overfit SP

Co-authored-by: Kevin Lin <42618777+kevin314@users.noreply.github.com>
Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com>
Co-authored-by: Will Lin <wlsaidhi@gmail.com>

.github/ISSUE_TEMPLATE/1-bug-report.yml

@@ -8,7 +8,7 @@ body:
   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.
+      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

.github/scripts/runpod_api.py

@@ -0,0 +1,250 @@
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+
+import requests
+
+
+def parse_arguments():
+    """Parse command line arguments"""
+    parser = argparse.ArgumentParser(description='Run tests on RunPod GPU')
+    parser.add_argument('--gpu-type', type=str, help='GPU type to use')
+    parser.add_argument('--gpu-count',
+                        type=int,
+                        help='Number of GPUs to use',
+                        default=1)
+    parser.add_argument('--test-command', type=str, help='Test command to run')
+    parser.add_argument('--disk-size',
+                        type=int,
+                        default=20,
+                        help='Container disk size in GB (default: 20)')
+    parser.add_argument('--volume-size',
+                        type=int,
+                        default=20,
+                        help='Persistent volume size in GB (default: 20)')
+    parser.add_argument(
+        '--image',
+        type=str,
+        required=True,
+        help='Docker image to use')
+    return parser.parse_args()
+
+
+args = parse_arguments()
+API_KEY = os.environ['RUNPOD_API_KEY']
+RUN_ID = os.environ['GITHUB_RUN_ID']
+JOB_ID = os.environ['JOB_ID']
+PODS_API = "https://rest.runpod.io/v1/pods"
+HEADERS = {
+    "Content-Type": "application/json",
+    "Authorization": f"Bearer {API_KEY}"
+}
+
+
+def create_pod():
+    """Create a RunPod instance"""
+    # Ensure image name is lowercase (Docker requirement)
+    image_name = args.image.lower()
+    print(f"Using specified image: {image_name}")
+
+    docker_start_cmd = [
+        "bash", 
+        "-c", 
+        "apt update;DEBIAN_FRONTEND=noninteractive apt-get install openssh-server -y;mkdir -p ~/.ssh;cd $_;chmod 700 ~/.ssh;echo \"$PUBLIC_KEY\" >> authorized_keys;chmod 700 authorized_keys;service ssh start;sleep infinity"
+    ]
+    
+    print(f"Creating RunPod instance with GPU: {args.gpu_type}...")
+    payload = {
+        "name": f"fastvideo-{JOB_ID}-{RUN_ID}",
+        "containerDiskInGb": args.disk_size,
+        "volumeInGb": args.volume_size,
+        "gpuTypeIds": [args.gpu_type],
+        "gpuCount": args.gpu_count,
+        "imageName": image_name,
+        "allowedCudaVersions": ["12.4"],
+        "dockerStartCmd": docker_start_cmd
+    }
+
+    response = requests.post(PODS_API, headers=HEADERS, json=payload)
+    response_data = response.json()
+    print(f"Response: {json.dumps(response_data, indent=2)}")
+
+    return response_data["id"]
+
+
+def wait_for_pod(pod_id):
+    """Wait for pod to be in RUNNING state and fully ready with SSH access"""
+    print("Waiting for RunPod to be ready...")
+
+    # First wait for RUNNING status
+    max_attempts = 10
+    attempts = 0
+    while attempts < max_attempts:
+        response = requests.get(f"{PODS_API}/{pod_id}", headers=HEADERS)
+        pod_data = response.json()
+        status = pod_data["desiredStatus"]
+
+        if status == "RUNNING":
+            print("RunPod is running! Now waiting for ports to be assigned...")
+            break
+
+        print(
+            f"Current status: {status}, waiting... (attempt {attempts+1}/{max_attempts})"
+        )
+        time.sleep(2)
+        attempts += 1
+
+    if attempts >= max_attempts:
+        raise TimeoutError(
+            "Timed out waiting for RunPod to reach RUNNING state")
+
+    # Wait for ports to be assigned
+    max_attempts = 50
+    attempts = 0
+    while attempts < max_attempts:
+        response = requests.get(f"{PODS_API}/{pod_id}", headers=HEADERS)
+        pod_data = response.json()
+        port_mappings = pod_data.get("portMappings")
+
+        if (port_mappings is not None and "22" in port_mappings
+                and pod_data.get("publicIp", "") != ""):
+            print("RunPod is ready with SSH access!")
+            print(f"SSH IP: {pod_data['publicIp']}")
+            print(f"SSH Port: {port_mappings['22']}")
+            break
+
+        print(
+            f"Waiting for SSH port and public IP to be available... (attempt {attempts+1}/{max_attempts})"
+        )
+        time.sleep(20)
+        attempts += 1
+
+    if attempts >= max_attempts:
+        raise TimeoutError("Timed out waiting for RunPod SSH access")
+
+
+def execute_command(pod_id):
+    """Execute command on the pod via SSH using system SSH client"""
+    print(f"Running command: {args.test_command}")
+
+    response = requests.get(f"{PODS_API}/{pod_id}", headers=HEADERS)
+    pod_data = response.json()
+    ssh_ip = pod_data["publicIp"]
+    ssh_port = pod_data["portMappings"]["22"]
+
+    # Copy the repository to the pod using scp
+    repo_dir = os.path.abspath(os.getcwd())
+    repo_name = os.path.basename(repo_dir)
+
+    print(f"Copying repository from {repo_dir} to RunPod...")
+
+    tar_command = [
+        "tar", "-czf", "/tmp/repo.tar.gz", "-C",
+        os.path.dirname(repo_dir), repo_name
+    ]
+    subprocess.run(tar_command, check=True)
+
+    # Copy the tarball to the pod
+    scp_command = [
+        "scp", "-o", "StrictHostKeyChecking=no", "-o",
+        "UserKnownHostsFile=/dev/null", "-o", "ServerAliveInterval=60", "-o",
+        "ServerAliveCountMax=10", "-P",
+        str(ssh_port), "/tmp/repo.tar.gz", f"root@{ssh_ip}:/tmp/"
+    ]
+    subprocess.run(scp_command, check=True)
+
+    # For custom image, we can use the pre-configured environment
+    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",
+        args.test_command
+    ]
+    
+    remote_command = " && ".join(setup_steps)
+
+    ssh_command = [
+        "ssh", "-o", "StrictHostKeyChecking=no", "-o",
+        "UserKnownHostsFile=/dev/null", "-o", "ServerAliveInterval=60", "-o",
+        "ServerAliveCountMax=10", "-p",
+        str(ssh_port), f"root@{ssh_ip}", remote_command
+    ]
+
+    print(f"Connecting to {ssh_ip}:{ssh_port}...")
+
+    try:
+        process = subprocess.Popen(ssh_command,
+                                   stdout=subprocess.PIPE,
+                                   stderr=subprocess.STDOUT,
+                                   universal_newlines=True,
+                                   bufsize=0)
+
+        stdout_lines = []
+
+        print("Command output:")
+
+        for line in iter(process.stdout.readline, ''):
+            print(line.strip())
+            stdout_lines.append(line)
+
+        process.wait()
+
+        return_code = process.returncode
+        success = return_code == 0
+
+        stdout_str = "".join(stdout_lines)
+
+        if success:
+            print("Command executed successfully")
+        else:
+            print(f"Command failed with exit code {return_code}")
+
+        result = {
+            "success": success,
+            "return_code": return_code,
+            "stdout": stdout_str,
+            "stderr": ""
+        }
+        return result
+
+    except Exception as e:
+        print(f"Error executing SSH command: {str(e)}")
+        result = {"success": False, "error": str(e), "stdout": "", "stderr": ""}
+        return result
+
+
+def terminate_pod(pod_id):
+    """Terminate the pod"""
+    print("Terminating RunPod...")
+    requests.delete(f"{PODS_API}/{pod_id}", headers=HEADERS)
+    print(f"Terminated pod {pod_id}")
+
+
+def main():
+    pod_id = None
+    try:
+        pod_id = create_pod()
+        wait_for_pod(pod_id)
+        result = execute_command(pod_id)
+
+        if result.get("error") is not None:
+            print(f"Error executing command: {result['error']}")
+            sys.exit(1)
+
+        if not result.get("success", False):
+            print(
+                "Tests failed - check the output above for details on which tests failed"
+            )
+            sys.exit(1)
+
+    finally:
+        if pod_id:
+            terminate_pod(pod_id)
+
+
+if __name__ == "__main__":
+    main()

.github/scripts/runpod_cleanup.py

@@ -0,0 +1,90 @@
+import json
+import os
+import sys
+import uuid
+
+import requests
+
+API_KEY = os.environ['RUNPOD_API_KEY']
+RUN_ID = os.environ.get('GITHUB_RUN_ID', str(uuid.uuid4()))
+PODS_API = "https://rest.runpod.io/v1/pods"
+HEADERS = {
+    "Content-Type": "application/json",
+    "Authorization": f"Bearer {API_KEY}"
+}
+
+
+def get_job_ids():
+    """Parse job IDs from environment variable"""
+    job_ids_str = os.environ.get('JOB_IDS')
+    try:
+        job_ids = json.loads(job_ids_str)
+        if not isinstance(job_ids, list):
+            print("Error: JOB_IDS is not a list.")
+            sys.exit(1)
+        return job_ids
+    except json.JSONDecodeError as e:
+        print(f"Error parsing JOB_IDS: {e}")
+        sys.exit(1)
+
+
+def cleanup_pods():
+    """Find and terminate RunPod instances"""
+    print(f"Run ID: {RUN_ID}")
+
+    single_job_id = os.environ.get('JOB_ID')
+    
+    if single_job_id:
+        job_ids = [single_job_id]
+        print(f"Job ID: {single_job_id}")
+    else:
+        job_ids = get_job_ids()
+        print(f"Job IDs: {job_ids}")
+
+    # Get all pods associated with RunPod API_KEY
+    try:
+        response = requests.get(PODS_API, headers=HEADERS)
+        response.raise_for_status()
+        pods = response.json()
+    except requests.exceptions.RequestException as e:
+        print(f"Error getting pods: {e}")
+        sys.exit(1)
+
+    # Find and terminate pods created by this workflow run
+    terminated_pods = []
+    for pod in pods:
+        pod_name = pod.get("name", "")
+        pod_id = pod.get("id")
+
+        # Check if this pod was created by one of our jobs
+        if any(f"{job_id}-{RUN_ID}" in pod_name for job_id in job_ids):
+            print(f"Found pod: {pod_id} ({pod_name})")
+            try:
+                print(f"Terminating pod {pod_id}...")
+                term_response = requests.delete(f"{PODS_API}/{pod_id}",
+                                                headers=HEADERS)
+                term_response.raise_for_status()
+                terminated_pods.append(pod_id)
+                print(f"Successfully terminated pod {pod_id}")
+            except requests.exceptions.RequestException as e:
+                print(f"Error terminating pod {pod_id}: {e}")
+                sys.exit(1)
+    
+    if terminated_pods:
+        if single_job_id:
+            print(f"Terminated pod: {terminated_pods[0]}")
+        else:
+            print(f"Terminated {len(terminated_pods)} pods: {terminated_pods}")
+    else:
+        if single_job_id:
+            print(f"No pod found matching pattern: {single_job_id}-{RUN_ID}")
+        else:
+            print("No pods found to terminate.")
+
+
+def main():
+    cleanup_pods()
+
+
+if __name__ == "__main__":
+    main()

.github/workflows/build-image-template.yml

@@ -0,0 +1,106 @@
+name: Build Image Template
+
+on:
+  workflow_call:
+    inputs:
+      python_version:
+        required: true
+        type: string
+      dockerfile_path:
+        required: true
+        type: string
+      tag_suffix:
+        required: true
+        type: string
+
+jobs:
+  build-and-push:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      packages: write
+    
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+      
+      - name: Free up disk space
+        run: |
+          # Display initial space
+          echo "Initial disk space:"
+          df -h
+          
+          # Remove large directories directly
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /usr/local/share/boost
+          sudo rm -rf /usr/share/swift
+          sudo rm -rf /usr/local/lib/node_modules
+          sudo rm -rf /usr/local/share/powershell
+          sudo rm -rf /usr/share/rust
+          sudo rm -rf /usr/local/.ghcup
+          
+          # Remove cached files
+          sudo rm -rf /var/lib/apt/lists/*
+          sudo rm -rf /var/cache/apt/archives/*
+          
+          # Clean Docker
+          docker system prune -af --volumes
+          
+          # Display available space after cleanup
+          echo "Disk space after cleanup:"
+          df -h
+      
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      
+      - name: Login to GitHub Container Registry
+        uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.repository_owner }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+      
+      - name: Prepare tags
+        id: prepare-tags
+        run: |
+          SHORT_SHA=$(echo ${{ github.sha }} | cut -c1-7)
+          
+          TAGS="type=raw,value=${{ inputs.tag_suffix }}-latest"
+          TAGS="${TAGS}\ntype=raw,value=${{ inputs.tag_suffix }}-sha-${SHORT_SHA}"
+          
+          # Set Python 3.10 as the default image
+          if [[ "${{ inputs.python_version }}" == "3.10" ]]; then
+            TAGS="${TAGS}\ntype=raw,value=latest"
+          fi
+          
+          {
+            echo "tags<<EOF"
+            echo -e "$TAGS"
+            echo "EOF" 
+          } >> $GITHUB_OUTPUT
+      
+      - name: Extract metadata for Docker
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ghcr.io/${{ github.repository }}/fastvideo-dev
+          tags: ${{ steps.prepare-tags.outputs.tags }}
+      
+      - name: Build and push Docker image
+        id: build-push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: ${{ inputs.dockerfile_path }}
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+      
+      - name: Success message
+        run: |
+          echo "✅ Python ${{ inputs.python_version }} image successfully built and pushed to ghcr.io/${{ github.repository }}/fastvideo-dev:${{ inputs.tag_suffix }}-latest"
+          echo "To run tests with this image, manually trigger the 'Run Tests' workflow." 

.github/workflows/build-image.yml

@@ -0,0 +1,52 @@
+name: Build and Push Docker Images
+
+on:
+  workflow_dispatch:
+    inputs:
+      python_3_10:
+        description: 'Build Python 3.10 image'
+        required: false
+        default: false
+        type: boolean
+      python_3_11:
+        description: 'Build Python 3.11 image'
+        required: false
+        default: false
+        type: boolean
+      python_3_12:
+        description: 'Build Python 3.12 image'
+        required: false
+        default: false
+        type: boolean
+
+permissions:
+  contents: read
+  packages: write
+
+jobs:
+  build-python-3-10:
+    if: ${{ github.event.inputs.python_3_10 == 'true' }}
+    uses: ./.github/workflows/build-image-template.yml
+    with:
+      python_version: '3.10'
+      dockerfile_path: docker/Dockerfile.python3.10
+      tag_suffix: py3.10
+    secrets: inherit
+
+  build-python-3-11:
+    if: ${{ github.event.inputs.python_3_11 == 'true' }}
+    uses: ./.github/workflows/build-image-template.yml
+    with:
+      python_version: '3.11'
+      dockerfile_path: docker/Dockerfile.python3.11
+      tag_suffix: py3.11
+    secrets: inherit
+
+  build-python-3-12:
+    if: ${{ github.event.inputs.python_3_12 == 'true' }}
+    uses: ./.github/workflows/build-image-template.yml
+    with:
+      python_version: '3.12'
+      dockerfile_path: docker/Dockerfile.python3.12
+      tag_suffix: py3.12
+    secrets: inherit

.github/workflows/codespell.yml

@@ -1,45 +0,0 @@
-name: codespell
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - "**/*.md"
-      - "**/*.rst"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/codespell.yml
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - "**/*.md"
-      - "**/*.rst"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/codespell.yml
-
-jobs:
-  codespell:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Check out repository
-        uses: actions/checkout@v3
-
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.12'  # or any version you need
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-lint.txt
-      - name: Spelling check with codespell
-        run: |
-          # Refer to the above environment variable here
-          codespell --toml pyproject.toml $CODESPELL_EXCLUDES

.github/workflows/docs.yml

@@ -0,0 +1,83 @@
+# Sample workflow for building and deploying a Hugo site to GitHub Pages
+name: Deploy FastVideo Docs to Pages
+
+on:
+  # Runs on pushes targeting the default branch
+  push:
+    branches:
+      - main
+    paths:
+      - "docs/**/*.md"
+      - "fastvideo/v1/examples/**/*.py"
+  pull_request:
+    branches: 
+      - main
+    types: [opened, ready_for_review, synchronize, reopened]
+    paths:
+      - "docs/**/*.md"
+      - "fastvideo/v1/examples/**/*.py"
+
+  # Allows you to run this workflow manually from the Actions tab
+  workflow_dispatch:
+
+# Sets permissions of the GITHUB_TOKEN to allow deployment to GitHub Pages
+permissions:
+  contents: read
+  pages: write
+  id-token: write
+
+# Allow only one concurrent deployment, skipping runs queued between the run in-progress and latest queued.
+# However, do NOT cancel in-progress runs as we want to allow these production deployments to complete.
+concurrency:
+  group: "pages"
+  cancel-in-progress: false
+
+# Default to bash
+defaults:
+  run:
+    shell: bash
+
+jobs:
+  pre-commit:
+    uses: ./.github/workflows/pre-commit.yml
+
+  # Build job
+  build:
+    runs-on: ubuntu-latest
+    needs: pre-commit
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Setup Pages
+        id: pages
+        uses: actions/configure-pages@v5
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+      - name: Install dependencies
+        run: |
+          cd docs
+          pip install -r requirements-docs.txt
+      - name: Build docs
+        run: |
+          cd docs
+          make clean
+          make html
+      - name: Upload artifact
+        uses: actions/upload-pages-artifact@v3
+        with:
+          path: ./docs/build/html
+
+  # Deployment job
+  deploy:
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    if: ${{ github.event_name == 'push' }}
+    runs-on: ubuntu-latest
+    needs: build
+    steps:
+      - name: Deploy to GitHub Pages
+        id: deployment
+        uses: actions/deploy-pages@v4

.github/workflows/fastvideo-publish.yml

@@ -0,0 +1,71 @@
+name: Publish FastVideo to PyPI on Version Change
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - 'pyproject.toml'  # Trigger when pyproject.toml changes
+  workflow_dispatch:
+
+jobs:
+  check-version-change:
+    runs-on: ubuntu-latest
+    outputs:
+      version-changed: ${{ steps.check-version.outputs.changed }}
+      new-version: ${{ steps.check-version.outputs.new-version }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 2
+
+      - name: Check if version changed
+        id: check-version
+        run: |
+          # Get current commit's version
+          NEW_VERSION=$(grep -oP "version\\s*=\\s*\"\\K[^\"]+\"" pyproject.toml)
+          echo "New version: $NEW_VERSION"
+          
+          # Get previous version from git history
+          OLD_VERSION=$(git show HEAD~1:./pyproject.toml | grep -oP "version\\s*=\\s*\"\\K[^\"]+\"" || echo "0.0.0")
+          echo "Old version: $OLD_VERSION"
+          
+          if [ "$NEW_VERSION" != "$OLD_VERSION" ]; then
+            echo "Version changed from $OLD_VERSION to $NEW_VERSION"
+            echo "changed=true" >> $GITHUB_OUTPUT
+            echo "new-version=$NEW_VERSION" >> $GITHUB_OUTPUT
+          else
+            echo "Version did not change"
+            echo "changed=false" >> $GITHUB_OUTPUT
+          fi
+
+  build-publish-main:
+    needs: check-version-change
+    if: ${{ needs.check-version-change.outputs.version-changed == 'true' || github.event_name == 'workflow_dispatch' }}
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write  # Needed for OIDC Trusted Publishing
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+
+      - name: Install build dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install build twine wheel
+
+      - name: Build package
+        run: |
+          python -m build
+          
+      - name: Publish release distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          packages-dir: dist/

.github/workflows/matchers/actionlint.json

@@ -0,0 +1,17 @@
+{
+  "problemMatcher": [
+    {
+      "owner": "actionlint",
+      "pattern": [
+        {
+          "regexp": "^(?:\\x1b\\[\\d+m)?(.+?)(?:\\x1b\\[\\d+m)*:(?:\\x1b\\[\\d+m)*(\\d+)(?:\\x1b\\[\\d+m)*:(?:\\x1b\\[\\d+m)*(\\d+)(?:\\x1b\\[\\d+m)*: (?:\\x1b\\[\\d+m)*(.+?)(?:\\x1b\\[\\d+m)* \\[(.+?)\\]$",
+          "file": 1,
+          "line": 2,
+          "column": 3,
+          "message": 4,
+          "code": 5
+        }
+      ]
+    }
+  ]
+}

.github/workflows/matchers/mypy.json

@@ -0,0 +1,16 @@
+{
+  "problemMatcher": [
+    {
+      "owner": "mypy",
+      "pattern": [
+        {
+          "regexp": "^(.+):(\\d+):\\s(error|warning):\\s(.+)$",
+          "file": 1,
+          "line": 2,
+          "severity": 3,
+          "message": 4
+        }
+      ]
+    }
+  ]
+}

.github/workflows/pr-test.yml

@@ -0,0 +1,185 @@
+name: PR Test
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - "fastvideo/**/*.py"
+      - ".github/workflows/pr-test.yml"
+  pull_request:
+    branches: [main]
+    types: [opened, ready_for_review, synchronize, reopened]
+    paths:
+      - "fastvideo/**/*.py"
+      - ".github/workflows/pr-test.yml"
+  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
+        default: false
+        type: boolean
+      run_vae_test:
+        description: "Run vae-test"
+        required: false
+        default: false
+        type: boolean
+      run_transformer_test:
+        description: "Run transformer-test"
+        required: false
+        default: false
+        type: boolean
+      run_ssim_test:
+        description: "Run ssim-test"
+        required: false
+        default: false
+        type: boolean
+
+env:
+  PYTHONUNBUFFERED: "1"
+
+concurrency:
+  group: pr-test-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  pre-commit:
+    uses: ./.github/workflows/pre-commit.yml
+
+  change-filter:
+    runs-on: ubuntu-latest
+    needs: pre-commit
+    if: ${{ github.event.pull_request.draft == false || github.event_name == 'workflow_dispatch' }}
+    outputs:
+      encoder-test: ${{ steps.filter.outputs.encoder-test }}
+      vae-test: ${{ steps.filter.outputs.vae-test }}
+      transformer-test: ${{ steps.filter.outputs.transformer-test }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dorny/paths-filter@v3
+        id: filter
+        with:
+          filters: |
+            encoder-test:
+              - 'fastvideo/v1/models/encoders/**'
+              - 'fastvideo/v1/models/loaders/**'
+              - 'fastvideo/v1/tests/encoders/**'
+            vae-test:
+              - 'fastvideo/v1/models/vaes/**'
+              - 'fastvideo/v1/models/loaders/**'
+              - 'fastvideo/v1/tests/vaes/**'
+            transformer-test:
+              - 'fastvideo/v1/models/dits/**'
+              - 'fastvideo/v1/models/loaders/**'
+              - 'fastvideo/v1/tests/transformers/**'
+              - 'fastvideo/v1/layers/**'
+              - 'fastvideo/v1/attention/**'
+
+  encoder-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && needs.change-filter.outputs.encoder-test == 'true') || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_encoder_test == 'true')
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "encoder-test"
+      gpu_type: "NVIDIA A40"
+      gpu_count: 1
+      volume_size: 100
+      image: "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
+      test_command: "pip install -e .[test] && pytest ./fastvideo/v1/tests/encoders -s"
+      timeout_minutes: 30
+    secrets:
+      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+
+  vae-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && needs.change-filter.outputs.vae-test == 'true') || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_vae_test == 'true')
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "vae-test"
+      gpu_type: "NVIDIA A40"
+      gpu_count: 1
+      volume_size: 100
+      image: "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
+      test_command: "pip install -e .[test] && pytest ./fastvideo/v1/tests/vaes -s"
+      timeout_minutes: 30
+    secrets:
+      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+
+  transformer-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && needs.change-filter.outputs.transformer-test == 'true') || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_transformer_test == 'true')
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "transformer-test"
+      gpu_type: "NVIDIA L40S"
+      gpu_count: 1
+      volume_size: 100
+      image: "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
+      test_command: "pip install -e .[test] && pytest ./fastvideo/v1/tests/transformers -s"
+      timeout_minutes: 30
+    secrets:
+      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+      
+  ssim-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_ssim_test == 'true')
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: [
+          # {version: "3.10", tag: "latest"},
+          # {version: "3.11", tag: "py3.11-latest"},
+          {version: "3.12", tag: "py3.12-latest"}
+        ]
+    uses: ./.github/workflows/runpod-test.yml
+    with:
+      job_id: "ssim-test-py${{ matrix.python-version.version }}"
+      gpu_type: "NVIDIA A40"
+      gpu_count: 2
+      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"
+      timeout_minutes: 60
+    secrets:
+      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+      RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+
+  runpod-cleanup:
+    needs: [encoder-test, vae-test, transformer-test, ssim-test] # Add other jobs to this list as you create them
+    if: ${{ always() && ((github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || github.event_name == 'workflow_dispatch') }}
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Install dependencies
+        run: pip install requests
+
+      - 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"]'
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+        run: python .github/scripts/runpod_cleanup.py

.github/workflows/pre-commit.yml

@@ -0,0 +1,18 @@
+name: pre-commit
+
+on:
+    workflow_call:
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v4
+    - uses: actions/setup-python@v5
+      with:
+        python-version: "3.10"
+    - 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
+      with:
+        extra_args: --all-files --hook-stage manual

.github/workflows/ruff.yml

@@ -1,50 +0,0 @@
-name: ruff
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/matchers/ruff.json
-      - .github/workflows/ruff.yml
-  pull_request:
-    branches:
-      - main
-    # This workflow is only relevant when one of the following files changes.
-    # However, we have github configured to expect and require this workflow
-    # to run and pass before github with auto-merge a pull request. Until github
-    # allows more flexible auto-merge policy, we can just run this on every PR.
-    # It doesn't take that long to run, anyway.
-    #paths:
-    #  - "**/*.py"
-    #  - pyproject.toml
-    #  - requirements-lint.txt
-    #  - .github/workflows/matchers/ruff.json
-    #  - .github/workflows/ruff.yml
-
-jobs:
-  ruff:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Check out repository
-        uses: actions/checkout@v3
-
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.12'  # or any version you need
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-lint.txt
-      - name: Analysing the code with ruff
-        run: |
-          ruff check .
-      - name: Run isort
-        run: |
-          isort . --check-only

.github/workflows/runpod-test.yml

@@ -0,0 +1,91 @@
+name: RunPod Test
+
+on:
+  workflow_call:
+    inputs:
+      job_id:
+        required: true
+        type: string
+        description: "Unique identifier for this test job"
+      gpu_type:
+        required: true
+        type: string
+        description: "GPU type to use (e.g. NVIDIA A40, NVIDIA L40S)"
+      gpu_count:
+        required: true
+        type: number
+        description: "Number of GPUs to use"
+      volume_size:
+        required: false
+        type: number
+        default: 20
+        description: "Volume size in GB"
+      disk_size:
+        required: false
+        type: number
+        default: 20
+        description: "Disk size in GB"
+      image:
+        required: true
+        type: string
+        description: "Docker image to use"
+      test_command:
+        required: true
+        type: string
+        description: "Command to run tests"
+      timeout_minutes:
+        required: false
+        type: number
+        default: 30
+        description: "Timeout in minutes"
+    secrets:
+      RUNPOD_API_KEY:
+        required: true
+      RUNPOD_PRIVATE_KEY:
+        required: true
+
+jobs:
+  run-test:
+    runs-on: ubuntu-latest
+    environment: runpod-runners
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Set up SSH key
+        run: |
+          mkdir -p ~/.ssh
+          echo "${{ secrets.RUNPOD_PRIVATE_KEY }}" > ~/.ssh/id_rsa
+          chmod 600 ~/.ssh/id_rsa
+          ssh-keygen -y -f ~/.ssh/id_rsa > ~/.ssh/id_rsa.pub
+
+      - name: Install dependencies
+        run: pip install requests
+
+      - name: Run tests on RunPod
+        env:
+          JOB_ID: ${{ inputs.job_id }}
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+        timeout-minutes: ${{ inputs.timeout_minutes }}
+        run: >-
+          python .github/scripts/runpod_api.py 
+          --gpu-type "${{ inputs.gpu_type }}"
+          --gpu-count ${{ inputs.gpu_count }}
+          --volume-size ${{ inputs.volume_size }}
+          --disk-size ${{ inputs.disk_size }}
+          --image "${{ inputs.image }}"
+          --test-command "${{ inputs.test_command }}"
+
+      - name: Terminate RunPod Instances
+        if: ${{ always() }}
+        env:
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+          JOB_ID: ${{ inputs.job_id }}
+        run: python .github/scripts/runpod_cleanup.py 

.github/workflows/sta-publish.yml

@@ -0,0 +1,245 @@
+name: Publish Sliding Tile Attention Kernel to PyPI on Version Change
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - "csrc/sliding_tile_attention/setup.py"
+  workflow_dispatch:
+
+jobs:
+  check-version-change:
+    runs-on: ubuntu-latest
+    outputs:
+      version-changed: ${{ steps.check-version.outputs.changed }}
+      new-version: ${{ steps.check-version.outputs.new-version }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 2
+
+      - name: Check if version changed
+        id: check-version
+        run: |
+          cd csrc/sliding_tile_attention
+          # Get current commit's version
+          NEW_VERSION=$(grep -oP 'VERSION\s*=\s*"\K[^"]+' setup.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")
+          echo "Old version: $OLD_VERSION"
+          
+          if [ "$NEW_VERSION" != "$OLD_VERSION" ]; then
+            echo "Version changed from $OLD_VERSION to $NEW_VERSION"
+            echo "changed=true" >> $GITHUB_OUTPUT
+            echo "new-version=$NEW_VERSION" >> $GITHUB_OUTPUT
+          else
+            echo "Version did not change"
+            echo "changed=false" >> $GITHUB_OUTPUT
+          fi
+
+  build_wheels:
+    name: Build Wheel
+    needs: check-version-change
+    if: ${{ needs.check-version-change.outputs.version-changed == 'true' || github.event_name == 'workflow_dispatch' }}
+    runs-on: ${{ matrix.os }}
+
+    strategy:
+      fail-fast: false
+      matrix:
+          # Using ubuntu-20.04 instead of 22.04 for more compatibility (glibc). Ideally we'd use the
+          # manylinux docker image, but I haven't figured out how to install CUDA on manylinux.
+          os: [ubuntu-22.04]
+          python-version: ['3.10', '3.11', '3.12', '3.13']
+          torch-version: ['2.5.1', '2.6.0']
+          cuda-version: ['12.4.1', '12.5.1', '12.6.3']
+
+    steps:
+      - name: Free up disk space
+        run: |
+          echo "Initial disk space:"
+          df -h
+          
+          # Remove large directories
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /usr/local/share/boost
+          sudo rm -rf /usr/share/swift
+          sudo rm -rf /usr/local/lib/node_modules
+          sudo rm -rf /usr/local/share/powershell
+          sudo rm -rf /usr/share/rust
+          sudo rm -rf /usr/local/.ghcup
+          
+          # Remove cached files
+          sudo rm -rf /var/lib/apt/lists/*
+          sudo rm -rf /var/cache/apt/archives/*
+          
+          echo "Disk space after cleanup:"
+          df -h
+
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install CUDA ${{ matrix.cuda-version }}
+        uses: Jimver/cuda-toolkit@v0.2.21
+        id: cuda-toolkit
+        with:
+          cuda: ${{ matrix.cuda-version }}
+          linux-local-args: '["--toolkit"]'
+          method: 'network'
+
+      - name: Install dependencies (GCC, Clang, CUDA Paths, Git)
+        run: |
+          sudo apt update
+          sudo apt install -y git patchelf gcc-11 g++-11 clang-11
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+
+          # Allow Git to Access Safe Directory
+          git config --global --add safe.directory /__w/FastVideo/FastVideo
+
+          # Set CUDA environment variables
+          export CUDA_HOME=/usr/local/cuda-${{ matrix.cuda-version }}
+          export PATH=${CUDA_HOME}/bin:${PATH}
+          export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+          
+          # Verify installation
+          gcc --version
+          g++ --version
+          clang-11 --version
+          nvcc --version
+
+      - name: Install PyTorch ${{ matrix.torch-version }}+cu${{ matrix.cuda-version }}
+        run: |
+          pip install --upgrade pip
+          # With python 3.13 and torch 2.5.1, unless we update typing-extensions, we get error
+          # AttributeError: attribute '__default__' of 'typing.ParamSpec' objects is not writable
+          pip install typing-extensions==4.12.2
+          # We want to figure out the CUDA version to download pytorch
+          # e.g. we can have system CUDA version being 11.7 but if torch==1.12 then we need to download the wheel from cu116
+          # see https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-compatibility-matrix
+          export TORCH_CUDA_VERSION=124
+          pip install --no-cache-dir torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
+          nvcc --version
+          python --version
+          python -c "import torch; print('PyTorch:', torch.__version__)"
+          python -c "import torch; print('CUDA:', torch.version.cuda)"
+          python -c "from torch.utils import cpp_extension; print (cpp_extension.CUDA_HOME)"
+
+      - name: Build wheel
+        run: |
+          # 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
+
+      - name: Rename wheel file
+        run: |
+          cd csrc/sliding_tile_attention
+
+          CUDA_SHORT_VERSION=$(echo ${{ matrix.cuda-version }} | cut -d. -f1,2 | sed 's/\.//g')
+          TORCH_SHORT_VERSION=$(echo ${{ matrix.torch-version }} | cut -d. -f1,2)
+          # Get the correct version format
+          tmpname=cu${CUDA_SHORT_VERSION}torch${TORCH_SHORT_VERSION}
+          wheel_name=$(ls dist/*whl | xargs -n 1 basename | sed "s/-/+$tmpname-/2")
+          # Rename with version information
+          ls dist/*whl |xargs -I {} mv {} dist/${wheel_name}
+          echo "wheel_name=${wheel_name}" >> $GITHUB_ENV
+
+      - name: Upload wheel artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: ${{ env.wheel_name }}
+          path: csrc/sliding_tile_attention/dist/*.whl
+          retention-days: 90
+
+  publish_package:
+    name: Publish package
+    needs: [build_wheels, check-version-change]
+    if: ${{ needs.check-version-change.outputs.version-changed == 'true' || github.event_name == 'workflow_dispatch' }}
+    runs-on: ubuntu-22.04
+    permissions:
+      id-token: write  # Needed for OIDC Trusted Publishing
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+      
+      - name: Install CUDA 12.4.1
+        uses: Jimver/cuda-toolkit@v0.2.21
+        id: cuda-toolkit
+        with:
+          cuda: 12.4.1
+          linux-local-args: '["--toolkit"]'
+          method: 'network'
+          sub-packages: '["nvcc"]'
+
+      - name: Install dependencies (GCC, Clang, CUDA Paths, Git)
+        run: |
+          sudo apt update
+          sudo apt install -y git patchelf gcc-11 g++-11 clang-11
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+
+          # Allow Git to Access Safe Directory
+          git config --global --add safe.directory /__w/FastVideo/FastVideo
+
+          # Set CUDA environment variables
+          export CUDA_HOME=/usr/local/cuda-12.4.1
+          export PATH=${CUDA_HOME}/bin:${PATH}
+          export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+          
+          # Verify installation
+          gcc --version
+          g++ --version
+          clang-11 --version
+          nvcc --version
+
+      - name: Install PyTorch 2.5.1+cu12.4.1
+        run: |
+          pip install --upgrade pip
+          # With python 3.13 and torch 2.5.1, unless we update typing-extensions, we get error
+          # AttributeError: attribute '__default__' of 'typing.ParamSpec' objects is not writable
+          pip install typing-extensions==4.12.2
+          # We want to figure out the CUDA version to download pytorch
+          # e.g. we can have system CUDA version being 11.7 but if torch==1.12 then we need to download the wheel from cu116
+          # see https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-compatibility-matrix
+          export TORCH_CUDA_VERSION=124
+          pip install --no-cache-dir torch==2.5.1 --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
+          nvcc --version
+          python --version
+          python -c "import torch; print('PyTorch:', torch.__version__)"
+          python -c "import torch; print('CUDA:', torch.version.cuda)"
+          python -c "from torch.utils import cpp_extension; print (cpp_extension.CUDA_HOME)"
+
+      - name: Build source distribution
+        run: |
+          # 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
+
+      - name: Publish release distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          packages-dir: csrc/sliding_tile_attention/dist/

.github/workflows/test.yml

@@ -11,10 +11,10 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Check out repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: '3.12'  # or any version you need
 
@@ -25,6 +25,7 @@ jobs:
           pip install packaging ninja
           pip install -e .
           pip install pytest
+
       - name: Run Pytest
         run: |
-          pytest
+          pytest --ignore csrc/sliding_tile_attention/test

.github/workflows/yapf.yml

@@ -1,38 +0,0 @@
-name: yapf
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - .github/workflows/yapf.yml
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - .github/workflows/yapf.yml
-
-jobs:
-  yapf:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Check out repository
-        uses: actions/checkout@v3
-
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.12'  # or any version you need
-
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install yapf==0.32.0
-          pip install toml==0.10.2
-      - name: Running yapf
-        run: |
-          yapf --diff --recursive .

.gitignore

@@ -1,17 +1,11 @@
-ucf101_stride4x4x4
 __pycache__
-*.mp4
 .ipynb_checkpoints
 *.pth
 UCF-101/
 results/
-build/
-fastvideo.egg-info/
 wandb/
-.idea
 *.ipynb
 *.jpg
-*.mp3
 *.safetensors
 *.mp4
 *.png
@@ -20,28 +14,6 @@ wandb/
 *.pt
 cache_dir/
 wandb/
-sample_video*
-sample_image*
-512*
-720*
-1024*
-debug*
-private*
-caption*
-*deepspeed*
-revised*
-129f*
-all*
-read*
-YSH*
-*pick*
-*ysh*
-hw*
-257f*
-513f*
-taming*
-221hw*
-65x512x512
 runs/
 samples/
 *validation/
@@ -51,3 +23,40 @@ outputs_video
 sbatch.sh
 *.out
 env
+*.o
+**/build/
+**.pyc
+**.txt
+**.json
+
+# Distribution / packaging
+build/
+dist/
+*.egg-info/
+*.egg
+eggs/
+.eggs/
+
+# Sphinx documentation
+docs/_build/
+docs/source/getting_started/examples/
+docs/source/inference/examples/
+
+# VSCode
+.vscode/
+
+# DS Store
+.DS_Store
+
+# vim swap files
+*.swo
+*.swp
+
+# Python pickle files
+*.pkl
+
+# Reference videos
+!fastvideo/v1/tests/ssim/reference_videos/**/*.mp4
+
+# Static images
+!docs/source/_static/images/**/*.png

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "csrc/sliding_tile_attention/tk"]
+	path = csrc/sliding_tile_attention/tk
+	url = https://github.com/HazyResearch/ThunderKittens.git

.pre-commit-config.yaml

@@ -0,0 +1,83 @@
+default_stages:
+  - pre-commit # Run locally
+  - manual # Run in CI
+exclude: |
+  (?x)(
+    fastvideo/v1/third_party/.*|
+    csrc/.*|
+    assets/.*|
+    tests/.*|
+    demo/.*|
+    predict\.py|
+    scripts/.*|
+    fastvideo/data_preprocess/.*|
+    fastvideo/dataset/.*|
+    fastvideo/distill/.*|
+    fastvideo/distill\.py|
+    fastvideo/distill_adv\.py|
+    fastvideo/models/.*|
+    fastvideo/sample/.*|
+    fastvideo/train\.py|
+    fastvideo/utils/.*|
+    examples/.*|
+    .github/workflows/fastvideo-publish.yml|
+    .github/workflows/sta-publish.yml|
+    .github/workflows/build-image-template.yml|
+    docs/source/inference/support_matrix.md
+  )
+repos:
+- repo: https://github.com/google/yapf
+  rev: v0.43.0
+  hooks:
+  - id: yapf
+    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
+  hooks:
+  - id: ruff
+    args: [--output-format, github, --fix]
+- repo: https://github.com/codespell-project/codespell
+  rev: v2.4.1
+  hooks:
+  - id: codespell
+    additional_dependencies: ['tomli']
+    args: ['--toml', 'pyproject.toml']
+- repo: https://github.com/PyCQA/isort
+  rev: 6.0.1
+  hooks:
+  - id: isort
+- repo: https://github.com/jackdewinter/pymarkdown
+  rev: v0.9.29
+  hooks:
+  - id: pymarkdown
+    args: [fix]
+- repo: https://github.com/rhysd/actionlint
+  rev: v1.7.7
+  hooks:
+  - id: actionlint
+- repo: https://github.com/pre-commit/mirrors-mypy
+  rev: v1.15.0
+  hooks:
+  - id: mypy
+    args: [--python-version, '3.10', --follow-imports, "skip", ]
+    additional_dependencies: [types-cachetools, types-setuptools, types-PyYAML, types-requests]
+- repo: local
+  hooks:
+  - id: check-filenames
+    name: Check for spaces in all filenames
+    entry: bash
+    args:
+      - -c
+      - 'git ls-files | grep -v "^fastvideo/v1/tests/ssim/" | grep " " && echo "Filenames should not contain spaces!" && exit 1 || exit 0'
+    language: system
+    always_run: true
+    pass_filenames: false
+  # Keep `suggestion` last
+  - id: suggestion
+    name: Suggestion
+    entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."'
+    language: system
+    verbose: true
+    pass_filenames: false
+  # Insert new entries above the `suggestion` entry

LICENSE

@@ -184,18 +184,4 @@
       comment syntax for the file format. We also recommend that a
       file or class name and description of purpose be included on the
       same "printed page" as the copyright notice for easier
-      identification within third-party archives.
-
-   Copyright [2023] Lightning AI
-
-   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.
+      identification within third-party archives.

README.md

@@ -2,186 +2,148 @@
 <img src=assets/logo.jpg width="30%"/>
 </div>
 
-FastVideo is a lightweight framework for accelerating large video diffusion models.
-
-https://github.com/user-attachments/assets/064ac1d2-11ed-4a0c-955b-4d412a96ef30
+**FastVideo is a unified framework for accelerated video generation.**
 
+It features a clean, consistent API that works across popular video models, making it easier for developers to author new models and incorporate system- or kernel-level optimizations.
+With FastVideo's optimizations, you can achieve more than 3x inference improvement compared to other systems.
 
 <p align="center">
-    🤗 <a href="https://huggingface.co/FastVideo/FastHunyuan"  target="_blank">FastHunyuan</a>  | 🤗 <a href="https://huggingface.co/FastVideo/FastMochi-diffusers" target="_blank">FastMochi</a> | 🎮 <a href="https://discord.gg/REBzDQTWWt" target="_blank"> Discord </a> | 🕹️ <a href="https://replicate.com/lucataco/fast-hunyuan-video" target="_blank"> Replicate </a> 
-</p> 
+    | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/FastVideo/FastHunyuan"  target="_blank"><b>FastHunyuan</b></a>  | 🤗 <a href="https://huggingface.co/FastVideo/FastMochi-diffusers" target="_blank"><b>FastMochi</b></a> | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-2zf6ru791-sRwI9lPIUJQq1mIeB_yjJg" target="_blank"> <b>Slack</b> </a> |
+</p>
 
+<div align="center">
+<img src=assets/perf.png width="90%"/>
+</div>
 
-FastVideo currently offers: (with more to come)
+## Key Features
 
-- FastHunyuan and FastMochi: consistency distilled video diffusion models for 8x inference speedup.
-- First open distillation recipes for video DiT, based on [PCM](https://github.com/G-U-N/Phased-Consistency-Model).
-- Support distilling/finetuning/inferencing state-of-the-art open video DiTs: 1. Mochi 2. Hunyuan.
+FastVideo has the following features:
+- State-of-the-art performance optimizations for inference
+  - [Sliding Tile Attention](https://arxiv.org/pdf/2502.04507)
+  - [TeaCache](https://arxiv.org/pdf/2411.19108)
+  - [Sage Attention](https://arxiv.org/abs/2410.02367)
+- Cutting edge models
+  - Wan2.1 T2V, I2V
+  - HunyuanVideo
+  - FastHunyuan: consistency distilled video diffusion models for 8x inference speedup.
+  - StepVideo T2V
+- Distillation support
+  - Recipes for video DiT, based on [PCM](https://github.com/G-U-N/Phased-Consistency-Model).
+  - Support distilling/finetuning/inferencing state-of-the-art open video DiTs: 1. Mochi 2. Hunyuan.
 - Scalable training with FSDP, sequence parallelism, and selective activation checkpointing, with near linear scaling to 64 GPUs.
 - Memory efficient finetuning with LoRA, precomputed latent, and precomputed text embeddings.
 
-Dev in progress and highly experimental.
-
-## 🎥 More Demos
-
-Fast-Mochi comparison with original Mochi, achieving an 8X diffusion speed boost with the FastVideo framework.
-
-https://github.com/user-attachments/assets/5fbc4596-56d6-43aa-98e0-da472cf8e26c
-
-Comparison between OpenAI Sora, original Hunyuan and FastHunyuan
-
-https://github.com/user-attachments/assets/d323b712-3f68-42b2-952b-94f6a49c4836
-
-Comparison between original FastHunyuan, LLM-INT8 quantized FastHunyuan and NF4 quantized FastHunyuan
-
-https://github.com/user-attachments/assets/cf89efb5-5f68-4949-a085-f41c1ef26c94
-
-## Change Log
-- ```2025/01/13```: Support Lora finetuning for HunyuanVideo.
-- ```2024/12/25```: Enable single 4090 inference for `FastHunyuan`, please rerun the installation steps to update the environment.
-- ```2024/12/17```: `FastVideo` v1.0 is released.
-
-
-## 🔧 Installation
-The code is tested on Python 3.10.0, CUDA 12.1 and H100.
-```
-./env_setup.sh fastvideo
-```
-
-## 🚀 Inference
-
-### Inference FastHunyuan on single RTX4090
-We now support NF4 and LLM-INT8 quantized inference using BitsAndBytes for FastHunyuan. With NF4 quantization, inference can be performed on a single RTX 4090 GPU, requiring just 20GB of VRAM.
-```bash
-# Download the model weight
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/FastHunyuan-diffusers --local_dir=data/FastHunyuan-diffusers --repo_type=model
-# CLI inference
-bash scripts/inference/inference_hunyuan_hf_quantization.sh
-```
-For more information about the VRAM requirements for BitsAndBytes quantization, please refer to the table below (timing measured on an H100 GPU):
-
-
-| Configuration                  | Memory to Init Transformer | Peak Memory After Init Pipeline (Denoise) | Diffusion Time | End-to-End Time |
-|--------------------------------|----------------------------|--------------------------------------------|----------------|-----------------|
-| BF16 + Pipeline CPU Offload    | 23.883G                   | 33.744G                                    | 81s            | 121.5s          |
-| INT8 + Pipeline CPU Offload    | 13.911G                   | 27.979G                                    | 88s            | 116.7s          |
-| NF4 + Pipeline CPU Offload     | 9.453G                    | 19.26G                                     | 78s            | 114.5s          |
-           
-
-
-For improved quality in generated videos, we recommend using a GPU with 80GB of memory to run the BF16 model with the original Hunyuan pipeline. To execute the inference, use the following section:
-
-### FastHunyuan
-```bash
-# Download the model weight
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/FastHunyuan --local_dir=data/FastHunyuan --repo_type=model
-# CLI inference
-bash scripts/inference/inference_hunyuan.sh
-```
-You can also inference FastHunyuan in the [official Hunyuan github](https://github.com/Tencent/HunyuanVideo).
-
-### FastMochi
+## Getting Started
+We recommend using an environment manager such as `Conda` to create a clean environment:
 
 ```bash
-# Download the model weight
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/FastMochi-diffusers --local_dir=data/FastMochi-diffusers --repo_type=model
-# CLI inference
-bash scripts/inference/inference_mochi_sp.sh
+# Create and activate a new conda environment
+conda create -n fastvideo python=3.12
+conda activate fastvideo
+
+# Install FastVideo
+pip install fastvideo
 ```
 
+Please see our [docs](https://hao-ai-lab.github.io/FastVideo/getting_started/installation.html) for more detailed installation instructions.
+
+## Inference
+### Generating Your First Video
+Here's a minimal example to generate a video using the default settings. Create a file called `example.py` with the following code:
+
+```python
+from fastvideo import VideoGenerator
+
+def main():
+    # Create a video generator with a pre-trained model
+    generator = VideoGenerator.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        num_gpus=1,  # Adjust based on your hardware
+    )
+
+    # Define a prompt for your video
+    prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest."
+
+    # Generate the video
+    video = generator.generate_video(
+        prompt,
+        return_frames=True,  # Also return frames from this call (defaults to False)
+        output_path="my_videos/",  # Controls where videos are saved
+        save_video=True
+    )
+
+if __name__ == '__main__':
+    main()
+```
+
+Run the script with:
 
-## 🎯 Distill
-Our distillation recipe is based on [Phased Consistency Model](https://github.com/G-U-N/Phased-Consistency-Model). We did not find significant improvement using multi-phase distillation, so we keep the one phase setup similar to the original latent consistency model's recipe.
-We use the [MixKit](https://huggingface.co/datasets/LanguageBind/Open-Sora-Plan-v1.1.0/tree/main/all_mixkit) dataset for distillation. To avoid running the text encoder and VAE during training, we preprocess all data to generate text embeddings and VAE latents.
-Preprocessing instructions can be found [data_preprocess.md](docs/data_preprocess.md). For convenience, we also provide preprocessed data that can be downloaded directly using the following command:
 ```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/HD-Mixkit-Finetune-Hunyuan --local_dir=data/HD-Mixkit-Finetune-Hunyuan --repo_type=dataset
+python example.py
 ```
-Next, download the original model weights with:
-```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/hunyuan --local_dir=data/hunyuan --repo_type=model # original hunyuan
-python scripts/huggingface/download_hf.py --repo_id=genmo/mochi-1-preview --local_dir=data/mochi --repo_type=model # original mochi
-```
-To launch the distillation process, use the following commands:
-```
-bash scripts/distill/distill_hunyuan.sh # for hunyuan
-bash scripts/distill/distill_mochi.sh # for mochi
-```
-We also provide an optional script for distillation with adversarial loss, located at `fastvideo/distill_adv.py`. Although we tried adversarial loss, we did not observe significant improvements.
-## Finetune
-### ⚡ Full Finetune
-Ensure your data is prepared and preprocessed in the format specified in [data_preprocess.md](docs/data_preprocess.md). For convenience, we also provide a mochi preprocessed Black Myth Wukong data that can be downloaded directly:
-```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/Mochi-Black-Myth --local_dir=data/Mochi-Black-Myth --repo_type=dataset
-```
-Download the original model weights as specified in [Distill Section](#-distill):
 
-Then you can run the finetune with:
-```
-bash scripts/finetune/finetune_mochi.sh # for mochi
-```
-**Note that for finetuning, we did not tune the hyperparameters in the provided script.**
-### ⚡ Lora Finetune 
+For a more detailed guide, please see our [inference quick start](https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html).
 
-Hunyuan supports Lora fine-tuning of videos up to 720p. Demos and prompts of Black-Myth-Wukong can be found in [here](https://huggingface.co/FastVideo/Hunyuan-Black-Myth-Wukong-lora-weight). You can download the Lora weight through:
-```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/Hunyuan-Black-Myth-Wukong-lora-weight --local_dir=data/Hunyuan-Black-Myth-Wukong-lora-weight --repo_type=model
-```
-#### Minimum Hardware Requirement
-- 40 GB GPU memory each for 2 GPUs with lora.
-- 30 GB GPU memory each for 2 GPUs with CPU offload and lora.  
+### Other docs:
 
+- [Design Overview](https://hao-ai-lab.github.io/FastVideo/design/overview.html)
+- [Contribution Guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation.html)
 
-Currently, both Mochi and Hunyuan models support Lora finetuning through diffusers. To generate personalized videos from your own dataset, you'll need to follow three main steps: dataset preparation, finetuning, and inference.
-
-#### Dataset Preparation
-We provide scripts to better help you get started to train on your own characters!  
-You can run this to organize your dataset to get the videos2caption.json before preprocess. Specify your video folder and corresponding caption folder (caption files should be .txt files and have the same name with its video):
-```
-python scripts/dataset_preparation/prepare_json_file.py --video_dir data/input_videos/ --prompt_dir data/captions/ --output_path data/output_folder/videos2caption.json --verbose
-```
-Also, we provide script to resize your videos:
-```
-python scripts/data_preprocess/resize_videos.py 
-```
-#### Finetuning
-After basic dataset preparation and preprocess, you can start to finetune your model using Lora:
-```
-bash scripts/finetune/finetune_hunyuan_hf_lora.sh
-```
-#### Inference
-For inference with Lora checkpoint, you can run the following scripts with additional parameter `--lora_checkpoint_dir`:
-```
-bash scripts/inference/inference_hunyuan_hf.sh 
-```
-**We also provide scripts for Mochi in the same directory.**
-
-#### Finetune with Both Image and Video
-Our codebase support finetuning with both image and video. 
-```bash
-bash scripts/finetune/finetune_hunyuan.sh
-bash scripts/finetune/finetune_mochi_lora_mix.sh
-```
-For Image-Video Mixture Fine-tuning, make sure to enable the `--group_frame` option in your script.
+## 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)
 
 ## 📑 Development Plan
 
-- More distillation methods
-  - [ ] Add Distribution Matching Distillation
+<!-- - More distillation methods -->
+  <!-- - [ ] Add Distribution Matching Distillation -->
 - More models support
-  - [ ] Add CogvideoX model
-- Code update
-  - [ ] fp8 support
-  - [ ] faster load model and save model support
+  <!-- - [ ] Add CogvideoX model -->
+  - [x] Add StepVideo to V1
+- Optimization features
+  - [x] Teacache in V1
+  - [x] SageAttention in V1
+- Code updates
+  - [x] V1 Configuration API
+  - [ ] Support Training in V1
+  <!-- - [ ] fp8 support -->
+  <!-- - [ ] faster load model and save model support -->
 
 ## 🤝 Contributing
 
-We welcome all contributions. Please run `bash format.sh` before submitting a pull request.
-
-## 🔧 Testing
-Run `pytest` to verify the data preprocessing, checkpoint saving, and sequence parallel pipelines. We recommend adding corresponding test cases in the `test` folder to support your contribution.
+We welcome all contributions. Please check out our guide [here](https://hao-ai-lab.github.io/FastVideo/developer_guide/overview.html)
 
 ## Acknowledgement
-We learned and reused code from the following projects: [PCM](https://github.com/G-U-N/Phased-Consistency-Model), [diffusers](https://github.com/huggingface/diffusers), [OpenSoraPlan](https://github.com/PKU-YuanGroup/Open-Sora-Plan), and [xDiT](https://github.com/xdit-project/xDiT).
+We learned and reused code from the following projects:
+- [PCM](https://github.com/G-U-N/Phased-Consistency-Model)
+- [diffusers](https://github.com/huggingface/diffusers)
+- [OpenSoraPlan](https://github.com/PKU-YuanGroup/Open-Sora-Plan)
+- [xDiT](https://github.com/xdit-project/xDiT)
+- [vLLM](https://github.com/vllm-project/vllm)
+- [SGLang](https://github.com/sgl-project/sglang)
 
-We thank MBZUAI and Anyscale for their support throughout this project.
+We thank MBZUAI and [Anyscale](https://www.anyscale.com/) for their support throughout this project.
+
+## Citation
+If you use FastVideo for your research, please cite our paper:
+
+```bibtex
+@misc{zhang2025fastvideogenerationsliding,
+      title={Fast Video Generation with Sliding Tile Attention},
+      author={Peiyuan Zhang and Yongqi Chen and Runlong Su and Hangliang Ding and Ion Stoica and Zhenghong Liu and Hao Zhang},
+      year={2025},
+      eprint={2502.04507},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2502.04507},
+}
+@misc{ding2025efficientvditefficientvideodiffusion,
+      title={Efficient-vDiT: Efficient Video Diffusion Transformers With Attention Tile},
+      author={Hangliang Ding and Dacheng Li and Runlong Su and Peiyuan Zhang and Zhijie Deng and Ion Stoica and Hao Zhang},
+      year={2025},
+      eprint={2502.06155},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2502.06155},
+}
+```

csrc/sliding_tile_attention/MANIFEST.in

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

csrc/sliding_tile_attention/README.md

@@ -0,0 +1,68 @@
+
+
+# Sliding Tile Atteniton Kernel
+
+
+## Installation
+We test our code on Pytorch 2.5.0 and CUDA>=12.4. Currently we only have implementation on H100.
+First, install C++20 for ThunderKittens:
+
+```bash
+sudo apt update
+sudo apt install gcc-11 g++-11
+
+sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+
+sudo apt update
+sudo apt install clang-11
+```
+Install STA:
+```bash
+export CUDA_HOME=/usr/local/cuda-12.4
+export PATH=${CUDA_HOME}/bin:${PATH} 
+export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+git submodule update --init --recursive
+python setup.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. 
+# q, k, v: [batch_size, num_heads, seq_length, head_dim], seq_length = T*H*W + 256
+# a tile is a cube of size (6, 8, 8)
+# window_size in tiles: [(window_t, window_h, window_w), (..)...]. For example, window size (3, 3, 3) means a query can attend to (3x6, 3x8, 3x8) = (18, 24, 24) tokens out of the total 30x48x80 video.
+# text_length: int ranging from 0 to 256
+# If your attention contains text token (Hunyuan)
+out = sliding_tile_attention(q, k, v, window_size, text_length)
+# If your attention does not contain text token (StepVideo)
+out = sliding_tile_attention(q, k, v, window_size, 0, False)
+
+```
+
+
+## Test
+```bash
+python test/test_sta.py
+```
+
+## How Does STA Work?
+We give a demo for 2D STA with window size (6,6) operating on a (10, 10) image. 
+
+
+https://github.com/user-attachments/assets/f3b6dd79-7b43-4b60-a0fa-3d6495ec5747
+
+## Why is STA Fast?
+2D/3D Sliding Window Attention (SWA) creates many mixed blocks in the attention map. Even though mixed blocks have less output value,a mixed block is significantly slower than a dense block due to the GPU-unfriendly masking operation. 
+
+STA removes mixed blocks.
+
+
+<div align="center">
+<img src=../../assets/sliding_tile_attn_map.png width="80%"/>
+</div>
+
+## Acknowledgement
+
+We learned or reuse code from FlexAtteniton, NATEN, and ThunderKittens.

csrc/sliding_tile_attention/config.py

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

csrc/sliding_tile_attention/setup.py

@@ -0,0 +1,76 @@
+import os
+import subprocess
+
+from config import kernels, sources, target
+from setuptools import find_packages, setup
+from torch.utils.cpp_extension import BuildExtension, CUDAExtension
+
+target = target.lower()
+
+# Package metadata
+PACKAGE_NAME = "st_attn"
+VERSION = "0.0.4"
+AUTHOR = "Hao AI Lab"
+DESCRIPTION = "Sliding Tile Atteniton Kernel Used in FastVideo"
+URL = "https://github.com/hao-ai-lab/FastVideo/tree/main/csrc/sliding_tile_attention"
+
+# Set environment variables
+tk_root = os.getenv('THUNDERKITTENS_ROOT', os.path.abspath(os.path.join(os.getcwd(), 'tk/')))
+python_include = subprocess.check_output(['python', '-c',
+                                          "import sysconfig; print(sysconfig.get_path('include'))"]).decode().strip()
+torch_include = subprocess.check_output([
+    'python', '-c',
+    "import torch; from torch.utils.cpp_extension import include_paths; print(' '.join(['-I' + p for p in include_paths()]))"
+]).decode().strip()
+print('st_attn root:', tk_root)
+print('Python include:', python_include)
+print('Torch include directories:', torch_include)
+
+# CUDA flags
+cuda_flags = [
+    '-DNDEBUG', '-Xcompiler=-Wno-psabi', '-Xcompiler=-fno-strict-aliasing', '--expt-extended-lambda',
+    '--expt-relaxed-constexpr', '-forward-unknown-to-host-compiler', '--use_fast_math', '-std=c++20', '-O3',
+    '-Xnvlink=--verbose', '-Xptxas=--verbose', '-Xptxas=--warn-on-spills', f'-I{tk_root}/include',
+    f'-I{tk_root}/prototype', f'-I{python_include}', '-DTORCH_COMPILE'
+] + torch_include.split()
+cpp_flags = ['-std=c++20', '-O3']
+
+if target == 'h100':
+    cuda_flags.append('-DKITTENS_HOPPER')
+    cuda_flags.append('-arch=sm_90a')
+else:
+    raise ValueError(f'Target {target} not supported')
+
+source_files = ['st_attn.cpp']
+for k in kernels:
+    if target not in sources[k]['source_files']:
+        raise KeyError(f'Target {target} not found in source files for kernel {k}')
+    if isinstance(sources[k]['source_files'][target], list):
+        source_files.extend(sources[k]['source_files'][target])
+    else:
+        source_files.append(sources[k]['source_files'][target])
+    cpp_flags.append(f'-DTK_COMPILE_{k.replace(" ", "_").upper()}')
+
+setup(name=PACKAGE_NAME,
+      version=VERSION,
+      author=AUTHOR,
+      description=DESCRIPTION,
+      url=URL,
+      packages=find_packages(),
+      ext_modules=[
+          CUDAExtension('st_attn_cuda',
+                        sources=source_files,
+                        extra_compile_args={
+                            'cxx': cpp_flags,
+                            'nvcc': cuda_flags
+                        },
+                        libraries=['cuda'])
+      ],
+      cmdclass={'build_ext': BuildExtension},
+      classifiers=[
+          "Programming Language :: Python :: 3",
+          "Environment :: GPU :: NVIDIA CUDA :: 12",
+          "License :: OSI Approved :: Apache Software License",
+      ],
+      python_requires='>=3.10',
+      install_requires=["torch>=2.5.0"])

csrc/sliding_tile_attention/st_attn.cpp

@@ -0,0 +1,24 @@
+#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_ATTN
+extern torch::Tensor sta_forward(
+    torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o, int kernel_t_size, int kernel_w_size, int kernel_h_size, int text_length, bool process_text, bool has_text, int kernel_aspect_ratio_flag
+); 
+#endif
+
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.doc() = "Sliding Block Attention Kernels"; // optional module docstring
+
+#ifdef TK_COMPILE_ATTN
+    m.def("sta_fwd",  torch::wrap_pybind_function(sta_forward), "sliding tile attention, assuming tile size is (6,8,8)");
+#endif
+}

csrc/sliding_tile_attention/st_attn/__init__.py

@@ -0,0 +1,46 @@
+import math
+
+import torch
+from st_attn_cuda import sta_fwd
+
+
+def sliding_tile_attention(q_all, k_all, v_all, window_size, text_length, has_text=True, img_latent_shape='30*48*80'):
+    seq_length = q_all.shape[2]
+    img_latent_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"
+        assert q_all.shape[1] == len(window_size), "Number of heads must match the number of window sizes"
+        target_size = math.ceil(seq_length / 384) * 384
+        pad_size = target_size - seq_length
+        if pad_size > 0:
+            q_all = torch.cat([q_all, q_all[:, :, -pad_size:]], dim=2)
+            k_all = torch.cat([k_all, k_all[:, :, -pad_size:]], dim=2)
+            v_all = torch.cat([v_all, v_all[:, :, -pad_size:]], dim=2)
+    else:
+        if img_latent_shape == '36x48x48': # Stepvideo 204x768x68
+            assert q_all.shape[2] == 82944
+        elif img_latent_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")
+
+    kernel_aspect_ratio_flag = img_latent_shape_mapping[img_latent_shape]
+    hidden_states = torch.empty_like(q_all)
+    # This for loop is ugly. but it is actually quite efficient. The sequence dimension alone can already oversubscribe SMs
+    for head_index, (t_kernel, h_kernel, w_kernel) in enumerate(window_size):
+        for batch in range(q_all.shape[0]):
+            q_head, k_head, v_head, o_head = (q_all[batch:batch + 1, head_index:head_index + 1],
+                                              k_all[batch:batch + 1,
+                                                    head_index:head_index + 1], v_all[batch:batch + 1,
+                                                                                      head_index:head_index + 1],
+                                              hidden_states[batch:batch + 1, head_index:head_index + 1])
+
+            _ = sta_fwd(q_head, k_head, v_head, o_head, t_kernel, h_kernel, w_kernel, text_length, False, has_text, kernel_aspect_ratio_flag)
+    if has_text:
+        _ = sta_fwd(q_all, k_all, v_all, hidden_states, 3, 3, 3, text_length, True, True, kernel_aspect_ratio_flag)
+    return hidden_states[:, :, :seq_length]

csrc/sliding_tile_attention/st_attn/st_attn_h100.cu

@@ -0,0 +1,831 @@
+// # Define TORCH_COMPILE macro
+
+#include "kittens.cuh"
+#include <cooperative_groups.h>
+#include <iostream>
+#include <stdio.h>
+
+#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); 
+constexpr int NUM_WARPGROUPS      = (CONSUMER_WARPGROUPS+PRODUCER_WARPGROUPS); 
+constexpr int NUM_WORKERS         = (NUM_WARPGROUPS*kittens::WARPGROUP_WARPS); 
+
+using namespace kittens;
+namespace cg = cooperative_groups;
+
+template<int D> struct fwd_attend_ker_tile_dims {};
+template<> struct fwd_attend_ker_tile_dims<64> {
+    constexpr static int tile_width = (64);
+    constexpr static int qo_height  = (4*16);
+    constexpr static int kv_height  = (8*16);
+    constexpr static int stages     = (4); 
+};
+template<> struct fwd_attend_ker_tile_dims<128> {
+    constexpr static int tile_width = (128);
+    constexpr static int qo_height  = (4*16);
+    constexpr static int kv_height  = (8*16);
+    constexpr static int stages     = (2); 
+};
+
+template<int D> struct fwd_globals {
+    using q_tile    =         st_bf<fwd_attend_ker_tile_dims<D>::qo_height, fwd_attend_ker_tile_dims<D>::tile_width>;
+    using k_tile    =         st_bf<fwd_attend_ker_tile_dims<D>::kv_height, fwd_attend_ker_tile_dims<D>::tile_width>;
+    using v_tile    =         st_bf<fwd_attend_ker_tile_dims<D>::kv_height, fwd_attend_ker_tile_dims<D>::tile_width>;
+    using l_col_vec = col_vec<st_fl<fwd_attend_ker_tile_dims<D>::qo_height, fwd_attend_ker_tile_dims<D>::tile_width>>;
+    using o_tile    =         st_bf<fwd_attend_ker_tile_dims<D>::qo_height, fwd_attend_ker_tile_dims<D>::tile_width>;
+
+    using q_gl = gl<bf16,  -1, -1, -1, -1, q_tile>;
+    using k_gl = gl<bf16,  -1, -1, -1, -1, k_tile>;
+    using v_gl = gl<bf16,  -1, -1, -1, -1, v_tile>;
+    using l_gl = gl<float, -1, -1, -1, -1, l_col_vec>;
+    using o_gl = gl<bf16,  -1, -1, -1, -1, o_tile>;
+
+    q_gl q;
+    k_gl k;
+    v_gl v;
+    l_gl l;
+    o_gl o;
+
+    const int N; 
+    const int text_L;
+    const int hr;
+};
+
+
+template<int D, bool is_causal, bool text_q, bool text_kv, int DT, int DH, int DW, int CT, int CH, int CW>
+__global__  __launch_bounds__((NUM_WORKERS)*kittens::WARP_THREADS, 1)
+void fwd_attend_ker(const __grid_constant__ fwd_globals<D> g) {
+    extern __shared__ int __shm[]; 
+    tma_swizzle_allocator al((int*)&__shm[0]);
+    int warpid = kittens::warpid(), warpgroupid = warpid/kittens::WARPGROUP_WARPS;
+
+    using K = fwd_attend_ker_tile_dims<D>;
+
+    using q_tile    =         st_bf<K::qo_height, K::tile_width>;
+    using k_tile    =         st_bf<K::kv_height, K::tile_width>;
+    using v_tile    =         st_bf<K::kv_height, K::tile_width>;
+    using l_col_vec = col_vec<st_fl<K::qo_height, K::tile_width>>;
+    using o_tile    =         st_bf<K::qo_height, K::tile_width>;
+    
+    q_tile    (&q_smem)[CONSUMER_WARPGROUPS] = al.allocate<q_tile, CONSUMER_WARPGROUPS>();
+    k_tile    (&k_smem)[K::stages]           = al.allocate<k_tile, K::stages          >();
+    v_tile    (&v_smem)[K::stages]           = al.allocate<v_tile, K::stages          >();
+    l_col_vec (&l_smem)[CONSUMER_WARPGROUPS] = al.allocate<l_col_vec, CONSUMER_WARPGROUPS>();
+    auto      (*o_smem)                      = reinterpret_cast<o_tile(*)>(q_smem);
+    int img_kv_blocks;
+    int kv_blocks   = g.N / (K::kv_height);
+    if constexpr (text_kv) {
+        img_kv_blocks = kv_blocks - 3;
+    } else {
+        img_kv_blocks = kv_blocks;
+    }
+    int kv_head_idx = blockIdx.y / g.hr;
+    int seq_idx;
+    if constexpr (text_q) {
+        seq_idx = CT * CH * CW * 6.0 + blockIdx.x * CONSUMER_WARPGROUPS;
+    } else {
+        seq_idx = blockIdx.x * CONSUMER_WARPGROUPS; 
+    }
+    __shared__ kittens::semaphore qsmem_semaphore, k_smem_arrived[K::stages], v_smem_arrived[K::stages], compute_done[K::stages];
+    if (threadIdx.x == 0) { 
+        init_semaphore(qsmem_semaphore, 0, 1); 
+        for(int j = 0; j < K::stages; j++) {
+            init_semaphore(k_smem_arrived[j], 0, 1); 
+            init_semaphore(v_smem_arrived[j], 0, 1); 
+            init_semaphore(compute_done[j], CONSUMER_WARPGROUPS, 0); 
+        }
+
+        tma::expect_bytes(qsmem_semaphore, sizeof(q_smem));
+
+        for (int wg = 0; wg < CONSUMER_WARPGROUPS; wg++) {
+            coord<q_tile> q_tile_idx = {blockIdx.z, blockIdx.y, (seq_idx) + wg, 0};
+            tma::load_async(q_smem[wg], g.q, q_tile_idx, qsmem_semaphore);
+        }
+
+        if constexpr (text_q){
+            for (int j = 0; j < K::stages - 1; j++) {
+                coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, j, 0};
+                tma::expect_bytes(k_smem_arrived[j], sizeof(k_tile));
+                tma::load_async(k_smem[j], g.k, kv_tile_idx, k_smem_arrived[j]);
+                tma::expect_bytes(v_smem_arrived[j], sizeof(v_tile));
+                tma::load_async(v_smem[j], g.v, kv_tile_idx, v_smem_arrived[j]);
+            }
+        } else {
+            int qt = seq_idx / 6 / (CH * CW);
+            int qh = (seq_idx / 6) % (CH * CW) / CW;
+            int qw = (seq_idx / 6) % CW;
+            qt = CLAMP(qt, DT, CT-DT-1);
+            qh = CLAMP(qh, DH, CH-DH-1);
+            qw = CLAMP(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);
+                if (mask){
+                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, j, 0};
+                    tma::expect_bytes(k_smem_arrived[count], sizeof(k_tile));
+                    tma::load_async(k_smem[count], g.k, kv_tile_idx, k_smem_arrived[count]);
+                    tma::expect_bytes(v_smem_arrived[count], sizeof(v_tile));
+                    tma::load_async(v_smem[count], g.v, kv_tile_idx, v_smem_arrived[count]);
+                    count += 1;
+                }
+                j += 1;
+            }
+        }
+    }
+    __syncthreads(); 
+
+    int pipe_idx = K::stages - 1; 
+    
+    if(warpgroupid == NUM_WARPGROUPS-1) {
+        warpgroup::decrease_registers<32>();      
+        
+        int kv_iters; 
+        if constexpr (is_causal) {
+            kv_iters = (seq_idx * (K::qo_height/kittens::TILE_ROW_DIM<bf16>)) - 1 + (CONSUMER_WARPGROUPS * (K::qo_height/kittens::TILE_ROW_DIM<bf16>)); 
+            kv_iters = ((kv_iters / (K::kv_height/kittens::TILE_ROW_DIM<bf16>)) == 0) ? (0) : ((kv_iters / (K::kv_height/kittens::TILE_ROW_DIM<bf16>)) - 1);
+        }
+        else { kv_iters = kv_blocks-2;}
+
+        if(warpid == NUM_WORKERS-4) {
+            if constexpr (text_q){
+                for (auto kv_idx = pipe_idx - 1; kv_idx <= kv_iters; kv_idx++) {
+                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, kv_idx + 1, 0};
+                    tma::expect_bytes(k_smem_arrived[(kv_idx+1)%K::stages], sizeof(k_tile));
+                    tma::load_async(k_smem[(kv_idx+1)%K::stages], g.k, kv_tile_idx, k_smem_arrived[(kv_idx+1)%K::stages]);
+                    tma::expect_bytes(v_smem_arrived[(kv_idx+1)%K::stages], sizeof(v_tile));
+                    tma::load_async(v_smem[(kv_idx+1)%K::stages], g.v, kv_tile_idx, v_smem_arrived[(kv_idx+1)%K::stages]);
+                    kittens::wait(compute_done[(kv_idx)%K::stages], (kv_idx/K::stages)%2);
+                }
+            } else {
+                int qt = seq_idx / 6 / (CH * CW);
+                int qh = (seq_idx / 6) % (CH * CW) / CW;
+                int qw = (seq_idx / 6) % CW;
+                qt = CLAMP(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);
+                int count = 0;
+                for (int kt = k_t_min; kt <= k_t_max; kt++) {
+                    for (int kh = k_h_min; kh <= k_h_max; kh++) {
+                        for (int kw = k_w_min; kw <= k_w_max; kw++) {
+                            for (int j = 0; j <= 2; j++){
+                                if (count >= K::stages - 1) {
+                                    int index = ((kt * (CH * CW)) + (kh * CW) + kw) * 3 + j;
+                                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, index, 0};
+                                    tma::expect_bytes(k_smem_arrived[count%K::stages], sizeof(k_tile));
+                                    tma::load_async(k_smem[count%K::stages], g.k, kv_tile_idx, k_smem_arrived[count%K::stages]);
+                                    tma::expect_bytes(v_smem_arrived[count%K::stages], sizeof(v_tile));
+                                    tma::load_async(v_smem[count%K::stages], g.v, kv_tile_idx, v_smem_arrived[count%K::stages]);
+                                    kittens::wait(compute_done[(count - 1)%K::stages], ((count - 1)/K::stages)%2);
+                                    count += 1;
+                                } else {
+                                    count += 1;
+                                }
+                            }
+                        }
+                    }
+                }
+                // for text 
+                for (int index = img_kv_blocks; index < kv_blocks; index++) {
+                    coord<k_tile> kv_tile_idx = {blockIdx.z, kv_head_idx, index, 0};
+                    tma::expect_bytes(k_smem_arrived[count%K::stages], sizeof(k_tile));
+                    tma::load_async(k_smem[count%K::stages], g.k, kv_tile_idx, k_smem_arrived[count%K::stages]);
+                    tma::expect_bytes(v_smem_arrived[count%K::stages], sizeof(v_tile));
+                    tma::load_async(v_smem[count%K::stages], g.v, kv_tile_idx, v_smem_arrived[count%K::stages]);
+                    kittens::wait(compute_done[(count - 1)%K::stages], ((count - 1)/K::stages)%2);
+                    count += 1;
+                }
+            }
+
+
+        }
+    }
+    else {
+        warpgroup::increase_registers<160>();
+
+        rt_fl<16, K::kv_height>  att_block;
+        rt_bf<16, K::kv_height>  att_block_mma;
+        rt_fl<16, K::tile_width> o_reg;
+        
+        col_vec<rt_fl<16, K::kv_height>> max_vec, norm_vec, max_vec_last_scaled, max_vec_scaled;
+        
+        neg_infty(max_vec);
+        zero(norm_vec);
+        zero(o_reg);
+
+        int kv_iters; 
+        if constexpr (is_causal) {
+            kv_iters = (seq_idx * 4) - 1 + (CONSUMER_WARPGROUPS * 4);
+            kv_iters = (kv_iters/8);
+        }
+        else if constexpr (text_q){ 
+            // the last three kv blocks are for text, we process them separately
+            kv_iters = img_kv_blocks - 1;
+        } else {
+            kv_iters = CLAMP(DT*2+1, 1, CT) * CLAMP(DH*2+1, 1, CH) * CLAMP(DW*2+1, 1, CW) * 3 - 1 ; 
+        }
+
+        kittens::wait(qsmem_semaphore, 0);
+        for (auto kv_idx = 0; kv_idx <= kv_iters; kv_idx++) {
+
+            kittens::wait(k_smem_arrived[(kv_idx)%K::stages], (kv_idx/K::stages)%2);
+            warpgroup::mm_ABt(att_block, q_smem[warpgroupid], k_smem[(kv_idx)%K::stages]);
+            
+            copy(max_vec_last_scaled, max_vec);
+            if constexpr (D == 64) { mul(max_vec_last_scaled, max_vec_last_scaled, 1.44269504089f*0.125f); }
+            else                   { mul(max_vec_last_scaled, max_vec_last_scaled, 1.44269504089f*0.08838834764f); }
+            
+            warpgroup::mma_async_wait();
+
+            row_max(max_vec, att_block, max_vec);
+            
+            if constexpr (D == 64) { 
+                mul(att_block, att_block,    1.44269504089f*0.125f); 
+                mul(max_vec_scaled, max_vec, 1.44269504089f*0.125f);
+            }
+            else                   { 
+                mul(att_block, att_block,    1.44269504089f*0.08838834764f); 
+                mul(max_vec_scaled, max_vec, 1.44269504089f*0.08838834764f);
+            }
+
+            sub_row(att_block, att_block, max_vec_scaled);
+            exp2(att_block, att_block);
+            sub(max_vec_last_scaled, max_vec_last_scaled, max_vec_scaled);
+            exp2(max_vec_last_scaled,       max_vec_last_scaled);
+            mul(norm_vec,            norm_vec,     max_vec_last_scaled);
+            row_sum(norm_vec,  att_block, norm_vec);
+            add(att_block, att_block, 0.f);
+            copy(att_block_mma, att_block); 
+            mul_row(o_reg, o_reg, max_vec_last_scaled); 
+
+            kittens::wait(v_smem_arrived[(kv_idx)%K::stages], (kv_idx/K::stages)%2); 
+
+            warpgroup::mma_AB(o_reg, att_block_mma, v_smem[(kv_idx)%K::stages]);
+            warpgroup::mma_async_wait();
+
+            if(warpgroup::laneid() == 0) arrive(compute_done[(kv_idx)%K::stages], 1);
+        }
+        // the last three kv blocks are for text, we process them separately
+        if constexpr(text_kv) {
+            for (auto kv_idx = kv_iters + 1; kv_idx <= kv_iters + 3; kv_idx++) {
+
+                kittens::wait(k_smem_arrived[(kv_idx)%K::stages], (kv_idx/K::stages)%2);
+                warpgroup::mm_ABt(att_block, q_smem[warpgroupid], k_smem[(kv_idx)%K::stages]);
+                
+                copy(max_vec_last_scaled, max_vec);
+                if constexpr (D == 64) { mul(max_vec_last_scaled, max_vec_last_scaled, 1.44269504089f*0.125f); }
+                else                   { mul(max_vec_last_scaled, max_vec_last_scaled, 1.44269504089f*0.08838834764f); }
+                
+                warpgroup::mma_async_wait();
+                // apply non-pad mask
+                int offset = g.text_L - (kv_idx - (kv_iters + 1)) * K::kv_height;
+                // printf("k_idx_start: %d, k_idx_end: %d, text_end: %d, offset: %d\n", k_idx_start, k_idx_end, text_end, offset);
+                right_fill(att_block, att_block, offset, base_types::constants<float>::neg_infty());
+
+
+                row_max(max_vec, att_block, max_vec);
+                
+                if constexpr (D == 64) { 
+                    mul(att_block, att_block,    1.44269504089f*0.125f); 
+                    mul(max_vec_scaled, max_vec, 1.44269504089f*0.125f);
+                }
+                else                   { 
+                    mul(att_block, att_block,    1.44269504089f*0.08838834764f); 
+                    mul(max_vec_scaled, max_vec, 1.44269504089f*0.08838834764f);
+                }
+
+                sub_row(att_block, att_block, max_vec_scaled);
+                exp2(att_block, att_block);
+                sub(max_vec_last_scaled, max_vec_last_scaled, max_vec_scaled);
+                exp2(max_vec_last_scaled,       max_vec_last_scaled);
+                mul(norm_vec,            norm_vec,     max_vec_last_scaled);
+                row_sum(norm_vec,  att_block, norm_vec);
+                add(att_block, att_block, 0.f);
+                copy(att_block_mma, att_block); 
+                mul_row(o_reg, o_reg, max_vec_last_scaled); 
+
+                kittens::wait(v_smem_arrived[(kv_idx)%K::stages], (kv_idx/K::stages)%2); 
+
+                warpgroup::mma_AB(o_reg, att_block_mma, v_smem[(kv_idx)%K::stages]);
+                warpgroup::mma_async_wait();
+
+                if(warpgroup::laneid() == 0) arrive(compute_done[(kv_idx)%K::stages], 1);
+            }
+        }
+
+        div_row(o_reg, o_reg, norm_vec);
+        warpgroup::store(o_smem[warpgroupid], o_reg); 
+        warpgroup::sync(warpgroupid+4);
+
+        if (warpid % 4 == 0) {
+            coord<o_tile> o_tile_idx = {blockIdx.z, blockIdx.y, (seq_idx) + warpgroupid, 0};
+            tma::store_async(g.o, o_smem[warpgroupid], o_tile_idx);
+        }
+
+        mul(max_vec_scaled,   max_vec_scaled, 0.69314718056f);
+        log(norm_vec, norm_vec);
+        add(norm_vec, norm_vec, max_vec_scaled);
+
+        if constexpr (D == 64) { mul(norm_vec, norm_vec, -8.0f); }
+        else                   { mul(norm_vec, norm_vec, -11.313708499f); }
+    
+        warpgroup::store(l_smem[warpgroupid], norm_vec);
+        warpgroup::sync(warpgroupid+4);
+
+        if (warpid % 4 == 0) {
+            coord<l_col_vec> tile_idx = {blockIdx.z, blockIdx.y, 0, (seq_idx) + warpgroupid};
+            tma::store_async(g.l, l_smem[warpgroupid], tile_idx);
+        }
+        tma::store_async_wait();
+    }
+}
+
+
+
+#include "pyutils/torch_helpers.cuh"
+#include <ATen/cuda/CUDAContext.h>
+#include <iostream>
+
+torch::Tensor 
+sta_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor o, int kernel_t_size, int kernel_h_size, int kernel_w_size, int text_length, bool process_text, bool has_text, int kernel_aspect_ratio_flag)
+{
+    CHECK_INPUT(q);
+    CHECK_INPUT(k);
+    CHECK_INPUT(v);
+
+    auto batch    = q.size(0);
+    auto seq_len  = q.size(2); 
+    auto head_dim = q.size(3);  
+    auto qo_heads = q.size(1);
+    auto kv_heads = k.size(1);
+
+    // check to see that these dimensions match for all inputs
+    TORCH_CHECK(q.size(0) == batch, "Q batch dimension - idx 0 - must match for all inputs");
+    TORCH_CHECK(k.size(0) == batch, "K batch dimension - idx 0 - must match for all inputs");
+    TORCH_CHECK(v.size(0) == batch, "V batch dimension - idx 0 - must match for all inputs");
+
+    TORCH_CHECK(q.size(2) == seq_len, "Q sequence length dimension - idx 2 - must match for all inputs");
+    TORCH_CHECK(k.size(2) == seq_len, "K sequence length dimension - idx 2 - must match for all inputs");
+    TORCH_CHECK(v.size(2) == seq_len, "V sequence length dimension - idx 2 - must match for all inputs");
+
+    TORCH_CHECK(q.size(3) == head_dim, "Q head dimension - idx 3 - must match for all non-vector inputs");
+    TORCH_CHECK(k.size(3) == head_dim, "K head dimension - idx 3 - must match for all non-vector inputs");
+    TORCH_CHECK(v.size(3) == head_dim, "V head dimension - idx 3 - must match for all non-vector inputs");
+
+    TORCH_CHECK(qo_heads >= kv_heads, "QO heads must be greater than or equal to KV heads");
+    TORCH_CHECK(qo_heads % kv_heads == 0, "QO heads must be divisible by KV heads");
+    TORCH_CHECK(q.size(1) == qo_heads, "QO head dimension - idx 1 - must match for all inputs");
+    TORCH_CHECK(k.size(1) == kv_heads, "KV head dimension - idx 1 - must match for all inputs");
+    TORCH_CHECK(v.size(1) == kv_heads, "KV head dimension - idx 1 - must match for all inputs");  
+
+    auto hr = qo_heads / kv_heads;
+
+    c10::BFloat16* q_ptr = q.data_ptr<c10::BFloat16>();
+    c10::BFloat16* k_ptr = k.data_ptr<c10::BFloat16>();
+    c10::BFloat16* v_ptr = v.data_ptr<c10::BFloat16>();
+
+    bf16*  d_q = reinterpret_cast<bf16*>(q_ptr);
+    bf16*  d_k = reinterpret_cast<bf16*>(k_ptr);
+    bf16*  d_v = reinterpret_cast<bf16*>(v_ptr);
+    
+
+    
+    torch::Tensor l_vec = torch::empty({static_cast<const uint>(batch), 
+                                        static_cast<const uint>(qo_heads), 
+                                        static_cast<const uint>(seq_len), 
+                                        static_cast<const uint>(1)}, 
+                                        torch::TensorOptions().dtype(torch::kFloat).device(q.device()).memory_format(at::MemoryFormat::Contiguous));
+        
+
+    bf16*  o_ptr = reinterpret_cast<bf16*>(o.data_ptr<c10::BFloat16>());
+    bf16*  d_o   = reinterpret_cast<bf16*>(o_ptr);
+
+    float* l_ptr = reinterpret_cast<float*>(l_vec.data_ptr<float>());
+    float* d_l   = reinterpret_cast<float*>(l_ptr);
+
+    cudaDeviceSynchronize();
+    auto stream = at::cuda::getCurrentCUDAStream().stream(); 
+
+
+    if (head_dim == 128) {
+        using q_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using k_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::kv_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using v_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::kv_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using l_col_vec = col_vec<st_fl<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>>;
+        using o_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+
+        using q_global = gl<bf16,  -1, -1, -1, -1, q_tile>;
+        using k_global = gl<bf16,  -1, -1, -1, -1, k_tile>;
+        using v_global = gl<bf16,  -1, -1, -1, -1, v_tile>;
+        using l_global = gl<float, -1, -1, -1, -1, l_col_vec>;
+        using o_global = gl<bf16,  -1, -1, -1, -1, o_tile>;
+
+        using globals      = fwd_globals<128>;
+
+        q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
+        k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 128U};
+        v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 128U};
+        l_global lg_arg{d_l, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,   static_cast<unsigned int>(seq_len)};
+        o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
+
+        globals g{qg_arg, kg_arg, vg_arg, lg_arg, og_arg, static_cast<int>(seq_len),  static_cast<int>(text_length), static_cast<int>(hr)};
+
+        auto mem_size = kittens::MAX_SHARED_MEMORY;
+        auto threads  = NUM_WORKERS * kittens::WARP_THREADS;
+        if (has_text) {
+            // TORCH_CHECK(seq_len % (CONSUMER_WARPGROUPS*kittens::TILE_DIM*4) == 0, "sequence length must be divisible by 192");
+            dim3 grid_image(seq_len/(CONSUMER_WARPGROUPS*kittens::TILE_ROW_DIM<bf16>*4)-2, qo_heads, batch);
+            dim3 grid_text(2, qo_heads, batch);
+            if (!process_text) {
+                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3) {
+
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 1, 1, 1, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 1, 1, 1, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                }  else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 5) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 1, 1, 2, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true,1, 1, 2, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 5 && kernel_h_size == 3 && kernel_w_size == 3) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 1, 1, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 1, 1, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                }else if (kernel_t_size ==3 && kernel_h_size == 5 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 1, 2, 2, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 1, 2, 2, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size ==5 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 3, 0, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 3, 0, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size ==5 && kernel_h_size == 3 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 1, 2, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 1, 2, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 5 && kernel_h_size == 5 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 2, 2, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 2, 2, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 5 && kernel_h_size == 5 && kernel_w_size == 7){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 2, 3, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 2, 3, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 5 && kernel_h_size == 6 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 3, 5, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 3, 5, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 5 && kernel_h_size == 1 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 0, 0, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 2, 0, 0, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 0, 3, 5, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true, 0, 3, 5, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 5 && kernel_h_size == 1 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, true, 2, 0, 5, 5, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, true,2, 0, 5, 5, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else {
+                    // print error
+                    std::cout << "Invalid kernel size" << std::endl;
+                    //print kernel size
+                    std::cout << "Kernel size: " << kernel_t_size << " " << kernel_h_size << " " << kernel_w_size << std::endl;
+                }
+            } else {
+                cudaFuncSetAttribute(
+                    fwd_attend_ker<128, false, true, true, 1, 1, 1, 5, 6, 10>,
+                    cudaFuncAttributeMaxDynamicSharedMemorySize,
+                    mem_size
+                );
+                fwd_attend_ker<128, false, true, true, 1, 1, 1, 5, 6, 10><<<grid_text, (32*NUM_WORKERS), mem_size, stream>>>(g);
+            }
+
+        } else {
+            dim3 grid_image(seq_len/(CONSUMER_WARPGROUPS*kittens::TILE_ROW_DIM<bf16>*4), qo_heads, batch);
+            if (kernel_aspect_ratio_flag == 2){
+                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 1, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 1, 1, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                }  else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 6) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,1, 1, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 6 && kernel_h_size == 3 && kernel_w_size == 3) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 1, 1, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 1, 1, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size ==3 && kernel_h_size == 6 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                }else if (kernel_t_size ==3 && kernel_h_size == 6 && kernel_w_size == 3){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 1, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 1, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size ==6 && kernel_h_size == 3 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 1, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 1, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 6 && kernel_h_size == 6 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 3, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 3, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 6 && kernel_h_size == 1 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 0, 0, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 0, 0, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 6 && kernel_h_size == 1 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 0, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 0, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                }  else if (kernel_t_size == 6 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 3, 0, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 3, 0, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 3, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                }  else if (kernel_t_size == 1 && kernel_h_size == 1 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 0, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 0, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 0, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 3, 0, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 6 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 3, 0, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 3, 0, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 6 && kernel_h_size == 1 && kernel_w_size == 6){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 3, 0, 3, 6, 6, 6>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 3, 0, 3, 6, 6, 6><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else {
+                    // print error
+                    std::cout << "Invalid kernel size" << std::endl;
+                    //print kernel size
+                    std::cout << "Kernel size: " << kernel_t_size << " " << kernel_h_size << " " << kernel_w_size << std::endl;
+                }
+            }
+            else if (kernel_aspect_ratio_flag == 3) {
+                if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 3) {
+
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 1, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 1, 1, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 5) {
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 2, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,1, 1, 2, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 2, 2, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 2, 2, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 0, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 0, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+
+                } else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 7){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 2, 3, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 2, 3, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 5 && kernel_w_size == 9){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 2, 4, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 2, 4, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 6 && kernel_w_size == 3){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 3, 1, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 3, 1, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 1 && kernel_w_size == 1){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 0, 0, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 1, 0, 0, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 3, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 2, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 2, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 7){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 3, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 3, 3, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 7){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 2, 3, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 2, 3, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 5 && kernel_w_size == 9){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 2, 4, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false, 0, 2, 4, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 1 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 0, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,1, 0, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 3 && kernel_h_size == 3 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 1, 1, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,1, 1, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 3 && kernel_w_size == 10){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 1, 5, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,0, 1, 5, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else if (kernel_t_size == 1 && kernel_h_size == 6 && kernel_w_size == 5){
+                    cudaFuncSetAttribute(
+                        fwd_attend_ker<128, false, false, false, 0, 3, 2, 3, 6, 10>,
+                        cudaFuncAttributeMaxDynamicSharedMemorySize,
+                        mem_size
+                    );
+                    fwd_attend_ker<128, false, false, false,0, 3, 2, 3, 6, 10><<<grid_image, (32*NUM_WORKERS), mem_size, stream>>>(g);
+                } else {
+                    // print error
+                    std::cout << "Invalid kernel size" << std::endl;
+                    //print kernel size
+                    std::cout << "Kernel size: " << kernel_t_size << " " << kernel_h_size << " " << kernel_w_size << std::endl;
+                }
+            }
+
+            else {
+                std::cout << "Unsupported kernel_aspect_ratio_flag: " << kernel_aspect_ratio_flag << std::endl;
+            }
+
+        }
+        CHECK_CUDA_ERROR(cudaGetLastError());
+        cudaStreamSynchronize(stream);
+    }
+
+    return o;
+    cudaDeviceSynchronize();
+}

csrc/sliding_tile_attention/test/bench.py

@@ -0,0 +1,151 @@
+import os
+from collections import defaultdict
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+from st_attn import sliding_tile_attention
+
+
+def flops(batch, seqlen, nheads, headdim, causal, mode="fwd"):
+    assert mode in ["fwd", "bwd", "fwd_bwd"]
+    f = 4 * batch * seqlen**2 * nheads * headdim // (2 if causal else 1)
+    return f if mode == "fwd" else (2.5 * f if mode == "bwd" else 3.5 * f)
+
+
+def efficiency(flop, time):
+    flop = flop / 1e12
+    time = time / 1e6
+    return flop / time
+
+
+def benchmark_attention(configurations):
+    results = {'fwd': defaultdict(list), 'bwd': defaultdict(list)}
+
+    for B, H, N, D, causal in configurations:
+        print("=" * 60)
+        print(f"Timing forward and backward pass for B={B}, H={H}, N={N}, D={D}, causal={causal}")
+
+        q = torch.randn(B, H, N, D, dtype=torch.bfloat16, device='cuda', requires_grad=False).contiguous()
+        k = torch.randn(B, H, N, D, dtype=torch.bfloat16, device='cuda', requires_grad=False).contiguous()
+        v = torch.randn(B, H, N, D, dtype=torch.bfloat16, device='cuda', requires_grad=False).contiguous()
+
+        grad_output = torch.randn_like(q, requires_grad=False).contiguous()
+
+        qg = torch.zeros_like(q, requires_grad=False, dtype=torch.float).contiguous()
+        kg = torch.zeros_like(k, requires_grad=False, dtype=torch.float).contiguous()
+        vg = torch.zeros_like(v, requires_grad=False, dtype=torch.float).contiguous()
+
+        # 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)]
+
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+        # Warmup for forward pass
+        for _ in range(10):
+            o = sliding_tile_attention(q, k, v, [[6, 6, 6]] * 24, 0, False)
+
+        # 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)
+        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("-" * 60)
+
+        # torch.cuda.empty_cache()
+        # torch.cuda.synchronize()
+
+        # # Prepare for timing backward pass
+        # start_events_bwd = [torch.cuda.Event(enable_timing=True) for _ in range(10)]
+        # end_events_bwd = [torch.cuda.Event(enable_timing=True) for _ in range(10)]
+
+        # # Warmup for backward pass
+        # for _ in range(10):
+        #     qg, kg, vg = tk.mha_backward(q, k, v, o, l_vec, grad_output, causal)
+
+        # # Time the backward pass
+        # for i in range(10):
+        #     start_events_bwd[i].record()
+        #     qg, kg, vg = tk.mha_backward(q, k, v, o, l_vec, grad_output, causal)
+        #     end_events_bwd[i].record()
+
+        # torch.cuda.synchronize()
+        # times_bwd = [s.elapsed_time(e) for s, e in zip(start_events_bwd, end_events_bwd)]
+        # time_us_bwd = np.mean(times_bwd) * 1000
+
+        # tflops_bwd = efficiency(flops(B, N, H, D, causal, 'bwd'), time_us_bwd)
+        # 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)
+
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+    return results
+
+
+def plot_results(results):
+    os.makedirs('benchmark_results', exist_ok=True)
+    for mode in ['fwd', 'bwd']:
+        for (D, causal), values in results[mode].items():
+            seq_lens = [x[0] for x in values]
+            tflops = [x[1] for x in values]
+
+            plt.figure(figsize=(10, 6))
+            bars = plt.bar(range(len(seq_lens)), tflops, tick_label=seq_lens)
+            plt.xlabel('Sequence Length')
+            plt.ylabel('TFLOPS')
+            plt.title(f'{mode.upper()} Pass - Head Dim: {D}, Causal: {causal}')
+            plt.grid(True)
+
+            # Adding the numerical y value on top of each bar
+            for bar in bars:
+                yval = bar.get_height()
+                plt.text(bar.get_x() + bar.get_width() / 2, yval, round(yval, 2), ha='center', va='bottom')
+
+            filename = f'benchmark_results/{mode}_D{D}_causal{causal}.png'
+            plt.savefig(filename)
+            plt.close()
+
+
+# Example list of configurations to test
+configurations = [
+    (2, 24, 82944, 128, False),
+    # (16, 16, 768*16,  128, False),
+    # (16, 16, 768*2,  128, False),
+    # (16, 16, 768*4,  128, False),
+    # (16, 16, 768*8,  128, False),
+    # (16, 16, 768*16, 128, False),
+    # (16, 16, 768,    128, True),
+    # (16, 16, 768*2,  128, True),
+    # (16, 16, 768*4,  128, True),
+    # (16, 16, 768*8,  128, True),
+    # (16, 16, 768*16, 128, True),
+    # (16, 32, 768,    64,  False),
+    # (16, 32, 768*2,  64,  False),
+    # (16, 32, 768*4,  64,  False),
+    # (16, 32, 768*8,  64,  False),
+    # (16, 32, 768*16, 64,  False),
+    # (16, 32, 768,    64,  True),
+    # (16, 32, 768*2,  64,  True),
+    # (16, 32, 768*4,  64,  True),
+    # (16, 32, 768*8,  64,  True),
+    # (16, 32, 768*16, 64,  True),
+]
+
+results = benchmark_attention(configurations)
+# plot_results(results)

csrc/sliding_tile_attention/test/flex_sta_ref.py

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

csrc/sliding_tile_attention/test/test_sta.py

@@ -0,0 +1,95 @@
+import torch
+from flex_sta_ref import get_sliding_tile_attention_mask
+from st_attn import sliding_tile_attention
+from torch.nn.attention.flex_attention import flex_attention
+from tqdm import tqdm
+
+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)
+    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)
+    return o
+
+
+def generate_tensor(shape, mean, std, dtype, device):
+    tensor = torch.randn(shape, dtype=dtype, device=device)
+
+    magnitude = torch.linalg.norm(tensor, dim=-1, keepdim=True)
+    scaled_tensor = tensor * (torch.randn(magnitude.shape, dtype=dtype, device=device) * std + mean) / magnitude
+
+    return scaled_tensor.contiguous()
+
+
+def check_correctness(b, h, n, d, causal, mean, std, num_iterations=50, error_mode='all'):
+    results = {
+        'TK vs FLEX': {
+            'sum_diff': 0,
+            'sum_abs': 0,
+            'max_diff': 0
+        },
+    }
+    kernel_size_ls = [(6, 1, 6), (6, 6, 1)]
+    from tqdm import tqdm
+    for kernel_size in tqdm(kernel_size_ls):
+        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')
+            tk_o = h100_fwd_kernel_test(Q, K, V, kernel_size)
+            pt_o = flex_test(Q, K, V, kernel_size)
+
+            diff = pt_o - tk_o
+            abs_diff = torch.abs(diff)
+            results['TK vs FLEX']['sum_diff'] += torch.sum(abs_diff).item()
+            results['TK vs FLEX']['max_diff'] = max(results['TK vs FLEX']['max_diff'], torch.max(abs_diff).item())
+
+            torch.cuda.empty_cache()
+        print("kernel_size", kernel_size)
+        print("max_diff", torch.max(abs_diff).item())
+        print(
+            "avg_diff",
+            torch.sum(abs_diff).item() / (b * h * n * d *
+                                          (1 if error_mode == 'output' else 3 if error_mode == 'backward' else 4)))
+
+    total_elements = b * h * n * d * num_iterations * (1 if error_mode == 'output' else
+                                                       3 if error_mode == 'backward' else 4) * len(kernel_size_ls)
+    for name, data in results.items():
+        avg_diff = data['sum_diff'] / total_elements
+        max_diff = data['max_diff']
+        results[name] = {'avg_diff': avg_diff, 'max_diff': max_diff}
+
+    return results
+
+
+def generate_error_graphs(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
+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.")

demo/gradio_web_demo.py

@@ -23,9 +23,7 @@ def init_args():
     parser.add_argument("--model_path", type=str, default="data/mochi")
     parser.add_argument("--seed", type=int, default=12345)
     parser.add_argument("--transformer_path", type=str, default=None)
-    parser.add_argument("--scheduler_type",
-                        type=str,
-                        default="pcm_linear_quadratic")
+    parser.add_argument("--scheduler_type", type=str, default="pcm_linear_quadratic")
     parser.add_argument("--lora_checkpoint_dir", type=str, default=None)
     parser.add_argument("--shift", type=float, default=8.0)
     parser.add_argument("--num_euler_timesteps", type=int, default=50)
@@ -50,15 +48,11 @@ def load_model(args):
         )
 
     if args.transformer_path:
-        transformer = MochiTransformer3DModel.from_pretrained(
-            args.transformer_path)
+        transformer = MochiTransformer3DModel.from_pretrained(args.transformer_path)
     else:
-        transformer = MochiTransformer3DModel.from_pretrained(
-            args.model_path, subfolder="transformer/")
+        transformer = MochiTransformer3DModel.from_pretrained(args.model_path, subfolder="transformer/")
 
-    pipe = MochiPipeline.from_pretrained(args.model_path,
-                                         transformer=transformer,
-                                         scheduler=scheduler)
+    pipe = MochiPipeline.from_pretrained(args.model_path, transformer=transformer, scheduler=scheduler)
     pipe.enable_vae_tiling()
     # pipe.to(device)
     # if args.cpu_offload:
@@ -137,11 +131,7 @@ with gr.Blocks() as demo:
                     step=32,
                     value=args.height,
                 )
-                width = gr.Slider(label="Width",
-                                  minimum=256,
-                                  maximum=1024,
-                                  step=32,
-                                  value=args.width)
+                width = gr.Slider(label="Width", minimum=256, maximum=1024, step=32, value=args.width)
 
             with gr.Row():
                 num_frames = gr.Slider(
@@ -164,8 +154,7 @@ with gr.Blocks() as demo:
                 )
 
             with gr.Row():
-                use_negative_prompt = gr.Checkbox(label="Use negative prompt",
-                                                  value=False)
+                use_negative_prompt = gr.Checkbox(label="Use negative prompt", value=False)
             negative_prompt = gr.Text(
                 label="Negative prompt",
                 max_lines=1,
@@ -173,11 +162,7 @@ with gr.Blocks() as demo:
                 visible=False,
             )
 
-            seed = gr.Slider(label="Seed",
-                             minimum=0,
-                             maximum=1000000,
-                             step=1,
-                             value=args.seed)
+            seed = gr.Slider(label="Seed", minimum=0, maximum=1000000, step=1, value=args.seed)
             randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
             seed_output = gr.Number(label="Used Seed")
 

demo/qualitative_results.md

@@ -0,0 +1,15 @@
+Fast-Hunyuan comparison with original Hunyuan, achieving an 8X diffusion speed boost with the FastVideo framework.
+
+https://github.com/user-attachments/assets/064ac1d2-11ed-4a0c-955b-4d412a96ef30
+
+Fast-Mochi comparison with original Mochi, achieving an 8X diffusion speed boost with the FastVideo framework.
+
+https://github.com/user-attachments/assets/5fbc4596-56d6-43aa-98e0-da472cf8e26c
+
+Comparison between OpenAI Sora, original Hunyuan and FastHunyuan
+
+https://github.com/user-attachments/assets/d323b712-3f68-42b2-952b-94f6a49c4836
+
+Comparison between original FastHunyuan, LLM-INT8 quantized FastHunyuan and NF4 quantized FastHunyuan
+
+https://github.com/user-attachments/assets/cf89efb5-5f68-4949-a085-f41c1ef26c94

docker/Dockerfile.python3.10

@@ -0,0 +1,48 @@
+FROM nvidia/cuda:12.4.1-devel-ubuntu20.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+WORKDIR /FastVideo
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    wget \
+    git \
+    ca-certificates \
+    openssh-server \
+    && 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
+
+ENV PATH=/opt/conda/bin:$PATH
+
+RUN conda create --name fastvideo-dev python=3.10.0 -y
+
+SHELL ["/bin/bash", "-c"]
+
+# Copy just the pyproject.toml first to leverage Docker cache
+COPY pyproject.toml ./
+
+# 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
+
+COPY . .
+
+RUN conda run -n fastvideo-dev pip install --no-cache-dir -e .[dev]
+
+# Remove authentication headers
+RUN git config --unset-all http.https://github.com/.extraheader || true
+
+# Set up automatic conda environment activation for all shells
+RUN echo 'source /opt/conda/etc/profile.d/conda.sh' >> /root/.bashrc && \
+    echo 'conda activate fastvideo-dev' >> /root/.bashrc && \
+    # Ensure .bashrc is sourced for SSH login shells
+    echo 'if [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
+
+EXPOSE 22

docker/Dockerfile.python3.11

@@ -0,0 +1,48 @@
+FROM nvidia/cuda:12.4.1-devel-ubuntu20.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+WORKDIR /FastVideo
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    wget \
+    git \
+    ca-certificates \
+    openssh-server \
+    && 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
+
+ENV PATH=/opt/conda/bin:$PATH
+
+RUN conda create --name fastvideo-dev python=3.11.11 -y
+
+SHELL ["/bin/bash", "-c"]
+
+# Copy just the pyproject.toml first to leverage Docker cache
+COPY pyproject.toml ./
+
+# 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
+
+COPY . .
+
+RUN conda run -n fastvideo-dev pip install --no-cache-dir -e .[dev]
+
+# Remove authentication headers
+RUN git config --unset-all http.https://github.com/.extraheader || true
+
+# Set up automatic conda environment activation for all shells
+RUN echo 'source /opt/conda/etc/profile.d/conda.sh' >> /root/.bashrc && \
+    echo 'conda activate fastvideo-dev' >> /root/.bashrc && \
+    # Ensure .bashrc is sourced for SSH login shells
+    echo 'if [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
+
+EXPOSE 22

docker/Dockerfile.python3.12

@@ -0,0 +1,48 @@
+FROM nvidia/cuda:12.4.1-devel-ubuntu20.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+WORKDIR /FastVideo
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    wget \
+    git \
+    ca-certificates \
+    openssh-server \
+    && 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
+
+ENV PATH=/opt/conda/bin:$PATH
+
+RUN conda create --name fastvideo-dev python=3.12.9 -y
+
+SHELL ["/bin/bash", "-c"]
+
+# Copy just the pyproject.toml first to leverage Docker cache
+COPY pyproject.toml ./
+
+# Create a dummy README to satisfy the installation
+RUN echo "# Placeholder" > README.md
+
+RUN conda run -n fastvideo-dev pip install --no-cache-dir --upgrade pip && \
+    conda run -n fastvideo-dev pip install --no-cache-dir .[dev] && \
+    conda run -n fastvideo-dev pip install --no-cache-dir flash-attn==2.7.4.post1 --no-build-isolation && \
+    conda clean -afy
+
+COPY . .
+
+RUN conda run -n fastvideo-dev pip install --no-cache-dir -e .[dev]
+
+# Remove authentication headers
+RUN git config --unset-all http.https://github.com/.extraheader || true
+
+# Set up automatic conda environment activation for all shells
+RUN echo 'source /opt/conda/etc/profile.d/conda.sh' >> /root/.bashrc && \
+    echo 'conda activate fastvideo-dev' >> /root/.bashrc && \
+    # Ensure .bashrc is sourced for SSH login shells
+    echo 'if [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile
+
+EXPOSE 22

docs/Makefile

@@ -0,0 +1,25 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line, and also
+# from the environment for the first two.
+SPHINXOPTS    ?=
+SPHINXBUILD   ?= sphinx-build
+SOURCEDIR     = source
+BUILDDIR      = build
+
+# Put it first so that "make" without argument is like "make help".
+help:
+	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+clean:
+	@$(SPHINXBUILD) -M clean "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+	rm -rf "$(SOURCEDIR)/getting_started/examples"
+	rm -rf "$(SOURCEDIR)/inference/examples"

docs/README.md

@@ -0,0 +1,20 @@
+# FastVideo documents
+
+## Build the docs
+
+```bash
+# Install dependencies.
+pip install -r requirements-docs.txt
+
+# Build the docs.
+make clean
+make html
+```
+
+## Open the docs with your browser
+
+```bash
+python -m http.server -d build/html/
+```
+
+Launch your browser and open localhost:8000.

docs/make.bat

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

docs/requirements-docs.txt

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

docs/seed_parameter_behavior.md

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

docs/source/_static/custom.css

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

docs/source/_static/custom.js

@@ -0,0 +1,18 @@
+// Update URL search params when tab is clicked
+  document.addEventListener("DOMContentLoaded", function () {
+    const tabs = document.querySelectorAll(".sd-tab-label");
+
+    function updateURL(tab) {
+      const syncGroup = tab.getAttribute("data-sync-group");
+      const syncId = tab.getAttribute("data-sync-id");
+      if (syncGroup && syncId) {
+          const url = new URL(window.location);
+          url.searchParams.set(syncGroup, syncId);
+          window.history.replaceState(null, "", url);
+      }
+    }
+
+    tabs.forEach(tab => {
+        tab.addEventListener("click", () => updateURL(tab));
+    });
+});

docs/source/_templates/sections/header.html

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

docs/source/api/summary.md

@@ -0,0 +1,19 @@
+# Summary
+
+## Video Generator
+
+```{autodoc2-summary}
+    fastvideo.VideoGenerator
+```
+
+## Initialization Configuration
+
+```{autodoc2-summary}
+    fastvideo.v1.configs.pipelines.PipelineConfig
+```
+
+## Sampling Configuration
+
+```{autodoc2-summary}
+    fastvideo.v1.configs.sample.SamplingParam
+```

docs/source/autodoc2_docstring_parser.py

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

docs/source/conf.py

@@ -0,0 +1,278 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Configuration file for the Sphinx documentation builder.
+#
+# This file only contains a selection of the most common options. For a full
+# list see the documentation:
+# https://www.sphinx-doc.org/en/master/usage/configuration.html
+
+# -- Path setup --------------------------------------------------------------
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute, like shown here.
+
+import datetime
+import logging
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Optional
+
+import requests
+
+logger = logging.getLogger(__name__)
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+print(os.path.abspath(REPO_ROOT))
+sys.path.append(os.path.abspath(REPO_ROOT))
+
+# -- Project information -----------------------------------------------------
+
+project = 'FastVideo'
+copyright = f'{datetime.datetime.now().year}, FastVideo Team'
+author = 'the FastVideo Team'
+
+# -- General configuration ---------------------------------------------------
+
+# Add any Sphinx extension module names here, as strings. They can be
+# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
+# ones.
+extensions = [
+    "sphinx.ext.napoleon",
+    "sphinx.ext.linkcode",
+    "sphinx.ext.intersphinx",
+    "sphinx_copybutton",
+    "autodoc2",
+    "myst_parser",
+    "sphinxarg.ext",
+    "sphinx_design",
+    "sphinx_togglebutton",
+]
+myst_enable_extensions = [
+    "colon_fence",
+    "fieldlist",
+]
+autodoc2_packages = [
+    {
+        "path": "../../fastvideo",
+        "exclude_dirs": ["__pycache__", "third_party"],
+    },
+]
+autodoc2_output_dir = "api"
+autodoc2_render_plugin = "myst"
+autodoc2_hidden_objects = ["dunder", "private", "inherited"]
+autodoc2_docstring_parser_regexes = [
+    (".*", "docs.source.autodoc2_docstring_parser"),
+]
+autodoc2_sort_names = True
+autodoc2_index_template = None
+autodoc2_skip_module_regexes = [
+    "fastvideo.dataset",
+    "fastvideo.distill",
+    "fastvideo.data_preprocess",
+    "fastvideo.models",
+    "fastvideo.sample",
+    "fastvideo.utils",
+    "fastvideo.distill_adv",
+    "fastvideo.train",
+]
+
+# Add any paths that contain templates here, relative to this directory.
+templates_path = ['_templates']
+
+# List of patterns, relative to source directory, that match files and
+# directories to ignore when looking for source files.
+# This pattern also affects html_static_path and html_extra_path.
+exclude_patterns: list[str] = ["**/*.template.md", "**/*.inc.md"]
+
+# Exclude the prompt "$" when copying code
+copybutton_prompt_text = r"\$ "
+copybutton_prompt_is_regexp = True
+
+# -- Options for HTML output -------------------------------------------------
+
+# The theme to use for HTML and HTML Help pages.  See the documentation for
+# a list of builtin themes.
+#
+html_title = project
+html_theme = 'sphinx_book_theme'
+html_logo = '../../assets/logo.jpg'
+#html_favicon = 'assets/logos/vllm-logo-only-light.ico'
+html_theme_options = {
+    'path_to_docs': 'docs/source',
+    'repository_url': 'https://github.com/hao-ai-lab/FastVideo/',
+    'use_repository_button': True,
+    'use_edit_page_button': True,
+    # Prevents the full API being added to the left sidebar of every page.
+    # Reduces build time by 2.5x and reduces build size from ~225MB to ~95MB.
+    'collapse_navbar': True,
+    # Makes API visible in the right sidebar on API reference pages.
+    'show_toc_level': 3,
+}
+# Add any paths that contain custom static files (such as style sheets) here,
+# relative to this directory. They are copied after the builtin static files,
+# so a file named "default.css" will overwrite the builtin "default.css".
+html_static_path = ["_static"]
+html_js_files = ["custom.js"]
+html_css_files = ["custom.css"]
+
+myst_url_schemes = {
+    'http': None,
+    'https': None,
+    'mailto': None,
+    'ftp': None,
+    "gh-issue": {
+        "url":
+        "https://github.com/hao-ai-lab/FastVideo/issues/{{path}}#{{fragment}}",
+        "title": "Issue #{{path}}",
+        "classes": ["github"],
+    },
+    "gh-pr": {
+        "url":
+        "https://github.com/hao-ai-lab/FastVideo/pull/{{path}}#{{fragment}}",
+        "title": "Pull Request #{{path}}",
+        "classes": ["github"],
+    },
+    "gh-dir": {
+        "url": "https://github.com/hao-ai-lab/FastVideo/tree/main/{{path}}",
+        "title": "{{path}}",
+        "classes": ["github"],
+    },
+    "gh-file": {
+        "url": "https://github.com/hao-ai-lab/FastVideo/blob/main/{{path}}",
+        "title": "{{path}}",
+        "classes": ["github"],
+    },
+}
+
+# see https://docs.readthedocs.io/en/stable/reference/environment-variables.html # noqa
+READTHEDOCS_VERSION_TYPE = os.environ.get('READTHEDOCS_VERSION_TYPE')
+if READTHEDOCS_VERSION_TYPE == "tag":
+    # remove the warning banner if the version is a tagged release
+    header_file = os.path.join(os.path.dirname(__file__),
+                               "_templates/sections/header.html")
+    # The file might be removed already if the build is triggered multiple times
+    # (readthedocs build both HTML and PDF versions separately)
+    if os.path.exists(header_file):
+        os.remove(header_file)
+
+
+# Generate additional rst documentation here.
+def setup(app):
+    from docs.source.generate_examples import generate_examples
+    generate_examples()
+
+
+_cached_base: str = ""
+_cached_branch: str = ""
+
+
+def get_repo_base_and_branch(
+        pr_number: str) -> tuple[Optional[str], Optional[str]]:
+    global _cached_base, _cached_branch
+    if _cached_base and _cached_branch:
+        return _cached_base, _cached_branch
+
+    url = f"https://api.github.com/repos/hao-ai-lab/FastVideo/pulls/{pr_number}"
+    response = requests.get(url)
+    if response.status_code == 200:
+        data = response.json()
+        _cached_base = data['head']['repo']['full_name']
+        _cached_branch = data['head']['ref']
+        return _cached_base, _cached_branch
+    else:
+        logger.error("Failed to fetch PR details: %s", response)
+        return None, None
+
+
+def linkcode_resolve(domain, info):
+    if domain != 'py':
+        return None
+    if not info['module']:
+        return None
+
+    # Get path from module name
+    file = Path(f"{info['module'].replace('.', '/')}.py")
+    path = REPO_ROOT / file
+    if not path.exists():
+        path = REPO_ROOT / file.with_suffix("") / "__init__.py"
+    if not path.exists():
+        return None
+
+    # Get the line number of the object
+    with open(path) as f:
+        lines = f.readlines()
+    name = info['fullname'].split(".")[-1]
+    pattern = fr"^( {{4}})*((def|class) )?{name}\b.*"
+    for lineno, line in enumerate(lines, 1):
+        if not line or line.startswith("#"):
+            continue
+        if re.match(pattern, line):
+            break
+
+    # If the line number is not found, return None
+    if lineno == len(lines):
+        return None
+
+    # If the line number is found, create the URL
+    filename = path.relative_to(REPO_ROOT)
+    if "checkouts" in path.parts:
+        # a PR build on readthedocs
+        pr_number = REPO_ROOT.name
+        base, branch = get_repo_base_and_branch(pr_number)
+        if base and branch:
+            return f"https://github.com/{base}/blob/{branch}/{filename}#L{lineno}"
+    # Otherwise, link to the source file on the main branch
+    return f"https://github.com/hao-ai-lab/FastVideo/blob/main/{filename}#L{lineno}"
+
+
+# Mock out external dependencies here, otherwise the autodoc pages may be blank.
+autodoc_mock_imports = [
+    "blake3",
+    "compressed_tensors",
+    "cpuinfo",
+    "cv2",
+    "torch",
+    "huggingface_hub",
+    "torchvision",
+    "transformers",
+    "psutil",
+    "prometheus_client",
+    "sentencepiece",
+    "vllm._C",
+    "PIL",
+    "numpy",
+    'triton',
+    "tqdm",
+    "tensorizer",
+    "pynvml",
+    "outlines",
+    "xgrammar",
+    "librosa",
+    "soundfile",
+    "gguf",
+    "lark",
+    "decord",
+]
+
+for mock_target in autodoc_mock_imports:
+    if mock_target in sys.modules:
+        logger.info(
+            "Potentially problematic mock target (%s) found; "
+            "autodoc_mock_imports cannot mock modules that have already "
+            "been loaded into sys.modules when the sphinx build starts.",
+            mock_target)
+
+intersphinx_mapping = {
+    "python": ("https://docs.python.org/3", None),
+    "typing_extensions":
+    ("https://typing-extensions.readthedocs.io/en/latest", None),
+    "aiohttp": ("https://docs.aiohttp.org/en/stable", None),
+    "pillow": ("https://pillow.readthedocs.io/en/stable", None),
+    "numpy": ("https://numpy.org/doc/stable", None),
+    "torch": ("https://pytorch.org/docs/stable", None),
+    "psutil": ("https://psutil.readthedocs.io/en/stable", None),
+}
+
+navigation_with_keys = False

docs/source/contributing/developer_env/docker.md

@@ -0,0 +1,32 @@
+(docker)=
+# 🐳 Using the FastVideo Docker Image
+
+If you prefer a containerized development environment or want to avoid managing dependencies manually, you can use our prebuilt Docker image:
+
+**Image:** [`ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:latest`](https://ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev)
+
+## Starting the container
+
+```bash
+docker run --gpus all -it ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:latest
+```
+
+This will:
+
+- Start the container with GPU access  
+- Drop you into a shell with the `fastvideo-dev` Conda environment preconfigured
+
+## Using the container
+
+```bash
+# Conda environment should already be active
+# FastVideo package installed in editable mode
+
+# Pull the latest changes from remote
+cd /FastVideo
+git pull
+
+# Run linters and tests
+pre-commit run --all-files
+pytest tests/
+```

docs/source/contributing/developer_env/index.md

@@ -0,0 +1,13 @@
+(developer-env)
+
+# 🧰 Developer Environment
+
+Accelerate your FastVideo development workflow by leveraging Docker images and cloud GPUs for efficient experimentation and reproducible environments.
+
+:::{toctree}
+:caption: Contents
+:maxdepth: 1
+
+docker
+runpod
+:::

docs/source/contributing/developer_env/runpod.md

@@ -0,0 +1,52 @@
+(runpod)=
+
+# 📦 Developing FastVideo on RunPod
+
+You can easily use the FastVideo Docker image as a custom container on [RunPod](https://www.runpod.io) for development or experimentation.
+
+## Creating a new pod
+
+Choose a GPU that supports CUDA 12.4
+
+![RunPod CUDA selection](../../_static/images/runpod_cuda.png)
+
+When creating your pod template, use this image:
+
+```
+ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:latest
+```
+
+Paste Container Start Command to support SSH ([RunPod Docs](https://docs.runpod.io/pods/configuration/use-ssh)):
+
+```bash
+bash -c "apt update;DEBIAN_FRONTEND=noninteractive apt-get install openssh-server -y;mkdir -p ~/.ssh;cd $_;chmod 700 ~/.ssh;echo \"$PUBLIC_KEY\" >> authorized_keys;chmod 700 authorized_keys;service ssh start;sleep infinity"
+```
+
+![RunPod template configuration](../../_static/images/runpod_template.png)
+
+After deploying, the pod will take a few minutes to pull the image and start the SSH service.
+
+![RunPod ssh](../../_static/images/runpod_ssh.png)
+
+## Working with the pod
+
+After SSH'ing into your pod, you'll find the `fastvideo-dev` Conda environment already activated.
+
+To pull in the latest changes from the GitHub repo:
+
+```bash
+cd /FastVideo
+git pull
+```
+
+`If you have a persistent volume and want to keep your code changes, you can move /FastVideo to /workspace/FastVideo, or simply clone the repository there.`
+
+Run your development workflows as usual:
+
+```bash
+# Run linters
+pre-commit run --all-files
+
+# Run tests
+pytest tests/
+```

docs/source/contributing/overview.md

@@ -0,0 +1,52 @@
+(developer-overview)=
+
+# 🛠️ Contributing to FastVideo
+
+Thank you for your interest in contributing to FastVideo. We want to make the process as smooth for you as possible and this is a guide to help get you started!
+
+Our community is open to everyone and welcomes any contributions no matter how large or small.
+
+# Developer Environment:
+Do make sure you have CUDA 12.4 installed and supported. FastVideo currently only support Linux and CUDA GPUs, but we hope to support other platforms in the future.
+
+We recommend using a fresh Python 3.10 Conda environment to develop FastVideo:
+
+Install Miniconda:
+
+```
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
+bash Miniconda3-latest-Linux-x86_64.sh
+source ~/.bashrc
+```
+
+Create and activate a Conda environment for FastVideo:
+
+```
+conda create -n fastvideo python=3.10 -y
+conda activate fastvideo
+```
+
+Clone the FastVideo repository and go to the FastVideo directory:
+
+```
+git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
+
+```
+
+Now you can install FastVideo and setup git hooks for running linting. By using `pre-commit`, the linters will run and have to pass before you'll be able to make a commit.
+
+```bash
+pip install -e .[dev]
+
+# Can also install flash-attn (optional)
+pip install flash-attn==2.7.4.post1 --no-build-isolation 
+
+# Linting, formatting and static type checking
+pre-commit install --hook-type pre-commit --hook-type commit-msg
+
+# You can manually run pre-commit with
+pre-commit run --all-files
+
+# Unit tests
+pytest tests/
+```

docs/source/design/overview.md

@@ -0,0 +1,410 @@
+# 🔍 FastVideo Overview
+
+This document outlines FastVideo's architecture for developers interested in framework internals or contributions. It serves as an onboarding guide for new contributors by providing an overview of the most important directories and files within the `fastvideo/v1/` codebase.
+
+## Table of Contents - V1 Directory Structure and Files
+
+- [`fastvideo/v1/pipelines/`](#design-pipeline-system) - Core diffusion pipeline components
+- [`fastvideo/v1/models/`](#design-model-components) - Model implementations
+  - [`dits/`](#design-transformer-models) - Transformer-based diffusion models
+  - [`vaes/`](#design-vae-variational-auto-encoder) - Variational autoencoders
+  - [`encoders/`](#design-text-and-image-encoders) - Text and image encoders
+  - [`schedulers/`](#design-schedulers) - Diffusion schedulers
+- [`fastvideo/v1/attention/`](#design-optimized-attention) - Optimized attention implementations
+- [`fastvideo/v1/distributed/`](#design-distributed-processing) - Distributed computing utilities
+- [`fastvideo/v1/layers/`](#design-tensor-parallelism) - Custom neural network layers
+- [`fastvideo/v1/platforms/`](#design-platforms) - Hardware platform abstractions
+- [`fastvideo/v1/worker/`](#design-executor-and-worker-abstractions) - Multi-GPU process management
+- [`fastvideo/v1/fastvideo_args.py`](#design-fastvideo-args) - Argument handling
+- [`fastvideo/v1/forward_context.py`](#design-forwardcontext) - Forward pass context management
+- `fastvideo/v1/utils.py` - Utility functions
+- [`fastvideo/v1/logger.py`](#design-logger) - Logging infrastructure
+
+## Core Architecture
+
+FastVideo separates model components from execution logic with these principles:
+- **Component Isolation**: Models (encoders, VAEs, transformers) are isolated from execution (pipelines, stages, distributed processing)
+- **Modular Design**: Components can be independently replaced
+- **Distributed Execution**: Supports various parallelism strategies (Tensor, Sequence)
+- **Custom Attention Backends**: Components can support and use different Attention implementations
+- **Pipeline Abstraction**: Consistent interface across diffusion models
+
+(design-fastvideo-args)=
+## FastVideoArgs
+
+The `FastVideoArgs` class in `fastvideo/v1/fastvideo_args.py` serves as the central configuration system for FastVideo. It contains all parameters needed to control model loading, inference configuration, performance optimization settings, and more.
+
+Key features include:
+- **Command-line Interface**: Automatic conversion between CLI arguments and dataclass fields
+- **Configuration Groups**: Organized by functional areas (model loading, video params, optimization settings)
+- **Context Management**: Global access to current settings via `get_current_fastvideo_args()`
+- **Parameter Validation**: Ensures valid combinations of settings
+
+Common configuration areas:
+- **Model paths and loading options**: `model_path`, `trust_remote_code`, `revision`
+- **Distributed execution settings**: `num_gpus`, `tp_size`, `sp_size`
+- **Video generation parameters**: `height`, `width`, `num_frames`, `num_inference_steps`
+- **Precision settings**: Control computation precision for different components
+
+Example usage:
+
+```python
+# Load arguments from command line
+fastvideo_args = prepare_fastvideo_args(sys.argv[1:])
+
+# Access parameters
+model = load_model(fastvideo_args.model_path)
+
+# Set as global context
+with set_current_fastvideo_args(fastvideo_args):
+    # Code that requires access to these arguments
+    result = generate_video()
+```
+
+(design-pipeline-system)=
+## Pipeline System
+
+### `ComposedPipelineBase`
+
+This foundational class provides:
+
+- **Model Loading**: Automatically loads components from HuggingFace-Diffusers-compatible model directories
+- **Stage Management**: Creates and orchestrates processing stages
+- **Data Flow Coordination**: Ensures proper state flow between stages
+
+```python
+class MyCustomPipeline(ComposedPipelineBase):
+    _required_config_modules = [
+        "text_encoder", "tokenizer", "vae", "transformer", "scheduler"
+    ]
+    
+    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
+        # Pipeline-specific initialization
+        pass
+        
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
+        self.add_stage("input_validation_stage", InputValidationStage())
+        self.add_stage("text_encoding_stage", CLIPTextEncodingStage(
+            text_encoder=self.get_module("text_encoder"),
+            tokenizer=self.get_module("tokenizer")
+        ))
+        # Additional stages...
+```
+
+### Pipeline Stages
+Each stage handles a specific diffusion process component:
+- **Input Validation**: Parameter verification
+- **Text Encoding**: CLIP, LLaMA, or T5-based encoding
+- **Image Encoding**: Image input processing
+- **Timestep & Latent Preparation**: Setup for diffusion
+- **Denoising**: Core diffusion loop
+- **Decoding**: Latent-to-pixel conversion
+
+Each stage implements a standard interface:
+
+```python
+def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+    # Process batch and update state
+    return batch
+```
+
+(design-forwardbatch)=
+### ForwardBatch
+
+Defined in `fastvideo/v1/pipelines/pipeline_batch_info.py`, `ForwardBatch` encapsulates the data payload passed between pipeline stages. It typically holds:
+
+- **Input Data**: Prompts, images, generation parameters
+- **Intermediate State**: Embeddings, latents, timesteps, accumulated during stage execution
+- **Output Storage**: Generated results and metadata
+- **Configuration**: Sampling parameters, precision settings
+
+This structure facilitates clear state transitions between stages.
+
+(design-model-components)=
+## Model Components
+
+The `fastvideo/v1/models/` directory contains implementations of the core neural network models used in video diffusion:
+
+(design-transformer-models)=
+### Transformer Models
+
+Transformer networks perform the actual denoising during diffusion:
+
+- **Location**: `fastvideo/v1/models/dits/`
+- **Examples**:
+  - `WanTransformer3DModel`
+  - `HunyuanVideoTransformer3DModel`
+
+Features include:
+- Text/image conditioning
+- Standardized interface for model-specific optimizations
+
+```python
+def forward(
+    self, 
+    latents,                    # [B, T, C, H, W]
+    encoder_hidden_states,      # Text embeddings
+    timestep,                   # Current diffusion timestep
+    encoder_hidden_states_image=None,  # Optional image embeddings
+    **kwargs
+):
+    # Perform denoising computation
+    return noise_pred  # Predicted noise residual
+```
+
+(design-vae-variational-auto-encoder)=
+### VAE (Variational Auto-Encoder)
+
+VAEs handle conversion between pixel space and latent space:
+
+- **Location**: `fastvideo/v1/models/vaes/`
+- **Examples**:
+  - `AutoencoderKLWan`
+  - `AutoencoderKLHunyuanVideo`
+
+These models compress image/video data to a more efficient latent representation (typically 4x-8x smaller in each dimension).
+
+FastVideo's VAE implementations include:
+- Efficient video batch processing
+- Memory optimization
+- Optional tiling for large frames
+- Distributed weight support
+
+(design-text-and-image-encoders)=
+### Text and Image Encoders
+
+Encoders process conditioning inputs into embeddings:
+
+- **Location**: `fastvideo/v1/models/encoders/`
+- **Text Encoders**:
+  - `CLIPTextModel`
+  - `LlamaModel`
+  - `UMT5EncoderModel`
+- **Image Encoders**:
+  - `CLIPVisionModel`
+
+FastVideo implements optimizations such as:
+- Vocab parallelism for distributed processing
+- Caching for common prompts
+- Precision-tuned computation
+
+(design-schedulers)=
+### Schedulers
+
+Schedulers manage the diffusion sampling process:
+
+- **Location**: `fastvideo/v1/models/schedulers/`
+- **Examples**:
+  - `UniPCMultistepScheduler`
+  - `FlowMatchEulerDiscreteScheduler`
+
+These components control:
+- Diffusion timestep sequences
+- Noise prediction to latent update conversions
+- Quality/speed trade-offs
+
+```python
+def step(
+    self, 
+    model_output: torch.Tensor,
+    timestep: torch.LongTensor,
+    sample: torch.Tensor,
+    **kwargs
+) -> torch.Tensor:
+    # Process model output and update latents
+    # Return updated latents
+    return prev_sample
+```
+
+(design-optimized-attention)=
+## Optimized Attention
+
+The `fastvideo/v1/attention/` directory contains optimized attention implementations crucial for efficient video diffusion:
+
+### Attention Backends
+Multiple implementations with automatic selection:
+- **FLASH_ATTN**: Optimized for supporting hardware
+- **TORCH_SDPA**: Built-in PyTorch scaled dot-product attention
+- **SLIDING_TILE_ATTN**: For very long sequences
+
+```python
+# Configure available attention backends for this layer
+self.attn = LocalAttention(
+    num_heads=num_heads,
+    head_size=head_dim,
+    causal=False,
+    supported_attention_backends=(_Backend.FLASH_ATTN, _Backend.TORCH_SDPA)
+)
+
+# Override via environment variable
+# export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+```
+
+### Attention Patterns
+Supports various patterns with memory optimization techniques:
+- **Cross/Self/Temporal/Global-Local Attention**
+- Chunking, progressive computation, optimized masking
+
+(design-distributed-processing)=
+## Distributed Processing
+
+The `fastvideo/v1/distributed/` directory contains implementations for distributed model execution:
+
+(design-tensor-parallelism)=
+### Tensor Parallelism
+
+Tensor parallelism splits model weights across devices:
+
+- **Implementation**: Through `RowParallelLinear` and `ColumnParallelLinear` layers
+- **Use cases**: Will be used by encoder models as their sequence lengths are shorter and enables efficient sharding.
+
+```python
+# Tensor-parallel layers in a transformer block
+from fastvideo.v1.layers.linear import ColumnParallelLinear, RowParallelLinear
+
+# Split along output dimension
+self.qkv_proj = ColumnParallelLinear(
+    input_size=hidden_size,
+    output_size=3 * hidden_size,
+    bias=True,
+    gather_output=False
+)
+
+# Split along input dimension
+self.out_proj = RowParallelLinear(
+    input_size=hidden_size,
+    output_size=hidden_size,
+    bias=True,
+    input_is_parallel=True
+)
+```
+
+### Sequence Parallelism
+
+Sequence parallelism splits sequences across devices:
+
+- **Implementation**: Through `DistributedAttention` and sequence splitting
+- **Use cases**: Long video sequences or high-resolution processing. Used by DiT models.
+
+```python
+# Distributed attention for long sequences
+from fastvideo.v1.attention import DistributedAttention
+
+self.attn = DistributedAttention(
+    num_heads=num_heads,
+    head_size=head_dim,
+    causal=False,
+    supported_attention_backends=(_Backend.SLIDING_TILE_ATTN, _Backend.FLASH_ATTN)
+)
+```
+
+### Communication Primitives
+Efficient distributed operations via AllGather, AllReduce, and synchronization mechanisms.
+
+Efficient communication primitives minimize distributed overhead:
+
+- **Sequence-Parallel AllGather**: Collects sequence chunks
+- **Tensor-Parallel AllReduce**: Combines partial results
+- **Distributed Synchronization**: Coordinates execution
+
+(design-forwardcontext)=
+## Forward Context Management
+
+### ForwardContext
+
+Defined in `fastvideo/v1/forward_context.py`, `ForwardContext` manages execution-specific state *within* a forward pass, particularly for low-level optimizations. It is accessed via `get_forward_context()`.
+
+- **Attention Metadata**: Configuration for optimized attention kernels (`attn_metadata`)
+- **Profiling Data**: Potential hooks for performance metrics collection
+
+This context-based approach enables:
+- Dynamic optimization based on execution state (e.g., attention backend selection)
+- Step-specific customizations within model components
+
+Usage example:
+
+```python
+with set_forward_context(current_timestep, attn_metadata, fastvideo_args):
+    # During this forward pass, components can access context
+    # through get_forward_context()
+    output = model(inputs)
+```
+
+(design-executor-and-worker-abstractions)=
+## Executor and Worker System
+
+The `fastvideo/v1/worker/` directory contains the distributed execution framework:
+
+### Executor Abstraction
+
+FastVideo implements a flexible execution model for distributed processing:
+
+- **Executor Base Class**: An abstract base class defining the interface for all executors
+- **MultiProcExecutor**: Primary implementation that spawns and manages worker processes
+- **GPU Workers**: Handle actual model execution on individual GPUs
+
+The MultiProcExecutor implementation:
+1. Spawns worker processes for each GPU
+2. Establishes communication channels via pipes
+3. Coordinates distributed operations across workers
+4. Handles graceful startup and shutdown of the process group
+
+Each GPU worker:
+1. Initializes the distributed environment
+2. Builds the pipeline for the specified model
+3. Executes requested operations on its assigned GPU
+4. Manages local resources and communicates results back to the executor
+
+This design allows FastVideo to efficiently utilize multiple GPUs while providing a simple, unified interface for model execution.
+
+(design-platforms)=
+## Platforms
+
+The `fastvideo/v1/platforms/` directory provides hardware platform abstractions that enable FastVideo to run efficiently on different hardware configurations:
+
+### Platform Abstraction
+
+FastVideo's platform abstraction layer enables:
+- **Hardware Detection**: Automatic detection of available hardware
+- **Backend Selection**: Appropriate selection of compute kernels
+- **Memory Management**: Efficient utilization of hardware-specific memory features
+
+The primary components include:
+- **Platform Interface**: Defines the common API for all platform implementations
+- **CUDA Platform**: Optimized implementation for NVIDIA GPUs
+- **Backend Enum**: Used throughout the codebase for feature selection
+
+Usage example:
+
+```python
+from fastvideo.v1.platforms import current_platform, _Backend
+
+# Check hardware capabilities
+if current_platform.supports_backend(_Backend.FLASH_ATTN):
+    # Use FlashAttention implementation
+else:
+    # Fall back to standard implementation
+```
+
+The platform system is designed to be extensible for future hardware targets.
+
+(design-logger)=
+## Logger
+See [PR](https://github.com/hao-ai-lab/FastVideo/pull/356)
+
+*TODO*: (help wanted) Add an environment variable that disables process-aware logging.
+
+## Contributing to FastVideo
+
+If you're a new contributor, here are some common areas to explore:
+
+1. **Adding a new model**: Implement new model types in the appropriate subdirectory of `fastvideo/v1/models/`
+2. **Optimizing performance**: Look at attention implementations or memory management
+3. **Adding a new pipeline**: Create a new pipeline subclass in `fastvideo/v1/pipelines/`
+4. **Hardware support**: Extend the `platforms` module for new hardware targets
+
+When adding code, follow these practices:
+- Use type hints for better code readability
+- Add appropriate docstrings
+- Maintain the separation between model components and execution logic
+- Follow existing patterns for distributed processing

docs/source/generate_examples.py

@@ -0,0 +1,274 @@
+# SPDX-License-Identifier: Apache-2.0
+# adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/docs/source/generate_examples.py
+
+import itertools
+import re
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Optional
+
+ROOT_DIR = Path(__file__).parent.parent.parent.resolve()
+ROOT_DIR_RELATIVE = '../../../..'
+EXAMPLE_DIR = ROOT_DIR / "examples"
+EXAMPLE_DOC_DIR = ROOT_DIR / "docs/source/getting_started/examples"
+
+
+def fix_case(text: str) -> str:
+    subs = {
+        "api": "API",
+        "cli": "CLI",
+        "cpu": "CPU",
+        "llm": "LLM",
+        "tpu": "TPU",
+        "aqlm": "AQLM",
+        "gguf": "GGUF",
+        "lora": "LoRA",
+        "rlhf": "RLHF",
+        "vllm": "vLLM",
+        "openai": "OpenAI",
+        "multilora": "MultiLoRA",
+        "mlpspeculator": "MLPSpeculator",
+        r"fp\d+": lambda x: x.group(0).upper(),  # e.g. fp16, fp32
+        r"int\d+": lambda x: x.group(0).upper(),  # e.g. int8, int16
+    }
+    for pattern, repl in subs.items():
+        text = re.sub(rf'\b{pattern}\b', repl, text,
+                      flags=re.IGNORECASE)  # type: ignore[call-overload]
+    return text
+
+
+@dataclass
+class Index:
+    """
+    Index class to generate a structured document index.
+
+    Attributes:
+        path (Path): The path save the index file to.
+        title (str): The title of the index.
+        description (str): A brief description of the index.
+        caption (str): An optional caption for the table of contents.
+        maxdepth (int): The maximum depth of the table of contents. Defaults to 1.
+        documents (list[str]): A list of document paths to include in the index. Defaults to an empty list.
+
+    Methods:
+        generate() -> str:
+            Generates the index content as a string in the specified format.
+    """ # noqa: E501
+    path: Path
+    title: str
+    description: str
+    caption: str
+    maxdepth: int = 1
+    documents: list[str] = field(default_factory=list)
+
+    def generate(self) -> str:
+        content = f"# {self.title}\n\n{self.description}\n\n"
+        content += ":::{toctree}\n"
+        content += f":caption: {self.caption}\n:maxdepth: {self.maxdepth}\n"
+        content += "\n".join(self.documents) + "\n:::\n"
+        return content
+
+
+@dataclass
+class Example:
+    """
+    Example class for generating documentation content from a given path.
+
+    Attributes:
+        path (Path): The path to the main directory or file.
+        category (str): The category of the document.
+        main_file (Path): The main file in the directory.
+        other_files (list[Path]): list of other files in the directory.
+        title (str): The title of the document.
+
+    Methods:
+        __post_init__(): Initializes the main_file, other_files, and title attributes.
+        determine_main_file() -> Path: Determines the main file in the given path.
+        determine_other_files() -> list[Path]: Determines other files in the directory excluding the main file.
+        determine_title() -> str: Determines the title of the document.
+        generate() -> str: Generates the documentation content.
+    """ # noqa: E501
+    path: Path
+    category: Optional[str] = None
+    main_file: Path = field(init=False)
+    other_files: list[Path] = field(init=False)
+    title: str = field(init=False)
+
+    def __post_init__(self):
+        self.main_file = self.determine_main_file()
+        self.other_files = self.determine_other_files()
+        self.title = self.determine_title()
+
+    def determine_main_file(self) -> Path:
+        """
+        Determines the main file in the given path.
+        If the path is a file, it returns the path itself. Otherwise, it searches
+        for Markdown files (*.md) in the directory and returns the first one found.
+        Returns:
+            Path: The main file path, either the original path if it's a file or the first
+            Markdown file found in the directory.
+        Raises:
+            IndexError: If no Markdown files are found in the directory.
+        """ # noqa: E501
+        return self.path if self.path.is_file() else list(
+            self.path.glob("*.md")).pop()
+
+    def determine_other_files(self) -> list[Path]:
+        """
+        Determine other files in the directory excluding the main file.
+
+        This method checks if the given path is a file. If it is, it returns an empty list.
+        Otherwise, it recursively searches through the directory and returns a list of all
+        files that are not the main file.
+
+        Returns:
+            list[Path]: A list of Path objects representing the other files in the directory.
+        """ # noqa: E501
+        if self.path.is_file():
+            return []
+        is_other_file = lambda file: file.is_file() and file != self.main_file
+        return [file for file in self.path.rglob("*")
+                if is_other_file(file)]  # type: ignore[no-untyped-call]
+
+    def determine_title(self) -> str:
+        return fix_case(self.path.stem.replace("_", " ").title())
+
+    def generate(self) -> str:
+        # Convert the path to a relative path from __file__
+        make_relative = lambda path: ROOT_DIR_RELATIVE / path.relative_to(
+            ROOT_DIR)
+
+        content = f"Source <gh-file:{self.path.relative_to(ROOT_DIR)}>.\n\n"
+        include = "include" if self.main_file.suffix == ".md" else \
+            "literalinclude"
+        if include == "literalinclude":
+            content += f"# {self.title}\n\n"
+        content += f":::{{{include}}} {make_relative(self.main_file)}\n"  # type: ignore[no-untyped-call]
+        if include == "literalinclude":
+            content += f":language: {self.main_file.suffix[1:]}\n"
+        content += ":::\n\n"
+
+        if not self.other_files:
+            return content
+
+        content += "## Example materials\n\n"
+        for file in sorted(self.other_files):
+            include = "include" if file.suffix == ".md" else "literalinclude"
+            content += f":::{{admonition}} {file.relative_to(self.path)}\n"
+            content += ":class: dropdown\n\n"
+            content += f":::{{{include}}} {make_relative(file)}\n:::\n"  # type: ignore[no-untyped-call]
+            content += ":::\n\n"
+
+        return content
+
+
+def generate_examples(generate_main_index=False):
+    """
+    Generate example documentation.
+    
+    Args:
+        generate_main_index (bool): Whether to generate the main examples index.
+            If False, only category-specific indices will be generated.
+    """
+    # Create empty indices with dynamic paths
+    main_index_dir = ROOT_DIR / "docs/source/examples"
+    if not main_index_dir.exists():
+        main_index_dir.mkdir(parents=True)
+
+    # Create the main examples index only if requested
+    examples_index = None
+    if generate_main_index:
+        examples_index = Index(
+            path=main_index_dir / "examples_index.md",
+            title="💡 Examples",
+            description=
+            "A collection of examples demonstrating usage of FastVideo.\nAll documented examples are autogenerated using <gh-file:docs/source/generate_examples.py> from examples found in <gh-file:examples>.",  # noqa: E501
+            caption="Examples",
+            maxdepth=2)
+
+    # Category indices with dynamic paths based on category names
+    category_indices = {
+        "inference":
+        Index(
+            path=ROOT_DIR /
+            "docs/source/inference/examples/examples_inference_index.md",
+            title="🚀 Examples",
+            description=
+            "Inference examples demonstrate how to use FastVideo in an offline setting, where the model is queried for predictions in batches. We recommend starting with <project:basic.md>.",  # noqa: E501
+            caption="Examples",
+        ),
+    }
+
+    # Ensure all category doc directories exist
+    for category, index in category_indices.items():
+        category_dir = index.path.parent
+        if not category_dir.exists():
+            category_dir.mkdir(parents=True)
+
+    examples = []
+    glob_patterns = ["*.py", "*.md", "*.sh"]
+    # Find categorised examples
+    for category in category_indices:
+        print(category)
+        category_dir = EXAMPLE_DIR / category
+        globs = [category_dir.glob(pattern) for pattern in glob_patterns]
+        for path in itertools.chain(*globs):
+            examples.append(Example(path, category))
+        # Find examples in subdirectories
+        for path in category_dir.glob("*/*.md"):
+            examples.append(Example(path.parent, category))
+
+    # Find uncategorised examples only if we're generating a main index
+    if generate_main_index:
+        globs = [EXAMPLE_DIR.glob(pattern) for pattern in glob_patterns]
+        for path in itertools.chain(*globs):
+            examples.append(Example(path))
+        # Find examples in subdirectories
+        for path in EXAMPLE_DIR.glob("*/*.md"):
+            # Skip categorised examples
+            if path.parent.name in category_indices:
+                continue
+            examples.append(Example(path.parent))
+
+    # Create document directories for each category based on category name and generate files
+    for example in sorted(examples, key=lambda e: e.path.stem):
+        print(example)
+
+        # Determine which index to use for this example
+        if example.category is not None and example.category in category_indices:
+            index = category_indices[example.category]
+        elif generate_main_index:
+            assert examples_index is not None
+            index = examples_index  # Default to main index if available
+        else:
+            # Skip examples without a category if no main index
+            print(f"Skipping {example.path} (no category and no main index)")
+            continue
+
+        # Place generated example markdown in the same directory as its index
+        doc_path = index.path.parent / f"{example.path.stem}.md"
+        with open(doc_path, "w+") as f:
+            f.write(example.generate())
+        # Add the example to the index
+        index.documents.append(example.path.stem)
+
+    # Generate the index files for categories
+    for category_index in category_indices.values():
+        if category_index.documents:
+            # Add to main index if it exists
+            if generate_main_index:
+                rel_path = category_index.path.relative_to(
+                    main_index_dir.parent)
+                assert examples_index is not None
+                examples_index.documents.insert(
+                    0,
+                    str(rel_path).replace(".md", ""))
+
+            # Write the category index file
+            with open(category_index.path, "w+") as f:
+                f.write(category_index.generate())
+
+    # Write the main index file if requested
+    if generate_main_index and examples_index:
+        with open(examples_index.path, "w+") as f:
+            f.write(examples_index.generate())

docs/source/getting_started/installation.md

@@ -0,0 +1,120 @@
+(fastvideo-installation)=
+
+# 🔧 Installation
+
+FastVideo currently only supports Linux and NVIDIA CUDA GPUs.
+
+## Requirements
+
+- **OS: Linux**
+- **Python: 3.10-3.12**
+- **CUDA 12.4**
+- **At least 1 NVIDIA GPU**
+
+## Set up using Python
+### Create a new Python environment
+
+#### Conda
+You can create a new python environment using [Conda](https://docs.conda.io/projects/conda/en/stable/user-guide/getting-started.html)
+##### 1. Install Miniconda (if not already installed)
+
+```bash
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
+bash Miniconda3-latest-Linux-x86_64.sh
+source ~/.bashrc
+```
+
+##### 2. Create and activate a Conda environment for FastVideo
+
+```bash
+# (Recommended) Create a new conda environment.
+conda create -n fastvideo python=3.12 -y
+conda activate fastvideo
+```
+
+:::{note}
+[PyTorch has deprecated the conda release channel](https://github.com/pytorch/pytorch/issues/138506). If you use `conda`, please only use it to create Python environment rather than installing packages.
+:::
+
+#### uv
+
+:::{tip}
+We highly recommend using `uv` to install FastVideo. In our experience, `uv` speeds up installation by at least 3x.
+:::
+
+Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
+
+```console
+# (Recommended) Create a new uv environment. Use `--seed` to install `pip` and `setuptools` in the environment.
+uv venv --python 3.12 --seed
+source .venv/bin/activate
+```
+
+### Installation
+
+```bash
+pip install fastvideo
+
+# or if you are using uv
+uv pip install fastvideo
+```
+
+Also optionally install flash-attn:
+
+```bash
+pip install flash-attn==2.7.4.post1 --no-build-isolation
+```
+
+### Installation from Source
+
+#### 1. Clone the FastVideo repository
+
+```bash
+git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
+```
+
+#### 2. Install FastVideo
+
+Basic installation:
+
+```bash
+pip install -e .
+
+# or if you are using uv
+uv pip install -e .
+```
+
+### Optional Dependencies
+
+#### Flash Attention
+
+```bash
+pip install flash-attn==2.7.4.post1 --no-build-isolation
+```
+
+## Set up using Docker
+We also have prebuilt docker images with FastVideo dependencies pre-installed:
+[Docker Images](#docker)
+
+## Development Environment Setup
+
+If you're planning to contribute to FastVideo please see the following page:
+[Contributor Guide](#developer-overview)
+
+## Hardware Requirements
+
+### For Basic Inference
+- NVIDIA GPU with CUDA 12.4 support
+
+### For Lora Finetuning
+- 40GB GPU memory each for 2 GPUs with lora
+- 30GB GPU memory each for 2 GPUs with CPU offload and lora
+
+### For Full Finetuning/Distillation
+- Multiple high-memory GPUs recommended (e.g., H100)
+
+## Troubleshooting
+
+If you encounter any issues during installation, please open an issue on our [GitHub repository](https://github.com/hao-ai-lab/FastVideo).
+
+You can also join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-2zf6ru791-sRwI9lPIUJQq1mIeB_yjJg) for additional support.

docs/source/getting_started/v1_api.md

@@ -0,0 +1,83 @@
+# V1 API
+
+FastVideo's V1 API provides a streamlined interface for video generation tasks with powerful customization options. This page documents the primary components of the API.
+
+## Video Generator
+
+This class will be the primary Python API for generating videos and images.
+
+```{autodoc2-summary}
+    fastvideo.VideoGenerator
+```
+
+`````{py:class} VideoGenerator(fastvideo_args: fastvideo.v1.fastvideo_args.FastVideoArgs, executor_class: type[fastvideo.v1.worker.executor.Executor], log_stats: bool)
+:canonical: fastvideo.v1.entrypoints.video_generator.VideoGenerator
+
+```{autodoc2-docstring} fastvideo.v1.entrypoints.video_generator.VideoGenerator
+:parser: docs.source.autodoc2_docstring_parser
+```
+
+`VideoGenerator.from_pretrained()` should be the primary way of creating a new video generator.
+
+````{py:method} from_pretrained(model_path: str, device: typing.Optional[str] = None, torch_dtype: typing.Optional[torch.dtype] = None, pipeline_config: typing.Optional[typing.Union[str | fastvideo.v1.configs.pipelines.PipelineConfig]] = None, **kwargs) -> fastvideo.v1.entrypoints.video_generator.VideoGenerator
+:canonical: fastvideo.v1.entrypoints.video_generator.VideoGenerator.from_pretrained
+:classmethod:
+
+```{autodoc2-docstring} fastvideo.v1.entrypoints.video_generator.VideoGenerator.from_pretrained
+:parser: docs.source.autodoc2_docstring_parser
+```
+
+
+## Configuring FastVideo
+
+The follow two classes `PipelineConfig` and `SamplingParam` are used to configure initialization and sampling parameters, respectively.
+
+### PipelineConfig
+```{autodoc2-summary}
+    fastvideo.PipelineConfig
+```
+
+`````{py:class} PipelineConfig
+:canonical: fastvideo.v1.configs.pipelines.base.PipelineConfig
+
+```{autodoc2-docstring} fastvideo.v1.configs.pipelines.base.PipelineConfig
+:parser: docs.source.autodoc2_docstring_parser
+```
+
+````{py:method} from_pretrained(model_path: str) -> fastvideo.v1.configs.pipelines.base.PipelineConfig
+:canonical: fastvideo.v1.configs.pipelines.base.PipelineConfig.from_pretrained
+:classmethod:
+
+```{autodoc2-docstring} fastvideo.v1.configs.pipelines.base.PipelineConfig.from_pretrained
+:parser: docs.source.autodoc2_docstring_parser
+```
+
+
+````{py:method} dump_to_json(file_path: str)
+:canonical: fastvideo.v1.configs.pipelines.base.PipelineConfig.dump_to_json
+
+```{autodoc2-docstring} fastvideo.v1.configs.pipelines.base.PipelineConfig.dump_to_json
+:parser: docs.source.autodoc2_docstring_parser
+```
+
+
+### SamplingParam
+
+```{autodoc2-summary}
+    fastvideo.SamplingParam
+```
+
+`````{py:class} SamplingParam
+:canonical: fastvideo.v1.configs.sample.base.SamplingParam
+
+```{autodoc2-docstring} fastvideo.v1.configs.sample.base.SamplingParam
+:parser: docs.source.autodoc2_docstring_parser
+```
+
+````{py:method} from_pretrained(model_path: str) -> fastvideo.v1.configs.sample.base.SamplingParam
+:canonical: fastvideo.v1.configs.sample.base.SamplingParam.from_pretrained
+:classmethod:
+
+```{autodoc2-docstring} fastvideo.v1.configs.sample.base.SamplingParam.from_pretrained
+:parser: docs.source.autodoc2_docstring_parser
+```

docs/source/index.md

@@ -0,0 +1,119 @@
+# Welcome to FastVideo
+
+:::{figure} ../../assets/logo.jpg
+:align: center
+:alt: FastVideo
+:class: no-scaled-link
+:width: 60%
+:::
+
+:::{raw} html
+<p style="text-align:center">
+<strong>FastVideo is a unified framework for accelerated video generation.
+</strong>
+</p>
+
+<p style="text-align:center">
+<script async defer src="https://buttons.github.io/buttons.js"></script>
+<a class="github-button" href="https://github.com/hao-ai-lab/FastVideo/" data-show-count="true" data-size="large" aria-label="Star">Star</a>
+<a class="github-button" href="https://github.com/hao-ai-lab/FastVideo/subscription" data-icon="octicon-eye" data-size="large" aria-label="Watch">Watch</a>
+<a class="github-button" href="https://github.com/hao-ai-lab/FastVideo/fork" data-icon="octicon-repo-forked" data-size="large" aria-label="Fork">Fork</a>
+</p>
+:::
+
+It features a clean, consistent API that works across popular video models, making it easier for developers to author new models and incorporate system- or kernel-level optimizations.
+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%"/>
+</div>
+
+## Key Features
+
+FastVideo has the following features:
+- State-of-the-art performance optimizations for inference
+  - [Sliding Tile Attention](https://arxiv.org/pdf/2502.04507)
+  - [TeaCache](https://arxiv.org/pdf/2411.19108)
+  - [Sage Attention](https://arxiv.org/abs/2410.02367)
+- Cutting edge models
+  - Wan2.1 T2V, I2V
+  - HunyuanVideo
+  - FastHunyuan: consistency distilled video diffusion models for 8x inference speedup.
+  - StepVideo T2V
+- Distillation support
+  - Recipes for video DiT, based on [PCM](https://github.com/G-U-N/Phased-Consistency-Model).
+  - Support distilling/finetuning/inferencing state-of-the-art open video DiTs: 1. Mochi 2. Hunyuan.
+- Scalable training with FSDP, sequence parallelism, and selective activation checkpointing, with near linear scaling to 64 GPUs.
+- Memory efficient finetuning with LoRA, precomputed latent, and precomputed text embeddings.
+
+## Documentation
+
+% How to start using FastVideo?
+
+:::{toctree}
+:caption: Getting Started
+:maxdepth: 1
+
+getting_started/installation
+<!-- getting_started/v1_api -->
+:::
+
+:::{toctree}
+:caption: Inference
+:maxdepth: 1
+
+inference/inference_quick_start
+inference/configuration
+inference/optimizations
+inference/support_matrix
+inference/examples/examples_inference_index
+inference/cli
+inference/add_pipeline
+inference/v0_inference
+:::
+
+:::{toctree}
+:caption: Training
+:maxdepth: 1
+
+training/data_preprocess
+training/distillation
+training/finetune
+:::
+
+% What is STA Kernel?
+
+:::{toctree}
+:caption: Sliding Tile Attention
+:maxdepth: 1
+
+sliding_tile_attention/installation
+sliding_tile_attention/demo
+:::
+
+:::{toctree}
+:caption: Design
+:maxdepth: 1
+design/overview
+:::
+
+:::{toctree}
+:caption: Developer Guide
+:maxdepth: 2
+
+contributing/overview
+contributing/developer_env/index
+:::
+
+:::{toctree}
+:caption: API Reference
+:maxdepth: 2
+
+<!-- api/summary -->
+api/fastvideo/fastvideo
+:::
+
+## Indices and tables
+
+- {ref}`genindex`
+- {ref}`modindex`

docs/source/inference/add_pipeline.md

@@ -0,0 +1,316 @@
+(add-pipeline)=
+
+# 🏗️ Adding a New Pipeline
+
+This guide explains how to implement a custom diffusion pipeline in FastVideo, leveraging the framework's modular architecture for high-performance video generation.
+
+## Implementation Process Overview
+
+1. **Port Required Modules** - Identify and implement necessary model components
+2. **Create Directory Structure** - Set up pipeline files and folders
+3. **Implement Pipeline Class** - Build the pipeline using existing or custom stages
+4. **Register Your Pipeline** - Make it discoverable by the framework
+5. **Configure Your Pipeline** - (Coming soon)
+
+Need help? Join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-2zf6ru791-sRwI9lPIUJQq1mIeB_yjJg).
+
+## Step 1: Pipeline Modules
+
+### Identifying Required Modules
+
+FastVideo uses the Hugging Face Diffusers format for model organization:
+
+1. Examine the `model_index.json` in the HF model repository:
+
+```json
+{
+    "_class_name": "WanImageToVideoPipeline",
+    "_diffusers_version": "0.33.0.dev0",
+    "image_encoder": ["transformers", "CLIPVisionModelWithProjection"],
+    "image_processor": ["transformers", "CLIPImageProcessor"],
+    "scheduler": ["diffusers", "UniPCMultistepScheduler"],
+    "text_encoder": ["transformers", "UMT5EncoderModel"],
+    "tokenizer": ["transformers", "T5TokenizerFast"],
+    "transformer": ["diffusers", "WanTransformer3DModel"],
+    "vae": ["diffusers", "AutoencoderKLWan"]
+}
+```
+
+1. For each component:
+   - Note the originating library (`transformers` or `diffusers`)
+   - Identify the class name
+   - Check if it's already available in FastVideo
+
+2. Review config files in each component's directory for architecture details
+
+### Implementing Modules
+
+Place new modules in the appropriate directories:
+- Encoders: `fastvideo/v1/models/encoders/`
+- VAEs: `fastvideo/v1/models/vaes/`
+- Transformer models: `fastvideo/v1/models/dits/`
+- Schedulers: `fastvideo/v1/models/schedulers/`
+
+### Adapting Model Layers
+
+#### Layer Replacements
+Replace standard PyTorch layers with FastVideo optimized versions:
+- nn.LayerNorm → fastvideo.v1.layers.layernorm.RMSNorm
+- Embedding layers → fastvideo.v1.layers.vocab_parallel_embedding modules
+- Activation functions → versions from fastvideo.v1.layers.activation
+
+#### Distributed Linear Layers
+Use appropriate parallel layers for distribution:
+
+```python
+# Output dimension parallelism
+from fastvideo.v1.layers.linear import ColumnParallelLinear
+self.q_proj = ColumnParallelLinear(
+    input_size=hidden_size,
+    output_size=head_size * num_heads,
+    bias=bias,
+    gather_output=False
+)
+
+# Fused QKV projection
+from fastvideo.v1.layers.linear import QKVParallelLinear
+self.qkv_proj = QKVParallelLinear(
+    hidden_size=hidden_size,
+    head_size=attention_head_dim,
+    total_num_heads=num_attention_heads,
+    bias=True
+)
+
+# Input dimension parallelism
+from fastvideo.v1.layers.linear import RowParallelLinear
+self.out_proj = RowParallelLinear(
+    input_size=head_size * num_heads,
+    output_size=hidden_size,
+    bias=bias,
+    input_is_parallel=True
+)
+```
+
+### Attention Layers
+Replace standard attention with FastVideo's optimized attention:
+
+```python
+# Local attention patterns
+from fastvideo.v1.attention import LocalAttention
+from fastvideo.v1.attention.backends.abstract import _Backend
+self.attn = LocalAttention(
+    num_heads=num_heads,
+    head_size=head_dim,
+    dropout_rate=0.0,
+    softmax_scale=None,
+    causal=False,
+    supported_attention_backends=(_Backend.FLASH_ATTN, _Backend.TORCH_SDPA)
+)
+
+# Distributed attention for long sequences
+from fastvideo.v1.attention import DistributedAttention
+self.attn = DistributedAttention(
+    num_heads=num_heads,
+    head_size=head_dim,
+    dropout_rate=0.0,
+    softmax_scale=None,
+    causal=False,
+    supported_attention_backends=(_Backend.SLIDING_TILE_ATTN, _Backend.FLASH_ATTN, _Backend.TORCH_SDPA)
+)
+```
+
+#### Define supported backend selection
+
+```python
+   _supported_attention_backends = (_Backend.FLASH_ATTN, _Backend.TORCH_SDPA)
+```
+
+### Registering Models
+
+Register implemented modules in the model registry:
+
+```python
+# In fastvideo/v1/models/registry.py
+_TEXT_TO_VIDEO_DIT_MODELS = {
+    "YourTransformerModel": ("dits", "yourmodule", "YourTransformerClass"),
+}
+
+_VAE_MODELS = {
+    "YourVAEModel": ("vaes", "yourvae", "YourVAEClass"),
+}
+```
+
+## Step 2: Directory Structure
+
+Create a new directory for your pipeline:
+
+```
+fastvideo/v1/pipelines/
+├── your_pipeline/
+│   ├── __init__.py
+│   └── your_pipeline.py
+```
+
+## Step 3: Implement Pipeline Class
+
+Pipelines are composed of stages, each handling a specific part of the diffusion process:
+
+- **InputValidationStage**: Validates input parameters
+- **Text Encoding Stages**: Handle text encoding (CLIP/Llama/T5)
+- **CLIPImageEncodingStage**: Processes image inputs
+- **TimestepPreparationStage**: Prepares diffusion timesteps
+- **LatentPreparationStage**: Manages latent representations
+- **ConditioningStage**: Processes conditioning inputs
+- **DenoisingStage**: Performs denoising diffusion
+- **DecodingStage**: Converts latents to pixels
+
+### Creating Your Pipeline
+
+```python
+from fastvideo.v1.pipelines.composed_pipeline_base import ComposedPipelineBase
+from fastvideo.v1.pipelines.stages import (
+    InputValidationStage, CLIPTextEncodingStage, TimestepPreparationStage,
+    LatentPreparationStage, DenoisingStage, DecodingStage
+)
+from fastvideo.v1.fastvideo_args import FastVideoArgs
+from fastvideo.v1.pipelines.pipeline_batch_info import ForwardBatch
+import torch
+
+class MyCustomPipeline(ComposedPipelineBase):
+    """Custom diffusion pipeline implementation."""
+    
+    # Define required model components from model_index.json
+    _required_config_modules = [
+        "text_encoder", "tokenizer", "vae", "transformer", "scheduler"
+    ]
+    
+    @property
+    def required_config_modules(self) -> List[str]:
+        return self._required_config_modules
+        
+    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
+        """Initialize pipeline-specific components."""
+        pass
+        
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
+        """Set up pipeline stages with proper dependency injection."""
+        self.add_stage(
+            stage_name="input_validation_stage",
+            stage=InputValidationStage()
+        )
+        
+        self.add_stage(
+            stage_name="prompt_encoding_stage",
+            stage=CLIPTextEncodingStage(
+                text_encoder=self.get_module("text_encoder"),
+                tokenizer=self.get_module("tokenizer")
+            )
+        )
+        
+        self.add_stage(
+            stage_name="timestep_preparation_stage",
+            stage=TimestepPreparationStage(
+                scheduler=self.get_module("scheduler")
+            )
+        )
+        
+        self.add_stage(
+            stage_name="latent_preparation_stage",
+            stage=LatentPreparationStage(
+                scheduler=self.get_module("scheduler"),
+                vae=self.get_module("vae")
+            )
+        )
+        
+        self.add_stage(
+            stage_name="denoising_stage",
+            stage=DenoisingStage(
+                transformer=self.get_module("transformer"),
+                scheduler=self.get_module("scheduler")
+            )
+        )
+        
+        self.add_stage(
+            stage_name="decoding_stage",
+            stage=DecodingStage(
+                vae=self.get_module("vae")
+            )
+        )
+    
+# Register the pipeline class
+EntryClass = MyCustomPipeline
+```
+
+### Creating Custom Stages (Optional)
+
+If existing stages don't meet your needs, create custom ones:
+
+```python
+from fastvideo.v1.pipelines.stages.base import PipelineStage
+
+class MyCustomStage(PipelineStage):
+    """Custom processing stage for the pipeline."""
+    
+    def __init__(self, custom_module, other_param=None):
+        super().__init__()
+        self.custom_module = custom_module
+        self.other_param = other_param
+        
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+        # Access input data
+        input_data = batch.some_attribute
+        
+        # Validate inputs
+        if input_data is None:
+            raise ValueError("Required input is missing")
+            
+        # Process with your module
+        result = self.custom_module(input_data)
+        
+        # Update batch with results
+        batch.some_output = result
+        
+        return batch
+```
+
+Add your custom stage to the pipeline:
+
+```python
+self.add_stage(
+    stage_name="my_custom_stage",
+    stage=MyCustomStage(
+        custom_module=self.get_module("custom_module"),
+        other_param="some_value"
+    )
+)
+```
+
+#### Stage Design Principles
+
+1. **Single Responsibility**: Focus on one specific task
+2. **Functional Pattern**: Receive and return a `ForwardBatch` object
+3. **Dependency Injection**: Pass dependencies through constructor
+4. **Input Validation**: Validate inputs for clear error messages
+
+## Step 4: Register Your Pipeline
+
+Define `EntryClass` at the end of your pipeline file:
+
+```python
+# Single pipeline class
+EntryClass = MyCustomPipeline
+
+# Or multiple pipeline classes
+EntryClass = [MyCustomPipeline, MyOtherPipeline]
+```
+
+The registry will automatically:
+1. Scan all packages under `fastvideo/v1/pipelines/`
+2. Look for `EntryClass` variables
+3. Register pipelines using their class names as identifiers
+
+## Best Practices
+
+- **Reuse Existing Components**: Leverage built-in stages and modules
+- **Follow Module Organization**: Place new modules in appropriate directories
+- **Match Model Patterns**: Follow existing code patterns and conventions

docs/source/inference/cli.md

@@ -0,0 +1,151 @@
+# FastVideo CLI Inference
+
+The FastVideo CLI provides a quick way to access the FastVideo inference pipeline for video generation. For more advanced usage,
+see the Python interface [here](https://hao-ai-lab.github.io/FastVideo/inference/examples/basic.html).
+
+## Basic Usage
+
+The basic command to generate a video is:
+
+```bash
+fastvideo generate --model-path {MODEL_PATH} --prompt {PROMPT}
+```
+
+### Required Parameters
+
+- `--model-path {MODEL_PATH}`: Path to the model or model ID
+- `--prompt {PROMPT}`: Text description for the video you want to generate
+
+## Common Arguments
+
+To see all the options, you can use the `--help` flag:
+
+```bash
+fastvideo generate --help
+```
+
+### Hardware Configuration
+
+- `--num-gpus {NUM_GPUS}`: Number of GPUs to use
+- `--tp-size {TP_SIZE}`: Tensor parallelism size (Typically should match the number of GPUs)
+- `--sp-size {SP_SIZE}`: Sequence parallelism size (Typically should match the number of GPUs)
+
+#### Video Configuration
+
+- `--height {HEIGHT}`: Height of the generated video
+- `--width {WIDTH}`: Width of the generated video
+- `--num-frames {NUM_FRAMES}`: Number of frames to generate
+- `--fps {FPS}`: Frames per second for the saved video
+
+#### Generation Parameters
+
+- `--num-inference-steps {STEPS}`: Number of denoising steps
+- `--negative-prompt {PROMPT}`: Negative prompt to guide generation away from certain concepts
+- `--seed {SEED}`: Random seed for reproducible generation
+
+#### Output Options
+
+- `--output-path {PATH}`: Directory to save the generated video
+- `--save-video`: Whether to save the video to disk
+- `--return-frames`: Whether to return the raw frames
+
+## Using Configuration Files
+
+Instead of specifying all parameters on the command line, you can use a configuration file:
+
+```bash
+fastvideo generate --config {CONFIG_FILE_PATH}
+```
+
+The config file should be in JSON or YAML format with the same parameter names as the CLI options. Command-line arguments will take precedence over settings in the configuration file, allowing you to override specific values while keeping the rest from the config file.
+
+Example configuration file (config.json):
+
+```json
+{
+    "model_path": "FastVideo/FastHunyuan-diffusers",
+    "prompt": "A beautiful woman in a red dress walking down a street",
+    "output_path": "outputs/",
+    "num_gpus": 2,
+    "sp_size": 2,
+    "tp_size": 2,
+    "num_frames": 45,
+    "height": 720,
+    "width": 1280,
+    "num_inference_steps": 6,
+    "seed": 1024,
+    "fps": 24,
+    "precision": "bf16",
+    "vae_precision": "fp16",
+    "vae_tiling": true,
+    "vae_sp": true,
+    "vae_config": {
+        "load_encoder": false,
+        "load_decoder": true,
+        "tile_sample_min_height": 256,
+        "tile_sample_min_width": 256
+    },
+    "text_encoder_precisions": [
+        "fp16",
+        "fp16"
+    ],
+    "mask_strategy_file_path": null,
+    "enable_torch_compile": false
+}
+```
+
+Or using YAML format (config.yaml):
+
+```yaml
+model_path: "FastVideo/FastHunyuan-diffusers"
+prompt: "A beautiful woman in a red dress walking down a street"
+output_path: "outputs/"
+num_gpus: 2
+sp_size: 2
+tp_size: 2
+num_frames: 45
+height: 720
+width: 1280
+num_inference_steps: 6
+seed: 1024
+fps: 24
+precision: "bf16"
+vae_precision: "fp16"
+vae_tiling: true
+vae_sp: true
+vae_config:
+  load_encoder: false
+  load_decoder: true
+  tile_sample_min_height: 256
+  tile_sample_min_width: 256
+text_encoder_precisions:
+  - "fp16"
+  - "fp16"
+mask_strategy_file_path: null
+enable_torch_compile: false
+```
+
+## Examples
+
+Generating a simple video:
+
+```bash
+fastvideo generate --model-path FastVideo/FastHunyuan-diffusers --prompt "A cat playing with a ball of yarn" --num-frames 45 --height 720 --width 1280 --num-inference-steps 6 --seed 1024 --output-path outputs/
+```
+
+Using a negative prompt to avoid certain elements:
+
+```bash
+fastvideo generate --model-path FastVideo/FastHunyuan-diffusers --prompt "A beautiful forest landscape" --negative-prompt "people, buildings, roads"
+```
+
+Combining command line arguments and a configuration file:
+
+```bash
+fastvideo generate --config config.json --prompt "A capybara lounging in a hammock"
+```
+
+## Troubleshooting
+
+- If you encounter CUDA out-of-memory errors, try reducing the video dimensions or number of frames, or the number of inference steps.
+- For reproducible results, set the same seed value between runs.

docs/source/inference/configuration.md

@@ -0,0 +1,77 @@
+(inference-configuration)=
+# Configuration
+
+## Multi-GPU Setup
+
+FastVideo automatically distributes the generation process when multiple GPUs are specified:
+
+```python
+# Will use 4 GPUs in parallel for faster generation
+generator = VideoGenerator.from_pretrained(
+    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+    num_gpus=4,
+)
+```
+
+## Customizing Generation
+
+- `PipelineConfig`: Initialization time parameters
+- `SamplingParam`: Generation time parameters
+
+You can customize various parameters when generating videos using the `PipelineConfig` and `SamplingParam` class:
+
+```python
+from fastvideo import VideoGenerator, SamplingParam, PipelineConfig
+
+def main():
+    model_name = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+    config = PipelineConfig.from_pretrained(model_name)
+    config.vae_precision = "fp16"
+    config.use_cpu_offload = True
+
+    # Create the generator
+    generator = VideoGenerator.from_pretrained(
+        model_name,
+        num_gpus=1,
+        pipeline_config=config
+    )
+
+    # Create and customize sampling parameters
+    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
+    
+    # How many frames to generate
+    sampling_param.num_frames = 45
+    
+    # Video resolution (width, height)
+    sampling_param.width = 1024
+    sampling_param.height = 576
+    
+    # How many steps we denoise the video (higher = better quality, slower generation)
+    sampling_param.num_inference_steps = 30
+    
+    # How strongly the video conforms to the prompt (higher = more faithful to prompt)
+    sampling_param.guidance_scale = 7.5
+    
+    # Random seed for reproducibility
+    sampling_param.seed = 42  # Optional, leave unset for random results
+
+    # Generate video with custom parameters
+    prompt = "A beautiful sunset over a calm ocean, with gentle waves."
+    video = generator.generate_video(
+        prompt, 
+        sampling_param=sampling_param, 
+        output_path="my_videos/",  # Controls where videos are saved
+        return_frames=True,  # Also return frames from this call (defaults to False)
+        save_video=True
+    )
+
+    # If return_frames=True, video contains the generated frames as a NumPy array
+    print(f"Generated {len(video)} frames")
+
+if __name__ == '__main__':
+    main()
+```
+
+## Performance Optimization
+
+For configuring optimizations, please see our [optimizations guide](#inference-optimizations)

docs/source/inference/inference_quick_start.md

@@ -0,0 +1,124 @@
+# Inference Quick Start
+
+This page contains step-by-step instructions to get you quickly started with video generation using FastVideo.
+
+## Requirements
+- **OS**: Linux (Tested on Ubuntu 22.04+)
+- **Python**: 3.10-3.12
+- **CUDA**: 12.4
+- **GPU**: At least one NVIDIA GPU
+
+## Installation
+
+We recommend using an environment manager such as `Conda` to create a clean environment:
+
+```bash
+# Create and activate a new conda environment
+conda create -n fastvideo python=3.12
+conda activate fastvideo
+
+# Install FastVideo
+pip install fastvideo
+```
+
+For advanced installation options, see the [Installation Guide](installation.md).
+
+## Generating Your First Video
+Here's a minimal example to generate a video using the default settings. Create a file called `example.py` with the following code:
+
+```python
+from fastvideo import VideoGenerator
+
+def main():
+    # Create a video generator with a pre-trained model
+    generator = VideoGenerator.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        num_gpus=1,  # Adjust based on your hardware
+    )
+
+    # Define a prompt for your video
+    prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest."
+
+    # Generate the video
+    video = generator.generate_video(
+        prompt,
+        return_frames=True,  # Also return frames from this call (defaults to False)
+        output_path="my_videos/",  # Controls where videos are saved
+        save_video=True
+    )
+
+if __name__ == '__main__':
+    main()
+```
+
+Run the script with:
+
+```bash
+python example.py
+```
+
+The generated video will be saved in the current directory under `my_videos/`.
+
+## Available Models
+
+Please see the [support matrix](#support-matrix) for the list of supported models and their available optimizations.
+
+## Image-to-Video Generation
+
+You can generate a video starting from an initial image:
+
+```python
+from fastvideo import VideoGenerator, SamplingParam
+
+def main():
+    # Create the generator
+    model_name = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+    generator = VideoGenerator.from_pretrained(model_name, num_gpus=1)
+
+    # Set up parameters with an initial image
+    sampling_param = SamplingParam.from_pretrained(model_name)
+    sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+    sampling_param.num_frames = 107
+    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"
+    generator.generate_video(prompt, sampling_param=sampling_param,
+                             output_path="my_videos/",
+                             save_video=True)
+
+if __name__ == '__main__':
+    main()
+```
+
+## Troubleshooting
+
+Common issues and their solutions:
+
+### Out of Memory Errors
+If you encounter CUDA out of memory errors:
+- Reduce `num_frames` or video resolution
+- Enable memory optimization with `enable_model_cpu_offload`
+- Try a smaller model or use distilled versions
+- Use `num_gpus` > 1 if multiple GPUs are available
+
+### Slow Generation
+To speed up generation:
+- Reduce `num_inference_steps` (20-30 is usually sufficient)
+- Use half precision (`fp16`) for the VAE
+- Use multiple GPUs if available
+
+### Unexpected Results
+If the generated video doesn't match your prompt:
+- Try increasing `guidance_scale` (7.0-9.0 works well)
+- Make your prompt more detailed and specific
+- Experiment with different random seeds
+- Try a different model
+
+## Next Steps
+
+- Learn about [Advanced Inference Configurations](#inference-configuration)
+- Learn about using [Optimizations](#inference-optimizations)
+- See [Examples](../examples/examples_inference_index.md) for more usage scenarios
+- Join our [Community Discord](https://discord.gg/JA7cksDz86).
+- Join our [Community Slack](https://join.slack.com/t/fastvideo/shared_invite/zt-2zf6ru791-sRwI9lPIUJQq1mIeB_yjJg).

docs/source/inference/optimizations.md

@@ -0,0 +1,148 @@
+(inference-optimizations)=
+# Optimizations
+
+This page describes the various options for speeding up generation times in FastVideo.
+
+## Table of Contents
+- Optimized Attention Backends
+  - [Flash Attention](#optimizations-flash)
+  - [Sliding Tile Attention](#optimizations-sta)
+  - [Sage Attention](#optimizations-sage)
+
+- Caching Techniques
+  - [TeaCache](#optimizations-teacache)
+
+(optimizations-backends)=
+## Attention Backends
+
+### Available Backends
+- Torch SDPA: `FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA`
+- Flash Attention 2 and 3: `FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN`
+- Sliding Tile Attention: `FASTVIDEO_ATTENTION_BACKEND=SLIDING_TILE_ATTN`
+- Sage Attention: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN`
+
+### Configuring Backends
+
+There are two ways to configure the attention backend in FastVideo.
+
+#### 1. In Python
+In python, set the `FASTVIDEO_ATTENTION_BACKEND` environment variable before instantiating `VideoGenerator` like this:
+
+```python
+os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLIDING_TILE_ATTN"
+```
+
+#### 2. In CLI
+You can also set the environment variable on the command line:
+
+```bash
+FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN python example.py
+```
+
+(optimizations-flash)=
+### Flash Attention
+
+**`FLASH_ATTN`**
+
+We recommend always installing [Flash Attention 2](https://github.com/Dao-AILab/flash-attention):
+
+```bash
+pip install flash-attn==2.7.4.post1 --no-build-isolation
+```
+
+And if using a Hopper+ GPU (ie H100), installing [Flash Attention 3](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#flashattention-3-beta-release) by compiling it from source (takes about 10 minutes for me):
+
+```bash
+git clone https://github.com/Dao-AILab/flash-attention.git && cd flash-attention
+
+cd hopper
+pip install ninja 
+python setup.py install
+```
+
+:::{note}
+FastVideo will automatically detect and use `FA3` if it is installed when using `FLASH_ATTN` backend.
+:::
+
+(optimizations-sta)=
+### Sliding Tile Attention
+**`SLIDING_TILE_ATTN`**
+
+```bash
+pip install st_attn==0.0.4
+```
+
+Please see [this page](#sta-installation) for more installation instructions.
+
+(optimizations-sage)=
+### Sage Attention
+**`SAGE_ATTN`**
+
+To use [SageAttention](https://github.com/thu-ml/SageAttention) 2.1.1, please compile from source:
+
+```bash
+git clone https://github.com/thu-ml/SageAttention.git
+cd sageattention 
+python setup.py install  # or pip install -e .
+```
+
+(optimizations-teacache)=
+## Teacache
+TeaCache is an optimization technique supported in FastVideo that can significantly speed up video generation by skipping redundant calculations across diffusion steps. This guide explains how to enable and configure TeaCache for optimal performance in FastVideo.
+
+### What is TeaCache?
+
+See the official [TeaCache](https://github.com/ali-vilab/TeaCache) repo and their paper for more details.
+
+### How to Enable TeaCache
+
+Enabling TeaCache is straightforward - simply add the `enable_teacache=True` parameter to your `generate_video()` call:
+
+```python
+# ... previous code
+generator.generate_video(
+    prompt="Your prompt here",
+    sampling_param=params,
+    enable_teacache=True
+)
+# more code ...
+```
+
+### Complete Example
+
+At the bottom is a complete example of using TeaCache for faster video generation. You can run it using the following command:
+
+```bash
+python examples/inference/optimizations/teacache_example.py
+```
+
+### Advanced Configuration
+
+While TeaCache works well with default settings, you can fine-tune its behavior by adjusting the threshold value:
+
+1. Lower threshold values (e.g., 0.1) will result in more skipped calculations and faster generation with slightly more potential for quality degradation
+2. Higher threshold values (e.g., 0.15-0.23) will skip fewer calculations but maintain quality closer to the original
+
+Note that the optimal threshold depends on your specific model and content.
+
+## Benchmarking different optimizations
+
+To benchmark the performance improvement, try generating the same video with and without TeaCache enabled and compare the generation times:
+
+```python
+# Without TeaCache
+start_time = time.perf_counter()
+generator.generate_video(prompt="Your prompt", enable_teacache=False)
+standard_time = time.perf_counter() - start_time
+
+# With TeaCache
+start_time = time.perf_counter()
+generator.generate_video(prompt="Your prompt", enable_teacache=True)
+teacache_time = time.perf_counter() - start_time
+
+print(f"Standard generation: {standard_time:.2f} seconds")
+print(f"TeaCache generation: {teacache_time:.2f} seconds")
+print(f"Speedup: {standard_time/teacache_time:.2f}x")
+```
+
+Note: If you want to benchmark different attention backends, you'll need to reinstantiate `VideoGenerator`.

docs/source/inference/support_matrix.md

@@ -0,0 +1,92 @@
+(support-matrix)=
+# Compatibility Matrix
+The table below shows every supported model and optimizations supported for them.
+
+The symbols used have the following meanings:
+
+- ✅ = Full compatibility
+- ❌ = No compatibility
+
+## Models x Optimization
+The `HuggingFace Model ID` can be directly pass to `from_pretrained()` methods and FastVideo will use the optimal default parameters when initializing and generating videos.
+
+:::{raw} html
+<style>
+  /* Make smaller to try to improve readability  */
+  td {
+    font-size: 0.9rem;
+    text-align: center;
+  }
+
+  th {
+    text-align: center;
+    font-size: 0.9rem;
+  }
+</style>
+:::
+
+:::{list-table}
+:header-rows: 1
+:stub-columns: 3
+:widths: auto
+:class: vertical-table-header
+
+- * Model Name
+  * HuggingFace Model ID
+  * Resolutions
+  * TeaCache
+  * Sliding Tile Attn
+  * Sage Attn
+- * HunyuanVideo
+  * `hunyuanvideo-community/HunyuanVideo`
+  * 720px1280p<br>544px960p
+  * ❌
+  * ✅
+  * ✅
+- * FastHunyuan
+  * `FastVideo/FastHunyuan-diffusers`
+  * 720px1280p<br>544px960p
+  * ❌
+  * ✅
+  * ✅
+- * Wan T2V 1.3B
+  * `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`
+  * 480P
+  * ✅
+  * ✅*
+  * ✅
+- * Wan T2V 14B
+  * `Wan-AI/Wan2.1-T2V-14B-Diffusers`
+  * 480P, 720P
+  * ✅
+  * ✅*
+  * ✅
+- * Wan I2V 480P
+  * `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers`
+  * 480P
+  * ✅
+  * ✅*
+  * ✅
+- * Wan I2V 720P
+  * `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers`
+  * 720P
+  * ✅
+  * ✅*
+  * ✅
+- * StepVideo T2V
+  * `FastVideo/stepvideo-t2v-diffusers`
+  * 768px768px204f<br>544px992px204f<br>544px992px136f
+  * ❌
+  * ❌
+  * ✅
+:::
+
+**Note**: there are some known quality issues with Wan2.1 + Sliding Tile Attn. We are working on fixing this issue.
+
+## Special requirements
+
+### StepVideo T2V
+- The self-attention in text-encoder (step_llm) only supports CUDA capabilities sm_80 sm_86 and sm_90
+
+### Sliding Tile Attention
+- Currently only Hopper GPUs (H100s) are supported.

docs/source/inference/v0_inference.md

@@ -0,0 +1,74 @@
+(v0-inference)=
+
+# [Deprecated] V0 Inference
+The following commands and APIs are deprecated but still supported until V1's API can completely replace all the features in this page.
+
+## Inference StepVideo with Sliding Tile Attention
+First, download the model:
+
+```
+python scripts/huggingface/download_hf.py --repo_id=stepfun-ai/stepvideo-t2v --local_dir=data/stepvideo-t2v --repo_type=model
+```
+
+Use the following scripts to run inference for StepVideo. When using STA for inference, the generated videos will have dimensions of 204×768×768 (currently, this is the only supported shape).
+
+```bash
+sh scripts/inference/inference_stepvideo_STA.sh # Inference stepvideo with STA
+sh scripts/inference/inference_stepvideo.sh # Inference original stepvideo
+```
+
+## Inference HunyuanVideo with Sliding Tile Attention
+First, download the model:
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/hunyuan --local_dir=data/hunyuan --repo_type=model
+```
+
+We provide two examples in the following script to run inference with STA + [TeaCache](https://github.com/ali-vilab/TeaCache) and STA only.
+
+```bash
+sh scripts/inference/inference_hunyuan_STA.sh
+```
+
+## Video Demos using STA + Teacache
+Visit our [demo website](https://fast-video.github.io/) to explore our complete collection of examples. We shorten a single video generation process from 945s to 317s on H100.
+
+## Inference FastHunyuan on single RTX4090
+We now support NF4 and LLM-INT8 quantized inference using BitsAndBytes for FastHunyuan. With NF4 quantization, inference can be performed on a single RTX 4090 GPU, requiring just 20GB of VRAM.
+
+```bash
+# Download the model weight
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/FastHunyuan-diffusers --local_dir=data/FastHunyuan-diffusers --repo_type=model
+# CLI inference
+bash scripts/inference/inference_hunyuan_hf_quantization.sh
+```
+
+For more information about the VRAM requirements for BitsAndBytes quantization, please refer to the table below (timing measured on an H100 GPU):
+
+| Configuration                  | Memory to Init Transformer | Peak Memory After Init Pipeline (Denoise) | Diffusion Time | End-to-End Time |
+|--------------------------------|----------------------------|--------------------------------------------|----------------|-----------------|
+| BF16 + Pipeline CPU Offload    | 23.883G                   | 33.744G                                    | 81s            | 121.5s          |
+| INT8 + Pipeline CPU Offload    | 13.911G                   | 27.979G                                    | 88s            | 116.7s          |
+| NF4 + Pipeline CPU Offload     | 9.453G                    | 19.26G                                     | 78s            | 114.5s          |
+
+For improved quality in generated videos, we recommend using a GPU with 80GB of memory to run the BF16 model with the original Hunyuan pipeline. To execute the inference, use the following section:
+
+## FastHunyuan
+
+```bash
+# Download the model weight
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/FastHunyuan --local_dir=data/FastHunyuan --repo_type=model
+# CLI inference
+bash scripts/inference/inference_hunyuan.sh
+```
+
+You can also inference FastHunyuan in the [official Hunyuan github](https://github.com/Tencent/HunyuanVideo).
+
+## FastMochi
+
+```bash
+# Download the model weight
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/FastMochi-diffusers --local_dir=data/FastMochi-diffusers --repo_type=model
+# CLI inference
+bash scripts/inference/inference_mochi_sp.sh
+```

docs/source/sliding_tile_attention/demo.md

@@ -0,0 +1,11 @@
+(sta-demo)=
+
+# 🔍 Demo
+There is a demo for 2D STA with window size (6,6) operating on a (10, 10) image.
+
+<div style="text-align: center;">
+  <video controls width="800">
+    <source src="https://github.com/user-attachments/assets/f3b6dd79-7b43-4b60-a0fa-3d6495ec5747" type="video/mp4">
+    Your browser does not support the video tag.
+  </video>
+</div>

docs/source/sliding_tile_attention/installation.md

@@ -0,0 +1,55 @@
+(sta-installation)=
+
+# 🔧 Installation
+You can install the Sliding Tile Attention package using
+
+```
+pip install st_attn==0.0.4
+``` 
+
+# Building from Source
+We test our code on Pytorch 2.5.0 and CUDA>=12.4. Currently we only have implementation on H100.
+First, install C++20 for ThunderKittens:
+
+```bash
+cd csrc/sliding_tile_attention/
+sudo apt update
+sudo apt install gcc-11 g++-11
+
+sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11
+
+sudo apt update
+sudo apt install clang-11
+```
+
+Install STA:
+
+```bash
+export CUDA_HOME=/usr/local/cuda-12.4
+export PATH=${CUDA_HOME}/bin:${PATH} 
+export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+git submodule update --init --recursive
+python setup.py install
+```
+
+# 🧪 Test
+
+```bash
+python test/test_sta.py
+```
+
+# 📋 Usage
+
+```python
+from st_attn import sliding_tile_attention
+# assuming video size (T, H, W) = (30, 48, 80), text tokens = 256 with padding. 
+# q, k, v: [batch_size, num_heads, seq_length, head_dim], seq_length = T*H*W + 256
+# a tile is a cube of size (6, 8, 8)
+# window_size in tiles: [(window_t, window_h, window_w), (..)...]. For example, window size (3, 3, 3) means a query can attend to (3x6, 3x8, 3x8) = (18, 24, 24) tokens out of the total 30x48x80 video.
+# text_length: int ranging from 0 to 256
+# If your attention contains text token (Hunyuan)
+out = sliding_tile_attention(q, k, v, window_size, text_length)
+# If your attention does not contain text token (StepVideo)
+out = sliding_tile_attention(q, k, v, window_size, 0, False)
+
+```

docs/source/training/data_preprocess.md

@@ -1,16 +1,17 @@
+(v0-data-preprocess)=
 
-
-
-## 🧱 Data Preprocess
+# 🧱 Data Preprocess
 
 To save GPU memory, we precompute text embeddings and VAE latents to eliminate the need to load the text encoder and VAE during training.
 
-
 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
 ```
+
 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
@@ -18,10 +19,11 @@ bash scripts/preprocess/preprocess_hunyuan_data.sh # for hunyuan
 
 The preprocessed dataset will be stored in `Image-Vid-Finetune-Mochi` or `Image-Vid-Finetune-HunYuan` correspondingly.
 
-### Process your own dataset
+## 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:
 
+```
 path_to_dataset_folder/
 ├── media/
 │   ├── 0.jpg
@@ -29,17 +31,21 @@ path_to_dataset_folder/
 │   ├── 2.jpg
 ├── video2caption.json
 └── merge.txt
+```
 
 Format the JSON file as a list, where each item represents a media source:
 
 For image media,
+
 ```
 {
     "path": "0.jpg",
     "cap": ["captions"]
 }
 ```
-For video media, 
+
+For video media,
+
 ```
 {
     "path": "1.mp4",
@@ -62,7 +68,9 @@ 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
 ```
+
 The preprocessed data will be put into the `OUTPUT_DIR` and the `videos2caption.json` can be used in finetune and distill scripts.

docs/source/training/distillation.md

@@ -0,0 +1,25 @@
+(v0-distill)=
+# 🎯 Distill
+Our distillation recipe is based on [Phased Consistency Model](https://github.com/G-U-N/Phased-Consistency-Model). We did not find significant improvement using multi-phase distillation, so we keep the one phase setup similar to the original latent consistency model's recipe.
+We use the [MixKit](https://huggingface.co/datasets/LanguageBind/Open-Sora-Plan-v1.1.0/tree/main/all_mixkit) dataset for distillation. To avoid running the text encoder and VAE during training, we prprocess all data to generate text embeddings and VAE latents.
+Preprocessing instructions can be found [data_preprocess.md](#v0-data-preprocess). For convenience, we also provide preprocessed data that can be downloaded directly using the following command:
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/HD-Mixkit-Finetune-Hunyuan --local_dir=data/HD-Mixkit-Finetune-Hunyuan --repo_type=dataset
+```
+
+Next, download the original model weights with:
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/hunyuan --local_dir=data/hunyuan --repo_type=model # original hunyuan
+python scripts/huggingface/download_hf.py --repo_id=genmo/mochi-1-preview --local_dir=data/mochi --repo_type=model # original mochi
+```
+
+To launch the distillation process, use the following commands:
+
+```
+bash scripts/distill/distill_hunyuan.sh # for hunyuan
+bash scripts/distill/distill_mochi.sh # for mochi
+```
+
+We also provide an optional script for distillation with adversarial loss, located at `fastvideo/distill_adv.py`. Although we tried adversarial loss, we did not observe significant improvements.

docs/source/training/finetune.md

@@ -0,0 +1,71 @@
+(v0-finetune)=
+# 🧠 Finetune
+## ⚡ Full Finetune
+Ensure your data is prepared and preprocessed in the format specified in [data_preprocess.md](#v0-data-preprocess). For convenience, we also provide a mochi preprocessed Black Myth Wukong data that can be downloaded directly:
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/Mochi-Black-Myth --local_dir=data/Mochi-Black-Myth --repo_type=dataset
+```
+
+Download the original model weights as specified in [Distill Section](#v0-distill):
+
+Then you can run the finetune with:
+
+```
+bash scripts/finetune/finetune_mochi.sh # for mochi
+```
+
+**Note that for finetuning, we did not tune the hyperparameters in the provided script.**
+## ⚡ 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:
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/Hunyuan-Black-Myth-Wukong-lora-weight --local_dir=data/Hunyuan-Black-Myth-Wukong-lora-weight --repo_type=model
+```
+
+### Minimum Hardware Requirement
+- 40 GB GPU memory each for 2 GPUs with lora.
+- 30 GB GPU memory each for 2 GPUs with CPU offload and lora.
+
+Currently, both Mochi and Hunyuan models support Lora finetuning through diffusers. To generate personalized videos from your own dataset, you'll need to follow three main steps: dataset preparation, finetuning, and inference.
+
+### Dataset Preparation
+We provide scripts to better help you get started to train on your own characters!
+You can run this to organize your dataset to get the videos2caption.json before preprocess. Specify your video folder and corresponding caption folder (caption files should be .txt files and have the same name with its video):
+
+```
+python scripts/dataset_preparation/prepare_json_file.py --video_dir data/input_videos/ --prompt_dir data/captions/ --output_path data/output_folder/videos2caption.json --verbose
+```
+
+Also, we provide script to resize your videos:
+
+```
+python scripts/data_preprocess/resize_videos.py
+```
+
+### Finetuning
+After basic dataset preparation and preprocess, you can start to finetune your model using Lora:
+
+```
+bash scripts/finetune/finetune_hunyuan_hf_lora.sh
+```
+
+### Inference
+For inference with Lora checkpoint, you can run the following scripts with additional parameter `--lora_checkpoint_dir`:
+
+```
+bash scripts/inference/inference_hunyuan_hf.sh
+```
+
+**We also provide scripts for Mochi in the same directory.**
+
+### Finetune with Both Image and Video
+Our codebase support finetuning with both image and video.
+
+```bash
+bash scripts/finetune/finetune_hunyuan.sh
+bash scripts/finetune/finetune_mochi_lora_mix.sh
+```
+
+For Image-Video Mixture Fine-tuning, make sure to enable the `--group_frame` option in your script.

env_setup.sh

@@ -1,12 +0,0 @@
-#!/bin/bash
-
-# install torch
-pip install torch==2.5.0 torchvision --index-url https://download.pytorch.org/whl/cu121
-
-# install FA2 and diffusers
-pip install packaging ninja && pip install flash-attn==2.7.0.post2 --no-build-isolation 
-
-pip install -r requirements-lint.txt
-
-# install fastvideo
-pip install -e .

examples/inference/basic/README.md

@@ -0,0 +1,41 @@
+# Basic Video Generation Tutorial
+The `VideoGenerator` class provides the primary Python interface for doing offline video generation, which is interacting with a diffusion pipeline without using a separate inference api server.
+
+## Requirements
+- At least a single NVIDIA GPU with CUDA 12.4.
+- Python 3.10-3.12
+
+## Installation
+If you have not installed FastVideo, please following these [instructions](https://hao-ai-lab.github.io/FastVideo/getting_started/installation.html) first.
+
+## Usage
+The first script in this example shows the most basic usage of FastVideo. If you are new to Python and FastVideo, you should start here.
+
+```bash
+# if you have not cloned the directory:
+git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
+
+python examples/inference/basic/basic.py
+```
+
+## Basic Walkthrough
+
+All you need to generate videos using multi-gpus from state-of-the-art diffusion pipelines is the following few lines!
+
+```python
+from fastvideo import VideoGenerator
+
+def main():
+    generator = VideoGenerator.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        num_gpus=1,
+    )
+
+    prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+             "wide with interest. The playful yet serene atmosphere is complemented by soft "
+             "natural light filtering through the petals. Mid-shot, warm and cheerful tones.")
+    video = generator.generate_video(prompt)
+
+if __name__ == "__main__":
+    main()
+```

examples/inference/basic/basic.py

@@ -0,0 +1,41 @@
+from fastvideo import VideoGenerator
+
+# from fastvideo.v1.configs.sample import SamplingParam
+
+
+def main():
+    # FastVideo will automatically use the optimal default arguments for the
+    # model.
+    # If a local path is provided, FastVideo will make a best effort
+    # attempt to identify the optimal arguments.
+    generator = VideoGenerator.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        # if num_gpus > 1, FastVideo will automatically handle distributed setup
+        num_gpus=1,
+    )
+
+    # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
+    # sampling_param.num_frames = 45
+    # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+    # Generate videos with the same simple API, regardless of GPU count
+    prompt = (
+        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+        "wide with interest. The playful yet serene atmosphere is complemented by soft "
+        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
+    )
+    video = generator.generate_video(prompt)
+    # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/")
+
+    # Generate another video with a different prompt, without reloading the
+    # model!
+    prompt2 = (
+        "A majestic lion strides across the golden savanna, its powerful frame "
+        "glistening under the warm afternoon sun. The tall grass ripples gently in "
+        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
+        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
+        "cinematic.")
+    video2 = generator.generate_video(prompt2)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/default_args.py

@@ -0,0 +1,62 @@
+from fastvideo import VideoGenerator
+
+
+def main():
+
+    # This is the config class for the model initialization
+    config = PipelineConfig.from_pretrained("FastVideo/FastHunyuan-Diffusers")
+    # can be used to dump the config to a yaml file
+    config.dump_to_yaml("config.yaml")
+    print(config)
+    # {
+    #    'vae_config': {
+    #        'scale_factor': 8,
+    #        'sp': True,
+    #        'tiling': True,
+    #        'precision': 'fp16'
+    #    },
+    #    'text_encoder_config': {
+    #        'precision': 'fp16'
+    #    },
+    #    'dit_config': {
+    #        'precision': 'fp16'
+    #    },
+    #    'inference_args': {
+    #        'guidance_scale': 7.5,
+    #        'num_inference_steps': 5,
+    #        'seed': 1024,
+    #        'guidance_rescale': 0.0,
+    #        'flow_shift': 17,
+    #        'num_inference_steps': 5,
+    #    }
+    # }
+
+    config.vae_config.scale_factor = 16
+
+    # FastVideo will automatically used the optimal default arguments for the model
+    # If a local path is provided, FastVideo will make a best effort attempt to
+    # identify the optimal arguments.
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/FastHunyuan-Diffusers",
+        num_gpus=4,
+        config=config,
+        # or
+        config_path="config.yaml",
+    )
+
+    sampling_param = SamplingParam.from_pretrained(
+        "FastVideo/FastHunyuan-Diffusers")
+    sampling_param.num_inference_steps = 5
+
+    # Generate videos with the same simple API, regardless of GPU count
+    prompt = "A beautiful woman in a red dress walking down a street"
+    video = generator.generate_video(prompt,
+                                     sampling_param=sampling_param,
+                                     num_inference_steps=6)
+
+    video2 = generator.generate_video(prompt2)
+    prompt2 = "A beautiful woman in a blue dress walking down a street"
+
+
+if __name__ == "__main__":
+    main()

examples/inference/gradio/README.md

@@ -0,0 +1,59 @@
+# FastVideo Gradio Demo
+
+This is a Gradio-based web interface for generating videos using the FastVideo framework. The demo allows users to create videos from text prompts with various customization options.
+
+## Overview
+
+The demo uses the FastVideo framework to generate videos based on text prompts. It provides a simple web interface built with Gradio that allows users to:
+
+- Enter text prompts to generate videos
+- Customize video parameters (dimensions, number of frames, etc.)
+- Use negative prompts to guide the generation process
+- Set or randomize seeds for reproducibility
+
+---
+
+## Usage
+
+Run the demo with:
+
+```bash
+python examples/inference/gradio/gradio_demo.py
+```
+
+This will start a web server at `http://0.0.0.0:7860` where you can access the interface.
+
+---
+
+## Model Initialization
+
+This demo initializes a `VideoGenerator` with the minimum required arguments for inference. Users can seamlessly adjust inference options between generations, including prompts, resolution, video length, or even the number of inference steps, *without ever needing to reload the model*.
+
+## Video Generation
+
+The core functionality is in the `generate_video` function, which:
+1. Processes user inputs
+2. Uses the FastVideo VideoGenerator from earlier to run inference (`generator.generate_video()`)
+3. Returns an output path that Gradio uses to display the generated video
+
+## Gradio Interface
+
+The interface is built with several components:
+- A text input for the prompt
+- A video display for the result
+- Inference options in a collapsible accordion:
+  - Height and width sliders
+  - Number of frames slider
+  - Guidance scale slider
+  - Inference steps slider
+  - Negative prompt options
+  - Seed controls
+
+### Inference Options
+
+- **Height/Width**: Control the resolution of the generated video
+- **Number of Frames**: Set how many frames to generate
+- **Guidance Scale**: Control how closely the generation follows the prompt
+- **Inference Steps**: More steps can improve quality but take longer
+- **Negative Prompt**: Specify what you don't want to see in the video
+- **Seed**: Control randomness for reproducible results

examples/inference/gradio/gradio_demo.py

@@ -0,0 +1,169 @@
+import argparse
+import os
+from copy import deepcopy
+
+import gradio as gr
+import torch
+
+from fastvideo import VideoGenerator
+from fastvideo.v1.configs.sample.base import SamplingParam
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="FastVideo Gradio Demo")
+    parser.add_argument("--model_path",
+                        type=str,
+                        default="FastVideo/FastHunyuan-diffusers",
+                        help="Path to the model")
+    parser.add_argument("--num_gpus",
+                        type=int,
+                        default=1,
+                        help="Number of GPUs to use")
+    parser.add_argument("--output_path",
+                        type=str,
+                        default="outputs",
+                        help="Path to save generated videos")
+    parsed_args = parser.parse_args()
+
+    # args = FastVideoArgs(model_path="FastVideo/FastHunyuan-Diffusers", num_gpus=2)
+
+    generator = VideoGenerator.from_pretrained(
+        model_path=parsed_args.model_path, num_gpus=parsed_args.num_gpus)
+
+    default_params = SamplingParam.from_pretrained(parsed_args.model_path)
+
+    def generate_video(
+        prompt,
+        negative_prompt,
+        use_negative_prompt,
+        seed,
+        guidance_scale,
+        num_frames,
+        height,
+        width,
+        num_inference_steps,
+        randomize_seed=False,
+    ):
+        params = deepcopy(default_params)
+        params.prompt = prompt
+        params.negative_prompt = negative_prompt
+        params.seed = seed
+        params.guidance_scale = guidance_scale
+        params.num_frames = num_frames
+        params.height = height
+        params.width = width
+        params.num_inference_steps = num_inference_steps
+
+        if randomize_seed:
+            params.seed = torch.randint(0, 1000000, (1, )).item()
+
+        if not use_negative_prompt:
+            params.negative_prompt = None
+
+        generator.generate_video(prompt=prompt, sampling_param=params)
+
+        output_path = os.path.join(parsed_args.output_path,
+                                   f"{params.prompt[:100]}.mp4")
+
+        return output_path, params.seed
+
+    examples = [
+        "A hand enters the frame, pulling a sheet of plastic wrap over three balls of dough placed on a wooden surface. The plastic wrap is stretched to cover the dough more securely. The hand adjusts the wrap, ensuring that it is tight and smooth over the dough. The scene focuses on the hand’s movements as it secures the edges of the plastic wrap. No new objects appear, and the camera remains stationary, focusing on the action of covering the dough.",
+        "A vintage train snakes through the mountains, its plume of white steam rising dramatically against the jagged peaks. The cars glint in the late afternoon sun, their deep crimson and gold accents lending a touch of elegance. The tracks carve a precarious path along the cliffside, revealing glimpses of a roaring river far below. Inside, passengers peer out the large windows, their faces lit with awe as the landscape unfolds.",
+        "A crowded rooftop bar buzzes with energy, the city skyline twinkling like a field of stars in the background. Strings of fairy lights hang above, casting a warm, golden glow over the scene. Groups of people gather around high tables, their laughter blending with the soft rhythm of live jazz. The aroma of freshly mixed cocktails and charred appetizers wafts through the air, mingling with the cool night breeze.",
+    ]
+
+    with gr.Blocks() as demo:
+        gr.Markdown("# FastVideo Inference Demo")
+
+        with gr.Group():
+            with gr.Row():
+                prompt = gr.Text(
+                    label="Prompt",
+                    show_label=False,
+                    max_lines=1,
+                    placeholder="Enter your prompt",
+                    container=False,
+                )
+                run_button = gr.Button("Run", scale=0)
+            result = gr.Video(label="Result", show_label=False)
+
+        with gr.Accordion("Advanced options", open=False):
+            with gr.Group():
+                with gr.Row():
+                    height = gr.Slider(
+                        label="Height",
+                        minimum=256,
+                        maximum=1024,
+                        step=32,
+                        value=default_params.height,
+                    )
+                    width = gr.Slider(label="Width",
+                                      minimum=256,
+                                      maximum=1024,
+                                      step=32,
+                                      value=default_params.width)
+
+                with gr.Row():
+                    num_frames = gr.Slider(
+                        label="Number of Frames",
+                        minimum=21,
+                        maximum=163,
+                        value=default_params.num_frames,
+                    )
+                    guidance_scale = gr.Slider(
+                        label="Guidance Scale",
+                        minimum=1,
+                        maximum=12,
+                        value=default_params.guidance_scale,
+                    )
+                    num_inference_steps = gr.Slider(
+                        label="Inference Steps",
+                        minimum=4,
+                        maximum=100,
+                        value=default_params.num_inference_steps,
+                    )
+
+                with gr.Row():
+                    use_negative_prompt = gr.Checkbox(
+                        label="Use negative prompt", value=False)
+                    negative_prompt = gr.Text(
+                        label="Negative prompt",
+                        max_lines=1,
+                        placeholder="Enter a negative prompt",
+                        visible=False,
+                    )
+
+                seed = gr.Slider(label="Seed",
+                                 minimum=0,
+                                 maximum=1000000,
+                                 step=1,
+                                 value=default_params.seed)
+                randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+                seed_output = gr.Number(label="Used Seed")
+
+        gr.Examples(examples=examples, inputs=prompt)
+
+        use_negative_prompt.change(
+            fn=lambda x: gr.update(visible=x),
+            inputs=use_negative_prompt,
+            outputs=default_params.negative_prompt,
+        )
+
+        run_button.click(
+            fn=generate_video,
+            inputs=[
+                prompt,
+                negative_prompt,
+                use_negative_prompt,
+                seed,
+                guidance_scale,
+                num_frames,
+                height,
+                width,
+                num_inference_steps,
+                randomize_seed,
+            ],
+            outputs=[result, seed_output],
+        )
+
+    demo.queue(max_size=20).launch(server_name="0.0.0.0", server_port=7860)

examples/inference/optimizations/README.md

@@ -0,0 +1,9 @@
+# Optimization Examples
+
+```bash
+python examples/inference/optimizations/attention_example.py
+```
+
+```bash
+python examples/inference/optimizations/teacache_example.py
+```

examples/inference/optimizations/attention_example.py

@@ -0,0 +1,33 @@
+import os
+import time
+
+from fastvideo import VideoGenerator
+
+def main():
+    # set the attention backend 
+    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "FLASH_ATTN"
+
+    start_time = time.perf_counter()
+    gen = VideoGenerator.from_pretrained(
+        model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        num_gpus=1,
+    )
+    load_time = time.perf_counter() - start_time
+    print(f"Model loading time: {load_time:.2f} seconds")
+
+    gen_start_time = time.perf_counter()
+
+    gen.generate_video(
+        prompt=
+        "Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of a bustling street food market. The scene captures a mix of humor and authenticity. Mid-shot framing, vibrant lighting.",
+        seed=1024,
+        output_path="example_outputs/")
+    
+    generation_time = time.perf_counter() - gen_start_time
+    print(f"Video generation time: {generation_time:.2f} seconds")
+
+    total_time = time.perf_counter() - start_time
+    print(f"Total execution time: {total_time:.2f} seconds")
+
+if __name__ == "__main__":
+    main()

examples/inference/optimizations/teacache_example.py

@@ -0,0 +1,44 @@
+import time
+
+from fastvideo import VideoGenerator, SamplingParam
+
+
+def main():
+    start_time = time.perf_counter()
+
+    gen = VideoGenerator.from_pretrained(
+        model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        num_gpus=1,
+        use_cpu_offload=False,
+    )
+    load_time = time.perf_counter() - start_time
+    print(f"Model loading time: {load_time:.2f} seconds")
+
+    gen_start_time = time.perf_counter()
+
+    params = SamplingParam.from_pretrained(
+        model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+    )
+    # this controls the threshold for the tea cache
+    params.teacache_params.teacache_thresh = 0.08
+    gen.generate_video(
+        prompt=
+        "Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of a bustling street food market. The scene captures a mix of humor and authenticity. Mid-shot framing, vibrant lighting.",
+        sampling_param=params,
+        height=480,
+        width=832,
+        num_frames=61,  # 85 ,77 
+        num_inference_steps=50,
+        enable_teacache=True,
+        seed=1024,
+        output_path="example_outputs/")
+    
+    generation_time = time.perf_counter() - gen_start_time
+    print(f"Video generation time: {generation_time:.2f} seconds")
+
+    total_time = time.perf_counter() - start_time
+    print(f"Total execution time: {total_time:.2f} seconds")
+
+
+if __name__ == "__main__":
+    main()

fastvideo/__init__.py

@@ -0,0 +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", "__version__"]

fastvideo/data_preprocess/preprocess.py

@@ -0,0 +1,111 @@
+import argparse
+import json
+import os
+
+import torch
+import torch.distributed as dist
+
+from fastvideo.v1.logger import init_logger
+from fastvideo.v1.utils import maybe_download_model, shallow_asdict
+from fastvideo.v1.distributed import init_distributed_environment, initialize_model_parallel
+from fastvideo.v1.fastvideo_args import FastVideoArgs
+from fastvideo.v1.configs.models.vaes import WanVAEConfig
+from fastvideo import PipelineConfig
+from fastvideo.v1.pipelines.preprocess_pipeline import PreprocessPipeline
+
+logger = init_logger(__name__)
+
+BASE_MODEL_PATH = "/workspace/data/Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
+                                  local_dir=os.path.join(
+                                      'data', BASE_MODEL_PATH))
+
+def main(args):
+    # Assume using torchrun
+    local_rank = int(os.getenv("RANK", 0))
+    rank = int(os.environ.get("RANK", 0))
+    world_size = int(os.getenv("WORLD_SIZE", 1))
+    init_distributed_environment(world_size=world_size, rank=rank, local_rank=local_rank)
+    initialize_model_parallel(tensor_model_parallel_size=world_size, sequence_model_parallel_size=world_size)
+    torch.cuda.set_device(local_rank)
+    if not dist.is_initialized():
+        dist.init_process_group(backend="nccl", init_method="env://", world_size=world_size, rank=local_rank)
+
+    pipeline_config = PipelineConfig.from_pretrained(MODEL_PATH)
+    kwargs = {
+        "use_cpu_offload": False,
+        "vae_precision": "fp32",
+        "vae_config": WanVAEConfig(load_encoder=True, load_decoder=False),
+    }
+    pipeline_config_args = shallow_asdict(pipeline_config)
+    pipeline_config_args.update(kwargs)
+    fastvideo_args = FastVideoArgs(model_path=MODEL_PATH,
+                                   num_gpus=world_size,
+                                   device_str="cuda",
+                                   **pipeline_config_args,
+                                   )
+    fastvideo_args.check_fastvideo_args()
+    fastvideo_args.device = torch.device(f"cuda:{local_rank}")
+
+    pipeline = PreprocessPipeline(MODEL_PATH, fastvideo_args)
+    pipeline.forward(batch=None, fastvideo_args=fastvideo_args, args=args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # dataset & dataloader
+    parser.add_argument("--model_path", type=str, default="data/mochi")
+    parser.add_argument("--model_type", type=str, default="mochi")
+    parser.add_argument("--data_merge_path", type=str, required=True)
+    parser.add_argument("--validation_prompt_txt", type=str)
+    parser.add_argument("--num_frames", type=int, default=163)
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=1,
+        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
+    )
+    parser.add_argument(
+        "--preprocess_video_batch_size",
+        type=int,
+        default=2,
+        help="Batch size (per device) for the training dataloader.",
+    )
+    parser.add_argument(
+        "--preprocess_text_batch_size",
+        type=int,
+        default=8,
+        help="Batch size (per device) for the training dataloader.",
+    )
+    parser.add_argument("--num_latent_t", type=int, default=28, help="Number of latent timesteps.")
+    parser.add_argument("--max_height", type=int, default=480)
+    parser.add_argument("--max_width", type=int, default=848)
+    parser.add_argument("--video_length_tolerance_range", type=int, default=2.0)
+    parser.add_argument("--group_frame", action="store_true")  # TODO
+    parser.add_argument("--group_resolution", action="store_true")  # TODO
+    parser.add_argument("--dataset", default="t2v")
+    parser.add_argument("--train_fps", type=int, default=30)
+    parser.add_argument("--use_image_num", type=int, default=0)
+    parser.add_argument("--text_max_length", type=int, default=256)
+    parser.add_argument("--speed_factor", type=float, default=1.0)
+    parser.add_argument("--drop_short_ratio", type=float, default=1.0)
+    # text encoder & vae & diffusion model
+    parser.add_argument("--text_encoder_name", type=str, default="google/t5-v1_1-xxl")
+    parser.add_argument("--cache_dir", type=str, default="./cache_dir")
+    parser.add_argument("--cfg", type=float, default=0.0)
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        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)

fastvideo/data_preprocess/preprocess_text_embeddings.py

@@ -27,8 +27,7 @@ class T5dataset(Dataset):
         self.vae_debug = vae_debug
         with open(self.json_path, "r") as f:
             train_dataset = json.load(f)
-            self.train_dataset = sorted(train_dataset,
-                                        key=lambda x: x["latent_path"])
+            self.train_dataset = sorted(train_dataset, key=lambda x: x["latent_path"])
 
     def __getitem__(self, idx):
         caption = self.train_dataset[idx]["caption"]
@@ -36,17 +35,13 @@ class T5dataset(Dataset):
         length = self.train_dataset[idx]["length"]
         if self.vae_debug:
             latents = torch.load(
-                os.path.join(args.output_dir, "latent",
-                             self.train_dataset[idx]["latent_path"]),
+                os.path.join(args.output_dir, "latent", self.train_dataset[idx]["latent_path"]),
                 map_location="cpu",
             )
         else:
             latents = []
 
-        return dict(caption=caption,
-                    latents=latents,
-                    filename=filename,
-                    length=length)
+        return dict(caption=caption, latents=latents, filename=filename, length=length)
 
     def __len__(self):
         return len(self.train_dataset)
@@ -60,31 +55,21 @@ def main(args):
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     torch.cuda.set_device(local_rank)
     if not dist.is_initialized():
-        dist.init_process_group(backend="nccl",
-                                init_method="env://",
-                                world_size=world_size,
-                                rank=local_rank)
+        dist.init_process_group(backend="nccl", init_method="env://", world_size=world_size, rank=local_rank)
 
     videoprocessor = VideoProcessor(vae_scale_factor=8)
     os.makedirs(args.output_dir, exist_ok=True)
     os.makedirs(os.path.join(args.output_dir, "video"), exist_ok=True)
     os.makedirs(os.path.join(args.output_dir, "latent"), exist_ok=True)
     os.makedirs(os.path.join(args.output_dir, "prompt_embed"), exist_ok=True)
-    os.makedirs(os.path.join(args.output_dir, "prompt_attention_mask"),
-                exist_ok=True)
+    os.makedirs(os.path.join(args.output_dir, "prompt_attention_mask"), exist_ok=True)
 
-    latents_json_path = os.path.join(args.output_dir,
-                                     "videos2caption_temp.json")
+    latents_json_path = os.path.join(args.output_dir, "videos2caption_temp.json")
     train_dataset = T5dataset(latents_json_path, args.vae_debug)
-    text_encoder = load_text_encoder(args.model_type,
-                                     args.model_path,
-                                     device=device)
+    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()
-    sampler = DistributedSampler(train_dataset,
-                                 rank=local_rank,
-                                 num_replicas=world_size,
-                                 shuffle=True)
+    sampler = DistributedSampler(train_dataset, rank=local_rank, num_replicas=world_size, shuffle=True)
     train_dataloader = DataLoader(
         train_dataset,
         sampler=sampler,
@@ -96,26 +81,19 @@ def main(args):
     for _, data in tqdm(enumerate(train_dataloader), disable=local_rank != 0):
         with torch.inference_mode():
             with torch.autocast("cuda", dtype=autocast_type):
-                prompt_embeds, prompt_attention_mask = text_encoder.encode_prompt(
-                    prompt=data["caption"], )
+                prompt_embeds, prompt_attention_mask = text_encoder.encode_prompt(prompt=data["caption"], )
                 if args.vae_debug:
                     latents = data["latents"]
-                    video = vae.decode(latents.to(device),
-                                       return_dict=False)[0]
+                    video = vae.decode(latents.to(device), return_dict=False)[0]
                     video = videoprocessor.postprocess_video(video)
                 for idx, video_name in enumerate(data["filename"]):
-                    prompt_embed_path = os.path.join(args.output_dir,
-                                                     "prompt_embed",
-                                                     video_name + ".pt")
-                    video_path = os.path.join(args.output_dir, "video",
-                                              video_name + ".mp4")
-                    prompt_attention_mask_path = os.path.join(
-                        args.output_dir, "prompt_attention_mask",
-                        video_name + ".pt")
+                    prompt_embed_path = os.path.join(args.output_dir, "prompt_embed", video_name + ".pt")
+                    video_path = os.path.join(args.output_dir, "video", video_name + ".mp4")
+                    prompt_attention_mask_path = os.path.join(args.output_dir, "prompt_attention_mask",
+                                                              video_name + ".pt")
                     # save latent
                     torch.save(prompt_embeds[idx], prompt_embed_path)
-                    torch.save(prompt_attention_mask[idx],
-                               prompt_attention_mask_path)
+                    torch.save(prompt_attention_mask[idx], prompt_attention_mask_path)
                     print(f"sample {video_name} saved")
                     if args.vae_debug:
                         export_to_video(video[idx], video_path, fps=fps)
@@ -133,8 +111,7 @@ def main(args):
     if local_rank == 0:
         # os.remove(latents_json_path)
         all_json_data = [item for sublist in gathered_data for item in sublist]
-        with open(os.path.join(args.output_dir, "videos2caption.json"),
-                  "w") as f:
+        with open(os.path.join(args.output_dir, "videos2caption.json"), "w") as f:
             json.dump(all_json_data, f, indent=4)
 
 
@@ -148,8 +125,7 @@ if __name__ == "__main__":
         "--dataloader_num_workers",
         type=int,
         default=1,
-        help=
-        "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
+        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
     )
     parser.add_argument(
         "--train_batch_size",
@@ -157,16 +133,13 @@ if __name__ == "__main__":
         default=1,
         help="Batch size (per device) for the training dataloader.",
     )
-    parser.add_argument("--text_encoder_name",
-                        type=str,
-                        default="google/t5-v1_1-xxl")
+    parser.add_argument("--text_encoder_name", type=str, default="google/t5-v1_1-xxl")
     parser.add_argument("--cache_dir", type=str, default="./cache_dir")
     parser.add_argument(
         "--output_dir",
         type=str,
         default=None,
-        help=
-        "The output directory where the model predictions and checkpoints will be written.",
+        help="The output directory where the model predictions and checkpoints will be written.",
     )
     parser.add_argument("--vae_debug", action="store_true")
     args = parser.parse_args()

fastvideo/data_preprocess/preprocess_vae_latents.py

@@ -20,10 +20,7 @@ def main(args):
     world_size = int(os.getenv("WORLD_SIZE", 1))
     print("world_size", world_size, "local rank", local_rank)
     train_dataset = getdataset(args)
-    sampler = DistributedSampler(train_dataset,
-                                 rank=local_rank,
-                                 num_replicas=world_size,
-                                 shuffle=True)
+    sampler = DistributedSampler(train_dataset, rank=local_rank, num_replicas=world_size, shuffle=True)
     train_dataloader = DataLoader(
         train_dataset,
         sampler=sampler,
@@ -31,14 +28,10 @@ def main(args):
         num_workers=args.dataloader_num_workers,
     )
 
-    encoder_device = torch.device(
-        "cuda" if torch.cuda.is_available() else "cpu")
+    encoder_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     torch.cuda.set_device(local_rank)
     if not dist.is_initialized():
-        dist.init_process_group(backend="nccl",
-                                init_method="env://",
-                                world_size=world_size,
-                                rank=local_rank)
+        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()
     os.makedirs(args.output_dir, exist_ok=True)
@@ -48,12 +41,10 @@ def main(args):
     for _, data in tqdm(enumerate(train_dataloader), disable=local_rank != 0):
         with torch.inference_mode():
             with torch.autocast("cuda", dtype=autocast_type):
-                latents = vae.encode(data["pixel_values"].to(
-                    encoder_device))["latent_dist"].sample()
+                latents = vae.encode(data["pixel_values"].to(encoder_device))["latent_dist"].sample()
             for idx, video_path in enumerate(data["path"]):
                 video_name = os.path.basename(video_path).split(".")[0]
-                latent_path = os.path.join(args.output_dir, "latent",
-                                           video_name + ".pt")
+                latent_path = os.path.join(args.output_dir, "latent", video_name + ".pt")
                 torch.save(latents[idx].to(torch.bfloat16), latent_path)
                 item = {}
                 item["length"] = latents[idx].shape[1]
@@ -67,8 +58,7 @@ def main(args):
     dist.all_gather_object(gathered_data, local_data)
     if local_rank == 0:
         all_json_data = [item for sublist in gathered_data for item in sublist]
-        with open(os.path.join(args.output_dir, "videos2caption_temp.json"),
-                  "w") as f:
+        with open(os.path.join(args.output_dir, "videos2caption_temp.json"), "w") as f:
             json.dump(all_json_data, f, indent=4)
 
 
@@ -83,8 +73,7 @@ if __name__ == "__main__":
         "--dataloader_num_workers",
         type=int,
         default=1,
-        help=
-        "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
+        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
     )
     parser.add_argument(
         "--train_batch_size",
@@ -92,15 +81,10 @@ if __name__ == "__main__":
         default=16,
         help="Batch size (per device) for the training dataloader.",
     )
-    parser.add_argument("--num_latent_t",
-                        type=int,
-                        default=28,
-                        help="Number of latent timesteps.")
+    parser.add_argument("--num_latent_t", type=int, default=28, help="Number of latent timesteps.")
     parser.add_argument("--max_height", type=int, default=480)
     parser.add_argument("--max_width", type=int, default=848)
-    parser.add_argument("--video_length_tolerance_range",
-                        type=int,
-                        default=2.0)
+    parser.add_argument("--video_length_tolerance_range", type=int, default=2.0)
     parser.add_argument("--group_frame", action="store_true")  # TODO
     parser.add_argument("--group_resolution", action="store_true")  # TODO
     parser.add_argument("--dataset", default="t2v")
@@ -110,25 +94,21 @@ if __name__ == "__main__":
     parser.add_argument("--speed_factor", type=float, default=1.0)
     parser.add_argument("--drop_short_ratio", type=float, default=1.0)
     # text encoder & vae & diffusion model
-    parser.add_argument("--text_encoder_name",
-                        type=str,
-                        default="google/t5-v1_1-xxl")
+    parser.add_argument("--text_encoder_name", type=str, default="google/t5-v1_1-xxl")
     parser.add_argument("--cache_dir", type=str, default="./cache_dir")
     parser.add_argument("--cfg", type=float, default=0.0)
     parser.add_argument(
         "--output_dir",
         type=str,
         default=None,
-        help=
-        "The output directory where the model predictions and checkpoints will be written.",
+        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***."),
+        help=("[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+              " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."),
     )
 
     args = parser.parse_args()

fastvideo/data_preprocess/preprocess_validation_text_embeddings.py

@@ -18,14 +18,9 @@ def main(args):
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     torch.cuda.set_device(local_rank)
     if not dist.is_initialized():
-        dist.init_process_group(backend="nccl",
-                                init_method="env://",
-                                world_size=world_size,
-                                rank=local_rank)
+        dist.init_process_group(backend="nccl", init_method="env://", world_size=world_size, rank=local_rank)
 
-    text_encoder = load_text_encoder(args.model_type,
-                                     args.model_path,
-                                     device=device)
+    text_encoder = load_text_encoder(args.model_type, args.model_path, device=device)
     autocast_type = torch.float16 if args.model_type == "hunyuan" else torch.bfloat16
     # output_dir/validation/prompt_attention_mask
     # output_dir/validation/prompt_embed
@@ -34,8 +29,7 @@ def main(args):
         os.path.join(args.output_dir, "validation", "prompt_attention_mask"),
         exist_ok=True,
     )
-    os.makedirs(os.path.join(args.output_dir, "validation", "prompt_embed"),
-                exist_ok=True)
+    os.makedirs(os.path.join(args.output_dir, "validation", "prompt_embed"), exist_ok=True)
 
     with open(args.validation_prompt_txt, "r", encoding="utf-8") as file:
         lines = file.readlines()
@@ -43,12 +37,9 @@ def main(args):
     for prompt in prompts:
         with torch.inference_mode():
             with torch.autocast("cuda", dtype=autocast_type):
-                prompt_embeds, prompt_attention_mask = text_encoder.encode_prompt(
-                    prompt)
+                prompt_embeds, prompt_attention_mask = text_encoder.encode_prompt(prompt)
                 file_name = prompt.split(".")[0]
-                prompt_embed_path = os.path.join(args.output_dir, "validation",
-                                                 "prompt_embed",
-                                                 f"{file_name}.pt")
+                prompt_embed_path = os.path.join(args.output_dir, "validation", "prompt_embed", f"{file_name}.pt")
                 prompt_attention_mask_path = os.path.join(
                     args.output_dir,
                     "validation",
@@ -56,8 +47,7 @@ def main(args):
                     f"{file_name}.pt",
                 )
                 torch.save(prompt_embeds[0], prompt_embed_path)
-                torch.save(prompt_attention_mask[0],
-                           prompt_attention_mask_path)
+                torch.save(prompt_attention_mask[0], prompt_attention_mask_path)
                 print(f"sample {file_name} saved")
 
 
@@ -71,8 +61,7 @@ if __name__ == "__main__":
         "--output_dir",
         type=str,
         default=None,
-        help=
-        "The output directory where the model predictions and checkpoints will be written.",
+        help="The output directory where the model predictions and checkpoints will be written.",
     )
     args = parser.parse_args()
     main(args)

fastvideo/dataset/__init__.py

@@ -3,16 +3,14 @@ from torchvision.transforms import Lambda
 from transformers import AutoTokenizer
 
 from fastvideo.dataset.t2v_datasets import T2V_dataset
-from fastvideo.dataset.transform import (CenterCropResizeVideo, Normalize255,
-                                         TemporalRandomCrop)
+from fastvideo.dataset.transform import CenterCropResizeVideo, Normalize255, TemporalRandomCrop
 
 
 def getdataset(args):
     temporal_sample = TemporalRandomCrop(args.num_frames)  # 16 x
     norm_fun = Lambda(lambda x: 2.0 * x - 1.0)
     resize_topcrop = [
-        CenterCropResizeVideo((args.max_height, args.max_width),
-                              top_crop=True),
+        CenterCropResizeVideo((args.max_height, args.max_width), top_crop=True),
     ]
     resize = [
         CenterCropResizeVideo((args.max_height, args.max_width)),
@@ -27,8 +25,7 @@ def getdataset(args):
         norm_fun,
     ])
     # tokenizer = AutoTokenizer.from_pretrained("/storage/ongoing/new/Open-Sora-Plan/cache_dir/mt5-xxl", cache_dir=args.cache_dir)
-    tokenizer = AutoTokenizer.from_pretrained(args.text_encoder_name,
-                                              cache_dir=args.cache_dir)
+    tokenizer = AutoTokenizer.from_pretrained(args.text_encoder_name, cache_dir=args.cache_dir)
     if args.dataset == "t2v":
         return T2V_dataset(
             args,
@@ -66,8 +63,7 @@ if __name__ == "__main__":
             "interpolation_scale_h": 1,
             "interpolation_scale_w": 1,
             "cache_dir": "../cache_dir",
-            "image_data":
-            "/storage/ongoing/new/Open-Sora-Plan-bak/7.14bak/scripts/train_data/image_data.txt",
+            "image_data": "/storage/ongoing/new/Open-Sora-Plan-bak/7.14bak/scripts/train_data/image_data.txt",
             "video_data": "1",
             "train_fps": 24,
             "drop_short_ratio": 1.0,
@@ -84,10 +80,7 @@ if __name__ == "__main__":
     zero = 0
     for idx in tqdm(range(num)):
         image_data = dataset_prog.img_cap_list[idx]
-        caps = [
-            i["cap"] if isinstance(i["cap"], list) else [i["cap"]]
-            for i in image_data
-        ]
+        caps = [i["cap"] if isinstance(i["cap"], list) else [i["cap"]] for i in image_data]
         try:
             caps = [[random.choice(i)] for i in caps]
         except Exception as e:

fastvideo/dataset/latent_datasets.py

@@ -20,10 +20,8 @@ class LatentDataset(Dataset):
         self.datase_dir_path = os.path.dirname(json_path)
         self.video_dir = os.path.join(self.datase_dir_path, "video")
         self.latent_dir = os.path.join(self.datase_dir_path, "latent")
-        self.prompt_embed_dir = os.path.join(self.datase_dir_path,
-                                             "prompt_embed")
-        self.prompt_attention_mask_dir = os.path.join(self.datase_dir_path,
-                                                      "prompt_attention_mask")
+        self.prompt_embed_dir = os.path.join(self.datase_dir_path, "prompt_embed")
+        self.prompt_attention_mask_dir = os.path.join(self.datase_dir_path, "prompt_attention_mask")
         with open(self.json_path, "r") as f:
             self.data_anno = json.load(f)
         # json.load(f) already keeps the order
@@ -33,16 +31,12 @@ class LatentDataset(Dataset):
         self.uncond_prompt_embed = torch.zeros(256, 4096).to(torch.float32)
         # 256 zeros
         self.uncond_prompt_mask = torch.zeros(256).bool()
-        self.lengths = [
-            data_item["length"] if "length" in data_item else 1
-            for data_item in self.data_anno
-        ]
+        self.lengths = [data_item["length"] if "length" in data_item else 1 for data_item in self.data_anno]
 
     def __getitem__(self, idx):
         latent_file = self.data_anno[idx]["latent_path"]
         prompt_embed_file = self.data_anno[idx]["prompt_embed_path"]
-        prompt_attention_mask_file = self.data_anno[idx][
-            "prompt_attention_mask"]
+        prompt_attention_mask_file = self.data_anno[idx]["prompt_attention_mask"]
         # load
         latent = torch.load(
             os.path.join(self.latent_dir, latent_file),
@@ -60,8 +54,7 @@ class LatentDataset(Dataset):
                 weights_only=True,
             )
             prompt_attention_mask = torch.load(
-                os.path.join(self.prompt_attention_mask_dir,
-                             prompt_attention_mask_file),
+                os.path.join(self.prompt_attention_mask_dir, prompt_attention_mask_file),
                 map_location="cpu",
                 weights_only=True,
             )
@@ -111,13 +104,8 @@ def latent_collate_function(batch):
 
 
 if __name__ == "__main__":
-    dataset = LatentDataset("data/Mochi-Synthetic-Data/merge.txt",
-                            num_latent_t=28)
-    dataloader = torch.utils.data.DataLoader(
-        dataset,
-        batch_size=2,
-        shuffle=False,
-        collate_fn=latent_collate_function)
+    dataset = LatentDataset("data/Mochi-Synthetic-Data/merge.txt", num_latent_t=28)
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=2, shuffle=False, collate_fn=latent_collate_function)
     for latent, prompt_embed, latent_attn_mask, prompt_attention_mask in dataloader:
         print(
             latent.shape,

fastvideo/dataset/t2v_datasets.py

@@ -46,8 +46,7 @@ class DataSetProg(metaclass=SingletonMeta):
 
         for i in range(self.num_workers):
             self.n_used_elements[i] = 0
-            per_worker = int(
-                math.ceil(len(self.elements) / float(self.num_workers)))
+            per_worker = int(math.ceil(len(self.elements) / float(self.num_workers)))
             start = i * per_worker
             end = min(start + per_worker, len(self.elements))
             self.worker_elements[i] = self.elements[start:end]
@@ -58,9 +57,7 @@ class DataSetProg(metaclass=SingletonMeta):
         else:
             worker_id = work_info.id
 
-        idx = self.worker_elements[worker_id][
-            self.n_used_elements[worker_id] %
-            len(self.worker_elements[worker_id])]
+        idx = self.worker_elements[worker_id][self.n_used_elements[worker_id] % len(self.worker_elements[worker_id])]
         self.n_used_elements[worker_id] += 1
         return idx
 
@@ -68,10 +65,7 @@ class DataSetProg(metaclass=SingletonMeta):
 dataset_prog = DataSetProg()
 
 
-def filter_resolution(h,
-                      w,
-                      max_h_div_w_ratio=17 / 16,
-                      min_h_div_w_ratio=8 / 16):
+def filter_resolution(h, w, max_h_div_w_ratio=17 / 16, min_h_div_w_ratio=8 / 16):
     if h / w <= max_h_div_w_ratio and h / w >= min_h_div_w_ratio:
         return True
     return False
@@ -79,8 +73,7 @@ def filter_resolution(h,
 
 class T2V_dataset(Dataset):
 
-    def __init__(self, args, transform, temporal_sample, tokenizer,
-                 transform_topcrop):
+    def __init__(self, args, transform, temporal_sample, tokenizer, transform_topcrop):
         self.data = args.data_merge_path
         self.num_frames = args.num_frames
         self.train_fps = args.train_fps
@@ -109,8 +102,7 @@ class T2V_dataset(Dataset):
         self.lengths = self.sample_num_frames
 
         n_elements = len(cap_list)
-        dataset_prog.set_cap_list(args.dataloader_num_workers, cap_list,
-                                  n_elements)
+        dataset_prog.set_cap_list(args.dataloader_num_workers, cap_list, n_elements)
 
         print(f"video length: {len(dataset_prog.cap_list)}", flush=True)
 
@@ -137,8 +129,7 @@ class T2V_dataset(Dataset):
         video_path = dataset_prog.cap_list[idx]["path"]
         assert os.path.exists(video_path), f"file {video_path} do not exist!"
         frame_indices = dataset_prog.cap_list[idx]["sample_frame_index"]
-        torchvision_video, _, metadata = torchvision.io.read_video(
-            video_path, output_format="TCHW")
+        torchvision_video, _, metadata = torchvision.io.read_video(video_path, output_format="TCHW")
         video = torchvision_video[frame_indices]
         video = self.transform(video)
         video = rearrange(video, "t c h w -> c t h w")
@@ -178,8 +169,7 @@ class T2V_dataset(Dataset):
         )
 
     def get_image(self, idx):
-        image_data = dataset_prog.cap_list[
-            idx]  # [{'path': path, 'cap': cap}, ...]
+        image_data = dataset_prog.cap_list[idx]  # [{'path': path, 'cap': cap}, ...]
 
         image = Image.open(image_data["path"]).convert("RGB")  # [h, w, c]
         image = torch.from_numpy(np.array(image))  # [h, w, c]
@@ -188,15 +178,13 @@ class T2V_dataset(Dataset):
         #     h, w = i.shape[-2:]
         #     assert h / w <= 17 / 16 and h / w >= 8 / 16, f'Only image with a ratio (h/w) less than 17/16 and more than 8/16 are supported. But found ratio is {round(h / w, 2)} with the shape of {i.shape}'
 
-        image = (self.transform_topcrop(image) if "human_images"
-                 in image_data["path"] else self.transform(image)
+        image = (self.transform_topcrop(image) if "human_images" in image_data["path"] else self.transform(image)
                  )  #  [1 C H W] -> num_img [1 C H W]
         image = image.transpose(0, 1)  # [1 C H W] -> [C 1 H W]
 
         image = image.float() / 127.5 - 1.0
 
-        caps = (image_data["cap"] if isinstance(image_data["cap"], list) else
-                [image_data["cap"]])
+        caps = (image_data["cap"] if isinstance(image_data["cap"], list) else [image_data["cap"]])
         caps = [random.choice(caps)]
         text = caps
         input_ids, cond_mask = [], []
@@ -250,12 +238,10 @@ class T2V_dataset(Dataset):
                     cnt_no_resolution += 1
                     continue
                 else:
-                    if (resolution.get("height", None) is None
-                            or resolution.get("width", None) is None):
+                    if (resolution.get("height", None) is None or resolution.get("width", None) is None):
                         cnt_no_resolution += 1
                         continue
-                    height, width = i["resolution"]["height"], i["resolution"][
-                        "width"]
+                    height, width = i["resolution"]["height"], i["resolution"]["width"]
                     aspect = self.max_height / self.max_width
                     hw_aspect_thr = 1.5
                     is_pick = filter_resolution(
@@ -273,34 +259,29 @@ class T2V_dataset(Dataset):
                 i["num_frames"] = math.ceil(fps * duration)
                 # max 5.0 and min 1.0 are just thresholds to filter some videos which have suitable duration.
                 if i["num_frames"] / fps > self.video_length_tolerance_range * (
-                        self.num_frames / self.train_fps * self.speed_factor
-                ):  # too long video is not suitable for this training stage (self.num_frames)
+                        self.num_frames / self.train_fps *
+                        self.speed_factor):  # too long video is not suitable for this training stage (self.num_frames)
                     cnt_too_long += 1
                     continue
 
                 # resample in case high fps, such as 50/60/90/144 -> train_fps(e.g, 24)
                 frame_interval = fps / self.train_fps
                 start_frame_idx = 0
-                frame_indices = np.arange(start_frame_idx, i["num_frames"],
-                                          frame_interval).astype(int)
+                frame_indices = np.arange(start_frame_idx, i["num_frames"], frame_interval).astype(int)
 
                 # comment out it to enable dynamic frames training
-                if (len(frame_indices) < self.num_frames
-                        and random.random() < self.drop_short_ratio):
+                if (len(frame_indices) < self.num_frames and random.random() < self.drop_short_ratio):
                     cnt_too_short += 1
                     continue
 
                 #  too long video will be temporal-crop randomly
                 if len(frame_indices) > self.num_frames:
-                    begin_index, end_index = self.temporal_sample(
-                        len(frame_indices))
+                    begin_index, end_index = self.temporal_sample(len(frame_indices))
                     frame_indices = frame_indices[begin_index:end_index]
                     # frame_indices = frame_indices[:self.num_frames]  # head crop
                 i["sample_frame_index"] = frame_indices.tolist()
                 new_cap_list.append(i)
-                i["sample_num_frames"] = len(
-                    i["sample_frame_index"]
-                )  # will use in dataloader(group sampler)
+                i["sample_num_frames"] = len(i["sample_frame_index"])  # will use in dataloader(group sampler)
                 sample_num_frames.append(i["sample_num_frames"])
             elif path.endswith(".jpg"):  # image
                 cnt_img += 1
@@ -309,32 +290,26 @@ class T2V_dataset(Dataset):
                 sample_num_frames.append(i["sample_num_frames"])
             else:
                 raise NameError(
-                    f"Unknown file extension {path.split('.')[-1]}, only support .mp4 for video and .jpg for image"
-                )
+                    f"Unknown file extension {path.split('.')[-1]}, only support .mp4 for video and .jpg for image")
         # import ipdb;ipdb.set_trace()
         main_print(
             f"no_cap: {cnt_no_cap}, too_long: {cnt_too_long}, too_short: {cnt_too_short}, "
             f"no_resolution: {cnt_no_resolution}, resolution_mismatch: {cnt_resolution_mismatch}, "
             f"Counter(sample_num_frames): {Counter(sample_num_frames)}, cnt_movie: {cnt_movie}, cnt_img: {cnt_img}, "
-            f"before filter: {len(cap_list)}, after filter: {len(new_cap_list)}"
-        )
+            f"before filter: {len(cap_list)}, after filter: {len(new_cap_list)}")
         return new_cap_list, sample_num_frames
 
     def decord_read(self, path, frame_indices):
         decord_vr = self.v_decoder(path)
         video_data = decord_vr.get_batch(frame_indices).asnumpy()
         video_data = torch.from_numpy(video_data)
-        video_data = video_data.permute(0, 3, 1,
-                                        2)  # (T, H, W, C) -> (T C H W)
+        video_data = video_data.permute(0, 3, 1, 2)  # (T, H, W, C) -> (T C H W)
         return video_data
 
     def read_jsons(self, data):
         cap_lists = []
         with open(data, "r") as f:
-            folder_anno = [
-                i.strip().split(",") for i in f.readlines()
-                if len(i.strip()) > 0
-            ]
+            folder_anno = [i.strip().split(",") for i in f.readlines() if len(i.strip()) > 0]
         print(folder_anno)
         for folder, anno in folder_anno:
             with open(anno, "r") as f:

fastvideo/dataset/transform.py

@@ -21,19 +21,15 @@ def center_crop_arr(pil_image, image_size):
     https://github.com/openai/guided-diffusion/blob/8fb3ad9197f16bbc40620447b2742e13458d2831/guided_diffusion/image_datasets.py#L126
     """
     while min(*pil_image.size) >= 2 * image_size:
-        pil_image = pil_image.resize(tuple(x // 2 for x in pil_image.size),
-                                     resample=Image.BOX)
+        pil_image = pil_image.resize(tuple(x // 2 for x in pil_image.size), resample=Image.BOX)
 
     scale = image_size / min(*pil_image.size)
-    pil_image = pil_image.resize(tuple(
-        round(x * scale) for x in pil_image.size),
-                                 resample=Image.BICUBIC)
+    pil_image = pil_image.resize(tuple(round(x * scale) for x in pil_image.size), resample=Image.BICUBIC)
 
     arr = np.array(pil_image)
     crop_y = (arr.shape[0] - image_size) // 2
     crop_x = (arr.shape[1] - image_size) // 2
-    return Image.fromarray(arr[crop_y:crop_y + image_size,
-                               crop_x:crop_x + image_size])
+    return Image.fromarray(arr[crop_y:crop_y + image_size, crop_x:crop_x + image_size])
 
 
 def crop(clip, i, j, h, w):
@@ -48,9 +44,7 @@ def crop(clip, i, j, h, w):
 
 def resize(clip, target_size, interpolation_mode):
     if len(target_size) != 2:
-        raise ValueError(
-            f"target size should be tuple (height, width), instead got {target_size}"
-        )
+        raise ValueError(f"target size should be tuple (height, width), instead got {target_size}")
     return torch.nn.functional.interpolate(
         clip,
         size=target_size,
@@ -62,9 +56,7 @@ def resize(clip, target_size, interpolation_mode):
 
 def resize_scale(clip, target_size, interpolation_mode):
     if len(target_size) != 2:
-        raise ValueError(
-            f"target size should be tuple (height, width), instead got {target_size}"
-        )
+        raise ValueError(f"target size should be tuple (height, width), instead got {target_size}")
     H, W = clip.size(-2), clip.size(-1)
     scale_ = target_size[0] / min(H, W)
     return torch.nn.functional.interpolate(
@@ -174,8 +166,7 @@ def normalize_video(clip):
     """
     _is_tensor_video_clip(clip)
     if not clip.dtype == torch.uint8:
-        raise TypeError("clip tensor should have data type uint8. Got %s" %
-                        str(clip.dtype))
+        raise TypeError("clip tensor should have data type uint8. Got %s" % str(clip.dtype))
     # return clip.float().permute(3, 0, 1, 2) / 255.0
     return clip.float() / 255.0
 
@@ -236,9 +227,7 @@ class RandomCropVideo:
         th, tw = self.size
 
         if h < th or w < tw:
-            raise ValueError(
-                f"Required crop size {(th, tw)} is larger than input image size {(h, w)}"
-            )
+            raise ValueError(f"Required crop size {(th, tw)} is larger than input image size {(h, w)}")
 
         if w == tw and h == th:
             return 0, 0, h, w
@@ -312,9 +301,7 @@ class LongSideResizeVideo:
         else:
             h = int(h * self.size / w)
             w = self.size
-        resize_clip = resize(clip,
-                             target_size=(h, w),
-                             interpolation_mode=self.interpolation_mode)
+        resize_clip = resize(clip, target_size=(h, w), interpolation_mode=self.interpolation_mode)
         return resize_clip
 
     def __repr__(self) -> str:
@@ -334,8 +321,7 @@ class CenterCropResizeVideo:
         interpolation_mode="bilinear",
     ):
         if len(size) != 2:
-            raise ValueError(
-                f"size should be tuple (height, width), instead got {size}")
+            raise ValueError(f"size should be tuple (height, width), instead got {size}")
         self.size = size
         self.top_crop = top_crop
         self.interpolation_mode = interpolation_mode
@@ -349,10 +335,7 @@ class CenterCropResizeVideo:
                 size is (T, C, crop_size, crop_size)
         """
         # clip_center_crop = center_crop_using_short_edge(clip)
-        clip_center_crop = center_crop_th_tw(clip,
-                                             self.size[0],
-                                             self.size[1],
-                                             top_crop=self.top_crop)
+        clip_center_crop = center_crop_th_tw(clip, self.size[0], self.size[1], top_crop=self.top_crop)
         # import ipdb;ipdb.set_trace()
         clip_center_crop_resize = resize(
             clip_center_crop,
@@ -378,9 +361,7 @@ class UCFCenterCropVideo:
     ):
         if isinstance(size, tuple):
             if len(size) != 2:
-                raise ValueError(
-                    f"size should be tuple (height, width), instead got {size}"
-                )
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
             self.size = size
         else:
             self.size = (size, size)
@@ -395,9 +376,7 @@ class UCFCenterCropVideo:
             torch.tensor: scale resized / center cropped video clip.
                 size is (T, C, crop_size, crop_size)
         """
-        clip_resize = resize_scale(clip=clip,
-                                   target_size=self.size,
-                                   interpolation_mode=self.interpolation_mode)
+        clip_resize = resize_scale(clip=clip, target_size=self.size, interpolation_mode=self.interpolation_mode)
         clip_center_crop = center_crop(clip_resize, self.size)
         return clip_center_crop
 
@@ -417,9 +396,7 @@ class KineticsRandomCropResizeVideo:
     ):
         if isinstance(size, tuple):
             if len(size) != 2:
-                raise ValueError(
-                    f"size should be tuple (height, width), instead got {size}"
-                )
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
             self.size = size
         else:
             self.size = (size, size)
@@ -428,8 +405,7 @@ class KineticsRandomCropResizeVideo:
 
     def __call__(self, clip):
         clip_random_crop = random_shift_crop(clip)
-        clip_resize = resize(clip_random_crop, self.size,
-                             self.interpolation_mode)
+        clip_resize = resize(clip_random_crop, self.size, self.interpolation_mode)
         return clip_resize
 
 
@@ -442,9 +418,7 @@ class CenterCropVideo:
     ):
         if isinstance(size, tuple):
             if len(size) != 2:
-                raise ValueError(
-                    f"size should be tuple (height, width), instead got {size}"
-                )
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
             self.size = size
         else:
             self.size = (size, size)
@@ -571,8 +545,7 @@ class DynamicSampleDuration(object):
     def __call__(self, t, h, w):
         if self.extra_1:
             t = t - 1
-        truncate_t_list = list(
-            range(t + 1))[t // 2:][::self.t_stride]  # need half at least
+        truncate_t_list = list(range(t + 1))[t // 2:][::self.t_stride]  # need half at least
         truncate_t = random.choice(truncate_t_list)
         if self.extra_1:
             truncate_t = truncate_t + 1
@@ -587,18 +560,14 @@ if __name__ == "__main__":
     from torchvision import transforms
     from torchvision.utils import save_image
 
-    vframes, aframes, info = io.read_video(filename="./v_Archery_g01_c03.avi",
-                                           pts_unit="sec",
-                                           output_format="TCHW")
+    vframes, aframes, info = io.read_video(filename="./v_Archery_g01_c03.avi", pts_unit="sec", output_format="TCHW")
 
     trans = transforms.Compose([
         Normalize255(),
         RandomHorizontalFlipVideo(),
         UCFCenterCropVideo(512),
         # NormalizeVideo(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
-        transforms.Normalize(mean=[0.5, 0.5, 0.5],
-                             std=[0.5, 0.5, 0.5],
-                             inplace=True),
+        transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
     ])
 
     target_video_len = 32
@@ -613,10 +582,7 @@ if __name__ == "__main__":
     # print(start_frame_ind)
     # print(end_frame_ind)
     assert end_frame_ind - start_frame_ind >= target_video_len
-    frame_indice = np.linspace(start_frame_ind,
-                               end_frame_ind - 1,
-                               target_video_len,
-                               dtype=int)
+    frame_indice = np.linspace(start_frame_ind, end_frame_ind - 1, target_video_len, dtype=int)
     print(frame_indice)
 
     select_vframes = vframes[frame_indice]
@@ -627,14 +593,11 @@ if __name__ == "__main__":
     print(select_vframes_trans.shape)
     print(select_vframes_trans.dtype)
 
-    select_vframes_trans_int = ((select_vframes_trans * 0.5 + 0.5) *
-                                255).to(dtype=torch.uint8)
+    select_vframes_trans_int = ((select_vframes_trans * 0.5 + 0.5) * 255).to(dtype=torch.uint8)
     print(select_vframes_trans_int.dtype)
     print(select_vframes_trans_int.permute(0, 2, 3, 1).shape)
 
-    io.write_video("./test.avi",
-                   select_vframes_trans_int.permute(0, 2, 3, 1),
-                   fps=8)
+    io.write_video("./test.avi", select_vframes_trans_int.permute(0, 2, 3, 1), fps=8)
 
     for i in range(target_video_len):
         save_image(

fastvideo/distill.py

@@ -21,23 +21,17 @@ from torch.utils.data import DataLoader
 from torch.utils.data.distributed import DistributedSampler
 from tqdm.auto import tqdm
 
-from fastvideo.dataset.latent_datasets import (LatentDataset,
-                                               latent_collate_function)
+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.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)
+from fastvideo.utils.checkpoint import (resume_lora_optimizer, save_checkpoint, save_lora_checkpoint)
+from fastvideo.utils.communications import (broadcast, sp_parallel_dataloader_wrapper)
 from fastvideo.utils.dataset_utils import LengthGroupedSampler
-from fastvideo.utils.fsdp_util import (apply_fsdp_checkpointing,
-                                       get_dit_fsdp_kwargs)
+from fastvideo.utils.fsdp_util import (apply_fsdp_checkpointing, get_dit_fsdp_kwargs)
 from fastvideo.utils.load import load_transformer
-from fastvideo.utils.parallel_states import (destroy_sequence_parallel_group,
-                                             get_sequence_parallel_state,
-                                             initialize_sequence_parallel_state
-                                             )
+from fastvideo.utils.parallel_states import (destroy_sequence_parallel_group, get_sequence_parallel_state,
+                                             initialize_sequence_parallel_state)
 from fastvideo.utils.validation import log_validation
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
@@ -59,18 +53,14 @@ def get_norm(model_pred, norms, gradient_accumulation_steps):
     fro_norm = (
         torch.linalg.matrix_norm(model_pred, ord="fro") /  # codespell:ignore
         gradient_accumulation_steps)
-    largest_singular_value = (torch.linalg.matrix_norm(model_pred, ord=2) /
-                              gradient_accumulation_steps)
-    absolute_mean = torch.mean(
-        torch.abs(model_pred)) / gradient_accumulation_steps
-    absolute_max = torch.max(
-        torch.abs(model_pred)) / gradient_accumulation_steps
+    largest_singular_value = (torch.linalg.matrix_norm(model_pred, ord=2) / gradient_accumulation_steps)
+    absolute_mean = torch.mean(torch.abs(model_pred)) / gradient_accumulation_steps
+    absolute_max = torch.max(torch.abs(model_pred)) / gradient_accumulation_steps
     dist.all_reduce(fro_norm, op=dist.ReduceOp.AVG)
     dist.all_reduce(largest_singular_value, op=dist.ReduceOp.AVG)
     dist.all_reduce(absolute_mean, op=dist.ReduceOp.AVG)
     norms["fro"] += torch.mean(fro_norm).item()  # codespell:ignore
-    norms["largest singular value"] += torch.mean(
-        largest_singular_value).item()
+    norms["largest singular value"] += torch.mean(largest_singular_value).item()
     norms["absolute mean"] += absolute_mean.item()
     norms["absolute max"] += absolute_max.item()
 
@@ -118,23 +108,18 @@ def distill_one_step(
         model_input = normalize_dit_input(model_type, latents)
         noise = torch.randn_like(model_input)
         bsz = model_input.shape[0]
-        index = torch.randint(0,
-                              num_euler_timesteps, (bsz, ),
-                              device=model_input.device).long()
+        index = torch.randint(0, num_euler_timesteps, (bsz, ), device=model_input.device).long()
         if sp_size > 1:
             broadcast(index)
         # Add noise according to flow matching.
         # sigmas = get_sigmas(start_timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
         sigmas = extract_into_tensor(solver.sigmas, index, model_input.shape)
-        sigmas_prev = extract_into_tensor(solver.sigmas_prev, index,
-                                          model_input.shape)
+        sigmas_prev = extract_into_tensor(solver.sigmas_prev, index, model_input.shape)
 
-        timesteps = (sigmas *
-                     noise_scheduler.config.num_train_timesteps).view(-1)
+        timesteps = (sigmas * noise_scheduler.config.num_train_timesteps).view(-1)
         # if squeeze to [], unsqueeze to [1]
 
-        timesteps_prev = (sigmas_prev *
-                          noise_scheduler.config.num_train_timesteps).view(-1)
+        timesteps_prev = (sigmas_prev * noise_scheduler.config.num_train_timesteps).view(-1)
         noisy_model_input = sigmas * noise + (1.0 - sigmas) * model_input
         # Predict the noise residual
         with torch.autocast("cuda", dtype=torch.bfloat16):
@@ -146,15 +131,13 @@ def distill_one_step(
                 "return_dict": False,
             }
             if hunyuan_teacher_disable_cfg:
-                teacher_kwargs["guidance"] = torch.tensor(
-                    [1000.0],
-                    device=noisy_model_input.device,
-                    dtype=torch.bfloat16)
+                teacher_kwargs["guidance"] = torch.tensor([1000.0],
+                                                          device=noisy_model_input.device,
+                                                          dtype=torch.bfloat16)
             model_pred = transformer(**teacher_kwargs)[0]
 
         # if accelerator.is_main_process:
-        model_pred, end_index = solver.euler_style_multiphase_pred(
-            noisy_model_input, model_pred, index, multiphase)
+        model_pred, end_index = solver.euler_style_multiphase_pred(noisy_model_input, model_pred, index, multiphase)
         with torch.no_grad():
             w = distill_cfg
             with torch.autocast("cuda", dtype=torch.bfloat16):
@@ -177,10 +160,8 @@ def distill_one_step(
                         uncond_prompt_mask.unsqueeze(0).expand(bsz, -1),
                         return_dict=False,
                     )[0].float()
-            teacher_output = cond_teacher_output + w * (cond_teacher_output -
-                                                        uncond_teacher_output)
-            x_prev = solver.euler_step(noisy_model_input, teacher_output,
-                                       index)
+            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():
@@ -202,33 +183,26 @@ def distill_one_step(
                         return_dict=False,
                     )[0]
 
-            target, end_index = solver.euler_style_multiphase_pred(
-                x_prev, target_pred, index, multiphase, True)
+            target, end_index = solver.euler_style_multiphase_pred(x_prev, target_pred, index, multiphase, True)
 
         huber_c = 0.001
         # loss = loss.mean()
-        loss = (torch.mean(
-            torch.sqrt((model_pred.float() - target.float())**2 + huber_c**2) -
-            huber_c) / gradient_accumulation_steps)
+        loss = (torch.mean(torch.sqrt((model_pred.float() - target.float())**2 + huber_c**2) - huber_c) /
+                gradient_accumulation_steps)
         if pred_decay_weight > 0:
             if pred_decay_type == "l1":
-                pred_decay_loss = (
-                    torch.mean(torch.sqrt(model_pred.float()**2)) *
-                    pred_decay_weight / gradient_accumulation_steps)
+                pred_decay_loss = (torch.mean(torch.sqrt(model_pred.float()**2)) * pred_decay_weight /
+                                   gradient_accumulation_steps)
                 loss += pred_decay_loss
             elif pred_decay_type == "l2":
                 # essnetially k2?
-                pred_decay_loss = (torch.mean(model_pred.float()**2) *
-                                   pred_decay_weight /
-                                   gradient_accumulation_steps)
+                pred_decay_loss = (torch.mean(model_pred.float()**2) * pred_decay_weight / gradient_accumulation_steps)
                 loss += pred_decay_loss
             else:
-                assert NotImplementedError(
-                    "pred_decay_type is not implemented")
+                assert NotImplementedError("pred_decay_type is not implemented")
 
         # calculate model_pred norm and mean
-        get_norm(model_pred.detach().float(), model_pred_norm,
-                 gradient_accumulation_steps)
+        get_norm(model_pred.detach().float(), model_pred_norm, gradient_accumulation_steps)
         loss.backward()
 
         avg_loss = loss.detach().clone()
@@ -238,12 +212,9 @@ def distill_one_step(
     # update ema
     if ema_transformer is not None:
         reshard_fsdp(ema_transformer)
-        for p_averaged, p_model in zip(ema_transformer.parameters(),
-                                       transformer.parameters()):
+        for p_averaged, p_model in zip(ema_transformer.parameters(), transformer.parameters()):
             with torch.no_grad():
-                p_averaged.copy_(
-                    torch.lerp(p_averaged.detach(), p_model.detach(),
-                               1 - ema_decay))
+                p_averaged.copy_(torch.lerp(p_averaged.detach(), p_model.detach(), 1 - ema_decay))
 
     grad_norm = transformer.clip_grad_norm_(max_grad_norm)
     optimizer.step()
@@ -306,11 +277,8 @@ def main(args):
         transformer.add_adapter(transformer_lora_config)
 
     main_print(
-        f"  Total training parameters = {sum(p.numel() for p in transformer.parameters() if p.requires_grad) / 1e6} M"
-    )
-    main_print(
-        f"--> Initializing FSDP with sharding strategy: {args.fsdp_sharding_startegy}"
-    )
+        f"  Total training parameters = {sum(p.numel() for p in transformer.parameters() if p.requires_grad) / 1e6} M")
+    main_print(f"--> Initializing FSDP with sharding strategy: {args.fsdp_sharding_startegy}")
     fsdp_kwargs, no_split_modules = get_dit_fsdp_kwargs(
         transformer,
         args.fsdp_sharding_startegy,
@@ -322,12 +290,9 @@ def main(args):
     if args.use_lora:
         transformer.config.lora_rank = args.lora_rank
         transformer.config.lora_alpha = args.lora_alpha
-        transformer.config.lora_target_modules = [
-            "to_k", "to_q", "to_v", "to_out.0"
-        ]
+        transformer.config.lora_target_modules = ["to_k", "to_q", "to_v", "to_out.0"]
         transformer._no_split_modules = no_split_modules
-        fsdp_kwargs["auto_wrap_policy"] = fsdp_kwargs["auto_wrap_policy"](
-            transformer)
+        fsdp_kwargs["auto_wrap_policy"] = fsdp_kwargs["auto_wrap_policy"](transformer)
 
     transformer = FSDP(
         transformer,
@@ -345,13 +310,10 @@ def main(args):
     main_print("--> model loaded")
 
     if args.gradient_checkpointing:
-        apply_fsdp_checkpointing(transformer, no_split_modules,
-                                 args.selective_checkpointing)
-        apply_fsdp_checkpointing(teacher_transformer, no_split_modules,
-                                 args.selective_checkpointing)
+        apply_fsdp_checkpointing(transformer, no_split_modules, args.selective_checkpointing)
+        apply_fsdp_checkpointing(teacher_transformer, no_split_modules, args.selective_checkpointing)
         if args.use_ema:
-            apply_fsdp_checkpointing(ema_transformer, no_split_modules,
-                                     args.selective_checkpointing)
+            apply_fsdp_checkpointing(ema_transformer, no_split_modules, args.selective_checkpointing)
     # Set model as trainable.
     transformer.train()
     teacher_transformer.requires_grad_(False)
@@ -359,8 +321,7 @@ def main(args):
         ema_transformer.requires_grad_(False)
     noise_scheduler = FlowMatchEulerDiscreteScheduler(shift=args.shift)
     if args.scheduler_type == "pcm_linear_quadratic":
-        linear_steps = int(noise_scheduler.config.num_train_timesteps *
-                           args.linear_range)
+        linear_steps = int(noise_scheduler.config.num_train_timesteps * args.linear_range)
         sigmas = linear_quadratic_schedule(
             noise_scheduler.config.num_train_timesteps,
             args.linear_quadratic_threshold,
@@ -376,8 +337,7 @@ def main(args):
     )
     solver.to(device)
     params_to_optimize = transformer.parameters()
-    params_to_optimize = list(
-        filter(lambda p: p.requires_grad, params_to_optimize))
+    params_to_optimize = list(filter(lambda p: p.requires_grad, params_to_optimize))
 
     optimizer = torch.optim.AdamW(
         params_to_optimize,
@@ -389,8 +349,8 @@ def main(args):
 
     init_steps = 0
     if args.resume_from_lora_checkpoint:
-        transformer, optimizer, init_steps = resume_lora_optimizer(
-            transformer, args.resume_from_lora_checkpoint, optimizer)
+        transformer, optimizer, init_steps = resume_lora_optimizer(transformer, args.resume_from_lora_checkpoint,
+                                                                   optimizer)
     main_print(f"optimizer: {optimizer}")
 
     # todo add lr scheduler
@@ -404,8 +364,7 @@ def main(args):
         last_epoch=init_steps - 1,
     )
 
-    train_dataset = LatentDataset(args.data_json_path, args.num_latent_t,
-                                  args.cfg)
+    train_dataset = LatentDataset(args.data_json_path, args.num_latent_t, args.cfg)
     uncond_prompt_embed = train_dataset.uncond_prompt_embed
     uncond_prompt_mask = train_dataset.uncond_prompt_mask
     sampler = (LengthGroupedSampler(
@@ -429,42 +388,33 @@ def main(args):
     )
 
     num_update_steps_per_epoch = math.ceil(
-        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)
+        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:
         project = args.tracker_project_name or "fastvideo"
         wandb.init(project=project, config=args)
 
     # Train!
-    total_batch_size = (world_size * args.gradient_accumulation_steps /
-                        args.sp_size * args.train_sp_batch_size)
+    total_batch_size = (world_size * args.gradient_accumulation_steps / args.sp_size * args.train_sp_batch_size)
     main_print("***** Running training *****")
     main_print(f"  Num examples = {len(train_dataset)}")
     main_print(f"  Dataloader size = {len(train_dataloader)}")
     main_print(f"  Num Epochs = {args.num_train_epochs}")
     main_print(f"  Resume training from step {init_steps}")
-    main_print(
-        f"  Instantaneous batch size per device = {args.train_batch_size}")
-    main_print(
-        f"  Total train batch size (w. data & sequence parallel, accumulation) = {total_batch_size}"
-    )
-    main_print(
-        f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    main_print(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    main_print(f"  Total train batch size (w. data & sequence parallel, accumulation) = {total_batch_size}")
+    main_print(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
     main_print(f"  Total optimization steps = {args.max_train_steps}")
     main_print(
         f"  Total training parameters per FSDP shard = {sum(p.numel() for p in transformer.parameters() if p.requires_grad) / 1e9} B"
     )
     # print dtype
-    main_print(
-        f"  Master weight dtype: {transformer.parameters().__next__().dtype}")
+    main_print(f"  Master weight dtype: {transformer.parameters().__next__().dtype}")
 
     # Potentially load in the weights and states from a previous save
     if args.resume_from_checkpoint:
-        assert NotImplementedError(
-            "resume_from_checkpoint is not supported now.")
+        assert NotImplementedError("resume_from_checkpoint is not supported now.")
         # TODO
 
     progress_bar = tqdm(
@@ -502,7 +452,7 @@ def main(args):
         return phase
 
     for step in range(init_steps + 1, args.max_train_steps + 1):
-        start_time = time.time()
+        start_time = time.perf_counter()
         assert args.multi_phased_distill_schedule is not None
         num_phases = get_num_phases(args.multi_phased_distill_schedule, step)
 
@@ -532,7 +482,7 @@ def main(args):
             args.hunyuan_teacher_disable_cfg,
         )
 
-        step_time = time.time() - start_time
+        step_time = time.perf_counter() - start_time
         step_times.append(step_time)
         avg_step_time = sum(step_times) / len(step_times)
 
@@ -546,37 +496,26 @@ def main(args):
         if rank <= 0:
             wandb.log(
                 {
-                    "train_loss":
-                    loss,
-                    "learning_rate":
-                    lr_scheduler.get_last_lr()[0],
-                    "step_time":
-                    step_time,
-                    "avg_step_time":
-                    avg_step_time,
-                    "grad_norm":
-                    grad_norm,
-                    "pred_fro_norm":
-                    pred_norm["fro"],  # codespell:ignore
-                    "pred_largest_singular_value":
-                    pred_norm["largest singular value"],
-                    "pred_absolute_mean":
-                    pred_norm["absolute mean"],
-                    "pred_absolute_max":
-                    pred_norm["absolute max"],
+                    "train_loss": loss,
+                    "learning_rate": lr_scheduler.get_last_lr()[0],
+                    "step_time": step_time,
+                    "avg_step_time": avg_step_time,
+                    "grad_norm": grad_norm,
+                    "pred_fro_norm": pred_norm["fro"],  # codespell:ignore
+                    "pred_largest_singular_value": pred_norm["largest singular value"],
+                    "pred_absolute_mean": pred_norm["absolute mean"],
+                    "pred_absolute_max": pred_norm["absolute max"],
                 },
                 step=step,
             )
         if step % args.checkpointing_steps == 0:
             if args.use_lora:
                 # Save LoRA weights
-                save_lora_checkpoint(transformer, optimizer, rank,
-                                     args.output_dir, step)
+                save_lora_checkpoint(transformer, optimizer, rank, args.output_dir, step)
             else:
                 # Your existing checkpoint saving code
                 if args.use_ema:
-                    save_checkpoint(ema_transformer, rank, args.output_dir,
-                                    step)
+                    save_checkpoint(ema_transformer, rank, args.output_dir, step)
                 else:
                     save_checkpoint(transformer, rank, args.output_dir, step)
             dist.barrier()
@@ -610,11 +549,9 @@ def main(args):
                 )
 
     if args.use_lora:
-        save_lora_checkpoint(transformer, optimizer, rank, args.output_dir,
-                             args.max_train_steps)
+        save_lora_checkpoint(transformer, optimizer, rank, args.output_dir, args.max_train_steps)
     else:
-        save_checkpoint(transformer, rank, args.output_dir,
-                        args.max_train_steps)
+        save_checkpoint(transformer, rank, args.output_dir, args.max_train_steps)
 
     if get_sequence_parallel_state():
         destroy_sequence_parallel_group()
@@ -623,10 +560,7 @@ def main(args):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
 
-    parser.add_argument("--model_type",
-                        type=str,
-                        default="mochi",
-                        help="The type of model to train.")
+    parser.add_argument("--model_type", type=str, default="mochi", help="The type of model to train.")
 
     # dataset & dataloader
     parser.add_argument("--data_json_path", type=str, required=True)
@@ -637,8 +571,7 @@ if __name__ == "__main__":
         "--dataloader_num_workers",
         type=int,
         default=10,
-        help=
-        "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
+        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
     )
     parser.add_argument(
         "--train_batch_size",
@@ -646,10 +579,7 @@ if __name__ == "__main__":
         default=16,
         help="Batch size (per device) for the training dataloader.",
     )
-    parser.add_argument("--num_latent_t",
-                        type=int,
-                        default=28,
-                        help="Number of latent timesteps.")
+    parser.add_argument("--num_latent_t", type=int, default=28, help="Number of latent timesteps.")
     parser.add_argument("--group_frame", action="store_true")  # TODO
     parser.add_argument("--group_resolution", action="store_true")  # TODO
 
@@ -671,16 +601,12 @@ if __name__ == "__main__":
     parser.add_argument("--validation_steps", type=float, default=64)
     parser.add_argument("--log_validation", action="store_true")
     parser.add_argument("--tracker_project_name", type=str, default=None)
-    parser.add_argument("--seed",
-                        type=int,
-                        default=None,
-                        help="A seed for reproducible training.")
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
     parser.add_argument(
         "--output_dir",
         type=str,
         default=None,
-        help=
-        "The output directory where the model predictions and checkpoints will be written.",
+        help="The output directory where the model predictions and checkpoints will be written.",
     )
     parser.add_argument(
         "--checkpoints_total_limit",
@@ -692,37 +618,31 @@ if __name__ == "__main__":
         "--checkpointing_steps",
         type=int,
         default=500,
-        help=
-        ("Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
-         " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
-         " training using `--resume_from_checkpoint`."),
+        help=("Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
+              " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
+              " training using `--resume_from_checkpoint`."),
     )
     parser.add_argument("--shift", type=float, default=1.0)
     parser.add_argument(
         "--resume_from_checkpoint",
         type=str,
         default=None,
-        help=
-        ("Whether training should be resumed from a previous checkpoint. Use a path saved by"
-         ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
-         ),
+        help=("Whether training should be resumed from a previous checkpoint. Use a path saved by"
+              ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'),
     )
     parser.add_argument(
         "--resume_from_lora_checkpoint",
         type=str,
         default=None,
-        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.'
-         ),
+        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***."),
+        help=("[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+              " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."),
     )
 
     # optimizer & scheduler & Training
@@ -731,29 +651,25 @@ if __name__ == "__main__":
         "--max_train_steps",
         type=int,
         default=None,
-        help=
-        "Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
     )
     parser.add_argument(
         "--gradient_accumulation_steps",
         type=int,
         default=1,
-        help=
-        "Number of updates steps to accumulate before performing a backward/update pass.",
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
     )
     parser.add_argument(
         "--learning_rate",
         type=float,
         default=1e-4,
-        help=
-        "Initial learning rate (after the potential warmup period) to use.",
+        help="Initial learning rate (after the potential warmup period) to use.",
     )
     parser.add_argument(
         "--scale_lr",
         action="store_true",
         default=False,
-        help=
-        "Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
     )
     parser.add_argument(
         "--lr_warmup_steps",
@@ -761,47 +677,36 @@ if __name__ == "__main__":
         default=10,
         help="Number of steps for the warmup in the lr scheduler.",
     )
-    parser.add_argument("--max_grad_norm",
-                        default=1.0,
-                        type=float,
-                        help="Max gradient norm.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
     parser.add_argument(
         "--gradient_checkpointing",
         action="store_true",
-        help=
-        "Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
     )
     parser.add_argument("--selective_checkpointing", type=float, default=1.0)
     parser.add_argument(
         "--allow_tf32",
         action="store_true",
-        help=
-        ("Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
-         " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
-         ),
+        help=("Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
+              " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"),
     )
     parser.add_argument(
         "--mixed_precision",
         type=str,
         default=None,
         choices=["no", "fp16", "bf16"],
-        help=
-        ("Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
-         " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
-         " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
-         ),
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."),
     )
     parser.add_argument(
         "--use_cpu_offload",
         action="store_true",
-        help=
-        "Whether to use CPU offload for param & gradient & optimizer states.",
+        help="Whether to use CPU offload for param & gradient & optimizer states.",
     )
 
-    parser.add_argument("--sp_size",
-                        type=int,
-                        default=1,
-                        help="For sequence parallel")
+    parser.add_argument("--sp_size", type=int, default=1, help="For sequence parallel")
     parser.add_argument(
         "--train_sp_batch_size",
         type=int,
@@ -815,14 +720,8 @@ if __name__ == "__main__":
         default=False,
         help="Whether to use LoRA for finetuning.",
     )
-    parser.add_argument("--lora_alpha",
-                        type=int,
-                        default=256,
-                        help="Alpha parameter for LoRA.")
-    parser.add_argument("--lora_rank",
-                        type=int,
-                        default=128,
-                        help="LoRA rank parameter. ")
+    parser.add_argument("--lora_alpha", type=int, default=256, help="Alpha parameter for LoRA.")
+    parser.add_argument("--lora_rank", type=int, default=128, help="LoRA rank parameter. ")
     parser.add_argument("--fsdp_sharding_startegy", default="full")
 
     # lr_scheduler
@@ -830,9 +729,8 @@ if __name__ == "__main__":
         "--lr_scheduler",
         type=str,
         default="constant",
-        help=
-        ('The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
-         ' "constant", "constant_with_warmup"]'),
+        help=('The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+              ' "constant", "constant_with_warmup"]'),
     )
     parser.add_argument("--num_euler_timesteps", type=int, default=100)
     parser.add_argument(
@@ -852,15 +750,9 @@ if __name__ == "__main__":
         action="store_true",
         help="Whether to apply the cfg_solver.",
     )
-    parser.add_argument("--distill_cfg",
-                        type=float,
-                        default=3.0,
-                        help="Distillation coefficient.")
+    parser.add_argument("--distill_cfg", type=float, default=3.0, help="Distillation coefficient.")
     # ["euler_linear_quadratic", "pcm", "pcm_linear_qudratic"]
-    parser.add_argument("--scheduler_type",
-                        type=str,
-                        default="pcm",
-                        help="The scheduler type to use.")
+    parser.add_argument("--scheduler_type", type=str, default="pcm", help="The scheduler type to use.")
     parser.add_argument(
         "--linear_quadratic_threshold",
         type=float,
@@ -873,16 +765,9 @@ if __name__ == "__main__":
         default=0.5,
         help="Range for linear quadratic scheduler.",
     )
-    parser.add_argument("--weight_decay",
-                        type=float,
-                        default=0.001,
-                        help="Weight decay to apply.")
-    parser.add_argument("--use_ema",
-                        action="store_true",
-                        help="Whether to use EMA.")
-    parser.add_argument("--multi_phased_distill_schedule",
-                        type=str,
-                        default=None)
+    parser.add_argument("--weight_decay", type=float, default=0.001, help="Weight decay to apply.")
+    parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA.")
+    parser.add_argument("--multi_phased_distill_schedule", type=str, default=None)
     parser.add_argument("--pred_decay_weight", type=float, default=0.0)
     parser.add_argument("--pred_decay_type", default="l1")
     parser.add_argument("--hunyuan_teacher_disable_cfg", action="store_true")

fastvideo/distill/discriminator.py

@@ -12,16 +12,12 @@ class DiscriminatorHead(nn.Module):
         self.conv1 = nn.Sequential(
             nn.Conv2d(input_channel, inner_channel, 1, 1, 0),
             nn.GroupNorm(32, inner_channel),
-            nn.LeakyReLU(
-                inplace=True
-            ),  # use LeakyReLu instead of GELU shown in the paper to save memory
+            nn.LeakyReLU(inplace=True),  # use LeakyReLu instead of GELU shown in the paper to save memory
         )
         self.conv2 = nn.Sequential(
             nn.Conv2d(inner_channel, inner_channel, 1, 1, 0),
             nn.GroupNorm(32, inner_channel),
-            nn.LeakyReLU(
-                inplace=True
-            ),  # use LeakyReLu instead of GELU shown in the paper to save memory
+            nn.LeakyReLU(inplace=True),  # use LeakyReLu instead of GELU shown in the paper to save memory
         )
 
         self.conv_out = nn.Conv2d(inner_channel, output_channel, 1, 1, 0)
@@ -53,10 +49,8 @@ class Discriminator(nn.Module):
         self.num_h_per_head = num_h_per_head
         self.head_num = len(adapter_channel_dims)
         self.heads = nn.ModuleList([
-            nn.ModuleList([
-                DiscriminatorHead(adapter_channel)
-                for _ in range(self.num_h_per_head)
-            ]) for adapter_channel in adapter_channel_dims
+            nn.ModuleList([DiscriminatorHead(adapter_channel) for _ in range(self.num_h_per_head)])
+            for adapter_channel in adapter_channel_dims
         ])
 
     def forward(self, features):

fastvideo/distill/solver.py

@@ -7,7 +7,7 @@ from diffusers.configuration_utils import ConfigMixin, register_to_config
 from diffusers.schedulers.scheduling_utils import SchedulerMixin
 from diffusers.utils import BaseOutput, logging
 
-from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
+# from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
@@ -38,29 +38,23 @@ class PCMFMScheduler(SchedulerMixin, ConfigMixin):
         linear_range=0.5,
     ):
         if linear_quadratic:
+            raise NotImplementedError("Linear quadratic schedule is not implemented")
             linear_steps = int(num_train_timesteps * linear_range)
-            sigmas = linear_quadratic_schedule(num_train_timesteps,
-                                               linear_quadratic_threshold,
-                                               linear_steps)
+            sigmas = linear_quadratic_schedule(num_train_timesteps, linear_quadratic_threshold, linear_steps)
             sigmas = torch.tensor(sigmas).to(dtype=torch.float32)
         else:
-            timesteps = np.linspace(1,
-                                    num_train_timesteps,
-                                    num_train_timesteps,
-                                    dtype=np.float32)[::-1].copy()
+            timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy()
             timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32)
             sigmas = timesteps / num_train_timesteps
             sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)
         self.euler_timesteps = (np.arange(1, pcm_timesteps + 1) *
-                                (num_train_timesteps //
-                                 pcm_timesteps)).round().astype(np.int64) - 1
+                                (num_train_timesteps // pcm_timesteps)).round().astype(np.int64) - 1
         self.sigmas = sigmas.numpy()[::-1][self.euler_timesteps]
         self.sigmas = torch.from_numpy((self.sigmas[::-1].copy()))
         self.timesteps = self.sigmas * num_train_timesteps
         self._step_index = None
         self._begin_index = None
-        self.sigmas = self.sigmas.to(
-            "cpu")  # to avoid too much CPU/GPU communication
+        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
         self.sigma_min = self.sigmas[-1].item()
         self.sigma_max = self.sigmas[0].item()
 
@@ -119,9 +113,7 @@ class PCMFMScheduler(SchedulerMixin, ConfigMixin):
     def _sigma_to_t(self, sigma):
         return sigma * self.config.num_train_timesteps
 
-    def set_timesteps(self,
-                      num_inference_steps: int,
-                      device: Union[str, torch.device] = None):
+    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
         """
         Sets the discrete timesteps used for the diffusion chain (to be run before inference).
 
@@ -132,19 +124,14 @@ class PCMFMScheduler(SchedulerMixin, ConfigMixin):
                 The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
         """
         self.num_inference_steps = num_inference_steps
-        inference_indices = np.linspace(0,
-                                        self.config.pcm_timesteps,
-                                        num=num_inference_steps,
-                                        endpoint=False)
+        inference_indices = np.linspace(0, self.config.pcm_timesteps, num=num_inference_steps, endpoint=False)
         inference_indices = np.floor(inference_indices).astype(np.int64)
         inference_indices = torch.from_numpy(inference_indices).long()
 
         self.sigmas_ = self.sigmas[inference_indices]
         timesteps = self.sigmas_ * self.config.num_train_timesteps
         self.timesteps = timesteps.to(device=device)
-        self.sigmas_ = torch.cat(
-            [self.sigmas_,
-             torch.zeros(1, device=self.sigmas_.device)])
+        self.sigmas_ = torch.cat([self.sigmas_, torch.zeros(1, device=self.sigmas_.device)])
         self._step_index = None
         self._begin_index = None
 
@@ -208,10 +195,9 @@ class PCMFMScheduler(SchedulerMixin, ConfigMixin):
 
         if (isinstance(timestep, int) or isinstance(timestep, torch.IntTensor)
                 or isinstance(timestep, torch.LongTensor)):
-            raise ValueError((
-                "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
-                " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
-                " one of the `scheduler.timesteps` as a timestep."), )
+            raise ValueError(("Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+                              " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
+                              " one of the `scheduler.timesteps` as a timestep."), )
 
         if self.step_index is None:
             self._init_step_index(timestep)
@@ -241,18 +227,14 @@ class EulerSolver:
 
     def __init__(self, sigmas, timesteps=1000, euler_timesteps=50):
         self.step_ratio = timesteps // euler_timesteps
-        self.euler_timesteps = (np.arange(1, euler_timesteps + 1) *
-                                self.step_ratio).round().astype(np.int64) - 1
-        self.euler_timesteps_prev = np.asarray(
-            [0] + self.euler_timesteps[:-1].tolist())
+        self.euler_timesteps = (np.arange(1, euler_timesteps + 1) * self.step_ratio).round().astype(np.int64) - 1
+        self.euler_timesteps_prev = np.asarray([0] + self.euler_timesteps[:-1].tolist())
         self.sigmas = sigmas[self.euler_timesteps]
-        self.sigmas_prev = np.asarray(
-            [sigmas[0]] + sigmas[self.euler_timesteps[:-1]].tolist()
-        )  # either use sigma0 or 0
+        self.sigmas_prev = np.asarray([sigmas[0]] +
+                                      sigmas[self.euler_timesteps[:-1]].tolist())  # either use sigma0 or 0
 
         self.euler_timesteps = torch.from_numpy(self.euler_timesteps).long()
-        self.euler_timesteps_prev = torch.from_numpy(
-            self.euler_timesteps_prev).long()
+        self.euler_timesteps_prev = torch.from_numpy(self.euler_timesteps_prev).long()
         self.sigmas = torch.from_numpy(self.sigmas)
         self.sigmas_prev = torch.from_numpy(self.sigmas_prev)
 
@@ -265,10 +247,8 @@ class EulerSolver:
         return self
 
     def euler_step(self, sample, model_pred, timestep_index):
-        sigma = extract_into_tensor(self.sigmas, timestep_index,
-                                    model_pred.shape)
-        sigma_prev = extract_into_tensor(self.sigmas_prev, timestep_index,
-                                         model_pred.shape)
+        sigma = extract_into_tensor(self.sigmas, timestep_index, model_pred.shape)
+        sigma_prev = extract_into_tensor(self.sigmas_prev, timestep_index, model_pred.shape)
         x_prev = sample + (sigma_prev - sigma) * model_pred
         return x_prev
 
@@ -280,29 +260,20 @@ class EulerSolver:
         multiphase,
         is_target=False,
     ):
-        inference_indices = np.linspace(0,
-                                        len(self.euler_timesteps),
-                                        num=multiphase,
-                                        endpoint=False)
+        inference_indices = np.linspace(0, len(self.euler_timesteps), num=multiphase, endpoint=False)
         inference_indices = np.floor(inference_indices).astype(np.int64)
-        inference_indices = (torch.from_numpy(inference_indices).long().to(
-            self.euler_timesteps.device))
-        expanded_timestep_index = timestep_index.unsqueeze(1).expand(
-            -1, inference_indices.size(0))
+        inference_indices = (torch.from_numpy(inference_indices).long().to(self.euler_timesteps.device))
+        expanded_timestep_index = timestep_index.unsqueeze(1).expand(-1, inference_indices.size(0))
         valid_indices_mask = expanded_timestep_index >= inference_indices
-        last_valid_index = valid_indices_mask.flip(dims=[1]).long().argmax(
-            dim=1)
+        last_valid_index = valid_indices_mask.flip(dims=[1]).long().argmax(dim=1)
         last_valid_index = inference_indices.size(0) - 1 - last_valid_index
         timestep_index_end = inference_indices[last_valid_index]
 
         if is_target:
-            sigma = extract_into_tensor(self.sigmas_prev, timestep_index,
-                                        sample.shape)
+            sigma = extract_into_tensor(self.sigmas_prev, timestep_index, sample.shape)
         else:
-            sigma = extract_into_tensor(self.sigmas, timestep_index,
-                                        sample.shape)
-        sigma_prev = extract_into_tensor(self.sigmas_prev, timestep_index_end,
-                                         sample.shape)
+            sigma = extract_into_tensor(self.sigmas, timestep_index, sample.shape)
+        sigma_prev = extract_into_tensor(self.sigmas_prev, timestep_index_end, sample.shape)
         x_prev = sample + (sigma_prev - sigma) * model_pred
 
         return x_prev, timestep_index_end