Complete Model Config design for VAEs

Co-authored-by: Will Lin <wlsaidhi@gmail.com>

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

@@ -0,0 +1,29 @@
+name: 🐞 Bug report
+description: Create a report to help us reproduce and fix the bug
+title: "[Bug] "
+labels: ['Bug']
+
+body:
+- type: textarea
+  attributes:
+    label: Environment
+    description: |
+      Please share your environment with us. You can run the command **python fastvideo/utils/env_utils.py** and copy-paste its output below.
+    placeholder: FastVideo version, platform, python version, cuda version...
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Describe the bug
+    description: A clear and concise description of what the bug is.
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Reproduction
+    description: |
+      What command or script did you run? Which **model** are you using?
+    placeholder: |
+      A placeholder for the command.
+  validations:
+    required: true

.github/ISSUE_TEMPLATE/2-feature-request.yml

@@ -0,0 +1,17 @@
+name: 🚀 Feature request
+description: Suggest an idea for this project
+title: "[Feature] "
+
+body:
+- type: textarea
+  attributes:
+    label: Motivation
+    description: |
+      A clear and concise description of the motivation of the feature.
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Related resources
+    description: |
+      If there is an official code release or third-party implementations, please also provide the information here, which would be very helpful.

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1 @@
+blank_issues_enabled: false

.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.yml

@@ -0,0 +1,78 @@
+name: Build and Push Docker Image
+
+on:
+  workflow_dispatch:  # Only manual triggers
+
+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: ${{ secrets.GITHUB_TOKEN }}
+      
+      - name: Extract metadata for Docker
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ghcr.io/${{ github.repository }}/fastvideo-dev
+          tags: |
+            type=raw,value=latest
+            type=sha,format=short
+      
+      - name: Build and push Docker image
+        id: build-push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+      
+      - name: Success message
+        run: |
+          echo "✅ Image successfully built and pushed to ghcr.io/${{ github.repository }}/fastvideo-dev:latest"
+          echo "To run tests with this image, manually trigger the 'Run Tests' workflow." 

.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 +1,295 @@
+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: ${{ steps.filter.outputs.transformer-test }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dorny/paths-filter@v3
+        id: filter
+        with:
+          filters: |
+            encoder-test:
+              - 'fastvideo/v1/models/encoders/**'
+              - 'fastvideo/v1/models/loaders/**'
+              - 'fastvideo/v1/tests/encoders/**'
+            vae-test:
+              - 'fastvideo/v1/models/vaes/**'
+              - 'fastvideo/v1/models/loaders/**'
+              - 'fastvideo/v1/tests/vaes/**'
+            transformer-test:
+              - 'fastvideo/v1/models/dits/**'
+              - 'fastvideo/v1/models/loaders/**'
+              - 'fastvideo/v1/tests/transformers/**'
+
+  encoder-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && needs.change-filter.outputs.encoder-test == 'true') || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_encoder_test == 'true')
+    runs-on: ubuntu-latest
+    environment: runpod-runners
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Set up SSH key
+        run: |
+          mkdir -p ~/.ssh
+          echo "${{ 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: "encoder-test"
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+        timeout-minutes: 30
+        run: >-
+          python .github/scripts/runpod_api.py 
+          --gpu-type "NVIDIA A40"
+          --gpu-count 1
+          --volume-size 100
+          --image "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
+          --test-command "pip install -e .[test] && pytest ./fastvideo/v1/tests/encoders -s"
+
+      - name: Terminate RunPod Instances
+        if: ${{ always() }}
+        env:
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+          JOB_ID: "encoder-test"
+        run: python .github/scripts/runpod_cleanup.py
+
+  vae-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && needs.change-filter.outputs.vae-test == 'true') || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_vae_test == 'true')
+    runs-on: ubuntu-latest
+    environment: runpod-runners
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Set up SSH key
+        run: |
+          mkdir -p ~/.ssh
+          echo "${{ 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: "vae-test"
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+        timeout-minutes: 30
+        run: >-
+          python .github/scripts/runpod_api.py 
+          --gpu-type "NVIDIA A40"
+          --gpu-count 1
+          --volume-size 100
+          --image "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
+          --test-command "pip install -e .[test] && pytest ./fastvideo/v1/tests/vaes -s"
+
+      - name: Terminate RunPod Instances
+        if: ${{ always() }}
+        env:
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+          JOB_ID: "vae-test"
+        run: python .github/scripts/runpod_cleanup.py
+
+  transformer-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && needs.change-filter.outputs.transformer-test == 'true') || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_transformer_test == 'true')
+    runs-on: ubuntu-latest
+    environment: runpod-runners
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Set up SSH key
+        run: |
+          mkdir -p ~/.ssh
+          echo "${{ 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: "transformer-test"
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+        timeout-minutes: 30
+        run: >-
+          python .github/scripts/runpod_api.py 
+          --gpu-type "NVIDIA L40S"
+          --gpu-count 1
+          --volume-size 100
+          --image "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
+          --test-command "pip install -e .[test] && pytest ./fastvideo/v1/tests/transformers -s"
+
+      - name: Terminate RunPod Instances
+        if: ${{ always() }}
+        env:
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+          JOB_ID: "transformer-test"
+        run: python .github/scripts/runpod_cleanup.py
+      
+  ssim-test:
+    needs: change-filter
+    if: >-
+      (github.event_name != 'workflow_dispatch' && github.event.pull_request.draft == false) || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_ssim_test == 'true')
+    runs-on: ubuntu-latest
+    environment: runpod-runners
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Set up SSH key
+        run: |
+          mkdir -p ~/.ssh
+          echo "${{ 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: "ssim-test"
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+        timeout-minutes: 45
+        run: >-
+          python .github/scripts/runpod_api.py 
+          --gpu-type "NVIDIA A40"
+          --gpu-count 2
+          --disk-size 200
+          --volume-size 200
+          --image "ghcr.io/${{ github.repository }}/${{ github.event.inputs.custom_image || 'fastvideo-dev:latest' }}"
+          --test-command "pip install -e .[test] && pytest ./fastvideo/v1/tests/ssim -vs"
+
+      - name: Terminate RunPod Instances
+        if: ${{ always() }}
+        env:
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+          JOB_ID: "ssim-test"
+        run: python .github/scripts/runpod_cleanup.py
+
+  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"]' # JSON array of job IDs
+          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/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

@@ -0,0 +1,31 @@
+name: Run Tests
+
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'  # or any version you need
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip setuptools wheel
+          pip install torch
+          pip install packaging ninja
+          pip install -e .
+          pip install pytest
+
+      - name: Run Pytest
+        run: |
+          pytest --ignore csrc/sliding_tile_attention/test

.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
@@ -20,29 +14,6 @@ wandb/
 *.pt
 cache_dir/
 wandb/
-test*
-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/
@@ -52,3 +23,39 @@ 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/
+
+# 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 +1,81 @@
+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/.*|
+    fastvideo/v1/examples/.*|
+    .github/workflows/fastvideo-publish.yml|
+    .github/workflows/sta-publish.yml
+  )
+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

Dockerfile

@@ -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==2.7.0.post2 --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

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

@@ -1,162 +1,263 @@
-# FastVideo
-
 <div align="center">
-  <a href=""><img src="https://img.shields.io/static/v1?label=API:H100&message=Replicate&color=pink"></a> &ensp;
-  <a href=""><img src="https://img.shields.io/static/v1?label=Discuss&message=Discord&color=purple&logo=discord"></a> &ensp;
-</div>
-<br>
-<div align="center">
-<img src=assets/logo.png width="50%"/>
+<img src=assets/logo.jpg width="30%"/>
 </div>
 
-FastVideo is a scalable framework for post-training video diffusion models, addressing the growing challenges of fine-tuning, distillation, and inference as model sizes and sequence lengths increase. As a first step, it provides an efficient script for distilling and fine-tuning the 10B Mochi model, with plans to expand features and support for more models.
+FastVideo is a lightweight framework for accelerating large video diffusion models.
 
-### Features
+<p align="center">
+    | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</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>
 
-- FastMochi, a distilled Mochi model that can generate videos with merely 8 sampling steps.
-- Finetuning with FSDP (both master weight and ema weight), sequence parallelism, and selective gradient checkpointing.
-- LoRA coupled with pecomputed the latents and text embedding for minumum memory consumption.
-- Finetuning with both image and videos.
+https://github.com/user-attachments/assets/79af5fb8-707c-4263-b153-9ab2a01d3ac1
+
+FastVideo currently offers: (with more to come)
+
+- [NEW!] [Sliding Tile Attention](https://hao-ai-lab.github.io/blogs/sta/).
+- 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.
+- 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.
 
 ## Change Log
+- ```2025/02/20```: FastVideo now supports STA on [StepVideo](https://github.com/stepfun-ai/Step-Video-T2V) with 3.4X speedup!
+- ```2025/02/18```: Release the inference code and kernel for [Sliding Tile Attention](https://hao-ai-lab.github.io/blogs/sta/).
+- ```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.
 
-
-- ```2024/12/06```: `FastMochi` v0.0.1 is released.
-
-
-## Fast and High-Quality Text-to-video Generation
-
-### 8-Step Results of FastMochi
-
-<table class="center">
-  <td><img src=assets/8steps/1.gif width="320"></td></td>
-  <td><img src=assets/8steps/2.gif width="320"></td></td></td>
-  <tr>
-  <td style="text-align:center;" width="320">tmp</td>
-  <td style="text-align:center;" width="320">tmp</td>
-  <tr>
-</table >
-
-
-## Table of Contents
-
-Jump to a specific section:
-
-- [🔧 Installation](#-installation)
-- [🚀 Inference](#-inference)
-- [🎯 Distill](#-distill)
-- [⚡ Finetune](#-lora-finetune)
-
-
-## 🔧 Installation
+## 🔧 Installation from source
+The code is tested on Python 3.10-3.12, CUDA 12.4 and H100.
 
 ```
-conda create -n fastmochi python=3.10.0 -y && conda activate fastmochi
-pip3 install torch==2.5.0 torchvision --index-url https://download.pytorch.org/whl/cu121
-pip install packaging ninja && pip install flash-attn==2.7.0.post2 --no-build-isolation 
-pip install "git+https://github.com/huggingface/diffusers.git@bf64b32652a63a1865a0528a73a13652b201698b"
-git clone https://github.com/hao-ai-lab/FastVideo.git
-cd FastVideo && pip install -e .
+# Clone FastVideo
+git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
+
+# Install FastVideo
+pip install -e .
+
+# Install Flash Attention (optional)
+pip install flash-attn==2.7.0.post2
 ```
 
+To try Sliding Tile Attention (optional), please follow the instruction in [csrc/sliding_tile_attention/README.md](csrc/sliding_tile_attention/README.md) to install STA.
 
+You can also install the Sliding Tile Attention package using
 
+```
+pip install st_attn==0.0.4
+```
 
 ## 🚀 Inference
+### 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).
 
-Use [scripts/download_hf.py](scripts/download_hf.py) to download the hugging-face style model to a local directory. Use it like this:
 ```bash
-python scripts/download_hf.py --repo_id=FastVideo/FastMochi --local_dir=data/FastMochi --repo_type=model
+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:
 
-Start the gradio UI with
-```
-python fastvideo/demo/gradio_web_demo.py --model_path data/FastMochi
+```bash
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/hunyuan --local_dir=data/hunyuan --repo_type=model
 ```
 
-We also provide CLI inference script featured with sequence parallelism.
+We provide two examples in the following script to run inference with STA + [TeaCache](https://github.com/ali-vilab/TeaCache) and STA only.
 
-```
-export NUM_GPUS=4
-
-torchrun --nnodes=1 --nproc_per_node=$NUM_GPUS \
-    fastvideo/sample/sample_t2v_mochi.py \
-    --model_path data/FastMochi \
-    --prompt_path assets/prompt.txt \
-    --num_frames 163 \
-    --height 480 \
-    --width 848 \
-    --num_inference_steps 8 \
-    --guidance_scale 1.5 \
-    --output_path outputs_video/demo_video \
-    --seed 12345 \
-    --scheduler_type "pcm_linear_quadratic" \
-    --linear_threshold 0.1 \
-    --linear_range 0.75
+```bash
+sh scripts/inference/inference_hunyuan_STA.sh
 ```
 
-For the mochi style, simply following the scripts list in mochi repo.
+### 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.
 
-```
-git clone https://github.com/genmoai/mochi.git
-cd mochi
+### 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.
 
-# install env
-...
-
-python3 ./demos/cli.py --model_dir weights/ --cpu_offload
+```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
+```
 
 ## 🎯 Distill
-
-## 💰Hardware requirement
-
--  VRAM is required for both distill 10B mochi model
-
-To launch distillation, you will first need to prepare data in the following formats
+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
-asset/example_data
-├── AAA.txt
-├── AAA.png
-├── BCC.txt
-├── BCC.png
-├── ......
-├── CCC.txt
-└── CCC.png
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/HD-Mixkit-Finetune-Hunyuan --local_dir=data/HD-Mixkit-Finetune-Hunyuan --repo_type=dataset
 ```
 
-We provide a dataset example here. First download testing data. Use [scripts/download_hf.py](scripts/download_hf.py) to download the data to a local directory. Use it like this:
+Next, download the original model weights with:
+
 ```bash
-python scripts/download_hf.py --repo_id=Stealths-Video/Merge-425-Data --local_dir=data/Merge-425-Data --repo_type=dataset
-python scripts/download_hf.py --repo_id=Stealths-Video/validation_embeddings --local_dir=data/validation_embeddings --repo_type=dataset
+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
 ```
 
-Then the distillation can be launched by:
+To launch the distillation process, use the following commands:
 
 ```
-bash scripts/distill_t2v.sh
+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:
 
-## ⚡ Lora Finetune
+```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):
 
-## 💰Hardware requirement
-
--  VRAM is required for both distill 10B mochi model
-
-To launch finetuning, you will first need to prepare data in the following formats.
-
-
-
-Then the finetuning can be launched by:
+Then you can run the finetune with:
 
 ```
-bash scripts/lora_finetune.sh
+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.
+
+## 📑 Development Plan
+
+- More distillation methods
+  - [ ] Add Distribution Matching Distillation
+- More models support
+  - [ ] Add CogvideoX model
+- Code update
+  - [ ] fp8 support
+  - [ ] faster load model and save model support
+
+## 🤝 Contributing
+
+We welcome all contributions. Please run `bash format.sh --all` 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.
+
 ## Acknowledgement
-We learned from and reused code from the following projects: [PCM](https://github.com/G-U-N/Phased-Consistency-Model), [diffusers](https://github.com/huggingface/diffusers), and [OpenSoraPlan](https://github.com/PKU-YuanGroup/Open-Sora-Plan).
+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 thank MBZUAI and Anyscale 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},
+}
+```

assets/prompt.txt

@@ -1,9 +1,8 @@
-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.
-En "The Matrix", Neo, interpretado por Keanu Reeves, personifica la lucha contra un sistema opresor a través de su icónica imagen, que incluye unos anteojos oscuros. Estos lentes no son solo un accesorio de moda; representan una barrera entre la realidad y la percepción. Al usar estos anteojos, Neo se sumerge en un mundo donde la verdad se oculta detrás de ilusiones y engaños. La oscuridad de los lentes simboliza la ignorancia y el control que las máquinas tienen sobre la humanidad, mientras que su propia búsqueda de la verdad lo lleva a descubrir sus auténticos poderes. La escena en que se los pone se convierte en un momento crucial, marcando su transformación de un simple programador a "El Elegido". Esta imagen se ha convertido en un ícono cultural, encapsulando el mensaje de que, al enfrentar la oscuridad, podemos encontrar la luz que nos guía hacia la libertad. Así, los anteojos de Neo se convierten en un símbolo de resistencia y autoconocimiento en un mundo manipulado.
-Medium close up. Low-angle shot. A woman in a 1950s retro dress sits in a diner bathed in neon light, surrounded by classic decor and lively chatter. The camera starts with a medium shot of her sitting at the counter, then slowly zooms in as she blows a shiny pink bubblegum bubble. The bubble swells dramatically before popping with a soft, playful burst. The scene is vibrant and nostalgic, evoking the fun and carefree spirit of the 1950s.
-Will Smith eats noodles.
-A short clip of the blonde woman taking a sip from her whiskey glass, her eyes locking with the camera as she smirks playfully. The background shows a group of people laughing and enjoying the party, with vibrant neon signs illuminating the space. The shot is taken in a way that conveys the feeling of a tipsy, carefree night out. The camera then zooms in on her face as she winks, creating a cheeky, flirtatious vibe.
-A superintelligent humanoid robot waking up. The robot has a sleek metallic body with futuristic design features. Its glowing red eyes are the focal point, emanating a sharp, intense light as it powers on. The scene is set in a dimly lit, high-tech laboratory filled with glowing control panels, robotic arms, and holographic screens. The setting emphasizes advanced technology and an atmosphere of mystery. The ambiance is eerie and dramatic, highlighting the moment of awakening and the robot's immense intelligence. Photorealistic style with a cinematic, dark sci-fi aesthetic. Aspect ratio: 16:9 --v 6.1
-A chimpanzee lead vocalist singing into a microphone on stage. The camera zooms in to show him singing. There is a spotlight on him.
+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.
+A lone hiker stands atop a towering cliff, silhouetted against the vast horizon. The rugged landscape stretches endlessly beneath, its earthy tones blending into the soft blues of the sky. The scene captures the spirit of exploration and human resilience. High angle, dynamic framing, with soft natural lighting emphasizing the grandeur of nature.
+A hand with delicate fingers picks up a bright yellow lemon from a wooden bowl filled with lemons and sprigs of mint against a peach-colored background. The hand gently tosses the lemon up and catches it, showcasing its smooth texture. A beige string bag sits beside the bowl, adding a rustic touch to the scene. Additional lemons, one halved, are scattered around the base of the bowl. The even lighting enhances the vibrant colors and creates a fresh, inviting atmosphere.
+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.
+A superintelligent humanoid robot waking up. The robot has a sleek metallic body with futuristic design features. Its glowing red eyes are the focal point, emanating a sharp, intense light as it powers on. The scene is set in a dimly lit, high-tech laboratory filled with glowing control panels, robotic arms, and holographic screens. The setting emphasizes advanced technology and an atmosphere of mystery. The ambiance is eerie and dramatic, highlighting the moment of awakening and the robots immense intelligence. Photorealistic style with a cinematic, dark sci-fi aesthetic. Aspect ratio: 16:9 --v 6.1
+fox in the forest close-up quickly turned its head to the left
+Man walking his dog in the woods on a hot sunny day
+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.

cog.yaml

@@ -0,0 +1,24 @@
+# Configuration for Cog ⚙️
+# Reference: https://cog.run/yaml
+
+build:
+  gpu: true
+  cuda: "12.1"
+  python_version: "3.10"
+  python_packages:
+    - "torch==2.4.0"
+    - "torchvision"
+    - "ninja==1.11.1.3"
+    - "transformers==4.46.1"
+    - "git+https://github.com/huggingface/diffusers.git@bf64b32652a63a1865a0528a73a13652b201698b"
+    - "accelerate==1.0.1"
+    - "safetensors==0.4.5"
+    - "peft==0.13.2"
+    - "packaging==24.2"
+    - "git+https://github.com/hao-ai-lab/FastVideo"
+
+  run:
+    - FLASH_ATTENTION_SKIP_CUDA_BUILD=TRUE pip install flash-attn --no-build-isolation
+    - curl -o /usr/local/bin/pget -L "https://github.com/replicate/pget/releases/latest/download/pget_$(uname -s)_$(uname -m)" && chmod +x /usr/local/bin/pget
+
+predict: "predict.py:Predictor"

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 +1,96 @@
+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 flash_attn_interface import flash_attn_func
+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.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

@@ -1,65 +1,85 @@
+import argparse
+import os
+import tempfile
+
 import gradio as gr
 import torch
-from fastvideo.model.pipeline_mochi import MochiPipeline
-from fastvideo.model.modeling_mochi import MochiTransformer3DModel
 from diffusers import FlowMatchEulerDiscreteScheduler
 from diffusers.utils import export_to_video
+
 from fastvideo.distill.solver import PCMFMScheduler
-import tempfile
-import os
-import argparse
+from fastvideo.models.mochi_hf.modeling_mochi import MochiTransformer3DModel
+from fastvideo.models.mochi_hf.pipeline_mochi import MochiPipeline
+
 
 def init_args():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--prompts", nargs='+', default=[])
-    parser.add_argument("--num_frames", type=int, default=163)
+    parser.add_argument("--prompts", nargs="+", default=[])
+    parser.add_argument("--num_frames", type=int, default=25)
     parser.add_argument("--height", type=int, default=480)
     parser.add_argument("--width", type=int, default=848)
-    parser.add_argument("--num_inference_steps", type=int, default=64)
+    parser.add_argument("--num_inference_steps", type=int, default=8)
     parser.add_argument("--guidance_scale", type=float, default=4.5)
     parser.add_argument("--model_path", type=str, default="data/mochi")
-    parser.add_argument("--seed", type=int, default=42)
+    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="euler")
+    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=100)
-    parser.add_argument("--linear_threshold", type=float, default=0.025)
-    parser.add_argument("--linear_range", type=float, default=0.5)
+    parser.add_argument("--num_euler_timesteps", type=int, default=50)
+    parser.add_argument("--linear_threshold", type=float, default=0.1)
+    parser.add_argument("--linear_range", type=float, default=0.75)
     parser.add_argument("--cpu_offload", action="store_true")
     return parser.parse_args()
 
+
 def load_model(args):
-    device = "cuda" if torch.cuda.is_available() else "cpu"
     if args.scheduler_type == "euler":
         scheduler = FlowMatchEulerDiscreteScheduler()
     else:
-        scheduler = PCMFMScheduler(1000, args.shift, args.num_euler_timesteps, False, args.linear_threshold, args.linear_range)
-    
+        linear_quadratic = True if "linear_quadratic" in args.scheduler_type else False
+        scheduler = PCMFMScheduler(
+            1000,
+            args.shift,
+            args.num_euler_timesteps,
+            linear_quadratic,
+            args.linear_threshold,
+            args.linear_range,
+        )
+
     if 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.enable_vae_tiling()
-    pipe.to(device)
-    if args.cpu_offload:
-        pipe.enable_model_cpu_offload()
+    # pipe.to(device)
+    # if args.cpu_offload:
+    pipe.enable_sequential_cpu_offload()
     return pipe
 
-def generate_video(prompt, negative_prompt, use_negative_prompt, seed, guidance_scale, 
-                  num_frames, height, width, num_inference_steps, randomize_seed=False):
+
+def generate_video(
+    prompt,
+    negative_prompt,
+    use_negative_prompt,
+    seed,
+    guidance_scale,
+    num_frames,
+    height,
+    width,
+    num_inference_steps,
+    randomize_seed=False,
+):
     if randomize_seed:
-        seed = torch.randint(0, 1000000, (1,)).item()
-    
-    pipe = load_model(args)
-    print("load model successfully")
+        seed = torch.randint(0, 1000000, (1, )).item()
+
     generator = torch.Generator(device="cuda").manual_seed(seed)
-    
+
     if not use_negative_prompt:
         negative_prompt = None
-    
+
     with torch.autocast("cuda", dtype=torch.bfloat16):
         output = pipe(
             prompt=[prompt],
@@ -71,22 +91,24 @@ def generate_video(prompt, negative_prompt, use_negative_prompt, seed, guidance_
             guidance_scale=guidance_scale,
             generator=generator,
         ).frames[0]
-    
+
     output_path = os.path.join(tempfile.mkdtemp(), "output.mp4")
     export_to_video(output, output_path, fps=30)
     return output_path, 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."
+    "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.",
 ]
 
 args = init_args()
-
+pipe = load_model(args)
+print("load model successfully")
 with gr.Blocks() as demo:
-    gr.Markdown("# Mochi Video Generation Demo")
-    
+    gr.Markdown("# Fastvideo Mochi Video Generation Demo")
+
     with gr.Group():
         with gr.Row():
             prompt = gr.Text(
@@ -98,33 +120,54 @@ with gr.Blocks() as demo:
             )
             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=args.height)
+                height = gr.Slider(
+                    label="Height",
+                    minimum=256,
+                    maximum=1024,
+                    step=32,
+                    value=args.height,
+                )
                 width = gr.Slider(label="Width", minimum=256, maximum=1024, step=32, value=args.width)
-            
+
             with gr.Row():
-                num_frames = gr.Slider(label="Number of Frames", minimum=8, maximum=256, value=args.num_frames)
-                guidance_scale = gr.Slider(label="Guidance Scale", minimum=1, maximum=20, value=args.guidance_scale)
-                num_inference_steps = gr.Slider(label="Inference Steps", minimum=10, maximum=100, value=args.num_inference_steps)
-            
+                num_frames = gr.Slider(
+                    label="Number of Frames",
+                    minimum=21,
+                    maximum=163,
+                    value=args.num_frames,
+                )
+                guidance_scale = gr.Slider(
+                    label="Guidance Scale",
+                    minimum=1,
+                    maximum=12,
+                    value=args.guidance_scale,
+                )
+                num_inference_steps = gr.Slider(
+                    label="Inference Steps",
+                    minimum=4,
+                    maximum=100,
+                    value=args.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
+                visible=False,
             )
-            
+
             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")
-    
+
     gr.Examples(examples=examples, inputs=prompt)
-    
+
     use_negative_prompt.change(
         fn=lambda x: gr.update(visible=x),
         inputs=use_negative_prompt,
@@ -133,10 +176,20 @@ with gr.Blocks() as demo:
 
     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]
+        inputs=[
+            prompt,
+            negative_prompt,
+            use_negative_prompt,
+            seed,
+            guidance_scale,
+            num_frames,
+            height,
+            width,
+            num_inference_steps,
+            randomize_seed,
+        ],
+        outputs=[result, seed_output],
     )
 
 if __name__ == "__main__":
-    demo.queue(max_size=20).launch(server_name="0.0.0.0", server_port=7860)
+    demo.queue(max_size=20).launch(server_name="0.0.0.0", server_port=7860)

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

docs/Makefile

@@ -0,0 +1,24 @@
+# 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"

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/data_preprocess.md

@@ -0,0 +1,73 @@
+
+## 🧱 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
+```
+
+The preprocessed dataset will be stored in `Image-Vid-Finetune-Mochi` or `Image-Vid-Finetune-HunYuan` correspondingly.
+
+### Process your own dataset
+
+If you wish to create your own dataset for finetuning or distillation, please structure you video dataset in the following format:
+
+path_to_dataset_folder/
+├── media/
+│   ├── 0.jpg
+│   ├── 1.mp4
+│   ├── 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,
+
+```
+{
+    "path": "1.mp4",
+    "resolution": {
+      "width": 848,
+      "height": 480
+    },
+    "fps": 30.0,
+    "duration": 6.033333333333333,
+    "cap": [
+      "caption"
+    ]
+  }
+```
+
+Use a txt file (merge.txt) to contain the source folder for media and the JSON file for meta information:
+
+```
+path_to_media_source_foder,path_to_json_file
+```
+
+Adjust the `DATA_MERGE_PATH` and `OUTPUT_DIR` in `scripts/preprocess/preprocess_****_data.sh` accordingly and run:
+
+```
+bash scripts/preprocess/preprocess_****_data.sh
+```
+
+The preprocessed data will be put into the `OUTPUT_DIR` and the `videos2caption.json` can be used in finetune and distill scripts.

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 +1,25 @@
+sphinx==6.2.1
+sphinx-argparse==0.4.0
+sphinx-book-theme==1.0.1
+sphinx-copybutton==0.5.2
+sphinx-design==0.6.1
+sphinx-togglebutton==0.3.2
+myst-parser==3.0.1
+msgspec
+cloudpickle
+
+# packages to install to build the documentation
+cachetools
+pydantic >= 2.8
+-f https://download.pytorch.org/whl/cpu
+torch
+py-cpuinfo
+transformers
+mistral_common >= 1.5.4
+aiohttp
+starlette
+openai # Required by docs/source/serving/openai_compatible_server.md's vllm.entrypoints.openai.cli_args
+fastapi # Required by docs/source/serving/openai_compatible_server.md's vllm.entrypoints.openai.cli_args
+partial-json-parser # Required by docs/source/serving/openai_compatible_server.md's vllm.entrypoints.openai.cli_args
+requests
+zmq

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/conf.py

@@ -0,0 +1,260 @@
+# 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 inspect
+import logging
+import os
+import sys
+from typing import Optional
+
+import requests
+from sphinx.ext import autodoc
+
+logger = logging.getLogger(__name__)
+sys.path.append(os.path.abspath("../.."))
+
+# -- 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",
+    "sphinx.ext.autodoc",
+    "sphinx.ext.autosummary",
+    "myst_parser",
+    "sphinxarg.ext",
+    "sphinx_design",
+    "sphinx_togglebutton",
+]
+myst_enable_extensions = [
+    "colon_fence",
+]
+
+# 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,
+}
+# 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
+    module = info['module']
+
+    # try to determine the correct file and line number to link to
+    obj = sys.modules[module]
+
+    # get as specific as we can
+    lineno: int = 0
+    filename: str = ""
+    try:
+        for part in info['fullname'].split('.'):
+            obj = getattr(obj, part)
+
+            if not (inspect.isclass(obj) or inspect.isfunction(obj)
+                    or inspect.ismethod(obj)):
+                obj = obj.__class__  # type: ignore[assignment]
+
+            lineno = inspect.getsourcelines(obj)[1]
+            filename = (inspect.getsourcefile(obj)
+                        or f"{filename}.py").split("FastVideo/", 1)[1]
+    except Exception:
+        # For some things, like a class member, won't work, so
+        # we'll use the line number of the parent (the class)
+        pass
+
+    if filename.startswith("checkouts/"):
+        # a PR build on readthedocs
+        pr_number = filename.split("/")[1]
+        filename = filename.split("/", 2)[2]
+        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",
+    "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)
+
+
+class MockedClassDocumenter(autodoc.ClassDocumenter):
+    """Remove note about base class when a class is derived from object."""
+
+    def add_line(self, line: str, source: str, *lineno: int) -> None:
+        if line == "   Bases: :py:class:`object`":
+            return
+        super().add_line(line, source, *lineno)
+
+
+autodoc.ClassDocumenter = MockedClassDocumenter
+
+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),
+}
+
+autodoc_preserve_defaults = True
+autodoc_warningiserror = True
+
+navigation_with_keys = False

docs/source/developer_guide/overview.md

@@ -0,0 +1,138 @@
+(developer-guide)=
+
+# 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==2.7.0.post2 --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/
+```
+
+---
+## 🐳 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/
+```
+
+---
+
+## 📦 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/generate_examples.py

@@ -0,0 +1,249 @@
+# 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 / "fastvideo/v1/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():
+    # Create the EXAMPLE_DOC_DIR if it doesn't exist
+    if not EXAMPLE_DOC_DIR.exists():
+        EXAMPLE_DOC_DIR.mkdir(parents=True)
+
+    # Create empty indices
+    examples_index = Index(
+        path=EXAMPLE_DOC_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 stored in reverse order because they are inserted into
+    # examples_index.documents at index 0 in order
+    category_indices = {
+        # "other":
+        # Index(
+        #     path=EXAMPLE_DOC_DIR / "examples_other_index.md",
+        #     title="Other",
+        #     description=
+        #     "Other examples that don't strongly fit into the online or offline serving categories.",  # noqa: E501
+        #     caption="Examples",
+        # ),
+        # "online_serving":
+        # Index(
+        #     path=EXAMPLE_DOC_DIR / "examples_online_serving_index.md",
+        #     title="Online Serving",
+        #     description=
+        #     "Online serving examples demonstrate how to use FastVideo in an online setting, where the model is queried for predictions in real-time.",  # noqa: E501
+        #     caption="Examples",
+        # ),
+        "inference":
+        Index(
+            path=EXAMPLE_DOC_DIR / "examples_inference_index.md",
+            title="Inference",
+            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",
+        ),
+    }
+
+    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
+    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))
+
+    # Generate the example documentation
+    for example in sorted(examples, key=lambda e: e.path.stem):
+        print(example)
+        doc_path = EXAMPLE_DOC_DIR / f"{example.path.stem}.md"
+        with open(doc_path, "w+") as f:
+            f.write(example.generate())
+        # Add the example to the appropriate index
+        assert example.category is not None
+        index = category_indices.get(example.category, examples_index)
+        index.documents.append(example.path.stem)
+
+    # Generate the index files
+    for category_index in category_indices.values():
+        if category_index.documents:
+            examples_index.documents.insert(0, category_index.path.name)
+            with open(category_index.path, "w+") as f:
+                f.write(category_index.generate())
+
+    with open(examples_index.path, "w+") as f:
+        f.write(examples_index.generate())

docs/source/getting_started/installation.md

@@ -0,0 +1,95 @@
+(fastvideo-installation)=
+
+# 🔧 Installation
+
+FastVideo currently only supports Linux and NVIDIA CUDA GPUs.
+
+FastVideo has been tested on the following GPUs, but it should work on any GPUs that supports CUDA 12.4+, please create an issue if you discover any issues:
+- RTX 4090
+- A40
+- L40S
+- A100
+- H100
+
+## Requirements
+
+- OS: Linux
+- Python: 3.10-3.12
+- CUDA 12.4+ (Untested on CUDA < 12.4)
+
+## Installation Options
+
+### Option 1: Quick Install
+
+```bash
+pip install fastvideo
+```
+
+### Option 2: Installation from Source
+
+We recommend using a Python environment such as Conda.
+
+#### 1. [Optional] 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. [Optional] Create and activate a Conda environment for FastVideo
+
+```bash
+conda create -n fastvideo python=3.10 -y
+conda activate fastvideo
+```
+
+#### 3. Clone the FastVideo repository
+
+```bash
+git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
+```
+
+#### 4. Install FastVideo
+
+Basic installation:
+
+```bash
+pip install -e .
+```
+
+## Optional Dependencies
+
+### Flash Attention
+
+```bash
+pip install flash-attn==2.7.0.post2 --no-build-isolation
+```
+
+### Sliding Tile Attention (STA) (Requires CUDA 12.4+ and H100)
+
+To try Sliding Tile Attention (optional), please follow the instructions in [csrc/sliding_tile_attention/README.md](#sta-installation) to install STA.
+
+## Development Environment Setup
+
+If you're planning to contribute to FastVideo please see the following page:
+[Contributor Guide](#developer-guide)
+
+## Hardware Requirements
+
+### For Basic Inference
+- NVIDIA GPU with CUDA support
+- Minimum 20GB VRAM for quantized models (e.g., single RTX 4090)
+
+### 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/index.md

@@ -0,0 +1,89 @@
+# 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 lightweight framework for accelerating large video diffusion models.
+</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>
+:::
+
+FastVideo is a lightweight framework for accelerating large video diffusion models developed by the [Hao AI Lab](https://hao-ai-lab.github.io/).
+
+<div style="text-align: center;">
+  <video controls width="800">
+    <source src="https://github.com/user-attachments/assets/79af5fb8-707c-4263-b153-9ab2a01d3ac1" type="video/mp4">
+    Your browser does not support the video tag.
+  </video>
+</div>
+
+FastVideo currently offers: (with more to come)
+
+- [NEW!] [Sliding Tile Attention](https://hao-ai-lab.github.io/blogs/sta/).
+- 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.
+- 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.
+
+## Documentation
+
+% How to start using vLLM?
+
+:::{toctree}
+:caption: Getting Started
+:maxdepth: 1
+
+getting_started/installation
+getting_started/examples/examples_index
+:::
+
+% What is STA Kernel?
+
+:::{toctree}
+:caption: Sliding Tile Attention
+:maxdepth: 1
+
+sliding_tile_attention/installation
+sliding_tile_attention/usage
+sliding_tile_attention/test
+sliding_tile_attention/demo
+:::
+
+:::{toctree}
+:caption: Inference
+:maxdepth: 1
+
+inference/wanvideo
+inference/stepvideo
+inference/hunyuanvideo
+inference/fasthunyuan
+inference/fastmochi
+:::
+
+:::{toctree}
+:caption: Developer Guide
+:maxdepth: 1
+
+developer_guide/overview
+:::
+
+## Indices and tables
+
+- {ref}`genindex`
+- {ref}`modindex`

docs/source/inference/fasthunyuan.md

@@ -0,0 +1,33 @@
+(fasthunyuan)=
+
+# FastHunyuan
+## 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).

docs/source/inference/fastmochi.md

@@ -0,0 +1,9 @@
+(fastmochi)=
+
+# 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/inference/hunyuanvideo.md

@@ -0,0 +1,18 @@
+(hunyuanvideo)=
+
+# HunyuanVideo
+## 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.

docs/source/inference/stepvideo.md

@@ -0,0 +1,16 @@
+(stepvideo)=
+
+# StepVideo
+## 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
+```

docs/source/inference/wanvideo.md

@@ -0,0 +1,44 @@
+(wanvideo)=
+
+# WanVideo
+## Inference T2V with WanVideo
+First, download the model:
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=Wan-AI/Wan2.1-T2V-1.3B-Diffusers --local_dir=YOUR_LOCAL_DIR --repo_type=model 
+```
+
+or
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=Wan-AI/Wan2.1-T2V-14B-Diffusers --local_dir=YOUR_LOCAL_DIR --repo_type=model 
+```
+
+Then run the inference using:
+
+```bash
+sh scripts/inference/v1_inference_wan.sh
+```
+
+Remember to set `MODEL_BASE` and `num_gpus` accordingly.
+
+## Inference I2V with WanVideo
+First, download the model:
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=Wan-AI/Wan2.1-I2V-14B-480P-Diffusers --local_dir=YOUR_LOCAL_DIR --repo_type=model
+```
+
+or
+
+```bash
+python scripts/huggingface/download_hf.py --repo_id=Wan-AI/Wan2.1-I2V-14B-720P-Diffusers --local_dir=YOUR_LOCAL_DIR --repo_type=model
+```
+
+Then run the inference using:
+
+```bash
+sh scripts/inference/v1_inference_wan_i2v.sh
+```
+
+Remember to set `MODEL_BASE` and `num_gpus` accordingly.

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 +1,25 @@
+(sta-installation)=
+
+# 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
+```

docs/source/sliding_tile_attention/test.md

@@ -0,0 +1,7 @@
+(sta-test)=
+
+# Test
+
+```bash
+python test/test_sta.py
+```

docs/source/sliding_tile_attention/usage.md

@@ -0,0 +1,17 @@
+(sta-usage)=
+
+# 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)
+
+```

examples/inference/basic/README.md

@@ -0,0 +1,3 @@
+# Basic
+
+The class provides the main python interface for using FastVideo's inference pipeline.

examples/inference/basic/basic.py

@@ -0,0 +1 @@
+print('Hello, world!')

fastvideo/__init__.py

@@ -0,0 +1,3 @@
+from fastvideo.v1.entrypoints.video_generator import VideoGenerator
+
+__all__ = ["VideoGenerator"]

fastvideo/data_preprocess/preprocess_text_embeddings.py

@@ -1,19 +1,23 @@
 import argparse
-import torch
-from accelerate.logging import get_logger
-from fastvideo.models.mochi_hf.pipeline_mochi import MochiPipeline
-from diffusers.utils import export_to_video
 import json
 import os
+
+import torch
 import torch.distributed as dist
+from accelerate.logging import get_logger
+from diffusers.utils import export_to_video
+from diffusers.video_processor import VideoProcessor
+from torch.utils.data import DataLoader, Dataset
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm
+
+from fastvideo.utils.load import load_text_encoder, load_vae
 
 logger = get_logger(__name__)
-from torch.utils.data import Dataset
-from torch.utils.data.distributed import DistributedSampler
-from torch.utils.data import DataLoader
 
 
 class T5dataset(Dataset):
+
     def __init__(
         self,
         json_path,
@@ -31,9 +35,7 @@ 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:
@@ -53,12 +55,9 @@ 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)
 
-    pipe = MochiPipeline.from_pretrained(args.model_path).to(device)
-    pipe.vae.enable_tiling()
+    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)
@@ -67,9 +66,10 @@ def main(args):
 
     latents_json_path = os.path.join(args.output_dir, "videos2caption_temp.json")
     train_dataset = T5dataset(latents_json_path, args.vae_debug)
-    sampler = DistributedSampler(
-        train_dataset, rank=local_rank, num_replicas=world_size, shuffle=True
-    )
+    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)
     train_dataloader = DataLoader(
         train_dataset,
         sampler=sampler,
@@ -78,32 +78,25 @@ def main(args):
     )
 
     json_data = []
-    for _, data in enumerate(train_dataloader):
+    for _, data in tqdm(enumerate(train_dataloader), disable=local_rank != 0):
         with torch.inference_mode():
-            with torch.autocast("cuda", dtype=torch.bfloat16):
-                prompt_embeds, prompt_attention_mask, _, _ = pipe.encode_prompt(
-                    prompt=data["caption"],
-                )
+            with torch.autocast("cuda", dtype=autocast_type):
+                prompt_embeds, prompt_attention_mask = text_encoder.encode_prompt(prompt=data["caption"], )
                 if args.vae_debug:
                     latents = data["latents"]
-                    video = pipe.vae.decode(latents.to(device), return_dict=False)[0]
-                    video = pipe.video_processor.postprocess_video(video)
+                    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)
                     print(f"sample {video_name} saved")
                     if args.vae_debug:
-                        export_to_video(video[idx], video_path, fps=30)
+                        export_to_video(video[idx], video_path, fps=fps)
                     item = {}
                     item["length"] = int(data["length"][idx])
                     item["latent_path"] = video_name + ".pt"
@@ -126,6 +119,7 @@ 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")
     # text encoder & vae & diffusion model
     parser.add_argument(
         "--dataloader_num_workers",

fastvideo/data_preprocess/preprocess_vae_latents.py

@@ -1,16 +1,16 @@
-from fastvideo.dataset import getdataset
-from torch.utils.data import DataLoader
-from fastvideo.utils.dataset_utils import Collate
 import argparse
-import torch
-from accelerate import Accelerator
-from accelerate.logging import get_logger
-from accelerate.utils import ProjectConfiguration
 import json
 import os
-from diffusers import AutoencoderKLMochi
+
+import torch
 import torch.distributed as dist
+from accelerate.logging import get_logger
+from torch.utils.data import DataLoader
 from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm
+
+from fastvideo.dataset import getdataset
+from fastvideo.utils.load import load_vae
 
 logger = get_logger(__name__)
 
@@ -19,25 +19,8 @@ def main(args):
     local_rank = int(os.getenv("RANK", 0))
     world_size = int(os.getenv("WORLD_SIZE", 1))
     print("world_size", world_size, "local rank", local_rank)
-    args.ae_stride_t, args.ae_stride_h, args.ae_stride_w = 4, 8, 8
-    args.ae_stride = args.ae_stride_h
-    patch_size_t, patch_size_h, patch_size_w = 1, 2, 2
-    args.patch_size = patch_size_h
-    args.patch_size_t, args.patch_size_h, args.patch_size_w = (
-        patch_size_t,
-        patch_size_h,
-        patch_size_w,
-    )
-    accelerator_project_config = ProjectConfiguration(
-        project_dir=args.output_dir, logging_dir=args.logging_dir
-    )
-    accelerator = Accelerator(
-        project_config=accelerator_project_config,
-    )
     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,
@@ -45,38 +28,30 @@ def main(args):
         num_workers=args.dataloader_num_workers,
     )
 
-    encoder_device = torch.device(f"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
-        )
-    vae = AutoencoderKLMochi.from_pretrained(args.model_path, subfolder="vae").to(
-        "cuda"
-    )
+        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)
     os.makedirs(os.path.join(args.output_dir, "latent"), exist_ok=True)
 
     json_data = []
-    for _, data in enumerate(train_dataloader):
+    for _, data in tqdm(enumerate(train_dataloader), disable=local_rank != 0):
         with torch.inference_mode():
-            with torch.autocast("cuda", dtype=torch.bfloat16):
-                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"
-                    )
-                    torch.save(latents[idx].to(torch.bfloat16), latent_path)
-                    item = {}
-                    item["length"] = latents[idx].shape[1]
-                    item["latent_path"] = video_name + ".pt"
-                    item["caption"] = data["text"][idx]
-                    json_data.append(item)
-                    print(f"{video_name} processed")
+            with torch.autocast("cuda", dtype=autocast_type):
+                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")
+                torch.save(latents[idx].to(torch.bfloat16), latent_path)
+                item = {}
+                item["length"] = latents[idx].shape[1]
+                item["latent_path"] = video_name + ".pt"
+                item["caption"] = data["text"][idx]
+                json_data.append(item)
+                print(f"{video_name} processed")
     dist.barrier()
     local_data = json_data
     gathered_data = [None] * world_size
@@ -91,6 +66,7 @@ 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("--num_frames", type=int, default=163)
     parser.add_argument(
@@ -105,9 +81,7 @@ 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("--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)
@@ -133,10 +107,8 @@ if __name__ == "__main__":
         "--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

@@ -0,0 +1,67 @@
+import argparse
+import os
+
+import torch
+import torch.distributed as dist
+from accelerate.logging import get_logger
+
+from fastvideo.utils.load import load_text_encoder
+
+logger = get_logger(__name__)
+
+
+def main(args):
+    local_rank = int(os.getenv("RANK", 0))
+    world_size = int(os.getenv("WORLD_SIZE", 1))
+    print("world_size", world_size, "local rank", local_rank)
+
+    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)
+
+    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
+    os.makedirs(os.path.join(args.output_dir, "validation"), exist_ok=True)
+    os.makedirs(
+        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)
+
+    with open(args.validation_prompt_txt, "r", encoding="utf-8") as file:
+        lines = file.readlines()
+    prompts = [line.strip() for line in lines]
+    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)
+                file_name = prompt.split(".")[0]
+                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",
+                    "prompt_attention_mask",
+                    f"{file_name}.pt",
+                )
+                torch.save(prompt_embeds[0], prompt_embed_path)
+                torch.save(prompt_attention_mask[0], prompt_attention_mask_path)
+                print(f"sample {file_name} saved")
+
+
+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("--validation_prompt_txt", type=str)
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default=None,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    args = parser.parse_args()
+    main(args)

fastvideo/dataset/__init__.py

@@ -1,14 +1,9 @@
-from transformers import AutoTokenizer
-
 from torchvision import transforms
 from torchvision.transforms import Lambda
+from transformers import AutoTokenizer
+
 from fastvideo.dataset.t2v_datasets import T2V_dataset
-from fastvideo.dataset.latent_datasets import LatentDataset
-from fastvideo.dataset.transform import (
-    Normalize255,
-    TemporalRandomCrop,
-    CenterCropResizeVideo,
-)
+from fastvideo.dataset.transform import CenterCropResizeVideo, Normalize255, TemporalRandomCrop
 
 
 def getdataset(args):
@@ -20,26 +15,17 @@ def getdataset(args):
     resize = [
         CenterCropResizeVideo((args.max_height, args.max_width)),
     ]
-    transform = transforms.Compose(
-        [
-            # Normalize255(),
-            *resize,
-            # RandomHorizontalFlipVideo(p=0.5),  # in case their caption have position decription
-            # norm_fun
-        ]
-    )
-    transform_topcrop = transforms.Compose(
-        [
-            Normalize255(),
-            *resize_topcrop,
-            # RandomHorizontalFlipVideo(p=0.5),  # in case their caption have position decription
-            norm_fun,
-        ]
-    )
+    transform = transforms.Compose([
+        # Normalize255(),
+        *resize,
+    ])
+    transform_topcrop = transforms.Compose([
+        Normalize255(),
+        *resize_topcrop,
+        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,
@@ -53,11 +39,13 @@ def getdataset(args):
 
 
 if __name__ == "__main__":
-    from accelerate import Accelerator
-    from fastvideo.dataset.t2v_datasets import dataset_prog
     import random
+
+    from accelerate import Accelerator
     from tqdm import tqdm
 
+    from fastvideo.dataset.t2v_datasets import dataset_prog
+
     args = type(
         "args",
         (),
@@ -92,9 +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

@@ -1,11 +1,13 @@
-import torch
-from torch.utils.data import Dataset
 import json
 import os
 import random
 
+import torch
+from torch.utils.data import Dataset
+
 
 class LatentDataset(Dataset):
+
     def __init__(
         self,
         json_path,
@@ -19,9 +21,7 @@ class LatentDataset(Dataset):
         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_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
@@ -31,10 +31,7 @@ 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"]
@@ -46,7 +43,7 @@ class LatentDataset(Dataset):
             map_location="cpu",
             weights_only=True,
         )
-        latent = latent.squeeze(0)[:, -self.num_latent_t :]
+        latent = latent.squeeze(0)[:, -self.num_latent_t:]
         if random.random() < self.cfg_rate:
             prompt_embed = self.uncond_prompt_embed
             prompt_attention_mask = self.uncond_prompt_mask
@@ -57,9 +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,
             )
@@ -92,16 +87,15 @@ def latent_collate_function(batch):
                 0,
                 max_w - latent.shape[3],
             ),
-        )
-        for latent in latents
+        ) for latent in latents
     ]
     # attn mask
     latent_attn_mask = torch.ones(len(latents), max_t, max_h, max_w)
     # set to 0 if padding
     for i, latent in enumerate(latents):
-        latent_attn_mask[i, latent.shape[1] :, :, :] = 0
-        latent_attn_mask[i, :, latent.shape[2] :, :] = 0
-        latent_attn_mask[i, :, :, latent.shape[3] :] = 0
+        latent_attn_mask[i, latent.shape[1]:, :, :] = 0
+        latent_attn_mask[i, :, latent.shape[2]:, :] = 0
+        latent_attn_mask[i, :, :, latent.shape[3]:] = 0
 
     prompt_embeds = torch.stack(prompt_embeds, dim=0)
     prompt_attention_masks = torch.stack(prompt_attention_masks, dim=0)
@@ -111,9 +105,7 @@ 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
-    )
+    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

@@ -1,21 +1,19 @@
 import json
-import os, io, csv, math, random
-import numpy as np
-from einops import rearrange
-from decord import VideoReader
-from os.path import join as opj
+import math
+import os
+import random
 from collections import Counter
+from os.path import join as opj
 
+import numpy as np
 import torch
-from torch.utils.data.dataset import Dataset
-from torch.utils.data import DataLoader, Dataset, get_worker_info
-from tqdm import tqdm
-from PIL import Image
-from accelerate.logging import get_logger
-from fastvideo.utils.dataset_utils import DecordInit
 import torchvision
+from einops import rearrange
+from PIL import Image
+from torch.utils.data import Dataset
 
-logger = get_logger(__name__)
+from fastvideo.utils.dataset_utils import DecordInit
+from fastvideo.utils.logging_ import main_print
 
 
 class SingletonMeta(type):
@@ -29,6 +27,7 @@ class SingletonMeta(type):
 
 
 class DataSetProg(metaclass=SingletonMeta):
+
     def __init__(self):
         self.cap_list = []
         self.elements = []
@@ -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
 
@@ -75,6 +72,7 @@ def filter_resolution(h, w, max_h_div_w_ratio=17 / 16, min_h_div_w_ratio=8 / 16)
 
 
 class T2V_dataset(Dataset):
+
     def __init__(self, args, transform, temporal_sample, tokenizer, transform_topcrop):
         self.data = args.data_merge_path
         self.num_frames = args.num_frames
@@ -94,7 +92,7 @@ class T2V_dataset(Dataset):
         self.v_decoder = DecordInit()
         self.video_length_tolerance_range = args.video_length_tolerance_range
         self.support_Chinese = True
-        if not ("mt5" in args.text_encoder_name):
+        if "mt5" not in args.text_encoder_name:
             self.support_Chinese = False
 
         cap_list = self.get_cap_list()
@@ -116,14 +114,9 @@ class T2V_dataset(Dataset):
         return dataset_prog.n_elements
 
     def __getitem__(self, idx):
-        try:
-            data = self.get_data(idx)
-            return data
-        except Exception as e:
-            logger.info(f"Error with {e}")
-            if idx in dataset_prog.cap_list:
-                logger.info(f"Caught an exception! {dataset_prog.cap_list[idx]}")
-            return self.__getitem__(random.randint(0, self.__len__() - 1))
+
+        data = self.get_data(idx)
+        return data
 
     def get_data(self, idx):
         path = dataset_prog.cap_list[idx]["path"]
@@ -136,9 +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")
@@ -187,24 +178,17 @@ 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)
-        )  #  [1 C H W] -> num_img [1 C H W]
+        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 = [], []
-        text = text if random.random() > self.cfg else ""
+        text = text[0] if random.random() > self.cfg else ""
         text_tokens_and_mask = self.tokenizer(
             text,
             max_length=self.text_max_length,
@@ -254,10 +238,7 @@ 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"]
@@ -277,26 +258,19 @@ class T2V_dataset(Dataset):
                 # import ipdb;ipdb.set_trace()
                 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)
+                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)
                     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
 
@@ -307,9 +281,7 @@ class T2V_dataset(Dataset):
                     # 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
@@ -318,15 +290,13 @@ class T2V_dataset(Dataset):
                 sample_num_frames.append(i["sample_num_frames"])
             else:
                 raise NameError(
-                    f"Unknown file extention {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()
-        logger.info(
+        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):
@@ -339,14 +309,11 @@ class T2V_dataset(Dataset):
     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:
                 sub_list = json.load(f)
-            logger.info(f"Building {anno}...")
             for i in range(len(sub_list)):
                 sub_list[i]["path"] = opj(folder, sub_list[i]["path"])
             cap_lists += sub_list

fastvideo/dataset/transform.py

@@ -1,7 +1,8 @@
-import torch
-import random
 import numbers
-from torchvision.transforms import RandomCrop, RandomResizedCrop
+import random
+
+import torch
+from PIL import Image
 
 
 def _is_tensor_video_clip(clip):
@@ -20,21 +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):
@@ -44,14 +39,12 @@ def crop(clip, i, j, h, w):
     """
     if len(clip.size()) != 4:
         raise ValueError("clip should be a 4D tensor")
-    return clip[..., i : i + h, j : j + w]
+    return clip[..., i:i + h, j:j + 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,
@@ -63,9 +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(
@@ -153,16 +144,14 @@ def random_shift_crop(clip):
     h, w = clip.size(-2), clip.size(-1)
 
     if h <= w:
-        long_edge = w
         short_edge = h
     else:
-        long_edge = h
         short_edge = w
 
     th, tw = short_edge, short_edge
 
-    i = torch.randint(0, h - th + 1, size=(1,)).item()
-    j = torch.randint(0, w - tw + 1, size=(1,)).item()
+    i = torch.randint(0, h - th + 1, size=(1, )).item()
+    j = torch.randint(0, w - tw + 1, size=(1, )).item()
     return crop(clip, i, j, th, tw)
 
 
@@ -177,9 +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
 
@@ -217,6 +204,7 @@ def hflip(clip):
 
 
 class RandomCropVideo:
+
     def __init__(self, size):
         if isinstance(size, numbers.Number):
             self.size = (int(size), int(size))
@@ -239,15 +227,13 @@ 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
 
-        i = torch.randint(0, h - th + 1, size=(1,)).item()
-        j = torch.randint(0, w - tw + 1, size=(1,)).item()
+        i = torch.randint(0, h - th + 1, size=(1, )).item()
+        j = torch.randint(0, w - tw + 1, size=(1, )).item()
 
         return i, j, th, tw
 
@@ -256,6 +242,7 @@ class RandomCropVideo:
 
 
 class SpatialStrideCropVideo:
+
     def __init__(self, stride):
         self.stride = stride
 
@@ -314,9 +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:
@@ -336,9 +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
@@ -352,9 +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,
@@ -380,9 +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)
@@ -397,9 +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
 
@@ -419,9 +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)
@@ -435,6 +410,7 @@ class KineticsRandomCropResizeVideo:
 
 
 class CenterCropVideo:
+
     def __init__(
         self,
         size,
@@ -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)
@@ -571,9 +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
@@ -581,27 +553,22 @@ class DynamicSampleDuration(object):
 
 
 if __name__ == "__main__":
-    from torchvision import transforms
-    import torchvision.io as io
-    import numpy as np
-    from torchvision.utils import save_image
     import os
 
-    vframes, aframes, info = io.read_video(
-        filename="./v_Archery_g01_c03.avi", pts_unit="sec", output_format="TCHW"
-    )
+    import numpy as np
+    import torchvision.io as io
+    from torchvision import transforms
+    from torchvision.utils import save_image
 
-    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
-            ),
-        ]
-    )
+    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),
+    ])
 
     target_video_len = 32
     frame_interval = 1
@@ -615,9 +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]
@@ -628,9 +593,7 @@ 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)
 

fastvideo/distill.py

@@ -1,86 +1,48 @@
+# !/bin/python3
+# isort: skip_file
 import argparse
 import math
 import os
-from fastvideo.utils.parallel_states import (
-    initialize_sequence_parallel_state,
-    destroy_sequence_parallel_group,
-    get_sequence_parallel_state,
-    nccl_info,
-)
-from fastvideo.utils.communications import sp_parallel_dataloader_wrapper, broadcast
-from fastvideo.models.mochi_hf.mochi_latents_utils import normalize_mochi_dit_input
-from fastvideo.utils.validation import log_validation
 import time
-from torch.utils.data import DataLoader
+from collections import deque
+from copy import deepcopy
+
 import torch
-from torch.distributed.fsdp import ShardingStrategy
-from torch.distributed.fsdp import (
-    FullyShardedDataParallel as FSDP,
-    StateDictType,
-    FullStateDictConfig,
-)
-from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
-import json
-from torch.utils.data.distributed import DistributedSampler
-from fastvideo.utils.dataset_utils import LengthGroupedSampler
+import torch.distributed as dist
 import wandb
 from accelerate.utils import set_seed
-from tqdm.auto import tqdm
-from fastvideo.fsdp_util import get_dit_fsdp_kwargs, apply_fsdp_checkpointing
-from diffusers import (
-    FlowMatchEulerDiscreteScheduler,
-)
-from fastvideo.distill.solver import EulerSolver, extract_into_tensor
-from copy import deepcopy
+from diffusers import FlowMatchEulerDiscreteScheduler
 from diffusers.optimization import get_scheduler
-from fastvideo.models.mochi_hf.modeling_mochi import MochiTransformer3DModel
 from diffusers.utils import check_min_version
-from fastvideo.dataset.latent_datasets import LatentDataset, latent_collate_function
-import torch.distributed as dist
-from safetensors.torch import save_file
 from peft import LoraConfig
-from torch.distributed.fsdp import (
-    FullyShardedDataParallel as FSDP,
-)
-from fastvideo.utils.checkpoint import (
-    save_checkpoint,
-    save_lora_checkpoint,
-    resume_lora_optimizer,
-)
+from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+from torch.distributed.fsdp import ShardingStrategy
+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.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.dataset_utils import LengthGroupedSampler
+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.validation import log_validation
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
 check_min_version("0.31.0")
-import time
-from collections import deque
+
 
 def main_print(content):
     if int(os.environ["LOCAL_RANK"]) <= 0:
         print(content)
 
 
-def save_checkpoint(transformer: MochiTransformer3DModel, rank, output_dir, step):
-    main_print(f"--> saving checkpoint at step {step}")
-    with FSDP.state_dict_type(
-        transformer,
-        StateDictType.FULL_STATE_DICT,
-        FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
-    ):
-        cpu_state = transformer.state_dict()
-    # todo move to get_state_dict
-    if rank <= 0:
-        save_dir = os.path.join(output_dir, f"checkpoint-{step}")
-        os.makedirs(save_dir, exist_ok=True)
-        # save using safetensors
-        weight_path = os.path.join(save_dir, "diffusion_pytorch_model.safetensors")
-        save_file(cpu_state, weight_path)
-        config_dict = dict(transformer.config)
-        config_path = os.path.join(save_dir, "config.json")
-        # save dict as json
-        with open(config_path, "w") as f:
-            json.dump(config_dict, f, indent=4)
-    main_print(f"--> checkpoint saved at step {step}")
-
-
 def reshard_fsdp(model):
     for m in FSDP.fsdp_modules(model):
         if m._has_params and m.sharding_strategy is not ShardingStrategy.NO_SHARD:
@@ -89,24 +51,23 @@ def reshard_fsdp(model):
 
 def get_norm(model_pred, norms, gradient_accumulation_steps):
     fro_norm = (
-        torch.linalg.matrix_norm(model_pred, ord="fro") / gradient_accumulation_steps
-    )
-    largest_singular_value = (
-        torch.linalg.matrix_norm(model_pred, ord=2) / gradient_accumulation_steps
-    )
+        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
     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()
+    norms["fro"] += torch.mean(fro_norm).item()  # codespell:ignore
     norms["largest singular value"] += torch.mean(largest_singular_value).item()
     norms["absolute mean"] += absolute_mean.item()
     norms["absolute max"] += absolute_max.item()
 
 
-def train_one_step_mochi(
+def distill_one_step(
     transformer,
+    model_type,
     teacher_transformer,
     ema_transformer,
     optimizer,
@@ -127,11 +88,12 @@ def train_one_step_mochi(
     ema_decay,
     pred_decay_weight,
     pred_decay_type,
+    hunyuan_teacher_disable_cfg,
 ):
     total_loss = 0.0
     optimizer.zero_grad()
     model_pred_norm = {
-        "fro": 0.0,
+        "fro": 0.0,  # codespell:ignore
         "largest singular value": 0.0,
         "absolute mean": 0.0,
         "absolute max": 0.0,
@@ -143,12 +105,10 @@ def train_one_step_mochi(
             latents_attention_mask,
             encoder_attention_mask,
         ) = next(loader)
-        model_input = normalize_mochi_dit_input(latents)
+        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.
@@ -159,25 +119,25 @@ def train_one_step_mochi(
         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):
-            model_pred = transformer(
-                noisy_model_input,
-                encoder_hidden_states,
-                timesteps,
-                encoder_attention_mask,  # B, L
-                return_dict=False,
-            )[0]
+            teacher_kwargs = {
+                "hidden_states": noisy_model_input,
+                "encoder_hidden_states": encoder_hidden_states,
+                "timestep": timesteps,
+                "encoder_attention_mask": encoder_attention_mask,  # B, L
+                "return_dict": False,
+            }
+            if hunyuan_teacher_disable_cfg:
+                teacher_kwargs["guidance"] = torch.tensor([1000.0],
+                                                          device=noisy_model_input.device,
+                                                          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):
@@ -200,9 +160,7 @@ def train_one_step_mochi(
                         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
-            )
+            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
@@ -225,65 +183,44 @@ def train_one_step_mochi(
                         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")
 
         # 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()
         dist.all_reduce(avg_loss, op=dist.ReduceOp.AVG)
-        dist.all_reduce(pred_decay_loss.detach(), op=dist.ReduceOp.AVG)
         total_loss += avg_loss.item()
 
     # 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()
     lr_scheduler.step()
 
-    return total_loss, grad_norm.item(), model_pred_norm, pred_decay_loss.item()
+    return total_loss, grad_norm.item(), model_pred_norm
 
 
 def main(args):
@@ -308,26 +245,20 @@ def main(args):
     if rank <= 0 and args.output_dir is not None:
         os.makedirs(args.output_dir, exist_ok=True)
 
-    # For mixed precision training we cast all non-trainable weigths to half-precision
+    # For mixed precision training we cast all non-trainable weights to half-precision
     # as these weights are only used for inference, keeping weights in full precision is not required.
 
     # Create model:
 
     main_print(f"--> loading model from {args.pretrained_model_name_or_path}")
-    # keep the master weight to float32
-    if args.dit_model_name_or_path:
-        transformer = transformer = MochiTransformer3DModel.from_pretrained(
-            args.dit_model_name_or_path,
-            torch_dtype=torch.float32,
-            # torch_dtype=torch.bfloat16 if args.use_lora else torch.float32,
-        )
-    else:
-        transformer = MochiTransformer3DModel.from_pretrained(
-            args.pretrained_model_name_or_path,
-            subfolder="transformer",
-            torch_dtype=torch.float32,
-            # torch_dtype=torch.bfloat16 if args.use_lora else torch.float32,
-        )
+
+    transformer = load_transformer(
+        args.model_type,
+        args.dit_model_name_or_path,
+        args.pretrained_model_name_or_path,
+        torch.float32 if args.master_weight_type == "fp32" else torch.bfloat16,
+    )
+
     teacher_transformer = deepcopy(transformer)
     if args.use_ema:
         ema_transformer = deepcopy(transformer)
@@ -335,6 +266,7 @@ def main(args):
         ema_transformer = None
 
     if args.use_lora:
+        assert args.model_type == "mochi", "LoRA is only supported for Mochi model."
         transformer.requires_grad_(False)
         transformer_lora_config = LoraConfig(
             r=args.lora_rank,
@@ -345,12 +277,10 @@ 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}"
-    )
-    fsdp_kwargs = get_dit_fsdp_kwargs(
+        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,
         args.use_lora,
         args.use_cpu_offload,
@@ -361,7 +291,7 @@ def main(args):
         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._no_split_modules = ["MochiTransformerBlock"]
+        transformer._no_split_modules = no_split_modules
         fsdp_kwargs["auto_wrap_policy"] = fsdp_kwargs["auto_wrap_policy"](transformer)
 
     transformer = FSDP(
@@ -377,13 +307,13 @@ def main(args):
             ema_transformer,
             **fsdp_kwargs,
         )
-    main_print(f"--> model loaded")
+    main_print("--> model loaded")
 
     if args.gradient_checkpointing:
-        apply_fsdp_checkpointing(transformer, args.selective_checkpointing)
-        apply_fsdp_checkpointing(teacher_transformer, 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, 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)
@@ -391,9 +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,
@@ -421,9 +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
@@ -440,20 +367,15 @@ def main(args):
     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(
-            args.train_batch_size,
-            rank=rank,
-            world_size=world_size,
-            lengths=train_dataset.lengths,
-            group_frame=args.group_frame,
-            group_resolution=args.group_resolution,
-        )
-        if (args.group_frame or args.group_resolution)
-        else DistributedSampler(
-            train_dataset, rank=rank, num_replicas=world_size, shuffle=False
-        )
-    )
+    sampler = (LengthGroupedSampler(
+        args.train_batch_size,
+        rank=rank,
+        world_size=world_size,
+        lengths=train_dataset.lengths,
+        group_frame=args.group_frame,
+        group_resolution=args.group_resolution,
+    ) if (args.group_frame or args.group_resolution) else DistributedSampler(
+        train_dataset, rank=rank, num_replicas=world_size, shuffle=False))
 
     train_dataloader = DataLoader(
         train_dataset,
@@ -466,11 +388,7 @@ 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
-    )
+        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:
@@ -478,22 +396,14 @@ def main(args):
         wandb.init(project=project, config=args)
 
     # Train!
-    total_batch_size = (
-        args.train_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"  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(
@@ -546,8 +456,9 @@ def main(args):
         assert args.multi_phased_distill_schedule is not None
         num_phases = get_num_phases(args.multi_phased_distill_schedule, step)
 
-        loss, grad_norm, pred_norm, aux_loss = train_one_step_mochi(
+        loss, grad_norm, pred_norm = distill_one_step(
             transformer,
+            args.model_type,
             teacher_transformer,
             ema_transformer,
             optimizer,
@@ -568,20 +479,19 @@ def main(args):
             args.ema_decay,
             args.pred_decay_weight,
             args.pred_decay_type,
+            args.hunyuan_teacher_disable_cfg,
         )
 
         step_time = time.time() - start_time
         step_times.append(step_time)
         avg_step_time = sum(step_times) / len(step_times)
 
-        progress_bar.set_postfix(
-            {
-                "loss": f"{loss:.4f}",
-                "step_time": f"{step_time:.2f}s",
-                "grad_norm": grad_norm,
-                "phases": num_phases,
-            }
-        )
+        progress_bar.set_postfix({
+            "loss": f"{loss:.4f}",
+            "step_time": f"{step_time:.2f}s",
+            "grad_norm": grad_norm,
+            "phases": num_phases,
+        })
         progress_bar.update(1)
         if rank <= 0:
             wandb.log(
@@ -591,20 +501,17 @@ def main(args):
                     "step_time": step_time,
                     "avg_step_time": avg_step_time,
                     "grad_norm": grad_norm,
-                    "pred_fro_norm": pred_norm["fro"],
+                    "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"],
-                    "aux_loss": aux_loss,
                 },
                 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:
@@ -642,9 +549,7 @@ 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)
 
@@ -655,8 +560,12 @@ 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.")
+
     # dataset & dataloader
     parser.add_argument("--data_json_path", type=str, required=True)
+    parser.add_argument("--num_height", type=int, default=480)
+    parser.add_argument("--num_width", type=int, default=848)
     parser.add_argument("--num_frames", type=int, default=163)
     parser.add_argument(
         "--dataloader_num_workers",
@@ -670,9 +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
 
@@ -694,9 +601,7 @@ 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,
@@ -713,39 +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
@@ -780,9 +677,7 @@ 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",
@@ -792,10 +687,8 @@ if __name__ == "__main__":
     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",
@@ -805,8 +698,7 @@ if __name__ == "__main__":
         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."
-        ),
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."),
     )
     parser.add_argument(
         "--use_cpu_offload",
@@ -828,12 +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
@@ -841,10 +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(
@@ -864,13 +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,
@@ -883,14 +765,12 @@ 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("--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("--finetune_weight", type=float, default=0.0)
     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")
     parser.add_argument(
         "--master_weight_type",
         type=str,

fastvideo/distill/discriminator.py

@@ -1,45 +1,23 @@
-from typing import Any, Dict, Optional, Union
-
-import torch
 import torch.nn as nn
-
-from diffusers.configuration_utils import ConfigMixin, register_to_config
-from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
-from diffusers.models.attention import JointTransformerBlock
-from diffusers.models.attention_processor import Attention, AttentionProcessor
-from diffusers.models.modeling_utils import ModelMixin
-from diffusers.models.normalization import AdaLayerNormContinuous
-from diffusers.utils import (
-    USE_PEFT_BACKEND,
-    is_torch_version,
-    logging,
-    scale_lora_layers,
-    unscale_lora_layers,
-)
-from diffusers.models.embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
-from diffusers.models.transformers.transformer_2d import Transformer2DModelOutput
-from diffusers.models.transformers.transformer_sd3 import SD3Transformer2DModel
+from diffusers.utils import logging
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
 class DiscriminatorHead(nn.Module):
+
     def __init__(self, input_channel, output_channel=1):
         super().__init__()
         inner_channel = 1024
         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)
@@ -57,41 +35,37 @@ class DiscriminatorHead(nn.Module):
 
 
 class Discriminator(nn.Module):
+
     def __init__(
         self,
         stride=8,
         num_h_per_head=1,
         adapter_channel_dims=[3072],
+        total_layers=48,
     ):
         super().__init__()
-        adapter_channel_dims = adapter_channel_dims * (48 // stride)
+        adapter_channel_dims = adapter_channel_dims * (total_layers // stride)
         self.stride = stride
         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
-            ]
-        )
+        self.heads = nn.ModuleList([
+            nn.ModuleList([DiscriminatorHead(adapter_channel) for _ in range(self.num_h_per_head)])
+            for adapter_channel in adapter_channel_dims
+        ])
 
     def forward(self, features):
         outputs = []
 
         def create_custom_forward(module):
+
             def custom_forward(*inputs):
                 return module(*inputs)
 
             return custom_forward
 
-        assert len(features) // self.stride == len(self.heads)
-        for i in range(0, len(features), self.stride):
-            for h in self.heads[i // self.stride]:
+        assert len(features) == len(self.heads)
+        for i in range(0, len(features)):
+            for h in self.heads[i]:
                 # out = torch.utils.checkpoint.checkpoint(
                 #     create_custom_forward(h),
                 #     features[i],

fastvideo/distill/solver.py

@@ -3,11 +3,10 @@ from typing import Optional, Tuple, Union
 
 import numpy as np
 import torch
-
 from diffusers.configuration_utils import ConfigMixin, register_to_config
-from diffusers.utils import BaseOutput, logging
-from diffusers.utils.torch_utils import randn_tensor
 from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils import BaseOutput, logging
+
 from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -21,7 +20,7 @@ class PCMFMSchedulerOutput(BaseOutput):
 def extract_into_tensor(a, t, x_shape):
     b, *_ = t.shape
     out = a.gather(-1, t)
-    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+    return out.reshape(b, *((1, ) * (len(x_shape) - 1)))
 
 
 class PCMFMScheduler(SchedulerMixin, ConfigMixin):
@@ -40,20 +39,15 @@ class PCMFMScheduler(SchedulerMixin, ConfigMixin):
     ):
         if linear_quadratic:
             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
+        self.euler_timesteps = (np.arange(1, pcm_timesteps + 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
@@ -118,9 +112,7 @@ 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).
 
@@ -131,18 +123,14 @@ 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
 
@@ -204,18 +192,11 @@ class PCMFMScheduler(SchedulerMixin, ConfigMixin):
                 returned, otherwise a tuple is returned where the first element is the sample tensor.
         """
 
-        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."
-                ),
-            )
+        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."), )
 
         if self.step_index is None:
             self._init_step_index(timestep)
@@ -233,7 +214,7 @@ class PCMFMScheduler(SchedulerMixin, ConfigMixin):
         self._step_index += 1
 
         if not return_dict:
-            return (prev_sample,)
+            return (prev_sample, )
 
         return PCMFMSchedulerOutput(prev_sample=prev_sample)
 
@@ -242,16 +223,14 @@ class PCMFMScheduler(SchedulerMixin, ConfigMixin):
 
 
 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 = (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()
@@ -268,9 +247,7 @@ class EulerSolver:
 
     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_prev = extract_into_tensor(self.sigmas_prev, timestep_index, model_pred.shape)
         x_prev = sample + (sigma_prev - sigma) * model_pred
         return x_prev
 
@@ -282,16 +259,10 @@ 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 = inference_indices.size(0) - 1 - last_valid_index
@@ -301,9 +272,7 @@ class EulerSolver:
             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_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

fastvideo/distill_adv.py

@@ -1,70 +1,43 @@
+# !/bin/python3
+# isort: skip_file
 import argparse
-from email.policy import strict
-import logging
 import math
 import os
-import shutil
-from pathlib import Path
-from fastvideo.utils.parallel_states import (
-    initialize_sequence_parallel_state,
-    destroy_sequence_parallel_group,
-    get_sequence_parallel_state,
-    nccl_info,
-)
-from fastvideo.utils.communications import sp_parallel_dataloader_wrapper, broadcast
-from fastvideo.models.mochi_hf.mochi_latents_utils import normalize_mochi_dit_input
-from fastvideo.utils.validation import log_validation
 import time
-from torch.utils.data import DataLoader
-import torch
-from torch.distributed.fsdp import (
-    FullyShardedDataParallel as FSDP,
-    StateDictType,
-    FullStateDictConfig,
-)
+from collections import deque
+from copy import deepcopy
 
-from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
-import json
-from torch.utils.data.distributed import DistributedSampler
-from fastvideo.utils.dataset_utils import LengthGroupedSampler
+import torch
+import torch.distributed as dist
 import wandb
 from accelerate.utils import set_seed
+from diffusers import FlowMatchEulerDiscreteScheduler
+from diffusers.optimization import get_scheduler
+from diffusers.utils import check_min_version
+from peft import LoraConfig
+from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
 from tqdm.auto import tqdm
-from fastvideo.fsdp_util import (
-    get_dit_fsdp_kwargs,
-    apply_fsdp_checkpointing,
-    get_discriminator_fsdp_kwargs,
-)
-import diffusers
-from diffusers import (
-    FlowMatchEulerDiscreteScheduler,
-)
+
+from fastvideo.dataset.latent_datasets import (LatentDataset, latent_collate_function)
 from fastvideo.distill.discriminator import Discriminator
 from fastvideo.distill.solver import EulerSolver, extract_into_tensor
-from copy import deepcopy
-from diffusers.optimization import get_scheduler
-from fastvideo.models.mochi_hf.modeling_mochi import MochiTransformer3DModel
-from diffusers.utils import check_min_version
-from fastvideo.dataset.latent_datasets import LatentDataset, latent_collate_function
-import torch.distributed as dist
-from peft import LoraConfig
-from torch.distributed.fsdp import (
-    FullyShardedDataParallel as FSDP,
-)
-from fastvideo.utils.checkpoint import (
-    save_checkpoint,
-    save_lora_checkpoint,
-    resume_lora_optimizer,
-    resume_training,
-    save_checkpoint_generator_discriminator,
-    resume_training_generator_discriminator,
-)
-from fastvideo.utils.logging import main_print
+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, resume_training_generator_discriminator, 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_discriminator_fsdp_kwargs, get_dit_fsdp_kwargs)
+from fastvideo.utils.load import load_transformer
+from fastvideo.utils.logging_ import main_print
+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.
 check_min_version("0.31.0")
-import time
-from collections import deque
 
 
 def gan_d_loss(
@@ -76,6 +49,7 @@ def gan_d_loss(
     encoder_hidden_states,
     encoder_attention_mask,
     weight,
+    discriminator_head_stride,
 ):
     loss = 0.0
     # collate sample_fake and sample_real
@@ -85,7 +59,8 @@ def gan_d_loss(
             encoder_hidden_states,
             timestep,
             encoder_attention_mask,
-            output_attn=True,
+            output_features=True,
+            output_features_stride=discriminator_head_stride,
             return_dict=False,
         )[1]
         real_features = teacher_transformer(
@@ -93,17 +68,16 @@ def gan_d_loss(
             encoder_hidden_states,
             timestep,
             encoder_attention_mask,
-            output_attn=True,
+            output_features=True,
+            output_features_stride=discriminator_head_stride,
             return_dict=False,
         )[1]
 
     fake_outputs = discriminator(fake_features)
     real_outputs = discriminator(real_features)
     for fake_output, real_output in zip(fake_outputs, real_outputs):
-        loss += (
-            torch.mean(weight * torch.relu(fake_output.float() + 1))
-            + torch.mean(weight * torch.relu(1 - real_output.float()))
-        ) / (discriminator.head_num * discriminator.num_h_per_head)
+        loss += (torch.mean(weight * torch.relu(fake_output.float() + 1)) + torch.mean(
+            weight * torch.relu(1 - real_output.float()))) / (discriminator.head_num * discriminator.num_h_per_head)
     return loss
 
 
@@ -115,6 +89,7 @@ def gan_g_loss(
     encoder_hidden_states,
     encoder_attention_mask,
     weight,
+    discriminator_head_stride,
 ):
     loss = 0.0
     features = teacher_transformer(
@@ -122,33 +97,30 @@ def gan_g_loss(
         encoder_hidden_states,
         timestep,
         encoder_attention_mask,
-        output_attn=True,
+        output_features=True,
+        output_features_stride=discriminator_head_stride,
         return_dict=False,
     )[1]
-    fake_outputs = discriminator(
-        features,
-    )
+    fake_outputs = discriminator(features, )
     for fake_output in fake_outputs:
-        loss += torch.mean(weight * torch.relu(1 - fake_output.float())) / (
-            discriminator.head_num * discriminator.num_h_per_head
-        )
+        loss += torch.mean(
+            weight * torch.relu(1 - fake_output.float())) / (discriminator.head_num * discriminator.num_h_per_head)
     return loss
 
 
-def train_one_step_mochi(
+def distill_one_step_adv(
     transformer,
+    model_type,
     teacher_transformer,
     optimizer,
     discriminator,
     discriminator_optimizer,
-    global_step,
     lr_scheduler,
     loader,
     noise_scheduler,
     solver,
     noise_random_generator,
     sp_size,
-    precondition_outputs,
     max_grad_norm,
     uncond_prompt_embed,
     uncond_prompt_mask,
@@ -157,6 +129,7 @@ def train_one_step_mochi(
     not_apply_cfg_solver,
     distill_cfg,
     adv_weight,
+    discriminator_head_stride,
 ):
     optimizer.zero_grad()
     discriminator_optimizer.zero_grad()
@@ -167,12 +140,10 @@ def train_one_step_mochi(
         latents_attention_mask,
         encoder_attention_mask,
     ) = next(loader)
-    model_input = normalize_mochi_dit_input(latents)
+    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.
@@ -197,11 +168,8 @@ def train_one_step_mochi(
         )[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)
 
-    weighting = 1.0
     # # simplified flow matching aka 0-rectified flow matching loss
     # # target = model_input - noise
     # target = model_input
@@ -210,14 +178,13 @@ def train_one_step_mochi(
         adv_index[i] = torch.randint(
             end_index[i].item(),
             end_index[i].item() + num_euler_timesteps // multiphase,
-            (1,),
+            (1, ),
             dtype=end_index.dtype,
             device=end_index.device,
         )
 
     sigmas_end = extract_into_tensor(solver.sigmas_prev, end_index, model_input.shape)
     sigmas_adv = extract_into_tensor(solver.sigmas_prev, adv_index, model_input.shape)
-    timesteps_end = (sigmas_end * noise_scheduler.config.num_train_timesteps).view(-1)
     timesteps_adv = (sigmas_adv * noise_scheduler.config.num_train_timesteps).view(-1)
 
     with torch.no_grad():
@@ -242,9 +209,7 @@ def train_one_step_mochi(
                     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
-        )
+        teacher_output = cond_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
@@ -258,22 +223,14 @@ def train_one_step_mochi(
                 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)
 
-    real_adv = (
-        (1 - sigmas_adv) * target + (sigmas_adv - sigmas_end) * torch.randn_like(target)
-    ) / (1 - sigmas_end)
-    fake_adv = (
-        (1 - sigmas_adv) * model_pred
-        + (sigmas_adv - sigmas_end) * torch.randn_like(model_pred)
-    ) / (1 - sigmas_end)
+    real_adv = ((1 - sigmas_adv) * target + (sigmas_adv - sigmas_end) * torch.randn_like(target)) / (1 - sigmas_end)
+    fake_adv = ((1 - sigmas_adv) * model_pred +
+                (sigmas_adv - sigmas_end) * torch.randn_like(model_pred)) / (1 - sigmas_end)
 
     huber_c = 0.001
-    g_loss = torch.mean(
-        torch.sqrt((model_pred.float() - target.float()) ** 2 + huber_c**2) - huber_c
-    )
+    g_loss = torch.mean(torch.sqrt((model_pred.float() - target.float())**2 + huber_c**2) - huber_c)
     discriminator.requires_grad_(False)
     with torch.autocast("cuda", dtype=torch.bfloat16):
         g_gan_loss = adv_weight * gan_g_loss(
@@ -284,6 +241,7 @@ def train_one_step_mochi(
             encoder_hidden_states.float(),
             encoder_attention_mask,
             1.0,
+            discriminator_head_stride,
         )
     g_loss += g_gan_loss
     g_loss.backward()
@@ -308,6 +266,7 @@ def train_one_step_mochi(
             encoder_hidden_states,
             encoder_attention_mask,
             1.0,
+            discriminator_head_stride,
         )
 
     d_loss.backward()
@@ -340,28 +299,24 @@ def main(args):
     if rank <= 0 and args.output_dir is not None:
         os.makedirs(args.output_dir, exist_ok=True)
 
-    # For mixed precision training we cast all non-trainable weigths to half-precision
+    # For mixed precision training we cast all non-trainable weights to half-precision
     # as these weights are only used for inference, keeping weights in full precision is not required.
 
     # Create model:
 
     main_print(f"--> loading model from {args.pretrained_model_name_or_path}")
     # keep the master weight to float32
-    if args.dit_model_name_or_path:
-        transformer = transformer = MochiTransformer3DModel.from_pretrained(
-            args.dit_model_name_or_path,
-            torch_dtype=torch.float32,
-            # torch_dtype=torch.bfloat16 if args.use_lora else torch.float32,
-        )
-    else:
-        transformer = MochiTransformer3DModel.from_pretrained(
-            args.pretrained_model_name_or_path,
-            subfolder="transformer",
-            torch_dtype=torch.float32,
-            # torch_dtype=torch.bfloat16 if args.use_lora else torch.float32,
-        )
+    transformer = load_transformer(
+        args.model_type,
+        args.dit_model_name_or_path,
+        args.pretrained_model_name_or_path,
+        torch.float32 if args.master_weight_type == "fp32" else torch.bfloat16,
+    )
     teacher_transformer = deepcopy(transformer)
-    discriminator = Discriminator(args.discriminator_head_stride)
+    discriminator = Discriminator(
+        args.discriminator_head_stride,
+        total_layers=48 if args.model_type == "mochi" else 40,
+    )
 
     if args.use_lora:
         transformer.requires_grad_(False)
@@ -380,18 +335,21 @@ def main(args):
     main_print(
         f"  Total discriminator parameters = {sum(p.numel() for p in discriminator.parameters() if p.requires_grad) / 1e6} M"
     )
-    main_print(
-        f"--> Initializing FSDP with sharding strategy: {args.fsdp_sharding_startegy}"
-    )
-    fsdp_kwargs = get_dit_fsdp_kwargs(
-        args.fsdp_sharding_startegy, args.use_lora, args.use_cpu_offload
+    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,
+        args.use_lora,
+        args.use_cpu_offload,
+        args.master_weight_type,
     )
     discriminator_fsdp_kwargs = get_discriminator_fsdp_kwargs(args.master_weight_type)
     if args.use_lora:
+        assert args.model_type == "mochi", "LoRA is only supported for Mochi model."
         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._no_split_modules = ["MochiTransformerBlock"]
+        transformer._no_split_modules = no_split_modules
         fsdp_kwargs["auto_wrap_policy"] = fsdp_kwargs["auto_wrap_policy"](transformer)
 
     transformer = FSDP(
@@ -406,19 +364,17 @@ def main(args):
         discriminator,
         **discriminator_fsdp_kwargs,
     )
-    main_print(f"--> model loaded")
+    main_print("--> model loaded")
 
     if args.gradient_checkpointing:
-        apply_fsdp_checkpointing(transformer, args.selective_checkpointing)
-        apply_fsdp_checkpointing(teacher_transformer, args.selective_checkpointing)
+        apply_fsdp_checkpointing(transformer, no_split_modules, args.selective_checkpointing)
+        apply_fsdp_checkpointing(teacher_transformer, no_split_modules, args.selective_checkpointing)
     # Set model as trainable.
     transformer.train()
     teacher_transformer.requires_grad_(False)
     noise_scheduler = FlowMatchEulerDiscreteScheduler(shift=args.shift)
     if args.scheduler_type == "pcm_linear_quadratic":
-        sigmas = linear_quadratic_schedule(
-            noise_scheduler.config.num_train_timesteps, args.linear_quadratic_threshold
-        )
+        sigmas = linear_quadratic_schedule(noise_scheduler.config.num_train_timesteps, args.linear_quadratic_threshold)
         sigmas = torch.tensor(sigmas).to(dtype=torch.float32)
     else:
         sigmas = noise_scheduler.sigmas
@@ -435,7 +391,7 @@ def main(args):
         params_to_optimize,
         lr=args.learning_rate,
         betas=(0.9, 0.999),
-        weight_decay=1e-3,
+        weight_decay=args.weight_decay,
         eps=1e-8,
     )
 
@@ -443,15 +399,14 @@ def main(args):
         discriminator.parameters(),
         lr=args.discriminator_learning_rate,
         betas=(0, 0.999),
-        weight_decay=1e-3,
+        weight_decay=args.weight_decay,
         eps=1e-8,
     )
 
     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)
     elif args.resume_from_checkpoint:
         (
             transformer,
@@ -483,20 +438,15 @@ def main(args):
     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(
-            args.train_batch_size,
-            rank=rank,
-            world_size=world_size,
-            lengths=train_dataset.lengths,
-            group_frame=args.group_frame,
-            group_resolution=args.group_resolution,
-        )
-        if (args.group_frame or args.group_resolution)
-        else DistributedSampler(
-            train_dataset, rank=rank, num_replicas=world_size, shuffle=False
-        )
-    )
+    sampler = (LengthGroupedSampler(
+        args.train_batch_size,
+        rank=rank,
+        world_size=world_size,
+        lengths=train_dataset.lengths,
+        group_frame=args.group_frame,
+        group_resolution=args.group_resolution,
+    ) if (args.group_frame or args.group_resolution) else DistributedSampler(
+        train_dataset, rank=rank, num_replicas=world_size, shuffle=False))
 
     train_dataloader = DataLoader(
         train_dataset,
@@ -509,11 +459,7 @@ def main(args):
     )
     assert args.gradient_accumulation_steps == 1
     num_update_steps_per_epoch = math.ceil(
-        len(train_dataloader)
-        / args.gradient_accumulation_steps
-        * args.sp_size
-        / args.train_sp_batch_size
-    )
+        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:
@@ -521,22 +467,14 @@ def main(args):
         wandb.init(project=project, config=args)
 
     # Train!
-    total_batch_size = (
-        args.train_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"  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(
@@ -561,55 +499,65 @@ def main(args):
     )
 
     step_times = deque(maxlen=100)
+
     # log_validation(args, transformer, device,
-    #         torch.bfloat16, init_steps, scheduler_type=args.scheduler_type, shift=args.shift, num_euler_timesteps=args.num_euler_timesteps, linear_quadratic_threshold=args.linear_quadratic_threshold, ema=False)
+    #             torch.bfloat16, 0, scheduler_type=args.scheduler_type, shift=args.shift, num_euler_timesteps=args.num_euler_timesteps, linear_quadratic_threshold=args.linear_quadratic_threshold,ema=False)
+    def get_num_phases(multi_phased_distill_schedule, step):
+        # step-phase,step-phase
+        multi_phases = multi_phased_distill_schedule.split(",")
+        phase = multi_phases[-1].split("-")[-1]
+        for step_phases in multi_phases:
+            phase_step, phase = step_phases.split("-")
+            if step <= int(phase_step):
+                return int(phase)
+        return phase
 
     for i in range(init_steps):
         _ = next(loader)
     for step in range(init_steps + 1, args.max_train_steps + 1):
+        assert args.multi_phased_distill_schedule is not None
+        num_phases = get_num_phases(args.multi_phased_distill_schedule, step)
         start_time = time.time()
         (
             generator_loss,
             generator_grad_norm,
             discriminator_loss,
             discriminator_grad_norm,
-        ) = train_one_step_mochi(
+        ) = distill_one_step_adv(
             transformer,
+            args.model_type,
             teacher_transformer,
             optimizer,
             discriminator,
             discriminator_optimizer,
-            step,
             lr_scheduler,
             loader,
             noise_scheduler,
             solver,
             noise_random_generator,
             args.sp_size,
-            args.precondition_outputs,
             args.max_grad_norm,
             uncond_prompt_embed,
             uncond_prompt_mask,
             args.num_euler_timesteps,
-            args.validation_sampling_steps,
+            num_phases,
             args.not_apply_cfg_solver,
             args.distill_cfg,
             args.adv_weight,
+            args.discriminator_head_stride,
         )
 
         step_time = time.time() - start_time
         step_times.append(step_time)
         avg_step_time = sum(step_times) / len(step_times)
 
-        progress_bar.set_postfix(
-            {
-                "g_loss": f"{generator_loss:.4f}",
-                "d_loss": f"{discriminator_loss:.4f}",
-                "g_grad_norm": generator_grad_norm,
-                "d_grad_norm": discriminator_grad_norm,
-                "step_time": f"{step_time:.2f}s",
-            }
-        )
+        progress_bar.set_postfix({
+            "g_loss": f"{generator_loss:.4f}",
+            "d_loss": f"{discriminator_loss:.4f}",
+            "g_grad_norm": generator_grad_norm,
+            "d_grad_norm": discriminator_grad_norm,
+            "step_time": f"{step_time:.2f}s",
+        })
         progress_bar.update(1)
         if rank <= 0:
             wandb.log(
@@ -628,20 +576,20 @@ def main(args):
             main_print(f"--> saving checkpoint at step {step}")
             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
-                save_checkpoint_generator_discriminator(
-                    transformer,
-                    optimizer,
-                    discriminator,
-                    discriminator_optimizer,
-                    rank,
-                    args.output_dir,
-                    step,
-                )
+                # TODO
+                # save_checkpoint_generator_discriminator(
+                #     transformer,
+                #     optimizer,
+                #     discriminator,
+                #     discriminator_optimizer,
+                #     rank,
+                #     args.output_dir,
+                #     step,
+                # )
+                save_checkpoint(transformer, rank, args.output_dir, step)
             main_print(f"--> checkpoint saved at step {step}")
             dist.barrier()
         if args.log_validation and step % args.validation_steps == 0:
@@ -655,25 +603,14 @@ def main(args):
                 shift=args.shift,
                 num_euler_timesteps=args.num_euler_timesteps,
                 linear_quadratic_threshold=args.linear_quadratic_threshold,
+                linear_range=args.linear_range,
                 ema=False,
             )
 
     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, optimizer, rank, args.output_dir, args.max_train_steps
-        )
-        save_checkpoint(
-            discriminator,
-            discriminator_optimizer,
-            rank,
-            args.output_dir,
-            step,
-            discriminator=True,
-        )
+        save_checkpoint(transformer, rank, args.output_dir, args.max_train_steps)
 
     if get_sequence_parallel_state():
         destroy_sequence_parallel_group()
@@ -682,8 +619,11 @@ 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.")
     # dataset & dataloader
     parser.add_argument("--data_json_path", type=str, required=True)
+    parser.add_argument("--num_height", type=int, default=480)
+    parser.add_argument("--num_width", type=int, default=848)
     parser.add_argument("--num_frames", type=int, default=163)
     parser.add_argument(
         "--dataloader_num_workers",
@@ -697,9 +637,7 @@ 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
 
@@ -712,22 +650,13 @@ if __name__ == "__main__":
     parser.add_argument("--ema_decay", type=float, default=0.999)
     parser.add_argument("--ema_start_step", type=int, default=0)
     parser.add_argument("--cfg", type=float, default=0.1)
-    parser.add_argument(
-        "--precondition_outputs",
-        action="store_true",
-        help="Whether to precondition the outputs of the model.",
-    )
-
     # validation & logs
-    parser.add_argument("--validation_prompt_dir", type=str)
-    parser.add_argument("--validation_sampling_steps", type=int, default=64)
-    parser.add_argument("--validation_guidance_scale", type=float, default=4.5)
+    parser.add_argument("--validation_sampling_steps", type=str, default="64")
+    parser.add_argument("--validation_guidance_scale", type=str, default="4.5")
     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,
@@ -744,39 +673,32 @@ 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("--validation_prompt_dir", type=str)
     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
@@ -811,9 +733,7 @@ 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",
@@ -823,10 +743,8 @@ if __name__ == "__main__":
     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",
@@ -836,8 +754,7 @@ if __name__ == "__main__":
         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."
-        ),
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."),
     )
     parser.add_argument(
         "--use_cpu_offload",
@@ -859,13 +776,10 @@ 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")
+    parser.add_argument("--multi_phased_distill_schedule", type=str, default=None)
     parser.add_argument(
         "--gradient_accumulation_steps",
         type=int,
@@ -878,10 +792,8 @@ 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(
@@ -901,13 +813,9 @@ 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(
         "--adv_weight",
         type=float,
@@ -920,6 +828,13 @@ if __name__ == "__main__":
         default=2,
         help="The stride of the discriminator head.",
     )
+    parser.add_argument(
+        "--linear_range",
+        type=float,
+        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(
         "--linear_quadratic_threshold",
         type=float,

fastvideo/models/flash_attn_no_pad.py

@@ -0,0 +1,28 @@
+from einops import rearrange
+from flash_attn import flash_attn_varlen_qkvpacked_func
+from flash_attn.bert_padding import pad_input, unpad_input
+
+
+def flash_attn_no_pad(qkv, key_padding_mask, causal=False, dropout_p=0.0, softmax_scale=None):
+    # adapted from https://github.com/Dao-AILab/flash-attention/blob/13403e81157ba37ca525890f2f0f2137edf75311/flash_attn/flash_attention.py#L27
+    batch_size = qkv.shape[0]
+    seqlen = qkv.shape[1]
+    nheads = qkv.shape[-2]
+    x = rearrange(qkv, "b s three h d -> b s (three h d)")
+    x_unpad, indices, cu_seqlens, max_s, used_seqlens_in_batch = unpad_input(x, key_padding_mask)
+
+    x_unpad = rearrange(x_unpad, "nnz (three h d) -> nnz three h d", three=3, h=nheads)
+    output_unpad = flash_attn_varlen_qkvpacked_func(
+        x_unpad,
+        cu_seqlens,
+        max_s,
+        dropout_p,
+        softmax_scale=softmax_scale,
+        causal=causal,
+    )
+    output = rearrange(
+        pad_input(rearrange(output_unpad, "nnz h d -> nnz (h d)"), indices, batch_size, seqlen),
+        "b s (h d) -> b s h d",
+        h=nheads,
+    )
+    return output

fastvideo/models/hunyuan/constants.py

@@ -0,0 +1,89 @@
+import os
+
+import torch
+
+__all__ = [
+    "C_SCALE",
+    "PROMPT_TEMPLATE",
+    "MODEL_BASE",
+    "PRECISIONS",
+    "NORMALIZATION_TYPE",
+    "ACTIVATION_TYPE",
+    "VAE_PATH",
+    "TEXT_ENCODER_PATH",
+    "TOKENIZER_PATH",
+    "TEXT_PROJECTION",
+    "DATA_TYPE",
+    "NEGATIVE_PROMPT",
+]
+
+PRECISION_TO_TYPE = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+# =================== Constant Values =====================
+# Computation scale factor, 1P = 1_000_000_000_000_000. Tensorboard will display the value in PetaFLOPS to avoid
+# overflow error when tensorboard logging values.
+C_SCALE = 1_000_000_000_000_000
+
+# When using decoder-only models, we must provide a prompt template to instruct the text encoder
+# on how to generate the text.
+# --------------------------------------------------------------------
+PROMPT_TEMPLATE_ENCODE = (
+    "<|start_header_id|>system<|end_header_id|>\n\nDescribe the image by detailing the color, shape, size, texture, "
+    "quantity, text, spatial relationships of the objects and background:<|eot_id|>"
+    "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>")
+PROMPT_TEMPLATE_ENCODE_VIDEO = (
+    "<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: "
+    "1. The main content and theme of the video."
+    "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects."
+    "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects."
+    "4. background environment, light, style and atmosphere."
+    "5. camera angles, movements, and transitions used in the video:<|eot_id|>"
+    "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>")
+
+NEGATIVE_PROMPT = "Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion"
+
+PROMPT_TEMPLATE = {
+    "dit-llm-encode": {
+        "template": PROMPT_TEMPLATE_ENCODE,
+        "crop_start": 36,
+    },
+    "dit-llm-encode-video": {
+        "template": PROMPT_TEMPLATE_ENCODE_VIDEO,
+        "crop_start": 95,
+    },
+}
+
+# ======================= Model ======================
+PRECISIONS = {"fp32", "fp16", "bf16"}
+NORMALIZATION_TYPE = {"layer", "rms"}
+ACTIVATION_TYPE = {"relu", "silu", "gelu", "gelu_tanh"}
+
+# =================== Model Path =====================
+MODEL_BASE = os.getenv("MODEL_BASE", "./data/hunyuan")
+
+# =================== Data =======================
+DATA_TYPE = {"image", "video", "image_video"}
+
+# 3D VAE
+VAE_PATH = {"884-16c-hy": f"{MODEL_BASE}/hunyuan-video-t2v-720p/vae"}
+
+# Text Encoder
+TEXT_ENCODER_PATH = {
+    "clipL": f"{MODEL_BASE}/text_encoder_2",
+    "llm": f"{MODEL_BASE}/text_encoder",
+}
+
+# Tokenizer
+TOKENIZER_PATH = {
+    "clipL": f"{MODEL_BASE}/text_encoder_2",
+    "llm": f"{MODEL_BASE}/text_encoder",
+}
+
+TEXT_PROJECTION = {
+    "linear",  # Default, an nn.Linear() layer
+    "single_refiner",  # Single TokenRefiner. Refer to LI-DiT
+}

fastvideo/models/hunyuan/diffusion/__init__.py

@@ -0,0 +1,3 @@
+# ruff: noqa: F401
+from .pipelines import HunyuanVideoPipeline
+from .schedulers import FlowMatchDiscreteScheduler

fastvideo/models/hunyuan/diffusion/pipelines/__init__.py

@@ -0,0 +1,2 @@
+# ruff: noqa: F401
+from .pipeline_hunyuan_video import HunyuanVideoPipeline

fastvideo/models/hunyuan/diffusion/pipelines/pipeline_hunyuan_video.py

@@ -0,0 +1,931 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+#
+# Modified from diffusers==0.29.2
+#
+# ==============================================================================
+import inspect
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.configuration_utils import FrozenDict
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from diffusers.models import AutoencoderKL
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (USE_PEFT_BACKEND, BaseOutput, deprecate, logging, replace_example_docstring,
+                             scale_lora_layers)
+from diffusers.utils.torch_utils import randn_tensor
+from einops import rearrange
+
+from fastvideo.utils.communications import all_gather
+from fastvideo.utils.parallel_states import get_sequence_parallel_state, nccl_info
+
+from ...constants import PRECISION_TO_TYPE
+from ...modules import HYVideoDiffusionTransformer
+from ...text_encoder import TextEncoder
+from ...vae.autoencoder_kl_causal_3d import AutoencoderKLCausal3D
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """"""
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = (guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg)
+    return noise_cfg
+
+
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler.")
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler.")
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+@dataclass
+class HunyuanVideoPipelineOutput(BaseOutput):
+    videos: Union[torch.Tensor, np.ndarray]
+
+
+class HunyuanVideoPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-video generation using HunyuanVideo.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        text_encoder ([`TextEncoder`]):
+            Frozen text-encoder.
+        text_encoder_2 ([`TextEncoder`]):
+            Frozen text-encoder_2.
+        transformer ([`HYVideoDiffusionTransformer`]):
+            A `HYVideoDiffusionTransformer` to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
+    _optional_components = ["text_encoder_2"]
+    _exclude_from_cpu_offload = ["transformer"]
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: TextEncoder,
+        transformer: HYVideoDiffusionTransformer,
+        scheduler: KarrasDiffusionSchedulers,
+        text_encoder_2: Optional[TextEncoder] = None,
+        progress_bar_config: Dict[str, Any] = None,
+        args=None,
+    ):
+        super().__init__()
+
+        # ==========================================================================================
+        if progress_bar_config is None:
+            progress_bar_config = {}
+        if not hasattr(self, "_progress_bar_config"):
+            self._progress_bar_config = {}
+        self._progress_bar_config.update(progress_bar_config)
+
+        self.args = args
+        # ==========================================================================================
+
+        if (hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1):
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file")
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if (hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True):
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+                " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+                " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+                " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file")
+            deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["clip_sample"] = False
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            transformer=transformer,
+            scheduler=scheduler,
+            text_encoder_2=text_encoder_2,
+        )
+        self.vae_scale_factor = 2**(len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    def encode_prompt(
+        self,
+        prompt,
+        device,
+        num_videos_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_attention_mask: Optional[torch.Tensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+        text_encoder: Optional[TextEncoder] = None,
+        data_type: Optional[str] = "image",
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_videos_per_prompt (`int`):
+                number of videos that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the video generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            attention_mask (`torch.Tensor`, *optional*):
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            negative_attention_mask (`torch.Tensor`, *optional*):
+            lora_scale (`float`, *optional*):
+                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            text_encoder (TextEncoder, *optional*):
+            data_type (`str`, *optional*):
+        """
+        if text_encoder is None:
+            text_encoder = self.text_encoder
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if not USE_PEFT_BACKEND:
+                adjust_lora_scale_text_encoder(text_encoder.model, lora_scale)
+            else:
+                scale_lora_layers(text_encoder.model, lora_scale)
+
+        if prompt_embeds is None:
+            # textual inversion: process multi-vector tokens if necessary
+            if isinstance(self, TextualInversionLoaderMixin):
+                prompt = self.maybe_convert_prompt(prompt, text_encoder.tokenizer)
+
+            text_inputs = text_encoder.text2tokens(prompt, data_type=data_type)
+            if clip_skip is None:
+                prompt_outputs = text_encoder.encode(text_inputs, data_type=data_type, device=device)
+                prompt_embeds = prompt_outputs.hidden_state
+            else:
+                prompt_outputs = text_encoder.encode(
+                    text_inputs,
+                    output_hidden_states=True,
+                    data_type=data_type,
+                    device=device,
+                )
+                # Access the `hidden_states` first, that contains a tuple of
+                # all the hidden states from the encoder layers. Then index into
+                # the tuple to access the hidden states from the desired layer.
+                prompt_embeds = prompt_outputs.hidden_states_list[-(clip_skip + 1)]
+                # We also need to apply the final LayerNorm here to not mess with the
+                # representations. The `last_hidden_states` that we typically use for
+                # obtaining the final prompt representations passes through the LayerNorm
+                # layer.
+                prompt_embeds = text_encoder.model.text_model.final_layer_norm(prompt_embeds)
+
+            attention_mask = prompt_outputs.attention_mask
+            if attention_mask is not None:
+                attention_mask = attention_mask.to(device)
+                bs_embed, seq_len = attention_mask.shape
+                attention_mask = attention_mask.repeat(1, num_videos_per_prompt)
+                attention_mask = attention_mask.view(bs_embed * num_videos_per_prompt, seq_len)
+
+        if text_encoder is not None:
+            prompt_embeds_dtype = text_encoder.dtype
+        elif self.transformer is not None:
+            prompt_embeds_dtype = self.transformer.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+        if prompt_embeds.ndim == 2:
+            bs_embed, _ = prompt_embeds.shape
+            # duplicate text embeddings for each generation per prompt, using mps friendly method
+            prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt)
+            prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, -1)
+        else:
+            bs_embed, seq_len, _ = prompt_embeds.shape
+            # duplicate text embeddings for each generation per prompt, using mps friendly method
+            prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+            prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+        return (
+            prompt_embeds,
+            negative_prompt_embeds,
+            attention_mask,
+            negative_attention_mask,
+        )
+
+    def decode_latents(self, latents, enable_tiling=True):
+        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+        if enable_tiling:
+            self.vae.enable_tiling()
+        image = self.vae.decode(latents, return_dict=False)[0]
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        if image.ndim == 4:
+            image = image.cpu().permute(0, 2, 3, 1).float()
+        else:
+            image = image.cpu().float()
+        return image
+
+    def prepare_extra_func_kwargs(self, func, kwargs):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        extra_step_kwargs = {}
+
+        for k, v in kwargs.items():
+            accepts = k in set(inspect.signature(func).parameters.keys())
+            if accepts:
+                extra_step_kwargs[k] = v
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        video_length,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        vae_ver="88-4c-sd",
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if video_length is not None:
+            if "884" in vae_ver:
+                if video_length != 1 and (video_length - 1) % 4 != 0:
+                    raise ValueError(f"`video_length` has to be 1 or a multiple of 4 but is {video_length}.")
+            elif "888" in vae_ver:
+                if video_length != 1 and (video_length - 1) % 8 != 0:
+                    raise ValueError(f"`video_length` has to be 1 or a multiple of 8 but is {video_length}.")
+
+        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+            raise ValueError(f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                             f" {type(callback_steps)}.")
+        if callback_on_step_end_tensor_inputs is not None and not all(k in self._callback_tensor_inputs
+                                                                      for k in callback_on_step_end_tensor_inputs):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two.")
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined.")
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                             f" {negative_prompt_embeds}. Please make sure to only forward one of the two.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}.")
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        video_length,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            video_length,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators.")
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # Check existence to make it compatible with FlowMatchEulerDiscreteScheduler
+        if hasattr(self.scheduler, "init_noise_sigma"):
+            # scale the initial noise by the standard deviation required by the scheduler
+            latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(
+        self,
+        w: torch.Tensor,
+        embedding_dim: int = 512,
+        dtype: torch.dtype = torch.float32,
+    ) -> torch.Tensor:
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            w (`torch.Tensor`):
+                Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
+            embedding_dim (`int`, *optional*, defaults to 512):
+                Dimension of the embeddings to generate.
+            dtype (`torch.dtype`, *optional*, defaults to `torch.float32`):
+                Data type of the generated embeddings.
+
+        Returns:
+            `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        # return self._guidance_scale > 1 and self.transformer.config.time_cond_proj_dim is None
+        return self._guidance_scale > 1
+
+    @property
+    def cross_attention_kwargs(self):
+        return self._cross_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: int,
+        width: int,
+        video_length: int,
+        data_type: str = "video",
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        sigmas: List[float] = None,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_videos_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Union[Callable[[int, int, Dict], None], PipelineCallback,
+                                             MultiPipelineCallbacks, ]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        vae_ver: str = "88-4c-sd",
+        enable_tiling: bool = False,
+        enable_vae_sp: bool = False,
+        n_tokens: Optional[int] = None,
+        embedded_guidance_scale: Optional[float] = None,
+        mask_strategy: Optional[Dict[str, list]] = None,
+        **kwargs,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            height (`int`):
+                The height in pixels of the generated image.
+            width (`int`):
+                The width in pixels of the generated image.
+            video_length (`int`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`HunyuanVideoPipelineOutput`] instead of a
+                plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
+                using zero terminal SNR.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~HunyuanVideoPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`HunyuanVideoPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list with the generated images and the
+                second element is a list of `bool`s indicating whether the corresponding generated image contains
+                "not-safe-for-work" (nsfw) content.
+        """
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 0. Default height and width to unet
+        # height = height or self.transformer.config.sample_size * self.vae_scale_factor
+        # width = width or self.transformer.config.sample_size * self.vae_scale_factor
+        # to deal with lora scaling and other possible forward hooks
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            video_length,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+            vae_ver=vae_ver,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = (torch.device(f"cuda:{dist.get_rank()}") if dist.is_initialized() else self._execution_device)
+
+        # 3. Encode input prompt
+        lora_scale = (self.cross_attention_kwargs.get("scale", None)
+                      if self.cross_attention_kwargs is not None else None)
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            prompt_mask,
+            negative_prompt_mask,
+        ) = self.encode_prompt(
+            prompt,
+            device,
+            num_videos_per_prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            attention_mask=attention_mask,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_attention_mask=negative_attention_mask,
+            lora_scale=lora_scale,
+            clip_skip=self.clip_skip,
+            data_type=data_type,
+        )
+        if self.text_encoder_2 is not None:
+            (
+                prompt_embeds_2,
+                negative_prompt_embeds_2,
+                prompt_mask_2,
+                negative_prompt_mask_2,
+            ) = self.encode_prompt(
+                prompt,
+                device,
+                num_videos_per_prompt,
+                self.do_classifier_free_guidance,
+                negative_prompt,
+                prompt_embeds=None,
+                attention_mask=None,
+                negative_prompt_embeds=None,
+                negative_attention_mask=None,
+                lora_scale=lora_scale,
+                clip_skip=self.clip_skip,
+                text_encoder=self.text_encoder_2,
+                data_type=data_type,
+            )
+        else:
+            prompt_embeds_2 = None
+            negative_prompt_embeds_2 = None
+            prompt_mask_2 = None
+            negative_prompt_mask_2 = None
+
+        # For classifier free guidance, we need to do two forward passes.
+        # Here we concatenate the unconditional and text embeddings into a single batch
+        # to avoid doing two forward passes
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+            if prompt_mask is not None:
+                prompt_mask = torch.cat([negative_prompt_mask, prompt_mask])
+            if prompt_embeds_2 is not None:
+                prompt_embeds_2 = torch.cat([negative_prompt_embeds_2, prompt_embeds_2])
+            if prompt_mask_2 is not None:
+                prompt_mask_2 = torch.cat([negative_prompt_mask_2, prompt_mask_2])
+
+        # 4. Prepare timesteps
+        extra_set_timesteps_kwargs = self.prepare_extra_func_kwargs(self.scheduler.set_timesteps,
+                                                                    {"n_tokens": n_tokens})
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas,
+            **extra_set_timesteps_kwargs,
+        )
+        if "884" in vae_ver:
+            video_length = (video_length - 1) // 4 + 1
+        elif "888" in vae_ver:
+            video_length = (video_length - 1) // 8 + 1
+        else:
+            video_length = video_length
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            video_length,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+        world_size, rank = nccl_info.sp_size, nccl_info.rank_within_group
+        if get_sequence_parallel_state():
+            latents = rearrange(latents, "b t (n s) h w -> b t n s h w", n=world_size).contiguous()
+            latents = latents[:, :, rank, :, :, :]
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_func_kwargs(
+            self.scheduler.step,
+            {
+                "generator": generator,
+                "eta": eta
+            },
+        )
+
+        target_dtype = PRECISION_TO_TYPE[self.args.precision]
+        autocast_enabled = (target_dtype != torch.float32) and not self.args.disable_autocast
+        vae_dtype = PRECISION_TO_TYPE[self.args.vae_precision]
+        vae_autocast_enabled = (vae_dtype != torch.float32) and not self.args.disable_autocast
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        def dict_to_3d_list(mask_strategy, t_max=50, l_max=60, h_max=24):
+            result = [[[None for _ in range(h_max)] for _ in range(l_max)] for _ in range(t_max)]
+            if mask_strategy is None:
+                return result
+            for key, value in mask_strategy.items():
+                t, l, h = map(int, key.split('_'))
+                result[t][l][h] = value
+            return result
+
+        mask_strategy = dict_to_3d_list(mask_strategy)
+        # if is_progress_bar:
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = (torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents)
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                t_expand = t.repeat(latent_model_input.shape[0])
+                guidance_expand = (torch.tensor(
+                    [embedded_guidance_scale] * latent_model_input.shape[0],
+                    dtype=torch.float32,
+                    device=device,
+                ).to(target_dtype) * 1000.0 if embedded_guidance_scale is not None else None)
+                # predict the noise residual
+                with torch.autocast(device_type="cuda", dtype=target_dtype, enabled=autocast_enabled):
+                    # concat prompt_embeds_2 and prompt_embeds. Mismatch fill with zeros
+                    if prompt_embeds_2.shape[-1] != prompt_embeds.shape[-1]:
+                        prompt_embeds_2 = F.pad(
+                            prompt_embeds_2,
+                            (0, prompt_embeds.shape[2] - prompt_embeds_2.shape[1]),
+                            value=0,
+                        ).unsqueeze(1)
+                    encoder_hidden_states = torch.cat([prompt_embeds_2, prompt_embeds], dim=1)
+                    noise_pred = self.transformer(  # For an input image (129, 192, 336) (1, 256, 256)
+                        latent_model_input,
+                        encoder_hidden_states,
+                        t_expand,
+                        prompt_mask,
+                        mask_strategy=mask_strategy[i],
+                        guidance=guidance_expand,
+                        return_dict=False,
+                    )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(
+                        noise_pred,
+                        noise_pred_text,
+                        guidance_rescale=self.guidance_rescale,
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    if progress_bar is not None:
+                        progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if get_sequence_parallel_state():
+            latents = all_gather(latents, dim=2)
+
+        if not output_type == "latent":
+            expand_temporal_dim = False
+            if len(latents.shape) == 4:
+                if isinstance(self.vae, AutoencoderKLCausal3D):
+                    latents = latents.unsqueeze(2)
+                    expand_temporal_dim = True
+            elif len(latents.shape) == 5:
+                pass
+            else:
+                raise ValueError(
+                    f"Only support latents with shape (b, c, h, w) or (b, c, f, h, w), but got {latents.shape}.")
+
+            if (hasattr(self.vae.config, "shift_factor") and self.vae.config.shift_factor):
+                latents = (latents / self.vae.config.scaling_factor + self.vae.config.shift_factor)
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            with torch.autocast(device_type="cuda", dtype=vae_dtype, enabled=vae_autocast_enabled):
+                if enable_tiling:
+                    self.vae.enable_tiling()
+                if enable_vae_sp:
+                    self.vae.enable_parallel()
+                image = self.vae.decode(latents, return_dict=False, generator=generator)[0]
+
+            if expand_temporal_dim or image.shape[2] == 1:
+                image = image.squeeze(2)
+
+        else:
+            image = latents
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().float()
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return image
+
+        return HunyuanVideoPipelineOutput(videos=image)

fastvideo/models/hunyuan/diffusion/schedulers/__init__.py

@@ -0,0 +1,2 @@
+# ruff: noqa: F401
+from .scheduling_flow_match_discrete import FlowMatchDiscreteScheduler

fastvideo/models/hunyuan/diffusion/schedulers/scheduling_flow_match_discrete.py

@@ -0,0 +1,239 @@
+# Copyright 2024 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+#
+# Modified from diffusers==0.29.2
+#
+# ==============================================================================
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils import BaseOutput, logging
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class FlowMatchDiscreteSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+    """
+
+    prev_sample: torch.FloatTensor
+
+
+class FlowMatchDiscreteScheduler(SchedulerMixin, ConfigMixin):
+    """
+    Euler scheduler.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+    methods the library implements for all schedulers such as loading and saving.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        timestep_spacing (`str`, defaults to `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        shift (`float`, defaults to 1.0):
+            The shift value for the timestep schedule.
+        reverse (`bool`, defaults to `True`):
+            Whether to reverse the timestep schedule.
+    """
+
+    _compatibles = []
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        shift: float = 1.0,
+        reverse: bool = True,
+        solver: str = "euler",
+        n_tokens: Optional[int] = None,
+    ):
+        sigmas = torch.linspace(1, 0, num_train_timesteps + 1)
+
+        if not reverse:
+            sigmas = sigmas.flip(0)
+
+        self.sigmas = sigmas
+        # the value fed to model
+        self.timesteps = (sigmas[:-1] * num_train_timesteps).to(dtype=torch.float32)
+
+        self._step_index = None
+        self._begin_index = None
+
+        self.supported_solver = ["euler"]
+        if solver not in self.supported_solver:
+            raise ValueError(f"Solver {solver} not supported. Supported solvers: {self.supported_solver}")
+
+    @property
+    def step_index(self):
+        """
+        The index counter for current timestep. It will increase 1 after each scheduler step.
+        """
+        return self._step_index
+
+    @property
+    def begin_index(self):
+        """
+        The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+        """
+        return self._begin_index
+
+    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+    def set_begin_index(self, begin_index: int = 0):
+        """
+        Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+        Args:
+            begin_index (`int`):
+                The begin index for the scheduler.
+        """
+        self._begin_index = begin_index
+
+    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,
+        n_tokens: int = None,
+    ):
+        """
+        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+        Args:
+            num_inference_steps (`int`):
+                The number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, *optional*):
+                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+            n_tokens (`int`, *optional*):
+                Number of tokens in the input sequence.
+        """
+        self.num_inference_steps = num_inference_steps
+
+        sigmas = torch.linspace(1, 0, num_inference_steps + 1)
+        sigmas = self.sd3_time_shift(sigmas)
+
+        if not self.config.reverse:
+            sigmas = 1 - sigmas
+
+        self.sigmas = sigmas
+        self.timesteps = (sigmas[:-1] * self.config.num_train_timesteps).to(dtype=torch.float32, device=device)
+
+        # Reset step index
+        self._step_index = None
+
+    def index_for_timestep(self, timestep, schedule_timesteps=None):
+        if schedule_timesteps is None:
+            schedule_timesteps = self.timesteps
+
+        indices = (schedule_timesteps == timestep).nonzero()
+
+        # The sigma index that is taken for the **very** first `step`
+        # is always the second index (or the last index if there is only 1)
+        # This way we can ensure we don't accidentally skip a sigma in
+        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+        pos = 1 if len(indices) > 1 else 0
+
+        return indices[pos].item()
+
+    def _init_step_index(self, timestep):
+        if self.begin_index is None:
+            if isinstance(timestep, torch.Tensor):
+                timestep = timestep.to(self.timesteps.device)
+            self._step_index = self.index_for_timestep(timestep)
+        else:
+            self._step_index = self._begin_index
+
+    def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor:
+        return sample
+
+    def sd3_time_shift(self, t: torch.Tensor):
+        return (self.config.shift * t) / (1 + (self.config.shift - 1) * t)
+
+    def step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+        sample: torch.FloatTensor,
+        return_dict: bool = True,
+    ) -> Union[FlowMatchDiscreteSchedulerOutput, Tuple]:
+        """
+        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+        process from the learned model outputs (most often the predicted noise).
+
+        Args:
+            model_output (`torch.FloatTensor`):
+                The direct output from learned diffusion model.
+            timestep (`float`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.FloatTensor`):
+                A current instance of a sample created by the diffusion process.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            n_tokens (`int`, *optional*):
+                Number of tokens in the input sequence.
+            return_dict (`bool`):
+                Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or
+                tuple.
+
+        Returns:
+            [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is
+                returned, otherwise a tuple is returned where the first element is the sample tensor.
+        """
+
+        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."), )
+
+        if self.step_index is None:
+            self._init_step_index(timestep)
+
+        # Upcast to avoid precision issues when computing prev_sample
+        sample = sample.to(torch.float32)
+
+        dt = self.sigmas[self.step_index + 1] - self.sigmas[self.step_index]
+
+        if self.config.solver == "euler":
+            prev_sample = sample + model_output.to(torch.float32) * dt
+        else:
+            raise ValueError(f"Solver {self.config.solver} not supported. Supported solvers: {self.supported_solver}")
+
+        # upon completion increase step index by one
+        self._step_index += 1
+
+        if not return_dict:
+            return (prev_sample, )
+
+        return FlowMatchDiscreteSchedulerOutput(prev_sample=prev_sample)
+
+    def __len__(self):
+        return self.config.num_train_timesteps

fastvideo/models/hunyuan/idle_config.py

@@ -0,0 +1,380 @@
+# ruff: noqa: F405, F403
+import argparse
+import re
+
+from .constants import *
+from .modules.models import HUNYUAN_VIDEO_CONFIG
+
+
+def parse_args(namespace=None):
+    parser = argparse.ArgumentParser(description="HunyuanVideo inference script")
+
+    parser = add_network_args(parser)
+    parser = add_extra_models_args(parser)
+    parser = add_denoise_schedule_args(parser)
+    parser = add_inference_args(parser)
+    parser = add_parallel_args(parser)
+
+    args = parser.parse_args(namespace=namespace)
+    args = sanity_check_args(args)
+
+    return args
+
+
+def add_network_args(parser: argparse.ArgumentParser):
+    group = parser.add_argument_group(title="HunyuanVideo network args")
+
+    # Main model
+    group.add_argument(
+        "--model",
+        type=str,
+        choices=list(HUNYUAN_VIDEO_CONFIG.keys()),
+        default="HYVideo-T/2-cfgdistill",
+    )
+    group.add_argument(
+        "--latent-channels",
+        type=str,
+        default=16,
+        help="Number of latent channels of DiT. If None, it will be determined by `vae`. If provided, "
+        "it still needs to match the latent channels of the VAE model.",
+    )
+    group.add_argument(
+        "--precision",
+        type=str,
+        default="bf16",
+        choices=PRECISIONS,
+        help="Precision mode. Options: fp32, fp16, bf16. Applied to the backbone model and optimizer.",
+    )
+
+    # RoPE
+    group.add_argument("--rope-theta", type=int, default=256, help="Theta used in RoPE.")
+    return parser
+
+
+def add_extra_models_args(parser: argparse.ArgumentParser):
+    group = parser.add_argument_group(title="Extra models args, including vae, text encoders and tokenizers)")
+
+    # - VAE
+    group.add_argument(
+        "--vae",
+        type=str,
+        default="884-16c-hy",
+        choices=list(VAE_PATH),
+        help="Name of the VAE model.",
+    )
+    group.add_argument(
+        "--vae-precision",
+        type=str,
+        default="fp16",
+        choices=PRECISIONS,
+        help="Precision mode for the VAE model.",
+    )
+    group.add_argument(
+        "--vae-tiling",
+        action="store_true",
+        help="Enable tiling for the VAE model to save GPU memory.",
+    )
+    group.set_defaults(vae_tiling=True)
+
+    group.add_argument(
+        "--text-encoder",
+        type=str,
+        default="llm",
+        choices=list(TEXT_ENCODER_PATH),
+        help="Name of the text encoder model.",
+    )
+    group.add_argument(
+        "--text-encoder-precision",
+        type=str,
+        default="fp16",
+        choices=PRECISIONS,
+        help="Precision mode for the text encoder model.",
+    )
+    group.add_argument(
+        "--text-states-dim",
+        type=int,
+        default=4096,
+        help="Dimension of the text encoder hidden states.",
+    )
+    group.add_argument("--text-len", type=int, default=256, help="Maximum length of the text input.")
+    group.add_argument(
+        "--tokenizer",
+        type=str,
+        default="llm",
+        choices=list(TOKENIZER_PATH),
+        help="Name of the tokenizer model.",
+    )
+    group.add_argument(
+        "--prompt-template",
+        type=str,
+        default="dit-llm-encode",
+        choices=PROMPT_TEMPLATE,
+        help="Image prompt template for the decoder-only text encoder model.",
+    )
+    group.add_argument(
+        "--prompt-template-video",
+        type=str,
+        default="dit-llm-encode-video",
+        choices=PROMPT_TEMPLATE,
+        help="Video prompt template for the decoder-only text encoder model.",
+    )
+    group.add_argument(
+        "--hidden-state-skip-layer",
+        type=int,
+        default=2,
+        help="Skip layer for hidden states.",
+    )
+    group.add_argument(
+        "--apply-final-norm",
+        action="store_true",
+        help="Apply final normalization to the used text encoder hidden states.",
+    )
+
+    # - CLIP
+    group.add_argument(
+        "--text-encoder-2",
+        type=str,
+        default="clipL",
+        choices=list(TEXT_ENCODER_PATH),
+        help="Name of the second text encoder model.",
+    )
+    group.add_argument(
+        "--text-encoder-precision-2",
+        type=str,
+        default="fp16",
+        choices=PRECISIONS,
+        help="Precision mode for the second text encoder model.",
+    )
+    group.add_argument(
+        "--text-states-dim-2",
+        type=int,
+        default=768,
+        help="Dimension of the second text encoder hidden states.",
+    )
+    group.add_argument(
+        "--tokenizer-2",
+        type=str,
+        default="clipL",
+        choices=list(TOKENIZER_PATH),
+        help="Name of the second tokenizer model.",
+    )
+    group.add_argument(
+        "--text-len-2",
+        type=int,
+        default=77,
+        help="Maximum length of the second text input.",
+    )
+
+    return parser
+
+
+def add_denoise_schedule_args(parser: argparse.ArgumentParser):
+    group = parser.add_argument_group(title="Denoise schedule args")
+
+    group.add_argument(
+        "--denoise-type",
+        type=str,
+        default="flow",
+        help="Denoise type for noised inputs.",
+    )
+
+    # Flow Matching
+    group.add_argument(
+        "--flow-shift",
+        type=float,
+        default=7.0,
+        help="Shift factor for flow matching schedulers.",
+    )
+    group.add_argument(
+        "--flow-reverse",
+        action="store_true",
+        help="If reverse, learning/sampling from t=1 -> t=0.",
+    )
+    group.add_argument(
+        "--flow-solver",
+        type=str,
+        default="euler",
+        help="Solver for flow matching.",
+    )
+    group.add_argument(
+        "--use-linear-quadratic-schedule",
+        action="store_true",
+        help="Use linear quadratic schedule for flow matching."
+        "Following MovieGen (https://ai.meta.com/static-resource/movie-gen-research-paper)",
+    )
+    group.add_argument(
+        "--linear-schedule-end",
+        type=int,
+        default=25,
+        help="End step for linear quadratic schedule for flow matching.",
+    )
+
+    return parser
+
+
+def add_inference_args(parser: argparse.ArgumentParser):
+    group = parser.add_argument_group(title="Inference args")
+
+    # ======================== Model loads ========================
+    group.add_argument(
+        "--model-base",
+        type=str,
+        default="ckpts",
+        help="Root path of all the models, including t2v models and extra models.",
+    )
+    group.add_argument(
+        "--dit-weight",
+        type=str,
+        default="ckpts/hunyuan-video-t2v-720p/transformers/mp_rank_00_model_states.pt",
+        help="Path to the HunyuanVideo model. If None, search the model in the args.model_root."
+        "1. If it is a file, load the model directly."
+        "2. If it is a directory, search the model in the directory. Support two types of models: "
+        "1) named `pytorch_model_*.pt`"
+        "2) named `*_model_states.pt`, where * can be `mp_rank_00`.",
+    )
+    group.add_argument(
+        "--model-resolution",
+        type=str,
+        default="540p",
+        choices=["540p", "720p"],
+        help="Root path of all the models, including t2v models and extra models.",
+    )
+    group.add_argument(
+        "--load-key",
+        type=str,
+        default="module",
+        help="Key to load the model states. 'module' for the main model, 'ema' for the EMA model.",
+    )
+    group.add_argument(
+        "--use-cpu-offload",
+        action="store_true",
+        help="Use CPU offload for the model load.",
+    )
+
+    # ======================== Inference general setting ========================
+    group.add_argument(
+        "--batch-size",
+        type=int,
+        default=1,
+        help="Batch size for inference and evaluation.",
+    )
+    group.add_argument(
+        "--infer-steps",
+        type=int,
+        default=50,
+        help="Number of denoising steps for inference.",
+    )
+    group.add_argument(
+        "--disable-autocast",
+        action="store_true",
+        help="Disable autocast for denoising loop and vae decoding in pipeline sampling.",
+    )
+    group.add_argument(
+        "--save-path",
+        type=str,
+        default="./results",
+        help="Path to save the generated samples.",
+    )
+    group.add_argument(
+        "--save-path-suffix",
+        type=str,
+        default="",
+        help="Suffix for the directory of saved samples.",
+    )
+    group.add_argument(
+        "--name-suffix",
+        type=str,
+        default="",
+        help="Suffix for the names of saved samples.",
+    )
+    group.add_argument(
+        "--num-videos",
+        type=int,
+        default=1,
+        help="Number of videos to generate for each prompt.",
+    )
+    # ---sample size---
+    group.add_argument(
+        "--video-size",
+        type=int,
+        nargs="+",
+        default=(720, 1280),
+        help="Video size for training. If a single value is provided, it will be used for both height "
+        "and width. If two values are provided, they will be used for height and width "
+        "respectively.",
+    )
+    group.add_argument(
+        "--video-length",
+        type=int,
+        default=129,
+        help="How many frames to sample from a video. if using 3d vae, the number should be 4n+1",
+    )
+    # --- prompt ---
+    group.add_argument(
+        "--prompt",
+        type=str,
+        default=None,
+        help="Prompt for sampling during evaluation.",
+    )
+    group.add_argument(
+        "--seed-type",
+        type=str,
+        default="auto",
+        choices=["file", "random", "fixed", "auto"],
+        help="Seed type for evaluation. If file, use the seed from the CSV file. If random, generate a "
+        "random seed. If fixed, use the fixed seed given by `--seed`. If auto, `csv` will use the "
+        "seed column if available, otherwise use the fixed `seed` value. `prompt` will use the "
+        "fixed `seed` value.",
+    )
+    group.add_argument("--seed", type=int, default=None, help="Seed for evaluation.")
+
+    # Classifier-Free Guidance
+    group.add_argument("--neg-prompt", type=str, default=None, help="Negative prompt for sampling.")
+    group.add_argument("--cfg-scale", type=float, default=1.0, help="Classifier free guidance scale.")
+    group.add_argument(
+        "--embedded-cfg-scale",
+        type=float,
+        default=6.0,
+        help="Embedded classifier free guidance scale.",
+    )
+
+    group.add_argument(
+        "--reproduce",
+        action="store_true",
+        help="Enable reproducibility by setting random seeds and deterministic algorithms.",
+    )
+
+    return parser
+
+
+def add_parallel_args(parser: argparse.ArgumentParser):
+    group = parser.add_argument_group(title="Parallel args")
+
+    # ======================== Model loads ========================
+    group.add_argument(
+        "--ulysses-degree",
+        type=int,
+        default=1,
+        help="Ulysses degree.",
+    )
+    group.add_argument(
+        "--ring-degree",
+        type=int,
+        default=1,
+        help="Ulysses degree.",
+    )
+
+    return parser
+
+
+def sanity_check_args(args):
+    # VAE channels
+    vae_pattern = r"\d{2,3}-\d{1,2}c-\w+"
+    if not re.match(vae_pattern, args.vae):
+        raise ValueError(f"Invalid VAE model: {args.vae}. Must be in the format of '{vae_pattern}'.")
+    vae_channels = int(args.vae.split("-")[1][:-1])
+    if args.latent_channels is None:
+        args.latent_channels = vae_channels
+    if vae_channels != args.latent_channels:
+        raise ValueError(f"Latent channels ({args.latent_channels}) must match the VAE channels ({vae_channels}).")
+    return args

fastvideo/models/hunyuan/inference.py

@@ -0,0 +1,482 @@
+import os
+import random
+import time
+from pathlib import Path
+
+import torch
+from loguru import logger
+from safetensors.torch import load_file as safetensors_load_file
+
+from fastvideo.models.hunyuan.constants import NEGATIVE_PROMPT, PRECISION_TO_TYPE, PROMPT_TEMPLATE
+from fastvideo.models.hunyuan.diffusion.pipelines import HunyuanVideoPipeline
+from fastvideo.models.hunyuan.diffusion.schedulers import FlowMatchDiscreteScheduler
+from fastvideo.models.hunyuan.modules import load_model
+from fastvideo.models.hunyuan.text_encoder import TextEncoder
+from fastvideo.models.hunyuan.utils.data_utils import align_to
+from fastvideo.models.hunyuan.vae import load_vae
+from fastvideo.utils.parallel_states import nccl_info
+
+
+class Inference(object):
+
+    def __init__(
+        self,
+        args,
+        vae,
+        vae_kwargs,
+        text_encoder,
+        model,
+        text_encoder_2=None,
+        pipeline=None,
+        use_cpu_offload=False,
+        device=None,
+        logger=None,
+        parallel_args=None,
+    ):
+        self.vae = vae
+        self.vae_kwargs = vae_kwargs
+
+        self.text_encoder = text_encoder
+        self.text_encoder_2 = text_encoder_2
+
+        self.model = model
+        self.pipeline = pipeline
+        self.use_cpu_offload = use_cpu_offload
+
+        self.args = args
+        self.device = (device if device is not None else "cuda" if torch.cuda.is_available() else "cpu")
+        self.logger = logger
+        self.parallel_args = parallel_args
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_path, args, device=None, **kwargs):
+        """
+        Initialize the Inference pipeline.
+
+        Args:
+            pretrained_model_path (str or pathlib.Path): The model path, including t2v, text encoder and vae checkpoints.
+            args (argparse.Namespace): The arguments for the pipeline.
+            device (int): The device for inference. Default is 0.
+        """
+        # ========================================================================
+        logger.info(f"Got text-to-video model root path: {pretrained_model_path}")
+
+        # ==================== Initialize Distributed Environment ================
+        if nccl_info.sp_size > 1:
+            device = torch.device(f"cuda:{os.environ['LOCAL_RANK']}")
+        if device is None:
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        parallel_args = None  # {"ulysses_degree": args.ulysses_degree, "ring_degree": args.ring_degree}
+
+        # ======================== Get the args path =============================
+
+        # Disable gradient
+        torch.set_grad_enabled(False)
+
+        # =========================== Build main model ===========================
+        logger.info("Building model...")
+        factor_kwargs = {"device": device, "dtype": PRECISION_TO_TYPE[args.precision]}
+        in_channels = args.latent_channels
+        out_channels = args.latent_channels
+
+        model = load_model(
+            args,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            factor_kwargs=factor_kwargs,
+        )
+        model = model.to(device)
+        model = Inference.load_state_dict(args, model, pretrained_model_path)
+        if args.enable_torch_compile:
+            model = torch.compile(model)
+        model.eval()
+
+        # ============================= Build extra models ========================
+        # VAE
+        vae, _, s_ratio, t_ratio = load_vae(
+            args.vae,
+            args.vae_precision,
+            logger=logger,
+            device=device if not args.use_cpu_offload else "cpu",
+        )
+        vae_kwargs = {"s_ratio": s_ratio, "t_ratio": t_ratio}
+
+        # Text encoder
+        if args.prompt_template_video is not None:
+            crop_start = PROMPT_TEMPLATE[args.prompt_template_video].get("crop_start", 0)
+        elif args.prompt_template is not None:
+            crop_start = PROMPT_TEMPLATE[args.prompt_template].get("crop_start", 0)
+        else:
+            crop_start = 0
+        max_length = args.text_len + crop_start
+
+        # prompt_template
+        prompt_template = (PROMPT_TEMPLATE[args.prompt_template] if args.prompt_template is not None else None)
+
+        # prompt_template_video
+        prompt_template_video = (PROMPT_TEMPLATE[args.prompt_template_video]
+                                 if args.prompt_template_video is not None else None)
+
+        text_encoder = TextEncoder(
+            text_encoder_type=args.text_encoder,
+            max_length=max_length,
+            text_encoder_precision=args.text_encoder_precision,
+            tokenizer_type=args.tokenizer,
+            prompt_template=prompt_template,
+            prompt_template_video=prompt_template_video,
+            hidden_state_skip_layer=args.hidden_state_skip_layer,
+            apply_final_norm=args.apply_final_norm,
+            reproduce=args.reproduce,
+            logger=logger,
+            device=device if not args.use_cpu_offload else "cpu",
+        )
+        text_encoder_2 = None
+        if args.text_encoder_2 is not None:
+            text_encoder_2 = TextEncoder(
+                text_encoder_type=args.text_encoder_2,
+                max_length=args.text_len_2,
+                text_encoder_precision=args.text_encoder_precision_2,
+                tokenizer_type=args.tokenizer_2,
+                reproduce=args.reproduce,
+                logger=logger,
+                device=device if not args.use_cpu_offload else "cpu",
+            )
+
+        return cls(
+            args=args,
+            vae=vae,
+            vae_kwargs=vae_kwargs,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            model=model,
+            use_cpu_offload=args.use_cpu_offload,
+            device=device,
+            logger=logger,
+            parallel_args=parallel_args,
+        )
+
+    @staticmethod
+    def load_state_dict(args, model, pretrained_model_path):
+        load_key = args.load_key
+        dit_weight = Path(args.dit_weight)
+
+        if dit_weight is None:
+            model_dir = pretrained_model_path / f"t2v_{args.model_resolution}"
+            files = list(model_dir.glob("*.pt"))
+            if len(files) == 0:
+                raise ValueError(f"No model weights found in {model_dir}")
+            if str(files[0]).startswith("pytorch_model_"):
+                model_path = dit_weight / f"pytorch_model_{load_key}.pt"
+                bare_model = True
+            elif any(str(f).endswith("_model_states.pt") for f in files):
+                files = [f for f in files if str(f).endswith("_model_states.pt")]
+                model_path = files[0]
+                if len(files) > 1:
+                    logger.warning(f"Multiple model weights found in {dit_weight}, using {model_path}")
+                bare_model = False
+            else:
+                raise ValueError(f"Invalid model path: {dit_weight} with unrecognized weight format: "
+                                 f"{list(map(str, files))}. When given a directory as --dit-weight, only "
+                                 f"`pytorch_model_*.pt`(provided by HunyuanDiT official) and "
+                                 f"`*_model_states.pt`(saved by deepspeed) can be parsed. If you want to load a "
+                                 f"specific weight file, please provide the full path to the file.")
+        else:
+            if dit_weight.is_dir():
+                files = list(dit_weight.glob("*.pt"))
+                if len(files) == 0:
+                    raise ValueError(f"No model weights found in {dit_weight}")
+                if str(files[0]).startswith("pytorch_model_"):
+                    model_path = dit_weight / f"pytorch_model_{load_key}.pt"
+                    bare_model = True
+                elif any(str(f).endswith("_model_states.pt") for f in files):
+                    files = [f for f in files if str(f).endswith("_model_states.pt")]
+                    model_path = files[0]
+                    if len(files) > 1:
+                        logger.warning(f"Multiple model weights found in {dit_weight}, using {model_path}")
+                    bare_model = False
+                else:
+                    raise ValueError(f"Invalid model path: {dit_weight} with unrecognized weight format: "
+                                     f"{list(map(str, files))}. When given a directory as --dit-weight, only "
+                                     f"`pytorch_model_*.pt`(provided by HunyuanDiT official) and "
+                                     f"`*_model_states.pt`(saved by deepspeed) can be parsed. If you want to load a "
+                                     f"specific weight file, please provide the full path to the file.")
+            elif dit_weight.is_file():
+                model_path = dit_weight
+                bare_model = "unknown"
+            else:
+                raise ValueError(f"Invalid model path: {dit_weight}")
+
+        if not model_path.exists():
+            raise ValueError(f"model_path not exists: {model_path}")
+        logger.info(f"Loading torch model {model_path}...")
+        if model_path.suffix == ".safetensors":
+            # Use safetensors library for .safetensors files
+            state_dict = safetensors_load_file(model_path)
+        elif model_path.suffix == ".pt":
+            # Use torch for .pt files
+            state_dict = torch.load(model_path, map_location=lambda storage, loc: storage)
+        else:
+            raise ValueError(f"Unsupported file format: {model_path}")
+
+        if bare_model == "unknown" and ("ema" in state_dict or "module" in state_dict):
+            bare_model = False
+        if bare_model is False:
+            if load_key in state_dict:
+                state_dict = state_dict[load_key]
+            else:
+                raise KeyError(f"Missing key: `{load_key}` in the checkpoint: {model_path}. The keys in the checkpoint "
+                               f"are: {list(state_dict.keys())}.")
+        model.load_state_dict(state_dict, strict=True)
+        return model
+
+    @staticmethod
+    def parse_size(size):
+        if isinstance(size, int):
+            size = [size]
+        if not isinstance(size, (list, tuple)):
+            raise ValueError(f"Size must be an integer or (height, width), got {size}.")
+        if len(size) == 1:
+            size = [size[0], size[0]]
+        if len(size) != 2:
+            raise ValueError(f"Size must be an integer or (height, width), got {size}.")
+        return size
+
+
+class HunyuanVideoSampler(Inference):
+
+    def __init__(
+        self,
+        args,
+        vae,
+        vae_kwargs,
+        text_encoder,
+        model,
+        text_encoder_2=None,
+        pipeline=None,
+        use_cpu_offload=False,
+        device=0,
+        logger=None,
+        parallel_args=None,
+    ):
+        super().__init__(
+            args,
+            vae,
+            vae_kwargs,
+            text_encoder,
+            model,
+            text_encoder_2=text_encoder_2,
+            pipeline=pipeline,
+            use_cpu_offload=use_cpu_offload,
+            device=device,
+            logger=logger,
+            parallel_args=parallel_args,
+        )
+
+        self.pipeline = self.load_diffusion_pipeline(
+            args=args,
+            vae=self.vae,
+            text_encoder=self.text_encoder,
+            text_encoder_2=self.text_encoder_2,
+            model=self.model,
+            device=self.device,
+        )
+
+        self.default_negative_prompt = NEGATIVE_PROMPT
+
+    def load_diffusion_pipeline(
+        self,
+        args,
+        vae,
+        text_encoder,
+        text_encoder_2,
+        model,
+        scheduler=None,
+        device=None,
+        progress_bar_config=None,
+        data_type="video",
+    ):
+        """Load the denoising scheduler for inference."""
+        if scheduler is None:
+            if args.denoise_type == "flow":
+                scheduler = FlowMatchDiscreteScheduler(
+                    shift=args.flow_shift,
+                    reverse=args.flow_reverse,
+                    solver=args.flow_solver,
+                )
+            else:
+                raise ValueError(f"Invalid denoise type {args.denoise_type}")
+
+        pipeline = HunyuanVideoPipeline(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            transformer=model,
+            scheduler=scheduler,
+            progress_bar_config=progress_bar_config,
+            args=args,
+        )
+        if self.use_cpu_offload:
+            pipeline.enable_sequential_cpu_offload()
+        else:
+            pipeline = pipeline.to(device)
+
+        return pipeline
+
+    @torch.no_grad()
+    def predict(
+        self,
+        prompt,
+        height=192,
+        width=336,
+        video_length=129,
+        seed=None,
+        negative_prompt=None,
+        infer_steps=50,
+        guidance_scale=6,
+        flow_shift=5.0,
+        embedded_guidance_scale=None,
+        batch_size=1,
+        num_videos_per_prompt=1,
+        mask_strategy=None,
+        **kwargs,
+    ):
+        """
+        Predict the image/video from the given text.
+
+        Args:
+            prompt (str or List[str]): The input text.
+            kwargs:
+                height (int): The height of the output video. Default is 192.
+                width (int): The width of the output video. Default is 336.
+                video_length (int): The frame number of the output video. Default is 129.
+                seed (int or List[str]): The random seed for the generation. Default is a random integer.
+                negative_prompt (str or List[str]): The negative text prompt. Default is an empty string.
+                guidance_scale (float): The guidance scale for the generation. Default is 6.0.
+                num_images_per_prompt (int): The number of images per prompt. Default is 1.
+                infer_steps (int): The number of inference steps. Default is 100.
+        """
+
+        out_dict = dict()
+
+        # ========================================================================
+        # Arguments: seed
+        # ========================================================================
+        if isinstance(seed, torch.Tensor):
+            seed = seed.tolist()
+        if seed is None:
+            seeds = [random.randint(0, 1_000_000) for _ in range(batch_size * num_videos_per_prompt)]
+        elif isinstance(seed, int):
+            seeds = [seed + i for _ in range(batch_size) for i in range(num_videos_per_prompt)]
+        elif isinstance(seed, (list, tuple)):
+            if len(seed) == batch_size:
+                seeds = [int(seed[i]) + j for i in range(batch_size) for j in range(num_videos_per_prompt)]
+            elif len(seed) == batch_size * num_videos_per_prompt:
+                seeds = [int(s) for s in seed]
+            else:
+                raise ValueError(
+                    f"Length of seed must be equal to number of prompt(batch_size) or "
+                    f"batch_size * num_videos_per_prompt ({batch_size} * {num_videos_per_prompt}), got {seed}.")
+        else:
+            raise ValueError(f"Seed must be an integer, a list of integers, or None, got {seed}.")
+        # Peiyuan: using GPU seed will cause A100 and H100 to generate different results...
+        generator = [torch.Generator("cpu").manual_seed(seed) for seed in seeds]
+        out_dict["seeds"] = seeds
+
+        # ========================================================================
+        # Arguments: target_width, target_height, target_video_length
+        # ========================================================================
+        if width <= 0 or height <= 0 or video_length <= 0:
+            raise ValueError(
+                f"`height` and `width` and `video_length` must be positive integers, got height={height}, width={width}, video_length={video_length}"
+            )
+        if (video_length - 1) % 4 != 0:
+            raise ValueError(f"`video_length-1` must be a multiple of 4, got {video_length}")
+
+        logger.info(f"Input (height, width, video_length) = ({height}, {width}, {video_length})")
+
+        target_height = align_to(height, 16)
+        target_width = align_to(width, 16)
+        target_video_length = video_length
+
+        out_dict["size"] = (target_height, target_width, target_video_length)
+
+        # ========================================================================
+        # Arguments: prompt, new_prompt, negative_prompt
+        # ========================================================================
+        if not isinstance(prompt, str):
+            raise TypeError(f"`prompt` must be a string, but got {type(prompt)}")
+        prompt = [prompt.strip()]
+
+        # negative prompt
+        if negative_prompt is None or negative_prompt == "":
+            negative_prompt = self.default_negative_prompt
+        if not isinstance(negative_prompt, str):
+            raise TypeError(f"`negative_prompt` must be a string, but got {type(negative_prompt)}")
+        negative_prompt = [negative_prompt.strip()]
+
+        # ========================================================================
+        # Scheduler
+        # ========================================================================
+        scheduler = FlowMatchDiscreteScheduler(
+            shift=flow_shift,
+            reverse=self.args.flow_reverse,
+            solver=self.args.flow_solver,
+        )
+        self.pipeline.scheduler = scheduler
+
+        if "884" in self.args.vae:
+            latents_size = [(video_length - 1) // 4 + 1, height // 8, width // 8]
+        elif "888" in self.args.vae:
+            latents_size = [(video_length - 1) // 8 + 1, height // 8, width // 8]
+        n_tokens = latents_size[0] * latents_size[1] * latents_size[2]
+
+        # ========================================================================
+        # Print infer args
+        # ========================================================================
+        debug_str = f"""
+                        height: {target_height}
+                         width: {target_width}
+                  video_length: {target_video_length}
+                        prompt: {prompt}
+                    neg_prompt: {negative_prompt}
+                          seed: {seed}
+                   infer_steps: {infer_steps}
+         num_videos_per_prompt: {num_videos_per_prompt}
+                guidance_scale: {guidance_scale}
+                      n_tokens: {n_tokens}
+                    flow_shift: {flow_shift}
+       embedded_guidance_scale: {embedded_guidance_scale}"""
+        logger.debug(debug_str)
+
+        # ========================================================================
+        # Pipeline inference
+        # ========================================================================
+        start_time = time.time()
+        samples = self.pipeline(
+            prompt=prompt,
+            height=target_height,
+            width=target_width,
+            video_length=target_video_length,
+            num_inference_steps=infer_steps,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            num_videos_per_prompt=num_videos_per_prompt,
+            generator=generator,
+            output_type="pil",
+            n_tokens=n_tokens,
+            embedded_guidance_scale=embedded_guidance_scale,
+            data_type="video" if target_video_length > 1 else "image",
+            is_progress_bar=True,
+            vae_ver=self.args.vae,
+            enable_tiling=self.args.vae_tiling,
+            enable_vae_sp=self.args.vae_sp,
+            mask_strategy=mask_strategy,
+        )[0]
+        out_dict["samples"] = samples
+        out_dict["prompts"] = prompt
+
+        gen_time = time.time() - start_time
+        logger.info(f"Success, time: {gen_time}")
+
+        return out_dict

fastvideo/models/hunyuan/modules/__init__.py

@@ -0,0 +1,25 @@
+from .models import HUNYUAN_VIDEO_CONFIG, HYVideoDiffusionTransformer
+
+
+def load_model(args, in_channels, out_channels, factor_kwargs):
+    """load hunyuan video model
+
+    Args:
+        args (dict): model args
+        in_channels (int): input channels number
+        out_channels (int): output channels number
+        factor_kwargs (dict): factor kwargs
+
+    Returns:
+        model (nn.Module): The hunyuan video model
+    """
+    if args.model in HUNYUAN_VIDEO_CONFIG.keys():
+        model = HYVideoDiffusionTransformer(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            **HUNYUAN_VIDEO_CONFIG[args.model],
+            **factor_kwargs,
+        )
+        return model
+    else:
+        raise NotImplementedError()

fastvideo/models/hunyuan/modules/activation_layers.py

@@ -0,0 +1,23 @@
+import torch.nn as nn
+
+
+def get_activation_layer(act_type):
+    """get activation layer
+
+    Args:
+        act_type (str): the activation type
+
+    Returns:
+        torch.nn.functional: the activation layer
+    """
+    if act_type == "gelu":
+        return lambda: nn.GELU()
+    elif act_type == "gelu_tanh":
+        # Approximate `tanh` requires torch >= 1.13
+        return lambda: nn.GELU(approximate="tanh")
+    elif act_type == "relu":
+        return nn.ReLU
+    elif act_type == "silu":
+        return nn.SiLU
+    else:
+        raise ValueError(f"Unknown activation type: {act_type}")

fastvideo/models/hunyuan/modules/attenion.py

@@ -0,0 +1,124 @@
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+
+try:
+    from st_attn import sliding_tile_attention
+except ImportError:
+    print("Could not load Sliding Tile Attention.")
+    sliding_tile_attention = None
+
+from fastvideo.models.flash_attn_no_pad import flash_attn_no_pad
+from fastvideo.utils.communications import all_gather, all_to_all_4D
+from fastvideo.utils.parallel_states import get_sequence_parallel_state, nccl_info
+
+
+def attention(
+    q,
+    k,
+    v,
+    drop_rate=0,
+    attn_mask=None,
+    causal=False,
+):
+
+    qkv = torch.stack([q, k, v], dim=2)
+
+    if attn_mask is not None and attn_mask.dtype != torch.bool:
+        attn_mask = attn_mask.bool()
+
+    x = flash_attn_no_pad(qkv, attn_mask, causal=causal, dropout_p=drop_rate, softmax_scale=None)
+
+    b, s, a, d = x.shape
+    out = x.reshape(b, s, -1)
+    return out
+
+
+def tile(x, sp_size):
+    x = rearrange(x, "b (sp t h w) head d -> b (t sp h w) head d", sp=sp_size, t=30 // sp_size, h=48, w=80)
+    return rearrange(x,
+                     "b (n_t ts_t n_h ts_h n_w ts_w) h d -> b (n_t n_h n_w ts_t ts_h ts_w) h d",
+                     n_t=5,
+                     n_h=6,
+                     n_w=10,
+                     ts_t=6,
+                     ts_h=8,
+                     ts_w=8)
+
+
+def untile(x, sp_size):
+    x = rearrange(x,
+                  "b (n_t n_h n_w ts_t ts_h ts_w) h d -> b (n_t ts_t n_h ts_h n_w ts_w) h d",
+                  n_t=5,
+                  n_h=6,
+                  n_w=10,
+                  ts_t=6,
+                  ts_h=8,
+                  ts_w=8)
+    return rearrange(x, "b (t sp h w) head d -> b (sp t h w) head d", sp=sp_size, t=30 // sp_size, h=48, w=80)
+
+
+def parallel_attention(q, k, v, img_q_len, img_kv_len, text_mask, mask_strategy=None):
+    query, encoder_query = q
+    key, encoder_key = k
+    value, encoder_value = v
+    text_length = text_mask.sum()
+
+    if get_sequence_parallel_state():
+        # batch_size, seq_len, attn_heads, head_dim
+        query = all_to_all_4D(query, scatter_dim=2, gather_dim=1)
+        key = all_to_all_4D(key, scatter_dim=2, gather_dim=1)
+        value = all_to_all_4D(value, scatter_dim=2, gather_dim=1)
+
+        def shrink_head(encoder_state, dim):
+            local_heads = encoder_state.shape[dim] // nccl_info.sp_size
+            return encoder_state.narrow(dim, nccl_info.rank_within_group * local_heads, local_heads)
+
+        encoder_query = shrink_head(encoder_query, dim=2)
+        encoder_key = shrink_head(encoder_key, dim=2)
+        encoder_value = shrink_head(encoder_value, dim=2)
+        # [b, s, h, d]
+
+    sequence_length = query.size(1)
+    encoder_sequence_length = encoder_query.size(1)
+
+    if mask_strategy[0] is not None:
+        query = torch.cat([tile(query, nccl_info.sp_size), encoder_query], dim=1).transpose(1, 2)
+        key = torch.cat([tile(key, nccl_info.sp_size), encoder_key], dim=1).transpose(1, 2)
+        value = torch.cat([tile(value, nccl_info.sp_size), encoder_value], dim=1).transpose(1, 2)
+
+        head_num = query.size(1)
+        current_rank = nccl_info.rank_within_group
+        start_head = current_rank * head_num
+        windows = [mask_strategy[head_idx + start_head] for head_idx in range(head_num)]
+
+        hidden_states = sliding_tile_attention(query, key, value, windows, text_length).transpose(1, 2)
+    else:
+        query = torch.cat([query, encoder_query], dim=1)
+        key = torch.cat([key, encoder_key], dim=1)
+        value = torch.cat([value, encoder_value], dim=1)
+        # B, S, 3, H, D
+        qkv = torch.stack([query, key, value], dim=2)
+
+        attn_mask = F.pad(text_mask, (sequence_length, 0), value=True)
+        hidden_states = flash_attn_no_pad(qkv, attn_mask, causal=False, dropout_p=0.0, softmax_scale=None)
+
+    hidden_states, encoder_hidden_states = hidden_states.split_with_sizes((sequence_length, encoder_sequence_length),
+                                                                          dim=1)
+
+    if mask_strategy[0] is not None:
+        hidden_states = untile(hidden_states, nccl_info.sp_size)
+
+    if get_sequence_parallel_state():
+        hidden_states = all_to_all_4D(hidden_states, scatter_dim=1, gather_dim=2)
+        encoder_hidden_states = all_gather(encoder_hidden_states, dim=2).contiguous()
+
+    hidden_states = hidden_states.to(query.dtype)
+    encoder_hidden_states = encoder_hidden_states.to(query.dtype)
+
+    attn = torch.cat([hidden_states, encoder_hidden_states], dim=1)
+
+    b, s, a, d = attn.shape
+    attn = attn.reshape(b, s, -1)
+
+    return attn

fastvideo/models/hunyuan/modules/embed_layers.py

@@ -0,0 +1,150 @@
+import math
+
+import torch
+import torch.nn as nn
+
+from ..utils.helpers import to_2tuple
+
+
+class PatchEmbed(nn.Module):
+    """2D Image to Patch Embedding
+
+    Image to Patch Embedding using Conv2d
+
+    A convolution based approach to patchifying a 2D image w/ embedding projection.
+
+    Based on the impl in https://github.com/google-research/vision_transformer
+
+    Hacked together by / Copyright 2020 Ross Wightman
+
+    Remove the _assert function in forward function to be compatible with multi-resolution images.
+    """
+
+    def __init__(
+        self,
+        patch_size=16,
+        in_chans=3,
+        embed_dim=768,
+        norm_layer=None,
+        flatten=True,
+        bias=True,
+        dtype=None,
+        device=None,
+    ):
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+        self.flatten = flatten
+
+        self.proj = nn.Conv3d(
+            in_chans,
+            embed_dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=bias,
+            **factory_kwargs,
+        )
+        nn.init.xavier_uniform_(self.proj.weight.view(self.proj.weight.size(0), -1))
+        if bias:
+            nn.init.zeros_(self.proj.bias)
+
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+    def forward(self, x):
+        x = self.proj(x)
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
+        x = self.norm(x)
+        return x
+
+
+class TextProjection(nn.Module):
+    """
+    Projects text embeddings. Also handles dropout for classifier-free guidance.
+
+    Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
+    """
+
+    def __init__(self, in_channels, hidden_size, act_layer, dtype=None, device=None):
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+        self.linear_1 = nn.Linear(
+            in_features=in_channels,
+            out_features=hidden_size,
+            bias=True,
+            **factory_kwargs,
+        )
+        self.act_1 = act_layer()
+        self.linear_2 = nn.Linear(
+            in_features=hidden_size,
+            out_features=hidden_size,
+            bias=True,
+            **factory_kwargs,
+        )
+
+    def forward(self, caption):
+        hidden_states = self.linear_1(caption)
+        hidden_states = self.act_1(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+
+def timestep_embedding(t, dim, max_period=10000):
+    """
+    Create sinusoidal timestep embeddings.
+
+    Args:
+        t (torch.Tensor): a 1-D Tensor of N indices, one per batch element. These may be fractional.
+        dim (int): the dimension of the output.
+        max_period (int): controls the minimum frequency of the embeddings.
+
+    Returns:
+        embedding (torch.Tensor): An (N, D) Tensor of positional embeddings.
+
+    .. ref_link: https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
+    """
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) /
+                      half).to(device=t.device)
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+
+    def __init__(
+        self,
+        hidden_size,
+        act_layer,
+        frequency_embedding_size=256,
+        max_period=10000,
+        out_size=None,
+        dtype=None,
+        device=None,
+    ):
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+        self.frequency_embedding_size = frequency_embedding_size
+        self.max_period = max_period
+        if out_size is None:
+            out_size = hidden_size
+
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True, **factory_kwargs),
+            act_layer(),
+            nn.Linear(hidden_size, out_size, bias=True, **factory_kwargs),
+        )
+        nn.init.normal_(self.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.mlp[2].weight, std=0.02)
+
+    def forward(self, t):
+        t_freq = timestep_embedding(t, self.frequency_embedding_size, self.max_period).type(self.mlp[0].weight.dtype)
+        t_emb = self.mlp(t_freq)
+        return t_emb

fastvideo/models/hunyuan/modules/mlp_layers.py

@@ -0,0 +1,107 @@
+# Modified from timm library:
+# https://github.com/huggingface/pytorch-image-models/blob/648aaa41233ba83eb38faf5ba9d415d574823241/timm/layers/mlp.py#L13
+
+from functools import partial
+
+import torch
+import torch.nn as nn
+
+from ..utils.helpers import to_2tuple
+from .modulate_layers import modulate
+
+
+class MLP(nn.Module):
+    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""
+
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        norm_layer=None,
+        bias=True,
+        drop=0.0,
+        use_conv=False,
+        device=None,
+        dtype=None,
+    ):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        out_features = out_features or in_channels
+        hidden_channels = hidden_channels or in_channels
+        bias = to_2tuple(bias)
+        drop_probs = to_2tuple(drop)
+        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
+
+        self.fc1 = linear_layer(in_channels, hidden_channels, bias=bias[0], **factory_kwargs)
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.norm = (norm_layer(hidden_channels, **factory_kwargs) if norm_layer is not None else nn.Identity())
+        self.fc2 = linear_layer(hidden_channels, out_features, bias=bias[1], **factory_kwargs)
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.norm(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+#
+class MLPEmbedder(nn.Module):
+    """copied from https://github.com/black-forest-labs/flux/blob/main/src/flux/modules/layers.py"""
+
+    def __init__(self, in_dim: int, hidden_dim: int, device=None, dtype=None):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True, **factory_kwargs)
+        self.silu = nn.SiLU()
+        self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True, **factory_kwargs)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.out_layer(self.silu(self.in_layer(x)))
+
+
+class FinalLayer(nn.Module):
+    """The final layer of DiT."""
+
+    def __init__(self, hidden_size, patch_size, out_channels, act_layer, device=None, dtype=None):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+
+        # Just use LayerNorm for the final layer
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+        if isinstance(patch_size, int):
+            self.linear = nn.Linear(
+                hidden_size,
+                patch_size * patch_size * out_channels,
+                bias=True,
+                **factory_kwargs,
+            )
+        else:
+            self.linear = nn.Linear(
+                hidden_size,
+                patch_size[0] * patch_size[1] * patch_size[2] * out_channels,
+                bias=True,
+            )
+        nn.init.zeros_(self.linear.weight)
+        nn.init.zeros_(self.linear.bias)
+
+        # Here we don't distinguish between the modulate types. Just use the simple one.
+        self.adaLN_modulation = nn.Sequential(
+            act_layer(),
+            nn.Linear(hidden_size, 2 * hidden_size, bias=True, **factory_kwargs),
+        )
+        # Zero-initialize the modulation
+        nn.init.zeros_(self.adaLN_modulation[1].weight)
+        nn.init.zeros_(self.adaLN_modulation[1].bias)
+
+    def forward(self, x, c):
+        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift=shift, scale=scale)
+        x = self.linear(x)
+        return x