update

[Kernel] Remove all syncs from STA & VSA kernels (#517 )
[Feature][Training]Add diffusers format checkpoint saving for inference (#542 )
2025-06-24 15:30:55 -07:00 · 2025-06-23 13:13:09 -07:00 · 2025-06-22 01:23:41 -04:00 · 2025-06-21 17:58:35 -07:00 · 2025-06-20 22:10:56 -07:00 · 2025-06-20 20:42:25 -07:00
261 changed files with 59399 additions and 45315 deletions
@@ -0,0 +1,66 @@
+env:
+  IMAGE_VERSION: "py3.12-latest"
+
+steps:
+  - block: "Start Build"
+    blocked_state: "running"
+    prompt: "Approve build?"
+
+  - label: "Trigger Tests"
+    command: |
+      echo "Current working directory: $(pwd)"
+      echo "Current branch:"
+      git branch --show-current
+      echo "Full diff:"
+      git diff --name-only $BUILDKITE_PULL_REQUEST_BASE_BRANCH...HEAD
+    plugins:
+      - monorepo-diff#v1.4.0:
+          diff: "git diff --name-only $BUILDKITE_PULL_REQUEST_BASE_BRANCH...HEAD"
+          watch:
+            - path: 
+                - "fastvideo/v1/models/encoders/**"
+                - "fastvideo/v1/models/loaders/**"
+                - "fastvideo/v1/tests/encoders/**"
+              config:
+                command: "timeout 30m .buildkite/scripts/pr_test.sh"
+                label: "Encoder Tests"
+                env:
+                  - BUILDKITE_CLEAN_CHECKOUT=true
+                  - TEST_TYPE=encoder
+                agents:
+                  queue: "default"
+            - path:
+                - "fastvideo/v1/models/vaes/**"
+                - "fastvideo/v1/models/loaders/**"
+                - "fastvideo/v1/tests/vaes/**"
+              config:
+                command: "timeout 30m .buildkite/scripts/pr_test.sh"
+                label: "VAE Tests"
+                env:
+                  - BUILDKITE_CLEAN_CHECKOUT=true
+                  - TEST_TYPE=vae
+                agents:
+                  queue: "default"
+            - path:
+                - "fastvideo/v1/models/dits/**"
+                - "fastvideo/v1/models/loaders/**"
+                - "fastvideo/v1/tests/transformers/**"
+                - "fastvideo/v1/layers/**"
+                - "fastvideo/v1/attention/**"
+              config:
+                command: "timeout 30m .buildkite/scripts/pr_test.sh"
+                label: "Transformer Tests"
+                env:
+                  - BUILDKITE_CLEAN_CHECKOUT=true
+                  - TEST_TYPE=transformer
+                agents:
+                  queue: "default"
+            - path: "fastvideo/v1/**/*.py"
+              config:
+                command: "timeout 60m .buildkite/scripts/pr_test.sh"
+                label: "SSIM Tests"
+                env:
+                  - BUILDKITE_CLEAN_CHECKOUT=true
+                  - TEST_TYPE=ssim
+                agents:
+                  queue: "default"
@@ -0,0 +1,91 @@
+#!/bin/bash
+set -uo pipefail
+
+log() {
+  echo "[$(date '+%Y-%m-%d %H:%M:%S')] $1"
+}
+
+log "=== Starting Modal test execution ==="
+
+# Change to the project directory
+cd "$(dirname "$0")/../.."
+PROJECT_ROOT=$(pwd)
+log "Project root: $PROJECT_ROOT"
+
+# Install Modal if not available
+if ! python3 -m modal --version &> /dev/null; then
+    log "Modal not found, installing..."
+    python3 -m pip install modal
+    
+    # Verify installation
+    if ! python3 -m modal --version &> /dev/null; then
+        log "Error: Failed to install modal. Please install it manually."
+        exit 1
+    fi
+fi
+
+log "modal version: $(python3 -m modal --version)"
+
+# Set up Modal authentication using Buildkite secrets
+log "Setting up Modal authentication from Buildkite secrets..."
+MODAL_TOKEN_ID=$(buildkite-agent secret get modal_token_id)
+MODAL_TOKEN_SECRET=$(buildkite-agent secret get modal_token_secret)
+
+if [ -n "$MODAL_TOKEN_ID" ] && [ -n "$MODAL_TOKEN_SECRET" ]; then
+    log "Retrieved Modal credentials from Buildkite secrets"
+    python3 -m modal token set --token-id "$MODAL_TOKEN_ID" --token-secret "$MODAL_TOKEN_SECRET" --profile buildkite-ci --activate --verify
+    if [ $? -eq 0 ]; then
+        log "Modal authentication successful"
+    else
+        log "Error: Failed to set Modal credentials"
+        exit 1
+    fi
+else
+    log "Error: Could not retrieve Modal credentials from Buildkite secrets."
+    log "Please ensure 'modal_token_id' and 'modal_token_secret' secrets are set in Buildkite."
+    exit 1
+fi
+
+MODAL_TEST_FILE="fastvideo/v1/tests/modal/pr_test.py"
+
+if [ -z "${TEST_TYPE:-}" ]; then
+    log "Error: TEST_TYPE environment variable is not set"
+    exit 1
+fi
+log "Test type: $TEST_TYPE"
+
+case "$TEST_TYPE" in
+    "encoder")
+        log "Running encoder tests..."
+        MODAL_COMMAND="python3 -m modal run $MODAL_TEST_FILE::run_encoder_tests"
+        ;;
+    "vae")
+        log "Running VAE tests..."
+        MODAL_COMMAND="python3 -m modal run $MODAL_TEST_FILE::run_vae_tests"
+        ;;
+    "transformer")
+        log "Running transformer tests..."
+        MODAL_COMMAND="python3 -m modal run $MODAL_TEST_FILE::run_transformer_tests"
+        ;;
+    "ssim")
+        log "Running SSIM tests..."
+        MODAL_COMMAND="python3 -m modal run $MODAL_TEST_FILE::run_ssim_tests"
+        ;;
+    *)
+        log "Error: Unknown test type: $TEST_TYPE"
+        exit 1
+        ;;
+esac
+
+log "Executing: $MODAL_COMMAND"
+eval "$MODAL_COMMAND"
+TEST_EXIT_CODE=$?
+
+if [ $TEST_EXIT_CODE -eq 0 ]; then
+    log "Modal test completed successfully"
+else
+    log "Error: Modal test failed with exit code: $TEST_EXIT_CODE"
+fi
+
+log "=== Test execution completed with exit code: $TEST_EXIT_CODE ==="
+exit $TEST_EXIT_CODE
@@ -4,14 +4,6 @@ title: "[Bug] "
 labels: ['Bug']

 body:
- type: textarea
-  attributes:
-    label: Environment
-    description: |
-      Please share your environment with us. You can run the command **python fastvideo/utils/env_utils.py** and copy-paste its output below.
-    placeholder: FastVideo version, platform, python version, cuda version...
-  validations:
-    required: true
 - type: textarea
  attributes:
    label: Describe the bug
@@ -25,5 +17,13 @@ body:
      What command or script did you run? Which **model** are you using?
    placeholder: |
      A placeholder for the command.
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Environment
+    description: |
+      Please share your environment with us. You can run the command **python fastvideo/utils/collect_env.py** and copy-paste its output below.
+    placeholder: FastVideo version, platform, python version, cuda version...
  validations:
    required: true
@@ -160,8 +160,7 @@ def execute_command(pod_id):
    setup_steps = [
        "tar -xzf /tmp/repo.tar.gz --no-same-owner -C /workspace/",
        f"cd /workspace/{repo_name}",
-        "source /opt/conda/etc/profile.d/conda.sh",
-        "conda activate fastvideo-dev",
+        "source $HOME/.local/bin/env && source /opt/venv/bin/activate",
        args.test_command
    ]
    
@@ -12,13 +12,11 @@ on:
    paths:
      - "fastvideo/**/*.py"
      - ".github/workflows/pr-test.yml"
+      - "pyproject.toml"
+      - "docker/Dockerfile.python3.12"
+      - "csrc/**"
  workflow_dispatch:
    inputs:
-      custom_image:
-        description: "Custom image from this repository (default: fastvideo-dev:latest)"
-        required: false
-        default: "fastvideo-dev:latest"
-        type: string
      run_encoder_test:
        description: "Run encoder-test"
        required: false
@@ -39,10 +37,41 @@ on:
        required: false
        default: false
        type: boolean
+      run_training_test:
+        description: "Run training-test"
+        required: false
+        default: false
+        type: boolean
+      run_training_test_VSA:
+        description: "Run training-test-VSA"
+        required: false
+        default: false
+        type: boolean
+      run_inference_test_STA:
+        description: "Run inference-test-STA"
+        required: false
+        default: false
+        type: boolean
+      run_precision_test_STA:
+        description: "Run precision-test-STA"
+        required: false
+        default: false
+        type: boolean
+      run_precision_test_VSA:
+        description: "Run precision-test-VSA"
+        required: false
+        default: false
+        type: boolean
+      run_nightly_test:
+        description: "Run nightly-test"
+        required: false
+        default: false
+        type: boolean

 env:
  PYTHONUNBUFFERED: "1"

+
 concurrency:
  group: pr-test-${{ github.ref }}
  cancel-in-progress: true
@@ -59,24 +88,72 @@ jobs:
      encoder-test: ${{ steps.filter.outputs.encoder-test }}
      vae-test: ${{ steps.filter.outputs.vae-test }}
      transformer-test: ${{ steps.filter.outputs.transformer-test }}
+      training-test: ${{ steps.filter.outputs.training-test }}
+      training-test-VSA: ${{ steps.filter.outputs.training-test-VSA }}
+      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 }}
    steps:
      - uses: actions/checkout@v4
      - uses: dorny/paths-filter@v3
        id: filter
        with:
          filters: |
+            # Define reusable path patterns
+            common-paths: &common-paths
+              - 'pyproject.toml'
+              - 'docker/Dockerfile.python3.12'
+            sta-kernel-paths: &sta-kernel-paths
+              - '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/vsa/**'
+              - 'csrc/attn/tk/**'
+              - 'csrc/attn/setup_vsa.py'
+              - 'csrc/attn/config_vsa.py'
+              - 'csrc/attn/vsa.cpp'
+            vsa-paths: &vsa-paths
+              - 'fastvideo/v1/**'
+              - *common-paths
+              - *vsa-kernel-paths
+            
+            # Actual tests
            encoder-test:
              - 'fastvideo/v1/models/encoders/**'
              - 'fastvideo/v1/models/loaders/**'
              - 'fastvideo/v1/tests/encoders/**'
+              - *common-paths
            vae-test:
              - 'fastvideo/v1/models/vaes/**'
              - 'fastvideo/v1/models/loaders/**'
              - 'fastvideo/v1/tests/vaes/**'
+              - *common-paths
            transformer-test:
              - 'fastvideo/v1/models/dits/**'
              - 'fastvideo/v1/models/loaders/**'
              - 'fastvideo/v1/tests/transformers/**'
+              - 'fastvideo/v1/layers/**'
+              - 'fastvideo/v1/attention/**'
+              - *common-paths
+            training-test:
+              - 'fastvideo/v1/**'
+              - *common-paths
+            training-test-VSA:
+              - 'fastvideo/v1/**'
+              - *common-paths
+              - *vsa-kernel-paths
+            inference-test-STA:
+              - 'fastvideo/v1/**'
+              - *common-paths
+              - *sta-kernel-paths
+            precision-test-STA:
+              - *common-paths
+              - *sta-kernel-paths
+            precision-test-VSA:
+              - *common-paths
+              - *vsa-kernel-paths

  encoder-test:
    needs: change-filter
@@ -89,8 +166,8 @@ jobs:
      gpu_type: "NVIDIA A40"
      gpu_count: 1
      volume_size: 100
-      image: "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
-      test_command: "pip install -e .[test] && pytest ./fastvideo/v1/tests/encoders -s"
+      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/encoders -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -107,8 +184,8 @@ jobs:
      gpu_type: "NVIDIA A40"
      gpu_count: 1
      volume_size: 100
-      image: "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
-      test_command: "pip install -e .[test] && pytest ./fastvideo/v1/tests/vaes -s"
+      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/vaes -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -125,8 +202,8 @@ jobs:
      gpu_type: "NVIDIA L40S"
      gpu_count: 1
      volume_size: 100
-      image: "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
-      test_command: "pip install -e .[test] && pytest ./fastvideo/v1/tests/transformers -s"
+      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/transformers -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -141,8 +218,8 @@ jobs:
      fail-fast: false
      matrix:
        python-version: [
-          {version: "3.10", tag: "latest"},
-          {version: "3.11", tag: "py3.11-latest"},
+          # {version: "3.10", tag: "latest"},
+          # {version: "3.11", tag: "py3.11-latest"},
          {version: "3.12", tag: "py3.12-latest"}
        ]
    uses: ./.github/workflows/runpod-test.yml
@@ -153,14 +230,130 @@ jobs:
      volume_size: 200
      disk_size: 200
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:${{ matrix.python-version.tag }}"
-      test_command: "pip install -e .[test] && pytest ./fastvideo/v1/tests/ssim -vs"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/ssim -vs"
      timeout_minutes: 60
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+  
+  training-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_training_test == 'true')
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "training-test"
+      gpu_type: "NVIDIA A40"
+      gpu_count: 4
+      volume_size: 100
+      disk_size: 100
+      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/v1/tests/training/Vanilla -srP"
+      timeout_minutes: 30
+    secrets:
+      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+      WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
+
+  training-test-VSA:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_training_test_VSA == 'true')
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "training-test-VSA"
+      gpu_type: "NVIDIA H100 NVL"
+      gpu_count: 1
+      volume_size: 100
+      disk_size: 100
+      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/v1/tests/training/VSA -srP"
+      timeout_minutes: 30
+    secrets:
+      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+      WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
+
+  inference-test-STA:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_inference_test_STA == 'true')
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "inference-test-STA"
+      gpu_type: "NVIDIA H100 NVL"
+      gpu_count: 1
+      volume_size: 100
+      disk_size: 100
+      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/inference/STA -srP"
+      timeout_minutes: 30
+    secrets:
+      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+
+  precision-test-STA:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_precision_test_STA == 'true')
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "precision-test-STA"
+      gpu_type: "NVIDIA H100 NVL"
+      gpu_count: 1
+      volume_size: 100
+      disk_size: 100
+      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+      test_command: "uv pip install -e .[test] && python csrc/attn/tests/test_sta.py"
+      timeout_minutes: 30
+    secrets:
+      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+
+  precision-test-VSA:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_precision_test_VSA == 'true')
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "precision-test-VSA"
+      gpu_type: "NVIDIA H100 NVL"
+      gpu_count: 1
+      volume_size: 100
+      disk_size: 100
+      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+      test_command: "uv pip install -e .[test] && python csrc/attn/tests/test_block_sparse.py"
+      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/v1/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 }}

  runpod-cleanup:
-    needs: [encoder-test, vae-test, transformer-test, ssim-test] # Add other jobs to this list as you create them
+    # Add other jobs to this list as you create them
+    needs: [encoder-test, vae-test, transformer-test, ssim-test, training-test, training-test-VSA, inference-test-STA, precision-test-STA, precision-test-VSA] 
    if: ${{ always() && ((github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || github.event_name == 'workflow_dispatch') }}
    runs-on: ubuntu-latest
    steps:
@@ -177,7 +370,7 @@ jobs:

      - name: Cleanup all RunPod instances
        env:
-          JOB_IDS: '["encoder-test", "vae-test", "transformer-test", "ssim-test-py3.10", "ssim-test-py3.11", "ssim-test-py3.12"]'
+          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
@@ -10,7 +10,7 @@ jobs:
    - uses: actions/checkout@v4
    - uses: actions/setup-python@v5
      with:
-        python-version: "3.10"
+        python-version: "3.12"
    - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
    - run: echo "::add-matcher::.github/workflows/matchers/mypy.json"
    - uses: pre-commit/action@v3.0.1
@@ -43,6 +43,8 @@ on:
        required: true
      RUNPOD_PRIVATE_KEY:
        required: true
+      WANDB_API_KEY:
+        required: false

 jobs:
  run-test:
@@ -55,7 +57,7 @@ jobs:
      - name: Set up Python
        uses: actions/setup-python@v5
        with:
-          python-version: "3.10"
+          python-version: "3.12"

      - name: Set up SSH key
        run: |
@@ -72,6 +74,7 @@ jobs:
          JOB_ID: ${{ inputs.job_id }}
          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
          GITHUB_RUN_ID: ${{ github.run_id }}
+          WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
        timeout-minutes: ${{ inputs.timeout_minutes }}
        run: >-
          python .github/scripts/runpod_api.py 
@@ -5,7 +5,7 @@ on:
    branches:
      - main
    paths:
-      - "csrc/sliding_tile_attention/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/sliding_tile_attention
+          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
@@ -136,19 +136,21 @@ jobs:

      - name: Build wheel
        run: |
+          export PYTHONPATH=$GITHUB_WORKSPACE:$PYTHONPATH
+          
          # We want setuptools >= 49.6.0 otherwise we can't compile the extension if system CUDA version is 11.7 and pytorch cuda version is 11.6
          # https://github.com/pytorch/pytorch/blob/664058fa83f1d8eede5d66418abff6e20bd76ca8/torch/utils/cpp_extension.py#L810
          # However this still fails so I'm using a newer version of setuptools
          pip install setuptools
          pip install ninja packaging wheel

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

      - name: Rename wheel file
        run: |
-          cd csrc/sliding_tile_attention
+          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)
@@ -163,7 +165,7 @@ jobs:
        uses: actions/upload-artifact@v4
        with:
          name: ${{ env.wheel_name }}
-          path: csrc/sliding_tile_attention/dist/*.whl
+          path: csrc/attn/dist/*.whl
          retention-days: 90

  publish_package:
@@ -229,17 +231,19 @@ jobs:

      - name: Build source distribution
        run: |
+          export PYTHONPATH=$GITHUB_WORKSPACE:$PYTHONPATH
+          
          # We want setuptools >= 49.6.0 otherwise we can't compile the extension if system CUDA version is 11.7 and pytorch cuda version is 11.6
          # https://github.com/pytorch/pytorch/blob/664058fa83f1d8eede5d66418abff6e20bd76ca8/torch/utils/cpp_extension.py#L810
          # However this still fails so I'm using a newer version of setuptools
          pip install setuptools
          pip install ninja packaging wheel

-          cd csrc/sliding_tile_attention  # Move into the correct folder
-          git submodule update --init --recursive tk # Ensure ThunderKittens submodule is initialized
-          python setup.py sdist --dist-dir=dist
+          cd csrc/attn  # Move into the correct folder
+          git submodule update --init --recursive # Ensure ThunderKittens submodule is initialized
+          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/sliding_tile_attention/dist/
+          packages-dir: csrc/attn/dist/
@@ -28,4 +28,4 @@ jobs:

      - name: Run Pytest
        run: |
-          pytest --ignore csrc/sliding_tile_attention/test
+          pytest --ignore csrc/attn/test
@@ -27,7 +27,6 @@ env
 **/build/
 **.pyc
 **.txt
-**.json

 # Distribution / packaging
 build/
@@ -1,3 +1,3 @@
-[submodule "csrc/sliding_tile_attention/tk"]
-	path = csrc/sliding_tile_attention/tk
+[submodule "csrc/attn/tk"]
+	path = csrc/attn/tk
 	url = https://github.com/HazyResearch/ThunderKittens.git
@@ -33,7 +33,7 @@ repos:
    args: [--in-place, --verbose]
    additional_dependencies: [toml] # TODO: Remove when yapf is upgraded
 - repo: https://github.com/astral-sh/ruff-pre-commit
-  rev: v0.11.4
+  rev: v0.11.12
  hooks:
  - id: ruff
    args: [--output-format, github, --fix]
@@ -48,7 +48,7 @@ repos:
  hooks:
  - id: isort
 - repo: https://github.com/jackdewinter/pymarkdown
-  rev: v0.9.29
+  rev: v0.9.30
  hooks:
  - id: pymarkdown
    args: [fix]
@@ -91,7 +91,7 @@ For a more detailed guide, please see our [inference quick start](https://hao-ai

 ## Distillation and Finetuning
 - [Distillation Guide](https://hao-ai-lab.github.io/FastVideo/training/distillation.html)
- [Finetuning Guide](https://hao-ai-lab.github.io/FastVideo/training/finetuning.html)
+- [Finetuning Guide](https://hao-ai-lab.github.io/FastVideo/training/finetune.html)

 ## 📑 Development Plan

@@ -111,7 +111,7 @@ For a more detailed guide, please see our [inference quick start](https://hao-ai

 ## 🤝 Contributing

-We welcome all contributions. Please check out our guide [here](https://hao-ai-lab.github.io/FastVideo/developer_guide/overview.html)
+We welcome all contributions. Please check out our guide [here](https://hao-ai-lab.github.io/FastVideo/contributing/overview.html)

 ## Acknowledgement
 We learned and reused code from the following projects:
@@ -4,9 +4,8 @@


 ## Installation
-We test our code on Pytorch 2.5.0 and CUDA>=12.4. Currently we only have implementation on H100.
+We test our code on Pytorch 2.5.0 and CUDA>=12.4. Currently we only support H100/H200, because ThunderKittens uses TMA but doesn't support Blackwell yet.
 First, install C++20 for ThunderKittens:
-
 ```bash
 sudo apt update
 sudo apt install gcc-11 g++-11
@@ -16,17 +15,27 @@ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave
 sudo apt update
 sudo apt install clang-11
 ```
-Install STA:
+
+## Environment Setup
+First, set up your CUDA environment:
 ```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
 git submodule update --init --recursive
-python setup.py install
+```
+
+## Install Sliding Tile Attention (STA)
+```bash
+python setup_sta.py install
+```
+
+## Install Video Sparse Attention (VSA)
+```bash
+python setup_vsa.py install
 ```

 ## Usage
-
 ```python
 from st_attn import sliding_tile_attention
 # assuming video size (T, H, W) = (30, 48, 80), text tokens = 256 with padding. 
@@ -44,8 +53,14 @@ out = sliding_tile_attention(q, k, v, window_size, 0, False)

 ## Test
 ```bash
-python test/test_sta.py
+python tests/test_sta.py # test STA
+python tests/test_block_sparse.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. 
@@ -5,6 +5,7 @@ import matplotlib.pyplot as plt
 import numpy as np
 import torch
 from st_attn import sliding_tile_attention
+from triton.testing import do_bench


 def flops(batch, seqlen, nheads, headdim, causal, mode="fwd"):
@@ -13,16 +14,16 @@ def flops(batch, seqlen, nheads, headdim, causal, mode="fwd"):
    return f if mode == "fwd" else (2.5 * f if mode == "bwd" else 3.5 * f)


-def efficiency(flop, time):
-    flop = flop / 1e12
-    time = time / 1e6
-    return flop / time
+def compute_TFLOPS(flops, ms):
+    flops = flops / 1e12
+    ms = ms / 1e3
+    return flops / ms


 def benchmark_attention(configurations):
    results = {'fwd': defaultdict(list), 'bwd': defaultdict(list)}

-    for B, H, N, D, causal in configurations:
+    for B, H, N, D, causal, dit_seq_shape, window_size in configurations:
        print("=" * 60)
        print(f"Timing forward and backward pass for B={B}, H={H}, N={N}, D={D}, causal={causal}")

@@ -30,38 +31,31 @@ def benchmark_attention(configurations):
        k = torch.randn(B, H, N, D, dtype=torch.bfloat16, device='cuda', requires_grad=False).contiguous()
        v = torch.randn(B, H, N, D, dtype=torch.bfloat16, device='cuda', requires_grad=False).contiguous()

-        grad_output = torch.randn_like(q, requires_grad=False).contiguous()
+        # grad_output = torch.randn_like(q, requires_grad=False).contiguous()
+        # qg = torch.zeros_like(q, requires_grad=False, dtype=torch.float).contiguous()
+        # kg = torch.zeros_like(k, requires_grad=False, dtype=torch.float).contiguous()
+        # vg = torch.zeros_like(v, requires_grad=False, dtype=torch.float).contiguous()

-        qg = torch.zeros_like(q, requires_grad=False, dtype=torch.float).contiguous()
-        kg = torch.zeros_like(k, requires_grad=False, dtype=torch.float).contiguous()
-        vg = torch.zeros_like(v, requires_grad=False, dtype=torch.float).contiguous()

-        # Prepare for timing forward pass
-        start_events_fwd = [torch.cuda.Event(enable_timing=True) for _ in range(10)]
-        end_events_fwd = [torch.cuda.Event(enable_timing=True) for _ in range(10)]
+        # # Warmup for forward pass
+        # for _ in range(10):
+        #     o = sliding_tile_attention(q, k, v, [[3, 6, 10]] * 24, 0, False, dit_seq_shape)

-        torch.cuda.empty_cache()
-        torch.cuda.synchronize()
+        # # Time the forward pass
+        # for i in range(10):
+        #     start_events_fwd[i].record()
+        #     o = sliding_tile_attention(q, k, v, [[3, 6, 10]] * 24, 0, False, dit_seq_shape)
+        #     end_events_fwd[i].record()
+        ms = do_bench(lambda: sliding_tile_attention(q, k, v, [window_size] * 24, 0, False, dit_seq_shape))

-        # Warmup for forward pass
-        for _ in range(10):
-            o = sliding_tile_attention(q, k, v, [[6, 6, 6]] * 24, 0, False)
+        # times_fwd = [s.elapsed_time(e) for s, e in zip(start_events_fwd, end_events_fwd)]
+        # time_us_fwd = np.mean(times_fwd) * 1000

-        # Time the forward pass
-        for i in range(10):
-            start_events_fwd[i].record()
-            o = sliding_tile_attention(q, k, v, [[6, 6, 6]] * 24, 0, False)
-            end_events_fwd[i].record()
-
-        torch.cuda.synchronize()
-        times_fwd = [s.elapsed_time(e) for s, e in zip(start_events_fwd, end_events_fwd)]
-        time_us_fwd = np.mean(times_fwd) * 1000
-
-        tflops_fwd = efficiency(flops(B, N, H, D, causal, 'fwd'), time_us_fwd)
+        tflops_fwd = compute_TFLOPS(flops(B, N, H, D, causal, 'fwd'), ms)
        results['fwd'][(D, causal)].append((N, tflops_fwd))

-        print(f"Average time for forward pass in us: {time_us_fwd:.2f}")
-        print(f"Average efficiency for forward pass in TFLOPS: {tflops_fwd}")
+        print(f"Average time for forward pass (ms): {ms:.2f}")
+        print(f"Average TFLOPS: {tflops_fwd}")
        print("-" * 60)

        # torch.cuda.empty_cache()
@@ -85,15 +79,14 @@ def benchmark_attention(configurations):
        # times_bwd = [s.elapsed_time(e) for s, e in zip(start_events_bwd, end_events_bwd)]
        # time_us_bwd = np.mean(times_bwd) * 1000

-        # tflops_bwd = efficiency(flops(B, N, H, D, causal, 'bwd'), time_us_bwd)
+        # tflops_bwd = compute_TFLOPS(flops(B, N, H, D, causal, 'bwd'), ms)
        # results['bwd'][(D, causal)].append((N, tflops_bwd))

-        # print(f"Average time for backward pass in us: {time_us_bwd:.2f}")
-        # print(f"Average efficiency for backward pass in TFLOPS: {tflops_bwd}")
-        print("=" * 60)
+        # print(f"Average time for backward pass(ms): {ms:.2f}")
+        # print(f"Average TFLOPS: {tflops_bwd}")
+        # print("=" * 60)

        torch.cuda.empty_cache()
-        torch.cuda.synchronize()

    return results

@@ -124,7 +117,10 @@ def plot_results(results):

 # Example list of configurations to test
 configurations = [
-    (2, 24, 82944, 128, False),
+    (2, 24, 69120, 128, False, '18x48x80', [3, 6, 10]),
+    (2, 24, 69120, 128, True, '18x48x80', [3, 6, 10]),
+    (2, 24, 82944, 128, False, '36x48x48', [3, 3, 6]), # Stepvideo
+    (2, 24, 82944, 128, True, '36x48x48', [3, 3, 6]),
    # (16, 16, 768*16,  128, False),
    # (16, 16, 768*2,  128, False),
    # (16, 16, 768*4,  128, False),
@@ -0,0 +1,225 @@
+import torch
+import argparse
+from flash_attn.utils.benchmark import benchmark_forward
+from vsa import block_sparse_attention_fwd, block_sparse_attention_backward
+from vsa import BLOCK_M, BLOCK_N
+
+import numpy as np
+import random
+
+def set_seed(seed: int = 42):
+    # Python random module
+    random.seed(seed)
+
+    # NumPy
+    np.random.seed(seed)
+
+    # PyTorch
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)  # if using multi-GPU
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(description='Benchmark Block Sparse Attention')
+    parser.add_argument('--batch_size', type=int, default=1, help='Batch size')
+    parser.add_argument('--num_heads', type=int, default=12, help='Number of heads')
+    parser.add_argument('--head_dim', type=int, default=64, help='Head dimension')
+    parser.add_argument('--topk', type=int, default=None, help='Number of kv blocks each q block attends to')
+    parser.add_argument('--seq_lengths', type=int, nargs='+', default=[49152], help='Sequence lengths to benchmark')
+    return parser.parse_args()
+
+def create_input_tensors(batch, head, seq_len, headdim):
+    """Create random input tensors for attention."""
+    q = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
+    k = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
+    v = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
+    return q, k, v
+
+def generate_block_sparse_pattern(bs, h, num_q_blocks, num_kv_blocks, k, device="cuda"):
+    """
+    Generate a block sparse pattern where each q block attends to exactly k kv blocks.
+    
+    Args:
+        bs: batch size
+        h: number of heads
+        num_q_blocks: number of query blocks
+        num_kv_blocks: number of key-value blocks
+        k: number of kv blocks each q block attends to
+        device: device to create tensors on
+        
+    Returns:
+        q2k_block_sparse_index: [bs, h, num_q_blocks, k]
+            Contains the indices of kv blocks that each q block attends to.
+        q2k_block_sparse_num: [bs, h, num_q_blocks]
+            Contains the number of kv blocks that each q block attends to (all equal to k).
+        k2q_block_sparse_index: [bs, h, num_kv_blocks, num_q_blocks]
+            Contains the indices of q blocks that attend to each kv block.
+        k2q_block_sparse_num: [bs, h, num_kv_blocks]
+            Contains the number of q blocks that attend to each kv block.
+        block_sparse_mask: [bs, h, num_q_blocks, num_kv_blocks]
+            Binary mask where 1 indicates attention connection.
+    """
+    # Ensure k is not larger than num_kv_blocks
+    k = min(k, num_kv_blocks)
+    
+    # Create random scores for sampling
+    scores = torch.rand(bs, h, num_q_blocks, num_kv_blocks, device=device)
+    
+    # Get top-k indices for each q block
+    _, q2k_block_sparse_index = torch.topk(scores, k, dim=-1)
+    q2k_block_sparse_index = q2k_block_sparse_index.to(torch.int32)
+    
+    # sort q2k_block_sparse_index
+    q2k_block_sparse_index, _ = torch.sort(q2k_block_sparse_index, dim=-1)
+
+    # All q blocks attend to exactly k kv blocks
+    q2k_block_sparse_num = torch.full((bs, h, num_q_blocks), k, dtype=torch.int32, device=device)
+    
+    # Create the corresponding mask
+    block_sparse_mask = torch.zeros(bs, h, num_q_blocks, num_kv_blocks, dtype=torch.bool, device=device)
+    
+    # Fill in the mask based on the indices
+    for b in range(bs):
+        for head in range(h):
+            for q_idx in range(num_q_blocks):
+                kv_indices = q2k_block_sparse_index[b, head, q_idx]
+                block_sparse_mask[b, head, q_idx, kv_indices] = True
+    
+    # Create the reverse mapping (k2q)
+    # First, initialize lists to collect q indices for each kv block
+    k2q_indices_list = [[[] for _ in range(num_kv_blocks)] for _ in range(bs * h)]
+    
+    # Populate the lists based on q2k mapping
+    for b in range(bs):
+        for head in range(h):
+            flat_idx = b * h + head
+            for q_idx in range(num_q_blocks):
+                kv_indices = q2k_block_sparse_index[b, head, q_idx].tolist()
+                for kv_idx in kv_indices:
+                    k2q_indices_list[flat_idx][kv_idx].append(q_idx)
+    
+    # Find the maximum number of q blocks that attend to any kv block
+    max_q_per_kv = 0
+    for flat_idx in range(bs * h):
+        for kv_idx in range(num_kv_blocks):
+            max_q_per_kv = max(max_q_per_kv, len(k2q_indices_list[flat_idx][kv_idx]))
+    
+    # Create tensors for k2q mapping
+    k2q_block_sparse_index = torch.full((bs, h, num_kv_blocks, max_q_per_kv), -1, 
+                                        dtype=torch.int32, device=device)
+    k2q_block_sparse_num = torch.zeros((bs, h, num_kv_blocks), 
+                                       dtype=torch.int32, device=device)
+    
+    # Fill the tensors
+    for b in range(bs):
+        for head in range(h):
+            flat_idx = b * h + head
+            for kv_idx in range(num_kv_blocks):
+                q_indices = k2q_indices_list[flat_idx][kv_idx]
+                num_q = len(q_indices)
+                k2q_block_sparse_num[b, head, kv_idx] = num_q
+                if num_q > 0:
+                    k2q_block_sparse_index[b, head, kv_idx, :num_q] = torch.tensor(
+                        q_indices, dtype=torch.int32, device=device)
+                
+    return q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask
+
+def benchmark_block_sparse_attention(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, flops):
+    """Benchmark block sparse attention forward and backward passes."""
+    print("\n=== BLOCK SPARSE ATTENTION BENCHMARK ===")
+    
+    # Forward pass
+    # Warm-up run
+    o, l_vec = block_sparse_attention_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)
+    torch.cuda.synchronize()
+    
+    # Benchmark forward
+    _, fwd_time = benchmark_forward(
+        block_sparse_attention_fwd,
+        q, k, v, q2k_block_sparse_index, q2k_block_sparse_num,
+        repeats=20,
+        verbose=False,
+        desc='Block Sparse Forward'
+    )
+    
+    sparse_tflops = flops / fwd_time.mean * 1e-12
+    print(f"Block Sparse Forward - TFLOPS: {sparse_tflops:.2f}")
+    
+    # Backward pass
+    grad_output = torch.randn_like(o)
+    
+    # Warm-up runs
+    for _ in range(5):
+        block_sparse_attention_backward(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num)
+    torch.cuda.synchronize()
+    
+    # Benchmark backward
+    _, bwd_time = benchmark_forward(
+        block_sparse_attention_backward,
+        q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num,
+        repeats=20,
+        verbose=False,
+        desc='Block Sparse Backward'
+    )
+    bwd_flops = 2.5 * flops  # Approximation
+
+    sparse_bwd_tflops = bwd_flops / bwd_time.mean * 1e-12
+    print(f"Block Sparse Backward - TFLOPS: {sparse_bwd_tflops:.2f}")
+    
+    return sparse_tflops, sparse_bwd_tflops
+
+def main():
+    args = parse_arguments()
+
+    set_seed(42)
+    
+    # Extract parameters
+    batch = args.batch_size
+    head = args.num_heads
+    headdim = args.head_dim
+    
+    print(f"Block Sparse Attention Benchmark")
+    print(f"batch: {batch}, head: {head}, headdim: {headdim}")
+    
+    # Test with different sequence lengths
+    for seq_len in args.seq_lengths:
+        # Skip very long sequences if they might cause OOM
+        if seq_len > 16384 and batch > 1:
+            continue
+            
+        print("="*100)
+        print(f"\nSequence length: {seq_len}")
+        
+        # Calculate theoretical FLOPs for attention
+        flops = 4 * batch * head * headdim * seq_len * seq_len
+        
+        # Create input tensors
+        q, k, v = create_input_tensors(batch, head, seq_len, headdim)
+        
+        # Setup block sparse parameters
+        num_q_blocks = seq_len // BLOCK_M
+        num_kv_blocks = seq_len // BLOCK_N
+        
+        # Determine k value (number of kv blocks per q block)
+        topk = args.topk
+        if topk is None:
+            topk = num_kv_blocks // 10  # Default to ~90% sparsity if k is not specified
+        topk = max(1, topk)       
+        print(f"Using topk={topk} kv blocks per q block (out of {num_kv_blocks} total kv blocks)")
+        
+        # Generate block sparse pattern
+        q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, _ = generate_block_sparse_pattern(
+            batch, head, num_q_blocks, num_kv_blocks, topk, device="cuda")
+        
+        # Benchmark block sparse attention
+        sparse_fwd, sparse_bwd = benchmark_block_sparse_attention(
+            q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, flops
+        )
+        
+        # Print results
+        print("\n=== PERFORMANCE RESULTS ===")
+        print(f"Block Sparse Forward - TFLOPS: {sparse_fwd:.2f}")
+        print(f"Block Sparse Backward - TFLOPS: {sparse_bwd:.2f}")
+
+if __name__ == "__main__":
+    main()
@@ -1,6 +1,6 @@
 ### ADD TO THIS TO REGISTER NEW KERNELS
 sources = {
-    'attn': {
+    'st_attn': {
        'source_files': {
            'h100': 'st_attn/st_attn_h100.cu'  # define these source files for each GPU target desired.
        }
@@ -9,7 +9,7 @@ sources = {

 ### WHICH KERNELS DO WE WANT TO BUILD?
 # (oftentimes during development work you don't need to redefine them all.)
-kernels = ['attn']
+kernels = ['st_attn']

 ### WHICH GPU TARGET DO WE WANT TO BUILD FOR?
 target = 'h100'
@@ -0,0 +1,15 @@
+### ADD TO THIS TO REGISTER NEW KERNELS
+sources = {
+    'block_sparse': {
+        'source_files': {
+            'h100': 'vsa/block_sparse_h100.cu'
+        }
+    }
+}
+
+### WHICH KERNELS DO WE WANT TO BUILD?
+# (oftentimes during development work you don't need to redefine them all.)
+kernels = ['block_sparse']
+
+### WHICH GPU TARGET DO WE WANT TO BUILD FOR?
+target = 'h100'
@@ -1,7 +1,7 @@
 import os
 import subprocess

-from config 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

@@ -0,0 +1,76 @@
+import os
+import subprocess
+
+from csrc.attn.config_vsa import kernels, sources, target
+from setuptools import find_packages, setup
+from torch.utils.cpp_extension import BuildExtension, CUDAExtension
+
+target = target.lower()
+
+# Package metadata
+PACKAGE_NAME = "vsa"
+VERSION = "0.0.1"
+AUTHOR = "Hao AI Lab"
+DESCRIPTION = "Video Sparse Attention Kernel Used in FastVideo"
+URL = "https://github.com/hao-ai-lab/FastVideo/tree/main/csrc/attn"
+
+# Set environment variables
+tk_root = os.getenv('THUNDERKITTENS_ROOT', os.path.abspath(os.path.join(os.getcwd(), 'tk/')))
+python_include = subprocess.check_output(['python', '-c',
+                                          "import sysconfig; print(sysconfig.get_path('include'))"]).decode().strip()
+torch_include = subprocess.check_output([
+    'python', '-c',
+    "import torch; from torch.utils.cpp_extension import include_paths; print(' '.join(['-I' + p for p in include_paths()]))"
+]).decode().strip()
+print('vsa root:', tk_root)
+print('Python include:', python_include)
+print('Torch include directories:', torch_include)
+
+# CUDA flags
+cuda_flags = [
+    '-DNDEBUG', '-Xcompiler=-Wno-psabi', '-Xcompiler=-fno-strict-aliasing', '--expt-extended-lambda',
+    '--expt-relaxed-constexpr', '-forward-unknown-to-host-compiler', '--use_fast_math', '-std=c++20', '-O3',
+    '-Xnvlink=--verbose', '-Xptxas=--verbose', '-Xptxas=--warn-on-spills', f'-I{tk_root}/include',
+    f'-I{tk_root}/prototype', f'-I{python_include}', '-DTORCH_COMPILE'
+] + torch_include.split()
+cpp_flags = ['-std=c++20', '-O3']
+
+if target == 'h100':
+    cuda_flags.append('-DKITTENS_HOPPER')
+    cuda_flags.append('-arch=sm_90a')
+else:
+    raise ValueError(f'Target {target} not supported')
+
+source_files = ['vsa.cpp']
+for k in kernels:
+    if target not in sources[k]['source_files']:
+        raise KeyError(f'Target {target} not found in source files for kernel {k}')
+    if isinstance(sources[k]['source_files'][target], list):
+        source_files.extend(sources[k]['source_files'][target])
+    else:
+        source_files.append(sources[k]['source_files'][target])
+    cpp_flags.append(f'-DTK_COMPILE_{k.replace(" ", "_").upper()}')
+
+setup(name=PACKAGE_NAME,
+      version=VERSION,
+      author=AUTHOR,
+      description=DESCRIPTION,
+      url=URL,
+      packages=find_packages(),
+      ext_modules=[
+          CUDAExtension('vsa_cuda',
+                        sources=source_files,
+                        extra_compile_args={
+                            'cxx': cpp_flags,
+                            'nvcc': cuda_flags
+                        },
+                        libraries=['cuda'])
+      ],
+      cmdclass={'build_ext': BuildExtension},
+      classifiers=[
+          "Programming Language :: Python :: 3",
+          "Environment :: GPU :: NVIDIA CUDA :: 12",
+          "License :: OSI Approved :: Apache Software License",
+      ],
+      python_requires='>=3.10',
+      install_requires=["torch>=2.5.0"])
@@ -7,8 +7,7 @@

 #include <cuda_runtime.h>

-
-#ifdef TK_COMPILE_ATTN
+#ifdef TK_COMPILE_ST_ATTN
 extern torch::Tensor sta_forward(
    torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o, int kernel_t_size, int kernel_w_size, int kernel_h_size, int text_length, bool process_text, bool has_text, int kernel_aspect_ratio_flag
 ); 
@@ -17,8 +16,8 @@ extern torch::Tensor sta_forward(

 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.doc() = "Sliding Block Attention Kernels"; // optional module docstring
-
-#ifdef TK_COMPILE_ATTN
+    
+#ifdef TK_COMPILE_ST_ATTN
    m.def("sta_fwd",  torch::wrap_pybind_function(sta_forward), "sliding tile attention, assuming tile size is (6,8,8)");
 #endif
-}
+}
@@ -1,19 +1,22 @@
 import math

 import torch
-from st_attn_cuda import sta_fwd
+from torch.utils.checkpoint import detach_variable
+try:
+    from st_attn_cuda import sta_fwd
+except ImportError:
+    sta_fwd = None

-
-def sliding_tile_attention(q_all, k_all, v_all, window_size, text_length, has_text=True, img_latent_shape='30*48*80'):
+def sliding_tile_attention(q_all, k_all, v_all, window_size, text_length, has_text=True, dit_seq_shape='30x48x80'):
    seq_length = q_all.shape[2]
-    img_latent_shape_mapping = {
+    dit_seq_shape_mapping = {
        '30x48x80':1,
        '36x48x48':2,
        '18x48x80':3, 
    }
    if has_text:
        assert q_all.shape[
-            2] >= 115200, "STA currently only supports video with latent size (30, 48, 80), which is 117 frames x 768 x 1280 pixels"
+            2] >= 115200 and q_all.shape[2] <= 115456, f"Unsupported {dit_seq_shape}, current shape is {q_all.shape}, only support '30x48x80' for HunyuanVideo"
        assert q_all.shape[1] == len(window_size), "Number of heads must match the number of window sizes"
        target_size = math.ceil(seq_length / 384) * 384
        pad_size = target_size - seq_length
@@ -22,14 +25,14 @@ def sliding_tile_attention(q_all, k_all, v_all, window_size, text_length, has_te
            k_all = torch.cat([k_all, k_all[:, :, -pad_size:]], dim=2)
            v_all = torch.cat([v_all, v_all[:, :, -pad_size:]], dim=2)
    else:
-        if img_latent_shape == '36x48x48': # Stepvideo 204x768x68
+        if dit_seq_shape == '36x48x48': # Stepvideo 204x768x68
            assert q_all.shape[2] == 82944
-        elif img_latent_shape == '18x48x80': # Wan 69x768x1280
+        elif dit_seq_shape == '18x48x80': # Wan 69x768x1280
            assert q_all.shape[2] == 69120
        else:
-            raise ValueError(f"Unsupported {img_latent_shape}, current shape is {q_all.shape}, only support '36x48x48' for Stepvideo and '18x48x80' for Wan")
+            raise ValueError(f"Unsupported {dit_seq_shape}, current shape is {q_all.shape}, only support '36x48x48' for Stepvideo and '18x48x80' for Wan")

-    kernel_aspect_ratio_flag = img_latent_shape_mapping[img_latent_shape]
+    kernel_aspect_ratio_flag = dit_seq_shape_mapping[dit_seq_shape]
    hidden_states = torch.empty_like(q_all)
    # This for loop is ugly. but it is actually quite efficient. The sequence dimension alone can already oversubscribe SMs
    for head_index, (t_kernel, h_kernel, w_kernel) in enumerate(window_size):
@@ -43,4 +46,4 @@ def sliding_tile_attention(q_all, k_all, v_all, window_size, text_length, has_te
            _ = sta_fwd(q_head, k_head, v_head, o_head, t_kernel, h_kernel, w_kernel, text_length, False, has_text, kernel_aspect_ratio_flag)
    if has_text:
        _ = sta_fwd(q_all, k_all, v_all, hidden_states, 3, 3, 3, text_length, True, True, kernel_aspect_ratio_flag)
-    return hidden_states[:, :, :seq_length]
+    return hidden_states[:, :, :seq_length]
@@ -4,9 +4,17 @@
 #include <cooperative_groups.h>
 #include <iostream>
 #include <stdio.h>
+#include <c10/cuda/CUDAGuard.h>
+
+// #define CLAMP(value, min, max) ((value) < (min) ? (min) : ((value) > (max) ? (max) : (value)))
+__device__ __forceinline__ int clamp_int(int value, int min, int max) {
+    return (value < min) ? min : ((value > max) ? max : value);
+}
+// #define ABS(x) ((x) < 0 ? -(x) : (x))
+__device__ __forceinline__ int abs_int(int value) {
+    return (value < 0) ? -value : value;
+}

-#define CLAMP(value, min, max) ((value) < (min) ? (min) : ((value) > (max) ? (max) : (value)))
-#define ABS(x) ((x) < 0 ? -(x) : (x))

 constexpr int CONSUMER_WARPGROUPS = (3); 
 constexpr int PRODUCER_WARPGROUPS = (1); 
@@ -117,16 +125,16 @@ void fwd_attend_ker(const __grid_constant__ fwd_globals<D> g) {
            int qt = seq_idx / 6 / (CH * CW);
            int qh = (seq_idx / 6) % (CH * CW) / CW;
            int qw = (seq_idx / 6) % CW;
-            qt = CLAMP(qt, DT, CT-DT-1);
-            qh = CLAMP(qh, DH, CH-DH-1);
-            qw = CLAMP(qw, DW, CW-DW-1);
+            qt = clamp_int(qt, DT, CT-DT-1);
+            qh = clamp_int(qh, DH, CH-DH-1);
+            qw = clamp_int(qw, DW, CW-DW-1);
            int count = 0;
            int j = 0;
            while (count < K::stages - 1) {
                int kt = j / 3 / (CH * CW);
                int kh = (j / 3) % (CH * CW) / CW;
                int kw = (j / 3) % CW;
-                bool mask = (ABS(qt - kt) <= DT) && (ABS(qh - kh) <= DH) && (ABS(qw - kw) <= DW);
+                bool mask = (abs_int(qt - kt) <= DT) && (abs_int(qh - kh) <= DH) && (abs_int(qw - kw) <= DW);
                if (mask){
                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, j, 0};
                    tma::expect_bytes(k_smem_arrived[count], sizeof(k_tile));
@@ -167,15 +175,15 @@ void fwd_attend_ker(const __grid_constant__ fwd_globals<D> g) {
                int qt = seq_idx / 6 / (CH * CW);
                int qh = (seq_idx / 6) % (CH * CW) / CW;
                int qw = (seq_idx / 6) % CW;
-                qt = CLAMP(qt, DT, CT-DT-1);
-                qh = CLAMP(qh, DH, CH-DH-1);
-                qw = CLAMP(qw, DW, CW-DW-1);
-                int k_t_min = CLAMP(qt-DT, 0, CT-1);
-                int k_t_max = CLAMP(qt+DT, 0, CT-1);
-                int k_h_min = CLAMP(qh-DH, 0, CH-1);
-                int k_h_max = CLAMP(qh+DH, 0, CH-1);
-                int k_w_min = CLAMP(qw-DW, 0, CW-1);
-                int k_w_max = CLAMP(qw+DW, 0, CW-1);
+                qt = clamp_int(qt, DT, CT-DT-1);
+                qh = clamp_int(qh, DH, CH-DH-1);
+                qw = clamp_int(qw, DW, CW-DW-1);
+                int k_t_min = clamp_int(qt-DT, 0, CT-1);
+                int k_t_max = clamp_int(qt+DT, 0, CT-1);
+                int k_h_min = clamp_int(qh-DH, 0, CH-1);
+                int k_h_max = clamp_int(qh+DH, 0, CH-1);
+                int k_w_min = clamp_int(qw-DW, 0, CW-1);
+                int k_w_max = clamp_int(qw+DW, 0, CW-1);
                int count = 0;
                for (int kt = k_t_min; kt <= k_t_max; kt++) {
                    for (int kh = k_h_min; kh <= k_h_max; kh++) {
@@ -234,7 +242,7 @@ void fwd_attend_ker(const __grid_constant__ fwd_globals<D> g) {
            // the last three kv blocks are for text, we process them separately
            kv_iters = img_kv_blocks - 1;
        } else {
-            kv_iters = CLAMP(DT*2+1, 1, CT) * CLAMP(DH*2+1, 1, CH) * CLAMP(DW*2+1, 1, CW) * 3 - 1 ; 
+            kv_iters = clamp_int(DT*2+1, 1, CT) * clamp_int(DH*2+1, 1, CH) * clamp_int(DW*2+1, 1, CW) * 3 - 1 ; 
        }

        kittens::wait(qsmem_semaphore, 0);
@@ -415,8 +423,9 @@ sta_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o,
    float* l_ptr = reinterpret_cast<float*>(l_vec.data_ptr<float>());
    float* d_l   = reinterpret_cast<float*>(l_ptr);

-    cudaDeviceSynchronize();
-    auto stream = at::cuda::getCurrentCUDAStream().stream(); 
+    //cudadevicesynchronize();
+    const c10::cuda::OptionalCUDAGuard device_guard(q.device());
+    const cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream(); 


    if (head_dim == 128) {
@@ -442,8 +451,8 @@ sta_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o,

        globals g{qg_arg, kg_arg, vg_arg, lg_arg, og_arg, static_cast<int>(seq_len),  static_cast<int>(text_length), static_cast<int>(hr)};

-        auto mem_size = kittens::MAX_SHARED_MEMORY;
-        auto threads  = NUM_WORKERS * kittens::WARP_THREADS;
+        constexpr int mem_size = kittens::MAX_SHARED_MEMORY;
+        int threads = NUM_WORKERS * kittens::WARP_THREADS;
        if (has_text) {
            // TORCH_CHECK(seq_len % (CONSUMER_WARPGROUPS*kittens::TILE_DIM*4) == 0, "sequence length must be divisible by 192");
            dim3 grid_image(seq_len/(CONSUMER_WARPGROUPS*kittens::TILE_ROW_DIM<bf16>*4)-2, qo_heads, batch);
@@ -823,9 +832,10 @@ sta_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o,

        }
        CHECK_CUDA_ERROR(cudaGetLastError());
-        cudaStreamSynchronize(stream);
+        // cudaStreamSynchronize(stream);
    }

    return o;
-    cudaDeviceSynchronize();
+    //cudadevicesynchronize();
 }
+
@@ -0,0 +1,289 @@
+import torch
+import argparse
+from flash_attn.utils.benchmark import benchmark_forward
+from flash_attn import flash_attn_func
+from vsa import block_sparse_attention_fwd, block_sparse_attention_backward, BlockSparseAttentionFunction
+from vsa import BLOCK_M, BLOCK_N
+
+import numpy as np
+import random
+import gc
+
+def set_seed(seed: int = 42):
+    # Python random module
+    random.seed(seed)
+
+    # NumPy
+    np.random.seed(seed)
+
+    # PyTorch
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)  # if using multi-GPU
+
+
+@torch.no_grad
+def precision_metric(quant_o, fa2_o): 
+    x, xx = quant_o.float(), fa2_o.float() 
+    sim = torch.nn.functional.cosine_similarity(x.reshape(1, -1), xx.reshape(1, -1)).item()
+    l1 =   ((x - xx).abs().sum() / xx.abs().sum() ).item()
+    rmse = torch.sqrt(torch.mean((x -xx) ** 2)).item()
+
+    return sim, l1, rmse
+
+def create_input_tensors(batch, head, seq_len, headdim):
+    """Create random input tensors for attention."""
+    q = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
+    k = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
+    v = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
+    return q, k, v
+
+def generate_block_sparse_pattern(bs, h, num_q_blocks, num_kv_blocks, k, device="cuda"):
+    """
+    Generate a block sparse pattern where each q block attends to exactly k kv blocks.
+    
+    Args:
+        bs: batch size
+        h: number of heads
+        num_q_blocks: number of query blocks
+        num_kv_blocks: number of key-value blocks
+        k: number of kv blocks each q block attends to
+        device: device to create tensors on
+        
+    Returns:
+        q2k_block_sparse_index: [bs, h, num_q_blocks, k]
+            Contains the indices of kv blocks that each q block attends to.
+        q2k_block_sparse_num: [bs, h, num_q_blocks]
+            Contains the number of kv blocks that each q block attends to (all equal to k).
+        k2q_block_sparse_index: [bs, h, num_kv_blocks, num_q_blocks]
+            Contains the indices of q blocks that attend to each kv block.
+        k2q_block_sparse_num: [bs, h, num_kv_blocks]
+            Contains the number of q blocks that attend to each kv block.
+        block_sparse_mask: [bs, h, num_q_blocks, num_kv_blocks]
+            Binary mask where 1 indicates attention connection.
+    """
+    # Ensure k is not larger than num_kv_blocks
+    k = min(k, num_kv_blocks)
+    
+    # Create random scores for sampling
+    scores = torch.rand(bs, h, num_q_blocks, num_kv_blocks, device=device)
+    
+    # Get top-k indices for each q block
+    _, q2k_block_sparse_index = torch.topk(scores, k, dim=-1)
+    q2k_block_sparse_index = q2k_block_sparse_index.to(torch.int32)
+
+    # sort q2k_block_sparse_index
+    q2k_block_sparse_index, _ = torch.sort(q2k_block_sparse_index, dim=-1)
+
+    # All q blocks attend to exactly k kv blocks
+    q2k_block_sparse_num = torch.full((bs, h, num_q_blocks), k, dtype=torch.int32, device=device)
+    
+    # Create the corresponding mask
+    block_sparse_mask = torch.zeros(bs, h, num_q_blocks, num_kv_blocks, dtype=torch.bool, device=device)
+    
+    # Fill in the mask based on the indices
+    for b in range(bs):
+        for head in range(h):
+            for q_idx in range(num_q_blocks):
+                kv_indices = q2k_block_sparse_index[b, head, q_idx]
+                block_sparse_mask[b, head, q_idx, kv_indices] = True
+    
+    # Create the reverse mapping (k2q)
+    # First, initialize lists to collect q indices for each kv block
+    k2q_indices_list = [[[] for _ in range(num_kv_blocks)] for _ in range(bs * h)]
+    
+    # Populate the lists based on q2k mapping
+    for b in range(bs):
+        for head in range(h):
+            flat_idx = b * h + head
+            for q_idx in range(num_q_blocks):
+                kv_indices = q2k_block_sparse_index[b, head, q_idx].tolist()
+                for kv_idx in kv_indices:
+                    k2q_indices_list[flat_idx][kv_idx].append(q_idx)
+    
+    # Find the maximum number of q blocks that attend to any kv block
+    max_q_per_kv = 0
+    for flat_idx in range(bs * h):
+        for kv_idx in range(num_kv_blocks):
+            max_q_per_kv = max(max_q_per_kv, len(k2q_indices_list[flat_idx][kv_idx]))
+    
+    # Create tensors for k2q mapping
+    k2q_block_sparse_index = torch.full((bs, h, num_kv_blocks, max_q_per_kv), -1, 
+                                        dtype=torch.int32, device=device)
+    k2q_block_sparse_num = torch.zeros((bs, h, num_kv_blocks), 
+                                       dtype=torch.int32, device=device)
+    
+    # Fill the tensors
+    for b in range(bs):
+        for head in range(h):
+            flat_idx = b * h + head
+            for kv_idx in range(num_kv_blocks):
+                q_indices = k2q_indices_list[flat_idx][kv_idx]
+                num_q = len(q_indices)
+                k2q_block_sparse_num[b, head, kv_idx] = num_q
+                if num_q > 0:
+                    k2q_block_sparse_index[b, head, kv_idx, :num_q] = torch.tensor(
+                        q_indices, dtype=torch.int32, device=device)
+                
+    return q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask
+
+def main(args):
+    set_seed(42)
+    
+    # Extract parameters
+    batch = args.batch_size
+    head = args.num_heads
+    headdim = args.head_dim
+    num_iterations = args.num_iterations
+    
+    print(f"Block Sparse Attention Benchmark")
+    print(f"batch: {batch}, head: {head}, headdim: {headdim}, iterations: {num_iterations}")
+    
+    # Test with different sequence lengths
+    for seq_len in args.seq_lengths:
+        # Skip very long sequences if they might cause OOM
+        # if seq_len > 16384 and batch > 1:
+        #     continue
+            
+        print("="*100)
+        print(f"\nSequence length: {seq_len}")
+        
+        # Collect metrics across iterations
+        forward_metrics = {'sim': [], 'l1': [], 'rmse': []}
+        grad_q_metrics = {'sim': [], 'l1': [], 'rmse': []}
+        grad_k_metrics = {'sim': [], 'l1': [], 'rmse': []}
+        grad_v_metrics = {'sim': [], 'l1': [], 'rmse': []}
+        
+        for iter_idx in range(num_iterations):
+            if num_iterations > 1:
+                print(f"\nIteration {iter_idx+1}/{num_iterations}")
+            
+            # Create input tensors
+            q, k, v = create_input_tensors(batch, head, seq_len, headdim)
+            
+            # Setup block sparse parameters
+            num_q_blocks = seq_len // BLOCK_M
+            num_kv_blocks = seq_len // BLOCK_N
+            
+            # Determine k value (number of kv blocks per q block)
+            topk = args.topk
+            if topk is None:
+                topk = num_kv_blocks // 10  # Default to ~90% sparsity if k is not specified
+            topk = max(1, topk)
+            if iter_idx == 0:  # Only print this once
+                print(f"Using topk={topk} kv blocks per q block (out of {num_kv_blocks} total kv blocks)")
+
+            # Generate block sparse pattern
+            q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask = generate_block_sparse_pattern(
+                batch, head, num_q_blocks, num_kv_blocks, topk, device="cuda")
+
+            # expand block_sparse_mask to full mask
+            block_mask_expanded = block_sparse_mask.unsqueeze(-1).unsqueeze(-2)  # [b, h, num_q_blocks, num_kv_blocks, 1, 1]
+            block_mask_expanded = block_mask_expanded.expand(-1, -1, -1, -1, BLOCK_M, BLOCK_N)  # [b, h, num_q_blocks, num_kv_blocks, BLOCK_M, BLOCK_N]
+            full_mask = block_mask_expanded.permute(0, 1, 2, 4, 3, 5).reshape(batch, head, seq_len, seq_len)
+
+            q.requires_grad = True
+            k.requires_grad = True
+            v.requires_grad = True
+
+
+            # testing forward
+            o = BlockSparseAttentionFunction.apply(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num)
+            del q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask, block_mask_expanded
+            grad_o = torch.randn_like(o)
+            o.backward(grad_o)
+            # clear memory
+            q_sdpa = q.detach().clone()
+            k_sdpa = k.detach().clone()
+            v_sdpa = v.detach().clone()
+            q_sdpa.requires_grad = True
+            k_sdpa.requires_grad = True
+            v_sdpa.requires_grad = True
+            q.data = torch.empty(0, device=q.device)
+            k.data = torch.empty(0, device=k.device)
+            v.data = torch.empty(0, device=v.device)
+            torch.cuda.empty_cache()
+
+            o_sdpa = torch.nn.functional.scaled_dot_product_attention(q_sdpa, k_sdpa, v_sdpa, attn_mask=full_mask)
+
+
+            sim, l1, rmse = precision_metric(o, o_sdpa)
+            assert sim > 0.9999, f"SSIM too low: {sim}"
+            assert l1 < 8e-5, f"l1 too large: {l1}"
+            assert rmse < 2e-5, f"RMSE too large: {rmse}"
+            forward_metrics['sim'].append(sim)
+            forward_metrics['l1'].append(l1)
+            forward_metrics['rmse'].append(rmse)
+
+            print(f"block_sparse_attention_fwd vs torch.nn.functional.scaled_dot_product_attention:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
+
+            # test backward
+            o_sdpa.backward(grad_o)
+
+            sim, l1, rmse = precision_metric(q.grad, q_sdpa.grad)
+            # Error bounds collected on H100
+            assert sim > 0.9999, f"SSIM too low: {sim}"
+            assert l1 < 4e-3, f"l1 too large: {l1}"
+            assert rmse < 3e-4, f"RMSE too large: {rmse}"
+            grad_q_metrics['sim'].append(sim)
+            grad_q_metrics['l1'].append(l1)
+            grad_q_metrics['rmse'].append(rmse)
+            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_q:\nsim: {sim}, l1: {l1}, rmse: {rmse}")        
+            
+            sim, l1, rmse = precision_metric(k.grad, k_sdpa.grad)
+            assert sim > 0.9999, f"SSIM too low: {sim}"
+            assert l1 < 4e-3, f"l1 too large: {l1}"
+            assert rmse < 2e-4, f"RMSE too large: {rmse}"
+            grad_k_metrics['sim'].append(sim)
+            grad_k_metrics['l1'].append(l1)
+            grad_k_metrics['rmse'].append(rmse)
+            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_k:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
+            
+            sim, l1, rmse = precision_metric(v.grad, v_sdpa.grad)
+            assert sim > 0.9999, f"SSIM too low: {sim}"
+            assert l1 < 1e-4, f"l1 too large: {l1}"
+            assert rmse < 2e-5, f"RMSE too large: {rmse}"
+            grad_v_metrics['sim'].append(sim)
+            grad_v_metrics['l1'].append(l1)
+            grad_v_metrics['rmse'].append(rmse)
+            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_v:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
+            
+            del o, o_sdpa, grad_o, q_sdpa, k_sdpa, v_sdpa
+            gc.collect()
+            torch.cuda.empty_cache()
+
+        # Print summary statistics if multiple iterations were run
+        if num_iterations > 1:
+            print("\n" + "="*50)
+            print(f"Summary Statistics (over {num_iterations} iterations):")
+            
+            print("\nForward metrics:")
+            print(f"Similarity: mean={np.mean(forward_metrics['sim']):.6f}, std={np.std(forward_metrics['sim']):.6f}, min={np.min(forward_metrics['sim']):.6f}")
+            print(f"L1 error: mean={np.mean(forward_metrics['l1']):.6f}, std={np.std(forward_metrics['l1']):.6f}, max={np.max(forward_metrics['l1']):.6f}")
+            print(f"RMSE: mean={np.mean(forward_metrics['rmse']):.6f}, std={np.std(forward_metrics['rmse']):.6f}, max={np.max(forward_metrics['rmse']):.6f}")
+            
+            print("\nGradient Q metrics:")
+            print(f"Similarity: mean={np.mean(grad_q_metrics['sim']):.6f}, std={np.std(grad_q_metrics['sim']):.6f}, min={np.min(grad_q_metrics['sim']):.6f}")
+            print(f"L1 error: mean={np.mean(grad_q_metrics['l1']):.6f}, std={np.std(grad_q_metrics['l1']):.6f}, max={np.max(grad_q_metrics['l1']):.6f}")
+            print(f"RMSE: mean={np.mean(grad_q_metrics['rmse']):.6f}, std={np.std(grad_q_metrics['rmse']):.6f}, max={np.max(grad_q_metrics['rmse']):.6f}")
+            
+            print("\nGradient K metrics:")
+            print(f"Similarity: mean={np.mean(grad_k_metrics['sim']):.6f}, std={np.std(grad_k_metrics['sim']):.6f}, min={np.min(grad_k_metrics['sim']):.6f}")
+            print(f"L1 error: mean={np.mean(grad_k_metrics['l1']):.6f}, std={np.std(grad_k_metrics['l1']):.6f}, max={np.max(grad_k_metrics['l1']):.6f}")
+            print(f"RMSE: mean={np.mean(grad_k_metrics['rmse']):.6f}, std={np.std(grad_k_metrics['rmse']):.6f}, max={np.max(grad_k_metrics['rmse']):.6f}")
+            
+            print("\nGradient V metrics:")
+            print(f"Similarity: mean={np.mean(grad_v_metrics['sim']):.6f}, std={np.std(grad_v_metrics['sim']):.6f}, min={np.min(grad_v_metrics['sim']):.6f}")
+            print(f"L1 error: mean={np.mean(grad_v_metrics['l1']):.6f}, std={np.std(grad_v_metrics['l1']):.6f}, max={np.max(grad_v_metrics['l1']):.6f}")
+            print(f"RMSE: mean={np.mean(grad_v_metrics['rmse']):.6f}, std={np.std(grad_v_metrics['rmse']):.6f}, max={np.max(grad_v_metrics['rmse']):.6f}")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Benchmark Block Sparse Attention')
+    parser.add_argument('--batch_size', type=int, default=4, help='Batch size')
+    parser.add_argument('--num_heads', type=int, default=6, help='Number of heads')
+    parser.add_argument('--head_dim', type=int, default=128, help='Head dimension')
+    parser.add_argument('--topk', type=int, default=64, help='Number of kv blocks each q block attends to')
+    parser.add_argument('--seq_lengths', type=int, nargs='+', default=[29120], help='Sequence lengths to benchmark')
+    parser.add_argument('--num_iterations', type=int, default=50, help='Number of test iterations to run')
+    args = parser.parse_args()
+    main(args)
@@ -0,0 +1,136 @@
+import torch
+from tqdm import tqdm
+import matplotlib.pyplot as plt
+import numpy as np
+
+def pytorch_test(Q, K, V, dO):
+    q_ = Q.to(torch.float64).requires_grad_()
+    k_ = K.to(torch.float64).requires_grad_()
+    v_ = V.to(torch.float64).requires_grad_()
+    dO_ = dO.to(torch.float64)
+    
+    # manual pytorch implementation of scaled dot product attention
+    QK = torch.matmul(q_, k_.transpose(-2, -1))
+    QK /= (q_.size(-1) ** 0.5)
+    
+    # Causal mask removed since causal is always false
+    
+    QK = torch.nn.functional.softmax(QK, dim=-1)
+    output = torch.matmul(QK, v_)
+    
+    output.backward(dO_)
+    
+    q_grad = q_.grad
+    k_grad = k_.grad
+    v_grad = v_.grad
+    
+    return output, q_grad, k_grad, v_grad
+
+def fa2_test(Q, K, V, dO):
+    Q.requires_grad = True
+    K.requires_grad = True
+    V.requires_grad = True
+    output = torch.nn.functional.scaled_dot_product_attention(Q, K, V, is_causal=False)
+    output.backward(dO)
+    
+    return output, Q.grad, K.grad, V.grad
+
+def generate_tensor(shape, mean, std, dtype, device):
+    tensor = torch.randn(shape, dtype=dtype, device=device)
+
+    magnitude = torch.norm(tensor, dim=-1, keepdim=True)
+    scaled_tensor = tensor * (torch.randn(magnitude.shape, dtype=dtype, device=device) * std + mean) / magnitude
+    
+    return scaled_tensor.contiguous()
+
+def check_correctness(b, h, n, d, mean, std, num_iterations=100, error_mode='all', test_mode='forward_backward'):
+    results = {
+        'FA2 vs PT': {'sum_diff': 0, 'sum_abs': 0, 'max_diff': 0},
+    }
+
+    for _ in range(num_iterations):
+        torch.manual_seed(0)
+        
+        Q  = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+        K  = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+        V  = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+        dO = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+        
+        pt_o, pt_qg, pt_kg, pt_vg = pytorch_test(Q, K, V, dO)
+        fa2_o, fa2_qg, fa2_kg, fa2_vg = fa2_test(Q, K, V, dO)
+        
+        if test_mode == 'forward_only':
+            tensors_fa2_pt = [(pt_o, fa2_o)]
+        else:  # 'forward_backward'
+            if error_mode == 'output':
+                tensors_fa2_pt = [(pt_o, fa2_o)]
+            elif error_mode == 'backward':
+                tensors_fa2_pt = [(pt_qg, fa2_qg),
+                               (pt_kg, fa2_kg),
+                               (pt_vg, fa2_vg)]
+            else:  # 'all'
+                tensors_fa2_pt = [(pt_o, fa2_o),
+                               (pt_qg, fa2_qg),
+                               (pt_kg, fa2_kg),
+                               (pt_vg, fa2_vg)]
+                
+        for pt, fa2 in tensors_fa2_pt:
+            diff = pt - fa2
+            abs_diff = torch.abs(diff)
+            results['FA2 vs PT']['sum_diff'] += torch.sum(abs_diff).item()
+            results['FA2 vs PT']['sum_abs'] += torch.sum(torch.abs(pt)).item()
+            results['FA2 vs PT']['max_diff'] = max(results['FA2 vs PT']['max_diff'], torch.max(abs_diff).item())
+                
+        torch.cuda.empty_cache()
+
+    # Calculate total elements based on test mode and error mode
+    if test_mode == 'forward_only':
+        total_elements = b * h * n * d * num_iterations
+    else:  # 'forward_backward'
+        total_elements = b * h * n * d * num_iterations * (1 if error_mode == 'output' else 3 if error_mode == 'backward' else 4)
+    
+    for name, data in results.items():
+        avg_diff = data['sum_diff'] / total_elements
+        max_diff = data['max_diff']
+        results[name] = {'avg_diff': avg_diff, 'max_diff': max_diff}
+
+    return results
+
+def generate_error_tables(b, h, d, mean, std, error_mode='all', test_mode='forward_backward'):
+    seq_lengths = [768 * (2**i) for i in range(1)]
+
+    print(f"\n{'='*80}")
+    print(f"ATTENTION ERROR COMPARISON TABLE (b={b}, h={h}, d={d}, mean={mean}, std={std})")
+    print(f"Mode: {error_mode}, Test: {test_mode}")
+    print(f"{'='*80}")
+    
+    # Print header
+    print(f"{'Seq Length':<12} | {'FA2 vs PT Avg':<15} | {'FA2 vs PT Max':<15}")
+    print(f"{'-'*12} | {'-'*15} | {'-'*15}")
+    
+    for n in seq_lengths:
+        results = check_correctness(b, h, n, d, mean, std, error_mode=error_mode, test_mode=test_mode)
+        
+        fa2_pt_avg = results['FA2 vs PT']['avg_diff']
+        fa2_pt_max = results['FA2 vs PT']['max_diff']
+        
+        # Print row
+        print(f"{n:<12} | {fa2_pt_avg:<15.6e} | {fa2_pt_max:<15.6e}")
+    
+    print(f"{'='*80}\n")
+
+# fix random seed
+torch.manual_seed(0)
+
+# Example usage
+b, h, d = 2, 2, 64
+mean = 1e-1
+std = 10
+
+# Test forward only
+generate_error_tables(b, h, d, mean, std, error_mode='output', test_mode='forward_only')
+
+# Test forward and backward
+generate_error_tables(b, h, d, mean, std, error_mode='all', test_mode='forward_backward')
+
+print("Attention error comparison completed.")
@@ -0,0 +1,175 @@
+import torch
+from flash_attn_interface import flash_attn_func
+from st_attn import mha_forward, mha_backward
+import random
+from tqdm import tqdm
+import matplotlib.pyplot as plt
+import numpy as np
+
+def pytorch_test(Q, K, V, dO):
+    q_ = Q.to(torch.float64).requires_grad_()
+    k_ = K.to(torch.float64).requires_grad_()
+    v_ = V.to(torch.float64).requires_grad_()
+    dO_ = dO.to(torch.float64)
+    
+    # manual pytorch implementation of scaled dot product attention
+    QK = torch.matmul(q_, k_.transpose(-2, -1))
+    QK /= (q_.size(-1) ** 0.5)
+    
+    # Causal mask removed since causal is always false
+    
+    QK = torch.nn.functional.softmax(QK, dim=-1)
+    output = torch.matmul(QK, v_)
+    
+    output.backward(dO_)
+    
+    q_grad = q_.grad
+    k_grad = k_.grad
+    v_grad = v_.grad
+    
+    return output, q_grad, k_grad, v_grad
+
+def fa2_test(Q, K, V, dO):
+    Q.requires_grad = True
+    K.requires_grad = True
+    V.requires_grad = True
+    output = torch.nn.functional.scaled_dot_product_attention(Q, K, V, is_causal=False)
+    output.backward(dO)
+    
+    return output, Q.grad, K.grad, V.grad
+
+
+def mha_kernel_test(Q, K, V, dO, mode): 
+    Q.requires_grad = True
+    K.requires_grad = True
+    V.requires_grad = True
+    
+    o, l_vec = mha_forward(Q, K, V)
+    
+    if mode == 'forward_only':
+        return o, None, None, None
+    else:  # 'forward_backward'
+        qg, kg, vg = mha_backward(Q, K, V, o, l_vec, dO)
+        return o, qg, kg, vg
+
+def generate_tensor(shape, mean, std, dtype, device):
+    tensor = torch.randn(shape, dtype=dtype, device=device)
+
+    magnitude = torch.norm(tensor, dim=-1, keepdim=True)
+    scaled_tensor = tensor * (torch.randn(magnitude.shape, dtype=dtype, device=device) * std + mean) / magnitude
+    
+    return scaled_tensor.contiguous()
+
+def check_correctness(b, h, n, d, mean, std, num_iterations=100, error_mode='all', test_mode='forward_backward'):
+    results = {
+        'MHA vs PT': {'sum_diff': 0, 'sum_abs': 0, 'max_diff': 0},
+        'FA2 vs PT': {'sum_diff': 0, 'sum_abs': 0, 'max_diff': 0},
+    }
+
+    for _ in range(num_iterations):
+        torch.manual_seed(0)
+        
+        Q  = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+        K  = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+        V  = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+        dO = generate_tensor((b, h, n, d), mean, std, torch.bfloat16, 'cuda')
+        
+        pt_o, pt_qg, pt_kg, pt_vg = pytorch_test(Q, K, V, dO)
+        fa2_o, fa2_qg, fa2_kg, fa2_vg = fa2_test(Q, K, V, dO)
+        
+        if test_mode == 'forward_only':
+            mha_o, _, _, _ = mha_kernel_test(Q, K, V, dO, 'forward_only')
+            tensors_mha_pt = [(pt_o, mha_o)]
+            tensors_fa2_pt = [(pt_o, fa2_o)]
+        else:  # 'forward_backward'
+            mha_o, mha_qg, mha_kg, mha_vg = mha_kernel_test(Q, K, V, dO, 'forward_backward')
+            
+            if error_mode == 'output':
+                tensors_mha_pt = [(pt_o, mha_o)]
+                tensors_fa2_pt = [(pt_o, fa2_o)]
+            elif error_mode == 'backward':
+                tensors_mha_pt = [(pt_qg, mha_qg),
+                               (pt_kg, mha_kg),
+                               (pt_vg, mha_vg)]
+                tensors_fa2_pt = [(pt_qg, fa2_qg),
+                               (pt_kg, fa2_kg),
+                               (pt_vg, fa2_vg)]
+            else:  # 'all'
+                tensors_mha_pt = [(pt_o, mha_o),
+                               (pt_qg, mha_qg),
+                               (pt_kg, mha_kg),
+                               (pt_vg, mha_vg)]
+                tensors_fa2_pt = [(pt_o, fa2_o),
+                               (pt_qg, fa2_qg),
+                               (pt_kg, fa2_kg),
+                               (pt_vg, fa2_vg)]
+                
+        for pt, mha in tensors_mha_pt:
+            diff = pt - mha
+            abs_diff = torch.abs(diff)
+            results['MHA vs PT']['sum_diff'] += torch.sum(abs_diff).item()
+            results['MHA vs PT']['sum_abs'] += torch.sum(torch.abs(pt)).item()
+            results['MHA vs PT']['max_diff'] = max(results['MHA vs PT']['max_diff'], torch.max(abs_diff).item())
+                
+        for pt, fa2 in tensors_fa2_pt:
+            diff = pt - fa2
+            abs_diff = torch.abs(diff)
+            results['FA2 vs PT']['sum_diff'] += torch.sum(abs_diff).item()
+            results['FA2 vs PT']['sum_abs'] += torch.sum(torch.abs(pt)).item()
+            results['FA2 vs PT']['max_diff'] = max(results['FA2 vs PT']['max_diff'], torch.max(abs_diff).item())
+                
+        torch.cuda.empty_cache()
+
+    # Calculate total elements based on test mode and error mode
+    if test_mode == 'forward_only':
+        total_elements = b * h * n * d * num_iterations
+    else:  # 'forward_backward'
+        total_elements = b * h * n * d * num_iterations * (1 if error_mode == 'output' else 3 if error_mode == 'backward' else 4)
+    
+    for name, data in results.items():
+        avg_diff = data['sum_diff'] / total_elements
+        max_diff = data['max_diff']
+        results[name] = {'avg_diff': avg_diff, 'max_diff': max_diff}
+
+    return results
+
+def generate_error_tables(b, h, d, mean, std, error_mode='all', test_mode='forward_backward'):
+    seq_lengths = [768 * (2**i) for i in range(1)]
+
+    print(f"\n{'='*80}")
+    print(f"MHA ERROR COMPARISON TABLE (b={b}, h={h}, d={d}, mean={mean}, std={std})")
+    print(f"Mode: {error_mode}, Test: {test_mode}")
+    print(f"{'='*80}")
+    
+    # Print header
+    print(f"{'Seq Length':<12} | {'MHA vs PT Avg':<15} | {'MHA vs PT Max':<15} | {'FA2 vs PT Avg':<15} | {'FA2 vs PT Max':<15}")
+    print(f"{'-'*12} | {'-'*15} | {'-'*15} | {'-'*15} | {'-'*15}")
+    
+    for n in seq_lengths:
+        results = check_correctness(b, h, n, d, mean, std, error_mode=error_mode, test_mode=test_mode)
+        
+        mha_pt_avg = results['MHA vs PT']['avg_diff']
+        mha_pt_max = results['MHA vs PT']['max_diff']
+        fa2_pt_avg = results['FA2 vs PT']['avg_diff']
+        fa2_pt_max = results['FA2 vs PT']['max_diff']
+        
+        # Print row
+        print(f"{n:<12} | {mha_pt_avg:<15.6e} | {mha_pt_max:<15.6e} | {fa2_pt_avg:<15.6e} | {fa2_pt_max:<15.6e}")
+    
+    print(f"{'='*80}\n")
+
+# fix random seed
+torch.manual_seed(0)
+
+# Example usage
+b, h, d = 2, 2, 64
+mean = 1e-1
+std = 10
+
+# Test forward only
+generate_error_tables(b, h, d, mean, std, error_mode='output', test_mode='forward_only')
+
+# Test forward and backward
+generate_error_tables(b, h, d, mean, std, error_mode='all', test_mode='forward_backward')
+
+print("MHA attention error comparison completed.")
@@ -9,14 +9,14 @@ flex_attention = torch.compile(flex_attention, dynamic=False)


 def flex_test(Q, K, V, kernel_size):
-    mask = get_sliding_tile_attention_mask(kernel_size, (6, 8, 8), (36, 48, 48), 39, 'cuda', 0)
+    mask = get_sliding_tile_attention_mask(kernel_size, (6, 8, 8), (18, 48, 80), 0, 'cuda', 0)
    output = flex_attention(Q, K, V, block_mask=mask)

    return output


 def h100_fwd_kernel_test(Q, K, V, kernel_size):
-    o = sliding_tile_attention(Q, K, V, [kernel_size] * 24, 39, False)
+    o = sliding_tile_attention(Q, K, V, [kernel_size] * 24, 0, False, '18x48x80')
    return o


@@ -37,7 +37,7 @@ def check_correctness(b, h, n, d, causal, mean, std, num_iterations=50, error_mo
            'max_diff': 0
        },
    }
-    kernel_size_ls = [(6, 1, 6), (6, 6, 1)]
+    kernel_size_ls = [(3, 3, 5), (3, 1, 10)]
    from tqdm import tqdm
    for kernel_size in tqdm(kernel_size_ls):
        for _ in range(num_iterations):
@@ -72,25 +72,16 @@ def check_correctness(b, h, n, d, causal, mean, std, num_iterations=50, error_mo
    return results


-def generate_error_graphs(b, h, d, causal, mean, std, error_mode='all'):
-    seq_lengths = [82944]
-
-    tk_avg_errors, tk_max_errors = [], []
-
-    for n in tqdm(seq_lengths, desc="Generating error data"):
-        results = check_correctness(b, h, n, d, causal, mean, std, error_mode=error_mode)
-
-        tk_avg_errors.append(results['TK vs FLEX']['avg_diff'])
-        tk_max_errors.append(results['TK vs FLEX']['max_diff'])
-
-
 # Example usage
 b, h, d = 2, 24, 128
+n = 69120  # Sequence length
 causal = False
 mean = 1e-1
 std = 10

-for mode in ['output']:
-    generate_error_graphs(b, h, d, causal, mean, std, error_mode=mode)
-
-print("Error graphs generated and saved for all modes.")
+# Run correctness check directly
+results = check_correctness(b, h, n, d, causal, mean, std, error_mode='output')
+assert results['TK vs FLEX']['avg_diff'] < 3e-6, f"Average difference: {results['TK vs FLEX']['avg_diff']} is too large"
+assert results['TK vs FLEX']['max_diff'] < 4e-2, f"Maximum difference: {results['TK vs FLEX']['max_diff']} is too large"
+print(f"Average difference: {results['TK vs FLEX']['avg_diff']}")
+print(f"Maximum difference: {results['TK vs FLEX']['max_diff']}")
@@ -0,0 +1,27 @@
+#include <torch/extension.h>
+#include <ATen/ATen.h>
+
+#include <vector>
+#include <cuda_fp16.h>
+#include <cuda_bf16.h>
+
+#include <cuda_runtime.h>
+
+#ifdef TK_COMPILE_BLOCK_SPARSE
+extern std::vector<torch::Tensor> block_sparse_attention_forward(
+    torch::Tensor q, torch::Tensor k, torch::Tensor v,  torch::Tensor q2k_block_sparse_index, torch::Tensor q2k_block_sparse_num
+); 
+extern std::vector<torch::Tensor> block_sparse_attention_backward(
+    torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o, torch::Tensor l_vec, torch::Tensor og, torch::Tensor k2q_block_sparse_index, torch::Tensor k2q_block_sparse_num
+);
+#endif
+
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.doc() = "Video Sparse Attention Kernels"; // optional module docstring
+
+#ifdef TK_COMPILE_BLOCK_SPARSE
+    m.def("block_sparse_fwd",  torch::wrap_pybind_function(block_sparse_attention_forward), "block sparse attention");
+    m.def("block_sparse_bwd",  torch::wrap_pybind_function(block_sparse_attention_backward), "block sparse attention backward");
+#endif
+}
@@ -0,0 +1,470 @@
+import math
+
+import torch
+from torch.utils.checkpoint import detach_variable
+from typing import Tuple
+try:
+    from vsa_cuda import block_sparse_fwd, block_sparse_bwd
+except ImportError:
+    block_sparse_fwd = None
+    block_sparse_bwd = None
+
+
+BLOCK_M = 64
+BLOCK_N = 64
+
+def video_sparse_attn(q, k, v, topk, block_size, compress_attn_weight=None):
+    """
+    q: [batch_size, num_heads, seq_len, head_dim]
+    k: [batch_size, num_heads, seq_len, head_dim]
+    v: [batch_size, num_heads, seq_len, head_dim]
+    topk: int
+    block_size: int or tuple of 3 ints
+    video_shape: tuple of (T, H, W)
+    compress_attn_weight: [batch_size, num_heads, seq_len, head_dim]
+    select_attn_weight: [batch_size, num_heads, seq_len, head_dim]
+    
+    V1 of sparse attention. Include compress attn and sparse attn branch, use average pooling to compress. 
+    Assume q, k, v is flattened in this way: [batch_size, num_heads, T//block_size[0], H//block_size[1], W//block_size[2], block_size[0], block_size[1], block_size[2]]
+    """
+
+    if isinstance(block_size, int):
+        block_size = (block_size, block_size, block_size)
+
+    block_elements = block_size[0] * block_size[1] * block_size[2]
+    assert block_elements % 64 == 0 and block_elements >= 64
+    assert q.shape[2] % block_elements == 0
+    batch_size, num_heads, seq_len, head_dim = q.shape
+    # compress attn
+    q_compress = q.view(batch_size, num_heads, seq_len // block_elements,
+                        block_elements, head_dim).mean(dim=3)
+    k_compress = k.view(batch_size, num_heads, seq_len // block_elements,
+                        block_elements, head_dim).mean(dim=3)
+    v_compress = v.view(batch_size, num_heads, seq_len // block_elements,
+                        block_elements, head_dim).mean(dim=3)
+
+    output_compress, block_attn_score = torch_attention(q_compress, k_compress,
+                                                        v_compress)
+
+    output_compress = output_compress.view(batch_size, num_heads,
+                                           seq_len // block_elements, 1,
+                                           head_dim)
+    output_compress = output_compress.repeat(1, 1, 1, block_elements,
+                                             1).view(batch_size, num_heads,
+                                                     seq_len, head_dim)
+
+    q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num = generate_topk_block_sparse_pattern(
+        block_attn_score, topk)
+
+    output_select = block_sparse_attn(q, k, v, q2k_block_sparse_index,
+                                      q2k_block_sparse_num,
+                                      k2q_block_sparse_index,
+                                      k2q_block_sparse_num)
+
+    if compress_attn_weight is not None:
+        final_output = output_compress * compress_attn_weight + output_select
+    else:
+        final_output = output_compress + output_select
+    return final_output
+
+def torch_attention(q, k, v) -> Tuple[torch.Tensor, torch.Tensor]:
+    QK = torch.matmul(q, k.transpose(-2, -1))
+    QK /= (q.size(-1)**0.5)
+
+    # Causal mask removed since causal is always false
+
+    QK = torch.nn.functional.softmax(QK, dim=-1)
+    output = torch.matmul(QK, v)
+    return output, QK
+
+def generate_topk_block_sparse_pattern(block_attn_score: torch.Tensor,
+                                       topk: int):
+    """
+    Generate a block sparse pattern where each q block attends to exactly topk kv blocks,
+    based on the provided attention scores.
+    
+    Args:
+        block_attn_score: [bs, h, num_q_blocks, num_kv_blocks]
+            Attention scores between query and key blocks
+        topk: int
+            Number of kv blocks each q block attends to
+        
+    Returns:
+        q2k_block_sparse_index: [bs, h, num_q_blocks, topk]
+            Contains the indices of kv blocks that each q block attends to.
+        q2k_block_sparse_num: [bs, h, num_q_blocks]
+            Contains the number of kv blocks that each q block attends to (all equal to topk).
+        k2q_block_sparse_index: [bs, h, num_kv_blocks, max_q_per_kv]
+            Contains the indices of q blocks that attend to each kv block.
+        k2q_block_sparse_num: [bs, h, num_kv_blocks]
+            Contains the number of q blocks that attend to each kv block.
+    """
+    device = block_attn_score.device
+    # Extract dimensions from block_attn_score
+    bs, h, num_q_blocks, num_kv_blocks = block_attn_score.shape
+
+    sorted_result = torch.sort(block_attn_score, dim=-1, descending=True)
+
+    sorted_indice = sorted_result.indices
+
+    q2k_block_sparse_index, _ = torch.sort(sorted_indice[:, :, :, :topk],
+                                           dim=-1)
+    q2k_block_sparse_index = q2k_block_sparse_index.to(dtype=torch.int32)
+    q2k_block_sparse_num = torch.full((bs, h, num_q_blocks),
+                                      topk,
+                                      device=device,
+                                      dtype=torch.int32)
+
+    block_map = topk_index_to_map(q2k_block_sparse_index,
+                                  num_kv_blocks,
+                                  transpose_map=True)
+    k2q_block_sparse_index, k2q_block_sparse_num = map_to_index(
+        block_map.transpose(2, 3))
+
+    return q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num
+
+@torch._dynamo.disable
+def block_sparse_attn(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num):
+    """
+    Differentiable block sparse attention function.
+    
+    Args:
+        q: Query tensor [batch_size, num_heads, seq_len_q, head_dim]
+        k: Key tensor [batch_size, num_heads, seq_len_kv, head_dim]
+        v: Value tensor [batch_size, num_heads, seq_len_kv, head_dim]
+        q2k_block_sparse_index: Indices for query-to-key sparse blocks
+        q2k_block_sparse_num: Number of sparse blocks for each query block
+        k2q_block_sparse_index: Indices for key-to-query sparse blocks (for backward pass)
+        k2q_block_sparse_num: Number of sparse blocks for each key block (for backward pass)
+    
+    Returns:
+        output: Attention output tensor [batch_size, num_heads, seq_len_q, head_dim]
+    """
+    return BlockSparseAttentionFunction.apply(
+        q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num
+    )
+
+def block_sparse_attention_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num):
+    """
+    block_sparse_mask: [bs, h, num_q_blocks, num_kv_blocks]. 
+        [*, *, i, j] = 1 means the i-th q block should attend to the j-th kv block.
+    """
+    # assert all elements in q2k_block_sparse_num can be devisible by 2
+    o, lse = block_sparse_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)
+    return o, lse
+
+def block_sparse_attention_backward(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num):
+    grad_output = grad_output.contiguous()
+    grad_q, grad_k, grad_v = block_sparse_bwd(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num)
+    return grad_q, grad_k, grad_v
+
+## pytorch sdpa version of block sparse ##
+import triton
+import triton.language as tl
+
+@triton.jit
+def index_to_mask_kernel(
+    q2k_block_sparse_index_ptr,
+    q2k_block_sparse_num_ptr,
+    mask_ptr,
+    batch_size: tl.constexpr,
+    num_heads: tl.constexpr,
+    num_q_blocks: tl.constexpr,
+    num_k_blocks: tl.constexpr,
+    max_kv_blocks: tl.constexpr,
+    BLOCK_Q: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+):
+    bh, q, id = tl.program_id(0).to(tl.int64), tl.program_id(1).to(tl.int64), tl.program_id(2).to(tl.int64)
+    b = bh // num_heads
+    h = bh % num_heads
+
+    num_valid_blocks = tl.load(q2k_block_sparse_num_ptr + b * num_heads * num_q_blocks + h * num_q_blocks + q)
+
+    if num_valid_blocks <= id:
+        return
+    k = tl.load(q2k_block_sparse_index_ptr + b * num_heads * num_q_blocks * max_kv_blocks + h * num_q_blocks * max_kv_blocks + q * max_kv_blocks + id)
+
+    full_mask = (tl.arange(0, BLOCK_Q)[:, None] < BLOCK_Q) & (tl.arange(0, BLOCK_K)[None, :] < BLOCK_K)
+
+    q_lengths = num_q_blocks * BLOCK_Q
+    k_lengths = num_k_blocks * BLOCK_K
+    mask_ptr_base = mask_ptr + b * num_heads * q_lengths * k_lengths + h * q_lengths * k_lengths + q * BLOCK_Q * k_lengths + k * BLOCK_K
+
+    tl.store(mask_ptr_base + tl.arange(0, BLOCK_Q)[:, None] * k_lengths + tl.arange(0, BLOCK_K)[None, :], full_mask)
+
+def index_to_mask(q2k_block_sparse_index, q2k_block_sparse_num, BLOCK_Q, BLOCK_K, num_k_blocks):
+    """
+    Convert block sparse indices to a mask.
+    
+    Args:
+        q2k_block_sparse_index: Indices for query-to-key sparse blocks
+        q2k_block_sparse_num: Number of sparse blocks for each query block
+    
+    Returns:
+        mask: Block sparse mask tensor
+    """
+    batch_size, num_heads, num_q_blocks, max_kv_blocks = q2k_block_sparse_index.shape
+    assert q2k_block_sparse_num.shape == (batch_size, num_heads, num_q_blocks)
+
+    mask = torch.zeros((batch_size, num_heads, num_q_blocks * BLOCK_Q, num_k_blocks * BLOCK_K), dtype=torch.bool, device=q2k_block_sparse_index.device)
+
+    grid = (batch_size * num_heads, num_q_blocks, max_kv_blocks)
+    index_to_mask_kernel[grid](
+        q2k_block_sparse_index,
+        q2k_block_sparse_num,
+        mask,
+        batch_size,
+        num_heads,
+        num_q_blocks,
+        num_k_blocks,
+        max_kv_blocks,
+        BLOCK_Q=BLOCK_Q,
+        BLOCK_K=BLOCK_K,
+    )
+
+    return mask
+
+@triton.jit
+def topk_index_to_map_kernel(
+    map_ptr,
+    index_ptr,
+    map_bs_stride,
+    map_h_stride,
+    map_q_stride,
+    map_kv_stride,
+    index_bs_stride,
+    index_h_stride,
+    index_q_stride,
+    index_kv_stride,
+    topk: tl.constexpr,
+):
+    b, h, q = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    index_ptr_base = index_ptr + b * index_bs_stride + h * index_h_stride + q * index_q_stride
+    map_ptr_base = map_ptr + b * map_bs_stride + h * map_h_stride + q * map_q_stride
+
+    for i in tl.static_range(topk):
+        index = tl.load(index_ptr_base + i * index_kv_stride)
+        tl.store(map_ptr_base + index * map_kv_stride, 1.0)
+
+@triton.jit
+def map_to_index_kernel(
+    map_ptr,
+    index_ptr,
+    index_num_ptr,
+    map_bs_stride,
+    map_h_stride,
+    map_q_stride,
+    map_kv_stride,
+    index_bs_stride,
+    index_h_stride,
+    index_q_stride,
+    index_kv_stride,
+    index_num_bs_stride,
+    index_num_h_stride,
+    index_num_q_stride,
+    num_kv_blocks: tl.constexpr,
+):
+    b, h, q = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    index_ptr_base = index_ptr + b * index_bs_stride + h * index_h_stride + q * index_q_stride
+    map_ptr_base = map_ptr + b * map_bs_stride + h * map_h_stride + q * map_q_stride
+
+    num = 0
+    for i in tl.static_range(num_kv_blocks):
+        map_entry = tl.load(map_ptr_base + i * map_kv_stride)
+        if map_entry:
+            tl.store(index_ptr_base + num * index_kv_stride, i)
+            num += 1
+
+    tl.store(
+        index_num_ptr + b * index_num_bs_stride + h * index_num_h_stride +
+        q * index_num_q_stride, num)
+
+def topk_index_to_map(index: torch.Tensor,
+                      num_kv_blocks: int,
+                      transpose_map: bool = False):
+    """
+    Convert topk indices to a map.
+    
+    Args:
+        index: [bs, h, num_q_blocks, topk]
+            The topk indices tensor.
+        num_kv_blocks: int
+            The number of key-value blocks in the block_map returned
+        transpose_map: bool
+            If True, the block_map will be transposed on the final two dimensions.
+    
+    Returns:
+        block_map: [bs, h, num_q_blocks, num_kv_blocks]
+            A binary map where 1 indicates that the q block attends to the kv block.
+    """
+    bs, h, num_q_blocks, topk = index.shape
+
+    if transpose_map is False:
+        block_map = torch.zeros((bs, h, num_q_blocks, num_kv_blocks),
+                                dtype=torch.bool,
+                                device=index.device)
+    else:
+        block_map = torch.zeros((bs, h, num_kv_blocks, num_q_blocks),
+                                dtype=torch.bool,
+                                device=index.device)
+        block_map = block_map.transpose(2, 3)
+
+    grid = (bs, h, num_q_blocks)
+    topk_index_to_map_kernel[grid](
+        block_map,
+        index,
+        block_map.stride(0),
+        block_map.stride(1),
+        block_map.stride(2),
+        block_map.stride(3),
+        index.stride(0),
+        index.stride(1),
+        index.stride(2),
+        index.stride(3),
+        topk=topk,
+    )
+
+    return block_map
+
+def map_to_index(block_map: torch.Tensor):
+    """
+    Convert a block map to indices and counts.
+    
+    Args:
+        block_map: [bs, h, num_q_blocks, num_kv_blocks]
+            The block map tensor.
+    
+    Returns:
+        index: [bs, h, num_q_blocks, num_kv_blocks]
+            The indices of the blocks.
+        index_num: [bs, h, num_q_blocks]
+            The number of blocks for each q block.
+    """
+    bs, h, num_q_blocks, num_kv_blocks = block_map.shape
+
+    index = torch.full((block_map.shape),
+                       -1,
+                       dtype=torch.int32,
+                       device=block_map.device)
+    index_num = torch.empty((bs, h, num_q_blocks),
+                            dtype=torch.int32,
+                            device=block_map.device)
+
+    grid = (bs, h, num_q_blocks)
+    map_to_index_kernel[grid](
+        block_map,
+        index,
+        index_num,
+        block_map.stride(0),
+        block_map.stride(1),
+        block_map.stride(2),
+        block_map.stride(3),
+        index.stride(0),
+        index.stride(1),
+        index.stride(2),
+        index.stride(3),
+        index_num.stride(0),
+        index_num.stride(1),
+        index_num.stride(2),
+        num_kv_blocks=num_kv_blocks,
+    )
+
+    return index, index_num
+
+class BlockSparseAttentionFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num):
+        o, lse = block_sparse_attention_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)
+        ctx.save_for_backward(q, k, v, o, lse, k2q_block_sparse_index, k2q_block_sparse_num)
+        return o
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        q, k, v, o, lse, k2q_block_sparse_index, k2q_block_sparse_num = ctx.saved_tensors
+        grad_q, grad_k, grad_v = block_sparse_attention_backward(
+            q, k, v, o, lse, grad_output, k2q_block_sparse_index, k2q_block_sparse_num
+        )
+        return grad_q, grad_k, grad_v, None, None, None, None
+
+
+class DummyOperator(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x):
+        return x
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return grad_output
+
+class CheckpointSDPA(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, obj, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k):
+        """Forward pass."""
+        with torch.no_grad():
+            mask = index_to_mask(q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k, k.shape[2] // block_k)
+            outputs = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask)
+        ctx.save_for_backward(*detach_variable((q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)))
+        ctx.block_q = block_q
+        ctx.block_k = block_k
+        # the obj is passed in, then it can access the saved input
+        # tensors later for recomputation
+        obj.ctx = ctx
+        return outputs
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        """Backward pass."""
+        inputs = ctx.saved_tensors
+        output = ctx.output
+        torch.autograd.backward(output, grad_output)
+        ctx.output = None
+        grads = tuple(inp.grad for inp in inputs)
+        return (None, ) + grads + (None, None)
+
+
+class BlockSparseAttnTorch:
+    def __init__(self):
+        self.ctx = None
+    
+    def recompute_mask(self, _):
+        recomputed_mask = index_to_mask(self.q2k_block_sparse_index, self.q2k_block_sparse_num, self.block_q, self.block_k, self.num_kv_blocks)
+        mask_size = recomputed_mask.untyped_storage().size()
+        self.mask.untyped_storage().resize_(mask_size)
+        self.mask.untyped_storage().copy_(recomputed_mask.untyped_storage()) 
+    
+    def recompute(self, _):
+        q, k, v, q2k_block_sparse_index, q2k_block_sparse_num = self.ctx.saved_tensors
+        block_q = self.ctx.block_q
+        block_k = self.ctx.block_k
+        mask = index_to_mask(q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k, k.shape[2] // block_k)
+        with torch.enable_grad():
+            output = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask)
+        self.ctx.output = output
+        self.ctx = None
+
+    @torch._dynamo.disable
+    def forward(self, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k):
+        """
+        Differentiable block sparse attention function using PyTorch.
+        
+        Args:
+            q: Query tensor [batch_size, num_heads, seq_len_q, head_dim]
+            k: Key tensor [batch_size, num_heads, seq_len_kv, head_dim]
+            v: Value tensor [batch_size, num_heads, seq_len_kv, head_dim]
+            q2k_block_sparse_index: Indices for query-to-key sparse blocks
+            q2k_block_sparse_num: Number of sparse blocks for each query block
+            block_q: Block size for query
+            block_k: Block size for key-value
+
+        Returns:
+            output: Attention output tensor [batch_size, num_heads, seq_len_q, head_dim]
+        """
+
+        output = CheckpointSDPA.apply(
+            self, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k
+        )
+
+        o = DummyOperator.apply(output)
+        o.register_hook(self.recompute)
+        return o
@@ -1,7 +1,9 @@
-FROM nvidia/cuda:12.4.1-devel-ubuntu20.04
+FROM nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04

 ENV DEBIAN_FRONTEND=noninteractive

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

 RUN apt-get update && apt-get install -y --no-install-recommends \
@@ -9,17 +11,25 @@ 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/*

-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 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

-ENV PATH=/opt/conda/bin:$PATH
+# Set CUDA environment variables
+ENV CUDA_HOME=/usr/local/cuda-12.8
+ENV PATH=${CUDA_HOME}/bin:${PATH}
+ENV LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH

-RUN conda create --name fastvideo-dev python=3.12.9 -y
-
-SHELL ["/bin/bash", "-c"]
+# 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 ./
@@ -27,22 +37,36 @@ COPY pyproject.toml ./
 # Create a dummy README to satisfy the installation
 RUN echo "# Placeholder" > README.md

-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
+# 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==2.8.0.post2 --no-build-isolation

 COPY . .

-RUN conda run -n fastvideo-dev pip install --no-cache-dir -e .[dev]
+# 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

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

 EXPOSE 22
@@ -25,7 +25,7 @@ It features a clean, consistent API that works across popular video models, maki
 With FastVideo's optimizations, you can achieve more than 3x inference improvement compared to other systems.

 <div style="text-align: center;">
-  <img src=../../_static/images/perf.png width="100%"/>
+  <img src=_static/images/perf.png width="100%"/>
 </div>

 ## Key Features
@@ -67,6 +67,7 @@ inference/configuration
 inference/optimizations
 inference/support_matrix
 inference/examples/examples_inference_index
+inference/cli
 inference/add_pipeline
 inference/v0_inference
 :::
@@ -57,8 +57,9 @@ Run the script with:
 python example.py
 ```

-The generated video will be saved in the current directory under `my_videos/`.
+The generated video will be saved in the current directory under `my_videos/`  

+More inference example scripts can be found in `scripts/inference/`
 ## Available Models

 Please see the [support matrix](#support-matrix) for the list of supported models and their available optimizations.
@@ -79,7 +80,6 @@ def main():
    sampling_param = SamplingParam.from_pretrained(model_name)
    sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
    sampling_param.num_frames = 107
-    sampling_param.image_strength = 0.8  # How much to preserve the original image (0-1)

    # Generate video based on the image
    prompt = "A photograph coming to life with gentle movement"
@@ -76,8 +76,8 @@ The `HuggingFace Model ID` can be directly pass to `from_pretrained()` methods a
 - * StepVideo T2V
  * `FastVideo/stepvideo-t2v-diffusers`
  * 768px768px204f<br>544px992px204f<br>544px992px136f
-  *
-  *
+  * ❌
+  * ❌
  * ✅
 :::

@@ -1,7 +1,7 @@
 (sta-demo)=

 # 🔍 Demo
-There is a demo for 2D STA with window size (6,6) operating on a (10, 10) image.
+This is is a demo for 2D STA with window size (6,6) operating on a (10, 10) image.

 <div style="text-align: center;">
  <video controls width="800">
@@ -9,3 +9,9 @@ There is a demo for 2D STA with window size (6,6) operating on a (10, 10) image.
    Your browser does not support the video tag.
  </video>
 </div>
+
+You can run STA using the following command:
+
+```bash
+bash scripts/inference/v1_inference_wan_STA.sh
+```
@@ -7,70 +7,40 @@ To save GPU memory, we precompute text embeddings and VAE latents to eliminate t
 We provide a sample dataset to help you get started. Download the source media using the following command:

 ```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/Image-Vid-Finetune-Src --local_dir=data/Image-Vid-Finetune-Src --repo_type=dataset
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/mini_i2v_dataset --local_dir=FastVideo/mini_i2v_dataset --repo_type=dataset
 ```

+The folder `crush-smol_raw/` contains raw videos and captions for testing preprocessing, while `crush-smol_preprocessed/` contains latents prepared for testing training.
+
 To preprocess the dataset for fine-tuning or distillation, run:

 ```
-bash scripts/preprocess/preprocess_mochi_data.sh # for mochi
-bash scripts/preprocess/preprocess_hunyuan_data.sh # for hunyuan
+bash scripts/preprocess/v1_preprocess_wan_data_t2v # for wan
 ```

-The preprocessed dataset will be stored in `Image-Vid-Finetune-Mochi` or `Image-Vid-Finetune-HunYuan` correspondingly.
-
 ## Process your own dataset

-If you wish to create your own dataset for finetuning or distillation, please structure you video dataset in the following format:
+If you wish to create your own dataset for finetuning or distillation, please refer `mini_i2v_dataset/crush-smol_raw/` to structure you video dataset in the following format:

 ```
-path_to_dataset_folder/
-├── media/
-│   ├── 0.jpg
+path_to_your_dataset_folder/
+├── videos/
+│   ├── 0.mp4
 │   ├── 1.mp4
-│   ├── 2.jpg
-├── video2caption.json
-└── merge.txt
+├── videos.txt
+└── prompt.txt
 ```

-Format the JSON file as a list, where each item represents a media source:
+To geranate the `videos2caption.json` and `merge.txt`, run

-For image media,
-
-```
-{
-    "path": "0.jpg",
-    "cap": ["captions"]
-}
+``` python
+python scripts/dataset_preparation/prepare_json_file.py --data_folder mini_i2v_dataset/crush-smol_raw/ --output your_output_folder
 ```

-For video media,
+Adjust the `DATA_MERGE_PATH` and `OUTPUT_DIR` in `scripts/preprocess/v1_preprocess_****.sh` accordingly and run:

 ```
-{
-    "path": "1.mp4",
-    "resolution": {
-      "width": 848,
-      "height": 480
-    },
-    "fps": 30.0,
-    "duration": 6.033333333333333,
-    "cap": [
-      "caption"
-    ]
-  }
-```
-
-Use a txt file (merge.txt) to contain the source folder for media and the JSON file for meta information:
-
-```
-path_to_media_source_foder,path_to_json_file
-```
-
-Adjust the `DATA_MERGE_PATH` and `OUTPUT_DIR` in `scripts/preprocess/preprocess_****_data.sh` accordingly and run:
-
-```
-bash scripts/preprocess/preprocess_****_data.sh
+bash scripts/preprocess/v1_preprocess_****.sh
 ```

 The preprocessed data will be put into the `OUTPUT_DIR` and the `videos2caption.json` can be used in finetune and distill scripts.
@@ -16,6 +16,13 @@ bash scripts/finetune/finetune_mochi.sh # for mochi
 ```

 **Note that for finetuning, we did not tune the hyperparameters in the provided script.**
+## ⚡ Finetune with VSA
+Follow [data_preprocess.md](#v0-data-preprocess) to get parquet files for preproccessed latent, and then run:
+
+```bash
+bash scripts/finetune/finetune_v1_VSA.sh
+```
+
 ## ⚡ Lora Finetune

 Hunyuan supports Lora fine-tuning of videos up to 720p. Demos and prompts of Black-Myth-Wukong can be found in [here](https://huggingface.co/FastVideo/Hunyuan-Black-Myth-Wukong-lora-weight). You can download the Lora weight through:
@@ -2,7 +2,7 @@ from fastvideo import VideoGenerator

 # from fastvideo.v1.configs.sample import SamplingParam

-
+OUTPUT_PATH = "video_samples"
 def main():
    # FastVideo will automatically use the optimal default arguments for the
    # model.
@@ -11,7 +11,9 @@ def main():
    generator = VideoGenerator.from_pretrained(
        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
        # if num_gpus > 1, FastVideo will automatically handle distributed setup
-        num_gpus=1,
+        num_gpus=2,
+        use_fsdp_inference=True,
+        use_cpu_offload=False
    )

    # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
@@ -23,7 +25,7 @@ def main():
        "wide with interest. The playful yet serene atmosphere is complemented by soft "
        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
    )
-    video = generator.generate_video(prompt)
+    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True)
    # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/")

    # Generate another video with a different prompt, without reloading the
@@ -34,7 +36,7 @@ def main():
        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
        "cinematic.")
-    video2 = generator.generate_video(prompt2)
+    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True)


 if __name__ == "__main__":
@@ -1,5 +1,5 @@
 from fastvideo import VideoGenerator
-
+from fastvideo.v1.configs.pipelines.base import PipelineConfig

 def main():

@@ -0,0 +1,45 @@
+from fastvideo import VideoGenerator
+from fastvideo.v1.configs.sample import SamplingParam
+
+OUTPUT_PATH = "./lora"
+def main():
+    # Initialize VideoGenerator with the Wan model
+    generator = VideoGenerator.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        num_gpus=2,
+        lora_path="benjamin-paine/steamboat-willie-1.3b",
+        lora_nickname="steamboat"
+    )
+    kwargs = {
+        "height": 480,
+        "width": 832,
+        "num_frames": 81,
+        "guidance_scale": 5.0,
+        "num_inference_steps": 32,
+    }
+    # Generate video with LoRA style
+    prompt = "steamboat willie style, golden era animation, close-up of a short fluffy monster  kneeling beside a melting red candle. the mood is one of wonder and curiosity,  as the monster gazes at the flame with wide eyes and open mouth. Its pose and expression  convey a sense of innocence and playfulness, as if it is exploring the world around it for the first time.  The use of warm colors and dramatic lighting further enhances the cozy atmosphere of the image."
+    negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
+
+    video = generator.generate_video(
+        prompt,
+        # sampling_param=sampling_param,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        negative_prompt=negative_prompt,
+        **kwargs
+    )
+
+    generator.set_lora_adapter(lora_nickname="flat_color", lora_path="motimalu/wan-flat-color-1.3b-v2")
+    prompt = "flat color, no lineart, blending, negative space, artist:[john kafka|ponsuke kaikai|hara id 21|yoneyama mai|fuzichoco],  1girl, sakura miko, pink hair, cowboy shot, white shirt, floral print, off shoulder, outdoors, cherry blossom, tree shade, wariza, looking up, falling petals, half-closed eyes, white sky, clouds,  live2d animation, upper body, high quality cinematic video of a woman sitting under a sakura tree. Dreamy and lonely, the camera close-ups on the face of the woman as she turns towards the viewer. The Camera is steady, This is a cowboy shot. The animation is smooth and fluid."
+    negative_prompt = "bad quality video,色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
+    video = generator.generate_video(
+        prompt,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        negative_prompt=negative_prompt,
+        **kwargs
+    )
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,5 @@
+# STA Mask Search Examples
+
+```bash
+bash examples/inference/sta_mask_search/inference_wan_sta.sh
+```
@@ -0,0 +1,39 @@
+#!/bin/bash
+
+export FASTVIDEO_ATTENTION_CONFIG=assets/mask_strategy_wan.json
+export FASTVIDEO_ATTENTION_BACKEND=SLIDING_TILE_ATTN
+export MODEL_BASE=Wan-AI/Wan2.1-T2V-14B-Diffusers
+
+base_port=29503
+num_gpu=1
+gpu_ids=$(seq 0 $((num_gpu-1)))
+skip_time_steps=12
+
+output_path="inference_results/sta/mask_search_full"
+STA_mode="STA_searching"
+for i in $gpu_ids; do
+    port=$((base_port+i))
+    CUDA_VISIBLE_DEVICES=$i MASTER_PORT=$port python examples/inference/sta_mask_search/wan_example.py \
+        --prompt_path ./assets/prompt_${i}.txt \
+        --output_path $output_path \
+        --STA_mode $STA_mode &
+    sleep 1
+done
+wait
+echo "STA searching completed"
+
+output_path="inference_results/sta/mask_search_sparse"
+STA_mode="STA_tuning"
+for i in $gpu_ids; do
+    port=$((base_port+i))
+    CUDA_VISIBLE_DEVICES=$i MASTER_PORT=$port python examples/inference/sta_mask_search/wan_example.py \
+        --prompt_path ./assets/prompt_${i}.txt \
+        --output_path $output_path \
+        --STA_mode $STA_mode \
+        --skip_time_steps $skip_time_steps &
+    sleep 1
+done
+wait
+echo "STA tuning completed"
+
+echo "All jobs completed"
@@ -0,0 +1,63 @@
+import os
+import argparse
+from fastvideo import VideoGenerator, SamplingParam
+
+def main(args):
+    os.makedirs(args.output_path, exist_ok=True)
+    # Create a video generator with a pre-trained model
+    generator = VideoGenerator.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-14B-Diffusers",
+        num_gpus=args.num_gpus,  # Adjust based on your hardware
+        STA_mode=args.STA_mode,
+        skip_time_steps=args.skip_time_steps
+    )
+
+    # Prompts for your video
+    prompt = args.prompt
+    prompt_path = args.prompt_path
+    negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+    if prompt_path is not None:
+        with open(prompt_path, "r") as f:
+            prompts = f.readlines()
+    else:
+        prompts = [prompt]
+
+    params = SamplingParam(
+        height=args.height,
+        width=args.width,
+        num_frames=args.num_frames,
+        num_inference_steps=args.num_inference_steps,
+        fps=args.fps,
+        guidance_scale=args.guidance_scale,
+        seed=args.seed,
+        return_frames=True,  # Also return frames from this call (defaults to False)
+        output_path=args.output_path,  # Controls where videos are saved
+        save_video=True,
+        negative_prompt=negative_prompt
+    )
+
+    # Generate the video
+    for prompt in prompts:
+        video = generator.generate_video(
+            prompt,
+            sampling_param=params,
+        )
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--prompt", type=str, default="A man is dancing.")
+    parser.add_argument("--prompt_path", type=str, default=None)
+    parser.add_argument("--height", type=int, default=768)
+    parser.add_argument("--width", type=int, default=1280)
+    parser.add_argument("--num_frames", type=int, default=69)
+    parser.add_argument("--num_inference_steps", type=int, default=50)
+    parser.add_argument("--fps", type=int, default=16)
+    parser.add_argument("--guidance_scale", type=float, default=5.0)
+    parser.add_argument("--seed", type=int, default=12345)
+    parser.add_argument("--output_path", type=str, default="my_videos/")
+    parser.add_argument("--num_gpus", type=int, default=1)
+    parser.add_argument("--STA_mode", type=str, default="STA_searching")
+    parser.add_argument("--skip_time_steps", type=int, default=12)
+    args = parser.parse_args()
+    main(args)
@@ -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": "examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation_dataset/yYcK4nANZz4-Scene-034.mp4",
+      "num_inference_steps": 50,
+      "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": "examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation_dataset/yYcK4nANZz4-Scene-027.mp4",
+      "num_inference_steps": 50,
+      "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": "examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation_dataset/yYcK4nANZz4-Scene-030.mp4",
+      "num_inference_steps": 50,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    }
+  ]
+}
@@ -1,5 +1,6 @@
 from fastvideo.v1.configs.pipelines import PipelineConfig
 from fastvideo.v1.configs.sample import SamplingParam
 from fastvideo.v1.entrypoints.video_generator import VideoGenerator
+from fastvideo.version import __version__

-__all__ = ["VideoGenerator", "PipelineConfig", "SamplingParam"]
+__all__ = ["VideoGenerator", "PipelineConfig", "SamplingParam", "__version__"]
@@ -68,7 +68,8 @@ def main(args):
    train_dataset = T5dataset(latents_json_path, args.vae_debug)
    text_encoder = load_text_encoder(args.model_type, args.model_path, device=device)
    vae, autocast_type, fps = load_vae(args.model_type, args.model_path)
-    vae.enable_tiling()
+    if args.model_type != "wan":
+        vae.enable_tiling()
    sampler = DistributedSampler(train_dataset, rank=local_rank, num_replicas=world_size, shuffle=True)
    train_dataloader = DataLoader(
        train_dataset,
@@ -33,7 +33,8 @@ def main(args):
    if not dist.is_initialized():
        dist.init_process_group(backend="nccl", init_method="env://", world_size=world_size, rank=local_rank)
    vae, autocast_type, fps = load_vae(args.model_type, args.model_path)
-    vae.enable_tiling()
+    if args.model_type != "wan":
+        vae.enable_tiling()
    os.makedirs(args.output_dir, exist_ok=True)
    os.makedirs(os.path.join(args.output_dir, "latent"), exist_ok=True)

@@ -103,13 +104,7 @@ if __name__ == "__main__":
        default=None,
        help="The output directory where the model predictions and checkpoints will be written.",
    )
-    parser.add_argument(
-        "--logging_dir",
-        type=str,
-        default="logs",
-        help=("[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
-              " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."),
-    )
+

    args = parser.parse_args()
    main(args)
@@ -12,6 +12,7 @@ import torch.distributed as dist
 import wandb
 from accelerate.utils import set_seed
 from diffusers import FlowMatchEulerDiscreteScheduler
+from fastvideo.distill.solver import PCMFMScheduler
 from diffusers.optimization import get_scheduler
 from diffusers.utils import check_min_version
 from peft import LoraConfig
@@ -23,7 +24,7 @@ from tqdm.auto import tqdm

 from fastvideo.dataset.latent_datasets import (LatentDataset, latent_collate_function)
 from fastvideo.distill.solver import EulerSolver, extract_into_tensor
-from fastvideo.models.mochi_hf.mochi_latents_utils import normalize_dit_input
+from fastvideo.utils.latents_utils import normalize_dit_input
 from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
 from fastvideo.utils.checkpoint import (resume_lora_optimizer, save_checkpoint, save_lora_checkpoint)
 from fastvideo.utils.communications import (broadcast, sp_parallel_dataloader_wrapper)
@@ -123,13 +124,21 @@ def distill_one_step(
        noisy_model_input = sigmas * noise + (1.0 - sigmas) * model_input
        # Predict the noise residual
        with torch.autocast("cuda", dtype=torch.bfloat16):
-            teacher_kwargs = {
-                "hidden_states": noisy_model_input,
-                "encoder_hidden_states": encoder_hidden_states,
-                "timestep": timesteps,
-                "encoder_attention_mask": encoder_attention_mask,  # B, L
-                "return_dict": False,
-            }
+            if args.model_type == "wan":
+                    teacher_kwargs = {
+                    "hidden_states": noisy_model_input,
+                    "encoder_hidden_states": encoder_hidden_states,
+                    "timestep": timesteps,
+                    "return_dict": True,
+                }
+            else:
+                teacher_kwargs = {
+                    "hidden_states": noisy_model_input,
+                    "encoder_hidden_states": encoder_hidden_states,
+                    "timestep": timesteps,
+                    "encoder_attention_mask": encoder_attention_mask,  # B, L
+                    "return_dict": False,
+                }
            if hunyuan_teacher_disable_cfg:
                teacher_kwargs["guidance"] = torch.tensor([1000.0],
                                                          device=noisy_model_input.device,
@@ -141,47 +150,70 @@ def distill_one_step(
        with torch.no_grad():
            w = distill_cfg
            with torch.autocast("cuda", dtype=torch.bfloat16):
-                cond_teacher_output = teacher_transformer(
-                    noisy_model_input,
-                    encoder_hidden_states,
-                    timesteps,
-                    encoder_attention_mask,  # B, L
-                    return_dict=False,
-                )[0].float()
+                if args.model_type == "wan":
+                    cond_teacher_kwargs ={
+                    "hidden_states": noisy_model_input,
+                    "encoder_hidden_states": encoder_hidden_states,
+                    "timestep": timesteps,
+                    "return_dict": True,
+                    }
+                else:
+                    cond_teacher_kwargs = {
+                    "hidden_states": noisy_model_input,
+                    "encoder_hidden_states": encoder_hidden_states,
+                    "timestep": timesteps,
+                    "encoder_attention_mask": encoder_attention_mask,  # B, L
+                    "return_dict": False,
+                    }
+                cond_teacher_output = teacher_transformer(**cond_teacher_kwargs)[0].float()
            if not_apply_cfg_solver:
                uncond_teacher_output = cond_teacher_output
            else:
                # Get teacher model prediction on noisy_latents and unconditional embedding
                with torch.autocast("cuda", dtype=torch.bfloat16):
-                    uncond_teacher_output = teacher_transformer(
-                        noisy_model_input,
-                        uncond_prompt_embed.unsqueeze(0).expand(bsz, -1, -1),
-                        timesteps,
-                        uncond_prompt_mask.unsqueeze(0).expand(bsz, -1),
-                        return_dict=False,
-                    )[0].float()
+                    if args.model_type == "wan":
+                        uncond_teacher_kwargs = {
+                        "hidden_states": noisy_model_input,
+                        "encoder_hidden_states":uncond_prompt_embed.unsqueeze(0).expand(bsz, -1, -1),
+                        "timestep": timesteps,
+                        "return_dict": True,
+                        }
+                    else:
+                        uncond_teacher_kwargs = {
+                        "hidden_states": noisy_model_input,
+                        "encoder_hidden_states":uncond_prompt_embed.unsqueeze(0).expand(bsz, -1, -1),
+                        "timestep": timesteps,
+                        "encoder_attention_mask": uncond_prompt_mask.unsqueeze(0).expand(bsz, -1),
+                        "return_dict": False,
+                        }
+                    
+                    uncond_teacher_output = teacher_transformer(**uncond_teacher_kwargs)[0].float()
+
            teacher_output = uncond_teacher_output + w * (cond_teacher_output - uncond_teacher_output)
            x_prev = solver.euler_step(noisy_model_input, teacher_output, index)

        # 20.4.12. Get target LCM prediction on x_prev, w, c, t_n
        with torch.no_grad():
            with torch.autocast("cuda", dtype=torch.bfloat16):
-                if ema_transformer is not None:
-                    target_pred = ema_transformer(
-                        x_prev.float(),
-                        encoder_hidden_states,
-                        timesteps_prev,
-                        encoder_attention_mask,  # B, L
-                        return_dict=False,
-                    )[0]
+                if args.model_type == "wan":
+                    target_pred_kwargs = {
+                        "hidden_states": x_prev.float(),
+                        "encoder_hidden_states": encoder_hidden_states,
+                        "timestep":timesteps_prev,
+                        "return_dict":True,
+                    }
                else:
-                    target_pred = transformer(
-                        x_prev.float(),
-                        encoder_hidden_states,
-                        timesteps_prev,
-                        encoder_attention_mask,  # B, L
-                        return_dict=False,
-                    )[0]
+                    target_pred_kwargs = {
+                        "hidden_states": x_prev.float(),
+                        "encoder_hidden_states": encoder_hidden_states,
+                        "timestep":timesteps_prev,
+                        "encoder_attention_mask":encoder_attention_mask,
+                        "return_dict":False,
+                    }
+                if ema_transformer is not None:
+                    target_pred = ema_transformer(**target_pred_kwargs)[0]
+                else:
+                    target_pred = transformer(**target_pred_kwargs)[0]

            target, end_index = solver.euler_style_multiphase_pred(x_prev, target_pred, index, multiphase, True)

@@ -242,7 +274,7 @@ def main(args):
    noise_random_generator = None

    # Handle the repository creation
-    if rank <= 0 and args.output_dir is not None:
+    if rank == 0 and args.output_dir is not None:
        os.makedirs(args.output_dir, exist_ok=True)

    # For mixed precision training we cast all non-trainable weights to half-precision
@@ -319,7 +351,9 @@ def main(args):
    teacher_transformer.requires_grad_(False)
    if args.use_ema:
        ema_transformer.requires_grad_(False)
-    noise_scheduler = FlowMatchEulerDiscreteScheduler(shift=args.shift)
+    
+    noise_scheduler = FlowMatchEulerDiscreteScheduler()
+
    if args.scheduler_type == "pcm_linear_quadratic":
        linear_steps = int(noise_scheduler.config.num_train_timesteps * args.linear_range)
        sigmas = linear_quadratic_schedule(
@@ -391,7 +425,7 @@ def main(args):
        len(train_dataloader) / args.gradient_accumulation_steps * args.sp_size / args.train_sp_batch_size)
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

-    if rank <= 0:
+    if rank == 0:
        project = args.tracker_project_name or "fastvideo"
        wandb.init(project=project, config=args)

@@ -493,7 +527,7 @@ def main(args):
            "phases": num_phases,
        })
        progress_bar.update(1)
-        if rank <= 0:
+        if rank == 0:
            wandb.log(
                {
                    "train_loss": loss,
@@ -637,13 +671,6 @@ if __name__ == "__main__":
        help=("Whether training should be resumed from a previous lora checkpoint. Use a path saved by"
              ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'),
    )
-    parser.add_argument(
-        "--logging_dir",
-        type=str,
-        default="logs",
-        help=("[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
-              " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."),
-    )

    # optimizer & scheduler & Training
    parser.add_argument("--num_train_epochs", type=int, default=100)
@@ -23,7 +23,7 @@ from tqdm.auto import tqdm
 from fastvideo.dataset.latent_datasets import (LatentDataset, latent_collate_function)
 from fastvideo.distill.discriminator import Discriminator
 from fastvideo.distill.solver import EulerSolver, extract_into_tensor
-from fastvideo.models.mochi_hf.mochi_latents_utils import normalize_dit_input
+from fastvideo.utils.latents_utils import normalize_dit_input
 from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
 from fastvideo.utils.checkpoint import (resume_lora_optimizer, resume_training_generator_discriminator, save_checkpoint,
                                        save_lora_checkpoint)
@@ -296,7 +296,7 @@ def main(args):
    noise_random_generator = None

    # Handle the repository creation
-    if rank <= 0 and args.output_dir is not None:
+    if rank == 0 and args.output_dir is not None:
        os.makedirs(args.output_dir, exist_ok=True)

    # For mixed precision training we cast all non-trainable weights to half-precision
@@ -462,7 +462,7 @@ def main(args):
        len(train_dataloader) / args.gradient_accumulation_steps * args.sp_size / args.train_sp_batch_size)
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

-    if rank <= 0:
+    if rank == 0:
        project = args.tracker_project_name or "fastvideo"
        wandb.init(project=project, config=args)

@@ -559,7 +559,7 @@ def main(args):
            "step_time": f"{step_time:.2f}s",
        })
        progress_bar.update(1)
-        if rank <= 0:
+        if rank == 0:
            wandb.log(
                {
                    "generator_loss": generator_loss,
@@ -693,13 +693,6 @@ if __name__ == "__main__":
        help=("Whether training should be resumed from a previous lora checkpoint. Use a path saved by"
              ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'),
    )
-    parser.add_argument(
-        "--logging_dir",
-        type=str,
-        default="logs",
-        help=("[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
-              " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."),
-    )

    # optimizer & scheduler & Training
    parser.add_argument("--num_train_epochs", type=int, default=100)
@@ -237,7 +237,7 @@ def add_inference_args(parser: argparse.ArgumentParser):
        type=str,
        default="540p",
        choices=["540p", "720p"],
-        help="Root path of all the models, including t2v models and extra models.",
+        help="The resolution of the model.",
    )
    group.add_argument(
        "--load-key",
@@ -361,7 +361,7 @@ def add_parallel_args(parser: argparse.ArgumentParser):
        "--ring-degree",
        type=int,
        default=1,
-        help="Ulysses degree.",
+        help="Ring degree.",
    )

    return parser
@@ -17,7 +17,7 @@ from fastvideo.models.hunyuan.vae import load_vae
 from fastvideo.utils.parallel_states import nccl_info


-class Inference(object):
+class Inference:

    def __init__(
        self,
@@ -41,7 +41,7 @@ def get_rewrite_prompt(ori_prompt, mode="Normal"):
    elif mode == "Master":
        prompt = master_mode_prompt.format(input=ori_prompt)
    else:
-        raise Exception("Only supports Normal and Normal", mode)
+        raise Exception("Only supports Normal and Master mode, but got {}".format(mode))
    return prompt


@@ -267,25 +267,25 @@ class Step1Model(PreTrainedModel):
 class STEP1TextEncoder(torch.nn.Module):

    def __init__(self, model_dir, max_length=320):
-        super(STEP1TextEncoder, self).__init__()
+        super()
        self.max_length = max_length
        self.text_tokenizer = Wrapped_StepChatTokenizer(os.path.join(model_dir, 'step1_chat_tokenizer.model'))
        text_encoder = Step1Model.from_pretrained(model_dir)
        self.text_encoder = text_encoder.eval().to(torch.bfloat16)

    @torch.no_grad
+    @torch.autocast(device_type='cuda', dtype=torch.bfloat16)
    def forward(self, prompts, with_mask=True, max_length=None):
        self.device = next(self.text_encoder.parameters()).device
-        with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
-            if type(prompts) is str:
-                prompts = [prompts]
+        if type(prompts) is str:
+            prompts = [prompts]

-            txt_tokens = self.text_tokenizer(prompts,
-                                             max_length=max_length or self.max_length,
-                                             padding="max_length",
-                                             truncation=True,
-                                             return_tensors="pt")
-            y = self.text_encoder(txt_tokens.input_ids.to(self.device),
+        txt_tokens = self.text_tokenizer(prompts,
+                                            max_length=max_length or self.max_length,
+                                            padding="max_length",
+                                            truncation=True,
+                                            return_tensors="pt")
+        y = self.text_encoder(txt_tokens.input_ids.to(self.device),
                                  attention_mask=txt_tokens.attention_mask.to(self.device) if with_mask else None)
-            y_mask = txt_tokens.attention_mask
+        y_mask = txt_tokens.attention_mask
        return y.transpose(0, 1), y_mask
@@ -0,0 +1,486 @@
+# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
+from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import Attention
+from diffusers.models.cache_utils import CacheMixin
+from diffusers.models.embeddings import PixArtAlphaTextProjection, TimestepEmbedding, Timesteps, get_1d_rotary_pos_embed
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import FP32LayerNorm
+
+from fastvideo.models.flash_attn_no_pad import flash_attn_no_pad
+from fastvideo.utils.communications import all_gather, all_to_all_4D
+from fastvideo.utils.parallel_states import get_sequence_parallel_state, nccl_info
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class WanAttnProcessor2_0:
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("WanAttnProcessor2_0 requires PyTorch 2.0. To use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        encoder_hidden_states_img = None
+        if attn.add_k_proj is not None:
+            # 512 is the context length of the text encoder, hardcoded for now
+            image_context_length = encoder_hidden_states.shape[1] - 512
+            encoder_hidden_states_img = encoder_hidden_states[:, :image_context_length]
+            encoder_hidden_states = encoder_hidden_states[:, image_context_length:]
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(hidden_states: torch.Tensor, freqs: torch.Tensor):
+                x_rotated = torch.view_as_complex(hidden_states.to(torch.float64).unflatten(3, (-1, 2)))
+                x_out = torch.view_as_real(x_rotated * freqs).flatten(3, 4)
+                return x_out.type_as(hidden_states)
+
+            query = apply_rotary_emb(query, rotary_emb)
+            key = apply_rotary_emb(key, rotary_emb)
+
+        # I2V task
+        hidden_states_img = None
+        if encoder_hidden_states_img is not None:
+            key_img = attn.add_k_proj(encoder_hidden_states_img)
+            key_img = attn.norm_added_k(key_img)
+            value_img = attn.add_v_proj(encoder_hidden_states_img)
+
+            key_img = key_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+            value_img = value_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+            hidden_states_img = F.scaled_dot_product_attention(
+                query, key_img, value_img, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+            hidden_states_img = hidden_states_img.transpose(1, 2).flatten(2, 3)
+            hidden_states_img = hidden_states_img.type_as(query)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        if hidden_states_img is not None:
+            hidden_states = hidden_states + hidden_states_img
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
+class WanImageEmbedding(torch.nn.Module):
+    def __init__(self, in_features: int, out_features: int, pos_embed_seq_len=None):
+        super().__init__()
+
+        self.norm1 = FP32LayerNorm(in_features)
+        self.ff = FeedForward(in_features, out_features, mult=1, activation_fn="gelu")
+        self.norm2 = FP32LayerNorm(out_features)
+        if pos_embed_seq_len is not None:
+            self.pos_embed = nn.Parameter(torch.zeros(1, pos_embed_seq_len, in_features))
+        else:
+            self.pos_embed = None
+
+    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+        if self.pos_embed is not None:
+            batch_size, seq_len, embed_dim = encoder_hidden_states_image.shape
+            encoder_hidden_states_image = encoder_hidden_states_image.view(-1, 2 * seq_len, embed_dim)
+            encoder_hidden_states_image = encoder_hidden_states_image + self.pos_embed
+
+        hidden_states = self.norm1(encoder_hidden_states_image)
+        hidden_states = self.ff(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+        return hidden_states
+
+
+class WanTimeTextImageEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        time_freq_dim: int,
+        time_proj_dim: int,
+        text_embed_dim: int,
+        image_embed_dim: Optional[int] = None,
+        pos_embed_seq_len: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.timesteps_proj = Timesteps(num_channels=time_freq_dim, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.time_embedder = TimestepEmbedding(in_channels=time_freq_dim, time_embed_dim=dim)
+        self.act_fn = nn.SiLU()
+        self.time_proj = nn.Linear(dim, time_proj_dim)
+        self.text_embedder = PixArtAlphaTextProjection(text_embed_dim, dim, act_fn="gelu_tanh")
+
+        self.image_embedder = None
+        if image_embed_dim is not None:
+            self.image_embedder = WanImageEmbedding(image_embed_dim, dim, pos_embed_seq_len=pos_embed_seq_len)
+
+    def forward(
+        self,
+        timestep: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+    ):
+        timestep = self.timesteps_proj(timestep)
+
+        time_embedder_dtype = next(iter(self.time_embedder.parameters())).dtype
+        if timestep.dtype != time_embedder_dtype and time_embedder_dtype != torch.int8:
+            timestep = timestep.to(time_embedder_dtype)
+        temb = self.time_embedder(timestep).type_as(encoder_hidden_states)
+        timestep_proj = self.time_proj(self.act_fn(temb))
+
+        encoder_hidden_states = self.text_embedder(encoder_hidden_states)
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)
+
+        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image
+
+
+class WanRotaryPosEmbed(nn.Module):
+    def __init__(
+        self, attention_head_dim: int, patch_size: Tuple[int, int, int], max_seq_len: int, theta: float = 10000.0
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.patch_size = patch_size
+        self.max_seq_len = max_seq_len
+
+        h_dim = w_dim = 2 * (attention_head_dim // 6)
+        t_dim = attention_head_dim - h_dim - w_dim
+
+        freqs = []
+        for dim in [t_dim, h_dim, w_dim]:
+            freq = get_1d_rotary_pos_embed(
+                dim, max_seq_len, theta, use_real=False, repeat_interleave_real=False, freqs_dtype=torch.float64
+            )
+            freqs.append(freq)
+        self.freqs = torch.cat(freqs, dim=1)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w
+
+        freqs = self.freqs.to(hidden_states.device)
+        freqs = freqs.split_with_sizes(
+            [
+                self.attention_head_dim // 2 - 2 * (self.attention_head_dim // 6),
+                self.attention_head_dim // 6,
+                self.attention_head_dim // 6,
+            ],
+            dim=1,
+        )
+
+        freqs_f = freqs[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_h = freqs[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_w = freqs[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+        freqs = torch.cat([freqs_f, freqs_h, freqs_w], dim=-1).reshape(1, 1, ppf * pph * ppw, -1)
+        return freqs
+
+
+class WanTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        ffn_dim: int,
+        num_heads: int,
+        qk_norm: str = "rms_norm_across_heads",
+        cross_attn_norm: bool = False,
+        eps: float = 1e-6,
+        added_kv_proj_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        # 1. Self-attention
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            processor=WanAttnProcessor2_0(),
+        )
+
+        # 2. Cross-attention
+        self.attn2 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            added_kv_proj_dim=added_kv_proj_dim,
+            added_proj_bias=True,
+            processor=WanAttnProcessor2_0(),
+        )
+        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+
+        # 3. Feed-forward
+        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
+        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        rotary_emb: torch.Tensor,
+    ) -> torch.Tensor:
+        shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+            self.scale_shift_table + temb.float()
+        ).chunk(6, dim=1)
+
+        # 1. Self-attention
+        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+        attn_output = self.attn1(hidden_states=norm_hidden_states, rotary_emb=rotary_emb)
+        hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+
+        # 2. Cross-attention
+        norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+        attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+        hidden_states = hidden_states + attn_output
+
+        # 3. Feed-forward
+        norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
+            hidden_states
+        )
+        ff_output = self.ffn(norm_hidden_states)
+        hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+
+        return hidden_states
+
+
+class WanTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+    r"""
+    A Transformer model for video-like data used in the Wan model.
+
+    Args:
+        patch_size (`Tuple[int]`, defaults to `(1, 2, 2)`):
+            3D patch dimensions for video embedding (t_patch, h_patch, w_patch).
+        num_attention_heads (`int`, defaults to `40`):
+            Fixed length for text embeddings.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_dim (`int`, defaults to `512`):
+            Input dimension for text embeddings.
+        freq_dim (`int`, defaults to `256`):
+            Dimension for sinusoidal time embeddings.
+        ffn_dim (`int`, defaults to `13824`):
+            Intermediate dimension in feed-forward network.
+        num_layers (`int`, defaults to `40`):
+            The number of layers of transformer blocks to use.
+        window_size (`Tuple[int]`, defaults to `(-1, -1)`):
+            Window size for local attention (-1 indicates global attention).
+        cross_attn_norm (`bool`, defaults to `True`):
+            Enable cross-attention normalization.
+        qk_norm (`bool`, defaults to `True`):
+            Enable query/key normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        add_img_emb (`bool`, defaults to `False`):
+            Whether to use img_emb.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["patch_embedding", "condition_embedder", "norm"]
+    _no_split_modules = ["WanTransformerBlock"]
+    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
+    _keys_to_ignore_on_load_unexpected = ["norm_added_q"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: Tuple[int] = (1, 2, 2),
+        num_attention_heads: int = 40,
+        attention_head_dim: int = 128,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        text_dim: int = 4096,
+        freq_dim: int = 256,
+        ffn_dim: int = 13824,
+        num_layers: int = 40,
+        cross_attn_norm: bool = True,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        eps: float = 1e-6,
+        image_dim: Optional[int] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        rope_max_seq_len: int = 1024,
+        pos_embed_seq_len: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        # 1. Patch & position embedding
+        self.rope = WanRotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len)
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Condition embeddings
+        # image_embedding_dim=1280 for I2V model
+        self.condition_embedder = WanTimeTextImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=freq_dim,
+            time_proj_dim=inner_dim * 6,
+            text_embed_dim=text_dim,
+            image_embed_dim=image_dim,
+            pos_embed_seq_len=pos_embed_seq_len,
+        )
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList(
+            [
+                WanTransformerBlock(
+                    inner_dim, ffn_dim, num_attention_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output norm & projection
+        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image
+        )
+        timestep_proj = timestep_proj.unflatten(1, (6, -1))
+
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+
+        # 4. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb
+                )
+        else:
+            for block in self.blocks:
+                hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb)
+
+        # 5. Output norm, projection & unpatchify
+        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
+
+        # Move the shift and scale tensors to the same device as hidden_states.
+        # When using multi-GPU inference via accelerate these will be on the
+        # first device rather than the last device, which hidden_states ends up
+        # on.
+        shift = shift.to(hidden_states.device)
+        scale = scale.to(hidden_states.device)
+
+        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)
@@ -0,0 +1,609 @@
+# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import regex as re
+import torch
+from transformers import AutoTokenizer, UMT5EncoderModel
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import WanLoraLoaderMixin
+from diffusers.models import AutoencoderKLWan, WanTransformer3DModel
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.wan.pipeline_output import WanPipelineOutput
+from einops import rearrange
+from transformers import UMT5EncoderModel, T5TokenizerFast
+
+from fastvideo.models.mochi_hf.modeling_wan import WanTransformer3DModel
+from fastvideo.utils.communications import all_gather
+from fastvideo.utils.parallel_states import get_sequence_parallel_state, nccl_info
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+if is_ftfy_available():
+    import ftfy
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from diffusers import AutoencoderKLWan, WanPipeline
+        >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+
+        >>> # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+        >>> model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
+        >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+        >>> pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+        >>> flow_shift = 5.0  # 5.0 for 720P, 3.0 for 480P
+        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+        >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+        >>> output = pipe(
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     height=720,
+        ...     width=1280,
+        ...     num_frames=81,
+        ...     guidance_scale=5.0,
+        ... ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=16)
+        ```
+"""
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+class WanPipeline(DiffusionPipeline, WanLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using Wan.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        tokenizer ([`T5Tokenizer`]):
+            Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer),
+            specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        transformer ([`WanTransformer3DModel`]):
+            Conditional Transformer to denoise the input latents.
+        scheduler ([`UniPCMultistepScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: UMT5EncoderModel,
+        transformer: WanTransformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt = [prompt_clean(u) for u in prompt]
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        negative_prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif negative_prompt is not None and (
+            not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list)
+        ):
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_latent_frames,
+            int(height) // self.vae_scale_factor_spatial,
+            int(width) // self.vae_scale_factor_spatial,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `480`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `832`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `81`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `5.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
+                `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
+                the text `prompt`, usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"np"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
+                The dtype to use for the torch.amp.autocast.
+
+        Examples:
+
+        Returns:
+            [`~WanPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where
+                the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        if num_frames % self.vae_scale_factor_temporal != 1:
+            logger.warning(
+                f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number."
+            )
+            num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
+        num_frames = max(num_frames, 1)
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        world_size, rank = nccl_info.sp_size, nccl_info.rank_within_group
+        if get_sequence_parallel_state():
+            latents = rearrange(latents, "b t (n s) h w -> b t n s h w", n=world_size).contiguous()
+            latents = latents[:, :, rank, :, :, :]
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+        
+        self._progress_bar_config = {"disable": nccl_info.rank_within_group != 0}
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = latents.to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if get_sequence_parallel_state():
+            latents = all_gather(latents, dim=2)
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return WanPipelineOutput(frames=video)
@@ -86,7 +86,7 @@ def inference(args):
                    num_inference_steps=args.num_inference_steps,
                    generator=generator,
                ).frames
-                if nccl_info.global_rank <= 0:
+                if nccl_info.global_rank == 0:
                    os.makedirs(args.output_path, exist_ok=True)
                    suffix = prompt.split(".")[0]
                    export_to_video(
@@ -107,7 +107,7 @@ def inference(args):
                generator=generator,
            ).frames

-        if nccl_info.global_rank <= 0:
+        if nccl_info.global_rank == 0:
            export_to_video(videos[0], args.output_path + ".mp4", fps=24)


@@ -94,7 +94,7 @@ def main(args):
                    guidance_scale=args.guidance_scale,
                    generator=generator,
                ).frames
-                if nccl_info.global_rank <= 0:
+                if nccl_info.global_rank == 0:
                    os.makedirs(args.output_path, exist_ok=True)
                    suffix = prompt.split(".")[0]
                    export_to_video(
@@ -116,7 +116,7 @@ def main(args):
                generator=generator,
            ).frames

-        if nccl_info.global_rank <= 0:
+        if nccl_info.global_rank == 0:
            export_to_video(videos[0], args.output_path + ".mp4", fps=30)


@@ -20,7 +20,7 @@ from torch.utils.data.distributed import DistributedSampler
 from tqdm.auto import tqdm

 from fastvideo.dataset.latent_datasets import (LatentDataset, latent_collate_function)
-from fastvideo.models.mochi_hf.mochi_latents_utils import normalize_dit_input
+from fastvideo.utils.latents_utils import normalize_dit_input
 from fastvideo.models.mochi_hf.pipeline_mochi import MochiPipeline
 from fastvideo.models.hunyuan_hf.pipeline_hunyuan import HunyuanVideoPipeline

@@ -185,7 +185,7 @@ def main(args):
    noise_random_generator = None

    # Handle the repository creation
-    if rank <= 0 and args.output_dir is not None:
+    if rank == 0 and args.output_dir is not None:
        os.makedirs(args.output_dir, exist_ok=True)

    # For mixed precision training we cast all non-trainable weights to half-precision
@@ -316,7 +316,7 @@ def main(args):
        len(train_dataloader) / args.gradient_accumulation_steps * args.sp_size / args.train_sp_batch_size)
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

-    if rank <= 0:
+    if rank == 0:
        project = args.tracker_project_name or "fastvideo"
        wandb.init(project=project, config=args)

@@ -393,7 +393,7 @@ def main(args):
            "grad_norm": grad_norm,
        })
        progress_bar.update(1)
-        if rank <= 0:
+        if rank == 0:
            wandb.log(
                {
                    "train_loss": loss,
@@ -520,13 +520,7 @@ if __name__ == "__main__":
        help=("Whether training should be resumed from a previous lora checkpoint. Use a path saved by"
              ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'),
    )
-    parser.add_argument(
-        "--logging_dir",
-        type=str,
-        default="logs",
-        help=("[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
-              " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."),
-    )
+

    # optimizer & scheduler & Training
    parser.add_argument("--num_train_epochs", type=int, default=100)
@@ -32,7 +32,7 @@ def save_checkpoint_optimizer(model, optimizer, rank, output_dir, step, discrimi
    save_dir = os.path.join(output_dir, f"checkpoint-{step}")
    os.makedirs(save_dir, exist_ok=True)
    # save using safetensors
-    if rank <= 0 and not discriminator:
+    if rank == 0 and not discriminator:
        weight_path = os.path.join(save_dir, "diffusion_pytorch_model.safetensors")
        save_file(cpu_state, weight_path)
        config_dict = dict(model.config)
@@ -60,7 +60,7 @@ def save_checkpoint(transformer, rank, output_dir, step):
    ):
        cpu_state = transformer.state_dict()
    # todo move to get_state_dict
-    if rank <= 0:
+    if rank == 0:
        save_dir = os.path.join(output_dir, f"checkpoint-{step}")
        os.makedirs(save_dir, exist_ok=True)
        # save using safetensors
@@ -98,7 +98,7 @@ def save_checkpoint_generator_discriminator(
    hf_weight_dir = os.path.join(save_dir, "hf_weights")
    os.makedirs(hf_weight_dir, exist_ok=True)
    # save using safetensors
-    if rank <= 0:
+    if rank == 0:
        config_dict = dict(model.config)
        config_path = os.path.join(hf_weight_dir, "config.json")
        # save dict as json
@@ -139,7 +139,7 @@ def save_checkpoint_generator_discriminator(
        optim_state = FSDP.optim_state_dict(discriminator, discriminator_optimizer)
        model_state = discriminator.state_dict()
        state_dict = {"optimizer": optim_state, "model": model_state}
-        if rank <= 0:
+        if rank == 0:
            discriminator_fsdp_state_fil = os.path.join(discriminator_fsdp_state_dir, "discriminator_state.pt")
            torch.save(state_dict, discriminator_fsdp_state_fil)

@@ -178,7 +178,7 @@ def load_full_state_model(model, optimizer, checkpoint_file, rank):
    ):
        discriminator_state = torch.load(checkpoint_file)
        model_state = discriminator_state["model"]
-        if rank <= 0:
+        if rank == 0:
            optim_state = discriminator_state["optimizer"]
        else:
            optim_state = None
@@ -241,7 +241,7 @@ def save_lora_checkpoint(transformer, optimizer, rank, output_dir, step, pipelin
            optimizer,
        )

-    if rank <= 0:
+    if rank == 0:
        save_dir = os.path.join(output_dir, f"lora-checkpoint-{step}")
        os.makedirs(save_dir, exist_ok=True)

@@ -70,6 +70,8 @@ DEFAULT_CONDA_PATTERNS = {
    "optree",
    "nccl",
    "transformers",
+    "accelerate",
+    "peft",
    "zmq",
    "nvidia",
    "pynvml",
@@ -85,6 +87,8 @@ DEFAULT_PIP_PATTERNS = {
    "onnx",
    "nccl",
    "transformers",
+    "accelerate",
+    "peft",
    "zmq",
    "nvidia",
    "pynvml",
@@ -1,38 +0,0 @@
-import platform
-
-import accelerate
-import peft
-import torch
-import transformers
-from transformers.utils import is_torch_cuda_available, is_torch_npu_available
-
-VERSION = "1.2.0"
-
-if __name__ == "__main__":
-    info = {
-        "FastVideo version": VERSION,
-        "Platform": platform.platform(),
-        "Python version": platform.python_version(),
-        "PyTorch version": torch.__version__,
-        "Transformers version": transformers.__version__,
-        "Accelerate version": accelerate.__version__,
-        "PEFT version": peft.__version__,
-    }
-
-    if is_torch_cuda_available():
-        info["PyTorch version"] += " (GPU)"
-        info["GPU type"] = torch.cuda.get_device_name()
-
-    if is_torch_npu_available():
-        info["PyTorch version"] += " (NPU)"
-        info["NPU type"] = torch.npu.get_device_name()
-        info["CANN version"] = torch.version.cann  # codespell:ignore
-
-    try:
-        import bitsandbytes
-
-        info["Bitsandbytes version"] = bitsandbytes.__version__
-    except Exception:
-        pass
-
-    print("\n" + "\n".join([f"- {key}: {value}" for key, value in info.items()]) + "\n")
@@ -0,0 +1,88 @@
+import torch
+
+mochi_latents_mean = torch.tensor([
+    -0.06730895953510081,
+    -0.038011381506090416,
+    -0.07477820912866141,
+    -0.05565264470995561,
+    0.012767231469026969,
+    -0.04703542746246419,
+    0.043896967884726704,
+    -0.09346305707025976,
+    -0.09918314763016893,
+    -0.008729793427399178,
+    -0.011931556316503654,
+    -0.0321993391887285,
+]).view(1, 12, 1, 1, 1)
+mochi_latents_std = torch.tensor([
+    0.9263795028493863,
+    0.9248894543193766,
+    0.9393059390890617,
+    0.959253732819592,
+    0.8244560132752793,
+    0.917259975397747,
+    0.9294154431013696,
+    1.3720942357788521,
+    0.881393668867029,
+    0.9168315692124348,
+    0.9185249279345552,
+    0.9274757570805041,
+]).view(1, 12, 1, 1, 1)
+mochi_scaling_factor = 1.0
+
+
+wan_latents_mean = torch.tensor([
+    -0.7571,
+    -0.7089,
+    -0.9113,
+    0.1075,
+    -0.1745,
+    0.9653,
+    -0.1517,
+    1.5508,
+    0.4134,
+    -0.0715,
+    0.5517,
+    -0.3632,
+    -0.1922,
+    -0.9497,
+    0.2503,
+    -0.2921,
+]).view(1, 16, 1, 1, 1)
+wan_latents_std = torch.tensor([
+    2.8184,
+    1.4541,
+    2.3275,
+    2.6558,
+    1.2196,
+    1.7708,
+    2.6052,
+    2.0743,
+    3.2687,
+    2.1526,
+    2.8652,
+    1.5579,
+    1.6382,
+    1.1253,
+    2.8251,
+    1.916,
+]).view(1, 16, 1, 1, 1)
+
+
+def normalize_dit_input(model_type, latents):
+    if model_type == "mochi":
+        latents_mean = mochi_latents_mean.to(latents.device, latents.dtype)
+        latents_std = mochi_latents_std.to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) / latents_std
+        return latents
+    elif model_type == "hunyuan_hf":
+        return latents * 0.476986
+    elif model_type == "hunyuan":
+        return latents * 0.476986
+    elif model_type == "wan":
+        latents_mean = wan_latents_mean.to(latents.device, latents.dtype)
+        latents_std = wan_latents_std.to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) / latents_std
+        return latents
+    else:
+        raise NotImplementedError(f"model_type {model_type} not supported")
@@ -3,9 +3,9 @@ from pathlib import Path

 import torch
 import torch.nn.functional as F
-from diffusers import AutoencoderKLHunyuanVideo, AutoencoderKLMochi
+from diffusers import AutoencoderKLHunyuanVideo, AutoencoderKLMochi, AutoencoderKLWan
 from torch import nn
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel

 from fastvideo.models.hunyuan.modules.models import (HYVideoDiffusionTransformer, MMDoubleStreamBlock,
                                                     MMSingleStreamBlock)
@@ -14,6 +14,7 @@ from fastvideo.models.hunyuan.vae.autoencoder_kl_causal_3d import AutoencoderKLC
 from fastvideo.models.hunyuan_hf.modeling_hunyuan import (HunyuanVideoSingleTransformerBlock,
                                                          HunyuanVideoTransformer3DModel, HunyuanVideoTransformerBlock)
 from fastvideo.models.mochi_hf.modeling_mochi import MochiTransformer3DModel, MochiTransformerBlock
+from fastvideo.models.wan_hf.modeling_wan import WanTransformer3DModel, WanTransformerBlock
 from fastvideo.utils.logging_ import main_print

 hunyuan_config = {
@@ -200,6 +201,48 @@ class MochiTextEncoderWrapper(nn.Module):

        return prompt_embeds, prompt_attention_mask

+class WanTextEncoderWrapper(nn.Module):
+
+    def __init__(self, pretrained_model_name_or_path, device):
+        super().__init__()
+        self.text_encoder = UMT5EncoderModel.from_pretrained(os.path.join(pretrained_model_name_or_path,
+                                                                        "text_encoder")).to(device)
+        self.tokenizer = AutoTokenizer.from_pretrained(os.path.join(pretrained_model_name_or_path, "tokenizer"))
+        self.max_sequence_length = 256
+
+    def encode_prompt(self, prompt):
+        device = self.text_encoder.device
+        dtype = self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.max_sequence_length - 1:-1])
+            main_print(f"Truncated text input: {prompt} to: {removed_text} for model input.")
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.view(batch_size, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+
+        return prompt_embeds, prompt_attention_mask

 def load_hunyuan_state_dict(model, dit_model_name_or_path):
    load_key = "module"
@@ -240,6 +283,20 @@ def load_transformer(
                torch_dtype=master_weight_type,
                # torch_dtype=torch.bfloat16 if args.use_lora else torch.float32,
            )
+    elif model_type == "wan":
+        if dit_model_name_or_path:
+            transformer = WanTransformer3DModel.from_pretrained(
+                dit_model_name_or_path,
+                torch_dtype=master_weight_type,
+                # torch_dtype=torch.bfloat16 if args.use_lora else torch.float32,
+            )
+        else:
+            transformer = WanTransformer3DModel.from_pretrained(
+                pretrained_model_name_or_path,
+                subfolder="transformer",
+                torch_dtype=master_weight_type,
+                # torch_dtype=torch.bfloat16 if args.use_lora else torch.float32,
+            )
    elif model_type == "hunyuan_hf":
        if dit_model_name_or_path:
            transformer = HunyuanVideoTransformer3DModel.from_pretrained(
@@ -283,6 +340,12 @@ def load_vae(model_type, pretrained_model_name_or_path):
                                                        torch_dtype=weight_dtype).to("cuda")
        autocast_type = torch.bfloat16
        fps = 24
+    elif model_type == "wan":
+        vae = AutoencoderKLWan.from_pretrained(pretrained_model_name_or_path,
+                                                 subfolder="vae",
+                                                 torch_dtype=weight_dtype).to("cuda")
+        autocast_type = torch.bfloat16
+        fps = 24
    elif model_type == "hunyuan":
        vae_precision = torch.float32
        vae_path = os.path.join(pretrained_model_name_or_path, "hunyuan-video-t2v-720p/vae")
@@ -311,6 +374,8 @@ def load_vae(model_type, pretrained_model_name_or_path):
 def load_text_encoder(model_type, pretrained_model_name_or_path, device):
    if model_type == "mochi":
        text_encoder = MochiTextEncoderWrapper(pretrained_model_name_or_path, device)
+    elif model_type == "wan":
+        text_encoder = WanTextEncoderWrapper(pretrained_model_name_or_path, device)
    elif model_type == "hunyuan" or "hunyuan_hf":
        text_encoder = HunyuanTextEncoderWrapper(pretrained_model_name_or_path, device)
    else:
@@ -322,6 +387,8 @@ def get_no_split_modules(transformer):
    # if of type MochiTransformer3DModel
    if isinstance(transformer, MochiTransformer3DModel):
        return (MochiTransformerBlock, )
+    elif isinstance(transformer, WanTransformer3DModel):
+        return (WanTransformerBlock, )
    elif isinstance(transformer, HunyuanVideoTransformer3DModel):
        return (HunyuanVideoSingleTransformerBlock, HunyuanVideoTransformerBlock)
    elif isinstance(transformer, HYVideoDiffusionTransformer):
@@ -129,13 +129,22 @@ def sample_validation_video(
            # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
            timestep = t.expand(latent_model_input.shape[0])
            with torch.autocast("cuda", dtype=torch.bfloat16):
-                noise_pred = transformer(
-                    hidden_states=latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep=timestep,
-                    encoder_attention_mask=prompt_attention_mask,
-                    return_dict=False,
-                )[0]
+                if model_type == "wan":
+                    pred_kwargs = {
+                        "hidden_states": latent_model_input,
+                        "encoder_hidden_states": prompt_embeds,
+                        "timestep":timestep,
+                        "return_dict":False,
+                    }
+                else:
+                    pred_kwargs = {
+                        "hidden_states": latent_model_input,
+                        "encoder_hidden_states": prompt_embeds,
+                        "timestep":timestep,
+                        "encoder_attention_mask":prompt_attention_mask,
+                        "return_dict":False,
+                    }
+                noise_pred = transformer(**pred_kwargs)[0]

            # Mochi CFG + Sampling runs in FP32
            noise_pred = noise_pred.to(torch.float32)
@@ -166,10 +175,12 @@ def sample_validation_video(
        # denormalize with the mean and std if available and not None
        has_latents_mean = (hasattr(vae.config, "latents_mean") and vae.config.latents_mean is not None)
        has_latents_std = (hasattr(vae.config, "latents_std") and vae.config.latents_std is not None)
+        if model_type == "wan":
+            vae.config.scaling_factor = 1
        if has_latents_mean and has_latents_std:
-            latents_mean = (torch.tensor(vae.config.latents_mean).view(1, 12, 1, 1,
+            latents_mean = (torch.tensor(vae.config.latents_mean).view(1, num_channels_latents, 1, 1,
                                                                       1).to(latents.device, latents.dtype))
-            latents_std = (torch.tensor(vae.config.latents_std).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype))
+            latents_std = (torch.tensor(vae.config.latents_std).view(1, num_channels_latents, 1, 1, 1).to(latents.device, latents.dtype))
            latents = latents * latents_std / vae.config.scaling_factor + latents_mean
        else:
            latents = latents / vae.config.scaling_factor
@@ -202,14 +213,15 @@ def log_validation(
        vae_spatial_scale_factor = 8
        vae_temporal_scale_factor = 6
        num_channels_latents = 12
-    elif args.model_type == "hunyuan" or "hunyuan_hf":
+    elif args.model_type == "hunyuan" or "hunyuan_hf" or "wan":
        vae_spatial_scale_factor = 8
        vae_temporal_scale_factor = 4
        num_channels_latents = 16
    else:
        raise ValueError(f"Model type {args.model_type} not supported")
    vae, autocast_type, fps = load_vae(args.model_type, args.pretrained_model_name_or_path)
-    vae.enable_tiling()
+    if args.model_type != "wan":
+        vae.enable_tiling()
    if scheduler_type == "euler":
        scheduler = FlowMatchEulerDiscreteScheduler(shift=shift)
    else:
@@ -0,0 +1,407 @@
+# SPDX-License-Identifier: Apache-2.0
+import json
+import os
+from collections import defaultdict
+from typing import Any, Dict, List, Optional, Tuple
+
+import numpy as np
+
+from fastvideo.v1.utils import dict_to_3d_list
+
+
+def configure_sta(mode: str = 'STA_searching',
+                  layer_num: int = 40,
+                  time_step_num: int = 50,
+                  head_num: int = 40,
+                  **kwargs) -> List[List[List[Any]]]:
+    """
+    Configure Sliding Tile Attention (STA) parameters based on the specified mode.
+    
+    Parameters:
+    ----------
+    mode : str
+        The STA mode to use. Options are:
+        - 'STA_searching': Generate a set of mask candidates for initial search
+        - 'STA_tuning': Select best mask strategy based on previously saved results
+        - 'STA_inference': Load and use a previously tuned mask strategy
+    layer_num: int, number of layers
+    time_step_num: int, number of timesteps
+    head_num: int, number of heads
+    
+    **kwargs : dict
+        Mode-specific parameters:
+        
+        For 'STA_searching':
+        - mask_candidates: list of str, optional, mask candidates to use
+        - mask_selected: list of int, optional, indices of selected masks
+        
+        For 'STA_tuning':
+        - mask_search_files_path: str, required, path to mask search results
+        - mask_candidates: list of str, optional, mask candidates to use
+        - mask_selected: list of int, optional, indices of selected masks
+        - skip_time_steps: int, optional, number of time steps to use full attention (default 12)
+        - save_dir: str, optional, directory to save mask strategy (default "mask_candidates")
+        
+        For 'STA_inference':
+        - load_path: str, optional, path to load mask strategy (default "mask_candidates/mask_strategy.json")
+    """
+    valid_modes = [
+        'STA_searching', 'STA_tuning', 'STA_inference', 'STA_tuning_cfg'
+    ]
+    if mode not in valid_modes:
+        raise ValueError(f"Mode must be one of {valid_modes}, got {mode}")
+
+    if mode == 'STA_searching':
+        # Get parameters with defaults
+        mask_candidates: Optional[List[str]] = kwargs.get('mask_candidates')
+        if mask_candidates is None:
+            raise ValueError(
+                "mask_candidates is required for STA_searching mode")
+        mask_selected: List[int] = kwargs.get('mask_selected',
+                                              list(range(len(mask_candidates))))
+
+        # Parse selected masks
+        selected_masks: List[List[int]] = []
+        for index in mask_selected:
+            mask = mask_candidates[index]
+            masks_list = [int(x) for x in mask.split(',')]
+            selected_masks.append(masks_list)
+
+        # Create 3D mask structure with fixed dimensions (t=50, l=60)
+        masks_3d: List[List[List[List[int]]]] = []
+        for i in range(time_step_num):  # Fixed t dimension = 50
+            row = []
+            for j in range(layer_num):  # Fixed l dimension = 60
+                row.append(selected_masks)  # Add all masks at each position
+            masks_3d.append(row)
+
+        return masks_3d
+
+    elif mode == 'STA_tuning':
+        # Get required parameters
+        mask_search_files_path: Optional[str] = kwargs.get(
+            'mask_search_files_path')
+        if not mask_search_files_path:
+            raise ValueError(
+                "mask_search_files_path is required for STA_tuning mode")
+
+        # Get optional parameters with defaults
+        mask_candidates_tuning: Optional[List[str]] = kwargs.get(
+            'mask_candidates')
+        if mask_candidates_tuning is None:
+            raise ValueError("mask_candidates is required for STA_tuning mode")
+        mask_selected_tuning: List[int] = kwargs.get(
+            'mask_selected', list(range(len(mask_candidates_tuning))))
+        skip_time_steps_tuning: Optional[int] = kwargs.get('skip_time_steps')
+        save_dir_tuning: Optional[str] = kwargs.get('save_dir',
+                                                    "mask_candidates")
+
+        # Parse selected masks
+        selected_masks_tuning: List[List[int]] = []
+        for index in mask_selected_tuning:
+            mask = mask_candidates_tuning[index]
+            masks_list = [int(x) for x in mask.split(',')]
+            selected_masks_tuning.append(masks_list)
+
+        # Read JSON results
+        results = read_specific_json_files(mask_search_files_path)
+        averaged_results = average_head_losses(results, selected_masks_tuning)
+
+        # Add full attention mask for specific cases
+        full_attention_mask_tuning: Optional[List[int]] = kwargs.get(
+            'full_attention_mask')
+        if full_attention_mask_tuning is not None:
+            selected_masks_tuning.append(full_attention_mask_tuning)
+
+        # Select best mask strategy
+        timesteps_tuning: int = kwargs.get('timesteps', time_step_num)
+        if skip_time_steps_tuning is None:
+            skip_time_steps_tuning = 12
+        mask_strategy, sparsity, strategy_counts = select_best_mask_strategy(
+            averaged_results, selected_masks_tuning, skip_time_steps_tuning,
+            timesteps_tuning, head_num)
+
+        # Save mask strategy
+        if save_dir_tuning is not None:
+            os.makedirs(save_dir_tuning, exist_ok=True)
+            file_path = os.path.join(
+                save_dir_tuning,
+                f'mask_strategy_s{skip_time_steps_tuning}.json')
+            with open(file_path, 'w') as f:
+                json.dump(mask_strategy, f, indent=4)
+            print(f"Successfully saved mask_strategy to {file_path}")
+
+        # Print sparsity and strategy counts for information
+        print(f"Overall sparsity: {sparsity:.4f}")
+        print("\nStrategy usage counts:")
+        total_heads = time_step_num * layer_num * head_num  # Fixed dimensions
+        for strategy, count in strategy_counts.items():
+            print(
+                f"Strategy {strategy}: {count} heads ({count/total_heads*100:.2f}%)"
+            )
+
+        # Convert dictionary to 3D list with fixed dimensions
+        mask_strategy_3d = dict_to_3d_list(mask_strategy,
+                                           t_max=time_step_num,
+                                           l_max=layer_num,
+                                           h_max=head_num)
+
+        return mask_strategy_3d
+    elif mode == 'STA_tuning_cfg':
+        # Get required parameters for both positive and negative paths
+        mask_search_files_path_pos: Optional[str] = kwargs.get(
+            'mask_search_files_path_pos')
+        mask_search_files_path_neg: Optional[str] = kwargs.get(
+            'mask_search_files_path_neg')
+        save_dir_cfg: Optional[str] = kwargs.get('save_dir')
+
+        if not mask_search_files_path_pos or not mask_search_files_path_neg or not save_dir_cfg:
+            raise ValueError(
+                "mask_search_files_path_pos, mask_search_files_path_neg, and save_dir are required for STA_tuning_cfg mode"
+            )
+
+        # Get optional parameters with defaults
+        mask_candidates_cfg: Optional[List[str]] = kwargs.get('mask_candidates')
+        if mask_candidates_cfg is None:
+            raise ValueError(
+                "mask_candidates is required for STA_tuning_cfg mode")
+        mask_selected_cfg: List[int] = kwargs.get(
+            'mask_selected', list(range(len(mask_candidates_cfg))))
+        skip_time_steps_cfg: Optional[int] = kwargs.get('skip_time_steps')
+
+        # Parse selected masks
+        selected_masks_cfg: List[List[int]] = []
+        for index in mask_selected_cfg:
+            mask = mask_candidates_cfg[index]
+            masks_list = [int(x) for x in mask.split(',')]
+            selected_masks_cfg.append(masks_list)
+
+        # Read JSON results for both positive and negative paths
+        pos_results = read_specific_json_files(mask_search_files_path_pos)
+        neg_results = read_specific_json_files(mask_search_files_path_neg)
+        # Combine positive and negative results into one list
+        combined_results = pos_results + neg_results
+
+        # Average the combined results
+        averaged_results = average_head_losses(combined_results,
+                                               selected_masks_cfg)
+
+        # Add full attention mask for specific cases
+        full_attention_mask_cfg: Optional[List[int]] = kwargs.get(
+            'full_attention_mask')
+        if full_attention_mask_cfg is not None:
+            selected_masks_cfg.append(full_attention_mask_cfg)
+
+        timesteps_cfg: int = kwargs.get('timesteps', time_step_num)
+        if skip_time_steps_cfg is None:
+            skip_time_steps_cfg = 12
+        # Select best mask strategy using combined results
+        mask_strategy, sparsity, strategy_counts = select_best_mask_strategy(
+            averaged_results, selected_masks_cfg, skip_time_steps_cfg,
+            timesteps_cfg, head_num)
+
+        # Save mask strategy
+        os.makedirs(save_dir_cfg, exist_ok=True)
+        file_path = os.path.join(save_dir_cfg,
+                                 f'mask_strategy_s{skip_time_steps_cfg}.json')
+        with open(file_path, 'w') as f:
+            json.dump(mask_strategy, f, indent=4)
+        print(f"Successfully saved mask_strategy to {file_path}")
+
+        # Print sparsity and strategy counts for information
+        print(f"Overall sparsity: {sparsity:.4f}")
+        print("\nStrategy usage counts:")
+        total_heads = time_step_num * layer_num * head_num  # Fixed dimensions
+        for strategy, count in strategy_counts.items():
+            print(
+                f"Strategy {strategy}: {count} heads ({count/total_heads*100:.2f}%)"
+            )
+
+        # Convert dictionary to 3D list with fixed dimensions
+        mask_strategy_3d = dict_to_3d_list(mask_strategy,
+                                           t_max=time_step_num,
+                                           l_max=layer_num,
+                                           h_max=head_num)
+
+        return mask_strategy_3d
+
+    else:  # STA_inference
+        # Get parameters with defaults
+        load_path: Optional[str] = kwargs.get(
+            'load_path', "mask_candidates/mask_strategy.json")
+        if load_path is None:
+            raise ValueError("load_path is required for STA_inference mode")
+
+        # Load previously saved mask strategy
+        with open(load_path) as f:
+            mask_strategy = json.load(f)
+
+        # Convert dictionary to 3D list with fixed dimensions
+        mask_strategy_3d = dict_to_3d_list(mask_strategy,
+                                           t_max=time_step_num,
+                                           l_max=layer_num,
+                                           h_max=head_num)
+
+        return mask_strategy_3d
+
+
+# Helper functions
+
+
+def read_specific_json_files(folder_path: str) -> List[Dict[str, Any]]:
+    """Read and parse JSON files containing mask search results."""
+    json_contents: List[Dict[str, Any]] = []
+
+    # List files only in the current directory (no walk)
+    files = os.listdir(folder_path)
+    # Filter files
+    matching_files = [f for f in files if 'mask' in f and f.endswith('.json')]
+    print(f"Found {len(matching_files)} matching files: {matching_files}")
+
+    for file_name in matching_files:
+        file_path = os.path.join(folder_path, file_name)
+        with open(file_path) as file:
+            data = json.load(file)
+            json_contents.append(data)
+
+    return json_contents
+
+
+def average_head_losses(
+        results: List[Dict[str, Any]],
+        selected_masks: List[List[int]]) -> Dict[str, Dict[str, np.ndarray]]:
+    """Average losses across all prompts for each mask strategy."""
+    # Initialize a dictionary to store the averaged results
+    averaged_losses: Dict[str, Dict[str, np.ndarray]] = {}
+    loss_type = 'L2_loss'
+    # Get all loss types (e.g., 'L2_loss')
+    averaged_losses[loss_type] = {}
+
+    for mask in selected_masks:
+        mask_str = str(mask)
+        data_shape = np.array(results[0][loss_type][mask_str]).shape
+        accumulated_data = np.zeros(data_shape)
+
+        # Sum across all prompts
+        for prompt_result in results:
+            accumulated_data += np.array(prompt_result[loss_type][mask_str])
+
+        # Average by dividing by number of prompts
+        averaged_data = accumulated_data / len(results)
+        averaged_losses[loss_type][mask_str] = averaged_data
+
+    return averaged_losses
+
+
+def select_best_mask_strategy(
+        averaged_results: Dict[str, Dict[str, np.ndarray]],
+        selected_masks: List[List[int]],
+        skip_time_steps: int = 12,
+        timesteps: int = 50,
+        head_num: int = 40
+) -> Tuple[Dict[str, List[int]], float, Dict[str, int]]:
+    """Select the best mask strategy for each head based on loss minimization."""
+    best_mask_strategy: Dict[str, List[int]] = {}
+    loss_type = 'L2_loss'
+    # Get the shape of time steps and layers
+    layers = len(averaged_results[loss_type][str(selected_masks[0])][0])
+
+    # Counter for sparsity calculation
+    total_tokens = 0  # total number of masked tokens
+    total_length = 0  # total sequence length
+
+    strategy_counts: Dict[str, int] = {
+        str(strategy): 0
+        for strategy in selected_masks
+    }
+    full_attn_strategy = selected_masks[-1]  # Last strategy is full attention
+    print(f"Strategy {full_attn_strategy}, skip first {skip_time_steps} steps ")
+
+    for t in range(timesteps):
+        for layer_idx in range(layers):
+            for h in range(head_num):
+                if t < skip_time_steps:  # First steps use full attention
+                    strategy = full_attn_strategy
+                else:
+                    # Get losses for this head across all strategies
+                    head_losses = []
+                    for strategy in selected_masks[:
+                                                   -1]:  # Exclude full attention
+                        head_losses.append(averaged_results[loss_type][str(
+                            strategy)][t][layer_idx][h])
+
+                    # Find which strategy gives minimum loss
+                    best_strategy_idx = np.argmin(head_losses)
+                    strategy = selected_masks[best_strategy_idx]
+
+                best_mask_strategy[f'{t}_{layer_idx}_{h}'] = strategy
+
+                # Calculate sparsity
+                nums = strategy  # strategy is already a list of numbers
+                total_tokens += nums[0] * nums[1] * nums[
+                    2]  # masked tokens for chosen strategy
+                total_length += full_attn_strategy[0] * full_attn_strategy[
+                    1] * full_attn_strategy[2]
+
+                # Count strategy usage
+                strategy_counts[str(strategy)] += 1
+
+    overall_sparsity = 1 - total_tokens / total_length
+
+    return best_mask_strategy, overall_sparsity, strategy_counts
+
+
+def save_mask_search_results(
+        mask_search_final_result: List[Dict[str, List[float]]],
+        prompt: str,
+        mask_strategies: List[str],
+        output_dir: str = 'output/mask_search_result/') -> Optional[str]:
+    if not mask_search_final_result:
+        print("No mask search results to save")
+        return None
+
+    # Create result dictionary with defaultdict for nested lists
+    mask_search_dict: Dict[str, Dict[str, List[List[float]]]] = {
+        "L2_loss": defaultdict(list),
+        "L1_loss": defaultdict(list)
+    }
+
+    mask_selected = list(range(len(mask_strategies)))
+    selected_masks: List[List[int]] = []
+    for index in mask_selected:
+        mask = mask_strategies[index]
+        masks_list = [int(x) for x in mask.split(',')]
+        selected_masks.append(masks_list)
+
+    # Process each mask strategy
+    for i, mask_strategy in enumerate(selected_masks):
+        mask_strategy_str = str(mask_strategy)
+        # Process L2 loss
+        step_results: List[List[float]] = []
+        for step_data in mask_search_final_result:
+            if isinstance(step_data, dict) and "L2_loss" in step_data:
+                layer_losses = [float(loss) for loss in step_data["L2_loss"]]
+                step_results.append(layer_losses)
+        mask_search_dict["L2_loss"][mask_strategy_str] = step_results
+
+        step_results = []
+        for step_data in mask_search_final_result:
+            if isinstance(step_data, dict) and "L1_loss" in step_data:
+                layer_losses = [float(loss) for loss in step_data["L1_loss"]]
+                step_results.append(layer_losses)
+        mask_search_dict["L1_loss"][mask_strategy_str] = step_results
+
+    # Create the output directory if it doesn't exist
+    os.makedirs(output_dir, exist_ok=True)
+
+    # Create a filename based on the first 20 characters of the prompt
+    filename = prompt[:50].replace(" ", "_")
+    filepath = os.path.join(output_dir, f'mask_search_{filename}.json')
+
+    # Save the results to a JSON file
+    with open(filepath, 'w') as f:
+        json.dump(mask_search_dict, f, indent=4)
+
+    print(f"Successfully saved mask research results to {filepath}")
+
+    return filepath
@@ -3,12 +3,15 @@
 from fastvideo.v1.attention.backends.abstract import (AttentionBackend,
                                                      AttentionMetadata,
                                                      AttentionMetadataBuilder)
-from fastvideo.v1.attention.layer import DistributedAttention, LocalAttention
+from fastvideo.v1.attention.layer import (DistributedAttention,
+                                          DistributedAttention_VSA,
+                                          LocalAttention)
 from fastvideo.v1.attention.selector import get_attn_backend

 __all__ = [
    "DistributedAttention",
    "LocalAttention",
+    "DistributedAttention_VSA",
    "AttentionBackend",
    "AttentionMetadata",
    "AttentionMetadataBuilder",
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from typing import List, Optional, Type

 import torch
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from typing import List, Optional, Type

 import torch
@@ -1,6 +1,7 @@
+# SPDX-License-Identifier: Apache-2.0
 import json
 from dataclasses import dataclass
-from typing import List, Optional, Type
+from typing import Any, List, Optional, Type

 import torch
 from einops import rearrange
@@ -13,43 +14,24 @@ from fastvideo.v1.attention.backends.abstract import (AttentionBackend,
                                                      AttentionMetadataBuilder)
 from fastvideo.v1.distributed import get_sp_group
 from fastvideo.v1.fastvideo_args import FastVideoArgs
+from fastvideo.v1.forward_context import ForwardContext, get_forward_context
 from fastvideo.v1.logger import init_logger
 from fastvideo.v1.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.v1.utils import dict_to_3d_list

 logger = init_logger(__name__)


-# TODO(will-refactor): move this to a utils file
-def dict_to_3d_list(mask_strategy) -> List[List[List[Optional[torch.Tensor]]]]:
-    indices = [tuple(map(int, key.split('_'))) for key in mask_strategy]
-
-    max_timesteps_idx = max(
-        timesteps_idx for timesteps_idx, layer_idx, head_idx in indices) + 1
-    max_layer_idx = max(layer_idx
-                        for timesteps_idx, layer_idx, head_idx in indices) + 1
-    max_head_idx = max(head_idx
-                       for timesteps_idx, layer_idx, head_idx in indices) + 1
-
-    result = [[[None for _ in range(max_head_idx)]
-               for _ in range(max_layer_idx)] for _ in range(max_timesteps_idx)]
-
-    for key, value in mask_strategy.items():
-        timesteps_idx, layer_idx, head_idx = map(int, key.split('_'))
-        result[timesteps_idx][layer_idx][head_idx] = value
-
-    return result
-
-
 class RangeDict(dict):

-    def __getitem__(self, item):
+    def __getitem__(self, item: int) -> str:
        for key in self.keys():
            if isinstance(key, tuple):
                low, high = key
                if low <= item <= high:
-                    return super().__getitem__(key)
+                    return str(super().__getitem__(key))
            elif key == item:
-                return super().__getitem__(key)
+                return str(super().__getitem__(key))
        raise KeyError(f"seq_len {item} not supported for STA")


@@ -82,6 +64,8 @@ class SlidingTileAttentionBackend(AttentionBackend):
@dataclass
 class SlidingTileAttentionMetadata(AttentionMetadata):
    current_timestep: int
+    STA_param: List[List[
+        Any]]  # each timestep with one metadata, shape [num_layers, num_heads]


 class SlidingTileAttentionMetadataBuilder(AttentionMetadataBuilder):
@@ -98,8 +82,12 @@ class SlidingTileAttentionMetadataBuilder(AttentionMetadataBuilder):
        forward_batch: ForwardBatch,
        fastvideo_args: FastVideoArgs,
    ) -> SlidingTileAttentionMetadata:
-
-        return SlidingTileAttentionMetadata(current_timestep=current_timestep, )
+        param = forward_batch.STA_param
+        if param is None:
+            return SlidingTileAttentionMetadata(
+                current_timestep=current_timestep, STA_param=[])
+        return SlidingTileAttentionMetadata(current_timestep=current_timestep,
+                                            STA_param=param[current_timestep])


 class SlidingTileAttentionImpl(AttentionImpl):
@@ -120,17 +108,17 @@ class SlidingTileAttentionImpl(AttentionImpl):
        if config_file is None:
            raise ValueError("FASTVIDEO_ATTENTION_CONFIG is not set")

+        # TODO(kevin): get mask strategy for different STA modes
        with open(config_file) as f:
            mask_strategy = json.load(f)
-        mask_strategy = dict_to_3d_list(mask_strategy)
+        self.mask_strategy = dict_to_3d_list(mask_strategy)

        self.prefix = prefix
-        self.mask_strategy = mask_strategy
        sp_group = get_sp_group()
        self.sp_size = sp_group.world_size
        # STA config
        self.STA_base_tile_size = [6, 8, 8]
-        self.img_latent_shape_mapping = RangeDict({
+        self.dit_seq_shape_mapping = RangeDict({
            (115200, 115456): '30x48x80',
            82944: '36x48x48',
            69120: '18x48x80',
@@ -145,9 +133,9 @@ class SlidingTileAttentionImpl(AttentionImpl):
        x = rearrange(x,
                      "b (sp t h w) head d -> b (t sp h w) head d",
                      sp=self.sp_size,
-                      t=self.img_latent_shape_int[0] // self.sp_size,
-                      h=self.img_latent_shape_int[1],
-                      w=self.img_latent_shape_int[2])
+                      t=self.dit_seq_shape_int[0] // self.sp_size,
+                      h=self.dit_seq_shape_int[1],
+                      w=self.dit_seq_shape_int[2])
        return rearrange(
            x,
            "b (n_t ts_t n_h ts_h n_w ts_w) h d -> b (n_t n_h n_w ts_t ts_h ts_w) h d",
@@ -171,9 +159,9 @@ class SlidingTileAttentionImpl(AttentionImpl):
        return rearrange(x,
                         "b (t sp h w) head d -> b (sp t h w) head d",
                         sp=self.sp_size,
-                         t=self.img_latent_shape_int[0] // self.sp_size,
-                         h=self.img_latent_shape_int[1],
-                         w=self.img_latent_shape_int[2])
+                         t=self.dit_seq_shape_int[0] // self.sp_size,
+                         h=self.dit_seq_shape_int[1],
+                         w=self.dit_seq_shape_int[2])

    def preprocess_qkv(
        self,
@@ -181,14 +169,12 @@ class SlidingTileAttentionImpl(AttentionImpl):
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        img_sequence_length = qkv.shape[1]
-        self.img_latent_shape_str = self.img_latent_shape_mapping[
-            img_sequence_length]
-        self.full_window_size = self.full_window_mapping[
-            self.img_latent_shape_str]
-        self.img_latent_shape_int = list(
-            map(int, self.img_latent_shape_str.split('x')))
-        self.img_seq_length = self.img_latent_shape_int[
-            0] * self.img_latent_shape_int[1] * self.img_latent_shape_int[2]
+        self.dit_seq_shape_str = self.dit_seq_shape_mapping[img_sequence_length]
+        self.full_window_size = self.full_window_mapping[self.dit_seq_shape_str]
+        self.dit_seq_shape_int = list(
+            map(int, self.dit_seq_shape_str.split('x')))
+        self.img_seq_length = self.dit_seq_shape_int[
+            0] * self.dit_seq_shape_int[1] * self.dit_seq_shape_int[2]
        return self.tile(qkv)

    def postprocess_output(
@@ -205,16 +191,24 @@ class SlidingTileAttentionImpl(AttentionImpl):
        v: torch.Tensor,
        attn_metadata: SlidingTileAttentionMetadata,
    ) -> torch.Tensor:
-
-        assert self.mask_strategy is not None, "mask_strategy cannot be None for SlidingTileAttention"
-        assert self.mask_strategy[
-            0] is not None, "mask_strategy[0] cannot be None for SlidingTileAttention"
+        if self.mask_strategy is None:
+            raise ValueError(
+                "mask_strategy cannot be None for SlidingTileAttention")
+        if self.mask_strategy[0] is None:
+            raise ValueError(
+                "mask_strategy[0] cannot be None for SlidingTileAttention")

        timestep = attn_metadata.current_timestep
+        forward_context: ForwardContext = get_forward_context()
+        forward_batch = forward_context.forward_batch
+        if forward_batch is None:
+            raise ValueError("forward_batch cannot be None")
        # pattern:'.double_blocks.0.attn.impl' or '.single_blocks.0.attn.impl'
        layer_idx = int(self.prefix.split('.')[-3])
-
-        # TODO: remove hardcode
+        if attn_metadata.STA_param is None or len(
+                attn_metadata.STA_param) <= layer_idx:
+            raise ValueError("Invalid STA_param")
+        STA_param = attn_metadata.STA_param[layer_idx]

        text_length = q.shape[1] - self.img_seq_length
        has_text = text_length > 0
@@ -227,15 +221,56 @@ class SlidingTileAttentionImpl(AttentionImpl):
        sp_group = get_sp_group()
        current_rank = sp_group.rank_in_group
        start_head = current_rank * head_num
-        windows = [
-            self.mask_strategy[timestep][layer_idx][head_idx + start_head]
-            for head_idx in range(head_num)
-        ]
-        # if has_text is False:
-        #     from IPython import embed
-        #     embed()
-        hidden_states = sliding_tile_attention(
-            query, key, value, windows, text_length, has_text,
-            self.img_latent_shape_str).transpose(1, 2)
+
+        # searching or tuning mode
+        if len(STA_param) < head_num * sp_group.world_size:
+            sparse_attn_hidden_states_all = []
+            full_mask_window = STA_param[-1]
+            for window_size in STA_param[:-1]:
+                sparse_hidden_states = sliding_tile_attention(
+                    query, key, value, [window_size] * head_num, text_length,
+                    has_text, self.dit_seq_shape_str).transpose(1, 2)
+                sparse_attn_hidden_states_all.append(sparse_hidden_states)
+
+            hidden_states = sliding_tile_attention(
+                query, key, value, [full_mask_window] * head_num, text_length,
+                has_text, self.dit_seq_shape_str).transpose(1, 2)
+
+            attn_L2_loss = []
+            attn_L1_loss = []
+            # average loss across all heads
+            for sparse_attn_hidden_states in sparse_attn_hidden_states_all:
+                # L2 loss
+                attn_L2_loss_ = torch.mean((sparse_attn_hidden_states.float() -
+                                            hidden_states.float())**2,
+                                           dim=[0, 1, 3]).cpu().numpy()
+                attn_L2_loss_ = [round(float(x), 6) for x in attn_L2_loss_]
+                attn_L2_loss.append(attn_L2_loss_)
+                # L1 loss
+                attn_L1_loss_ = torch.mean(
+                    torch.abs(sparse_attn_hidden_states.float() -
+                              hidden_states.float()),
+                    dim=[0, 1, 3]).cpu().numpy()
+                attn_L1_loss_ = [round(float(x), 6) for x in attn_L1_loss_]
+                attn_L1_loss.append(attn_L1_loss_)
+
+            layer_loss_save = {"L2_loss": attn_L2_loss, "L1_loss": attn_L1_loss}
+
+            if forward_batch.is_cfg_negative:
+                if forward_batch.mask_search_final_result_neg is not None:
+                    forward_batch.mask_search_final_result_neg[timestep].append(
+                        layer_loss_save)
+            else:
+                if forward_batch.mask_search_final_result_pos is not None:
+                    forward_batch.mask_search_final_result_pos[timestep].append(
+                        layer_loss_save)
+        else:
+            windows = [
+                STA_param[head_idx + start_head] for head_idx in range(head_num)
+            ]
+
+            hidden_states = sliding_tile_attention(
+                query, key, value, windows, text_length, has_text,
+                self.dit_seq_shape_str).transpose(1, 2)

        return hidden_states
@@ -0,0 +1,198 @@
+# SPDX-License-Identifier: Apache-2.0
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Type
+
+import torch
+from einops import rearrange
+
+try:
+    from vsa import video_sparse_attn
+except ImportError:
+    video_sparse_attn = None
+
+from fastvideo.v1.attention.backends.abstract import (AttentionBackend,
+                                                      AttentionImpl,
+                                                      AttentionMetadata,
+                                                      AttentionMetadataBuilder)
+from fastvideo.v1.distributed import get_sp_group
+from fastvideo.v1.fastvideo_args import FastVideoArgs
+from fastvideo.v1.logger import init_logger
+from fastvideo.v1.pipelines.pipeline_batch_info import ForwardBatch
+
+logger = init_logger(__name__)
+
+
+class VideoSparseAttentionBackend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> List[int]:
+        return [64, 128]
+
+    @staticmethod
+    def get_name() -> str:
+        return "VIDEO_SPARSE_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> Type["VideoSparseAttentionImpl"]:
+        return VideoSparseAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["VideoSparseAttentionMetadata"]:
+        return VideoSparseAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["VideoSparseAttentionMetadataBuilder"]:
+        return VideoSparseAttentionMetadataBuilder
+
+
+@dataclass
+class VideoSparseAttentionMetadata(AttentionMetadata):
+    current_timestep: int
+    dit_seq_shape: List[int]
+    VSA_sparsity: float
+
+
+class VideoSparseAttentionMetadataBuilder(AttentionMetadataBuilder):
+
+    def __init__(self):
+        pass
+
+    def prepare(self):
+        pass
+
+    def build(
+        self,
+        current_timestep: int,
+        forward_batch: ForwardBatch,
+        fastvideo_args: FastVideoArgs,
+    ) -> VideoSparseAttentionMetadata:
+        if forward_batch.latents is None:
+            raise ValueError("latents cannot be None")
+
+        raw_latent_shape = forward_batch.raw_latent_shape
+        if raw_latent_shape is None:
+            raise ValueError("raw_latent_shape cannot be None")
+
+        patch_size = fastvideo_args.pipeline_config.dit_config.patch_size
+        dit_seq_shape = [
+            raw_latent_shape[2] // patch_size[0],
+            raw_latent_shape[3] // patch_size[1],
+            raw_latent_shape[4] // patch_size[2]
+        ]
+        VSA_sparsity = forward_batch.VSA_sparsity
+
+        return VideoSparseAttentionMetadata(current_timestep=current_timestep,
+                                            dit_seq_shape=dit_seq_shape,
+                                            VSA_sparsity=VSA_sparsity)
+
+
+class VideoSparseAttentionImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: Optional[int] = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.prefix = prefix
+        sp_group = get_sp_group()
+        self.sp_size = sp_group.world_size
+        self.VSA_base_tile_size = [4, 4, 4]
+        self.dit_seq_shape: List[int]
+        self.full_window_size: List[int]
+        self.img_seq_length: int
+
+    def tile(self, x: torch.Tensor) -> torch.Tensor:
+        x = rearrange(x,
+                      "b (sp t h w) head d -> b (t sp h w) head d",
+                      sp=self.sp_size,
+                      t=self.dit_seq_shape[0] // self.sp_size,
+                      h=self.dit_seq_shape[1],
+                      w=self.dit_seq_shape[2])
+
+        return rearrange(
+            x,
+            "b (n_t ts_t n_h ts_h n_w ts_w) h d -> b (n_t n_h n_w ts_t ts_h ts_w) h d",
+            n_t=self.full_window_size[0],
+            n_h=self.full_window_size[1],
+            n_w=self.full_window_size[2],
+            ts_t=self.VSA_base_tile_size[0],
+            ts_h=self.VSA_base_tile_size[1],
+            ts_w=self.VSA_base_tile_size[2])
+
+    def untile(self, x: torch.Tensor) -> torch.Tensor:
+        x = rearrange(
+            x,
+            "b (n_t n_h n_w ts_t ts_h ts_w) h d -> b (n_t ts_t n_h ts_h n_w ts_w) h d",
+            n_t=self.full_window_size[0],
+            n_h=self.full_window_size[1],
+            n_w=self.full_window_size[2],
+            ts_t=self.VSA_base_tile_size[0],
+            ts_h=self.VSA_base_tile_size[1],
+            ts_w=self.VSA_base_tile_size[2])
+        return rearrange(x,
+                         "b (t sp h w) head d -> b (sp t h w) head d",
+                         sp=self.sp_size,
+                         t=self.dit_seq_shape[0] // self.sp_size,
+                         h=self.dit_seq_shape[1],
+                         w=self.dit_seq_shape[2])
+
+    def preprocess_qkv(
+        self,
+        qkv: torch.Tensor,
+        attn_metadata: VideoSparseAttentionMetadata,
+    ) -> torch.Tensor:
+        self.dit_seq_shape = attn_metadata.dit_seq_shape
+        self.full_window_size = [
+            self.dit_seq_shape[0] // self.VSA_base_tile_size[0],
+            self.dit_seq_shape[1] // self.VSA_base_tile_size[1],
+            self.dit_seq_shape[2] // self.VSA_base_tile_size[2]
+        ]
+        self.img_seq_length = math.prod(self.dit_seq_shape)
+        return self.tile(qkv)
+
+    def postprocess_output(
+        self,
+        output: torch.Tensor,
+        attn_metadata: VideoSparseAttentionMetadata,
+    ) -> torch.Tensor:
+        return self.untile(output)
+
+    def forward(  # type: ignore[override]
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        gate_compress: torch.Tensor,
+        attn_metadata: VideoSparseAttentionMetadata,
+    ) -> torch.Tensor:
+        query = query.transpose(1, 2).contiguous()
+        key = key.transpose(1, 2).contiguous()
+        value = value.transpose(1, 2).contiguous()
+        gate_compress = gate_compress.transpose(1, 2).contiguous()
+
+        VSA_sparsity = attn_metadata.VSA_sparsity
+
+        cur_topk = math.ceil(
+            (1 - VSA_sparsity) *
+            (self.img_seq_length / math.prod(self.VSA_base_tile_size)))
+
+        if video_sparse_attn is None:
+            raise NotImplementedError("video_sparse_attn is not installed")
+
+        hidden_states = video_sparse_attn(
+            query,
+            key,
+            value,
+            topk=cur_topk,
+            block_size=(4, 4, 4),
+            compress_attn_weight=gate_compress).transpose(1, 2)
+
+        return hidden_states
@@ -9,10 +9,11 @@ from fastvideo.v1.attention.selector import (backend_name_to_enum,
                                             get_attn_backend)
 from fastvideo.v1.distributed.communication_op import (
    sequence_model_parallel_all_gather, sequence_model_parallel_all_to_all_4D)
-from fastvideo.v1.distributed.parallel_state import (
-    get_sequence_model_parallel_rank, get_sequence_model_parallel_world_size)
+from fastvideo.v1.distributed.parallel_state import (get_sp_parallel_rank,
+                                                     get_sp_world_size)
 from fastvideo.v1.forward_context import ForwardContext, get_forward_context
-from fastvideo.v1.platforms import _Backend
+from fastvideo.v1.platforms import AttentionBackendEnum
+from fastvideo.v1.utils import get_compute_dtype


 class DistributedAttention(nn.Module):
@@ -25,8 +26,8 @@ class DistributedAttention(nn.Module):
                 num_kv_heads: Optional[int] = None,
                 softmax_scale: Optional[float] = None,
                 causal: bool = False,
-                 supported_attention_backends: Optional[Tuple[_Backend,
-                                                              ...]] = None,
+                 supported_attention_backends: Optional[Tuple[
+                     AttentionBackendEnum, ...]] = None,
                 prefix: str = "",
                 **extra_impl_args) -> None:
        super().__init__()
@@ -38,19 +39,19 @@ class DistributedAttention(nn.Module):
        if num_kv_heads is None:
            num_kv_heads = num_heads

-        dtype = torch.get_default_dtype()
+        dtype = get_compute_dtype()
        attn_backend = get_attn_backend(
            head_size,
            dtype,
            supported_attention_backends=supported_attention_backends)
        impl_cls = attn_backend.get_impl_cls()
-        self.impl = impl_cls(num_heads=num_heads,
-                             head_size=head_size,
-                             causal=causal,
-                             softmax_scale=self.softmax_scale,
-                             num_kv_heads=num_kv_heads,
-                             prefix=f"{prefix}.impl",
-                             **extra_impl_args)
+        self.attn_impl = impl_cls(num_heads=num_heads,
+                                  head_size=head_size,
+                                  causal=causal,
+                                  softmax_scale=self.softmax_scale,
+                                  num_kv_heads=num_kv_heads,
+                                  prefix=f"{prefix}.impl",
+                                  **extra_impl_args)
        self.num_heads = num_heads
        self.head_size = head_size
        self.num_kv_heads = num_kv_heads
@@ -84,12 +85,9 @@ class DistributedAttention(nn.Module):
        # Check input shapes
        assert q.dim() == 4 and k.dim() == 4 and v.dim(
        ) == 4, "Expected 4D tensors"
-        # assert bs = 1
-        assert q.shape[
-            0] == 1, "Batch size must be 1, and there should be no padding tokens"
        batch_size, seq_len, num_heads, head_dim = q.shape
-        local_rank = get_sequence_model_parallel_rank()
-        world_size = get_sequence_model_parallel_world_size()
+        local_rank = get_sp_parallel_rank()
+        world_size = get_sp_world_size()

        forward_context: ForwardContext = get_forward_context()
        ctx_attn_metadata = forward_context.attn_metadata
@@ -102,7 +100,7 @@ class DistributedAttention(nn.Module):
                                                    scatter_dim=2,
                                                    gather_dim=1)
        # Apply backend-specific preprocess_qkv
-        qkv = self.impl.preprocess_qkv(qkv, ctx_attn_metadata)
+        qkv = self.attn_impl.preprocess_qkv(qkv, ctx_attn_metadata)

        # Concatenate with replicated QKV if provided
        if replicated_q is not None:
@@ -118,7 +116,7 @@ class DistributedAttention(nn.Module):

        q, k, v = qkv.chunk(3, dim=0)

-        output = self.impl.forward(q, k, v, ctx_attn_metadata)
+        output = self.attn_impl.forward(q, k, v, ctx_attn_metadata)

        # Redistribute back if using sequence parallelism
        replicated_output = None
@@ -129,7 +127,73 @@ class DistributedAttention(nn.Module):
            replicated_output = sequence_model_parallel_all_gather(
                replicated_output.contiguous(), dim=2)
        # Apply backend-specific postprocess_output
-        output = self.impl.postprocess_output(output, ctx_attn_metadata)
+        output = self.attn_impl.postprocess_output(output, ctx_attn_metadata)
+
+        output = sequence_model_parallel_all_to_all_4D(output,
+                                                       scatter_dim=1,
+                                                       gather_dim=2)
+        return output, replicated_output
+
+
+class DistributedAttention_VSA(DistributedAttention):
+    """Distributed attention layer with VSA support.
+    """
+
+    def forward(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        replicated_q: Optional[torch.Tensor] = None,
+        replicated_k: Optional[torch.Tensor] = None,
+        replicated_v: Optional[torch.Tensor] = None,
+        gate_compress: Optional[torch.Tensor] = None,
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Forward pass for distributed attention.
+        
+        Args:
+            q (torch.Tensor): Query tensor [batch_size, seq_len, num_heads, head_dim]
+            k (torch.Tensor): Key tensor [batch_size, seq_len, num_heads, head_dim]
+            v (torch.Tensor): Value tensor [batch_size, seq_len, num_heads, head_dim]
+            gate_compress (torch.Tensor): Gate compress tensor [batch_size, seq_len, num_heads, head_dim]
+            replicated_q (Optional[torch.Tensor]): Replicated query tensor, typically for text tokens
+            replicated_k (Optional[torch.Tensor]): Replicated key tensor
+            replicated_v (Optional[torch.Tensor]): Replicated value tensor
+            
+        Returns:
+            Tuple[torch.Tensor, Optional[torch.Tensor]]: A tuple containing:
+                - o (torch.Tensor): Output tensor after attention for the main sequence
+                - replicated_o (Optional[torch.Tensor]): Output tensor for replicated tokens, if provided
+        """
+        # Check text tokens are not supported for VSA now
+        assert replicated_q is None and replicated_k is None and replicated_v is None, "Replicated QKV is not supported for VSA now"
+        # Check input shapes
+        assert q.dim() == 4 and k.dim() == 4 and v.dim(
+        ) == 4, "Expected 4D tensors"
+
+        forward_context: ForwardContext = get_forward_context()
+        ctx_attn_metadata = forward_context.attn_metadata
+
+        # Stack QKV
+        qkvg = torch.cat([q, k, v, gate_compress],
+                         dim=0)  # [3, seq_len, num_heads, head_dim]
+
+        # Redistribute heads across sequence dimension
+        qkvg = sequence_model_parallel_all_to_all_4D(qkvg,
+                                                     scatter_dim=2,
+                                                     gather_dim=1)
+
+        qkvg = self.attn_impl.preprocess_qkv(qkvg, ctx_attn_metadata)
+
+        q, k, v, gate_compress = qkvg.chunk(4, dim=0)
+        output = self.attn_impl.forward(
+            q, k, v, gate_compress, ctx_attn_metadata)  # type: ignore[call-arg]
+
+        # Redistribute back if using sequence parallelism
+        replicated_output = None
+
+        # Apply backend-specific postprocess_output
+        output = self.attn_impl.postprocess_output(output, ctx_attn_metadata)

        output = sequence_model_parallel_all_to_all_4D(output,
                                                       scatter_dim=1,
@@ -147,8 +211,8 @@ class LocalAttention(nn.Module):
                 num_kv_heads: Optional[int] = None,
                 softmax_scale: Optional[float] = None,
                 causal: bool = False,
-                 supported_attention_backends: Optional[Tuple[_Backend,
-                                                              ...]] = None,
+                 supported_attention_backends: Optional[Tuple[
+                     AttentionBackendEnum, ...]] = None,
                 **extra_impl_args) -> None:
        super().__init__()
        if softmax_scale is None:
@@ -158,18 +222,18 @@ class LocalAttention(nn.Module):
        if num_kv_heads is None:
            num_kv_heads = num_heads

-        dtype = torch.get_default_dtype()
+        dtype = get_compute_dtype()
        attn_backend = get_attn_backend(
            head_size,
            dtype,
            supported_attention_backends=supported_attention_backends)
        impl_cls = attn_backend.get_impl_cls()
-        self.impl = impl_cls(num_heads=num_heads,
-                             head_size=head_size,
-                             softmax_scale=self.softmax_scale,
-                             num_kv_heads=num_kv_heads,
-                             causal=causal,
-                             **extra_impl_args)
+        self.attn_impl = impl_cls(num_heads=num_heads,
+                                  head_size=head_size,
+                                  softmax_scale=self.softmax_scale,
+                                  num_kv_heads=num_kv_heads,
+                                  causal=causal,
+                                  **extra_impl_args)
        self.num_heads = num_heads
        self.head_size = head_size
        self.num_kv_heads = num_kv_heads
@@ -200,5 +264,5 @@ class LocalAttention(nn.Module):
        forward_context: ForwardContext = get_forward_context()
        ctx_attn_metadata = forward_context.attn_metadata

-        output = self.impl.forward(q, k, v, ctx_attn_metadata)
+        output = self.attn_impl.forward(q, k, v, ctx_attn_metadata)
        return output
@@ -11,13 +11,13 @@ import torch
 import fastvideo.v1.envs as envs
 from fastvideo.v1.attention.backends.abstract import AttentionBackend
 from fastvideo.v1.logger import init_logger
-from fastvideo.v1.platforms import _Backend, current_platform
+from fastvideo.v1.platforms import AttentionBackendEnum, current_platform
 from fastvideo.v1.utils import STR_BACKEND_ENV_VAR, resolve_obj_by_qualname

 logger = init_logger(__name__)


-def backend_name_to_enum(backend_name: str) -> Optional[_Backend]:
+def backend_name_to_enum(backend_name: str) -> Optional[AttentionBackendEnum]:
    """
    Convert a string backend name to a _Backend enum value.

@@ -27,11 +27,11 @@ def backend_name_to_enum(backend_name: str) -> Optional[_Backend]:
            loaded.
    """
    assert backend_name is not None
-    return _Backend[backend_name] if backend_name in _Backend.__members__ else \
+    return AttentionBackendEnum[backend_name] if backend_name in AttentionBackendEnum.__members__ else \
          None


-def get_env_variable_attn_backend() -> Optional[_Backend]:
+def get_env_variable_attn_backend() -> Optional[AttentionBackendEnum]:
    '''
    Get the backend override specified by the FastVideo attention
    backend environment variable, if one is specified.
@@ -53,10 +53,11 @@ def get_env_variable_attn_backend() -> Optional[_Backend]:
 #
 # THIS SELECTION TAKES PRECEDENCE OVER THE
 # FASTVIDEO ATTENTION BACKEND ENVIRONMENT VARIABLE
-forced_attn_backend: Optional[_Backend] = None
+forced_attn_backend: Optional[AttentionBackendEnum] = None


-def global_force_attn_backend(attn_backend: Optional[_Backend]) -> None:
+def global_force_attn_backend(
+        attn_backend: Optional[AttentionBackendEnum]) -> None:
    '''
    Force all attention operations to use a specified backend.

@@ -71,7 +72,7 @@ def global_force_attn_backend(attn_backend: Optional[_Backend]) -> None:
    forced_attn_backend = attn_backend


-def get_global_forced_attn_backend() -> Optional[_Backend]:
+def get_global_forced_attn_backend() -> Optional[AttentionBackendEnum]:
    '''
    Get the currently-forced choice of attention backend,
    or None if auto-selection is currently enabled.
@@ -82,7 +83,8 @@ def get_global_forced_attn_backend() -> Optional[_Backend]:
 def get_attn_backend(
    head_size: int,
    dtype: torch.dtype,
-    supported_attention_backends: Optional[Tuple[_Backend, ...]] = None,
+    supported_attention_backends: Optional[Tuple[AttentionBackendEnum,
+                                                 ...]] = None,
 ) -> Type[AttentionBackend]:
    return _cached_get_attn_backend(head_size, dtype,
                                    supported_attention_backends)
@@ -92,7 +94,8 @@ def get_attn_backend(
 def _cached_get_attn_backend(
    head_size: int,
    dtype: torch.dtype,
-    supported_attention_backends: Optional[Tuple[_Backend, ...]] = None,
+    supported_attention_backends: Optional[Tuple[AttentionBackendEnum,
+                                                 ...]] = None,
 ) -> Type[AttentionBackend]:
    # Check whether a particular choice of backend was
    # previously forced.
@@ -102,7 +105,7 @@ def _cached_get_attn_backend(
    if not supported_attention_backends:
        raise ValueError("supported_attention_backends is empty")
    selected_backend = None
-    backend_by_global_setting: Optional[_Backend] = (
+    backend_by_global_setting: Optional[AttentionBackendEnum] = (
        get_global_forced_attn_backend())
    if backend_by_global_setting is not None:
        selected_backend = backend_by_global_setting
@@ -125,7 +128,7 @@ def _cached_get_attn_backend(

@contextmanager
 def global_force_attn_backend_context_manager(
-        attn_backend: _Backend) -> Generator[None, None, None]:
+        attn_backend: AttentionBackendEnum) -> Generator[None, None, None]:
    '''
    Globally force a FastVideo attention backend override within a
    context manager, reverting the global attention backend
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field, fields
 from typing import Any, Dict

@@ -1,9 +1,10 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
-from typing import Any, Optional, Tuple
+from typing import Any, List, Optional, Tuple

 from fastvideo.v1.configs.models.base import ArchConfig, ModelConfig
-from fastvideo.v1.configs.quantization import QuantizationConfig
-from fastvideo.v1.platforms import _Backend
+from fastvideo.v1.layers.quantization import QuantizationConfig
+from fastvideo.v1.platforms import AttentionBackendEnum


@dataclass
@@ -11,15 +12,17 @@ class DiTArchConfig(ArchConfig):
    _fsdp_shard_conditions: list = field(default_factory=list)
    _compile_conditions: list = field(default_factory=list)
    _param_names_mapping: dict = field(default_factory=dict)
-    _supported_attention_backends: Tuple[_Backend,
-                                         ...] = (_Backend.SLIDING_TILE_ATTN,
-                                                 _Backend.SAGE_ATTN,
-                                                 _Backend.FLASH_ATTN,
-                                                 _Backend.TORCH_SDPA)
+    _reverse_param_names_mapping: dict = field(default_factory=dict)
+    _lora_param_names_mapping: dict = field(default_factory=dict)
+    _supported_attention_backends: Tuple[AttentionBackendEnum, ...] = (
+        AttentionBackendEnum.SLIDING_TILE_ATTN, AttentionBackendEnum.SAGE_ATTN,
+        AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA,
+        AttentionBackendEnum.VIDEO_SPARSE_ATTN)

    hidden_size: int = 0
    num_attention_heads: int = 0
    num_channels_latents: int = 0
+    exclude_lora_layers: List[str] = field(default_factory=list)

    def __post_init__(self) -> None:
        if not self._compile_conditions:
@@ -1,5 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
-from typing import Optional, Tuple
+from typing import List, Optional, Tuple

 import torch

@@ -146,6 +147,9 @@ class HunyuanVideoArchConfig(DiTArchConfig):
            r"final_layer.linear.\1",
        })

+    # Reverse mapping for saving checkpoints: training -> diffusers
+    _reverse_param_names_mapping: dict = field(default_factory=lambda: {})
+
    patch_size: int = 2
    patch_size_t: int = 1
    in_channels: int = 16
@@ -163,6 +167,8 @@ class HunyuanVideoArchConfig(DiTArchConfig):
    pooled_projection_dim: int = 768
    rope_theta: int = 256
    qk_norm: str = "rms_norm"
+    exclude_lora_layers: List[str] = field(
+        default_factory=lambda: ["img_in", "txt_in", "time_in", "vector_in"])

    def __post_init__(self):
        super().__post_init__()
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import List, Optional, Tuple, Union

@@ -51,6 +52,7 @@ class StepVideoArchConfig(DiTArchConfig):
        default_factory=lambda: [6144, 1024])
    attention_type: Optional[str] = "torch"
    use_additional_conditions: Optional[bool] = False
+    exclude_lora_layers: List[str] = field(default_factory=lambda: [])

    def __post_init__(self):
        self.hidden_size = self.num_attention_heads * self.attention_head_dim
@@ -1,5 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
-from typing import Optional, Tuple
+from typing import List, Optional, Tuple

 from fastvideo.v1.configs.models.dits.base import DiTArchConfig, DiTConfig

@@ -48,10 +49,31 @@ class WanVideoArchConfig(DiTArchConfig):
            r"blocks.\1.ffn.fc_in.\2",
            r"^blocks\.(\d+)\.ffn\.net\.2\.(.*)$":
            r"blocks.\1.ffn.fc_out.\2",
-            r"blocks\.(\d+)\.norm2\.(.*)$":
+            r"^blocks\.(\d+)\.norm2\.(.*)$":
            r"blocks.\1.self_attn_residual_norm.norm.\2",
        })

+    # Reverse mapping for saving checkpoints: training -> diffusers
+    _reverse_param_names_mapping: dict = field(default_factory=lambda: {})
+
+    # Some LoRA adapters use the original official layer names instead of hf layer names,
+    # so apply this before the param_names_mapping
+    _lora_param_names_mapping: dict = field(
+        default_factory=lambda: {
+            r"^blocks\.(\d+)\.self_attn\.q\.(.*)$": r"blocks.\1.attn1.to_q.\2",
+            r"^blocks\.(\d+)\.self_attn\.k\.(.*)$": r"blocks.\1.attn1.to_k.\2",
+            r"^blocks\.(\d+)\.self_attn\.v\.(.*)$": r"blocks.\1.attn1.to_v.\2",
+            r"^blocks\.(\d+)\.self_attn\.o\.(.*)$":
+            r"blocks.\1.attn1.to_out.0.\2",
+            r"^blocks\.(\d+)\.cross_attn\.q\.(.*)$": r"blocks.\1.attn2.to_q.\2",
+            r"^blocks\.(\d+)\.cross_attn\.k\.(.*)$": r"blocks.\1.attn2.to_k.\2",
+            r"^blocks\.(\d+)\.cross_attn\.v\.(.*)$": r"blocks.\1.attn2.to_v.\2",
+            r"^blocks\.(\d+)\.cross_attn\.o\.(.*)$":
+            r"blocks.\1.attn2.to_out.0.\2",
+            r"^blocks\.(\d+)\.ffn\.0\.(.*)$": r"blocks.\1.ffn.fc_in.\2",
+            r"^blocks\.(\d+)\.ffn\.2\.(.*)$": r"blocks.\1.ffn.fc_out.\2",
+        })
+
    patch_size: Tuple[int, int, int] = (1, 2, 2)
    text_len = 512
    num_attention_heads: int = 40
@@ -68,6 +90,7 @@ class WanVideoArchConfig(DiTArchConfig):
    image_dim: Optional[int] = None
    added_kv_proj_dim: Optional[int] = None
    rope_max_seq_len: int = 1024
+    exclude_lora_layers: List[str] = field(default_factory=lambda: ["embedder"])

    def __post_init__(self):
        super().__post_init__()
@@ -1,18 +1,19 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import Any, Dict, List, Optional, Tuple

 import torch

 from fastvideo.v1.configs.models.base import ArchConfig, ModelConfig
-from fastvideo.v1.configs.quantization import QuantizationConfig
-from fastvideo.v1.platforms import _Backend
+from fastvideo.v1.layers.quantization import QuantizationConfig
+from fastvideo.v1.platforms import AttentionBackendEnum


@dataclass
 class EncoderArchConfig(ArchConfig):
    architectures: List[str] = field(default_factory=lambda: [])
-    _supported_attention_backends: Tuple[_Backend, ...] = (_Backend.FLASH_ATTN,
-                                                           _Backend.TORCH_SDPA)
+    _supported_attention_backends: Tuple[AttentionBackendEnum, ...] = (
+        AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA)
    output_hidden_states: bool = False
    use_return_dict: bool = True

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import Optional

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import Optional

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import Optional

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+import argparse
+import dataclasses
 from dataclasses import dataclass, field
 from typing import Any, Union

@@ -128,3 +131,12 @@ class VAEConfig(ModelConfig):
        )

        return parser
+
+    @classmethod
+    def from_cli_args(cls, args: argparse.Namespace) -> "VAEConfig":
+        kwargs = {}
+        for attr in dataclasses.fields(cls):
+            value = getattr(args, attr.name, None)
+            if value is not None:
+                kwargs[attr.name] = value
+        return cls(**kwargs)
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import Tuple

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field

 from fastvideo.v1.configs.models.vaes.base import VAEArchConfig, VAEConfig
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import Tuple

@@ -63,7 +64,7 @@ class WanVAEArchConfig(VAEArchConfig):

@dataclass
 class WanVAEConfig(VAEConfig):
-    arch_config: VAEArchConfig = field(default_factory=WanVAEArchConfig)
+    arch_config: WanVAEArchConfig = field(default_factory=WanVAEArchConfig)
    use_feature_cache: bool = True

    use_tiling: bool = False
@@ -3,7 +3,7 @@ from fastvideo.v1.configs.pipelines.base import (PipelineConfig,
 from fastvideo.v1.configs.pipelines.hunyuan import (FastHunyuanConfig,
                                                    HunyuanConfig)
 from fastvideo.v1.configs.pipelines.registry import (
-    get_pipeline_config_cls_for_name)
+    get_pipeline_config_cls_from_name)
 from fastvideo.v1.configs.pipelines.stepvideo import StepVideoT2VConfig
 from fastvideo.v1.configs.pipelines.wan import (WanI2V480PConfig,
                                                WanI2V720PConfig,
@@ -14,5 +14,5 @@ __all__ = [
    "HunyuanConfig", "FastHunyuanConfig", "PipelineConfig",
    "SlidingTileAttnConfig", "WanT2V480PConfig", "WanI2V480PConfig",
    "WanT2V720PConfig", "WanI2V720PConfig", "StepVideoT2VConfig",
-    "get_pipeline_config_cls_for_name"
+    "get_pipeline_config_cls_from_name"
 ]
@@ -1,18 +1,31 @@
+# SPDX-License-Identifier: Apache-2.0
 import json
 from dataclasses import asdict, dataclass, field, fields
-from typing import Any, Callable, Dict, Optional, Tuple, cast
+from enum import Enum
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union, cast

 import torch

 from fastvideo.v1.configs.models import (DiTConfig, EncoderConfig, ModelConfig,
                                         VAEConfig)
 from fastvideo.v1.configs.models.encoders import BaseEncoderOutput
+from fastvideo.v1.configs.utils import update_config_from_args
 from fastvideo.v1.logger import init_logger
-from fastvideo.v1.utils import shallow_asdict
+from fastvideo.v1.utils import (FlexibleArgumentParser, StoreBoolean,
+                                shallow_asdict)

 logger = init_logger(__name__)


+class STA_Mode(str, Enum):
+    """STA (Sliding Tile Attention) modes."""
+    STA_INFERENCE = "STA_inference"
+    STA_SEARCHING = "STA_searching"
+    STA_TUNING = "STA_tuning"
+    STA_TUNING_CFG = "STA_tuning_cfg"
+    NONE = None
+
+
 def preprocess_text(prompt: str) -> str:
    return prompt

@@ -21,58 +34,282 @@ def postprocess_text(output: BaseEncoderOutput) -> torch.tensor:
    raise NotImplementedError


+# config for a single pipeline
@dataclass
 class PipelineConfig:
    """Base configuration for all pipeline architectures."""
+    model_path: str = ""
+    pipeline_config_path: Optional[str] = None
+
    # Video generation parameters
    embedded_cfg_scale: float = 6.0
    flow_shift: Optional[float] = None
-    use_cpu_offload: bool = False
    disable_autocast: bool = False

    # Model configuration
-    precision: str = "bf16"
+    dit_config: DiTConfig = field(default_factory=DiTConfig)
+    dit_precision: str = "bf16"

    # VAE configuration
+    vae_config: VAEConfig = field(default_factory=VAEConfig)
    vae_precision: str = "fp16"
    vae_tiling: bool = True
    vae_sp: bool = True
-    vae_config: VAEConfig = field(default_factory=VAEConfig)

-    # DiT configuration
-    dit_config: DiTConfig = field(default_factory=DiTConfig)
+    # Image encoder configuration
+    image_encoder_config: EncoderConfig = field(default_factory=EncoderConfig)
+    image_encoder_precision: str = "fp32"

    # Text encoder configuration
-    text_encoder_precisions: Tuple[str, ...] = field(
-        default_factory=lambda: ("fp16", ))
+    DEFAULT_TEXT_ENCODER_PRECISIONS = ("fp16", )
    text_encoder_configs: Tuple[EncoderConfig, ...] = field(
        default_factory=lambda: (EncoderConfig(), ))
+    text_encoder_precisions: Tuple[str, ...] = field(
+        default_factory=lambda: ("fp16", ))
    preprocess_text_funcs: Tuple[Callable[[str], str], ...] = field(
        default_factory=lambda: (preprocess_text, ))
    postprocess_text_funcs: Tuple[Callable[[BaseEncoderOutput], torch.tensor],
                                  ...] = field(default_factory=lambda:
                                               (postprocess_text, ))

-    # STA (Spatial-Temporal Attention) parameters
+    # LoRA parameters
+    lora_path: Optional[str] = None
+    lora_nickname: Optional[
+        str] = "default"  # for swapping adapters in the pipeline
+    lora_target_names: Optional[List[
+        str]] = None  # can restrict list of layers to adapt, e.g. ["q_proj"]
+
+    # StepVideo specific parameters
+    pos_magic: Optional[str] = None
+    neg_magic: Optional[str] = None
+    timesteps_scale: Optional[bool] = None
+
+    # STA (Sliding Tile Attention) parameters
    mask_strategy_file_path: Optional[str] = None
+    STA_mode: STA_Mode = STA_Mode.STA_INFERENCE
+    skip_time_steps: int = 15

    # Compilation
-    enable_torch_compile: bool = False
+    # enable_torch_compile: bool = False
+
+    @staticmethod
+    def add_cli_args(parser: FlexibleArgumentParser,
+                     prefix: str = "") -> FlexibleArgumentParser:
+        prefix_with_dot = f"{prefix}." if (prefix.strip() != "") else ""
+
+        # model_path will be conflicting with the model_path in FastVideoArgs,
+        # so we add it separately if prefix is not empty
+        if prefix_with_dot != "":
+            parser.add_argument(
+                f"--{prefix_with_dot}model-path",
+                type=str,
+                dest=f"{prefix_with_dot.replace('-', '_')}model_path",
+                default=PipelineConfig.model_path,
+                help="Path to the pretrained model",
+            )
+
+        parser.add_argument(
+            f"--{prefix_with_dot}pipeline-config-path",
+            type=str,
+            dest=f"{prefix_with_dot.replace('-', '_')}pipeline_config_path",
+            default=PipelineConfig.pipeline_config_path,
+            help="Path to the pipeline config",
+        )
+        parser.add_argument(
+            f"--{prefix_with_dot}embedded-cfg-scale",
+            type=float,
+            dest=f"{prefix_with_dot.replace('-', '_')}embedded_cfg_scale",
+            default=PipelineConfig.embedded_cfg_scale,
+            help="Embedded CFG scale",
+        )
+        parser.add_argument(
+            f"--{prefix_with_dot}flow-shift",
+            type=float,
+            dest=f"{prefix_with_dot.replace('-', '_')}flow_shift",
+            default=PipelineConfig.flow_shift,
+            help="Flow shift parameter",
+        )
+
+        # DiT configuration
+        parser.add_argument(
+            f"--{prefix_with_dot}dit-precision",
+            type=str,
+            dest=f"{prefix_with_dot.replace('-', '_')}dit_precision",
+            default=PipelineConfig.dit_precision,
+            choices=["fp32", "fp16", "bf16"],
+            help="Precision for the DiT model",
+        )
+
+        # VAE configuration
+        parser.add_argument(
+            f"--{prefix_with_dot}vae-precision",
+            type=str,
+            dest=f"{prefix_with_dot.replace('-', '_')}vae_precision",
+            default=PipelineConfig.vae_precision,
+            choices=["fp32", "fp16", "bf16"],
+            help="Precision for VAE",
+        )
+        parser.add_argument(
+            f"--{prefix_with_dot}vae-tiling",
+            action=StoreBoolean,
+            dest=f"{prefix_with_dot.replace('-', '_')}vae_tiling",
+            default=PipelineConfig.vae_tiling,
+            help="Enable VAE tiling",
+        )
+        parser.add_argument(
+            f"--{prefix_with_dot}vae-sp",
+            action=StoreBoolean,
+            dest=f"{prefix_with_dot.replace('-', '_')}vae_sp",
+            help="Enable VAE spatial parallelism",
+        )
+
+        # Text encoder configuration
+        parser.add_argument(
+            f"--{prefix_with_dot}text-encoder-precisions",
+            nargs="+",
+            type=str,
+            dest=f"{prefix_with_dot.replace('-', '_')}text_encoder_precisions",
+            default=PipelineConfig.DEFAULT_TEXT_ENCODER_PRECISIONS,
+            choices=["fp32", "fp16", "bf16"],
+            help="Precision for each text encoder",
+        )
+
+        # Image encoder configuration
+        parser.add_argument(
+            f"--{prefix_with_dot}image-encoder-precision",
+            type=str,
+            dest=f"{prefix_with_dot.replace('-', '_')}image_encoder_precision",
+            default=PipelineConfig.image_encoder_precision,
+            choices=["fp32", "fp16", "bf16"],
+            help="Precision for image encoder",
+        )
+        parser.add_argument(
+            f"--{prefix_with_dot}pos_magic",
+            type=str,
+            dest=f"{prefix_with_dot.replace('-', '_')}pos_magic",
+            default=PipelineConfig.pos_magic,
+            help="Positive magic prompt for sampling, used in stepvideo",
+        )
+        parser.add_argument(
+            f"--{prefix_with_dot}neg_magic",
+            type=str,
+            dest=f"{prefix_with_dot.replace('-', '_')}neg_magic",
+            default=PipelineConfig.neg_magic,
+            help="Negative magic prompt for sampling, used in stepvideo",
+        )
+        parser.add_argument(
+            f"--{prefix_with_dot}timesteps_scale",
+            type=bool,
+            dest=f"{prefix_with_dot.replace('-', '_')}timesteps_scale",
+            default=PipelineConfig.timesteps_scale,
+            help=
+            "Bool for applying scheduler scale in set_timesteps, used in stepvideo",
+        )
+
+        # Add VAE configuration arguments
+        from fastvideo.v1.configs.models.vaes.base import VAEConfig
+        VAEConfig.add_cli_args(parser, prefix=f"{prefix_with_dot}vae-config")
+
+        # Add DiT configuration arguments
+        from fastvideo.v1.configs.models.dits.base import DiTConfig
+        DiTConfig.add_cli_args(parser, prefix=f"{prefix_with_dot}dit-config")
+
+        return parser
+
+    def update_config_from_dict(self,
+                                args: Dict[str, Any],
+                                prefix: str = "") -> None:
+        prefix_with_dot = f"{prefix}." if (prefix.strip() != "") else ""
+        update_config_from_args(self, args, prefix, pop_args=True)
+        update_config_from_args(self.vae_config,
+                                args,
+                                f"{prefix_with_dot}vae_config",
+                                pop_args=True)
+        update_config_from_args(self.dit_config,
+                                args,
+                                f"{prefix_with_dot}dit_config",
+                                pop_args=True)

    @classmethod
    def from_pretrained(cls, model_path: str) -> "PipelineConfig":
+        """
+        use the pipeline class setting from model_path to match the pipeline config
+        """
        from fastvideo.v1.configs.pipelines.registry import (
-            get_pipeline_config_cls_for_name)
-        pipeline_config_cls = get_pipeline_config_cls_for_name(model_path)
-        if pipeline_config_cls is not None:
-            pipeline_config = pipeline_config_cls()
-        else:
+            get_pipeline_config_cls_from_name)
+        pipeline_config_cls = get_pipeline_config_cls_from_name(model_path)
+
+        return cast(PipelineConfig, pipeline_config_cls(model_path=model_path))
+
+    @classmethod
+    def from_kwargs(cls,
+                    kwargs: Dict[str, Any],
+                    config_cli_prefix: str = "") -> "PipelineConfig":
+        """
+        Load PipelineConfig from kwargs Dictionary.
+        kwargs: dictionary of kwargs
+        config_cli_prefix: prefix of CLI arguments for this PipelineConfig instance
+        """
+        from fastvideo.v1.configs.pipelines.registry import (
+            get_pipeline_config_cls_from_name)
+
+        prefix_with_dot = f"{config_cli_prefix}." if (config_cli_prefix.strip()
+                                                      != "") else ""
+        model_path: Optional[str] = kwargs.get(prefix_with_dot + 'model_path',
+                                               None) or kwargs.get('model_path')
+        pipeline_config_or_path: Optional[Union[str, PipelineConfig, Dict[
+            str, Any]]] = kwargs.get(prefix_with_dot + 'pipeline_config',
+                                     None) or kwargs.get('pipeline_config')
+        if model_path is None:
+            raise ValueError("model_path is required in kwargs")
+
+        # 1. Get the pipeline config class from the registry
+        pipeline_config_cls = get_pipeline_config_cls_from_name(model_path)
+
+        # 2. Instantiate PipelineConfig
+        if pipeline_config_cls is None:
            logger.warning(
-                "Couldn't find an optimal sampling param for %s. Using the default sampling param.",
+                "Couldn't find pipeline config for %s. Using the default pipeline config.",
                model_path)
            pipeline_config = cls()
+        else:
+            pipeline_config = pipeline_config_cls()

-        return cast(PipelineConfig, pipeline_config)
+        # 3. Load PipelineConfig from a json file or a PipelineConfig object if provided
+        if isinstance(pipeline_config_or_path, str):
+            pipeline_config.load_from_json(pipeline_config_or_path)
+            kwargs[prefix_with_dot +
+                   'pipeline_config_path'] = pipeline_config_or_path
+        elif isinstance(pipeline_config_or_path, PipelineConfig):
+            pipeline_config = pipeline_config_or_path
+        elif isinstance(pipeline_config_or_path, dict):
+            pipeline_config.update_pipeline_config(pipeline_config_or_path)
+
+        # 4. Update PipelineConfig from CLI arguments if provided
+        kwargs[prefix_with_dot + 'model_path'] = model_path
+        pipeline_config.update_config_from_dict(kwargs, config_cli_prefix)
+        return pipeline_config
+
+    def check_pipeline_config(self) -> None:
+        if self.vae_sp and not self.vae_tiling:
+            raise ValueError(
+                "Currently enabling vae_sp requires enabling vae_tiling, please set --vae-tiling to True."
+            )
+
+        if len(self.text_encoder_configs) != len(self.text_encoder_precisions):
+            raise ValueError(
+                f"Length of text encoder configs ({len(self.text_encoder_configs)}) must be equal to length of text encoder precisions ({len(self.text_encoder_precisions)})"
+            )
+
+        if len(self.text_encoder_configs) != len(self.preprocess_text_funcs):
+            raise ValueError(
+                f"Length of text encoder configs ({len(self.text_encoder_configs)}) must be equal to length of text preprocessing functions ({len(self.preprocess_text_funcs)})"
+            )
+
+        if len(self.preprocess_text_funcs) != len(self.postprocess_text_funcs):
+            raise ValueError(
+                f"Length of text postprocess functions ({len(self.postprocess_text_funcs)}) must be equal to length of text preprocessing functions ({len(self.preprocess_text_funcs)})"
+            )

    def dump_to_json(self, file_path: str):
        output_dict = shallow_asdict(self)
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import Callable, Tuple, TypedDict

@@ -68,9 +69,6 @@ class HunyuanConfig(PipelineConfig):
    embedded_cfg_scale: int = 6
    flow_shift: int = 7

-    # Video parameters
-    use_cpu_offload: bool = True
-
    # Text encoding stage
    text_encoder_configs: Tuple[EncoderConfig, ...] = field(
        default_factory=lambda: (LlamaConfig(), CLIPTextConfig()))
@@ -82,7 +80,7 @@ class HunyuanConfig(PipelineConfig):
                     (llama_postprocess_text, clip_postprocess_text))

    # Precision for each component
-    precision: str = "bf16"
+    dit_precision: str = "bf16"
    vae_precision: str = "fp16"
    text_encoder_precisions: Tuple[str, ...] = field(
        default_factory=lambda: ("fp16", "fp16"))
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Registry for pipeline weight-specific configurations."""

 import os
@@ -18,7 +19,7 @@ from fastvideo.v1.utils import (maybe_download_model_index,
 logger = init_logger(__name__)

 # Registry maps specific model weights to their config classes
-WEIGHT_CONFIG_REGISTRY: Dict[str, Type[PipelineConfig]] = {
+PIPE_NAME_TO_CONFIG: Dict[str, Type[PipelineConfig]] = {
    "FastVideo/FastHunyuan-diffusers": FastHunyuanConfig,
    "hunyuanvideo-community/HunyuanVideo": HunyuanConfig,
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers": WanT2V480PConfig,
@@ -50,37 +51,74 @@ PIPELINE_FALLBACK_CONFIG: Dict[str, Type[PipelineConfig]] = {
 }


-def get_pipeline_config_cls_for_name(
-        pipeline_name_or_path: str) -> Optional[type[PipelineConfig]]:
-    """Get the appropriate config class for specific pretrained weights."""
+def get_pipeline_config_cls_from_name(
+        pipeline_name_or_path: str) -> Type[PipelineConfig]:
+    """Get the appropriate configuration class for a given pipeline name or path.

-    if os.path.exists(pipeline_name_or_path):
-        config = verify_model_config_and_directory(pipeline_name_or_path)
-        logger.warning(
-            "FastVideo may not correctly identify the optimal config for this model, as the local directory may have been renamed."
-        )
-    else:
-        config = maybe_download_model_index(pipeline_name_or_path)
+    This function implements a multi-step lookup process to find the most suitable
+    configuration class for a given pipeline. It follows this order:
+    1. Exact match in the PIPE_NAME_TO_CONFIG
+    2. Partial match in the PIPE_NAME_TO_CONFIG
+    3. Fallback to class name in the model_index.json
+    4. else raise an error

-    pipeline_name = config["_class_name"]
+    Args:
+        pipeline_name_or_path (str): The name or path of the pipeline. This can be:
+            - A registered model ID (e.g., "FastVideo/FastHunyuan-diffusers")
+            - A local path to a model directory
+            - A model ID that will be downloaded
+
+    Returns:
+        Type[PipelineConfig]: The configuration class that best matches the pipeline.
+            This will be one of:
+            - A specific weight configuration class if an exact match is found
+            - A fallback configuration class based on the pipeline architecture
+            - The base PipelineConfig class if no matches are found
+
+    Note:
+        - For local paths, the function will verify the model configuration
+        - For remote models, it will attempt to download the model index
+        - Warning messages are logged when falling back to less specific configurations
+    """
+
+    pipeline_config_cls: Optional[Type[PipelineConfig]] = None

    # First try exact match for specific weights
-    if pipeline_name_or_path in WEIGHT_CONFIG_REGISTRY:
-        return WEIGHT_CONFIG_REGISTRY[pipeline_name_or_path]
+    if pipeline_name_or_path in PIPE_NAME_TO_CONFIG:
+        pipeline_config_cls = PIPE_NAME_TO_CONFIG[pipeline_name_or_path]

    # Try partial matches (for local paths that might include the weight ID)
-    for registered_id, config_class in WEIGHT_CONFIG_REGISTRY.items():
+    for registered_id, config_class in PIPE_NAME_TO_CONFIG.items():
        if registered_id in pipeline_name_or_path:
-            return config_class
-
-    # If no match, try to use the fallback config
-    fallback_config = None
-    # Try to determine pipeline architecture for fallback
-    for pipeline_type, detector in PIPELINE_DETECTOR.items():
-        if detector(pipeline_name.lower()):
-            fallback_config = PIPELINE_FALLBACK_CONFIG.get(pipeline_type)
+            pipeline_config_cls = config_class
            break

-    logger.warning("No match found for pipeline %s, using fallback config %s.",
-                   pipeline_name_or_path, fallback_config)
-    return fallback_config
+    # If no match, try to use the fallback config
+    if pipeline_config_cls is None:
+        if os.path.exists(pipeline_name_or_path):
+            config = verify_model_config_and_directory(pipeline_name_or_path)
+        else:
+            config = maybe_download_model_index(pipeline_name_or_path)
+        logger.warning(
+            "Trying to use the config from the model_index.json. FastVideo may not correctly identify the optimal config for this model in this situation."
+        )
+
+        pipeline_name = config["_class_name"]
+        # Try to determine pipeline architecture for fallback
+        for pipeline_type, detector in PIPELINE_DETECTOR.items():
+            if detector(pipeline_name.lower()):
+                pipeline_config_cls = PIPELINE_FALLBACK_CONFIG.get(
+                    pipeline_type)
+                break
+
+        if pipeline_config_cls is not None:
+            logger.warning(
+                "No match found for pipeline %s, using fallback config %s.",
+                pipeline_name_or_path, pipeline_config_cls)
+
+    if pipeline_config_cls is None:
+        raise ValueError(
+            f"No match found for pipeline {pipeline_name_or_path}, please check the pipeline name or path."
+        )
+
+    return pipeline_config_cls
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field

 from fastvideo.v1.configs.models import DiTConfig, VAEConfig
@@ -18,9 +19,6 @@ class StepVideoT2VConfig(PipelineConfig):
    vae_tiling: bool = False
    vae_sp: bool = False

-    # Video parameters
-    use_cpu_offload: bool = True
-
    # Denoising stage
    flow_shift: int = 13
    timesteps_scale: bool = False
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
 from typing import Callable, Tuple

@@ -37,9 +38,6 @@ class WanT2V480PConfig(PipelineConfig):
    vae_tiling: bool = False
    vae_sp: bool = False

-    # Video parameters
-    use_cpu_offload: bool = True
-
    # Denoising stage
    flow_shift: int = 3

@@ -1,3 +0,0 @@
-from fastvideo.v1.configs.quantization.base import QuantizationConfig
-
-__all__ = ["QuantizationConfig"]
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
SolitaryThinker	480868bef9	update	2025-06-24 15:30:55 -07:00
Wenxuan Tan	aab74c1271	[Kernel] Remove all syncs from STA & VSA kernels (#517 )	2025-06-23 13:13:09 -07:00
Yongqi Chen	f89d86944f	[Feature][Training]Add diffusers format checkpoint saving for inference (#542 )	2025-06-22 01:23:41 -04:00
William Lin	8741d204a5	[Training] Refactor and improve validation datasets (#539 )	2025-06-21 17:58:35 -07:00
Wenxuan Tan	cdc85f58a8	[chore] Bump torch to 2.7.1 to support Blackwell (#483 )	2025-06-20 22:10:56 -07:00
William Lin	0262d2f089	[misc] [training] Reorganize training pipeline (#533 )	2025-06-20 20:42:25 -07:00
William Lin	62c0343465	[bugfix] [VSA] Fix layernorm type for VSA Wan2.1 TransformerBlock (#534 )	2025-06-20 00:24:51 -07:00
William Lin	1e1a023fb0	[bugfix] Fix stage validator for multi text encoder models (#535 )	2025-06-19 22:49:16 -07:00
William Lin	1d2517ad8e	[misc] Remove gradient checking code (#532 )	2025-06-18 23:29:25 -07:00
William Lin	d41186cb4a	[Feat] Add Stage input and output verification (#523 )	2025-06-18 23:29:11 -07:00
Kevin Lin Edenzzzz Wenxuan Tan	78e0c7eec9	Specify cu128 Pytorch installation (#530 ) Co-authored-by: Edenzzzz <wtan45@wisc.edu> Co-authored-by: Wenxuan Tan <wenxuan.tan@wisc.edu>	2025-06-18 20:02:50 -05:00
Wenxuan Tan	1c41a94b62	[Refactor] Move dict_to_3d_list under utils (#507 )	2025-06-18 13:34:37 -07:00
Yongqi Chen	2e66aafe20	[Bugfix][Readme]Fix readme website bugs and add VSA finetune docs (#531 )	2025-06-17 22:48:29 -07:00
Yongqi ChenandWill Lin	55074bda76	[CI] Add STA-inference/VSA-training test (#527 ) Co-authored-by: Will Lin <wlsaidhi@gmail.com>	2025-06-17 21:13:06 -07:00
William Lin	de65bec2b7	[Ci] add sta and vsa install to docker image (#528 )	2025-06-17 18:09:48 -07:00
Yongqi Chen	7664dd0de3	[Bugfix][Inference]Fix envs.attn_backend (#525 )	2025-06-17 18:38:06 -05:00
William Linandkevin314	019a88ced4	[CI][bugfix] Use new 3.12 docker image (#526 ) Co-authored-by: kevin314 <kevin.lin.cs1@gmail.com>	2025-06-17 15:37:08 -07:00
Kevin Lin	72de11abcc	[CI] Add current PR test workflow to Buildkite/Modal (#512 )	2025-06-17 13:29:22 -07:00
Kevin Lin	d71a4ebffc	[CI] Update Docker image to flash-attn 2.8.0 / CUDA 12.8 (#524 )	2025-06-16 17:48:23 -07:00
William Lin	1089ab43bf	[bugfix] [Training] use diffusers fp32layernorm for wan2.1 (#490 )	2025-06-15 22:45:48 -07:00
William Lin	97d4b984c9	[misc] [ci] fix e2e preprocess+training data path (#521 )	2025-06-14 22:37:51 -07:00
Wenxuan Tan	2a8953d74d	[Refactor] Fix attn backend selection not correctly setting env variable (#516 )	2025-06-15 00:04:54 -05:00
Yongqi Chen	8801b10da7	[Bugfix][Preprocess]fix mini dataset name (#520 )	2025-06-14 22:03:22 -07:00
William Lin	6b413f2ec4	[CI] [Training] drop negative prompt in validation dataset and CI test for preprocess + training overfit (#519 )	2025-06-14 18:50:17 -07:00
Yongqi Chen	28b72694aa	[Feature][Preprocess]Add Readme doc for preprocess (#518 )	2025-06-14 20:41:13 -04:00
Yongqi Chen	4afb0cfe4f	[Feature][Training]vsa for t2v training ready (#513 )	2025-06-14 01:08:00 -04:00
Zhang Peiyuan	3eec1281cf	[misc] Fix preprocessing and dataloader extra padding (#514 )	2025-06-13 15:15:33 -07:00
Wenxuan Tan	0660489e38	[CI] Restrict training CI to v1 (#508 )	2025-06-12 15:26:05 -07:00
Zhang Peiyuan	dd871a17bf	fix logging (#509 )	2025-06-12 15:24:12 -07:00
Jinzhe Pan Peiyuan Zhang Will Lin	dc11529862	[Refactor][Configurations] clean config orgnization (#505 ) Co-authored-by: Peiyuan Zhang <a1286225768@gmail.com> Co-authored-by: Will Lin <wlsaidhi@gmail.com>	2025-06-12 13:27:08 -07:00
Zhang Peiyuan	ffabf85e31	[feat] Add parquet iterable dataset. (#506 )	2025-06-12 04:30:56 -04:00
William Lin	c0026ca5ba	[CI] [Training] Initial e2e small training test (#504 )	2025-06-11 13:53:36 -07:00
Zhang Peiyuan	0f2bbe71ac	[misc] rename dp_size to hdsp_replicate_dim (#491 )	2025-06-10 16:36:56 -07:00
Yongqi ChenandJerryZhou54	2a46902ecb	[Feature][VSA]Update STA publish workflow (#498 ) Co-authored-by: JerryZhou54 <zhouw.jerry2017@outlook.com>	2025-06-10 19:33:34 -04:00
Zhang Peiyuan	66012d3a4c	[Feat][Dataloader] 1/n Refactor parquet map-style dataloader (#492 )	2025-06-10 16:00:13 -07:00
William Lin	f666b9de41	[misc] Add missing license headers (#499 )	2025-06-10 14:25:32 -07:00
Yongqi Chen	7e3c073b55	[Feature] Adding VSA inference (#478 )	2025-06-10 16:03:53 -04:00
Wei Zhou	a6aa21bd07	[bugfix][Cli Inference] Resolve runtime errors when running fastvideo generate (#495 )	2025-06-10 02:50:21 -04:00
Wenxuan Tan	6519b57aab	[chore] Fix main pre-commit CI failure (#494 )	2025-06-10 00:54:37 -05:00
Wei Zhou	675aea6ece	[bugfix][Cli Inference] Resolve runtime errors when running fastvideo generate (#493 )	2025-06-09 19:49:34 -07:00
Zhang Peiyuan	46e7a15e0d	[misc] Improve distributed related env variables and setup (#487 )	2025-06-08 09:14:48 -07:00
Yongqi ChenandJerryZhou54	e4f702d7ec	[Bug] Fix multi gpus issues in v1 scripts (#489 ) Co-authored-by: JerryZhou54 <zhouw.jerry2017@outlook.com>	2025-06-07 21:55:49 -07:00
Wenxuan Tan	bb68fcc809	Revert "Add torch.compile for all small ops" (#484 )	2025-06-07 07:21:17 -05:00
Wenxuan Tan	b392e6a874	Add torch.compile for all small ops (#432 )	2025-06-06 21:10:42 -07:00
Zhang PeiyuanandWill Lin	0991003905	[bugfix] [misc] fix denoising stage init; rename distributed env function; fix logging. (#481 ) Co-authored-by: Will Lin <wlsaidhi@gmail.com>	2025-06-06 20:01:10 -07:00
Zhang PeiyuanandWill Lin	8f8ce6d9e1	[bugfix] [training] Add negative prompt to preprocessing and validation (#479 ) Co-authored-by: Will Lin <wlsaidhi@gmail.com>	2025-06-06 11:21:21 -07:00
KevinzzandBrianChen1129	e3d0cbe185	[STA] Implement mask search for V1's Wan2.1 (#415 ) Co-authored-by: BrianChen1129 <yongqich@umich.edu>	2025-06-05 15:03:31 -04:00
eeecho	d5ec468d43	[V0] [Distill] support distill in V0 for wan (#444 )	2025-06-04 16:20:14 -07:00
Yongqi Chen	e55fa6e5dc	[Preprocess] I2V dataset (#473 )	2025-06-04 16:19:20 -07:00
Wenxuan Tan	61b6ddeee1	[Issue template] Move env report to the end for readability (#476 )	2025-06-04 16:18:22 -07:00
William Lin	a9a000f45d	[bugfix] fix bz >1 for training (#477 )	2025-06-04 16:03:39 -07:00
Wenxuan Tan	66b8b8561e	[LoRA] Support V1 LoRA inference (#451 )	2025-06-04 16:19:29 -05:00
Wenxuan Tan	6684872616	[misc] Find unused port in distributed init (#475 )	2025-06-04 14:15:12 -05:00
Wenxuan Tan	7f654e3332	[misc] Polish V1 training code (#469 )	2025-06-03 22:02:03 -05:00
William Lin	8631c1b806	[Training] Support Multi-Node training with FSDP + SP (#459 )	2025-06-03 17:40:22 -07:00
Wei Zhouand“BrianChen1129”	5357e12b5a	Update v1 inference scripts (#467 ) Co-authored-by: “BrianChen1129” <yongqich@umich.edu>	2025-06-03 02:54:12 -04:00
Kevin Lin	bdfdf1dfee	[Training] Add distributed checkpointing (#458 )	2025-06-02 15:15:45 -07:00
Wei Zhou	d156461785	Fix WanVideo (#461 )	2025-06-01 23:08:11 -07:00
Yongqi Chen	6edf113838	[Misc] Bring back mask files under asset/ and update new Wan mask strategy file (#462 )	2025-06-01 19:46:11 -07:00
William Lin	dcf7738cbc	[Misc] disable cast_forward_inputs (#460 )	2025-05-31 22:06:29 -07:00
Wenxuan Tan	b2ebaaf865	[Misc] Remove InferenceEngine (#455 )	2025-05-31 11:12:05 -07:00
Wenxuan Tan	7768bb80f6	misc: add remote pdb for debugging workers (#456 )	2025-05-30 20:44:27 -05:00
William Lin	a335811869	[Training] [8/n] SP Training (#450 )	2025-05-29 17:02:26 -07:00
William LinandZihang-He	357b0533fe	[Training] [7/n] gradient clipping (#449 ) Co-authored-by: Zihang-He <z6he@ucsd.edu>	2025-05-29 15:07:29 -07:00
William Lin	2ec3732758	[Training] [6/n]Mixed precision training (#448 )	2025-05-29 14:34:28 -07:00
Wei Zhouand“BrianChen1129”	a004408a93	[Training] [0/n] Add preprocessing pipeline (#442 ) Co-authored-by: “BrianChen1129” <yongqich@umich.edu>	2025-05-29 14:30:09 -07:00
William Lin JerryZhou54 Wei Zhou Kevin Lin “BrianChen1129”	007e237e69	[Training] [5/n] Add single gpu training pipeline (#447 ) Co-authored-by: JerryZhou54 <zhouw.jerry2017@outlook.com> Co-authored-by: Wei Zhou <69577934+JerryZhou54@users.noreply.github.com> Co-authored-by: Kevin Lin <42618777+kevin314@users.noreply.github.com> Co-authored-by: “BrianChen1129” <yongqich@umich.edu>	2025-05-29 11:49:46 -07:00
Yongqi Chen	8e18dc9f71	Update STA mask strategy downloading (#445 )	2025-05-28 12:49:22 -07:00
William Lin Wei Zhou JerryZhou54 “BrianChen1129”	7ab32539af	[Training] [1/n] Add latent datasets (#438 ) Co-authored-by: Wei Zhou <wzhou322@gatech.edu> Co-authored-by: JerryZhou54 <zhouw.jerry2017@outlook.com> Co-authored-by: “BrianChen1129” <yongqich@umich.edu>	2025-05-28 11:01:52 -07:00
William Lin	6ef8fcb61d	[Training] [4/n] add training save checkpoint (#441 )	2025-05-27 17:53:53 -07:00
William Lin	016e24da63	[Training] [3/n] Add training args and dependencies (#440 )	2025-05-27 17:53:39 -07:00
William Lin	85b8717545	[Training] [2/n] add bwd for all2all and all_gather (#439 )	2025-05-27 14:27:54 -07:00
Wenxuan Tan	657fd745e1	misc: Trigger transformers CI for layers and attention code change (#434 )	2025-05-27 11:43:23 -07:00
applesaucethebunandBrayden Zhong	12647457a7	[Misc] Small fixes to Torch code (#395 ) Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>	2025-05-23 14:40:24 -07:00
Kevin Lin	298f74f956	Set device for encode (#420 )	2025-05-23 14:19:45 -07:00
Wenxuan Tan	ee8babb298	Unify env report script in issue template (#423 )	2025-05-23 14:19:18 -07:00
Wenxuan Tan	60295cc03f	Use version.py (#424 )	2025-05-23 14:17:33 -07:00
William Lin	1572e13b6e	[Tests] don't run 3.10 and 3.11 for SSIM (#427 )	2025-05-23 12:59:50 -07:00
Wenxuan Tan	a157275b4c	Fix version number (#422 )	2025-05-22 12:37:34 -07:00
William Lin	c4dbe7dac3	[bug] fix bs > 1 (#418 )	2025-05-21 21:07:01 -07:00
Kevin Lin	d39591108e	Fulfill worker response on interrupt (#417 )	2025-05-21 20:59:11 -07:00
William Lin	ace6e971e5	[V1] Remove vLLM dependency (#413 )	2025-05-18 00:57:28 -07:00
William Lin	b4f6758253	[Teacache] allow None for forward_context batch when using teacache (#412 )	2025-05-17 18:43:14 -07:00
William Lin	535d29b392	[Docs] Fix image (#407 )	2025-05-12 14:30:05 -07:00
William Lin	b4255517e0	[Docs] Add CLI docs (#406 )	2025-05-12 14:16:48 -07:00