk_m_coord, 0);
+    float scale_b = BS(blk_n_coord, 0);
+
+    CUTLASS_PRAGMA_NO_UNROLL
+    for (int64_t mma_b_tile = 0; mma_b_tile < ntile; ++mma_b_tile) {
+      s2r_tile_in_block = (s2r_tile_in_block + 1) % NumTilePerBlock;
+      cp_async_wait<Stage - 2>();
+      __syncthreads();
+
+      // do mma first
+      CUTLASS_PRAGMA_UNROLL
+      for (int ik = 0; ik < nk; ++ik) {
+        cute::copy(s2r_tiled_copy_a, tCsA(_, _, ik, s2r_a_smem),
+                    tCrA_view(_, _, ik));
+        cute::copy(s2r_tiled_copy_b, tCsB(_, _, ik, s2r_b_smem),
+                    tCrB_view(_, _, ik));
+        cute::gemm(tiled_mma, tDrC, tCrA(_, _, ik), tCrB(_, _, ik), tDrC);
+      }
+
+      // a s2r increase anyway
+      s2r_a_smem = (s2r_a_smem + 1) % Stage;
+
+      // get next s2r b tile int64_t
+        // end of a block
+
+        // dequant first
+      dequant<AccumulatorDtype, TileM * TileN / ThreadNum, FastInt2Float>(
+        tDrC.data(), tDrD.data(), scale_a * scale_b
+      );
+
+      s2r_b_smem = (s2r_b_smem + 1) % Stage;
+      // b advance
+      ++mma_block_b;
+      if (mma_block_b < size<1>(BS)) scale_b = BS(blk_n_coord, mma_block_b);
+      mma_block_a = mma_block_b;
+      if (mma_block_a < size<1>(AS)) scale_a = AS(blk_m_coord, mma_block_a);
+
+      // load next stage
+      if (g2s_b_tile < ntile) {
+        g2s_tile_in_block = (g2s_tile_in_block + 1) % NumTilePerBlock;
+        copy_AB<IsEvenM>(g2s_tiled_copy_a, tAgA, tAsA, tAcA, g2s_a_tile, g2s_a_smem, m_limit);
+        copy_AB<IsEvenN>(g2s_tiled_copy_b, tBgB, tBsB, tBcB, g2s_b_tile, g2s_b_smem, n_limit);
+        ++g2s_b_tile;
+        g2s_b_smem = (g2s_b_smem + 1) % Stage;
+        ++g2s_block;
+        g2s_a_tile = g2s_block * NumTilePerBlock;
+        g2s_a_smem = (g2s_a_smem + 1) % Stage;
+      }
+      cp_async_fence();
+    }
+
+
+    // epilogue
+
+    Tensor sD = make_tensor(
+      make_smem_ptr<OutputDtype>(shared_storage.storage_D.smem_D.data()),
+      SmemLayoutD{}
+    );
+
+    R2STiledCopyD r2s_tiled_copy_d;
+    auto r2s_thr_copy_d = r2s_tiled_copy_d.get_slice(t_idx);
+    auto tDrD_r2s = r2s_thr_copy_d.retile_S(tDrD);
+    auto tDsD_r2s = r2s_thr_copy_d.partition_D(sD);
+
+    S2GCopyD s2g_tiled_copy_d;
+    auto s2g_thr_copy_d = s2g_tiled_copy_d.get_slice(t_idx);
+    auto tDsD_s2g = s2g_thr_copy_d.partition_S(sD);
+    auto tDgD_s2g = s2g_thr_copy_d.partition_D(gD);
+    Tensor cD = make_identity_tensor(make_shape(Int<TileM>{}, Int<TileN>{}));
+    auto tDcD_s2g = s2g_thr_copy_d.partition_D(cD);
+
+    auto tDgD_s2gx = group_modes<1, 3>(tDgD_s2g);  // (CPY_, CPY_MN)
+    auto tDrD_r2sx = group_modes<1, 3>(tDrD_r2s);  // (CPY_, CPY_MN)
+    auto tDcD_s2gx = group_modes<1, 3>(tDcD_s2g);
+    
+    int32_t step = size<3>(tDsD_r2s);  // pipe
+    CUTLASS_PRAGMA_UNROLL
+    for (int32_t i = 0; i < size<1>(tDrD_r2sx); i += step) {
+      CUTLASS_PRAGMA_UNROLL
+      for (int32_t j = 0; j < step; ++j) {
+        if constexpr (std::is_same<OutputDtype, float>::value) {
+            cute::copy(r2s_tiled_copy_d, tDrD_r2sx(_, i + j), tDsD_r2s(_, 0, 0, j));
+        } else {
+            auto t = make_tensor_like<OutputDtype>(tDrD_r2sx(_, i + j));
+            cute::copy(tDrD_r2sx(_, i + j), t);
+            cute::copy(r2s_tiled_copy_d, t, tDsD_r2s(_, 0, 0, j));
+        }
+      }
+
+      __syncthreads();
+
+      // shm -> global
+      if constexpr (IsEvenM && IsEvenN) {
+        CUTLASS_PRAGMA_UNROLL
+        for (int32_t j = 0; j < step; ++j)
+          cute::copy(s2g_tiled_copy_d, tDsD_s2g(_, 0, 0, j), tDgD_s2gx(_, i + j));
+      } else if constexpr (IsEvenN) {
+          CUTLASS_PRAGMA_UNROLL
+          for (int32_t j = 0; j < step; ++j) {
+            if (get<0>(tDcD_s2gx(0, i + j)) < m_limit)
+              cute::copy(s2g_tiled_copy_d, tDsD_s2g(_, 0, 0, j), tDgD_s2gx(_, i + j));
+          }
+      } else if constexpr (IsEvenM) {
+          CUTLASS_PRAGMA_UNROLL
+          for (int32_t j = 0; j < step; ++j)
+            if (get<1>(tDcD_s2gx(size<0>(tDsD_s2g) - 1, i + j)) < n_limit) {
+              cute::copy(s2g_tiled_copy_d, tDsD_s2g(_, 0, 0, j), tDgD_s2gx(_, i + j));
+            } else {
+              CUTLASS_PRAGMA_UNROLL
+              for (int k = 0; k < size<0>(tDsD_s2g); ++k)
+                if (get<1>(tDcD_s2gx(k, i + j)) < n_limit)
+                  tDgD_s2gx(k, i + j) = tDsD_s2g(k, 0, 0, j);
+            }
+      } else {
+          CUTLASS_PRAGMA_UNROLL
+          for (int32_t j = 0; j < step; ++j)
+            if (get<0>(tDcD_s2gx(0, i + j)) < m_limit) {
+              if (get<1>(tDcD_s2gx(size<0>(tDsD_s2g) - 1, i + j)) < n_limit) {
+                cute::copy(s2g_tiled_copy_d, tDsD_s2g(_, 0, 0, j), tDgD_s2gx(_, i + j));
+              } else {
+                for (int32_t k = 0; k < size<0>(tDsD_s2g); ++k)
+                  if (get<1>(tDcD_s2gx(k, i + j)) < n_limit)
+                    tDgD_s2gx(k, i + j) = tDsD_s2g(k, 0, 0, j);
+              }
+            }
+      }
+      __syncthreads();
+    }
+  }
+
+};
+  

fastvideo-kernel/csrc/turbodiffusion/gemm/launch.hpp

@@ -0,0 +1,64 @@
+/* 
+ * Copyright (c) 2025 by TurboDiffusion team.
+ * 
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * 
+ * Citation (please cite if you use this code):
+ * 
+ * @article{zhang2025turbodiffusion,
+ *   title={TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times},
+ *   author={Zhang, Jintao and Zheng, Kaiwen and Jiang, Kai and Wang, Haoxu and Stoica, Ion and Gonzalez, Joseph E and Chen, Jianfei and Zhu, Jun},
+ *   journal={arXiv preprint arXiv:2512.16093},
+ *   year={2025}
+ * }
+ */
+
+#pragma once
+
+#include "common/common.hpp"
+#include "common/launch.hpp"
+#include "gemm/kernel.hpp"
+
+
+template <class OutputDtype>
+bool int8_gemm_(
+  int8_t const *Aptr, float const *ASptr,
+  int8_t const *Bptr, float const *BSptr,
+  OutputDtype* Dptr, int64_t m, int64_t n, int64_t k,
+  int swizzle_dir = 1, int swizzle_size_log = 0,
+  cudaStream_t stream = nullptr
+) {
+  BOOL_SWITCH(m % 128 == 0, IsEvenM, [&] {
+    BOOL_SWITCH(n % 128 == 0, IsEvenN, [&] {
+      using Kernel = GemmKernel<OutputDtype, IsEvenM, IsEvenN>;
+      if (!Kernel::can_implement(m, n, k))
+        return false;
+      using Args = typename Kernel::Arguments;
+      Args args {
+        (void*)Aptr, (void*)ASptr,
+        (void*)Bptr, (void*)BSptr, (void*)Dptr,
+        m, n, k, swizzle_dir,
+        swizzle_size_log
+      };
+
+      auto params = Kernel::to_underlying_arguments(args);
+
+      static constexpr size_t ShmSize = Kernel::ShmSize;
+      dim3 grid_shape = Kernel::get_grid_size(m, n);
+      dim3 block_shape = dim3(Kernel::ThreadNum);
+      auto func = device_kernel<Kernel>;
+      if (ShmSize >= 48 * 1024) {
+        cudaFuncSetAttribute(
+          func,
+          cudaFuncAttributeMaxDynamicSharedMemorySize,
+          ShmSize
+        );
+      }
+      func<<<grid_shape, block_shape, ShmSize, stream>>>(
+        params
+      );
+      return true;
+    });
+  });
+  return true;
+}

fastvideo-kernel/csrc/turbodiffusion/gemm/utils.hpp

@@ -0,0 +1,129 @@
+/* 
+ * Copyright (c) 2025 by TurboDiffusion team.
+ * 
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * 
+ * Citation (please cite if you use this code):
+ * 
+ * @article{zhang2025turbodiffusion,
+ *   title={TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times},
+ *   author={Zhang, Jintao and Zheng, Kaiwen and Jiang, Kai and Wang, Haoxu and Stoica, Ion and Gonzalez, Joseph E and Chen, Jianfei and Zhu, Jun},
+ *   journal={arXiv preprint arXiv:2512.16093},
+ *   year={2025}
+ * }
+ */
+
+#pragma once
+
+#include <cuda.h>
+#include "cute/tensor.hpp"
+
+template <
+  bool IsEven, 
+  class TiledCopy,
+  class SrcTensor,
+  class DstTensor,
+  class PrdTensor
+>
+CUTLASS_DEVICE void
+copy_AB(
+    TiledCopy const& _copy, 
+    SrcTensor const &S, 
+    DstTensor &D, 
+    PrdTensor const &ID, 
+    const int64_t &i_read, 
+    const int64_t &i_write, 
+    const int64_t &limit
+) {
+  using namespace cute;
+  if constexpr (IsEven)
+    cute::copy(_copy, S(_, _, _, i_read), D(_, _, _, i_write));
+  else {
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < size<1>(ID); ++i)
+      if (get<0>(ID(0, i, 0)) < limit)
+        cute::copy(_copy, S(_, i, _, i_read), D(_, i, _, i_write));
+  }
+}
+
+template <int N>
+CUTLASS_DEVICE void copy_async(
+  void const* gmem_src,
+  void* smem_dst
+) {
+  uint32_t smem_int_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(smem_dst));;
+  asm volatile("cp.async.ca.shared.global.L2::128B [%0], [%1], %2;\n"
+      :: "r"(smem_int_ptr),
+          "l"(gmem_src),
+          "n"(N));
+}
+
+template<class LoadType, class T, int NumThreads>
+CUTLASS_DEVICE void copy_aligned(const void* src, void* dst, size_t N, int64_t thread_idx) {
+  static constexpr int NumElementPerLoad = sizeof(LoadType) / sizeof(T);
+  for (int64_t i = thread_idx * NumElementPerLoad; i < N; i += NumElementPerLoad * NumThreads) {
+    if (i + NumElementPerLoad <= N) {
+      copy_async<sizeof(LoadType)>(
+        (void*)((T*)src + i),
+        (void*)((T*)dst + i)
+      );
+    } else {
+      for (int64_t j = 0; j < N - i; ++j)
+        copy_async<sizeof(T)>(
+          (void*)((T*)src + i + j),
+          (void*)((T*)dst + i + j)
+        );
+    }
+  }
+}
+
+template<class T, int NumThreads, bool Wait = true, bool Commit = true>
+CUTLASS_DEVICE void g2s_vector_copy(const void* src, void* dst, size_t N, int64_t thread_idx) {
+  
+  uintptr_t src_addr = reinterpret_cast<uintptr_t>(src);
+
+  if (src_addr % 16 == 0) {
+      copy_aligned<int4, T, NumThreads>(src, dst, N, thread_idx);
+  } else if (src_addr % 8 == 0) {
+      copy_aligned<int2, T, NumThreads>(src, dst, N, thread_idx);
+  } else if (src_addr % 4 == 0) {
+      copy_aligned<int, T, NumThreads>(src, dst, N, thread_idx);
+  } else {
+    assert(0);
+  }
+  if constexpr (Commit) {
+    asm volatile("cp.async.commit_group;\n" ::);
+  }
+  if constexpr (Wait) {
+    asm volatile("cp.async.wait_all;\n" ::);
+  }
+}
+
+template<class T, int N, bool FastInt2Float>
+CUTLASS_DEVICE
+static void dequant(
+  T* mma_accum_ptr,
+  float* float_accum_ptr,
+  float scale
+) {
+  static int const ic = 0x4B400000;
+  if constexpr (FastInt2Float && std::is_same_v<T, int32_t>) {
+    CUTLASS_PRAGMA_UNROLL
+    for (size_t i = 0; i < N; ++i) {
+      *(float_accum_ptr + i) += (__int_as_float(*(mma_accum_ptr + i)) - __int_as_float(ic)) * scale;
+      *(mma_accum_ptr + i) = ic;
+    }
+  } else if constexpr (std::is_same_v<T, int32_t>) {
+    CUTLASS_PRAGMA_UNROLL
+    for (size_t i = 0; i < N; ++i) {
+      *(float_accum_ptr + i) += __int2float_rn(*(mma_accum_ptr + i)) * scale;
+      *(mma_accum_ptr + i) = 0;
+    }
+  } else {
+    CUTLASS_PRAGMA_UNROLL
+    for (size_t i = 0; i < N; ++i) {
+      *(float_accum_ptr + i) += (*(mma_accum_ptr + i)) * scale;
+      *(mma_accum_ptr + i) = 0;
+    }
+  }
+}