fix cm

fix preprocessing before vae deocde
init lpips
2025-08-21 05:18:30 +00:00 · 2025-08-20 01:10:53 +00:00 · 2025-08-19 23:48:06 +00:00 · 2025-08-19 16:40:42 -07:00 · 2025-08-19 15:40:29 -07:00 · 2025-08-19 07:19:40 +00:00
259 changed files with 3540 additions and 20881 deletions
@@ -117,11 +117,11 @@ 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"
+                - "csrc/attn/vsa/**"
+                - "csrc/attn/tk/**"
+                - "csrc/attn/setup_vsa.py"
+                - "csrc/attn/config_vsa.py"
+                - "csrc/attn/vsa.cpp"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -133,10 +133,10 @@ 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"
+                - "csrc/attn/st_attn/**"
+                - "csrc/attn/setup_sta.py"
+                - "csrc/attn/config_sta.py"
+                - "csrc/attn/st_attn.cpp"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -147,10 +147,10 @@ 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"
+                - "csrc/attn/st_attn/**"
+                - "csrc/attn/setup_sta.py"
+                - "csrc/attn/config_sta.py"
+                - "csrc/attn/st_attn.cpp"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -161,12 +161,12 @@ steps:
                agents:
                  queue: "default"
            - path:
-                - "csrc/attn/video_sparse_attn/**"
-                - "csrc/attn/video_sparse_attn/tk/**"
+                - "csrc/attn/vsa/**"
+                - "csrc/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"
+                - "csrc/attn/setup_vsa.py"
+                - "csrc/attn/config_vsa.py"
+                - "csrc/attn/vsa.cpp"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -176,37 +176,3 @@ steps:
                  - TEST_TYPE=precision_vsa
                agents:
                  queue: "default"
-            - path:
-                - "csrc/attn/vmoba_attn/**"
-                - "pyproject.toml"
-                - "docker/Dockerfile.python3.12"
-              config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "Precision Tests VMoBA"
-                env:
-                  - TEST_TYPE=precision_vmoba
-                agents:
-                  queue: "default"
-            - path:
-                - "csrc/attn/vmoba_attn/vmoba/**"
-                - "fastvideo/attention/backends/vmoba.py"
-                - "pyproject.toml"
-                - "docker/Dockerfile.python3.12"
-              config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "Inference Tests VMoBA"
-                env:  
-                  - TEST_TYPE=inference_vmoba
-                agents:
-                  queue: "default"
-            - path:
-                - "fastvideo/**"
-                - "pyproject.toml"
-                - "docker/Dockerfile.python3.12"
-              config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "Unit Tests"
-                env:
-                  - TEST_TYPE=unit_test
-                agents:
-                  queue: "default"
@@ -109,19 +109,6 @@ case "$TEST_TYPE" in
        log "Running distillation DMD tests..."
        MODAL_COMMAND="$MODAL_ENV WANDB_API_KEY=$WANDB_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_distill_dmd_tests"
        ;;
-        # run_inference_tests_vmoba
-    "inference_vmoba")
-        log "Running V-MoBA inference tests..."
-        MODAL_COMMAND="$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"
-        ;;
    *)
        log "Error: Unknown test type: $TEST_TYPE"
        exit 1
@@ -18,12 +18,6 @@ on:
        required: false
        default: false
        type: boolean
-      python_3_12_cuda_12_9:
-        description: 'Build Python 3.12 image Cuda 12.9'
-        required: false
-        default: false
-        type: boolean
-

 permissions:
  contents: read
@@ -55,13 +49,4 @@ jobs:
      python_version: '3.12'
      dockerfile_path: docker/Dockerfile.python3.12
      tag_suffix: py3.12
-    secrets: inherit
-
-  build-python-3-12-cuda-12-9:
-    if: ${{ github.event.inputs.python_3_12_cuda_12_9 == 'true' }}
-    uses: ./.github/workflows/build-image-template.yml
-    with:
-      python_version: '3.12'
-      dockerfile_path: docker/Dockerfile.python3.12.cuda12.9.1
-      tag_suffix: py3.12-cuda12.9.1
    secrets: inherit
@@ -62,8 +62,8 @@ on:
        required: false
        default: false
        type: boolean
-      run_unit_test:
-        description: "Run unit-test"
+      run_nightly_test:
+        description: "Run nightly-test"
        required: false
        default: false
        type: boolean
@@ -93,7 +93,6 @@ jobs:
      inference-test-STA: ${{ steps.filter.outputs.inference-test-STA }}
      precision-test-STA: ${{ steps.filter.outputs.precision-test-STA }}
      precision-test-VSA: ${{ steps.filter.outputs.precision-test-VSA }}
-      unit-test: ${{ steps.filter.outputs.unit-test }}
    steps:
      - uses: actions/checkout@v4
      - uses: dorny/paths-filter@v3
@@ -103,21 +102,18 @@ jobs:
            # Define reusable path patterns
            common-paths: &common-paths
              - 'pyproject.toml'
-              - 'docker/Dockerfile.python3.10'
-              - 'docker/Dockerfile.python3.11'
              - 'docker/Dockerfile.python3.12'
            sta-kernel-paths: &sta-kernel-paths
-              - 'csrc/attn/sliding_tile_attn/**'
-              - 'csrc/attn/sliding_tile_attn/tk/**'
-              - 'csrc/attn/sliding_tile_attn/setup.py'
-              - 'csrc/attn/sliding_tile_attn/config_sta.py'
-              - 'csrc/attn/sliding_tile_attn/st_attn.cpp'
+              - 'csrc/attn/st_attn/**'
+              - 'csrc/attn/setup_sta.py'
+              - 'csrc/attn/config_sta.py'
+              - 'csrc/attn/st_attn.cpp'
            vsa-kernel-paths: &vsa-kernel-paths
-              - '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'
+              - 'csrc/attn/vsa/**'
+              - 'csrc/attn/tk/**'
+              - 'csrc/attn/setup_vsa.py'
+              - 'csrc/attn/config_vsa.py'
+              - 'csrc/attn/vsa.cpp'
            vsa-paths: &vsa-paths
              - 'fastvideo/**'
              - *common-paths
@@ -158,9 +154,6 @@ jobs:
            precision-test-VSA:
              - *common-paths
              - *vsa-kernel-paths
-            unit-test:
-              - 'fastvideo/**'
-              - *common-paths

  encoder-test:
    needs: change-filter
@@ -241,7 +234,7 @@ jobs:
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
-
+  
  training-test:
    needs: change-filter
    if: >-
@@ -339,42 +332,23 @@ jobs:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}

-  unit-test:
-    needs: change-filter
+  nightly-test:
    if: >-
-      (github.event_name != 'workflow_dispatch' && needs.change-filter.outputs.unit-test == 'true') || 
-      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_unit_test == 'true')
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_nightly_test == 'true')
    uses: ./.github/workflows/runpod-test.yml
    with:
-      job_id: "unit-test"
-      gpu_type: "NVIDIA L40S"
-      gpu_count: 1
+      job_id: "nightly-test"
+      gpu_type: "NVIDIA A40"
+      gpu_count: 4
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/dataset/ -vs && pytest ./fastvideo/workflow/ -vs"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/nightly/test_e2e_overfit_single_sample.py -vs"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
-
-  # nightly-test:
-  #   if: >-
-  #     (github.event_name == 'workflow_dispatch' && github.event.inputs.run_nightly_test == 'true')
-  #   uses: ./.github/workflows/runpod-test.yml
-  #   with:
-  #     job_id: "nightly-test"
-  #     gpu_type: "NVIDIA A40"
-  #     gpu_count: 4
-  #     volume_size: 100
-  #     disk_size: 100
-  #     image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-  #     test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/nightly/test_e2e_overfit_single_sample.py -vs"
-  #     timeout_minutes: 30
-  #   secrets:
-  #     RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
-  #     RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
-  #     WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
+      WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}

  runpod-cleanup:
    # Add other jobs to this list as you create them
@@ -398,4 +372,4 @@ jobs:
          JOB_IDS: '["encoder-test", "vae-test", "transformer-test", "ssim-test-py3.10", "ssim-test-py3.11", "ssim-test-py3.12", "training-test", "training-test-VSA", "inference-test-STA", "precision-test-STA", "precision-test-VSA"]'
          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
          GITHUB_RUN_ID: ${{ github.run_id }}
-        run: python .github/scripts/runpod_cleanup.py
+        run: python .github/scripts/runpod_cleanup.py
@@ -5,7 +5,7 @@ on:
    branches:
      - main
    paths:
-      - "csrc/attn/sliding_tile_attn/setup.py"
+      - "csrc/attn/setup_sta.py"
  workflow_dispatch:

 jobs:
@@ -23,13 +23,13 @@ jobs:
      - name: Check if version changed
        id: check-version
        run: |
-          cd csrc/attn/sliding_tile_attn
+          cd csrc/attn
          # Get current commit's version
-          NEW_VERSION=$(grep -oP 'VERSION\s*=\s*"\K[^"]+' setup.py)
+          NEW_VERSION=$(grep -oP 'VERSION\s*=\s*"\K[^"]+' setup_sta.py)
          echo "New version: $NEW_VERSION"
          
          # Get previous version from git history
-          OLD_VERSION=$(git show HEAD~1:./setup.py | grep -oP 'VERSION\s*=\s*"\K[^"]+' || echo "0.0.0")
+          OLD_VERSION=$(git show HEAD~1:./setup_sta.py | grep -oP 'VERSION\s*=\s*"\K[^"]+' || echo "0.0.0")
          echo "Old version: $OLD_VERSION"
          
          if [ "$NEW_VERSION" != "$OLD_VERSION" ]; then
@@ -144,13 +144,13 @@ jobs:
          pip install setuptools
          pip install ninja packaging wheel

-          cd csrc/attn/sliding_tile_attn  # Move into the correct folder
+          cd csrc/attn  # Move into the correct folder
          git submodule update --init --recursive # Ensure ThunderKittens submodule is initialized
-          python setup.py bdist_wheel --dist-dir=dist
+          python setup_sta.py bdist_wheel --dist-dir=dist

      - name: Rename wheel file
        run: |
-          cd csrc/attn/sliding_tile_attn
+          cd csrc/attn

          CUDA_SHORT_VERSION=$(echo ${{ matrix.cuda-version }} | cut -d. -f1,2 | sed 's/\.//g')
          TORCH_SHORT_VERSION=$(echo ${{ matrix.torch-version }} | cut -d. -f1,2)
@@ -165,7 +165,7 @@ jobs:
        uses: actions/upload-artifact@v4
        with:
          name: ${{ env.wheel_name }}
-          path: csrc/attn/sliding_tile_attn/dist/*.whl
+          path: csrc/attn/dist/*.whl
          retention-days: 90

  publish_package:
@@ -239,11 +239,11 @@ jobs:
          pip install setuptools
          pip install ninja packaging wheel

-          cd csrc/attn/sliding_tile_attn  # Move into the correct folder
+          cd csrc/attn  # Move into the correct folder
          git submodule update --init --recursive # Ensure ThunderKittens submodule is initialized
-          python setup.py sdist --dist-dir=dist
+          python setup_sta.py sdist --dist-dir=dist

      - name: Publish release distributions to PyPI
        uses: pypa/gh-action-pypi-publish@release/v1
        with:
-          packages-dir: csrc/attn/sliding_tile_attn/dist/
+          packages-dir: csrc/attn/dist/
@@ -5,7 +5,7 @@ on:
    branches:
      - main
    paths:
-      - "csrc/attn/video_sparse_attn/setup.py"
+      - "csrc/attn/setup_vsa.py"
  workflow_dispatch:

 jobs:
@@ -23,13 +23,13 @@ jobs:
      - name: Check if version changed
        id: check-version
        run: |
-          cd csrc/attn/video_sparse_attn
+          cd csrc/attn
          # Get current commit's version
-          NEW_VERSION=$(grep -oP 'VERSION\s*=\s*"\K[^"]+' setup.py)
+          NEW_VERSION=$(grep -oP 'VERSION\s*=\s*"\K[^"]+' setup_vsa.py)
          echo "New version: $NEW_VERSION"
          
          # Get previous version from git history
-          OLD_VERSION=$(git show HEAD~1:./setup.py | grep -oP 'VERSION\s*=\s*"\K[^"]+' || echo "0.0.0")
+          OLD_VERSION=$(git show HEAD~1:./setup_vsa.py | grep -oP 'VERSION\s*=\s*"\K[^"]+' || echo "0.0.0")
          echo "Old version: $OLD_VERSION"
          
          if [ "$NEW_VERSION" != "$OLD_VERSION" ]; then
@@ -152,13 +152,13 @@ jobs:
          pip install setuptools
          pip install ninja packaging wheel

-          cd csrc/attn/video_sparse_attn  # Move into the correct folder
+          cd csrc/attn  # Move into the correct folder
          git submodule update --init --recursive # Ensure ThunderKittens submodule is initialized
-          python setup.py bdist_wheel --dist-dir=dist
+          python setup_vsa.py bdist_wheel --dist-dir=dist

      - name: Rename wheel file
        run: |
-          cd csrc/attn/video_sparse_attn
+          cd csrc/attn

          CUDA_SHORT_VERSION=$(echo ${{ matrix.torch-cuda.cuda-version }} | cut -d. -f1,2 | sed 's/\.//g')
          TORCH_SHORT_VERSION=$(echo ${{ matrix.torch-cuda.torch-version }} | cut -d. -f1,2)
@@ -173,7 +173,7 @@ jobs:
        uses: actions/upload-artifact@v4
        with:
          name: ${{ env.wheel_name }}
-          path: csrc/attn/video_sparse_attn/dist/*.whl
+          path: csrc/attn/dist/*.whl
          retention-days: 90

  publish_package:
@@ -247,11 +247,11 @@ jobs:
          pip install setuptools
          pip install ninja packaging wheel

-          cd csrc/attn/video_sparse_attn  # Move into the correct folder
+          cd csrc/attn  # Move into the correct folder
          git submodule update --init --recursive # Ensure ThunderKittens submodule is initialized
-          python setup.py sdist --dist-dir=dist
+          python setup_vsa.py sdist --dist-dir=dist

      - name: Publish release distributions to PyPI
        uses: pypa/gh-action-pypi-publish@release/v1
        with:
-          packages-dir: csrc/attn/video_sparse_attn/dist/
+          packages-dir: csrc/attn/dist/
@@ -64,6 +64,3 @@ docs/source/distillation/examples/
 !docs/source/_static/images/**/*.png
 !comfyui/assets/**/*.png
 !comfyui/assets/**/*.gif
-
-dmd_t2v_output/
-preprocess_output_text/
@@ -1,7 +1,3 @@
-[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 "csrc/attn/tk"]
+	path = csrc/attn/tk
 	url = https://github.com/HazyResearch/ThunderKittens.git
@@ -22,7 +22,6 @@ exclude: |
    examples/.*|
    .github/workflows/fastvideo-publish.yml|
    .github/workflows/sta-publish.yml|
-    .github/workflows/vsa-publish.yml|
    .github/workflows/build-image-template.yml|
    docs/source/inference/support_matrix.md
  )
@@ -1,5 +1,5 @@
 <div align="center">
-<img src=assets/logos/logo.svg width="30%"/>
+<img src=assets/logo.png width="30%"/>
 </div>

 **FastVideo is a unified post-training and inference framework for accelerated video generation.**
@@ -7,7 +7,7 @@
 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/S7HLCSTh" 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.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-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/qqPzbrw" target="_blank"> <b> WeChat </b> </a> |
 </p>

 <div align="center">
@@ -1,6 +0,0 @@
-<svg width="160" height="93" viewBox="0 0 160 93" fill="none" xmlns="http://www.w3.org/2000/svg">
-<path d="M28.8511 91.66L57.6319 1.86368H64.5394L35.7585 91.66H28.8511Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
-<path d="M15.0376 91.66L43.8185 1.86368H46.1209L17.3401 91.66H15.0376Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
-<path d="M1.22217 91.66L30.003 1.86366H31.1543L2.3734 91.66H1.22217Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.15122"/>
-<path d="M71.4465 1.86483L42.666 91.6599H69.144L78.3538 58.2746H123.251L129.007 39.855H84.1099L89.866 22.5868H152.032L157.788 1.86483H71.4465Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
-</svg>
@@ -1,18 +0,0 @@
-<svg width="252" height="105" viewBox="0 0 252 105" fill="none" xmlns="http://www.w3.org/2000/svg">
-<path d="M89.4843 55.5457H101.361L87.7028 101H74.638L89.4843 55.5457Z" fill="#356CFF"/>
-<path fill-rule="evenodd" clip-rule="evenodd" d="M96.0167 1.00057H112.645L118.583 48.273H104.924L103.737 39.7882H79.9827L67.5117 55.5457H85.3273L43.1638 101H28.3174L22.3789 55.5457H33.6621L38.4129 91.3031L58.604 68.2729H44.3515L96.0167 1.00057ZM100.768 13.1217L87.7028 29.4852H103.143L100.768 13.1217Z" fill="#356CFF"/>
-<path d="M37.2252 1.00057L22.3789 48.273H36.0375L40.7884 30.6974L62.6727 30.6974L69.6727 21.0005L43.7576 21.0004L46.7269 11.9096L77.6727 11.9096L86 1.00057L37.2252 1.00057Z" fill="#356CFF"/>
-<path fill-rule="evenodd" clip-rule="evenodd" d="M108.488 55.5457L94.2351 101C94.2351 101 105.518 101 120.959 101C136.399 101 144.078 93.0133 148.276 79.788C152.432 68.0157 153.027 55.5457 136.399 55.5457C119.771 55.5457 108.488 55.5457 108.488 55.5457ZM109.081 90.697L116.802 65.8487C116.802 65.8487 120.959 65.8487 132.242 65.8487C143.525 65.8487 137.586 78.5759 135.211 84.0304C133.307 88.4021 127.491 90.697 122.74 90.697C117.989 90.697 109.081 90.697 109.081 90.697Z" fill="#356CFF"/>
-<path d="M173.188 1.00056L168.625 11.9096C168.625 11.9096 149.386 11.9092 142.525 11.9095C135.664 11.9098 136.586 20.3944 141.337 20.3944H159.747C168.654 20.3944 166.961 33.6899 163.904 38.5761C160.467 44.0675 157.371 48.273 148.463 48.273L125.188 48.273L124 37.97L147.87 37.97C153.808 37.97 156.184 29.4852 151.433 29.4852H131.836C120.142 29.4852 125.897 1.00043 141.337 1.00043L173.188 1.00056Z" fill="#356CFF"/>
-<path d="M179.938 1.00056L175.688 11.9096L191.221 11.9096L179.938 48.273H192.409L203.692 11.9096L219.132 11.9095L223.289 1.00043L179.938 1.00056Z" fill="#356CFF"/>
-<path d="M161.341 55.5457H202.845L198.5 65.8487H169.654L167.279 73.7268H188.5L184.749 82.8177H164.31L161.934 90.697H190.251L186.624 101H146.494L161.341 55.5457Z" fill="#356CFF"/>
-<path fill-rule="evenodd" clip-rule="evenodd" d="M230.821 54.9391C255.169 54.9391 251.776 67.0602 249.231 77.9692C246.686 88.8783 240.917 101 217.757 101C194.596 101 195.606 88.8783 199.347 77.9692C203.089 67.0602 206.473 54.9391 230.821 54.9391ZM237.948 77.9692C239.984 70.6965 240.917 65.242 228.446 65.242C215.975 65.242 211.818 71.9087 210.037 77.9692C208.255 84.0298 208.255 91.3025 219.538 91.3025C230.821 91.3025 235.911 85.2419 237.948 77.9692Z" fill="#356CFF"/>
-<path d="M173.188 1.00056L168.625 11.9096C168.625 11.9096 149.386 11.9092 142.525 11.9095C135.664 11.9098 136.586 20.3944 141.337 20.3944M173.188 1.00056C173.188 1.00056 156.777 1.00043 141.337 1.00043M173.188 1.00056L141.337 1.00043M141.337 20.3944C146.088 20.3944 150.839 20.3944 159.747 20.3944M141.337 20.3944H159.747M159.747 20.3944C168.654 20.3944 166.961 33.6899 163.904 38.5761C160.467 44.0675 157.371 48.273 148.463 48.273M148.463 48.273C139.556 48.273 125.188 48.273 125.188 48.273M148.463 48.273L125.188 48.273M125.188 48.273L124 37.97M124 37.97C124 37.97 141.931 37.97 147.87 37.97M124 37.97L147.87 37.97M147.87 37.97C153.808 37.97 156.184 29.4852 151.433 29.4852M151.433 29.4852C146.682 29.4852 138.962 29.4852 131.836 29.4852M151.433 29.4852H131.836M131.836 29.4852C120.142 29.4852 125.897 1.00043 141.337 1.00043M37.2252 1.00057L22.3789 48.273H36.0375L40.7884 30.6974L62.6727 30.6974L69.6727 21.0005L43.7576 21.0004L46.7269 11.9096L77.6727 11.9096L86 1.00057L37.2252 1.00057ZM96.0167 1.00057H112.645L118.583 48.273H104.924L103.737 39.7882H79.9827L67.5117 55.5457H85.3273L43.1638 101H28.3174L22.3789 55.5457H33.6621L38.4129 91.3031L58.604 68.2729H44.3515L96.0167 1.00057ZM87.7028 29.4852L100.768 13.1217L103.143 29.4852H87.7028ZM89.4843 55.5457H101.361L87.7028 101H74.638L89.4843 55.5457ZM108.488 55.5457L94.2351 101C94.2351 101 105.518 101 120.959 101C136.399 101 144.078 93.0133 148.276 79.788C152.432 68.0157 153.027 55.5457 136.399 55.5457C119.771 55.5457 108.488 55.5457 108.488 55.5457ZM116.802 65.8487L109.081 90.697C109.081 90.697 117.989 90.697 122.74 90.697C127.491 90.697 133.307 88.4021 135.211 84.0304C137.586 78.5759 143.525 65.8487 132.242 65.8487C120.959 65.8487 116.802 65.8487 116.802 65.8487ZM179.938 1.00056L175.688 11.9096L191.221 11.9096L179.938 48.273H192.409L203.692 11.9096L219.132 11.9095L223.289 1.00043L179.938 1.00056ZM161.341 55.5457H202.845L198.5 65.8487H169.654L167.279 73.7268H188.5L184.749 82.8177H164.31L161.934 90.697H190.251L186.624 101H146.494L161.341 55.5457ZM230.821 54.9391C255.169 54.9391 251.776 67.0602 249.231 77.9692C246.686 88.8783 240.917 101 217.757 101C194.596 101 195.606 88.8783 199.347 77.9692C203.089 67.0602 206.473 54.9391 230.821 54.9391ZM228.446 65.242C240.917 65.242 239.984 70.6965 237.948 77.9692C235.911 85.2419 230.821 91.3025 219.538 91.3025C208.255 91.3025 208.255 84.0298 210.037 77.9692C211.818 71.9087 215.975 65.242 228.446 65.242Z" stroke="#356CFF" stroke-width="1.18771"/>
-<path d="M15.2524 55.5451L21.191 100.999L24.7541 100.999L18.8156 55.5451L15.2524 55.5451Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.18771"/>
-<path d="M8.12646 55.5451L14.065 100.999L15.2527 100.999L9.31417 55.5451L8.12646 55.5451Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.18771"/>
-<path d="M1 55.5451L6.93853 100.999L7.53239 100.999L1.59385 55.5451L1 55.5451Z" fill="#356CFF" stroke="#356CFF" stroke-width="0.593853"/>
-<path d="M15.2524 48.2724L30.0988 1H33.6619L18.8156 48.2724H15.2524Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.18771"/>
-<path d="M8.12646 48.2724L22.9728 1H24.1605L9.31417 48.2724H8.12646Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.18771"/>
-<path d="M1 48.2724L15.8463 1H16.4402L1.59385 48.2724H1Z" fill="#356CFF" stroke="#356CFF" stroke-width="0.593853"/>
-<path d="M85.3271 55.5457H67.5116L87 12.7363L44.3513 68.2729H58.6038L43.1636 101L85.3271 55.5457Z" fill="#FDC717" stroke="#FDC717" stroke-width="1.18771" stroke-miterlimit="16"/>
-</svg>
@@ -1,6 +0,0 @@
-<svg width="160" height="93" viewBox="0 0 160 93" fill="none" xmlns="http://www.w3.org/2000/svg">
-<path d="M28.8511 91.66L57.6319 1.86368H64.5394L35.7585 91.66H28.8511Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
-<path d="M15.0376 91.66L43.8185 1.86368H46.1209L17.3401 91.66H15.0376Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
-<path d="M1.22217 91.66L30.003 1.86366H31.1543L2.3734 91.66H1.22217Z" fill="#356CFF" stroke="#356CFF" stroke-width="1.15122"/>
-<path d="M71.4465 1.86483L42.666 91.6599H69.144L78.3538 58.2746H123.251L129.007 39.855H84.1099L89.866 22.5868H152.032L157.788 1.86483H71.4465Z" fill="#356CFF" stroke="#356CFF" stroke-width="2.30244"/>
-</svg>
@@ -1,2 +1,2 @@
 recursive-include tk *
-include config_vsa.py
+include config.py
@@ -25,9 +25,9 @@ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave
 sudo apt update
 sudo apt install clang-11
 ```
-(If you use CUDA12.8)
+(If you use CUDA12.4)
 ```bash
-export CUDA_HOME=/usr/local/cuda-12.8
+export CUDA_HOME=/usr/local/cuda-12.4
 export PATH=${CUDA_HOME}/bin:${PATH} 
 export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
 ```
@@ -0,0 +1,4 @@
+    off_hz = tl.program_id(2)
+    b = off_hz // H
+    h = off_hz % H
+    meta_base = ((b * H + h) * q_tiles + q_blk)
@@ -1,7 +1,7 @@
 import os
 import subprocess

-from config_sta import kernels, sources, target
+from csrc.attn.config_sta import kernels, sources, target
 from setuptools import find_packages, setup
 from torch.utils.cpp_extension import BuildExtension, CUDAExtension

@@ -9,7 +9,7 @@ target = target.lower()

 # Package metadata
 PACKAGE_NAME = "st_attn"
-VERSION = "0.0.6"
+VERSION = "0.0.4"
 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"
@@ -9,10 +9,10 @@ target = target.lower()

 # Package metadata
 PACKAGE_NAME = "vsa"
-VERSION = "0.0.3"
+VERSION = "0.0.1"
 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"
+URL = "https://github.com/hao-ai-lab/FastVideo/tree/main/csrc/attn"

 # Set environment variables
 tk_root = os.getenv('THUNDERKITTENS_ROOT', os.path.abspath(os.path.join(os.getcwd(), 'tk/')))
@@ -1,2 +0,0 @@
-recursive-include tk *
-include config_sta.py
@@ -1,87 +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
-
-## 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.
@@ -13,9 +13,9 @@ 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_()
+    q_ = Q.clone().requires_grad_()
+    k_ = K.clone().requires_grad_()
+    v_ = V.clone().requires_grad_()

    QK = torch.matmul(q_, k_.transpose(-2, -1))
    QK /= (q_.size(-1) ** 0.5)
@@ -35,9 +35,9 @@ def pytorch_test(Q, K, V, block_sparse_mask, dO):


 def block_sparse_kernel_test(Q, K, V, block_sparse_mask, variable_block_sizes, non_pad_index, dO):
-    Q = Q.detach().requires_grad_()
-    K = K.detach().requires_grad_()
-    V = V.detach().requires_grad_()
+    Q = Q.clone().requires_grad_()
+    K = K.clone().requires_grad_()
+    V = V.clone().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)
@@ -60,9 +60,11 @@ def get_non_pad_index(

    return index_pad[index_mask]

-def generate_tensor(shape, dtype, device):
+def generate_tensor(shape, mean, std, dtype, device):
    tensor = torch.randn(shape, dtype=dtype, device=device)
-    return tensor
+    magnitude = torch.norm(tensor, dim=-1, keepdim=True)
+    scaled_tensor = tensor * (torch.randn(magnitude.shape, dtype=dtype, device=device) * std + mean) / magnitude
+    return scaled_tensor.contiguous()

 def 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)
@@ -73,7 +75,7 @@ def vsa_pad(x, non_pad_index, num_blocks, block_size):
    padded_x[:, :, non_pad_index, :] = x
    return padded_x

-def check_correctness(h, d, num_blocks, k,  num_iterations=20, error_mode='all'):
+def check_correctness(h, d, num_blocks, k, mean, std, 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},
@@ -89,10 +91,10 @@ def check_correctness(h, d, num_blocks, k,  num_iterations=20, error_mode='all')
    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)
+        Q = generate_tensor((1, h, S, d), mean, std, torch.bfloat16, device)
+        K = generate_tensor((1, h, S, d), mean, std, torch.bfloat16, device)
+        V = generate_tensor((1, h, S, d), mean, std, torch.bfloat16, device)
+        dO = generate_tensor((1, h, S, d), mean, std, torch.bfloat16, device)

        # dO_padded = torch.zeros_like(dO_padded)
        # dO_padded[:, :, non_pad_index, :] = dO
@@ -105,8 +107,7 @@ def check_correctness(h, d, num_blocks, k,  num_iterations=20, error_mode='all')
                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())
+                results[name]['max_diff'] = max(results[name]['max_diff'], torch.max(abs_diff).item())
        if torch.cuda.is_available():
            torch.cuda.empty_cache()

@@ -118,27 +119,27 @@ def check_correctness(h, d, num_blocks, k,  num_iterations=20, error_mode='all')

    return results

-def generate_error_graphs(h, d, error_mode='all'):
+def generate_error_graphs(h, d, mean, std, 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(f"\nError Analysis for h={h}, d={d}, mean={mean}, std={std}, 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}")
+          f"{'gQ Avg':<12} {'gQ Max':<12} "
+          f"{'gK Avg':<12} {'gK Max':<12} "
+          f"{'gV Avg':<12} {'gV Max':<12} "
+          f"{'gO Avg':<12} {'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)
+        results = check_correctness(h, d, num_blocks, k, mean, std, 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} "
@@ -149,8 +150,10 @@ def generate_error_graphs(h, d, error_mode='all'):

 if __name__ == "__main__":
    h, d = 16, 128
+    mean = 0.0
+    std = 1
    print("Block Sparse Attention with Variable Block Sizes Analysis")
    print("=" * 60)
    for mode in ['backward']:
-        generate_error_graphs(h, d, error_mode=mode)
+        generate_error_graphs(h, d, mean, std, error_mode=mode)
    print("\nAnalysis completed for all modes.")
@@ -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.
@@ -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
-```
-
@@ -1,24 +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=[]
-)
@@ -1,97 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-import torch
-import pytest
-import random
-from csrc.attn.vmoba_attn.vmoba  import moba_attn_varlen
-
-def generate_test_data(batch_size, total_seqlen, num_heads, head_dim, dtype, device="cuda"):
-    """
-    Generates random data for testing the variable-length attention function.
-    """
-    torch.manual_seed(42)
-    random.seed(42)
-    torch.cuda.manual_seed_all(42)
-
-    # Generate sequence lengths for each item in the batch
-    if batch_size > 1:
-        # Ensure sequence lengths are reasonably distributed
-        avg_seqlen = total_seqlen // batch_size
-        seqlens = [random.randint(avg_seqlen // 2, avg_seqlen + avg_seqlen // 2) for _ in range(batch_size - 1)]
-        remaining_len = total_seqlen - sum(seqlens)
-        if remaining_len > 0:
-            seqlens.append(remaining_len)
-        else: # Adjust if sum exceeds total_seqlen
-            seqlens.append(avg_seqlen)
-            current_sum = sum(seqlens)
-            seqlens[-1] -= (current_sum - total_seqlen)
-        # Ensure all lengths are positive
-        seqlens = [max(1, s) for s in seqlens]
-        # Final adjustment to match total_seqlen
-        seqlens[-1] += total_seqlen - sum(seqlens)
-
-    else:
-        seqlens = [total_seqlen]
-
-    cu_seqlens = torch.tensor([0] + list(torch.cumsum(torch.tensor(seqlens), 0)), device=device, dtype=torch.int32)
-    max_seqlen = max(seqlens) if seqlens else 0
-
-    q = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
-    k = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
-    v = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
-
-    return q, k, v, cu_seqlens, max_seqlen
-
-
-@pytest.mark.parametrize("batch_size", [1, 2])
-@pytest.mark.parametrize("total_seqlen", [512, 1024])
-@pytest.mark.parametrize("num_heads", [8])
-@pytest.mark.parametrize("head_dim", [64])
-@pytest.mark.parametrize("moba_chunk_size", [64])
-@pytest.mark.parametrize("moba_topk", [2, 4])
-@pytest.mark.parametrize("select_mode", ["topk", "threshold"])
-@pytest.mark.parametrize("threshold_type", ["query_head", "head_global", "overall"])
-@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
-def test_moba_attn_varlen_forward(
-    batch_size, total_seqlen, num_heads, head_dim, moba_chunk_size, moba_topk, select_mode, threshold_type, dtype
-):
-    """
-    Tests the forward pass of moba_attn_varlen for basic correctness.
-    It checks output shape, dtype, and for the presence of NaNs/Infs.
-    """
-    if dtype == torch.float32:
-        pytest.skip("float32 is not supported in flash attention")
-
-    q, k, v, cu_seqlens, max_seqlen = generate_test_data(
-        batch_size, total_seqlen, num_heads, head_dim, dtype
-    )
-
-    # Ensure chunk size is not larger than the smallest sequence length
-    min_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).min().item()
-    if moba_chunk_size > min_seqlen:
-        pytest.skip("moba_chunk_size is larger than the minimum sequence length in the batch")
-
-    try:
-        output = moba_attn_varlen(
-            q=q,
-            k=k,
-            v=v,
-            cu_seqlens=cu_seqlens,
-            max_seqlen=max_seqlen,
-            moba_chunk_size=moba_chunk_size,
-            moba_topk=moba_topk,
-            select_mode=select_mode,
-            threshold_type=threshold_type,
-            simsum_threshold=0.5, # A reasonable default for threshold mode
-        )
-    except Exception as e:
-        pytest.fail(f"moba_attn_varlen forward pass failed with exception: {e}")
-
-    # 1. Check output shape
-    assert output.shape == q.shape, f"Expected output shape {q.shape}, but got {output.shape}"
-
-    # 2. Check output dtype
-    assert output.dtype == q.dtype, f"Expected output dtype {q.dtype}, but got {output.dtype}"
-
-    # 3. Check for NaNs or Infs in the output
-    assert torch.all(torch.isfinite(output)), "Output contains NaN or Inf values"
@@ -1,2 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from .vmoba import moba_attn_varlen, process_moba_input, process_moba_output
@@ -1,860 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# Adapt from https://github.com/KwaiVGI/VMoBA/blob/main/src/vmoba.py
-
-import random
-import time
-import os
-import torch
-from typing import Tuple
-from flash_attn import flash_attn_varlen_func  # Use the new flash attention function
-from flash_attn.flash_attn_interface import _flash_attn_varlen_forward, _flash_attn_varlen_backward
-from functools import lru_cache
-from einops import rearrange
-
-
-@lru_cache(maxsize=16)
-def calc_chunks(cu_seqlen, moba_chunk_size):
-    """
-    Calculate chunk boundaries.
-
-    For vision tasks we include all chunks (even the last one which might be shorter)
-    so that every chunk can be selected.
-    """
-    batch_sizes = cu_seqlen[1:] - cu_seqlen[:-1]
-    batch_num_chunk = (batch_sizes + (moba_chunk_size - 1)) // moba_chunk_size
-    cu_num_chunk = torch.ones(
-        batch_num_chunk.numel() + 1,
-        device=cu_seqlen.device,
-        dtype=batch_num_chunk.dtype,
-    )
-    cu_num_chunk[1:] = batch_num_chunk.cumsum(dim=0)
-    num_chunk = cu_num_chunk[-1]
-    chunk_sizes = torch.full(
-        (num_chunk + 1,), moba_chunk_size, dtype=torch.int32, device=cu_seqlen.device
-    )
-    chunk_sizes[0] = 0
-    batch_last_chunk_size = batch_sizes - (batch_num_chunk - 1) * moba_chunk_size
-    chunk_sizes[cu_num_chunk[1:]] = batch_last_chunk_size
-    cu_chunk = chunk_sizes.cumsum(dim=-1, dtype=torch.int32)
-    chunk_to_batch = torch.zeros(
-        (num_chunk,), dtype=torch.int32, device=cu_seqlen.device
-    )
-    chunk_to_batch[cu_num_chunk[1:-1]] = 1
-    chunk_to_batch = chunk_to_batch.cumsum(dim=0, dtype=torch.int32)
-
-    # Do not filter out any chunk
-    filtered_chunk_indices = torch.arange(
-        num_chunk, device=cu_seqlen.device, dtype=torch.int32
-    )
-    num_filtered_chunk = num_chunk
-
-    return cu_chunk, filtered_chunk_indices, num_filtered_chunk, chunk_to_batch
-
-
-# --- Threshold Selection Helper Functions ---
-
-def _select_threshold_query_head(
-    gate: torch.Tensor,
-    valid_gate_mask: torch.Tensor,
-    gate_self_chunk_mask: torch.Tensor,
-    simsum_threshold: float
-) -> torch.Tensor:
-    """
-    Selects chunks for each <query, head> pair based on threshold.
-    Normalization and sorting happen along the chunk dimension (dim=0).
-    """
-    C, H, S = gate.shape
-    eps = 1e-6
-
-    # LSE‐style normalization per <head, query> (across chunks)
-    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
-    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
-
-    row_min = gate_min_val.amin(dim=0)        # (H, S)
-    row_max = gate_masked.amax(dim=0)         # (H, S)
-    denom = row_max - row_min
-    denom = torch.where(denom <= eps, torch.ones_like(denom), denom) # avoid divide‑by‑zero
-
-    gate_norm = (gate - row_min.unsqueeze(0)) / denom.unsqueeze(0)
-    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
-
-    # 1) pull out the self‐chunk’s normalized weight for each <head,seq>
-    self_norm = (gate_norm * gate_self_chunk_mask).sum(dim=0)  # (H, S)
-
-    # 2) compute how much more normalized weight we need beyond self
-    total_norm_sum = gate_norm.sum(dim=0) # (H, S)
-    remain_ratio = simsum_threshold - self_norm / (total_norm_sum + eps)  # (H, S)
-    remain_ratio = torch.clamp(remain_ratio, min=0.0) # if already ≥ thresh, no extra needed
-
-    # 3) zero out the self‐chunk in a copy, so we only sort “others”
-    others_norm = gate_norm.clone()
-    others_norm[gate_self_chunk_mask] = 0.0
-
-    # 4) sort the other chunks by descending norm, per <head,seq>
-    sorted_norm, sorted_idx = torch.sort(others_norm, descending=True, dim=0) # (C, H, S)
-
-    # 5) cumulative‑sum the sorted norms per <head,seq>
-    cumsum_others = sorted_norm.cumsum(dim=0) # (C, H, S)
-
-    # 6) for each <head,seq>, find the smallest k where cumsum_ratio ≥ remain_ratio
-    ratio = cumsum_others / (total_norm_sum.unsqueeze(0) + eps) # (C, H, S)
-    cond = ratio >= remain_ratio.unsqueeze(0) # (C, H, S) boolean mask
-    any_cond = cond.any(dim=0) # (H, S)
-    # Find the index of the first True value along dim 0. If none, use C-1.
-    cutoff = torch.where(any_cond, cond.float().argmax(dim=0), torch.full_like(any_cond, fill_value=C - 1)) # (H, S)
-
-    # 7) build a mask in sorted order up to that cutoff
-    idx_range = torch.arange(C, device=gate.device).view(-1, 1, 1) # (C, 1, 1)
-    sorted_mask = idx_range <= cutoff.unsqueeze(0) # (C, H, S)
-
-    # 8) scatter it back to original chunk order
-    others_mask = torch.zeros_like(gate, dtype=torch.bool)
-    others_mask.scatter_(0, sorted_idx, sorted_mask)
-
-    # 9) finally, include every self‐chunk plus all selected others
-    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
-
-    return final_gate_mask
-
-
-def _select_threshold_block(
-    gate: torch.Tensor,
-    valid_gate_mask: torch.Tensor,
-    gate_self_chunk_mask: torch.Tensor,
-    simsum_threshold: float
-) -> torch.Tensor:
-    """
-    Selects <query, head> pairs for each block based on threshold.
-    Normalization and sorting happen across the head and sequence dimensions (dim=1, 2).
-    """
-    C, H, S = gate.shape
-    HS = H * S
-    eps = 1e-6
-
-    # LSE‐style normalization per block (across heads and queries)
-    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
-    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
-
-    block_max = gate_masked.amax(dim=(1, 2), keepdim=True) # (C, 1, 1)
-    block_min = gate_min_val.amin(dim=(1, 2), keepdim=True) # (C, 1, 1)
-    block_denom = block_max - block_min
-    block_denom = torch.where(block_denom <= eps, torch.ones_like(block_denom), block_denom) # (C, 1, 1)
-
-    gate_norm = (gate - block_min) / block_denom # (C, H, S)
-    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
-
-    # 1) identify normalized weights of entries that *are* self-chunks (from query perspective)
-    self_norm_entries = gate_norm * gate_self_chunk_mask # (C, H, S)
-    # Sum these weights *per block*
-    self_norm_sum_per_block = self_norm_entries.sum(dim=(1, 2)) # (C,)
-
-    # 2) compute how much more normalized weight each block needs beyond its self-chunk contributions
-    total_norm_sum_per_block = gate_norm.sum(dim=(1, 2)) # (C,)
-    remain_ratio = simsum_threshold - self_norm_sum_per_block / (total_norm_sum_per_block + eps) # (C,)
-    remain_ratio = torch.clamp(remain_ratio, min=0.0) # (C,)
-
-    # 3) zero out the self‐chunk entries in a copy, so we only sort “others”
-    others_norm = gate_norm.clone()
-    others_norm[gate_self_chunk_mask] = 0.0 # Zero out self entries
-
-    # 4) sort the other <head, seq> pairs by descending norm, per block
-    others_flat = others_norm.contiguous().view(C, HS) # (C, H*S)
-    sorted_others_flat, sorted_indices_flat = torch.sort(others_flat, dim=1, descending=True) # (C, H*S)
-
-    # 5) cumulative‑sum the sorted norms per block
-    cumsum_others_flat = sorted_others_flat.cumsum(dim=1) # (C, H*S)
-
-    # 6) for each block, find the smallest k where cumsum_ratio ≥ remain_ratio
-    ratio_flat = cumsum_others_flat / (total_norm_sum_per_block.unsqueeze(1) + eps) # (C, H*S)
-    cond_flat = ratio_flat >= remain_ratio.unsqueeze(1) # (C, H*S) boolean mask
-    any_cond = cond_flat.any(dim=1) # (C,)
-    # Find the index of the first True value along dim 1. If none, use HS-1.
-    cutoff_flat = torch.where(any_cond, cond_flat.float().argmax(dim=1), torch.full_like(any_cond, fill_value=HS - 1)) # (C,)
-
-    # 7) build a mask in sorted order up to that cutoff per block
-    idx_range_flat = torch.arange(HS, device=gate.device).unsqueeze(0) # (1, H*S)
-    sorted_mask_flat = idx_range_flat <= cutoff_flat.unsqueeze(1) # (C, H*S)
-
-    # 8) scatter it back to original <head, seq> order per block
-    others_mask_flat = torch.zeros_like(others_flat, dtype=torch.bool) # (C, H*S)
-    others_mask_flat.scatter_(1, sorted_indices_flat, sorted_mask_flat)
-    others_mask = others_mask_flat.view(C, H, S) # (C, H, S)
-
-    # 9) finally, include every self‐chunk entry plus all selected others
-    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
-
-    return final_gate_mask
-
-
-def _select_threshold_overall(
-    gate: torch.Tensor,
-    valid_gate_mask: torch.Tensor,
-    gate_self_chunk_mask: torch.Tensor,
-    simsum_threshold: float
-) -> torch.Tensor:
-    """
-    Selects <chunk, query, head> triplets globally based on threshold.
-    Normalization and sorting happen across all valid entries.
-    """
-    C, H, S = gate.shape
-    CHS = C * H * S
-    eps = 1e-6
-
-    # LSE‐style normalization globally across all valid entries
-    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
-    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
-
-    overall_max = gate_masked.max() # scalar
-    overall_min = gate_min_val.min() # scalar
-    overall_denom = overall_max - overall_min
-    overall_denom = torch.where(overall_denom <= eps, torch.tensor(1.0, device=gate.device, dtype=gate.dtype), overall_denom)
-
-    gate_norm = (gate - overall_min) / overall_denom # (C, H, S)
-    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
-
-    # 1) identify normalized weights of entries that *are* self-chunks
-    self_norm_entries = gate_norm * gate_self_chunk_mask # (C, H, S)
-    # Sum these weights globally
-    self_norm_sum_overall = self_norm_entries.sum() # scalar
-
-    # 2) compute how much more normalized weight is needed globally beyond self-chunk contributions
-    total_norm_sum_overall = gate_norm.sum() # scalar
-    remain_ratio = simsum_threshold - self_norm_sum_overall / (total_norm_sum_overall + eps) # scalar
-    remain_ratio = torch.clamp(remain_ratio, min=0.0) # scalar
-
-    # 3) zero out the self‐chunk entries in a copy, so we only sort “others”
-    others_norm = gate_norm.clone()
-    others_norm[gate_self_chunk_mask] = 0.0 # Zero out self entries
-
-    # 4) sort all other entries by descending norm, globally
-    others_flat = others_norm.flatten() # (C*H*S,)
-    valid_others_mask_flat = valid_gate_mask.flatten() & ~gate_self_chunk_mask.flatten() # Mask for valid, non-self entries
-    
-    # Only sort the valid 'other' entries
-    valid_others_indices = torch.where(valid_others_mask_flat)[0]
-    valid_others_values = others_flat[valid_others_indices]
-
-    sorted_others_values, sort_perm = torch.sort(valid_others_values, descending=True) # (N_valid_others,)
-    sorted_original_indices = valid_others_indices[sort_perm] # Original indices in C*H*S space, sorted by value
-
-    # 5) cumulative‑sum the sorted valid 'other' norms globally
-    cumsum_others_values = sorted_others_values.cumsum(dim=0) # (N_valid_others,)
-
-    # 6) find the smallest k where cumsum_ratio ≥ remain_ratio globally
-    ratio_values = cumsum_others_values / (total_norm_sum_overall + eps) # (N_valid_others,)
-    cond_values = ratio_values >= remain_ratio # (N_valid_others,) boolean mask
-    any_cond = cond_values.any() # scalar
-    
-    # Find the index of the first True value in the *sorted* list. If none, use all valid others.
-    cutoff_idx_in_sorted = torch.where(
-        any_cond, 
-        cond_values.float().argmax(dim=0), 
-        torch.tensor(len(sorted_others_values) - 1, device=gate.device, dtype=torch.long)
-    )
-
-    # 7) build a mask selecting the top-k others based on the cutoff
-    # Select the original indices corresponding to the top entries in the sorted list
-    selected_other_indices = sorted_original_indices[:cutoff_idx_in_sorted + 1]
-
-    # 8) create the mask in the original flat shape
-    others_mask_flat = torch.zeros_like(others_flat, dtype=torch.bool) # (C*H*S,)
-    if selected_other_indices.numel() > 0: # Check if any 'other' indices were selected
-        others_mask_flat[selected_other_indices] = True
-    others_mask = others_mask_flat.view(C, H, S) # (C, H, S)
-
-    # 9) finally, include every self‐chunk entry plus all selected others
-    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
-
-    return final_gate_mask
-
-
-def _select_threshold_head_global(
-    gate: torch.Tensor,
-    valid_gate_mask: torch.Tensor,
-    gate_self_chunk_mask: torch.Tensor,
-    simsum_threshold: float
-) -> torch.Tensor:
-    """
-    Selects <chunk, query> globally for each head based on threshold.
-    """
-    C, H, S = gate.shape
-    eps = 1e-6
-
-    # 1) LSE‐style normalization per head (across chunks and sequence dims)
-    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf)
-    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf)
-
-    max_per_head = gate_masked.amax(dim=(0, 2), keepdim=True)  # (1, H, 1)
-    min_per_head = gate_min_val.amin(dim=(0, 2), keepdim=True)  # (1, H, 1)
-    denom = max_per_head - min_per_head
-    denom = torch.where(denom <= eps, torch.ones_like(denom), denom)
-
-    gate_norm = (gate - min_per_head) / denom
-    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0)  # (C, H, S)
-
-    # 2) sum normalized self‐chunk contributions per head
-    self_norm_sum = (gate_norm * gate_self_chunk_mask).sum(dim=(0, 2))  # (H,)
-
-    # 3) total normalized sum per head
-    total_norm_sum = gate_norm.sum(dim=(0, 2))  # (H,)
-
-    # 4) how much more normalized weight needed per head
-    remain_ratio = simsum_threshold - self_norm_sum / (total_norm_sum + eps)  # (H,)
-    remain_ratio = torch.clamp(remain_ratio, min=0.0)
-
-    # 5) zero out self‐chunk entries to focus on "others"
-    others_norm = gate_norm.clone()
-    others_norm[gate_self_chunk_mask] = 0.0  # (C, H, S)
-
-    # 6) flatten chunk and sequence dims, per head
-    CS = C * S
-    others_flat = others_norm.permute(1, 0, 2).reshape(H, CS)  # (H, C*S)
-    valid_flat  = (valid_gate_mask & ~gate_self_chunk_mask) \
-                   .permute(1, 0, 2).reshape(H, CS)           # (H, C*S)
-
-    # 7) vectorized selection of “others” per head
-    masked_flat = torch.where(valid_flat, others_flat, torch.zeros_like(others_flat))
-    sorted_vals, sorted_idx = torch.sort(masked_flat, dim=1, descending=True)  # (H, C*S)
-
-    cumsum_vals = sorted_vals.cumsum(dim=1)                                    # (H, C*S)
-    ratio_vals  = cumsum_vals / (total_norm_sum.unsqueeze(1) + eps)            # (H, C*S)
-    cond        = ratio_vals >= remain_ratio.unsqueeze(1)                      # (H, C*S)
-
-    has_cutoff = cond.any(dim=1)                                               # (H,)
-    default    = torch.full((H,), CS - 1, device=gate.device, dtype=torch.long)
-    cutoff     = torch.where(has_cutoff, cond.float().argmax(dim=1), default)  # (H,)
-
-    idx_range   = torch.arange(CS, device=gate.device).unsqueeze(0)            # (1, C*S)
-    sorted_mask = idx_range <= cutoff.unsqueeze(1)                             # (H, C*S)
-
-    selected_flat = torch.zeros_like(valid_flat)                               # (H, C*S)
-    selected_flat.scatter_(1, sorted_idx, sorted_mask)                         # (H, C*S)
-
-    # 8) reshape selection mask back to (C, H, S)
-    others_mask = selected_flat.reshape(H, C, S).permute(1, 0, 2)               # (C, H, S)
-
-    # 9) include self‐chunks plus selected others, and obey valid mask
-    final_gate_mask = valid_gate_mask & (gate_self_chunk_mask | others_mask)
-
-    return final_gate_mask
-
-
-class MixedAttention(torch.autograd.Function):
-    @staticmethod
-    def forward(
-        ctx,
-        q,
-        k,
-        v,
-        self_attn_cu_seqlen,
-        moba_q,
-        moba_kv,
-        moba_cu_seqlen_q,
-        moba_cu_seqlen_kv,
-        max_seqlen,
-        moba_chunk_size,
-        moba_q_sh_indices,
-    ):
-        ctx.max_seqlen = max_seqlen
-        ctx.moba_chunk_size = moba_chunk_size
-        ctx.softmax_scale = softmax_scale = q.shape[-1] ** (-0.5)
-
-        # Non-causal self-attention branch
-        # return out, softmax_lse, S_dmask, rng_state
-        self_attn_out_sh, self_attn_lse_hs, _, _ = _flash_attn_varlen_forward(
-            q=q,
-            k=k,
-            v=v,
-            cu_seqlens_q=self_attn_cu_seqlen,
-            cu_seqlens_k=self_attn_cu_seqlen,
-            max_seqlen_q=max_seqlen,
-            max_seqlen_k=max_seqlen,
-            softmax_scale=softmax_scale,
-            causal=False,
-            dropout_p=0.0,
-        )
-        # MOBA attention branch (non-causal)
-        moba_attn_out, moba_attn_lse_hs, _, _ = _flash_attn_varlen_forward(
-            q=moba_q,
-            k=moba_kv[:, 0],
-            v=moba_kv[:, 1],
-            cu_seqlens_q=moba_cu_seqlen_q,
-            cu_seqlens_k=moba_cu_seqlen_kv,
-            max_seqlen_q=max_seqlen,
-            max_seqlen_k=moba_chunk_size,
-            softmax_scale=softmax_scale,
-            causal=False,
-            dropout_p=0.0,
-        )
-
-        self_attn_lse_sh = self_attn_lse_hs.t().contiguous()
-        moba_attn_lse = moba_attn_lse_hs.t().contiguous()
-
-        output = torch.zeros((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
-        output_2d = output.view(-1, q.shape[2])
-
-        max_lse_1d = self_attn_lse_sh.view(-1)
-        max_lse_1d = max_lse_1d.index_reduce(
-            0, moba_q_sh_indices, moba_attn_lse.view(-1), "amax"
-        )
-        self_attn_lse_sh = self_attn_lse_sh - max_lse_1d.view_as(self_attn_lse_sh)
-        moba_attn_lse = (
-            moba_attn_lse.view(-1)
-            .sub(max_lse_1d.index_select(0, moba_q_sh_indices))
-            .reshape_as(moba_attn_lse)
-        )
-
-        mixed_attn_se_sh = self_attn_lse_sh.exp()
-        moba_attn_se = moba_attn_lse.exp()
-
-        mixed_attn_se_sh.view(-1).index_add_(
-            0, moba_q_sh_indices, moba_attn_se.view(-1)
-        )
-        mixed_attn_lse_sh = mixed_attn_se_sh.log()
-
-        # Combine self-attention output
-        factor = (self_attn_lse_sh - mixed_attn_lse_sh).exp()  # [S, H]
-        self_attn_out_sh = self_attn_out_sh * factor.unsqueeze(-1)
-        output_2d += self_attn_out_sh.reshape_as(output_2d)
-
-        # Combine MOBA attention output
-        mixed_attn_lse = (
-            mixed_attn_lse_sh.view(-1)
-            .index_select(0, moba_q_sh_indices)
-            .view_as(moba_attn_lse)
-        )
-        factor = (moba_attn_lse - mixed_attn_lse).exp()  # [S, H]
-        moba_attn_out = moba_attn_out * factor.unsqueeze(-1)
-        raw_attn_out = moba_attn_out.view(-1, moba_attn_out.shape[-1])
-        output_2d.index_add_(0, moba_q_sh_indices, raw_attn_out)
-        output = output.to(q.dtype)
-        mixed_attn_lse_sh = mixed_attn_lse_sh + max_lse_1d.view_as(mixed_attn_se_sh)
-        ctx.save_for_backward(
-            output,
-            mixed_attn_lse_sh,
-            q,
-            k,
-            v,
-            self_attn_cu_seqlen,
-            moba_q,
-            moba_kv,
-            moba_cu_seqlen_q,
-            moba_cu_seqlen_kv,
-            moba_q_sh_indices,
-        )
-
-        return output
-
-    @staticmethod
-    def backward(ctx, d_output):
-
-        max_seqlen = ctx.max_seqlen
-        moba_chunk_size = ctx.moba_chunk_size
-        softmax_scale = ctx.softmax_scale
-
-        (
-            output,
-            mixed_attn_vlse_sh,
-            q,
-            k,
-            v,
-            self_attn_cu_seqlen,
-            moba_q,
-            moba_kv,
-            moba_cu_seqlen_q,
-            moba_cu_seqlen_kv,
-            moba_q_sh_indices,
-        ) = ctx.saved_tensors
-
-        d_output = d_output.contiguous()
-
-        dq = torch.empty_like(q)
-        dk = torch.empty_like(k)
-        dv = torch.empty_like(v)
-        _ = _flash_attn_varlen_backward(
-            dout=d_output,
-            q=q,
-            k=k,
-            v=v,
-            out=output,
-            softmax_lse=mixed_attn_vlse_sh.t().contiguous(),
-            dq=dq,
-            dk=dk,
-            dv=dv,
-            cu_seqlens_q=self_attn_cu_seqlen,
-            cu_seqlens_k=self_attn_cu_seqlen,
-            max_seqlen_q=max_seqlen,
-            max_seqlen_k=max_seqlen,
-            softmax_scale=softmax_scale,
-            causal=False,
-            dropout_p=0.0,
-            softcap=0.0,
-            alibi_slopes=None,
-            deterministic=True,
-            window_size_left=-1,
-            window_size_right=-1
-        )
-
-        headdim = q.shape[-1]
-        d_moba_output = (
-            d_output.view(-1, headdim).index_select(0, moba_q_sh_indices).unsqueeze(1)
-        )
-        moba_output = (
-            output.view(-1, headdim).index_select(0, moba_q_sh_indices).unsqueeze(1)
-        )
-
-        mixed_attn_vlse = (
-            mixed_attn_vlse_sh.view(-1).index_select(0, moba_q_sh_indices).view(1, -1)
-        )
-
-        dmq = torch.empty_like(moba_q)
-        dmkv = torch.empty_like(moba_kv)
-        _ = _flash_attn_varlen_backward(
-            dout=d_moba_output,
-            q=moba_q,
-            k=moba_kv[:, 0],
-            v=moba_kv[:, 1],
-            out=moba_output,
-            softmax_lse=mixed_attn_vlse,
-            dq=dmq,
-            dk=dmkv[:,0],
-            dv=dmkv[:,1],
-            cu_seqlens_q=moba_cu_seqlen_q,
-            cu_seqlens_k=moba_cu_seqlen_kv,
-            max_seqlen_q=max_seqlen,
-            max_seqlen_k=moba_chunk_size,
-            softmax_scale=softmax_scale,
-            causal=False,
-            dropout_p=0.0,
-            softcap=0.0,
-            alibi_slopes=None,
-            deterministic=True,
-            window_size_left=-1,
-            window_size_right=-1
-        )
-
-        return dq, dk, dv, None, dmq, dmkv, None, None, None, None, None
-
-
-def moba_attn_varlen(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    v: torch.Tensor,
-    cu_seqlens: torch.Tensor,
-    max_seqlen: int,
-    moba_chunk_size: int,
-    moba_topk: int,
-    select_mode: str = 'threshold', # "topk" or "threshold"
-    simsum_threshold: float = 0.25,
-    threshold_type: str = 'query_head',
-) -> torch.Tensor:
-    """
-    Accelerated MOBA attention for vision tasks with proper LSE normalization.
-
-    This version:
-      - Splits KV into chunks.
-      - For each query head, selects the top-k relevant KV chunks (including the self chunk)
-        by amplifying the diagonal (self-chunk) logits.
-      - Aggregates the attention outputs from the selected chunks using a log-sum-exp
-        reduction so that attending to each query over the selected chunks is equivalent
-        to the original algorithm.
-    """
-    # Stack keys and values.
-    kv = torch.stack((k, v), dim=1)
-    seqlen, num_head, head_dim = q.shape
-
-    # Compute chunk boundaries.
-    cu_chunk, filtered_chunk_indices, num_filtered_chunk, chunk_to_batch = calc_chunks(
-        cu_seqlens, moba_chunk_size
-    )
-
-    self_attn_cu_seqlen = cu_chunk
-
-    # Update top-k selection to include the self chunk.
-    moba_topk = min(moba_topk, num_filtered_chunk)
-
-    # --- Build filtered KV from chunks ---
-    chunk_starts = cu_chunk[filtered_chunk_indices]          # [num_filtered_chunk]
-    chunk_ends = cu_chunk[filtered_chunk_indices + 1]          # [num_filtered_chunk]
-    chunk_lengths = chunk_ends - chunk_starts                  # [num_filtered_chunk]
-    max_chunk_len = int(chunk_lengths.max().item())
-
-    range_tensor = torch.arange(max_chunk_len, device=kv.device, dtype=chunk_starts.dtype).unsqueeze(0)
-    indices = chunk_starts.unsqueeze(1) + range_tensor
-    indices = torch.clamp(indices, max=kv.shape[0] - 1)
-    valid_mask = range_tensor < chunk_lengths.unsqueeze(1)
-    gathered = kv[indices.view(-1)].view(num_filtered_chunk, max_chunk_len, *kv.shape[1:])
-    gathered = gathered * valid_mask.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).type_as(gathered)
-
-    # Compute key_gate_weight over valid tokens.
-    key_values = gathered[:, :, 0].float()  # [num_filtered_chunk, max_chunk_len, num_head, head_dim]
-    valid_mask_exp = valid_mask.unsqueeze(-1).unsqueeze(-1)
-    key_sum = (key_values * valid_mask_exp).sum(dim=1)
-    divisor = valid_mask.sum(dim=1).unsqueeze(-1).unsqueeze(-1)
-    key_gate_weight = key_sum / divisor  # [num_filtered_chunk, num_head, head_dim]
-
-    # Compute gate logits between key_gate_weight and queries.
-    q_float = q.float()
-    # gate = torch.einsum("nhd,shd->nhs", key_gate_weight, q_float)  # [num_filtered_chunk, num_head, seqlen]
-    gate = torch.bmm(key_gate_weight.permute(1, 0, 2), q_float.permute(1, 0, 2).transpose(1, 2)).permute(1, 0, 2)
-
-    # Amplify the diagonal (self chunk) contributions.
-    gate_seq_idx = torch.arange(seqlen, device=q.device, dtype=torch.int32).unsqueeze(0).expand(num_filtered_chunk, seqlen)
-    chunk_start = cu_chunk[filtered_chunk_indices]  # [num_filtered_chunk]
-    chunk_end = cu_chunk[filtered_chunk_indices + 1]  # [num_filtered_chunk]
-    gate_self_chunk_mask = ((gate_seq_idx >= chunk_start.unsqueeze(1)) &
-                            (gate_seq_idx < chunk_end.unsqueeze(1))).unsqueeze(1).expand(-1, num_head, -1)
-    amplification_factor = 1e9  # Example factor; adjust as needed.
-    origin_gate = gate.clone()
-    gate = gate.clone()
-    if select_mode == "topk":
-        gate[gate_self_chunk_mask] += amplification_factor
-
-    # Exclude positions that are outside the valid batch boundaries.
-    batch_starts = cu_seqlens[chunk_to_batch[filtered_chunk_indices]]
-    batch_ends = cu_seqlens[chunk_to_batch[filtered_chunk_indices] + 1]
-    gate_batch_start_mask = gate_seq_idx < batch_starts.unsqueeze(1)
-    gate_batch_end_mask = gate_seq_idx >= batch_ends.unsqueeze(1)
-    gate_inf_mask = gate_batch_start_mask | gate_batch_end_mask
-    gate.masked_fill_(gate_inf_mask.unsqueeze(1), -float("inf"))
-
-    if select_mode == 'topk':
-        # We amplify self‐chunk in gate already, so self entries will rank highest.
-        valid_gate_mask = gate != -float("inf")
-        if threshold_type == 'query_head':
-            # === per‐<head,seq> top-k across chunks (original behavior) ===
-            # gate: (C, H, S)
-            _, gate_topk_idx = torch.topk(gate, k=moba_topk, dim=0, largest=True, sorted=False)
-            gate_idx_mask = torch.zeros_like(gate, dtype=torch.bool)
-            gate_idx_mask.scatter_(0, gate_topk_idx, True)
-            gate_mask = valid_gate_mask & gate_idx_mask
-        elif threshold_type == 'overall':
-            # === global top-k across all (chunk, head, seq) entries ===
-            C, H, S = gate.shape
-            flat_gate = gate.flatten()
-            flat_mask = valid_gate_mask.flatten()
-            flat_gate_masked = torch.where(flat_mask, flat_gate, -float("inf"))
-            # pick topk global entries
-            vals, idx = torch.topk(flat_gate_masked, k=moba_topk * H * S, largest=True, sorted=False)
-            others_mask_flat = torch.zeros_like(flat_mask, dtype=torch.bool)
-            others_mask_flat[idx] = True
-            gate_mask = (valid_gate_mask.flatten() & others_mask_flat).view(gate.shape)
-        elif threshold_type == 'head_global':
-            # per-head top-k across all chunks and sequence positions
-            C, H, S = gate.shape
-            CS = C * S
-            flat_gate  = gate.permute(1, 0, 2).reshape(H, CS)
-            flat_valid = valid_gate_mask.permute(1, 0, 2).reshape(H, CS)
-            flat_gate_masked = torch.where(flat_valid, flat_gate, torch.full_like(flat_gate, -float('inf')))
-            # pick top-k indices per head
-            _, topk_idx = torch.topk(flat_gate_masked, k=moba_topk * S, dim=1, largest=True, sorted=False)
-            gate_idx_flat = torch.zeros_like(flat_valid, dtype=torch.bool)
-            gate_idx_flat.scatter_(1, topk_idx, True)
-            gate_mask = gate_idx_flat.reshape(H, C, S).permute(1, 0, 2)
-        else:
-            raise ValueError(
-                f"Invalid threshold_type for topk: {threshold_type}. "
-                "Choose 'query_head', 'block', or 'overall'."
-            )
-    elif select_mode == 'threshold':
-        # Delegate to the specific thresholding function
-        valid_gate_mask = gate != -float("inf")  # (num_chunk, num_head, seqlen)
-        if threshold_type == 'query_head':
-            gate_mask = _select_threshold_query_head(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
-        elif threshold_type == 'block':
-            gate_mask = _select_threshold_block(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
-        elif threshold_type == 'overall':
-            gate_mask = _select_threshold_overall(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
-        elif threshold_type == 'head_global':
-            gate_mask = _select_threshold_head_global(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
-        else:
-            raise ValueError(f"Invalid threshold_type: {threshold_type}. Choose 'query_head', 'block', or 'overall'.")
-    else:
-        raise ValueError(f"Invalid select_mode: {select_mode}. Choose 'topk' or 'threshold'.")
-    
-    # eliminate self_chunk in MoBA branch
-    gate_mask = gate_mask & ~gate_self_chunk_mask
-    # if gate_mask is all false, perform flash_attn instead
-    if gate_mask.sum() == 0:
-        return flash_attn_varlen_func(
-            q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen, causal=False
-        )
-
-    # Determine which query positions are selected.
-    # nonzero_indices has shape [N, 3] where each row is [chunk_index, head_index, seq_index].
-    moba_q_indices = gate_mask.reshape(gate_mask.shape[0], -1).nonzero(as_tuple=True)[-1] # [(h s k)]
-    moba_q_sh_indices = (moba_q_indices % seqlen) * num_head + (moba_q_indices // seqlen)
-    moba_q = rearrange(q, "s h d -> (h s) d").index_select(0, moba_q_indices).unsqueeze(1)
-
-    # Build cumulative sequence lengths for the selected queries.
-    moba_seqlen_q = gate_mask.sum(dim=-1).flatten()
-    q_zero_mask = moba_seqlen_q == 0
-    valid_expert_mask = ~q_zero_mask
-    if q_zero_mask.sum() > 0:
-        moba_seqlen_q = moba_seqlen_q[valid_expert_mask]
-    moba_cu_seqlen_q = torch.cat(
-        (
-            torch.tensor([0], device=q.device, dtype=moba_seqlen_q.dtype),
-            moba_seqlen_q.cumsum(dim=0),
-        ),
-        dim=0,
-    ).to(torch.int32)
-
-    # Rearrange gathered KV for the MOBA branch.
-    experts_tensor = rearrange(gathered, "nc cl two h d -> (nc h) cl two d")
-    valid_expert_lengths = chunk_lengths.unsqueeze(1).expand(num_filtered_chunk, num_head).reshape(-1).to(torch.int32)
-    if q_zero_mask.sum() > 0:
-        experts_tensor = experts_tensor[valid_expert_mask]
-        valid_expert_lengths = valid_expert_lengths[valid_expert_mask]
-
-    seq_range = torch.arange(experts_tensor.shape[1], device=experts_tensor.device).unsqueeze(0)
-    mask = seq_range < valid_expert_lengths.unsqueeze(1)
-    moba_kv = experts_tensor[mask]  # Shape: ((nc h cl_valid) two d)
-    moba_kv = moba_kv.unsqueeze(2)   # Shape: ((nc h cl_valid) two 1 d)
-
-    moba_cu_seqlen_kv = torch.cat(
-        [torch.zeros(1, device=experts_tensor.device, dtype=torch.int32),
-         valid_expert_lengths.cumsum(dim=0)],
-        dim=0,
-    ).to(torch.int32)
-
-    assert (
-        moba_cu_seqlen_kv.shape == moba_cu_seqlen_q.shape
-    ), f"Mismatch between moba_cu_seqlen_kv.shape and moba_cu_seqlen_q.shape: {moba_cu_seqlen_kv.shape} vs {moba_cu_seqlen_q.shape}"
-
-    return MixedAttention.apply(
-        q,
-        k,
-        v,
-        self_attn_cu_seqlen,
-        moba_q,
-        moba_kv,
-        moba_cu_seqlen_q,
-        moba_cu_seqlen_kv,
-        max_seqlen,
-        moba_chunk_size,
-        moba_q_sh_indices,
-    )
-
-
-def process_moba_input(
-    x,
-    patch_resolution,
-    chunk_size,
-):
-    """
-    Process inputs for the attention function.
-
-    Args:
-        x (torch.Tensor): Input tensor with shape [batch_size, num_patches, num_heads, head_dim].
-        patch_resolution (tuple): Tuple containing the patch resolution (t, h, w).
-        chunk_size (int): Size of the chunk. (maybe tuple or int, according to chunk type)
-
-    Returns:
-        torch.Tensor: Processed input tensor.
-    """
-    if isinstance(chunk_size, float) or isinstance(chunk_size, int):
-        moba_chunk_size = int(chunk_size * patch_resolution[1] * patch_resolution[2])
-    else:
-        assert isinstance(chunk_size, (Tuple, list)), f"chunk_size should be a tuple, list, or int, now it is: {type(chunk_size)}"
-        if len(chunk_size) == 2:
-            assert patch_resolution[1] % chunk_size[0] == 0 and patch_resolution[2] % chunk_size[1] == 0, f"spatial patch_resolution {patch_resolution[1:]} should be divisible by 2d chunk_size {chunk_size}"
-            nch, ncw = patch_resolution[1] // chunk_size[0], patch_resolution[2] // chunk_size[1]
-            x = rearrange(x, "b (t nch ch ncw cw) n d -> b (nch ncw t ch cw) n d", t=patch_resolution[0], nch=nch, ncw=ncw, ch=chunk_size[0], cw=chunk_size[1])
-            moba_chunk_size = patch_resolution[0] * chunk_size[0] * chunk_size[1]
-        elif len(chunk_size) == 3:
-            assert patch_resolution[0] % chunk_size[0] == 0 and patch_resolution[1] % chunk_size[1] == 0 and patch_resolution[2] % chunk_size[2] == 0, f"patch_resolution {patch_resolution} should be divisible by 3d chunk_size {chunk_size}"
-            nct, nch, ncw = patch_resolution[0] // chunk_size[0], patch_resolution[1] // chunk_size[1], patch_resolution[2] // chunk_size[2]
-            x = rearrange(x, "b (nct ct nch ch ncw cw) n d -> b (nct nch ncw ct ch cw) n d", nct=nct, nch=nch, ncw=ncw, ct=chunk_size[0], ch=chunk_size[1], cw=chunk_size[2])
-            moba_chunk_size = chunk_size[0] * chunk_size[1] * chunk_size[2]
-        else:
-            raise ValueError(f"chunk_size should be a int, or a tuple of length 2 or 3, now it is: {len(chunk_size)}")
-        
-    return x, moba_chunk_size
-
-
-def process_moba_output(
-    x,
-    patch_resolution,
-    chunk_size,
-):
-    if isinstance(chunk_size, float) or isinstance(chunk_size, int):
-        pass
-    elif len(chunk_size) == 2:
-        x = rearrange(x, "b (nch ncw t ch cw) n d -> b (t nch ch ncw cw) n d", nch=patch_resolution[1] // chunk_size[0], ncw=patch_resolution[2] // chunk_size[1], t=patch_resolution[0], ch=chunk_size[0], cw=chunk_size[1])
-    elif len(chunk_size) == 3:
-        x = rearrange(x, "b (nct nch ncw ct ch cw) n d -> b (nct ct nch ch ncw cw) n d", nct=patch_resolution[0] // chunk_size[0], nch=patch_resolution[1] // chunk_size[1], ncw=patch_resolution[2] // chunk_size[2], ct=chunk_size[0], ch=chunk_size[1], cw=chunk_size[2])
-
-    return x
-
-
-# TEST
-def generate_data(batch_size, seqlen, num_head, head_dim, dtype):
-    random.seed(0)
-    torch.manual_seed(0)
-    torch.cuda.manual_seed(0)
-    device = torch.cuda.current_device()
-
-    q = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
-    k = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
-    v = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
-    print(f"q.shape: {q.shape}, k.shape: {k.shape}, v.shape: {v.shape}")
-    cu_seqlens = torch.arange(0, q.shape[0] * q.shape[1] + 1, q.shape[1], dtype=torch.int32, device='cuda')
-    max_seqlen = q.shape[1]
-    q = rearrange(q, "b s ... -> (b s) ...")
-    k = rearrange(k, "b s ... -> (b s) ...")
-    v = rearrange(v, "b s ... -> (b s) ...")
-
-    return q, k, v, cu_seqlens, max_seqlen
-
-
-def test_attn_varlen_moba_speed(batch, head, seqlen, head_dim, moba_chunk_size, moba_topk, dtype=torch.bfloat16, select_mode='threshold', simsum_threshold=0.25, threshold_type='query_head'):
-    """Speed test comparing flash_attn vs moba_attention"""
-    # Get data
-    q, k, v, cu_seqlen, max_seqlen = generate_data(batch, seqlen, head, head_dim, dtype)
-    print(f"batch:{batch} head:{head} seqlen:{seqlen} chunk:{moba_chunk_size} topk:{moba_topk} select_mode: {select_mode} simsum_threshold:{simsum_threshold}")
-    vo_grad = torch.randn_like(q)
-    
-    # Warmup
-    warmup_iters = 3
-    perf_test_iters = 10
-
-    # Warmup
-    for _ in range(warmup_iters):
-        o = flash_attn_varlen_func(q, k, v, cu_seqlen, cu_seqlen, max_seqlen, max_seqlen, causal=False)
-        torch.autograd.backward(o, vo_grad)
-    
-    torch.cuda.synchronize()
-    start_flash = time.perf_counter()
-    for _ in range(perf_test_iters):
-        o = flash_attn_varlen_func(q, k, v, cu_seqlen, cu_seqlen, max_seqlen, max_seqlen, causal=False)
-        torch.autograd.backward(o, vo_grad)
-        
-    torch.cuda.synchronize()
-    time_flash = (time.perf_counter() - start_flash) / perf_test_iters * 1000
-
-    # Warmup
-    for _ in range(warmup_iters):
-        om = moba_attn_varlen(q, k, v, cu_seqlen, max_seqlen, moba_chunk_size=moba_chunk_size, moba_topk=moba_topk, select_mode=select_mode, simsum_threshold=simsum_threshold, threshold_type=threshold_type)
-        torch.autograd.backward(om, vo_grad)
-
-    torch.cuda.synchronize()
-    start_moba = time.perf_counter()
-    for _ in range(perf_test_iters):
-        om = moba_attn_varlen(q, k, v, cu_seqlen, max_seqlen, moba_chunk_size=moba_chunk_size, moba_topk=moba_topk, select_mode=select_mode, simsum_threshold=simsum_threshold, threshold_type=threshold_type)
-        torch.autograd.backward(om, vo_grad)
-    
-    torch.cuda.synchronize()
-    time_moba = (time.perf_counter() - start_moba) / perf_test_iters * 1000
-    
-    print(f"Flash: {time_flash:.2f}ms, MoBA: {time_moba:.2f}ms")
-    print(f"Speedup:  {time_flash / time_moba:.2f}x")
-
-
-
-if __name__ == "__main__":
-    """
-    CUDA_VISIBLE_DEVICES=1 \
-    python -u csrc/attn/vmoba_attn/vmoba/vmoba.py
-    """
-    test_attn_varlen_moba_speed(batch=1, head=12, seqlen=32760, head_dim=128, moba_chunk_size=32760 // 3 // 6 // 4, moba_topk=3, select_mode='threshold', simsum_threshold=0.3, threshold_type='query_head')
@@ -568,6 +568,7 @@ void bwd_attend_ker(const __grid_constant__ bwd_globals<D> g) {
        __syncthreads(); // wait for sd_smem shared memory write
        warpgroup::mm_AtB(qg_reg, ds_smem_t[0], k_smem[0]); //delat dQ = dSK
        warpgroup::mma_commit_group();
+        tma::store_async_wait();
        warpgroup::mma_async_wait();
        // store qg to shared memory
        warpgroup::store(qg_smem, qg_reg);
@@ -577,7 +578,6 @@ void bwd_attend_ker(const __grid_constant__ bwd_globals<D> g) {
        if (threadIdx.x / 32 == 0) {
            coord<qg_tile> tile_idx = {blockIdx.z, blockIdx.y, store_qg_block_index, 0};
            tma::store_add_async(g.qg, qg_smem, tile_idx);
-            tma::store_async_wait();
        }
    }

@@ -624,6 +624,7 @@ void bwd_attend_ker(const __grid_constant__ bwd_globals<D> g) {
        __syncthreads(); // wait for sd_smem shared memory write
        warpgroup::mm_AtB(qg_reg, ds_smem_t[0], k_smem[0]); //delat dQ = dSK
        warpgroup::mma_commit_group();
+        tma::store_async_wait();
        warpgroup::mma_async_wait();
        // store qg to shared memory
        warpgroup::store(qg_smem, qg_reg);
@@ -633,14 +634,13 @@ void bwd_attend_ker(const __grid_constant__ bwd_globals<D> g) {
        if (threadIdx.x / 32 == 0) {
            coord<qg_tile> tile_idx = {blockIdx.z, blockIdx.y, store_qg_block_index, 0};
            tma::store_add_async(g.qg, qg_smem, tile_idx);
-            tma::store_async_wait();
        }
    }

    // store kq and vq

    // ! the following two line seems unnecessary.
-    // tma::store_async_wait(); // ensure qg is finished
+    tma::store_async_wait(); // ensure qg is finished
    __syncthreads();

    warpgroup::store(kg_smem[0], kg_reg);
@@ -1174,4 +1174,4 @@ block_sparse_attention_backward(torch::Tensor q,

    return {qg, kg, vg};
    //cudadevicesynchronize();
-}
+}
@@ -1,9 +1,7 @@
-FROM nvidia/cuda:12.8.0-devel-ubuntu22.04
+FROM nvidia/cuda:12.4.1-devel-ubuntu20.04

 ENV DEBIAN_FRONTEND=noninteractive

-SHELL ["/bin/bash", "-c"]
-
 WORKDIR /FastVideo

 RUN apt-get update && apt-get install -y --no-install-recommends \
@@ -11,25 +9,17 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
    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
+RUN wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
+    bash Miniconda3-latest-Linux-x86_64.sh -b -p /opt/conda && \
+    rm Miniconda3-latest-Linux-x86_64.sh

-# 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
+ENV PATH=/opt/conda/bin:$PATH

-# Install uv and source its environment
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh && \
-    echo 'source $HOME/.local/bin/env' >> /root/.bashrc
+RUN conda create --name fastvideo-dev python=3.10.0 -y
+
+SHELL ["/bin/bash", "-c"]

 # Copy just the pyproject.toml first to leverage Docker cache
 COPY pyproject.toml ./
@@ -37,36 +27,22 @@ COPY 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 && \
-    uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
+RUN conda run -n fastvideo-dev pip install --no-cache-dir --upgrade pip && \
+    conda run -n fastvideo-dev pip install --no-cache-dir .[dev] && \
+    conda run -n fastvideo-dev pip install --no-cache-dir flash-attn==2.7.4.post1 --no-build-isolation && \
+    conda clean -afy

 COPY . .

-# Install dependencies using uv and set up shell configuration
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    uv pip install --no-cache-dir -e .[dev] && \
-    git config --unset-all http.https://github.com/.extraheader || true && \
-    echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
-    echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
+RUN conda run -n fastvideo-dev pip install --no-cache-dir -e .[dev]

-# Install STA (Sliding Tile Attention)
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
+# Remove authentication headers
+RUN git config --unset-all http.https://github.com/.extraheader || true

-# 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
+# Set up automatic conda environment activation for all shells
+RUN echo 'source /opt/conda/etc/profile.d/conda.sh' >> /root/.bashrc && \
+    echo 'conda activate fastvideo-dev' >> /root/.bashrc && \
+    # Ensure .bashrc is sourced for SSH login shells
+    echo 'if [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile

-EXPOSE 22
+EXPOSE 22
@@ -1,9 +1,7 @@
-FROM nvidia/cuda:12.8.0-devel-ubuntu22.04
+FROM nvidia/cuda:12.4.1-devel-ubuntu20.04

 ENV DEBIAN_FRONTEND=noninteractive

-SHELL ["/bin/bash", "-c"]
-
 WORKDIR /FastVideo

 RUN apt-get update && apt-get install -y --no-install-recommends \
@@ -11,25 +9,17 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
    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
+RUN wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
+    bash Miniconda3-latest-Linux-x86_64.sh -b -p /opt/conda && \
+    rm Miniconda3-latest-Linux-x86_64.sh

-# 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
+ENV PATH=/opt/conda/bin:$PATH

-# Install uv and source its environment
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh && \
-    echo 'source $HOME/.local/bin/env' >> /root/.bashrc
+RUN conda create --name fastvideo-dev python=3.11.11 -y
+
+SHELL ["/bin/bash", "-c"]

 # Copy just the pyproject.toml first to leverage Docker cache
 COPY pyproject.toml ./
@@ -37,36 +27,22 @@ COPY 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 3.11 --seed /opt/venv && \
-    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
+RUN conda run -n fastvideo-dev pip install --no-cache-dir --upgrade pip && \
+    conda run -n fastvideo-dev pip install --no-cache-dir .[dev] && \
+    conda run -n fastvideo-dev pip install --no-cache-dir flash-attn==2.7.4.post1 --no-build-isolation && \
+    conda clean -afy

 COPY . .

-# Install dependencies using uv and set up shell configuration
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    uv pip install --no-cache-dir -e .[dev] && \
-    git config --unset-all http.https://github.com/.extraheader || true && \
-    echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
-    echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
+RUN conda run -n fastvideo-dev pip install --no-cache-dir -e .[dev]

-# Install STA (Sliding Tile Attention)
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
+# Remove authentication headers
+RUN git config --unset-all http.https://github.com/.extraheader || true

-# 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
+# Set up automatic conda environment activation for all shells
+RUN echo 'source /opt/conda/etc/profile.d/conda.sh' >> /root/.bashrc && \
+    echo 'conda activate fastvideo-dev' >> /root/.bashrc && \
+    # Ensure .bashrc is sourced for SSH login shells
+    echo 'if [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile

-EXPOSE 22
+EXPOSE 22
@@ -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 flash-attn==2.8.0.post2 --no-build-isolation

 COPY . .

@@ -58,15 +58,15 @@ RUN source $HOME/.local/bin/env && \
 # Install STA (Sliding Tile Attention)
 RUN source $HOME/.local/bin/env && \
    source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
+    cd csrc/attn && \
    git submodule update --init --recursive && \
-    python setup.py install
+    python setup_sta.py install

 # Install VSA 
 RUN source $HOME/.local/bin/env && \
    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
+    cd csrc/attn && \
    git submodule update --init --recursive && \
-    python setup.py install
+    python setup_vsa.py install

-EXPOSE 22
+EXPOSE 22
@@ -1,72 +0,0 @@
-FROM nvidia/cuda:12.9.1-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-12.9
-ENV PATH=${CUDA_HOME}/bin:${PATH}
-ENV LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
-
-# 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.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 3.12 --seed /opt/venv && \
-    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
-
-COPY . .
-
-# Install dependencies using uv and set up shell configuration
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    uv pip install --no-cache-dir -e .[dev] && \
-    git config --unset-all http.https://github.com/.extraheader || true && \
-    echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
-    echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
-
-# Install STA (Sliding Tile Attention)
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/sliding_tile_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
-
-# Install VSA 
-RUN source $HOME/.local/bin/env && \
-    source /opt/venv/bin/activate && \
-    cd csrc/attn/video_sparse_attn && \
-    git submodule update --init --recursive && \
-    python setup.py install
-
-EXPOSE 22
@@ -96,7 +96,8 @@ copybutton_prompt_is_regexp = True
 #
 html_title = project
 html_theme = 'sphinx_book_theme'
-html_logo = '../../assets/logos/icon_simple.svg'
+html_logo = '../../assets/logo.jpg'
+#html_favicon = 'assets/logos/vllm-logo-only-light.ico'
 html_theme_options = {
    'path_to_docs': 'docs/source',
    'repository_url': 'https://github.com/hao-ai-lab/FastVideo/',
@@ -3,12 +3,12 @@

 If you prefer a containerized development environment or want to avoid managing dependencies manually, you can use our prebuilt Docker image:

-**Images:** [`ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:py3.12-latest`](https://ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev)
+**Image:** [`ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:latest`](https://ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev)

 ## Starting the container

 ```bash
-docker run --gpus all -it ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:py3.12-latest
+docker run --gpus all -it ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:latest
 ```

 This will:
@@ -6,7 +6,7 @@ You can easily use the FastVideo Docker image as a custom container on [RunPod](

 ## Creating a new pod

-Choose a GPU that supports CUDA 12.8
+Choose a GPU that supports CUDA 12.4

 Pick 1 or 2 L40S GPU(s)

@@ -22,20 +22,10 @@ source ~/.bashrc
 Create and activate a Conda environment for FastVideo:

 ```
-conda create -n fastvideo python=3.12 -y
+conda create -n fastvideo python=3.10 -y
 conda activate fastvideo
 ```

-Install `uv` (optional, but recommended):
-
-From instructions on [uv](https://astral.sh/uv/):
-
-```
-curl -LsSf https://astral.sh/uv/install.sh | sh
-# or 
-wget -qO- https://astral.sh/uv/install.sh | sh
-```
-
 Clone the FastVideo repository and go to the FastVideo directory:

 ```
@@ -46,10 +36,10 @@ git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
 Now you can install FastVideo and setup git hooks for running linting. By using `pre-commit`, the linters will run and have to pass before you'll be able to make a commit.

 ```bash
-uv pip install -e .[dev]
+pip install -e .[dev]

 # Can also install flash-attn (optional)
-uv pip install flash-attn --no-build-isolation 
+pip install flash-attn==2.7.4.post1 --no-build-isolation 

 # Linting, formatting and static type checking
 pre-commit install --hook-type pre-commit --hook-type commit-msg
@@ -60,14 +50,3 @@ pre-commit run --all-files
 # Unit tests
 pytest tests/
 ```
-
-If you are on a Hopper GPU, you should also install [FA3](https://github.com/Dao-AILab/flash-attention) for much better performance:
-
-```
-git clone https://github.com/Dao-AILab/flash-attention.git && cd flash-attention/hopper
-
-# make sure you have ninja installed
-uv pip install ninja
-
-python setup.py install
-```
@@ -6,7 +6,7 @@ Instructions to install FastVideo for NVIDIA CUDA GPUs.

 - **OS: Linux or Windows WSL**
 - **Python: 3.10-3.12**
- **CUDA 12.8**
+- **CUDA 12.4**
 - **At least 1 NVIDIA GPU**

 ## Set up using Python
@@ -38,7 +38,6 @@ conda activate fastvideo

 :::{tip}
 We highly recommend using `uv` to install FastVideo. In our experience, `uv` speeds up installation by at least 3x.
-Note that you can also use `uv` to install FastVideo in a Conda environment.
 :::

 Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
@@ -61,7 +60,7 @@ uv pip install fastvideo
 Also optionally install flash-attn:

 ```bash
-pip install flash-attn --no-build-isolation
+pip install flash-attn==2.7.4.post1 --no-build-isolation
 ```

 ### Installation from Source
@@ -88,7 +87,7 @@ uv pip install -e .
 #### Flash Attention

 ```bash
-pip install flash-attn --no-build-isolation
+pip install flash-attn==2.7.4.post1 --no-build-isolation
 ```

 ## Set up using Docker
@@ -103,7 +102,7 @@ If you're planning to contribute to FastVideo please see the following page:
 ## Hardware Requirements

 ### For Basic Inference
- NVIDIA GPU with CUDA 12.8 support
+- NVIDIA GPU with CUDA 12.4 support

 ### For Lora Finetuning
 - 40GB GPU memory each for 2 GPUs with lora
@@ -39,7 +39,6 @@ conda activate fastvideo

 :::{tip}
 We highly recommend using `uv` to install FastVideo. In our experience, `uv` speeds up installation by at least 3x.
-Note that you can also use `uv` to install FastVideo in a Conda environment.
 :::

 Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
@@ -1,6 +1,6 @@
 # Welcome to FastVideo

-:::{figure} ../../assets/logos/logo.svg
+:::{figure} ../../assets/logo.png
 :align: center
 :alt: FastVideo
 :class: no-scaled-link
@@ -101,6 +101,7 @@ sliding_tile_attention/demo
 :maxdepth: 1

 video_sparse_attention/installation
+video_sparse_attention/demo
 :::

 :::{toctree}
@@ -5,7 +5,7 @@ This page contains step-by-step instructions to get you quickly started with vid
 ## Requirements
 - **OS**: Linux (Tested on Ubuntu 22.04+)
 - **Python**: 3.10-3.12
- **CUDA**: 12.8
+- **CUDA**: 12.4
 - **GPU**: At least one NVIDIA GPU

 ## Installation
@@ -6,7 +6,6 @@ The symbols used have the following meanings:

 - ✅ = Full compatibility
 - ❌ = No compatibility
- ⭕ = Does not apply to this model

 ## Models x Optimization
 The `HuggingFace Model ID` can be directly pass to `from_pretrained()` methods and FastVideo will use the optimal default parameters when initializing and generating videos.
@@ -38,94 +37,51 @@ The `HuggingFace Model ID` can be directly pass to `from_pretrained()` methods a
  * TeaCache
  * Sliding Tile Attn
  * Sage Attn
-  * Video Sparse Attention (VSA)
- * FastWan2.1 T2V 1.3B
-  * `FastVideo/FastWan2.1-T2V-1.3B-Diffusers`
-  * 480P
-  * ⭕
-  * ⭕
-  * ⭕
-  * ✅
- * FastWan2.2 TI2V 5B Full Attn*
-  * `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers`
-  * 720P
-  * ⭕
-  * ⭕
-  * ⭕
-  * ✅
- * Wan2.2 TI2V 5B 
-  * `Wan-AI/Wan2.2-TI2V-5B-Diffusers`
-  * 720P
-  * ⭕
-  * ⭕
-  * ✅
-  * ⭕
- * Wan2.2 T2V A14B
-  * `Wan-AI/Wan2.2-T2V-A14B-Diffusers`
-  * 480P<br>720P
-  * ❌
-  * ❌
-  * ✅
-  * ⭕
- * Wan2.2 I2V A14B
-  * `Wan-AI/Wan2.2-I2V-A14B-Diffusers`
-  * 480P<br>720P
-  * ❌
-  * ❌
-  * ✅
-  * ⭕
 - * HunyuanVideo
  * `hunyuanvideo-community/HunyuanVideo`
  * 720px1280p<br>544px960p
  * ❌
  * ✅
  * ✅
-  * ⭕
 - * FastHunyuan
  * `FastVideo/FastHunyuan-diffusers`
  * 720px1280p<br>544px960p
  * ❌
  * ✅
  * ✅
-  * ⭕
- * Wan2.1 T2V 1.3B
+- * Wan T2V 1.3B
  * `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`
  * 480P
  * ✅
  * ✅*
  * ✅
-  * ⭕
- * Wan2.1 T2V 14B
+- * Wan T2V 14B
  * `Wan-AI/Wan2.1-T2V-14B-Diffusers`
  * 480P, 720P
  * ✅
  * ✅*
  * ✅
-  * ⭕
- * Wan2.1 I2V 480P
+- * Wan I2V 480P
  * `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers`
  * 480P
  * ✅
  * ✅*
  * ✅
-  * ⭕
- * Wan2.1 I2V 720P
+- * Wan I2V 720P
  * `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers`
  * 720P
  * ✅
+  * ✅*
  * ✅
-  * ✅
-  * ⭕
 - * StepVideo T2V
  * `FastVideo/stepvideo-t2v-diffusers`
  * 768px768px204f<br>544px992px204f<br>544px992px136f
  * ❌
  * ❌
  * ✅
-  * ⭕
 :::

-**Note**: Wan2.2 TI2V 5B has some quality issues when performing I2V generation. We are working on fixing this issue.
+**Note**: there are some known quality issues with Wan2.1 + Sliding Tile Attn. We are working on fixing this issue.

 ## Special requirements

@@ -4,7 +4,7 @@
 You can install the Sliding Tile Attention package using

 ```
-pip install st_attn
+pip install st_attn==0.0.4
 ``` 

 # Building from Source
@@ -12,6 +12,7 @@ We test our code on Pytorch 2.5.0 and CUDA>=12.4. Currently we only have impleme
 First, install C++20 for ThunderKittens:

 ```bash
+cd csrc/sliding_tile_attention/
 sudo apt update
 sudo apt install gcc-11 g++-11

@@ -21,20 +22,14 @@ sudo apt update
 sudo apt install clang-11
 ```

-Set up CUDA environment (if using CUDA 12.4):
+Install STA:

 ```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
+python setup_sta.py install
 ```

 # 🧪 Test
@@ -0,0 +1,3 @@
+(vsa-demo)=
+
+# 🎬 Demo
@@ -4,7 +4,8 @@
 You can install the Video Sparse Attention package using

 ```bash
-pip install vsa
+git submodule update --init --recursive
+python setup_vsa.py install
 ```

 # Building from Source
@@ -22,10 +23,10 @@ sudo apt update
 sudo apt install clang-11
 ```

-Set up CUDA environment (if using CUDA 12.8):
+Set up CUDA environment (if using CUDA 12.4):

 ```bash
-export CUDA_HOME=/usr/local/cuda-12.8
+export CUDA_HOME=/usr/local/cuda-12.4
 export PATH=${CUDA_HOME}/bin:${PATH} 
 export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
 ```
@@ -33,9 +34,9 @@ export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
 Install VSA:

 ```bash
-cd csrc/attn/video_sparse_attn/
+cd csrc/attn/
 git submodule update --init --recursive
-python setup.py install
+python setup_vsa.py install
 ```

 # 🧪 Test
@@ -1,9 +0,0 @@
-# VidProm Dataset
-
-From [Self-Forcing](https://github.com/gdhe17/Self-Forcing) repository.
-
-## Download the dataset
-
-```bash
-./download_dataset.sh
-```
@@ -1,3 +0,0 @@
-#! /bin/bash
-
-huggingface-cli download gdhe17/Self-Forcing vidprom_filtered_extended.txt --local-dir prompts
@@ -1,151 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=t2v
-#SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:1
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=dmd_t2v_output/t2v_%j.out
-#SBATCH --error=dmd_t2v_output/t2v_%j.err
-#SBATCH --exclusive
-
-# Basic Info
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-# different cache dir for different processes
-export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29501
-export TOKENIZERS_PARALLELISM=false
-export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=offline
-export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
-
-# Configs
-NUM_GPUS=4
-
-# Model paths for Self-Forcing DMD distillation:
-GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
-FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
-
-DATA_DIR="data/mixkit-64_processed/Node_0_GPU_1_File_1/combined_parquet_dataset"
-VALIDATION_DATASET_FILE="data/mixkit-64_processed/validation.json"
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
-  --output_dir "/mnt/sharefs/users/hao.zhang/SFwan_t2v_finetune"
-  --max_train_steps 4000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 81  # Must be divisible by num_frame_per_block (81 % 3 = 0 ✓)
-  --enable_gradient_checkpointing_type "full"
-  --log_visualization
-  --simulate_generator_forward
-  --num_frame_per_block 3  # Frame generation block size for self-forcing
-  --enable_gradient_masking
-  --gradient_mask_last_n_frames 21
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS # 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim $NUM_GPUS
-)
-
-# Model arguments
-model_args=(
-  --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
-  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
-  --generator_model_path $GENERATOR_MODEL_PATH
-  --real_score_model_path $REAL_SCORE_MODEL_PATH
-  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR"
-  --dataloader_num_workers 4
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 100
-  --validation_sampling_steps "4"
-  --validation_guidance_scale "6.0" # not used for dmd inference
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --training_state_checkpointing_steps 500
-  --weight_only_checkpointing_steps 500
-  --weight_decay 0.01
-  --betas '0.0,0.999'
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.0
-  --dit_precision "fp32"
-  --flow_shift 5
-  --seed 1000
-  --use_ema True
-  --ema_decay 0.99
-  --ema_start_step 100
-  --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
-)
-
-# Self-forcing DMD arguments
-dmd_args=(
-  --dmd_denoising_steps '1000,750,500,250'
-  --min_timestep_ratio 0.02
-  --max_timestep_ratio 0.98
-  --dfake_gen_update_ratio 5
-  --real_score_guidance_scale 3.0
-  --fake_score_learning_rate 8e-6
-  --fake_score_betas '0.0,0.999'
-  --warp_denoising_step
-)
-
-# Self-forcing specific arguments
-self_forcing_args=(
-  --independent_first_frame False  # Whether to treat first frame independently
-  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
-  --last_step_only False  # Whether to only use the last denoising step
-  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
-  --validate_cache_structure False  # Set to True for debugging KV cache issues
-)
-
-torchrun \
--nnodes 1 \
--master_port $MASTER_PORT \
--nproc_per_node $NUM_GPUS \
-    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}" \
-    "${self_forcing_args[@]}"
@@ -0,0 +1,16 @@
+# Wan2.1-I2V-1.3B-InP Crush-Smol Example
+These are e2e example scripts for finetuning Wan2.1 T2V 1.3B InP on the crush-smol dataset.
+
+## Execute the following commands from `FastVideo/` to run training:
+
+### Download crush-smol dataset:
+
+`bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/download_dataset.sh`
+
+### Preprocess the videos and captions into latents:
+
+`bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/preprocess_wan_data_i2v.sh`
+
+### Edit the following file and run finetuning:
+
+`bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/finetune_i2v.sh`
@@ -0,0 +1,111 @@
+#!/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="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
+DATA_DIR="data/crush-smol_processed_i2v_1_3b_inp_single/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/distill/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
+NUM_GPUS=8
+export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_i2v_distill"
+  --output_dir "checkpoints/wan_i2v_distill"
+  --wandb_run_name "wan_i2v_distill"
+  --max_train_steps 2000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 2
+  --num_latent_t 8
+  --num_height 480
+  --num_width 832
+  --num_frames 77
+  # --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim 8
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 1
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  # --log_visualization
+  --log_visualization
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "4"
+  --validation_guidance_scale "1.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 4e-6
+  --lr_scheduler "constant"
+  --fake_score_learning_rate 8e-7
+  --fake_score_lr_scheduler "constant"
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 2000
+  --weight_only_checkpointing_steps 2000
+  --weight_decay 0.01
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --ema_start_step 0
+  --flow_shift 8
+  --seed 1000
+  # --enable_gradient_checkpointing_type "full"
+)
+
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --generator_update_interval 5
+  --simulate_generator_forward
+  --real_score_guidance_scale 5
+  --VSA_sparsity 0.8
+)
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_i2v_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}"
@@ -4,28 +4,26 @@ export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
 export TOKENIZERS_PARALLELISM=false
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'

-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/weka/home/hao.zhang/wl/FastVideo2/data/crush-smol_processed_t2v_1_3b_ode_init_single"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
-NUM_GPUS=1
+MODEL_PATH="Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+DATA_DIR="data/crush-smol_processed_i2v_1_3b_inp/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/distill/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
+NUM_GPUS=8
+export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+export HF_HUB_ENABLE_HF_TRANSFER=1
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_crush_smol"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "overfitwan_ode_init_crush_smol"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 2001
-  # --warp_denoising_step
+  --tracker_project_name "wan_i2v_distill"
+  --output_dir "checkpoints/wan_i2v_distill"
+  --wandb_run_name "wan_i2v_distill"
+  --max_train_steps 2000
  --train_batch_size 1
  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
+  --gradient_accumulation_steps 2
+  --num_latent_t 8
  --num_height 480
  --num_width 832
  --num_frames 77
@@ -38,7 +36,7 @@ parallel_args=(
  --sp_size 1
  --tp_size 1
  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 1
+  --hsdp_shard_dim 8
 )

 # Model arguments
@@ -49,25 +47,30 @@ model_args=(

 # Dataset arguments
 dataset_args=(
-  --data_path "$DATA_DIR" 
+  --data_path "$DATA_DIR"
  --dataloader_num_workers 1
 )

 # Validation arguments
 validation_args=(
  --log_validation
+  # --log_visualization
+  --log_visualization
  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 20
+  --validation_steps 50
  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
+  --validation_guidance_scale "1.0" # not used for dmd inference
 )

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 4e-6
+  --lr_scheduler "constant"
+  --fake_score_learning_rate 8e-7
+  --fake_score_lr_scheduler "constant"
  --mixed_precision "bf16"
-  --checkpointing_steps 500
-  --weight_decay 1e-4
+  --checkpointing_steps 2000
+  --weight_decay 0.01
  --max_grad_norm 1.0
 )

@@ -75,24 +78,34 @@ optimizer_args=(
 miscellaneous_args=(
  --inference_mode False
  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
+  --training_cfg_rate 0.0
  --dit_precision "fp32"
-  --num_euler_timesteps 50
  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
+  --flow_shift 8
+  --seed 1000
+  # --enable_gradient_checkpointing_type "full"
+)
+
+dmd_args=(
+  --dmd_denoising_steps '1000,757,522'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --generator_update_interval 5
+  --simulate_generator_forward
+  --real_score_guidance_scale 3.5
+  --VSA_sparsity 0.8
 )

 # If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
 torchrun \
  --nnodes 1 \
  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/ode_causal_pipeline.py \
+    fastvideo/training/wan_i2v_training_pipeline.py \
    "${parallel_args[@]}" \
    "${model_args[@]}" \
    "${dataset_args[@]}" \
    "${training_args[@]}" \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}"
@@ -1,21 +1,22 @@
 #!/bin/bash
-#SBATCH --job-name=ode_vidprom2k
+#SBATCH --job-name=i2v
 #SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
+#SBATCH --nodes=4
+#SBATCH --ntasks=4
 #SBATCH --ntasks-per-node=1
 #SBATCH --gres=gpu:8
 #SBATCH --cpus-per-task=128
 #SBATCH --mem=1440G
-#SBATCH --output=ode_vidprom2k_output/ode_vidprom2k.out
-#SBATCH --error=ode_vidprom2k_output/ode_vidprom2k.err
+#SBATCH --output=i2v_output/i2v_%j.out
+#SBATCH --error=i2v_output/i2v_%j.err
 #SBATCH --exclusive
 set -e -x

 # Environment Setup
 source ~/conda/miniconda/bin/activate
-conda activate will-fv2
+conda activate will-fv

+# Basic Info
 export WANDB_MODE="online"
 export NCCL_P2P_DISABLE=1
 export TORCH_NCCL_ENABLE_MONITORING=0
@@ -29,46 +30,44 @@ export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
 export TOKENIZERS_PARALLELISM=false
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA

 echo "MASTER_ADDR: $MASTER_ADDR"
 echo "NODE_RANK: $NODE_RANK"

-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
+# Configs
 NUM_GPUS=8
+
+MODEL_PATH="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
+DATA_DIR="data/crush-smol_processed_i2v_1_3b_inp/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+
 # Training arguments
 training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_vidprom2k"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "vidprom2k_wan_ode_init_5e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 6001
-  --train_batch_size 1
+  --tracker_project_name wan_i2v_finetune
+  --output_dir "checkpoints/wan_i2v_finetune"
+  --max_train_steps 2000
+  --train_batch_size 2
  --train_sp_batch_size 1
  --gradient_accumulation_steps 1
-  --num_latent_t 21
+  --num_latent_t 8
  --num_height 480
  --num_width 832
  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
+  --enable_gradient_checkpointing_type "full"
 )

 # Parallel arguments
 parallel_args=(
  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 8
-  --hsdp_shard_dim 1
+  --sp_size $NUM_GPUS
+  --tp_size $NUM_GPUS
+  --hsdp_replicate_dim $SLURM_JOB_NUM_NODES
+  --hsdp_shard_dim $NUM_GPUS
 )

 # Model arguments
@@ -79,8 +78,8 @@ model_args=(

 # Dataset arguments
 dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 10
 )

 # Validation arguments
@@ -88,15 +87,15 @@ validation_args=(
  --log_validation
  --validation_dataset_file "$VALIDATION_DATASET_FILE"
  --validation_steps 100
-  --validation_sampling_steps "50"
+  --validation_sampling_steps "40"
  --validation_guidance_scale "6.0"
 )

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 5e-6
+  --learning_rate 1e-5
  --mixed_precision "bf16"
-  --checkpointing_steps 2000
+  --checkpointing_steps 1000
  --weight_decay 1e-4
  --max_grad_norm 1.0
 )
@@ -114,18 +113,17 @@ miscellaneous_args=(
  --enable_gradient_checkpointing_type "full"
 )

-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
 srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_i2v_training_pipeline.py \
    "${parallel_args[@]}" \
    "${model_args[@]}" \
    "${dataset_args[@]}" \
    "${training_args[@]}" \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
+    "${miscellaneous_args[@]}"
@@ -1,10 +1,10 @@
 #!/bin/bash

 GPU_NUM=1 # 2,4,8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+MODEL_PATH="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
 MODEL_TYPE="wan"
 DATA_MERGE_PATH="data/crush-smol/merge.txt"
-OUTPUT_DIR="data/crush-smol_processed_t2v/"
+OUTPUT_DIR="data/crush-smol_processed_i2v_1_3b_inp/"

 torchrun --nproc_per_node=$GPU_NUM \
    fastvideo/pipelines/preprocess/v1_preprocess.py \
@@ -14,11 +14,11 @@ torchrun --nproc_per_node=$GPU_NUM \
    --seed 42 \
    --max_height 480 \
    --max_width 832 \
-    --num_frames 81 \
+    --num_frames 77 \
    --dataloader_num_workers 0 \
    --output_dir=$OUTPUT_DIR \
    --train_fps 16 \
    --samples_per_file 8 \
    --flush_frequency 8 \
    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "i2v" 
@@ -0,0 +1,31 @@
+{
+  "data": [
+    {
+      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
+      "image_path": null,
+      "video_path": "validation_dataset/yYcK4nANZz4-Scene-034.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.",
+      "image_path": null,
+      "video_path": "validation_dataset/yYcK4nANZz4-Scene-027.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
+      "image_path": null,
+      "video_path": "validation_dataset/yYcK4nANZz4-Scene-030.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    }
+  ]
+}
@@ -0,0 +1,16 @@
+# Wan2.1-I2V-1.3B-InP Crush-Smol Example
+These are e2e example scripts for finetuning Wan2.1 T2V 1.3B InP on the crush-smol dataset.
+
+## Execute the following commands from `FastVideo/` to run training:
+
+### Download crush-smol dataset:
+
+`bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/download_dataset.sh`
+
+### Preprocess the videos and captions into latents:
+
+`bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/preprocess_wan_data_i2v.sh`
+
+### Edit the following file and run finetuning:
+
+`bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/finetune_i2v.sh`
@@ -0,0 +1,122 @@
+#!/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-I2V-14B-480P-Diffusers"
+# MODEL_PATH="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
+DATA_DIR="data/crush-smol_processed_wan21_i2v_14b/combined_parquet_dataset/"
+# DATA_DIR="data/crush-smol_processed_i2v_1_3b_inp/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/distill/Wan2.1-I2V/crush_smol/validation_orig.json"
+NUM_GPUS=8
+export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_i2v_distill"
+  --output_dir "checkpoints/wan_i2v_distill"
+  --wandb_run_name "14b_i2v_cm"
+  --max_train_steps 1500
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 8
+  --num_height 480
+  --num_width 832
+  --num_frames 77
+  --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim 8
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 1
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  # --log_visualization
+  --log_visualization
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 20
+  --validation_sampling_steps "3"
+  --validation_guidance_scale "1.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 4e-6
+  --lr_scheduler "constant"
+  # --min_lr_ratio 0.5
+  # --lr_warmup_steps 50
+  --fake_score_learning_rate 7e-7
+  --fake_score_lr_scheduler "constant"
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 2000
+  --weight_only_checkpointing_steps 2000
+  --weight_decay 0.01
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --ema_start_step 0
+  --flow_shift 8
+  --seed 1000
+  # --enable_gradient_checkpointing_type "full"
+)
+
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --warp_denoising_step True
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.96
+  --generator_update_interval 5
+  --simulate_generator_forward
+  --simulate_forward_interval 1
+  --real_score_guidance_scale 5
+  --VSA_sparsity 0.8
+  --regression_loss_weight 0.01
+  --use_regression_loss False
+  --cm_loss_weight 1
+  --ema_decay 0.999
+  --cm_use_ema_teacher
+)
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_i2v_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}"
@@ -5,41 +5,38 @@ export WANDB_MODE=online
 export TOKENIZERS_PARALLELISM=false
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA

-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-# DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing-extended-t2v-1-3b/"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
-NUM_GPUS=1
+MODEL_PATH="Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+DATA_DIR="data/crush-smol_processed_wan21_i2v_14b/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/distill/Wan2.1-I2V/crush_smol/validation.json"
+NUM_GPUS=8
+export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+export HF_HUB_ENABLE_HF_TRANSFER=1
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_70k"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "fixed_wan_ode_init_70k_6e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 6000
+  --tracker_project_name "wan_i2v_distill"
+  --output_dir "checkpoints/wan_i2v_distill"
+  --wandb_run_name "wan21_i2v_14b_ode_50steps"
+  --max_train_steps 2000
  --train_batch_size 1
  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
+  --gradient_accumulation_steps 2
+  --num_latent_t 8
  --num_height 480
  --num_width 832
  --num_frames 77
-  --warp_denoising_step
  # --enable_gradient_checkpointing_type "full"
 )

 # Parallel arguments
 parallel_args=(
  --num_gpus $NUM_GPUS
-  --sp_size 1
+  --sp_size 8
  --tp_size 1
  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 1
+  --hsdp_shard_dim 8
 )

 # Model arguments
@@ -50,25 +47,30 @@ model_args=(

 # Dataset arguments
 dataset_args=(
-  --data_path "$DATA_DIR" 
+  --data_path "$DATA_DIR"
  --dataloader_num_workers 1
 )

 # Validation arguments
 validation_args=(
  --log_validation
+  # --log_visualization
+  --log_visualization
  --validation_dataset_file "$VALIDATION_DATASET_FILE"
  --validation_steps 50
  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
+  --validation_guidance_scale "1.0" # not used for dmd inference
 )

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 6e-6
+  --learning_rate 4e-6
+  --lr_scheduler "constant"
+  --fake_score_learning_rate 8e-7
+  --fake_score_lr_scheduler "constant"
  --mixed_precision "bf16"
-  --checkpointing_steps 1000
-  --weight_decay 1e-4
+  --checkpointing_steps 2000
+  --weight_decay 0.01
  --max_grad_norm 1.0
 )

@@ -76,24 +78,34 @@ optimizer_args=(
 miscellaneous_args=(
  --inference_mode False
  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
+  --training_cfg_rate 0.0
  --dit_precision "fp32"
-  --num_euler_timesteps 50
  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
+  --flow_shift 8
+  --seed 1000
+  # --enable_gradient_checkpointing_type "full"
+)
+
+dmd_args=(
+  --dmd_denoising_steps '1000,757,522'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --generator_update_interval 5
+  --simulate_generator_forward
+  --real_score_guidance_scale 3.5
+  --VSA_sparsity 0.8
 )

 # If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
 torchrun \
  --nnodes 1 \
  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/ode_causal_pipeline.py \
+    fastvideo/training/wan_i2v_training_pipeline.py \
    "${parallel_args[@]}" \
    "${model_args[@]}" \
    "${dataset_args[@]}" \
    "${training_args[@]}" \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}"
@@ -1,21 +1,22 @@
 #!/bin/bash
-#SBATCH --job-name=ode_crush
+#SBATCH --job-name=i2v
 #SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
+#SBATCH --nodes=4
+#SBATCH --ntasks=4
 #SBATCH --ntasks-per-node=1
 #SBATCH --gres=gpu:8
 #SBATCH --cpus-per-task=128
 #SBATCH --mem=1440G
-#SBATCH --output=ode_crush_output/ode_crush.out
-#SBATCH --error=ode_crush_output/ode_crush.err
+#SBATCH --output=i2v_output/i2v_%j.out
+#SBATCH --error=i2v_output/i2v_%j.err
 #SBATCH --exclusive
 set -e -x

 # Environment Setup
 source ~/conda/miniconda/bin/activate
-conda activate will-fv2
+conda activate will-fv

+# Basic Info
 export WANDB_MODE="online"
 export NCCL_P2P_DISABLE=1
 export TORCH_NCCL_ENABLE_MONITORING=0
@@ -29,47 +30,44 @@ export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
 export TOKENIZERS_PARALLELISM=false
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA

 echo "MASTER_ADDR: $MASTER_ADDR"
 echo "NODE_RANK: $NODE_RANK"

+# Configs
+NUM_GPUS=8

-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
-NUM_GPUS=2
+MODEL_PATH="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
+DATA_DIR="data/crush-smol_processed_i2v_1_3b_inp/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+
 # Training arguments
 training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_warp_2"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "2warp_fixed_wan_ode_init_5e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  # --warp_denoising_step
-  --max_train_steps 6001
-  --train_batch_size 1
+  --tracker_project_name wan_i2v_finetune
+  --output_dir "checkpoints/wan_i2v_finetune"
+  --max_train_steps 2000
+  --train_batch_size 2
  --train_sp_batch_size 1
  --gradient_accumulation_steps 1
-  --num_latent_t 21
+  --num_latent_t 8
  --num_height 480
  --num_width 832
  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
+  --enable_gradient_checkpointing_type "full"
 )

 # Parallel arguments
 parallel_args=(
  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim $NUM_GPUS
-  --hsdp_shard_dim 1
+  --sp_size $NUM_GPUS
+  --tp_size $NUM_GPUS
+  --hsdp_replicate_dim $SLURM_JOB_NUM_NODES
+  --hsdp_shard_dim $NUM_GPUS
 )

 # Model arguments
@@ -80,24 +78,24 @@ model_args=(

 # Dataset arguments
 dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 10
 )

 # Validation arguments
 validation_args=(
  --log_validation
  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 20
-  --validation_sampling_steps "50"
+  --validation_steps 100
+  --validation_sampling_steps "40"
  --validation_guidance_scale "6.0"
 )

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 5e-6
+  --learning_rate 1e-5
  --mixed_precision "bf16"
-  --checkpointing_steps 2000
+  --checkpointing_steps 1000
  --weight_decay 1e-4
  --max_grad_norm 1.0
 )
@@ -115,18 +113,17 @@ miscellaneous_args=(
  --enable_gradient_checkpointing_type "full"
 )

-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
 srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_i2v_training_pipeline.py \
    "${parallel_args[@]}" \
    "${model_args[@]}" \
    "${dataset_args[@]}" \
    "${training_args[@]}" \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
+    "${miscellaneous_args[@]}"
@@ -1,16 +1,18 @@
 #!/bin/bash

 GPU_NUM=1 # 2,4,8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"
 MODEL_TYPE="wan"
+# DATA_MERGE_PATH="data/single/merge.txt"
 DATA_MERGE_PATH="data/crush-smol/merge.txt"
-OUTPUT_DIR="data/crush-smol_processed_t2v_i2v_1_3b/"
+OUTPUT_DIR="data/crush-smol_processed_wan21_t2v_14b/"
+# OUTPUT_DIR="data/single_processed_i2v_14b/"

 torchrun --nproc_per_node=$GPU_NUM \
    fastvideo/pipelines/preprocess/v1_preprocess.py \
    --model_path $MODEL_PATH \
    --data_merge_path $DATA_MERGE_PATH \
-    --preprocess_video_batch_size 2 \
+    --preprocess_video_batch_size 8 \
    --seed 42 \
    --max_height 480 \
    --max_width 832 \
@@ -18,7 +20,7 @@ torchrun --nproc_per_node=$GPU_NUM \
    --dataloader_num_workers 0 \
    --output_dir=$OUTPUT_DIR \
    --train_fps 16 \
-    --samples_per_file 8 \
+    --samples_per_file 8\
    --flush_frequency 8 \
    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v_ode_trajectory" 
+    --preprocess_task "t2v" 
@@ -0,0 +1,76 @@
+{
+  "data": [
+    {
+      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
+      "image_path": null,
+      "video_path": "validation_dataset/yYcK4nANZz4-Scene-034.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.",
+      "image_path": null,
+      "video_path": "validation_dataset/yYcK4nANZz4-Scene-027.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
+      "image_path": null,
+      "video_path": "validation_dataset/yYcK4nANZz4-Scene-030.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
+      "image_path": null,
+      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-016.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "The video shows a green and orange object being flattened as if it were under a hydraulic press, with the press moving down and compressing the object.",
+      "image_path": null,
+      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-056.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "The video shows a cylindrical object with a cityscape image being flattened as if it were under a hydraulic press. The object is placed on a metal platform, and a large, striped cylinder presses down on it, causing it to collapse and release a liquid inside. The background features a green wall with a yellow and red warning sign.",
+      "image_path": null,
+      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-059.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "The video shows a close-up of an orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.",
+      "image_path": null,
+      "video_path": "validation_dataset/EJqsC21GSBY-Scene-059.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A colorful puzzle ball is being crushed by a large metal cylinder, which flattens the objects as if they were under a hydraulic press.",
+      "image_path": null,
+      "video_path": "validation_dataset/GBSfpTcKegk-Scene-003.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    }
+  ]
+}
@@ -4,7 +4,9 @@ These are end-to-end example scripts for distilling Wan2.1 T2V 1.3B model using
 ### 0. Make sure you have installed VSA

 ```bash
-pip install vsa
+cd csrc/attn
+git submodule update --init --recursive
+python setup_vsa.py install
 ```

 ### 1. Download dataset:
@@ -4,7 +4,9 @@ These are end-to-end example scripts for distilling Wan2.2 TI2V 5B model DMD+VSA
 ### 0. Make sure you have installed VSA

 ```bash
-pip install vsa
+cd csrc/attn
+git submodule update --init --recursive
+python setup_vsa.py install
 ```

 ### Data-free Distillation
@@ -4,7 +4,9 @@ These are end-to-end example scripts for distilling Wan2.2 TI2V 5B model DMD+VSA
 ### 0. Make sure you have installed VSA

 ```bash
-pip install vsa
+cd csrc/attn
+git submodule update --init --recursive
+python setup_vsa.py install
 ```

 ### 1. Download dataset:
@@ -98,7 +98,6 @@ dmd_args=(
 torchrun \
 --nnodes 1 \
 --nproc_per_node $NUM_GPUS \
--master_port $MASTER_PORT \
    fastvideo/training/wan_distillation_pipeline.py \
    "${parallel_args[@]}" \
    "${model_args[@]}" \
@@ -1,112 +0,0 @@
-#!/bin/bash
-
-# Basic Info
-export WANDB_MODE="online"
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-export MASTER_PORT=29501
-export TOKENIZERS_PARALLELISM=false
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-# Configs
-NUM_GPUS=1
-MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
-DATA_DIR="data/crush-smol_processed_ti2v/combined_parquet_dataset/"
-VALIDATION_DATASET_FILE="examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/validation.json"
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir="checkpoints/wan_t2v_finetune"
-  --max_train_steps=4000
-  --train_batch_size=1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps=1
-  --num_latent_t 31
-  --num_height 704
-  --num_width 1280
-  --num_frames 121
-  --enable_gradient_checkpointing_type "full"
-  --training_state_checkpointing_steps=500
-  --weight_only_checkpointing_steps=500
-  --lora_rank 32
-  --lora_training True
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus 1
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR"
-  --dataloader_num_workers 4
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 200
-  --validation_sampling_steps "3"
-  --validation_guidance_scale "6.0" # not used for dmd inference
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate=1e-4
-  --mixed_precision="bf16"
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.0
-  --dit_precision "fp32"
-  --ema_start_step 0
-  --flow_shift 8
-  --seed 1000
-)
-
-# DMD arguments
-dmd_args=(
-  --dmd_denoising_steps '1000,757,522'
-  --min_timestep_ratio 0.02
-  --max_timestep_ratio 0.98
-  --generator_update_interval 5
-  --real_score_guidance_scale 3.5
-  --VSA_sparsity 0.8
-)
-
-torchrun \
--nnodes 1 \
--nproc_per_node $NUM_GPUS \
--master_port $MASTER_PORT \
-    fastvideo/training/wan_distillation_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
@@ -0,0 +1,118 @@
+#!/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-14B-Diffusers"
+# MODEL_PATH="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
+DATA_DIR="data/crush-smol_processed_wan21_t2v_14b/combined_parquet_dataset/"
+# DATA_DIR="data/crush-smol_processed_i2v_1_3b_inp/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/distill/Wan2.1-I2V/crush_smol/validation_orig.json"
+NUM_GPUS=8
+export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_i2v_distill"
+  --output_dir "checkpoints/wan_i2v_distill"
+  --wandb_run_name "cm_t2v"
+  --max_train_steps 1500
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 8
+  --num_height 480
+  --num_width 832
+  --num_frames 77
+  --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim 8
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 1
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  # --log_visualization
+  --log_visualization
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 20
+  --validation_sampling_steps "3"
+  --validation_guidance_scale "1.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 4e-6
+  --lr_scheduler "constant"
+  # --min_lr_ratio 0.5
+  # --lr_warmup_steps 50
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 2000
+  --weight_only_checkpointing_steps 2000
+  --weight_decay 0.01
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --ema_start_step 0
+  --flow_shift 8
+  --seed 1000
+  # --enable_gradient_checkpointing_type "full"
+)
+
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --warp_denoising_step True
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.96
+  --real_score_guidance_scale 5
+  --VSA_sparsity 0.8
+  --regression_loss_weight 0.01
+  --use_regression_loss False
+  --cm_loss_weight 1
+  --ema_decay 0.999
+  --cm_weighing_function "constant"
+  # --cm_use_ema_teacher
+)
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_cm_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}"
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
SolitaryThinker	083c2258ea	fix cm	2025-08-21 05:18:30 +00:00
SolitaryThinker	60260a6be9	fix preprocessing before vae deocde	2025-08-20 01:10:53 +00:00
SolitaryThinker	06cc80ee93	init lpips	2025-08-19 23:48:06 +00:00
Matthew Noto	c6b3475f77	LPIPs fix	2025-08-19 16:40:42 -07:00
Matthew Noto	1d34380c58	LPIPS	2025-08-19 15:40:29 -07:00
SolitaryThinker	ddf8198e65	remove validation back sim; reg loss for back sim	2025-08-19 07:19:40 +00:00
SolitaryThinker	c38ee88294	cm	2025-08-18 23:22:37 +00:00
SolitaryThinker	ff1b77bdd6	update	2025-08-18 21:44:47 +00:00
SolitaryThinker	22ba3759ee	cm	2025-08-18 03:14:18 +00:00
SolitaryThinker	2c3a3893c0	cm	2025-08-18 02:46:24 +00:00
SolitaryThinker	6e9ba7ba9d	sim interval arg	2025-08-17 02:49:03 +00:00
SolitaryThinker	64f1295878	fix	2025-08-15 03:45:32 +00:00
SolitaryThinker	d9cc3b7e42	reg loss	2025-08-15 02:08:15 +00:00
SolitaryThinker	cb3230f5e9	reg loss	2025-08-15 01:30:46 +00:00
SolitaryThinker	1e91cec93d	fix	2025-08-13 07:05:51 +00:00
SolitaryThinker	5f0688084c	update	2025-08-13 07:05:24 +00:00
SolitaryThinker	f3023db253	fix	2025-08-13 07:01:18 +00:00
SolitaryThinker	c405941f21	fix	2025-08-13 04:05:01 +00:00
SolitaryThinker	f4e818305b	fix	2025-08-13 03:11:22 +00:00
SolitaryThinker	4c4ab65020	fix	2025-08-13 03:10:12 +00:00
SolitaryThinker	6b8bf495ee	fix	2025-08-13 02:19:52 +00:00
SolitaryThinker	ebd23ba8e7	add orig val	2025-08-12 20:00:33 +00:00
SolitaryThinker	0c1cb0c19d	runnable	2025-08-12 19:58:36 +00:00
SolitaryThinker	d87ebe051b	fix	2025-08-11 03:42:08 +00:00
SolitaryThinker	0ae669cd4c	fix	2025-08-11 03:31:02 +00:00
SolitaryThinker	63c124fec8	update	2025-08-10 23:37:01 +00:00
SolitaryThinker	ececbe1839	wip	2025-08-10 08:33:03 +00:00