separate critic lr

.buildkite/pipeline.yml

@@ -13,12 +13,12 @@ steps:
   - label: "Trigger Tests"
     plugins:
       - monorepo-diff#v1.4.0:
-          diff: 'git fetch origin "$BUILDKITE_PULL_REQUEST_BASE_BRANCH" && git diff --name-only origin/"$BUILDKITE_PULL_REQUEST_BASE_BRANCH"...HEAD'
+          diff: "git diff --name-only $BUILDKITE_PULL_REQUEST_BASE_BRANCH...HEAD"
           watch:
             - path: 
-                - "fastvideo/models/encoders/**"
-                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/encoders/**"
+                - "fastvideo/v1/models/encoders/**"
+                - "fastvideo/v1/models/loader/**"
+                - "fastvideo/v1/tests/encoders/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -29,9 +29,9 @@ steps:
                 agents:
                   queue: "default"
             - path:
-                - "fastvideo/models/vaes/**"
-                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/vaes/**"
+                - "fastvideo/v1/models/vaes/**"
+                - "fastvideo/v1/models/loader/**"
+                - "fastvideo/v1/tests/vaes/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -42,11 +42,11 @@ steps:
                 agents:
                   queue: "default"
             - path:
-                - "fastvideo/models/dits/**"
-                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/transformers/**"
-                - "fastvideo/layers/**"
-                - "fastvideo/attention/**"
+                - "fastvideo/v1/models/dits/**"
+                - "fastvideo/v1/models/loader/**"
+                - "fastvideo/v1/tests/transformers/**"
+                - "fastvideo/v1/layers/**"
+                - "fastvideo/v1/attention/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -57,33 +57,16 @@ steps:
                 agents:
                   queue: "default"
             - path: 
-                - "fastvideo/**/*.py"
-                - "pyproject.toml"
-                - "docker/Dockerfile.python3.12"
+                - "fastvideo/v1/**/*.py"
               config:
-                command: "timeout 45m .buildkite/scripts/pr_test.sh"
+                command: "timeout 30m .buildkite/scripts/pr_test.sh"
                 label: "SSIM Tests"
                 env:
                   - TEST_TYPE=ssim
                 agents:
                   queue: "default"
             - path:
-                - "fastvideo/tests/lora/**"
-                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/transformers/**"
-                - "fastvideo/pipelines/**"
-                - "fastvideo/layers/lora/**"
-                - "pyproject.toml"
-                - "docker/Dockerfile.python3.12"
-              config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "LoRA Inference Tests"
-                env:
-                  - TEST_TYPE=inference_lora
-                agents:
-                  queue: "default"
-            - path:
-                - "fastvideo/**"
+                - "fastvideo/v1/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -94,18 +77,7 @@ steps:
                 agents:
                   queue: "default"
             - path:
-                - "fastvideo/**"
-                - "pyproject.toml"
-                - "docker/Dockerfile.python3.12"
-              config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
-                label: "LoRA Training Tests"
-                env:
-                  - TEST_TYPE=training_lora
-                agents:
-                  queue: "default"
-            - path:
-                - "fastvideo/**"
+                - "fastvideo/v1/**"
                 - "csrc/attn/vsa/**"
                 - "csrc/attn/tk/**"
                 - "csrc/attn/setup_vsa.py"
@@ -121,7 +93,7 @@ steps:
                 agents:
                   queue: "default"
             - path:
-                - "fastvideo/**"
+                - "fastvideo/v1/**"
                 - "csrc/attn/st_attn/**"
                 - "csrc/attn/setup_sta.py"
                 - "csrc/attn/config_sta.py"

.buildkite/scripts/pr_test.sh

@@ -50,7 +50,7 @@ else
     exit 1
 fi
 
-MODAL_TEST_FILE="fastvideo/tests/modal/pr_test.py"
+MODAL_TEST_FILE="fastvideo/v1/tests/modal/pr_test.py"
 
 if [ -z "${TEST_TYPE:-}" ]; then
     log "Error: TEST_TYPE environment variable is not set"
@@ -81,10 +81,6 @@ case "$TEST_TYPE" in
         log "Running training tests..."
         MODAL_COMMAND="$MODAL_ENV WANDB_API_KEY=$WANDB_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_training_tests"
         ;;
-    "training_lora")
-        log "Running LoRA training tests..."
-        MODAL_COMMAND="$MODAL_ENV WANDB_API_KEY=$WANDB_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_training_lora_tests"
-        ;;
     "training_vsa")
         log "Running training VSA tests..."
         MODAL_COMMAND="$MODAL_ENV WANDB_API_KEY=$WANDB_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_training_tests_VSA"
@@ -101,10 +97,6 @@ case "$TEST_TYPE" in
         log "Running precision VSA tests..."
         MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_precision_tests_VSA"
         ;;
-    "inference_lora")
-        log "Running LoRA tests..."
-        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_inference_lora_tests"
-        ;;
     *)
         log "Error: Unknown test type: $TEST_TYPE"
         exit 1

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

@@ -23,7 +23,7 @@ body:
   attributes:
     label: Environment
     description: |
-      Please share your environment with us. You can run the command **python collect_env.py** and copy-paste its output below.
+      Please share your environment with us. You can run the command **python fastvideo/utils/collect_env.py** and copy-paste its output below.
     placeholder: FastVideo version, platform, python version, cuda version...
   validations:
     required: true

.github/ISSUE_TEMPLATE/3-RFC.yml

@@ -1,56 +0,0 @@
-name: 💬 Request for comments (RFC).
-description: Ask for feedback on major architectural changes or design choices.
-title: "[RFC]: "
-labels: ["RFC"]
-
-body:
-- type: markdown
-  attributes:
-    value: >
-      #### Please take a look at previous [RFCs](https://github.com/hao-ai-lab/FastVideo/issues?q=label%3ARFC+sort%3Aupdated-desc) for reference.
-- type: textarea
-  attributes:
-    label: Motivation.
-    description: >
-      The motivation of the RFC.
-  validations:
-    required: true
-- type: textarea
-  attributes:
-    label: Proposed Change.
-    description: >
-      The proposed change of the RFC.
-  validations:
-    required: true
-- type: textarea
-  attributes:
-    label: Feedback Period.
-    description: >
-      The feedback period of the RFC. Usually at least one week.
-  validations:
-    required: false
-- type: textarea
-  attributes:
-    label: CC List.
-    description: >
-      The list of people you want to CC.
-  validations:
-    required: false
-- type: textarea
-  attributes:
-    label: Any Other Things.
-    description: >
-      Any other things you would like to mention.
-  validations:
-    required: false
-- type: markdown
-  attributes:
-    value: >
-      Thanks for contributing 🎉!
-- type: checkboxes
-  id: askllm
-  attributes:
-    label: Before submitting a new issue...
-    options:
-      - label: Make sure you already searched for relevant issues.
-        required: true

.github/workflows/docs.yml

@@ -8,14 +8,14 @@ on:
       - main
     paths:
       - "docs/**/*.md"
-      - "fastvideo/examples/**/*.py"
+      - "fastvideo/v1/examples/**/*.py"
   pull_request:
     branches: 
       - main
     types: [opened, ready_for_review, synchronize, reopened]
     paths:
       - "docs/**/*.md"
-      - "fastvideo/examples/**/*.py"
+      - "fastvideo/v1/examples/**/*.py"
 
   # Allows you to run this workflow manually from the Actions tab
   workflow_dispatch:

.github/workflows/matchers/mypy.json

@@ -13,4 +13,4 @@
       ]
     }
   ]
-}
+}

.github/workflows/pr-test.yml

@@ -115,37 +115,37 @@ jobs:
               - 'csrc/attn/config_vsa.py'
               - 'csrc/attn/vsa.cpp'
             vsa-paths: &vsa-paths
-              - 'fastvideo/**'
+              - 'fastvideo/v1/**'
               - *common-paths
               - *vsa-kernel-paths
             
             # Actual tests
             encoder-test:
-              - 'fastvideo/models/encoders/**'
-              - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/encoders/**'
+              - 'fastvideo/v1/models/encoders/**'
+              - 'fastvideo/v1/models/loader/**'
+              - 'fastvideo/v1/tests/encoders/**'
               - *common-paths
             vae-test:
-              - 'fastvideo/models/vaes/**'
-              - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/vaes/**'
+              - 'fastvideo/v1/models/vaes/**'
+              - 'fastvideo/v1/models/loader/**'
+              - 'fastvideo/v1/tests/vaes/**'
               - *common-paths
             transformer-test:
-              - 'fastvideo/models/dits/**'
-              - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/transformers/**'
-              - 'fastvideo/layers/**'
-              - 'fastvideo/attention/**'
+              - 'fastvideo/v1/models/dits/**'
+              - 'fastvideo/v1/models/loader/**'
+              - 'fastvideo/v1/tests/transformers/**'
+              - 'fastvideo/v1/layers/**'
+              - 'fastvideo/v1/attention/**'
               - *common-paths
             training-test:
-              - 'fastvideo/**'
+              - 'fastvideo/v1/**'
               - *common-paths
             training-test-VSA:
-              - 'fastvideo/**'
+              - 'fastvideo/v1/**'
               - *common-paths
               - *vsa-kernel-paths
             inference-test-STA:
-              - 'fastvideo/**'
+              - 'fastvideo/v1/**'
               - *common-paths
               - *sta-kernel-paths
             precision-test-STA:
@@ -167,7 +167,7 @@ jobs:
       gpu_count: 1
       volume_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/encoders -s"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/encoders -s"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -185,7 +185,7 @@ jobs:
       gpu_count: 1
       volume_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/vaes -s"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/vaes -s"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -203,7 +203,7 @@ jobs:
       gpu_count: 1
       volume_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/transformers -s"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/transformers -s"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -229,7 +229,7 @@ jobs:
       volume_size: 200
       disk_size: 200
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:${{ matrix.python-version.tag }}"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/ssim -vs"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/ssim -vs"
       timeout_minutes: 60
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -248,7 +248,7 @@ jobs:
       volume_size: 100
       disk_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/training/Vanilla -srP"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/v1/tests/training/Vanilla -srP"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -268,7 +268,7 @@ jobs:
       volume_size: 100
       disk_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/training/VSA -srP"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/v1/tests/training/VSA -srP"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -288,7 +288,7 @@ jobs:
       volume_size: 100
       disk_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/inference/STA -srP"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/v1/tests/inference/STA -srP"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -343,7 +343,7 @@ jobs:
       volume_size: 100
       disk_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/nightly/test_e2e_overfit_single_sample.py -vs"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/v1/tests/nightly/test_e2e_overfit_single_sample.py -vs"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -372,4 +372,4 @@ jobs:
           JOB_IDS: '["encoder-test", "vae-test", "transformer-test", "ssim-test-py3.10", "ssim-test-py3.11", "ssim-test-py3.12", "training-test", "training-test-VSA", "inference-test-STA", "precision-test-STA", "precision-test-VSA"]'
           RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
           GITHUB_RUN_ID: ${{ github.run_id }}
-        run: python .github/scripts/runpod_cleanup.py
+        run: python .github/scripts/runpod_cleanup.py

.github/workflows/publish-comfyui.yml

@@ -1,28 +0,0 @@
-name: Publish to Comfy registry
-on:
-  workflow_dispatch:
-  push:
-    branches:
-      - main
-      - master
-    paths:
-      - "pyproject.toml"
-
-permissions:
-  issues: write
-
-jobs:
-  publish-node:
-    name: Publish Custom Node to registry
-    runs-on: ubuntu-latest
-    if: ${{ github.repository_owner == 'hao-ai-lab' }}
-    steps:
-      - name: Check out code
-        uses: actions/checkout@v4
-        with:
-          submodules: true
-      - name: Publish Custom Node
-        uses: Comfy-Org/publish-node-action@v1
-        with:
-          ## Add your own personal access token to your Github Repository secrets and reference it here.
-          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}

.gitignore

@@ -20,7 +20,6 @@ samples/
 data/
 outputs/
 outputs_video
-checkpoints/
 sbatch.sh
 *.out
 env
@@ -41,7 +40,6 @@ eggs/
 docs/_build/
 docs/source/getting_started/examples/
 docs/source/inference/examples/
-docs/source/training/examples/
 
 # VSCode
 .vscode/
@@ -57,9 +55,12 @@ docs/source/training/examples/
 *.pkl
 
 # Reference videos
-!fastvideo/tests/ssim/reference_videos/**/*.mp4
+!fastvideo/v1/tests/ssim/reference_videos/**/*.mp4
 
 # Static images
 !docs/source/_static/images/**/*.png
+
+# Local scripts (keep local but don't track in git)
+local_scripts/
 !comfyui/assets/**/*.png
 !comfyui/assets/**/*.gif

.pre-commit-config.yaml

@@ -3,7 +3,7 @@ default_stages:
   - manual # Run in CI
 exclude: |
   (?x)(
-    fastvideo/third_party/.*|
+    fastvideo/v1/third_party/.*|
     csrc/.*|
     assets/.*|
     tests/.*|
@@ -60,7 +60,7 @@ repos:
   rev: v1.15.0
   hooks:
   - id: mypy
-    args: [--python-version, '3.10', --follow-imports, "skip", "--disable-error-code", "union-attr", "--disable-error-code", "override" ]
+    args: [--python-version, '3.10', --follow-imports, "skip", ]
     additional_dependencies: [types-cachetools, types-setuptools, types-PyYAML, types-requests]
 - repo: local
   hooks:
@@ -69,7 +69,7 @@ repos:
     entry: bash
     args:
       - -c
-      - 'git ls-files | grep -v "^fastvideo/tests/ssim/" | grep -v "^fastvideo/tests/inference/lora/L40S_reference_videos/" | grep " " && echo "Filenames should not contain spaces!" && exit 1 || exit 0'
+      - '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

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"

comfyui/README.md

@@ -74,8 +74,8 @@ After installation, the following nodes will be available in the ComfyUI interfa
 You may have noticed many arguments on the nodes have 'auto' as the default value. This is because FastVideo will automatically detect the best values for these parameters based on the model and the hardware. However, you can also manually configure these parameters to get the best performance for your specific use case. We plan on releasing more optimized workflow files for different models and hardware configurations in the future.
 
 You can see what some of the default configurations are by looking at the FastVideo repo:
-- [Wan2.1-I2V-14B-480P-Diffusers](https://github.com/hao-ai-lab/FastVideo/blob/main/fastvideo/configs/wan_14B_i2v_480p_pipeline.json)
-- [FastHunyuan-diffusers](https://github.com/hao-ai-lab/FastVideo/blob/main/fastvideo/configs/fasthunyuan_t2v.json)
+- [Wan2.1-I2V-14B-480P-Diffusers](https://github.com/hao-ai-lab/FastVideo/blob/main/fastvideo/v1/configs/wan_14B_i2v_480p_pipeline.json)
+- [FastHunyuan-diffusers](https://github.com/hao-ai-lab/FastVideo/blob/main/fastvideo/v1/configs/fasthunyuan_t2v.json)
 
 ### Node Configuration
 

comfyui/examples/FastHunyuan-diffusers.json

@@ -310,7 +310,7 @@
           "value": -99999,
           "cachedValue": "fp16"
         },
-        "dit_cpu_offload": {
+        "use_cpu_offload": {
           "isAuto": true,
           "value": -99999,
           "cachedValue": true

comfyui/examples/Wan2.1-I2V-14B-480P-Diffusers.json

@@ -517,7 +517,7 @@
           "value": "fp16",
           "cachedValue": "fp16"
         },
-        "dit_cpu_offload": {
+        "use_cpu_offload": {
           "isAuto": true,
           "value": true,
           "cachedValue": true

comfyui/video_generator/video_generator.py

@@ -105,7 +105,7 @@ class VideoGenerator:
                 "precision": (["fp16", "bf16"], {
                     "default": "fp16"
                 }),
-                "dit_cpu_offload": ([True, False], {
+                "use_cpu_offload": ([True, False], {
                     "default": False
                 }),
             }
@@ -204,7 +204,7 @@ class VideoGenerator:
         vae_config=None,
         text_encoder_config=None,
         dit_config=None,
-        dit_cpu_offload=None,
+        use_cpu_offload=None,
     ):
         print('Running FastVideo inference')
 
@@ -259,8 +259,8 @@ class VideoGenerator:
             raw_generation_args['tp_size'] = tp_size
         if sp_size is not None:
             raw_generation_args['sp_size'] = sp_size
-        if dit_cpu_offload is not None:
-            raw_generation_args['dit_cpu_offload'] = dit_cpu_offload
+        if use_cpu_offload is not None:
+            raw_generation_args['use_cpu_offload'] = use_cpu_offload
 
         generation_args = {
             k: v

comfyui/web/js/FastVideo.core.js

@@ -552,7 +552,7 @@ app.registerExtension({
                 ]
                 const floatWidgetNames = ["embedded_cfg_scale", "guidance_scale"]
                 const comboWidgetNames = ["vae_tiling", "vae_precision", "vae_sp", "text_encoder_precision", "precision",
-                    "load_encoder", "load_decoder", "use_tiling", "use_temporal_tiling", "use_parallel_tiling", "dit_cpu_offload", "enable_teacache"
+                    "load_encoder", "load_decoder", "use_tiling", "use_temporal_tiling", "use_parallel_tiling", "use_cpu_offload", "enable_teacache"
                 ]
                 const stringWidgetNames = ["prefix", "quant_config", "lora_config", "image_path"]
 

csrc/attn/README.md

@@ -1,23 +1,11 @@
 
 
-# Attention Kernel Used in FastVideo
+# Sliding Tile Atteniton Kernel
 
-## Sliding Tile Attention (STA)
-We only support H100 for STA.
-```bash
-git submodule update --init --recursive
-python setup_sta.py install
-```
 
-## Video Sparse Attention (VSA)
-We support H100 (via TK) and RTX 4090 (via triton) for VSA.
-```bash
-git submodule update --init --recursive
-python setup_vsa.py install
-```
-
-If you encounter error during installation, try below:
-Install C++20 for ThunderKittens:
+## Installation
+We test our code on Pytorch 2.5.0 and CUDA>=12.4. Currently we only support H100/H200, because ThunderKittens uses TMA but doesn't support Blackwell yet.
+First, install C++20 for ThunderKittens:
 ```bash
 sudo apt update
 sudo apt install gcc-11 g++-11
@@ -27,23 +15,27 @@ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave
 sudo apt update
 sudo apt install clang-11
 ```
-(If you use CUDA12.4)
+
+## Environment Setup
+First, set up your CUDA environment:
 ```bash
 export CUDA_HOME=/usr/local/cuda-12.4
 export PATH=${CUDA_HOME}/bin:${PATH} 
 export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
+git submodule update --init --recursive
 ```
 
+## Install Sliding Tile Attention (STA)
+```bash
+python setup_sta.py install
+```
 
+## Install Video Sparse Attention (VSA)
+```bash
+python setup_vsa.py install
+```
 
 ## Usage
-### STA
-End-2-end inference with FastVideo:
-```bash
-bash scripts/inference/v1_inference_wan_STA.sh
-```
-
-If you want to use sliding tile attention in your custom model:
 ```python
 from st_attn import sliding_tile_attention
 # assuming video size (T, H, W) = (30, 48, 80), text tokens = 256 with padding. 
@@ -55,10 +47,8 @@ from st_attn import sliding_tile_attention
 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)
-```
 
-### VSA
-We do not officially supoort end-2-end inference with VSA in FastVideo yet. Stay tuned.
+```
 
 
 ## Test

csrc/attn/benchmarks/bench_vsa.py

@@ -1,7 +1,7 @@
 import torch
 import argparse
 from flash_attn.utils.benchmark import benchmark_forward
-from vsa import block_sparse_fwd, block_sparse_bwd
+from vsa import block_sparse_attention_fwd, block_sparse_attention_backward
 from vsa import BLOCK_M, BLOCK_N
 
 import numpy as np
@@ -130,12 +130,12 @@ def benchmark_block_sparse_attention(q, k, v, q2k_block_sparse_index, q2k_block_
     
     # Forward pass
     # Warm-up run
-    o, l_vec = block_sparse_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)
+    o, l_vec = block_sparse_attention_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)
     torch.cuda.synchronize()
     
     # Benchmark forward
     _, fwd_time = benchmark_forward(
-        block_sparse_fwd,
+        block_sparse_attention_fwd,
         q, k, v, q2k_block_sparse_index, q2k_block_sparse_num,
         repeats=20,
         verbose=False,
@@ -150,12 +150,12 @@ def benchmark_block_sparse_attention(q, k, v, q2k_block_sparse_index, q2k_block_
     
     # Warm-up runs
     for _ in range(5):
-        block_sparse_bwd(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num)
+        block_sparse_attention_backward(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num)
     torch.cuda.synchronize()
     
     # Benchmark backward
     _, bwd_time = benchmark_forward(
-        block_sparse_bwd,
+        block_sparse_attention_backward,
         q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num,
         repeats=20,
         verbose=False,

csrc/attn/benchmarks/bench_vsa_triton.py

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

csrc/attn/scratch.sh

@@ -1,4 +0,0 @@
-    off_hz = tl.program_id(2)
-    b = off_hz // H
-    h = off_hz % H
-    meta_base = ((b * H + h) * q_tiles + q_blk)

csrc/attn/setup_vsa.py

@@ -1,7 +1,7 @@
 import os
 import subprocess
 
-from config_vsa import kernels, sources, target
+from csrc.attn.config_vsa import kernels, sources, target
 from setuptools import find_packages, setup
 from torch.utils.cpp_extension import BuildExtension, CUDAExtension
 
@@ -51,29 +51,21 @@ for k in kernels:
         source_files.append(sources[k]['source_files'][target])
     cpp_flags.append(f'-DTK_COMPILE_{k.replace(" ", "_").upper()}')
 
-ext_modules = []
-import torch
-major, minor = torch.cuda.get_device_capability(0)
-if major == 9 and minor == 0:# check if H100
-    ext_modules = [
-        CUDAExtension('vsa_cuda',
-            sources=source_files,
-            extra_compile_args={
-                'cxx': cpp_flags,
-                'nvcc': cuda_flags
-            },
-            libraries=['cuda'])
-    ]
-
-
-
 setup(name=PACKAGE_NAME,
       version=VERSION,
       author=AUTHOR,
       description=DESCRIPTION,
       url=URL,
       packages=find_packages(),
-      ext_modules=ext_modules,
+      ext_modules=[
+          CUDAExtension('vsa_cuda',
+                        sources=source_files,
+                        extra_compile_args={
+                            'cxx': cpp_flags,
+                            'nvcc': cuda_flags
+                        },
+                        libraries=['cuda'])
+      ],
       cmdclass={'build_ext': BuildExtension},
       classifiers=[
           "Programming Language :: Python :: 3",

csrc/attn/tests/test_block_sparse.py

@@ -2,7 +2,7 @@ import torch
 import argparse
 from flash_attn.utils.benchmark import benchmark_forward
 from flash_attn import flash_attn_func
-from vsa import block_sparse_attn
+from vsa import block_sparse_attention_fwd, block_sparse_attention_backward, BlockSparseAttentionFunction
 from vsa import BLOCK_M, BLOCK_N
 
 import numpy as np
@@ -188,7 +188,7 @@ def main(args):
 
 
             # testing forward
-            o = block_sparse_attn(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num)
+            o = BlockSparseAttentionFunction.apply(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num)
             del q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask, block_mask_expanded
             grad_o = torch.randn_like(o)
             o.backward(grad_o)
@@ -215,7 +215,7 @@ def main(args):
             forward_metrics['l1'].append(l1)
             forward_metrics['rmse'].append(rmse)
 
-            print(f"block_sparse_fwd vs torch.nn.functional.scaled_dot_product_attention:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
+            print(f"block_sparse_attention_fwd vs torch.nn.functional.scaled_dot_product_attention:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
 
             # test backward
             o_sdpa.backward(grad_o)
@@ -228,7 +228,7 @@ def main(args):
             grad_q_metrics['sim'].append(sim)
             grad_q_metrics['l1'].append(l1)
             grad_q_metrics['rmse'].append(rmse)
-            print(f"block_sparse_bwd vs torch.nn.functional.scaled_dot_product_attention grad_q:\nsim: {sim}, l1: {l1}, rmse: {rmse}")        
+            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_q:\nsim: {sim}, l1: {l1}, rmse: {rmse}")        
             
             sim, l1, rmse = precision_metric(k.grad, k_sdpa.grad)
             assert sim > 0.9999, f"SSIM too low: {sim}"
@@ -237,7 +237,7 @@ def main(args):
             grad_k_metrics['sim'].append(sim)
             grad_k_metrics['l1'].append(l1)
             grad_k_metrics['rmse'].append(rmse)
-            print(f"block_sparse_bwd vs torch.nn.functional.scaled_dot_product_attention grad_k:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
+            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_k:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
             
             sim, l1, rmse = precision_metric(v.grad, v_sdpa.grad)
             assert sim > 0.9999, f"SSIM too low: {sim}"
@@ -246,7 +246,7 @@ def main(args):
             grad_v_metrics['sim'].append(sim)
             grad_v_metrics['l1'].append(l1)
             grad_v_metrics['rmse'].append(rmse)
-            print(f"block_sparse_bwd vs torch.nn.functional.scaled_dot_product_attention grad_v:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
+            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_v:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
             
             del o, o_sdpa, grad_o, q_sdpa, k_sdpa, v_sdpa
             gc.collect()
@@ -279,11 +279,11 @@ def main(args):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description='Benchmark Block Sparse Attention')
-    parser.add_argument('--batch_size', type=int, default=4, help='Batch size')
-    parser.add_argument('--num_heads', type=int, default=6, help='Number of heads')
-    parser.add_argument('--head_dim', type=int, default=128, help='Head dimension')
-    parser.add_argument('--topk', type=int, default=64, help='Number of kv blocks each q block attends to')
+    parser.add_argument('--batch_size', type=int, default=2, help='Batch size')
+    parser.add_argument('--num_heads', type=int, default=12, help='Number of heads')
+    parser.add_argument('--head_dim', type=int, default=64, help='Head dimension')
+    parser.add_argument('--topk', type=int, default=32, help='Number of kv blocks each q block attends to')
     parser.add_argument('--seq_lengths', type=int, nargs='+', default=[29120], help='Sequence lengths to benchmark')
-    parser.add_argument('--num_iterations', type=int, default=50, help='Number of test iterations to run')
+    parser.add_argument('--num_iterations', type=int, default=100, help='Number of test iterations to run')
     args = parser.parse_args()
     main(args)

csrc/attn/tests/test_block_sparse_triton.py

@@ -1,289 +0,0 @@
-import torch
-import argparse
-from flash_attn.utils.benchmark import benchmark_forward
-from flash_attn import flash_attn_func
-from vsa import triton_attention_sparse
-from vsa import BLOCK_M, BLOCK_N
-
-import numpy as np
-import random
-import gc
-
-def set_seed(seed: int = 42):
-    # Python random module
-    random.seed(seed)
-
-    # NumPy
-    np.random.seed(seed)
-
-    # PyTorch
-    torch.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)  # if using multi-GPU
-
-
-@torch.no_grad
-def precision_metric(quant_o, fa2_o): 
-    x, xx = quant_o.float(), fa2_o.float() 
-    sim = torch.nn.functional.cosine_similarity(x.reshape(1, -1), xx.reshape(1, -1)).item()
-    l1 =   ((x - xx).abs().sum() / xx.abs().sum() ).item()
-    rmse = torch.sqrt(torch.mean((x -xx) ** 2)).item()
-
-    return sim, l1, rmse
-
-def create_input_tensors(batch, head, seq_len, headdim):
-    """Create random input tensors for attention."""
-    q = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    k = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    v = torch.randn(batch, head, seq_len, headdim, dtype=torch.bfloat16, device="cuda")
-    return q, k, v
-
-def generate_block_sparse_pattern(bs, h, num_q_blocks, num_kv_blocks, k, device="cuda"):
-    """
-    Generate a block sparse pattern where each q block attends to exactly k kv blocks.
-    
-    Args:
-        bs: batch size
-        h: number of heads
-        num_q_blocks: number of query blocks
-        num_kv_blocks: number of key-value blocks
-        k: number of kv blocks each q block attends to
-        device: device to create tensors on
-        
-    Returns:
-        q2k_block_sparse_index: [bs, h, num_q_blocks, k]
-            Contains the indices of kv blocks that each q block attends to.
-        q2k_block_sparse_num: [bs, h, num_q_blocks]
-            Contains the number of kv blocks that each q block attends to (all equal to k).
-        k2q_block_sparse_index: [bs, h, num_kv_blocks, num_q_blocks]
-            Contains the indices of q blocks that attend to each kv block.
-        k2q_block_sparse_num: [bs, h, num_kv_blocks]
-            Contains the number of q blocks that attend to each kv block.
-        block_sparse_mask: [bs, h, num_q_blocks, num_kv_blocks]
-            Binary mask where 1 indicates attention connection.
-    """
-    # Ensure k is not larger than num_kv_blocks
-    k = min(k, num_kv_blocks)
-    
-    # Create random scores for sampling
-    scores = torch.rand(bs, h, num_q_blocks, num_kv_blocks, device=device)
-    
-    # Get top-k indices for each q block
-    _, q2k_block_sparse_index = torch.topk(scores, k, dim=-1)
-    q2k_block_sparse_index = q2k_block_sparse_index.to(torch.int32)
-
-    # sort q2k_block_sparse_index
-    q2k_block_sparse_index, _ = torch.sort(q2k_block_sparse_index, dim=-1)
-
-    # All q blocks attend to exactly k kv blocks
-    q2k_block_sparse_num = torch.full((bs, h, num_q_blocks), k, dtype=torch.int32, device=device)
-    
-    # Create the corresponding mask
-    block_sparse_mask = torch.zeros(bs, h, num_q_blocks, num_kv_blocks, dtype=torch.bool, device=device)
-    
-    # Fill in the mask based on the indices
-    for b in range(bs):
-        for head in range(h):
-            for q_idx in range(num_q_blocks):
-                kv_indices = q2k_block_sparse_index[b, head, q_idx]
-                block_sparse_mask[b, head, q_idx, kv_indices] = True
-    
-    # Create the reverse mapping (k2q)
-    # First, initialize lists to collect q indices for each kv block
-    k2q_indices_list = [[[] for _ in range(num_kv_blocks)] for _ in range(bs * h)]
-    
-    # Populate the lists based on q2k mapping
-    for b in range(bs):
-        for head in range(h):
-            flat_idx = b * h + head
-            for q_idx in range(num_q_blocks):
-                kv_indices = q2k_block_sparse_index[b, head, q_idx].tolist()
-                for kv_idx in kv_indices:
-                    k2q_indices_list[flat_idx][kv_idx].append(q_idx)
-    
-    # Find the maximum number of q blocks that attend to any kv block
-    max_q_per_kv = 0
-    for flat_idx in range(bs * h):
-        for kv_idx in range(num_kv_blocks):
-            max_q_per_kv = max(max_q_per_kv, len(k2q_indices_list[flat_idx][kv_idx]))
-    
-    # Create tensors for k2q mapping
-    k2q_block_sparse_index = torch.full((bs, h, num_kv_blocks, max_q_per_kv), -1, 
-                                        dtype=torch.int32, device=device)
-    k2q_block_sparse_num = torch.zeros((bs, h, num_kv_blocks), 
-                                       dtype=torch.int32, device=device)
-    
-    # Fill the tensors
-    for b in range(bs):
-        for head in range(h):
-            flat_idx = b * h + head
-            for kv_idx in range(num_kv_blocks):
-                q_indices = k2q_indices_list[flat_idx][kv_idx]
-                num_q = len(q_indices)
-                k2q_block_sparse_num[b, head, kv_idx] = num_q
-                if num_q > 0:
-                    k2q_block_sparse_index[b, head, kv_idx, :num_q] = torch.tensor(
-                        q_indices, dtype=torch.int32, device=device)
-                
-    return q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask
-
-def main(args):
-    set_seed(42)
-    
-    # Extract parameters
-    batch = args.batch_size
-    head = args.num_heads
-    headdim = args.head_dim
-    num_iterations = args.num_iterations
-    
-    print(f"Block Sparse Attention Benchmark")
-    print(f"batch: {batch}, head: {head}, headdim: {headdim}, iterations: {num_iterations}")
-    
-    # Test with different sequence lengths
-    for seq_len in args.seq_lengths:
-        # Skip very long sequences if they might cause OOM
-        # if seq_len > 16384 and batch > 1:
-        #     continue
-            
-        print("="*100)
-        print(f"\nSequence length: {seq_len}")
-        
-        # Collect metrics across iterations
-        forward_metrics = {'sim': [], 'l1': [], 'rmse': []}
-        grad_q_metrics = {'sim': [], 'l1': [], 'rmse': []}
-        grad_k_metrics = {'sim': [], 'l1': [], 'rmse': []}
-        grad_v_metrics = {'sim': [], 'l1': [], 'rmse': []}
-        
-        for iter_idx in range(num_iterations):
-            if num_iterations > 1:
-                print(f"\nIteration {iter_idx+1}/{num_iterations}")
-            
-            # Create input tensors
-            q, k, v = create_input_tensors(batch, head, seq_len, headdim)
-            
-            # Setup block sparse parameters
-            num_q_blocks = seq_len // BLOCK_M
-            num_kv_blocks = seq_len // BLOCK_N
-            
-            # Determine k value (number of kv blocks per q block)
-            topk = args.topk
-            if topk is None:
-                topk = num_kv_blocks // 10  # Default to ~90% sparsity if k is not specified
-            topk = max(1, topk)
-            if iter_idx == 0:  # Only print this once
-                print(f"Using topk={topk} kv blocks per q block (out of {num_kv_blocks} total kv blocks)")
-
-            # Generate block sparse pattern
-            q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask = generate_block_sparse_pattern(
-                batch, head, num_q_blocks, num_kv_blocks, topk, device="cuda")
-
-            # expand block_sparse_mask to full mask
-            block_mask_expanded = block_sparse_mask.unsqueeze(-1).unsqueeze(-2)  # [b, h, num_q_blocks, num_kv_blocks, 1, 1]
-            block_mask_expanded = block_mask_expanded.expand(-1, -1, -1, -1, BLOCK_M, BLOCK_N)  # [b, h, num_q_blocks, num_kv_blocks, BLOCK_M, BLOCK_N]
-            full_mask = block_mask_expanded.permute(0, 1, 2, 4, 3, 5).reshape(batch, head, seq_len, seq_len)
-
-            q.requires_grad = True
-            k.requires_grad = True
-            v.requires_grad = True
-
-
-            # testing forward
-            o = triton_attention_sparse(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num)
-            del q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num, block_sparse_mask, block_mask_expanded
-            grad_o = torch.randn_like(o)
-            o.backward(grad_o)
-            # clear memory
-            q_sdpa = q.detach().clone()
-            k_sdpa = k.detach().clone()
-            v_sdpa = v.detach().clone()
-            q_sdpa.requires_grad = True
-            k_sdpa.requires_grad = True
-            v_sdpa.requires_grad = True
-            q.data = torch.empty(0, device=q.device)
-            k.data = torch.empty(0, device=k.device)
-            v.data = torch.empty(0, device=v.device)
-            torch.cuda.empty_cache()
-
-            o_sdpa = torch.nn.functional.scaled_dot_product_attention(q_sdpa, k_sdpa, v_sdpa, attn_mask=full_mask)
-
-
-            sim, l1, rmse = precision_metric(o, o_sdpa)
-            assert sim > 0.9999, f"SSIM too low: {sim}"
-            assert l1 < 8e-5, f"l1 too large: {l1}"
-            assert rmse < 5e-5, f"RMSE too large: {rmse}"
-            forward_metrics['sim'].append(sim)
-            forward_metrics['l1'].append(l1)
-            forward_metrics['rmse'].append(rmse)
-
-            print(f"block_sparse_attention_fwd vs torch.nn.functional.scaled_dot_product_attention:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
-
-            # test backward
-            o_sdpa.backward(grad_o)
-
-            sim, l1, rmse = precision_metric(q.grad, q_sdpa.grad)
-            # Error bounds collected on H100
-            assert sim > 0.9999, f"SSIM too low: {sim}"
-            assert l1 < 4e-3, f"l1 too large: {l1}"
-            assert rmse < 5e-4, f"RMSE too large: {rmse}"
-            grad_q_metrics['sim'].append(sim)
-            grad_q_metrics['l1'].append(l1)
-            grad_q_metrics['rmse'].append(rmse)
-            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_q:\nsim: {sim}, l1: {l1}, rmse: {rmse}")        
-            
-            sim, l1, rmse = precision_metric(k.grad, k_sdpa.grad)
-            assert sim > 0.9999, f"SSIM too low: {sim}"
-            assert l1 < 4e-3, f"l1 too large: {l1}"
-            assert rmse < 5e-4, f"RMSE too large: {rmse}"
-            grad_k_metrics['sim'].append(sim)
-            grad_k_metrics['l1'].append(l1)
-            grad_k_metrics['rmse'].append(rmse)
-            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_k:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
-            
-            sim, l1, rmse = precision_metric(v.grad, v_sdpa.grad)
-            assert sim > 0.9999, f"SSIM too low: {sim}"
-            assert l1 < 4e-3, f"l1 too large: {l1}"
-            assert rmse < 5e-4, f"RMSE too large: {rmse}"
-            grad_v_metrics['sim'].append(sim)
-            grad_v_metrics['l1'].append(l1)
-            grad_v_metrics['rmse'].append(rmse)
-            print(f"block_sparse_attention_bwd vs torch.nn.functional.scaled_dot_product_attention grad_v:\nsim: {sim}, l1: {l1}, rmse: {rmse}")
-            
-            del o, o_sdpa, grad_o, q_sdpa, k_sdpa, v_sdpa
-            gc.collect()
-            torch.cuda.empty_cache()
-
-        # Print summary statistics if multiple iterations were run
-        if num_iterations > 1:
-            print("\n" + "="*50)
-            print(f"Summary Statistics (over {num_iterations} iterations):")
-            
-            print("\nForward metrics:")
-            print(f"Similarity: mean={np.mean(forward_metrics['sim']):.6f}, std={np.std(forward_metrics['sim']):.6f}, min={np.min(forward_metrics['sim']):.6f}")
-            print(f"L1 error: mean={np.mean(forward_metrics['l1']):.6f}, std={np.std(forward_metrics['l1']):.6f}, max={np.max(forward_metrics['l1']):.6f}")
-            print(f"RMSE: mean={np.mean(forward_metrics['rmse']):.6f}, std={np.std(forward_metrics['rmse']):.6f}, max={np.max(forward_metrics['rmse']):.6f}")
-            
-            print("\nGradient Q metrics:")
-            print(f"Similarity: mean={np.mean(grad_q_metrics['sim']):.6f}, std={np.std(grad_q_metrics['sim']):.6f}, min={np.min(grad_q_metrics['sim']):.6f}")
-            print(f"L1 error: mean={np.mean(grad_q_metrics['l1']):.6f}, std={np.std(grad_q_metrics['l1']):.6f}, max={np.max(grad_q_metrics['l1']):.6f}")
-            print(f"RMSE: mean={np.mean(grad_q_metrics['rmse']):.6f}, std={np.std(grad_q_metrics['rmse']):.6f}, max={np.max(grad_q_metrics['rmse']):.6f}")
-            
-            print("\nGradient K metrics:")
-            print(f"Similarity: mean={np.mean(grad_k_metrics['sim']):.6f}, std={np.std(grad_k_metrics['sim']):.6f}, min={np.min(grad_k_metrics['sim']):.6f}")
-            print(f"L1 error: mean={np.mean(grad_k_metrics['l1']):.6f}, std={np.std(grad_k_metrics['l1']):.6f}, max={np.max(grad_k_metrics['l1']):.6f}")
-            print(f"RMSE: mean={np.mean(grad_k_metrics['rmse']):.6f}, std={np.std(grad_k_metrics['rmse']):.6f}, max={np.max(grad_k_metrics['rmse']):.6f}")
-            
-            print("\nGradient V metrics:")
-            print(f"Similarity: mean={np.mean(grad_v_metrics['sim']):.6f}, std={np.std(grad_v_metrics['sim']):.6f}, min={np.min(grad_v_metrics['sim']):.6f}")
-            print(f"L1 error: mean={np.mean(grad_v_metrics['l1']):.6f}, std={np.std(grad_v_metrics['l1']):.6f}, max={np.max(grad_v_metrics['l1']):.6f}")
-            print(f"RMSE: mean={np.mean(grad_v_metrics['rmse']):.6f}, std={np.std(grad_v_metrics['rmse']):.6f}, max={np.max(grad_v_metrics['rmse']):.6f}")
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description='Benchmark Block Sparse Attention')
-    parser.add_argument('--batch_size', type=int, default=4, help='Batch size')
-    parser.add_argument('--num_heads', type=int, default=6, help='Number of heads')
-    parser.add_argument('--head_dim', type=int, default=128, help='Head dimension')
-    parser.add_argument('--topk', type=int, default=4, help='Number of kv blocks each q block attends to')
-    parser.add_argument('--seq_lengths', type=int, nargs='+', default=[4096], help='Sequence lengths to benchmark')
-    parser.add_argument('--num_iterations', type=int, default=10, help='Number of test iterations to run')
-    args = parser.parse_args()
-    main(args)

csrc/attn/tests/test_fa2.py

@@ -3,15 +3,10 @@ from tqdm import tqdm
 import matplotlib.pyplot as plt
 import numpy as np
 
-from vsa import triton_attention
-
 def pytorch_test(Q, K, V, dO):
     q_ = Q.to(torch.float64).requires_grad_()
     k_ = K.to(torch.float64).requires_grad_()
     v_ = V.to(torch.float64).requires_grad_()
-    q_.grad = None
-    k_.grad = None
-    v_.grad = None
     dO_ = dO.to(torch.float64)
     
     # manual pytorch implementation of scaled dot product attention
@@ -35,33 +30,11 @@ def fa2_test(Q, K, V, dO):
     Q.requires_grad = True
     K.requires_grad = True
     V.requires_grad = True
-    Q.grad = None
-    K.grad = None
-    V.grad = None
-    
     output = torch.nn.functional.scaled_dot_product_attention(Q, K, V, is_causal=False)
     output.backward(dO)
     
     return output, Q.grad, K.grad, V.grad
 
-def triton_test(Q, K, V, dO):
-    Q.requires_grad = True
-    K.requires_grad = True
-    V.requires_grad = True
-    Q.grad = None
-    K.grad = None
-    V.grad = None
-
-    output = triton_attention(Q, K, V)
-
-    output.backward(dO)
-    
-    q_grad = Q.grad
-    k_grad = K.grad
-    v_grad = V.grad
-    
-    return output.to(Q.dtype) if output is not None else None, q_grad, k_grad, v_grad
-
 def generate_tensor(shape, mean, std, dtype, device):
     tensor = torch.randn(shape, dtype=dtype, device=device)
 
@@ -72,8 +45,7 @@ def generate_tensor(shape, mean, std, dtype, device):
 
 def check_correctness(b, h, n, d, mean, std, num_iterations=100, error_mode='all', test_mode='forward_backward'):
     results = {
-        'FA2 vs PT': {'sum_diff': 0.0, 'sum_abs': 0.0, 'max_diff': 0.0},
-        'Triton vs PT': {'sum_diff': 0.0, 'sum_abs': 0.0, 'max_diff': 0.0},
+        'FA2 vs PT': {'sum_diff': 0, 'sum_abs': 0, 'max_diff': 0},
     }
 
     for _ in range(num_iterations):
@@ -86,31 +58,21 @@ def check_correctness(b, h, n, d, mean, std, num_iterations=100, error_mode='all
         
         pt_o, pt_qg, pt_kg, pt_vg = pytorch_test(Q, K, V, dO)
         fa2_o, fa2_qg, fa2_kg, fa2_vg = fa2_test(Q, K, V, dO)
-        triton_o, triton_qg, triton_kg, triton_vg = triton_test(Q, K, V, dO)
         
         if test_mode == 'forward_only':
             tensors_fa2_pt = [(pt_o, fa2_o)]
-            tensors_triton_pt = [(pt_o, triton_o)]
         else:  # 'forward_backward'
             if error_mode == 'output':
                 tensors_fa2_pt = [(pt_o, fa2_o)]
-                tensors_triton_pt = [(pt_o, triton_o)]
             elif error_mode == 'backward':
                 tensors_fa2_pt = [(pt_qg, fa2_qg),
                                (pt_kg, fa2_kg),
                                (pt_vg, fa2_vg)]
-                tensors_triton_pt = [(pt_qg, triton_qg),
-                                 (pt_kg, triton_kg),
-                                 (pt_vg, triton_vg)]
             else:  # 'all'
                 tensors_fa2_pt = [(pt_o, fa2_o),
                                (pt_qg, fa2_qg),
                                (pt_kg, fa2_kg),
                                (pt_vg, fa2_vg)]
-                tensors_triton_pt = [(pt_o, triton_o),
-                                 (pt_qg, triton_qg),
-                                 (pt_kg, triton_kg),
-                                 (pt_vg, triton_vg)]
                 
         for pt, fa2 in tensors_fa2_pt:
             diff = pt - fa2
@@ -118,13 +80,6 @@ def check_correctness(b, h, n, d, mean, std, num_iterations=100, error_mode='all
             results['FA2 vs PT']['sum_diff'] += torch.sum(abs_diff).item()
             results['FA2 vs PT']['sum_abs'] += torch.sum(torch.abs(pt)).item()
             results['FA2 vs PT']['max_diff'] = max(results['FA2 vs PT']['max_diff'], torch.max(abs_diff).item())
-        
-        for pt, triton in tensors_triton_pt:
-            diff = pt - triton
-            abs_diff = torch.abs(diff)
-            results['Triton vs PT']['sum_diff'] += torch.sum(abs_diff).item()
-            results['Triton vs PT']['sum_abs'] += torch.sum(torch.abs(pt)).item()
-            results['Triton vs PT']['max_diff'] = max(results['Triton vs PT']['max_diff'], torch.max(abs_diff).item())
                 
         torch.cuda.empty_cache()
 
@@ -144,40 +99,38 @@ def check_correctness(b, h, n, d, mean, std, num_iterations=100, error_mode='all
 def generate_error_tables(b, h, d, mean, std, error_mode='all', test_mode='forward_backward'):
     seq_lengths = [768 * (2**i) for i in range(1)]
 
-    print(f"\n{'='*100}")
+    print(f"\n{'='*80}")
     print(f"ATTENTION ERROR COMPARISON TABLE (b={b}, h={h}, d={d}, mean={mean}, std={std})")
     print(f"Mode: {error_mode}, Test: {test_mode}")
-    print(f"{'='*100}")
+    print(f"{'='*80}")
     
     # Print header
-    print(f"{'Seq Length':<12} | {'FA2 vs PT Avg':<15} | {'FA2 vs PT Max':<15} | {'Triton vs PT Avg':<15} | {'Triton vs PT Max':<15}")
-    print(f"{'-'*12} | {'-'*15} | {'-'*15} | {'-'*15} | {'-'*15}")
+    print(f"{'Seq Length':<12} | {'FA2 vs PT Avg':<15} | {'FA2 vs PT Max':<15}")
+    print(f"{'-'*12} | {'-'*15} | {'-'*15}")
     
     for n in seq_lengths:
         results = check_correctness(b, h, n, d, mean, std, error_mode=error_mode, test_mode=test_mode)
         
         fa2_pt_avg = results['FA2 vs PT']['avg_diff']
         fa2_pt_max = results['FA2 vs PT']['max_diff']
-        triton_pt_avg = results['Triton vs PT']['avg_diff']
-        triton_pt_max = results['Triton vs PT']['max_diff']
         
-        # Print row with both comparisons
-        print(f"{n:<12} | {fa2_pt_avg:<15.6e} | {fa2_pt_max:<15.6e} | {triton_pt_avg:<15.6e} | {triton_pt_max:<15.6e}")
+        # Print row
+        print(f"{n:<12} | {fa2_pt_avg:<15.6e} | {fa2_pt_max:<15.6e}")
     
-    print(f"{'='*100}\n")
+    print(f"{'='*80}\n")
 
 # fix random seed
 torch.manual_seed(0)
 
+# Example usage
+b, h, d = 2, 2, 64
 mean = 1e-1
 std = 10
-configs = [
-    (4, 1, 128),  # Larger batch, single head, larger dim
-    (2, 8, 64),   # Medium batch, many heads, medium dim
-]
 
-for b, h, d in configs:
-    print(f"\nConfiguration: batch={b}, heads={h}, dim={d}")
-    generate_error_tables(b, h, d, mean, std, error_mode='backward', test_mode='forward_backward')
+# Test forward only
+generate_error_tables(b, h, d, mean, std, error_mode='output', test_mode='forward_only')
+
+# Test forward and backward
+generate_error_tables(b, h, d, mean, std, error_mode='all', test_mode='forward_backward')
 
 print("Attention error comparison completed.")

csrc/attn/vsa/__init__.py

@@ -1,27 +1,18 @@
+import math
+
 import torch
+from torch.utils.checkpoint import detach_variable
 from typing import Tuple
-from vsa.vsa import block_sparse_attn
 try:
     from vsa_cuda import block_sparse_fwd, block_sparse_bwd
 except ImportError:
     block_sparse_fwd = None
     block_sparse_bwd = None
-from vsa.block_sparse_attn_triton import attention as triton_attention, attention_sparse as triton_attention_sparse
+
+
 BLOCK_M = 64
 BLOCK_N = 64
 
-
-def torch_attention(q, k, v) -> Tuple[torch.Tensor, torch.Tensor]:
-    QK = torch.matmul(q, k.transpose(-2, -1))
-    QK /= (q.size(-1)**0.5)
-
-    # Causal mask removed since causal is always false
-
-    QK = torch.nn.functional.softmax(QK, dim=-1)
-    output = torch.matmul(QK, v)
-    return output, QK
-
-
 def video_sparse_attn(q, k, v, topk, block_size, compress_attn_weight=None):
     """
     q: [batch_size, num_heads, seq_len, head_dim]
@@ -76,6 +67,15 @@ def video_sparse_attn(q, k, v, topk, block_size, compress_attn_weight=None):
         final_output = output_compress + output_select
     return final_output
 
+def torch_attention(q, k, v) -> Tuple[torch.Tensor, torch.Tensor]:
+    QK = torch.matmul(q, k.transpose(-2, -1))
+    QK /= (q.size(-1)**0.5)
+
+    # Causal mask removed since causal is always false
+
+    QK = torch.nn.functional.softmax(QK, dim=-1)
+    output = torch.matmul(QK, v)
+    return output, QK
 
 def generate_topk_block_sparse_pattern(block_attn_score: torch.Tensor,
                                        topk: int):
@@ -123,11 +123,107 @@ def generate_topk_block_sparse_pattern(block_attn_score: torch.Tensor,
 
     return q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num
 
+@torch._dynamo.disable
+def block_sparse_attn(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num):
+    """
+    Differentiable block sparse attention function.
+    
+    Args:
+        q: Query tensor [batch_size, num_heads, seq_len_q, head_dim]
+        k: Key tensor [batch_size, num_heads, seq_len_kv, head_dim]
+        v: Value tensor [batch_size, num_heads, seq_len_kv, head_dim]
+        q2k_block_sparse_index: Indices for query-to-key sparse blocks
+        q2k_block_sparse_num: Number of sparse blocks for each query block
+        k2q_block_sparse_index: Indices for key-to-query sparse blocks (for backward pass)
+        k2q_block_sparse_num: Number of sparse blocks for each key block (for backward pass)
+    
+    Returns:
+        output: Attention output tensor [batch_size, num_heads, seq_len_q, head_dim]
+    """
+    return BlockSparseAttentionFunction.apply(
+        q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num
+    )
+
+def block_sparse_attention_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num):
+    """
+    block_sparse_mask: [bs, h, num_q_blocks, num_kv_blocks]. 
+        [*, *, i, j] = 1 means the i-th q block should attend to the j-th kv block.
+    """
+    # assert all elements in q2k_block_sparse_num can be devisible by 2
+    o, lse = block_sparse_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)
+    return o, lse
+
+def block_sparse_attention_backward(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num):
+    grad_output = grad_output.contiguous()
+    grad_q, grad_k, grad_v = block_sparse_bwd(q, k, v, o, l_vec, grad_output, k2q_block_sparse_index, k2q_block_sparse_num)
+    return grad_q, grad_k, grad_v
 
 ## pytorch sdpa version of block sparse ##
 import triton
 import triton.language as tl
 
+@triton.jit
+def index_to_mask_kernel(
+    q2k_block_sparse_index_ptr,
+    q2k_block_sparse_num_ptr,
+    mask_ptr,
+    batch_size: tl.constexpr,
+    num_heads: tl.constexpr,
+    num_q_blocks: tl.constexpr,
+    num_k_blocks: tl.constexpr,
+    max_kv_blocks: tl.constexpr,
+    BLOCK_Q: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+):
+    bh, q, id = tl.program_id(0).to(tl.int64), tl.program_id(1).to(tl.int64), tl.program_id(2).to(tl.int64)
+    b = bh // num_heads
+    h = bh % num_heads
+
+    num_valid_blocks = tl.load(q2k_block_sparse_num_ptr + b * num_heads * num_q_blocks + h * num_q_blocks + q)
+
+    if num_valid_blocks <= id:
+        return
+    k = tl.load(q2k_block_sparse_index_ptr + b * num_heads * num_q_blocks * max_kv_blocks + h * num_q_blocks * max_kv_blocks + q * max_kv_blocks + id)
+
+    full_mask = (tl.arange(0, BLOCK_Q)[:, None] < BLOCK_Q) & (tl.arange(0, BLOCK_K)[None, :] < BLOCK_K)
+
+    q_lengths = num_q_blocks * BLOCK_Q
+    k_lengths = num_k_blocks * BLOCK_K
+    mask_ptr_base = mask_ptr + b * num_heads * q_lengths * k_lengths + h * q_lengths * k_lengths + q * BLOCK_Q * k_lengths + k * BLOCK_K
+
+    tl.store(mask_ptr_base + tl.arange(0, BLOCK_Q)[:, None] * k_lengths + tl.arange(0, BLOCK_K)[None, :], full_mask)
+
+def index_to_mask(q2k_block_sparse_index, q2k_block_sparse_num, BLOCK_Q, BLOCK_K, num_k_blocks):
+    """
+    Convert block sparse indices to a mask.
+    
+    Args:
+        q2k_block_sparse_index: Indices for query-to-key sparse blocks
+        q2k_block_sparse_num: Number of sparse blocks for each query block
+    
+    Returns:
+        mask: Block sparse mask tensor
+    """
+    batch_size, num_heads, num_q_blocks, max_kv_blocks = q2k_block_sparse_index.shape
+    assert q2k_block_sparse_num.shape == (batch_size, num_heads, num_q_blocks)
+
+    mask = torch.zeros((batch_size, num_heads, num_q_blocks * BLOCK_Q, num_k_blocks * BLOCK_K), dtype=torch.bool, device=q2k_block_sparse_index.device)
+
+    grid = (batch_size * num_heads, num_q_blocks, max_kv_blocks)
+    index_to_mask_kernel[grid](
+        q2k_block_sparse_index,
+        q2k_block_sparse_num,
+        mask,
+        batch_size,
+        num_heads,
+        num_q_blocks,
+        num_k_blocks,
+        max_kv_blocks,
+        BLOCK_Q=BLOCK_Q,
+        BLOCK_K=BLOCK_K,
+    )
+
+    return mask
 
 @triton.jit
 def topk_index_to_map_kernel(
@@ -276,3 +372,99 @@ def map_to_index(block_map: torch.Tensor):
 
     return index, index_num
 
+class BlockSparseAttentionFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num):
+        o, lse = block_sparse_attention_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)
+        ctx.save_for_backward(q, k, v, o, lse, k2q_block_sparse_index, k2q_block_sparse_num)
+        return o
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        q, k, v, o, lse, k2q_block_sparse_index, k2q_block_sparse_num = ctx.saved_tensors
+        grad_q, grad_k, grad_v = block_sparse_attention_backward(
+            q, k, v, o, lse, grad_output, k2q_block_sparse_index, k2q_block_sparse_num
+        )
+        return grad_q, grad_k, grad_v, None, None, None, None
+
+
+class DummyOperator(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x):
+        return x
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return grad_output
+
+class CheckpointSDPA(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, obj, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k):
+        """Forward pass."""
+        with torch.no_grad():
+            mask = index_to_mask(q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k, k.shape[2] // block_k)
+            outputs = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask)
+        ctx.save_for_backward(*detach_variable((q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)))
+        ctx.block_q = block_q
+        ctx.block_k = block_k
+        # the obj is passed in, then it can access the saved input
+        # tensors later for recomputation
+        obj.ctx = ctx
+        return outputs
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        """Backward pass."""
+        inputs = ctx.saved_tensors
+        output = ctx.output
+        torch.autograd.backward(output, grad_output)
+        ctx.output = None
+        grads = tuple(inp.grad for inp in inputs)
+        return (None, ) + grads + (None, None)
+
+
+class BlockSparseAttnTorch:
+    def __init__(self):
+        self.ctx = None
+    
+    def recompute_mask(self, _):
+        recomputed_mask = index_to_mask(self.q2k_block_sparse_index, self.q2k_block_sparse_num, self.block_q, self.block_k, self.num_kv_blocks)
+        mask_size = recomputed_mask.untyped_storage().size()
+        self.mask.untyped_storage().resize_(mask_size)
+        self.mask.untyped_storage().copy_(recomputed_mask.untyped_storage()) 
+    
+    def recompute(self, _):
+        q, k, v, q2k_block_sparse_index, q2k_block_sparse_num = self.ctx.saved_tensors
+        block_q = self.ctx.block_q
+        block_k = self.ctx.block_k
+        mask = index_to_mask(q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k, k.shape[2] // block_k)
+        with torch.enable_grad():
+            output = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask)
+        self.ctx.output = output
+        self.ctx = None
+
+    @torch._dynamo.disable
+    def forward(self, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k):
+        """
+        Differentiable block sparse attention function using PyTorch.
+        
+        Args:
+            q: Query tensor [batch_size, num_heads, seq_len_q, head_dim]
+            k: Key tensor [batch_size, num_heads, seq_len_kv, head_dim]
+            v: Value tensor [batch_size, num_heads, seq_len_kv, head_dim]
+            q2k_block_sparse_index: Indices for query-to-key sparse blocks
+            q2k_block_sparse_num: Number of sparse blocks for each query block
+            block_q: Block size for query
+            block_k: Block size for key-value
+
+        Returns:
+            output: Attention output tensor [batch_size, num_heads, seq_len_q, head_dim]
+        """
+
+        output = CheckpointSDPA.apply(
+            self, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, block_q, block_k
+        )
+
+        o = DummyOperator.apply(output)
+        o.register_hook(self.recompute)
+        return o

csrc/attn/vsa/block_sparse_attn_triton.py

@@ -1,707 +0,0 @@
-"""
-Fused Attention
-===============
-
-This is a Triton implementation of the Flash Attention v2 algorithm from Tri Dao
-(https://tridao.me/publications/flash2/flash2.pdf)
-
-Credits: OpenAI kernel team
-"""
-
-import pytest
-import torch
-import triton
-import triton.language as tl
-
-# ──────────────────────────── SPARSE ADDITION BEGIN ───────────────────────────
-import math  # small utility needed by the sparse wrapper
-# ──────────────────────────── SPARSE ADDITION END ─────────────────────────────
-
-
-@triton.jit
-def _attn_fwd_inner(acc, l_i, m_i, q,  #
-                    K_block_ptr, V_block_ptr,  #
-                    start_m, qk_scale,  #
-                    BLOCK_M: tl.constexpr, HEAD_DIM: tl.constexpr, BLOCK_N: tl.constexpr,  #
-                    STAGE: tl.constexpr, offs_m: tl.constexpr, offs_n: tl.constexpr,  #
-                    N_CTX: tl.constexpr, fp8_v: tl.constexpr):
-
-    # loop over k, v and update accumulator
-    for start_n in range(0, N_CTX, BLOCK_N):
-        # -- compute qk ----
-        k = tl.load(K_block_ptr)
-        qk = tl.dot(q, k)
-
-        m_ij = tl.maximum(m_i, tl.max(qk, 1) * qk_scale)
-        qk = qk * qk_scale - m_ij[:, None]
-        p = tl.math.exp2(qk)
-        l_ij = tl.sum(p, 1)
-        # -- update m_i and l_i
-        alpha = tl.math.exp2(m_i - m_ij)
-        l_i = l_i * alpha + l_ij
-        # -- update output accumulator --
-        acc = acc * alpha[:, None]
-        # update acc
-        v = tl.load(V_block_ptr)
-        if fp8_v:
-            p = p.to(tl.float8e5)
-        else:
-            p = p.to(tl.bfloat16)
-        acc = tl.dot(p, v, acc)
-        # update m_i and l_i
-        m_i = m_ij
-        V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
-        K_block_ptr = tl.advance(K_block_ptr, (0, BLOCK_N))
-    return acc, l_i, m_i
-
-
-# We don't run auto-tuning every time to keep the tutorial fast. Keeping
-# the code below and commenting out the equivalent parameters is convenient for
-# re-tuning.
-configs = [
-    triton.Config({'BLOCK_M': BM, 'BLOCK_N': BN}, num_stages=s, num_warps=w) \
-    for BM in [64]\
-    for BN in [64]\
-    for s in [3, 4, 7]\
-    for w in [4, 8]\
-]
-
-# ──────────────────────────── SPARSE ADDITION BEGIN ───────────────────────────
-@triton.autotune(configs, key=["N_CTX", "HEAD_DIM"])
-@triton.jit
-def _attn_fwd_sparse(Q, K, V, sm_scale,                         #
-                     q2k_index, q2k_num, max_kv_blks,           #
-                     M, Out,                                    #
-                     stride_qz, stride_qh, stride_qm, stride_qk,
-                     stride_kz, stride_kh, stride_kn, stride_kk,
-                     stride_vz, stride_vh, stride_vk, stride_vn,
-                     stride_oz, stride_oh, stride_om, stride_on,
-                     Z, H, N_CTX,                               #
-                     HEAD_DIM: tl.constexpr,                    #
-                     BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr,
-                     STAGE: tl.constexpr):
-    """
-    64×64 **block-sparse** forward kernel. Back-prop kernels remain dense
-    (32×64 and 64×32) – memory footprint unchanged.
-    """
-
-    # ----- program-id mapping -----
-    q_blk   = tl.program_id(0)          # Q-tile index
-    off_hz  = tl.program_id(1)          # fused (batch, head)
-    b       = off_hz // H
-    h       = off_hz %  H
-    q_tiles = N_CTX // BLOCK_M
-    meta_base = ((b * H + h) * q_tiles + q_blk)
-
-    kv_blocks = tl.load(q2k_num  + meta_base)                 # int32
-    kv_ptr    = q2k_index + meta_base * max_kv_blks           # ptr to list
-
-    # ----- base pointers -----
-    qvk_off = (b.to(tl.int64) * stride_qz +
-               h.to(tl.int64) * stride_qh)
-
-    Q_ptr = tl.make_block_ptr(
-        base=Q + qvk_off, shape=(N_CTX, HEAD_DIM),
-        strides=(stride_qm, stride_qk),
-        offsets=(q_blk * BLOCK_M, 0),
-        block_shape=(BLOCK_M, HEAD_DIM), order=(1, 0))
-
-    K_base = tl.make_block_ptr(
-        base=K + qvk_off, shape=(HEAD_DIM, N_CTX),
-        strides=(stride_kk, stride_kn),
-        offsets=(0, 0),
-        block_shape=(HEAD_DIM, BLOCK_N), order=(0, 1))
-
-    v_order: tl.constexpr = (0, 1) if V.dtype.element_ty == tl.float8e5 else (1, 0)
-    V_base = tl.make_block_ptr(
-        base=V + qvk_off, shape=(N_CTX, HEAD_DIM),
-        strides=(stride_vk, stride_vn),
-        offsets=(0, 0),
-        block_shape=(BLOCK_N, HEAD_DIM), order=v_order)
-
-    O_ptr = tl.make_block_ptr(
-        base=Out + qvk_off, shape=(N_CTX, HEAD_DIM),
-        strides=(stride_om, stride_on),
-        offsets=(q_blk * BLOCK_M, 0),
-        block_shape=(BLOCK_M, HEAD_DIM), order=(1, 0))
-
-    # ----- accumulators -----
-    offs_m = q_blk * BLOCK_M + tl.arange(0, BLOCK_M)
-    m_i = tl.full([BLOCK_M], -float("inf"), tl.float32)
-    l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0
-    acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32)
-    qk_scale = sm_scale * 1.44269504  # 1/ln2
-    q = tl.load(Q_ptr)
-
-    # ----- sparse loop over valid K/V tiles -----
-    for i in range(0, kv_blocks):
-        kv_idx = tl.load(kv_ptr + i).to(tl.int32)
-
-        K_ptr = tl.advance(K_base, (0, kv_idx * BLOCK_N))
-        V_ptr = tl.advance(V_base, (kv_idx * BLOCK_N, 0))
-
-        k = tl.load(K_ptr)
-        qk = tl.dot(q, k)
-
-        m_ij = tl.maximum(m_i, tl.max(qk, 1) * qk_scale)
-        p = tl.math.exp2(qk * qk_scale - m_ij[:, None])
-        l_ij = tl.sum(p, 1)
-
-        alpha = tl.math.exp2(m_i - m_ij)
-        l_i = l_i * alpha + l_ij
-        acc = acc * alpha[:, None]
-
-        v = tl.load(V_ptr)
-        acc = tl.dot(p.to(tl.bfloat16), v, acc)
-        m_i = m_ij
-
-    # ----- epilogue -----
-    m_i += tl.math.log2(l_i)
-    acc = acc / l_i[:, None]
-    tl.store(M + off_hz * N_CTX + offs_m, m_i)
-    tl.store(O_ptr, acc.to(Out.type.element_ty))
-# ──────────────────────────── SPARSE ADDITION END ─────────────────────────────
-
-
-@triton.autotune(configs, key=["N_CTX", "HEAD_DIM"])
-@triton.jit
-def _attn_fwd(Q, K, V, sm_scale, M, Out,  #
-              stride_qz, stride_qh, stride_qm, stride_qk,  #
-              stride_kz, stride_kh, stride_kn, stride_kk,  #
-              stride_vz, stride_vh, stride_vk, stride_vn,  #
-              stride_oz, stride_oh, stride_om, stride_on,  #
-              Z, H, N_CTX,  #
-              HEAD_DIM: tl.constexpr,  #
-              BLOCK_M: tl.constexpr,  #
-              BLOCK_N: tl.constexpr,  #
-              STAGE: tl.constexpr  #
-              ):
-    tl.static_assert(BLOCK_N <= HEAD_DIM)
-    start_m = tl.program_id(0)
-    off_hz = tl.program_id(1)
-    off_z = off_hz // H
-    off_h = off_hz % H
-    qvk_offset = off_z.to(tl.int64) * stride_qz + off_h.to(tl.int64) * stride_qh
-
-    # block pointers
-    Q_block_ptr = tl.make_block_ptr(
-        base=Q + qvk_offset,
-        shape=(N_CTX, HEAD_DIM),
-        strides=(stride_qm, stride_qk),
-        offsets=(start_m * BLOCK_M, 0),
-        block_shape=(BLOCK_M, HEAD_DIM),
-        order=(1, 0),
-    )
-    v_order: tl.constexpr = (0, 1) if V.dtype.element_ty == tl.float8e5 else (1, 0)
-    V_block_ptr = tl.make_block_ptr(
-        base=V + qvk_offset,
-        shape=(N_CTX, HEAD_DIM),
-        strides=(stride_vk, stride_vn),
-        offsets=(0, 0),
-        block_shape=(BLOCK_N, HEAD_DIM),
-        order=v_order,
-    )
-    K_block_ptr = tl.make_block_ptr(
-        base=K + qvk_offset,
-        shape=(HEAD_DIM, N_CTX),
-        strides=(stride_kk, stride_kn),
-        offsets=(0, 0),
-        block_shape=(HEAD_DIM, BLOCK_N),
-        order=(0, 1),
-    )
-    O_block_ptr = tl.make_block_ptr(
-        base=Out + qvk_offset,
-        shape=(N_CTX, HEAD_DIM),
-        strides=(stride_om, stride_on),
-        offsets=(start_m * BLOCK_M, 0),
-        block_shape=(BLOCK_M, HEAD_DIM),
-        order=(1, 0),
-    )
-    # initialize offsets
-    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
-    offs_n = tl.arange(0, BLOCK_N)
-    # initialize pointer to m and l
-    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
-    l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0
-    acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32)
-    # load scales
-    qk_scale = sm_scale
-    qk_scale *= 1.44269504  # 1/log(2)
-    # load q: it will stay in SRAM throughout
-    q = tl.load(Q_block_ptr)
-    acc, l_i, m_i = _attn_fwd_inner(acc, l_i, m_i, q, K_block_ptr, V_block_ptr,  #
-                                    start_m, qk_scale,  #
-                                    BLOCK_M, HEAD_DIM, BLOCK_N,  #
-                                    3, offs_m, offs_n, N_CTX, V.dtype.element_ty == tl.float8e5  #
-                                    )
-
-    # epilogue
-    m_i += tl.math.log2(l_i)
-    acc = acc / l_i[:, None]
-    m_ptrs = M + off_hz * N_CTX + offs_m
-    tl.store(m_ptrs, m_i)
-    tl.store(O_block_ptr, acc.to(Out.type.element_ty))
-
-
-@triton.jit
-def _attn_bwd_preprocess(O, DO,  #
-                         Delta,  #
-                         Z, H, N_CTX,  #
-                         BLOCK_M: tl.constexpr, HEAD_DIM: tl.constexpr  #
-                         ):
-    off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
-    off_hz = tl.program_id(1)
-    off_n = tl.arange(0, HEAD_DIM)
-    # load
-    o = tl.load(O + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :])
-    do = tl.load(DO + off_hz * HEAD_DIM * N_CTX + off_m[:, None] * HEAD_DIM + off_n[None, :]).to(tl.float32)
-    delta = tl.sum(o * do, axis=1)
-    # write-back
-    tl.store(Delta + off_hz * N_CTX + off_m, delta)
-
-
-# The main inner-loop logic for computing dK and dV.
-@triton.jit
-def _attn_bwd_dkdv(dk, dv,  #
-                   Q, k, v, sm_scale,  #
-                   DO,  #
-                   M, D,  #
-                   k2q_index, k2q_num, max_q_blks,
-                   # shared by Q/K/V/DO.
-                   stride_tok, stride_d,  #
-                   H, N_CTX, BLOCK_M1: tl.constexpr,  #
-                   BLOCK_N1: tl.constexpr,  #
-                   HEAD_DIM: tl.constexpr,  #
-                   # Filled in by the wrapper.
-                   start_n, start_m, num_steps):
-    offs_m = start_m + tl.arange(0, BLOCK_M1)
-    offs_n = start_n + tl.arange(0, BLOCK_N1)
-    offs_k = tl.arange(0, HEAD_DIM)
-    qT_ptrs = Q + offs_m[None, :] * stride_tok + offs_k[:, None] * stride_d
-    do_ptrs = DO + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d
-    # BLOCK_N1 must be a multiple of BLOCK_M1, otherwise the code wouldn't work.
-    tl.static_assert(BLOCK_N1 % BLOCK_M1 == 0)
-    step_m = BLOCK_M1
-    kv_blk   = tl.program_id(0)          # Q-tile index
-    off_hz  = tl.program_id(2)          # fused (batch, head)
-    b       = off_hz // H
-    h       = off_hz %  H
-    q_tiles = N_CTX // BLOCK_N1
-    meta_base = ((b * H + h) * q_tiles + kv_blk)
-
-    q_blocks = tl.load(k2q_num  + meta_base)                 # int32
-    q_ptr    = k2q_index + meta_base * max_q_blks           # ptr to list
-    
-        
-    for blk_idx in range(q_blocks*2):
-        block_sparse_offset = (tl.load(q_ptr + blk_idx//2).to(tl.int32)*2 + blk_idx%2) *step_m
-        qT = tl.load(qT_ptrs + block_sparse_offset * stride_tok)
-        # Load m before computing qk to reduce pipeline stall.
-        offs_m = start_m + block_sparse_offset + tl.arange(0, BLOCK_M1)
-        m = tl.load(M + offs_m)
-        qkT = tl.dot(k, qT)
-        pT = tl.math.exp2(qkT - m[None, :])
-        do = tl.load(do_ptrs + block_sparse_offset * stride_tok)
-        # Compute dV.
-        ppT = pT
-        ppT = ppT.to(tl.bfloat16)
-        dv += tl.dot(ppT, do)
-        # D (= delta) is pre-divided by ds_scale.
-        Di = tl.load(D + offs_m)
-        # Compute dP and dS.
-        dpT = tl.dot(v, tl.trans(do)).to(tl.float32)
-        dsT = pT * (dpT - Di[None, :])
-        dsT = dsT.to(tl.bfloat16)
-        dk += tl.dot(dsT, tl.trans(qT))
-        # Increment pointers.
-    return dk, dv
-
-
-
-# the main inner-loop logic for computing dQ
-@triton.jit
-def _attn_bwd_dq(dq, q, K, V,  #
-                 do, m, D,
-                 # shared by Q/K/V/DO.
-                 q2k_index, q2k_num, max_kv_blks,
-                 stride_tok, stride_d,  #
-                 H, N_CTX,  #
-                 BLOCK_M2: tl.constexpr,  #
-                 BLOCK_N2: tl.constexpr,  #
-                 HEAD_DIM: tl.constexpr,
-                 # Filled in by the wrapper.
-                 start_m, start_n, num_steps):
-    offs_m = start_m + tl.arange(0, BLOCK_M2)
-    offs_n = start_n + tl.arange(0, BLOCK_N2)
-    offs_k = tl.arange(0, HEAD_DIM)
-    kT_ptrs = K + offs_n[None, :] * stride_tok + offs_k[:, None] * stride_d
-    vT_ptrs = V + offs_n[None, :] * stride_tok + offs_k[:, None] * stride_d
-    # D (= delta) is pre-divided by ds_scale.
-    Di = tl.load(D + offs_m)
-    # BLOCK_M2 must be a multiple of BLOCK_N2, otherwise the code wouldn't work.
-    tl.static_assert(BLOCK_M2 % BLOCK_N2 == 0)
-    step_n = BLOCK_N2
-    
-    q_blk   = tl.program_id(0)          # Q-tile index
-    off_hz  = tl.program_id(2)          # fused (batch, head)
-    b       = off_hz // H
-    h       = off_hz %  H
-    q_tiles = N_CTX // BLOCK_M2
-    meta_base = ((b * H + h) * q_tiles + q_blk)
-
-    kv_blocks = tl.load(q2k_num  + meta_base)                 # int32
-    kv_ptr    = q2k_index + meta_base * max_kv_blks           # ptr to list
-    
-    
-    for blk_idx in range(kv_blocks*2):
-        block_sparse_offset = (tl.load(kv_ptr + blk_idx//2).to(tl.int32)*2 + blk_idx%2) *step_n * stride_tok
-        kT = tl.load(kT_ptrs + block_sparse_offset)
-        vT = tl.load(vT_ptrs + block_sparse_offset)
-        qk = tl.dot(q, kT)
-        p = tl.math.exp2(qk - m)
-        # Compute dP and dS.
-        dp = tl.dot(do, vT).to(tl.float32)
-        ds = p * (dp - Di[:, None])
-        ds = ds.to(tl.bfloat16)
-        # Compute dQ.
-        # NOTE: We need to de-scale dq in the end, because kT was pre-scaled.
-        dq += tl.dot(ds, tl.trans(kT))
-        # Increment pointers.
-    return dq
-
-
-
-@triton.jit
-def _attn_bwd(Q, K, V, sm_scale,  #
-              DO,  #
-              DQ, DK, DV,  #
-              M, D,
-              q2k_index, q2k_num, max_kv_blks,
-              k2q_index, k2q_num, max_q_blks,
-              # shared by Q/K/V/DO.
-              stride_z, stride_h, stride_tok, stride_d,  #
-              H, N_CTX,  #
-              BLOCK_M1: tl.constexpr,  #
-              BLOCK_N1: tl.constexpr,  #
-              BLOCK_M2: tl.constexpr,  #
-              BLOCK_N2: tl.constexpr,  #
-              HEAD_DIM: tl.constexpr):
-    LN2 = 0.6931471824645996  # = ln(2)
-
-    bhid = tl.program_id(2)
-    off_chz = (bhid * N_CTX).to(tl.int64)
-    adj = (stride_h * (bhid % H) + stride_z * (bhid // H)).to(tl.int64)
-    pid = tl.program_id(0)
-
-    # offset pointers for batch/head
-    Q += adj
-    K += adj
-    V += adj
-    DO += adj
-    DQ += adj
-    DK += adj
-    DV += adj
-    M += off_chz
-    D += off_chz
-
-    # load scales
-    offs_k = tl.arange(0, HEAD_DIM)
-
-    start_n = pid * BLOCK_N1
-    start_m = 0
-
-    offs_n = start_n + tl.arange(0, BLOCK_N1)
-
-    dv = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32)
-    dk = tl.zeros([BLOCK_N1, HEAD_DIM], dtype=tl.float32)
-
-    # load K and V: they stay in SRAM throughout the inner loop.
-    k = tl.load(K + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d)
-    v = tl.load(V + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d)
-
-
-    num_steps = N_CTX // BLOCK_M1
-
-    dk, dv = _attn_bwd_dkdv(  #
-        dk, dv,  #
-        Q, k, v, sm_scale,  #
-        DO,  #
-        M, D,  #
-        k2q_index, k2q_num, max_q_blks,
-        stride_tok, stride_d,  #
-        H, N_CTX,  #
-        BLOCK_M1, BLOCK_N1, HEAD_DIM,  #
-        start_n, start_m, num_steps  #
-    )
-
-    dv_ptrs = DV + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d
-    tl.store(dv_ptrs, dv)
-
-    # Write back dK.
-    dk *= sm_scale
-    dk_ptrs = DK + offs_n[:, None] * stride_tok + offs_k[None, :] * stride_d
-    tl.store(dk_ptrs, dk)
-
-    # THIS BLOCK DOES DQ:
-    start_m = pid * BLOCK_M2
-    end_n = 0
-
-    offs_m = start_m + tl.arange(0, BLOCK_M2)
-
-    q = tl.load(Q + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d)
-    dq = tl.zeros([BLOCK_M2, HEAD_DIM], dtype=tl.float32)
-    do = tl.load(DO + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d)
-
-    m = tl.load(M + offs_m)
-    m = m[:, None]
-
-    num_steps = N_CTX // BLOCK_N2
-    dq = _attn_bwd_dq(dq, q, K, V,  #
-                      do, m, D,  #
-                      q2k_index, q2k_num, max_kv_blks,
-                      stride_tok, stride_d,  #
-                      H, N_CTX,  #
-                      BLOCK_M2, BLOCK_N2, HEAD_DIM,  #
-                      start_m, end_n, num_steps  #
-                      )
-    # Write back dQ.
-    dq_ptrs = DQ + offs_m[:, None] * stride_tok + offs_k[None, :] * stride_d
-    dq *= LN2
-    tl.store(dq_ptrs, dq)
-    
-    
-
-class _attention(torch.autograd.Function):
-
-    @staticmethod
-    def forward(ctx, q, k, v):
-        # shape constraints
-        sm_scale = 1.0 / (q.shape[-1] ** 0.5)
-        HEAD_DIM_Q, HEAD_DIM_K = q.shape[-1], k.shape[-1]
-        # when v is in float8_e5m2 it is transposed.
-        HEAD_DIM_V = v.shape[-1]
-        assert HEAD_DIM_Q == HEAD_DIM_K and HEAD_DIM_K == HEAD_DIM_V
-        assert HEAD_DIM_K in {16, 32, 64, 128, 256}
-        o = torch.empty_like(q)
-        stage = 1
-        extra_kern_args = {}
-
-
-        grid = lambda args: (triton.cdiv(q.shape[2], args["BLOCK_M"]), q.shape[0] * q.shape[1], 1)
-        M = torch.empty((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
-        _attn_fwd[grid](
-            q, k, v, sm_scale, M, o,  #
-            q.stride(0), q.stride(1), q.stride(2), q.stride(3),  #
-            k.stride(0), k.stride(1), k.stride(2), k.stride(3),  #
-            v.stride(0), v.stride(1), v.stride(2), v.stride(3),  #
-            o.stride(0), o.stride(1), o.stride(2), o.stride(3),  #
-            q.shape[0], q.shape[1],  #
-            N_CTX=q.shape[2],  #
-            HEAD_DIM=HEAD_DIM_K,  #
-            STAGE=stage,  #
-            **extra_kern_args)
-
-        ctx.save_for_backward(q, k, v, o, M)
-        ctx.grid = grid
-        ctx.sm_scale = sm_scale
-        ctx.HEAD_DIM = HEAD_DIM_K
-        return o
-
-    @staticmethod
-    def backward(ctx, do):
-        q, k, v, o, M = ctx.saved_tensors
-        assert do.is_contiguous()
-        assert q.stride() == k.stride() == v.stride() == o.stride() == do.stride()
-        dq = torch.empty_like(q)
-        dk = torch.empty_like(k)
-        dv = torch.empty_like(v)
-        BATCH, N_HEAD, N_CTX = q.shape[:3]
-        PRE_BLOCK = 128
-        BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 32, 64, 64, 32
-        RCP_LN2 = 1.4426950408889634  # = 1.0 / ln(2)
-        arg_k = k
-        arg_k = arg_k * (ctx.sm_scale * RCP_LN2)
-        PRE_BLOCK = 128
-        assert N_CTX % PRE_BLOCK == 0
-        pre_grid = (N_CTX // PRE_BLOCK, BATCH * N_HEAD)
-        delta = torch.empty_like(M)
-        _attn_bwd_preprocess[pre_grid](
-            o, do,  #
-            delta,  #
-            BATCH, N_HEAD, N_CTX,  #
-            BLOCK_M=PRE_BLOCK, HEAD_DIM=ctx.HEAD_DIM  #
-        )
-        grid = (N_CTX // BLOCK_N1, 1, BATCH * N_HEAD)
-        _attn_bwd[grid](
-            q, arg_k, v, ctx.sm_scale, do, dq, dk, dv,  #
-            M, delta,  #
-            q.stride(0), q.stride(1), q.stride(2), q.stride(3),  #
-            N_HEAD, N_CTX,  #
-            BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1,  #
-            BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2,  #
-            HEAD_DIM=ctx.HEAD_DIM  #
-        )
-
-        return dq, dk, dv, None, None
-
-
-attention = _attention.apply
-
-# ──────────────────────────── SPARSE ADDITION BEGIN ───────────────────────────
-class _attention_sparse(torch.autograd.Function):
-    """
-    Thin autograd wrapper that uses the sparse forward kernel above and the
-    standard dense backward kernels defined earlier (no extra memory use).
-    """
-
-    @staticmethod
-    def forward(ctx, q, k, v, q2k_index, q2k_num, k2q_index, k2q_num):
-        B, H, T, D = q.shape
-        sm_scale = 1.0 / math.sqrt(D)
-        max_kv_blks = q2k_index.shape[-1]
-        assert T % 64 == 0, f"T must be a multiple of 64, but got {T}"
-        assert T // 64 == q2k_num.shape[-1], f"shape mismatch, T // 64 = {T // 64}, q2k_num.shape[-2] = {q2k_num.shape[-2]}"
-        o = torch.empty_like(q)
-        M = torch.empty((B, H, T), dtype=torch.float32, device=q.device)
-
-        grid = lambda _: (triton.cdiv(T, 64), B * H, 1)
-        _attn_fwd_sparse[grid](
-            q, k, v, sm_scale,
-            q2k_index, q2k_num, max_kv_blks,
-            M, o,
-            q.stride(0), q.stride(1), q.stride(2), q.stride(3),
-            k.stride(0), k.stride(1), k.stride(2), k.stride(3),
-            v.stride(0), v.stride(1), v.stride(2), v.stride(3),
-            o.stride(0), o.stride(1), o.stride(2), o.stride(3),
-            B, H, T,
-            HEAD_DIM=D, STAGE=3
-        )
-
-        ctx.save_for_backward(q, k, v, o, M, q2k_index, q2k_num, k2q_index, k2q_num)
-        ctx.grid = None
-        ctx.sm_scale = sm_scale
-        ctx.HEAD_DIM = D
-        return o
-
-    @staticmethod
-    def backward(ctx, do):
-        q, k, v, o, M, q2k_index, q2k_num, k2q_index, k2q_num = ctx.saved_tensors
-        assert do.is_contiguous()
-        assert q.stride() == k.stride() == v.stride() == o.stride() == do.stride()
-        dq = torch.empty_like(q)
-        dk = torch.empty_like(k)
-        dv = torch.empty_like(v)
-        BATCH, N_HEAD, N_CTX = q.shape[:3]
-        PRE_BLOCK = 128
-        BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 32, 64, 64, 32
-        RCP_LN2 = 1.4426950408889634  # = 1.0 / ln(2)
-        arg_k = k
-        arg_k = arg_k * (ctx.sm_scale * RCP_LN2)
-        PRE_BLOCK = 128
-        assert N_CTX % PRE_BLOCK == 0
-        pre_grid = (N_CTX // PRE_BLOCK, BATCH * N_HEAD)
-        delta = torch.empty_like(M)
-        _attn_bwd_preprocess[pre_grid](
-            o, do,  #
-            delta,  #
-            BATCH, N_HEAD, N_CTX,  #
-            BLOCK_M=PRE_BLOCK, HEAD_DIM=ctx.HEAD_DIM  #
-        )
-        
-        
-        max_q_blks = k2q_index.shape[-1]
-        max_kv_blks = q2k_index.shape[-1]
-        
-        
-        grid = (N_CTX // BLOCK_N1, 1, BATCH * N_HEAD)
-        _attn_bwd[grid](
-            q, arg_k, v, ctx.sm_scale, do, dq, dk, dv,  #
-            M, delta,  #
-            q2k_index, q2k_num, max_kv_blks,
-            k2q_index, k2q_num, max_q_blks,
-            q.stride(0), q.stride(1), q.stride(2), q.stride(3),  #
-            N_HEAD, N_CTX,  #
-            BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1,  #
-            BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2,  #
-            HEAD_DIM=ctx.HEAD_DIM  #
-        )
-
-        return dq, dk, dv, None, None, None, None
-
-attention_sparse = _attention_sparse.apply
-# ──────────────────────────── SPARSE ADDITION END ─────────────────────────────
-
-
-try:
-    from flash_attn.flash_attn_interface import \
-        flash_attn_qkvpacked_func as flash_attn_func
-    HAS_FLASH = True
-except BaseException:
-    HAS_FLASH = False
-
-TORCH_HAS_FP8 = hasattr(torch, 'float8_e5m2')
-BATCH, N_HEADS, HEAD_DIM = 4, 32, 128
-# vary seq length for fixed head and batch=4
-configs = []
-for mode in ["fwd", "bwd"]:
-        configs.append(
-            triton.testing.Benchmark(
-                x_names=["N_CTX"],
-                x_vals=[2**i for i in range(10, 15)],
-                line_arg="provider",
-                line_vals=["triton-fp16"] + (["triton-fp8"] if TORCH_HAS_FP8 else []) +
-                (["flash"] if HAS_FLASH else []),
-                line_names=["Triton [FP16]"] + (["Triton [FP8]"] if TORCH_HAS_FP8 else []) +
-                (["Flash-2"] if HAS_FLASH else []),
-                styles=[("red", "-"), ("blue", "-"), ("green", "-")],
-                ylabel="ms",
-                plot_name=f"fused-attention-batch{BATCH}-head{N_HEADS}-d{HEAD_DIM}-{mode}",
-                args={
-                    "H": N_HEADS,
-                    "BATCH": BATCH,
-                    "HEAD_DIM": HEAD_DIM,
-                    "mode": mode,
-                },
-            ))
-
-
-@triton.testing.perf_report(configs)
-def bench_flash_attention(BATCH, H, N_CTX, HEAD_DIM, mode, provider, device="cuda"):
-    assert mode in ["fwd", "bwd"]
-    warmup = 25
-    rep = 100
-    dtype = torch.bfloat16
-    if "triton" in provider:
-        q = torch.randn((BATCH, H, N_CTX, HEAD_DIM), dtype=dtype, device=device, requires_grad=True)
-        k = torch.randn((BATCH, H, N_CTX, HEAD_DIM), dtype=dtype, device=device, requires_grad=True)
-        v = torch.randn((BATCH, H, N_CTX, HEAD_DIM), dtype=dtype, device=device, requires_grad=True)
-        if mode == "fwd" and "fp8" in provider:
-            q = q.to(torch.float8_e5m2)
-            k = k.to(torch.float8_e5m2)
-            v = v.permute(0, 1, 3, 2).contiguous()
-            v = v.permute(0, 1, 3, 2)
-            v = v.to(torch.float8_e5m2)
-        fn = lambda: attention(q, k, v)
-        if mode == "bwd":
-            o = fn()
-            do = torch.randn_like(o)
-            fn = lambda: o.backward(do, retain_graph=True)
-        ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
-    if provider == "flash":
-        qkv = torch.randn((BATCH, N_CTX, 3, H, HEAD_DIM), dtype=dtype, device=device, requires_grad=True)
-        fn = lambda: flash_attn_func(qkv, causal=False)
-        if mode == "bwd":
-            o = fn()
-            do = torch.randn_like(o)
-            fn = lambda: o.backward(do, retain_graph=True)
-        ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
-    flops_per_matmul = 2.0 * BATCH * H * N_CTX * N_CTX * HEAD_DIM
-    total_flops = 2 * flops_per_matmul
-    if mode == "bwd":
-        total_flops *= 2.5  # 2.0(bwd) + 0.5(recompute)
-    return total_flops / ms * 1e-9
-
-
-if __name__ == "__main__":
-    # only works on post-Ampere GPUs right now
-    bench_flash_attention.run(save_path=".", print_data=True)

csrc/attn/vsa/vsa.py

@@ -1,47 +0,0 @@
-import torch
-try:
-    from vsa_cuda import block_sparse_fwd, block_sparse_bwd
-except ImportError:
-    block_sparse_fwd = None
-    block_sparse_bwd = None
-from .block_sparse_attn_triton import attention_sparse as block_sparse_attn_triton
-
-class BlockSparseAttentionFunction(torch.autograd.Function):
-    @staticmethod
-    def forward(ctx, q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num):
-        o, lse = block_sparse_fwd(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num)
-        ctx.save_for_backward(q, k, v, o, lse, k2q_block_sparse_index, k2q_block_sparse_num)
-        return o
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        q, k, v, o, lse, k2q_block_sparse_index, k2q_block_sparse_num = ctx.saved_tensors
-        grad_q, grad_k, grad_v = block_sparse_bwd(
-            q, k, v, o, lse, grad_output, k2q_block_sparse_index, k2q_block_sparse_num
-        )
-        return grad_q, grad_k, grad_v, None, None, None, None
-
-
-@torch._dynamo.disable
-def block_sparse_attn(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num):
-    """
-    Differentiable block sparse attention function.
-    
-    Args:
-        q: Query tensor [batch_size, num_heads, seq_len_q, head_dim]
-        k: Key tensor [batch_size, num_heads, seq_len_kv, head_dim]
-        v: Value tensor [batch_size, num_heads, seq_len_kv, head_dim]
-        q2k_block_sparse_index: Indices for query-to-key sparse blocks
-        q2k_block_sparse_num: Number of sparse blocks for each query block
-        k2q_block_sparse_index: Indices for key-to-query sparse blocks (for backward pass)
-        k2q_block_sparse_num: Number of sparse blocks for each key block (for backward pass)
-    
-    Returns:
-        output: Attention output tensor [batch_size, num_heads, seq_len_q, head_dim]
-    """
-    if block_sparse_fwd is not None:
-        return BlockSparseAttentionFunction.apply(
-            q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num
-        )
-    else:
-        return block_sparse_attn_triton(q, k, v, q2k_block_sparse_index, q2k_block_sparse_num, k2q_block_sparse_index, k2q_block_sparse_num)

docs/Makefile

@@ -23,4 +23,3 @@ clean:
 	@$(SPHINXBUILD) -M clean "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
 	rm -rf "$(SOURCEDIR)/getting_started/examples"
 	rm -rf "$(SOURCEDIR)/inference/examples"
-	rm -rf "$(SOURCEDIR)/training/examples"

docs/requirements-docs.txt

@@ -11,5 +11,5 @@ commonmark # Required by sphinx-argparse when using :markdownhelp:
 
 # packages to install to build the documentation
 cachetools
-# -f https://download.pytorch.org/whl/cpu
+-f https://download.pytorch.org/whl/cpu
 torch

docs/source/api/summary.md

@@ -9,11 +9,11 @@
 ## Initialization Configuration
 
 ```{autodoc2-summary}
-    fastvideo.configs.pipelines.PipelineConfig
+    fastvideo.v1.configs.pipelines.PipelineConfig
 ```
 
 ## Sampling Configuration
 
 ```{autodoc2-summary}
-    fastvideo.configs.sample.SamplingParam
+    fastvideo.v1.configs.sample.SamplingParam
 ```

docs/source/design/overview.md

@@ -1,24 +1,24 @@
 # 🔍 FastVideo Overview
 
-This document outlines FastVideo's architecture for developers interested in framework internals or contributions. It serves as an onboarding guide for new contributors by providing an overview of the most important directories and files within the `fastvideo/` codebase.
+This document outlines FastVideo's architecture for developers interested in framework internals or contributions. It serves as an onboarding guide for new contributors by providing an overview of the most important directories and files within the `fastvideo/v1/` codebase.
 
-## Table of Contents - Directory Structure and Files
+## Table of Contents - V1 Directory Structure and Files
 
-- [`fastvideo/pipelines/`](#design-pipeline-system) - Core diffusion pipeline components
-- [`fastvideo/models/`](#design-model-components) - Model implementations
+- [`fastvideo/v1/pipelines/`](#design-pipeline-system) - Core diffusion pipeline components
+- [`fastvideo/v1/models/`](#design-model-components) - Model implementations
   - [`dits/`](#design-transformer-models) - Transformer-based diffusion models
   - [`vaes/`](#design-vae-variational-auto-encoder) - Variational autoencoders
   - [`encoders/`](#design-text-and-image-encoders) - Text and image encoders
   - [`schedulers/`](#design-schedulers) - Diffusion schedulers
-- [`fastvideo/attention/`](#design-optimized-attention) - Optimized attention implementations
-- [`fastvideo/distributed/`](#design-distributed-processing) - Distributed computing utilities
-- [`fastvideo/layers/`](#design-tensor-parallelism) - Custom neural network layers
-- [`fastvideo/platforms/`](#design-platforms) - Hardware platform abstractions
-- [`fastvideo/worker/`](#design-executor-and-worker-abstractions) - Multi-GPU process management
-- [`fastvideo/fastvideo_args.py`](#design-fastvideo-args) - Argument handling
-- [`fastvideo/forward_context.py`](#design-forwardcontext) - Forward pass context management
-- `fastvideo/utils.py` - Utility functions
-- [`fastvideo/logger.py`](#design-logger) - Logging infrastructure
+- [`fastvideo/v1/attention/`](#design-optimized-attention) - Optimized attention implementations
+- [`fastvideo/v1/distributed/`](#design-distributed-processing) - Distributed computing utilities
+- [`fastvideo/v1/layers/`](#design-tensor-parallelism) - Custom neural network layers
+- [`fastvideo/v1/platforms/`](#design-platforms) - Hardware platform abstractions
+- [`fastvideo/v1/worker/`](#design-executor-and-worker-abstractions) - Multi-GPU process management
+- [`fastvideo/v1/fastvideo_args.py`](#design-fastvideo-args) - Argument handling
+- [`fastvideo/v1/forward_context.py`](#design-forwardcontext) - Forward pass context management
+- `fastvideo/v1/utils.py` - Utility functions
+- [`fastvideo/v1/logger.py`](#design-logger) - Logging infrastructure
 
 ## Core Architecture
 
@@ -32,7 +32,7 @@ FastVideo separates model components from execution logic with these principles:
 (design-fastvideo-args)=
 ## FastVideoArgs
 
-The `FastVideoArgs` class in `fastvideo/fastvideo_args.py` serves as the central configuration system for FastVideo. It contains all parameters needed to control model loading, inference configuration, performance optimization settings, and more.
+The `FastVideoArgs` class in `fastvideo/v1/fastvideo_args.py` serves as the central configuration system for FastVideo. It contains all parameters needed to control model loading, inference configuration, performance optimization settings, and more.
 
 Key features include:
 - **Command-line Interface**: Automatic conversion between CLI arguments and dataclass fields
@@ -111,7 +111,7 @@ def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> Forward
 (design-forwardbatch)=
 ### ForwardBatch
 
-Defined in `fastvideo/pipelines/pipeline_batch_info.py`, `ForwardBatch` encapsulates the data payload passed between pipeline stages. It typically holds:
+Defined in `fastvideo/v1/pipelines/pipeline_batch_info.py`, `ForwardBatch` encapsulates the data payload passed between pipeline stages. It typically holds:
 
 - **Input Data**: Prompts, images, generation parameters
 - **Intermediate State**: Embeddings, latents, timesteps, accumulated during stage execution
@@ -123,14 +123,14 @@ This structure facilitates clear state transitions between stages.
 (design-model-components)=
 ## Model Components
 
-The `fastvideo/models/` directory contains implementations of the core neural network models used in video diffusion:
+The `fastvideo/v1/models/` directory contains implementations of the core neural network models used in video diffusion:
 
 (design-transformer-models)=
 ### Transformer Models
 
 Transformer networks perform the actual denoising during diffusion:
 
-- **Location**: `fastvideo/models/dits/`
+- **Location**: `fastvideo/v1/models/dits/`
 - **Examples**:
   - `WanTransformer3DModel`
   - `HunyuanVideoTransformer3DModel`
@@ -157,7 +157,7 @@ def forward(
 
 VAEs handle conversion between pixel space and latent space:
 
-- **Location**: `fastvideo/models/vaes/`
+- **Location**: `fastvideo/v1/models/vaes/`
 - **Examples**:
   - `AutoencoderKLWan`
   - `AutoencoderKLHunyuanVideo`
@@ -175,7 +175,7 @@ FastVideo's VAE implementations include:
 
 Encoders process conditioning inputs into embeddings:
 
-- **Location**: `fastvideo/models/encoders/`
+- **Location**: `fastvideo/v1/models/encoders/`
 - **Text Encoders**:
   - `CLIPTextModel`
   - `LlamaModel`
@@ -193,7 +193,7 @@ FastVideo implements optimizations such as:
 
 Schedulers manage the diffusion sampling process:
 
-- **Location**: `fastvideo/models/schedulers/`
+- **Location**: `fastvideo/v1/models/schedulers/`
 - **Examples**:
   - `UniPCMultistepScheduler`
   - `FlowMatchEulerDiscreteScheduler`
@@ -219,7 +219,7 @@ def step(
 (design-optimized-attention)=
 ## Optimized Attention
 
-The `fastvideo/attention/` directory contains optimized attention implementations crucial for efficient video diffusion:
+The `fastvideo/v1/attention/` directory contains optimized attention implementations crucial for efficient video diffusion:
 
 ### Attention Backends
 Multiple implementations with automatic selection:
@@ -248,7 +248,7 @@ Supports various patterns with memory optimization techniques:
 (design-distributed-processing)=
 ## Distributed Processing
 
-The `fastvideo/distributed/` directory contains implementations for distributed model execution:
+The `fastvideo/v1/distributed/` directory contains implementations for distributed model execution:
 
 (design-tensor-parallelism)=
 ### Tensor Parallelism
@@ -260,7 +260,7 @@ Tensor parallelism splits model weights across devices:
 
 ```python
 # Tensor-parallel layers in a transformer block
-from fastvideo.layers.linear import ColumnParallelLinear, RowParallelLinear
+from fastvideo.v1.layers.linear import ColumnParallelLinear, RowParallelLinear
 
 # Split along output dimension
 self.qkv_proj = ColumnParallelLinear(
@@ -288,7 +288,7 @@ Sequence parallelism splits sequences across devices:
 
 ```python
 # Distributed attention for long sequences
-from fastvideo.attention import DistributedAttention
+from fastvideo.v1.attention import DistributedAttention
 
 self.attn = DistributedAttention(
     num_heads=num_heads,
@@ -312,7 +312,7 @@ Efficient communication primitives minimize distributed overhead:
 
 ### ForwardContext
 
-Defined in `fastvideo/forward_context.py`, `ForwardContext` manages execution-specific state *within* a forward pass, particularly for low-level optimizations. It is accessed via `get_forward_context()`.
+Defined in `fastvideo/v1/forward_context.py`, `ForwardContext` manages execution-specific state *within* a forward pass, particularly for low-level optimizations. It is accessed via `get_forward_context()`.
 
 - **Attention Metadata**: Configuration for optimized attention kernels (`attn_metadata`)
 - **Profiling Data**: Potential hooks for performance metrics collection
@@ -333,7 +333,7 @@ with set_forward_context(current_timestep, attn_metadata, fastvideo_args):
 (design-executor-and-worker-abstractions)=
 ## Executor and Worker System
 
-The `fastvideo/worker/` directory contains the distributed execution framework:
+The `fastvideo/v1/worker/` directory contains the distributed execution framework:
 
 ### Executor Abstraction
 
@@ -360,7 +360,7 @@ This design allows FastVideo to efficiently utilize multiple GPUs while providin
 (design-platforms)=
 ## Platforms
 
-The `fastvideo/platforms/` directory provides hardware platform abstractions that enable FastVideo to run efficiently on different hardware configurations:
+The `fastvideo/v1/platforms/` directory provides hardware platform abstractions that enable FastVideo to run efficiently on different hardware configurations:
 
 ### Platform Abstraction
 
@@ -377,7 +377,7 @@ The primary components include:
 Usage example:
 
 ```python
-from fastvideo.platforms import current_platform, _Backend
+from fastvideo.v1.platforms import current_platform, _Backend
 
 # Check hardware capabilities
 if current_platform.supports_backend(_Backend.FLASH_ATTN):
@@ -398,9 +398,9 @@ See [PR](https://github.com/hao-ai-lab/FastVideo/pull/356)
 
 If you're a new contributor, here are some common areas to explore:
 
-1. **Adding a new model**: Implement new model types in the appropriate subdirectory of `fastvideo/models/`
+1. **Adding a new model**: Implement new model types in the appropriate subdirectory of `fastvideo/v1/models/`
 2. **Optimizing performance**: Look at attention implementations or memory management
-3. **Adding a new pipeline**: Create a new pipeline subclass in `fastvideo/pipelines/`
+3. **Adding a new pipeline**: Create a new pipeline subclass in `fastvideo/v1/pipelines/`
 4. **Hardware support**: Extend the `platforms` module for new hardware targets
 
 When adding code, follow these practices:

docs/source/generate_examples.py

@@ -27,15 +27,6 @@ def fix_case(text: str) -> str:
         "openai": "OpenAI",
         "multilora": "MultiLoRA",
         "mlpspeculator": "MLPSpeculator",
-        "finetune": "Finetune",
-        "distillation": "Distillation",
-        "wan": "Wan",
-        "i2v": "I2V",
-        "t2v": "T2V",
-        "1.3b": "1.3B",
-        "14b": "14B",
-        "480p": "480P",
-        "720p": "720P",
         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
     }
@@ -170,35 +161,31 @@ class Example:
         return content
 
 
-@dataclass
-class NestedStructure:
-    """Helper class to manage nested documentation structures for training/distillation."""
-    category: str
-    method: str
-    model: str
-    dataset: str
-    example: Example
-
-    @property
-    def filename(self) -> str:
-        return f"{self.model}_{self.dataset}"
-
-    @property
-    def title(self) -> str:
-        return fix_case(self.dataset.replace('_', ' '))
-
-    @property
-    def description(self) -> str:
-        category_name = self.category.title()
-        return f"{category_name} example using the {self.dataset} dataset with the {self.model} model."
-
-
-def create_category_indices() -> dict[str, Index]:
-    """Create category indices with their respective configurations."""
+def generate_examples(generate_main_index=False):
+    """
+    Generate example documentation.
+    
+    Args:
+        generate_main_index (bool): Whether to generate the main examples index.
+            If False, only category-specific indices will be generated.
+    """
+    # Create empty indices with dynamic paths
     main_index_dir = ROOT_DIR / "docs/source/examples"
     if not main_index_dir.exists():
         main_index_dir.mkdir(parents=True)
 
+    # Create the main examples index only if requested
+    examples_index = None
+    if generate_main_index:
+        examples_index = Index(
+            path=main_index_dir / "examples_index.md",
+            title="💡 Examples",
+            description=
+            "A collection of examples demonstrating usage of FastVideo.\nAll documented examples are autogenerated using <gh-file:docs/source/generate_examples.py> from examples found in <gh-file:examples>.",  # noqa: E501
+            caption="Examples",
+            maxdepth=2)
+
+    # Category indices with dynamic paths based on category names
     category_indices = {
         "inference":
         Index(
@@ -206,54 +193,28 @@ def create_category_indices() -> dict[str, Index]:
             "docs/source/inference/examples/examples_inference_index.md",
             title="🚀 Examples",
             description=
-            "Inference examples demonstrate how to use FastVideo inference. We recommend starting with <project:basic.md>.",
+            "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",
-            maxdepth=1,
-        ),
-        "training":
-        Index(
-            path=ROOT_DIR /
-            "docs/source/training/examples/examples_training_index.md",
-            title="🚀 Examples",
-            description=
-            "Training examples demonstrate how to use FastVideo training.",
-            caption="Examples",
-            maxdepth=3,
-        ),
-        "distillation":
-        Index(
-            path=ROOT_DIR /
-            "docs/source/distillation/examples/examples_distillation_index.md",
-            title="🚀 Examples",
-            description=
-            "Distillation examples demonstrate how to use FastVideo distillation.",
-            caption="Examples",
-            maxdepth=3,
         ),
     }
 
     # Ensure all category doc directories exist
-    for index in category_indices.values():
-        if not index.path.parent.exists():
-            index.path.parent.mkdir(parents=True)
+    for category, index in category_indices.items():
+        category_dir = index.path.parent
+        if not category_dir.exists():
+            category_dir.mkdir(parents=True)
 
-    return category_indices
-
-
-def find_examples(category_indices: dict[str, Index],
-                  generate_main_index: bool) -> list[Example]:
-    """Find all examples from the examples directory."""
     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 (recursively)
-        for path in category_dir.glob("**/*.md"):
+        # Find examples in subdirectories
+        for path in category_dir.glob("*/*.md"):
             examples.append(Example(path.parent, category))
 
     # Find uncategorised examples only if we're generating a main index
@@ -268,173 +229,36 @@ def find_examples(category_indices: dict[str, Index],
                 continue
             examples.append(Example(path.parent))
 
-    return examples
-
-
-def create_nested_structures(
-    examples: list[Example]
-) -> dict[str, dict[str, dict[str, dict[str, NestedStructure]]]]:
-    """Create nested structures for training and distillation categories."""
-    nested_structures: dict[str, dict[str, dict[str,
-                                                dict[str,
-                                                     NestedStructure]]]] = {}
-
-    for example in examples:
-        if example.category not in ["training", "distillation"]:
-            continue
-
-        category_dir = EXAMPLE_DIR / example.category
-        relative_path = example.path.relative_to(category_dir)
-        path_parts = relative_path.parts
-
-        # For nested examples like finetune/wan_i2v_14b_480p/crush_smol
-        if len(path_parts) >= 3:
-            method = path_parts[0]  # e.g., "finetune"
-            model = path_parts[1]  # e.g., "wan_i2v_14b_480p"
-            dataset = path_parts[2]  # e.g., "crush_smol"
-
-            # Initialize nested structure
-            if example.category not in nested_structures:
-                nested_structures[example.category] = {}
-            if method not in nested_structures[example.category]:
-                nested_structures[example.category][method] = {}
-            if model not in nested_structures[example.category][method]:
-                nested_structures[example.category][method][model] = {}
-
-            # Store the nested structure
-            nested_structures[
-                example.category][method][model][dataset] = NestedStructure(
-                    category=example.category,
-                    method=method,
-                    model=model,
-                    dataset=dataset,
-                    example=example)
-
-    return nested_structures
-
-
-def generate_flat_examples(examples: list[Example],
-                           category_indices: dict[str, Index],
-                           examples_index: Index | None,
-                           generate_main_index: bool) -> None:
-    """Generate documentation for flat structure examples (inference, etc.)."""
-    for example in examples:
-        if example.category in ["training", "distillation"]:
-            continue  # Skip nested structure examples
+    # Create document directories for each category based on category name and generate files
+    for example in sorted(examples, key=lambda e: e.path.stem):
+        print(example)
 
         # Determine which index to use for this example
         if example.category is not None and example.category in category_indices:
             index = category_indices[example.category]
         elif generate_main_index:
             assert examples_index is not None
-            index = examples_index
+            index = examples_index  # Default to main index if available
         else:
+            # Skip examples without a category if no main index
+            print(f"Skipping {example.path} (no category and no main index)")
             continue
 
-        # Generate the example documentation
+        # Place generated example markdown in the same directory as its index
         doc_path = index.path.parent / f"{example.path.stem}.md"
         with open(doc_path, "w+") as f:
             f.write(example.generate())
+        # Add the example to the index
         index.documents.append(example.path.stem)
 
-
-def generate_nested_examples(nested_structures: dict[str, dict[str, dict[
-    str, dict[str, NestedStructure]]]], category_indices: dict[str,
-                                                               Index]) -> None:
-    """Generate documentation for nested structure examples (training, distillation)."""
-    for category_name in ["training", "distillation"]:
-        if category_name not in category_indices or category_name not in nested_structures:
-            continue
-
-        category_index = category_indices[category_name]
-        category_base_dir = category_index.path.parent
-
-        for method, models in nested_structures[category_name].items():
-            # Create method-level index
-            method_index = Index(path=category_base_dir / f"{method}.md",
-                                 title=fix_case(method),
-                                 description=f"Examples using {method}.",
-                                 caption=f"{fix_case(method)} Examples",
-                                 maxdepth=2)
-
-            for model, datasets in models.items():
-                # Generate dataset examples using the Example class
-                for dataset, nested_struct in datasets.items():
-                    doc_path = category_base_dir / f"{nested_struct.filename}.md"
-                    with open(doc_path, "w+") as f:
-                        f.write(nested_struct.example.generate())
-
-                # Create model-level index
-                model_index = Index(
-                    path=category_base_dir / f"{model}.md",
-                    title=fix_case(model.replace('_', ' ')),
-                    description=f"Examples for the {model} model.",
-                    caption=f"{fix_case(model.replace('_', ' '))} Datasets",
-                    maxdepth=1)
-
-                # Add dataset indices to model index
-                for dataset, nested_struct in datasets.items():
-                    model_index.documents.append(nested_struct.filename)
-
-                # Write model index
-                with open(model_index.path, "w+") as f:
-                    f.write(model_index.generate())
-
-                # Add model to method index
-                method_index.documents.append(model)
-
-            # Write method index
-            with open(method_index.path, "w+") as f:
-                f.write(method_index.generate())
-
-            # Add method to main category index
-            category_index.documents.append(method)
-
-
-def generate_examples(generate_main_index=False):
-    """
-    Generate example documentation.
-    
-    Args:
-        generate_main_index (bool): Whether to generate the main examples index.
-            If False, only category-specific indices will be generated.
-    """
-    # Create category indices
-    category_indices = create_category_indices()
-
-    # Create the main examples index only if requested
-    examples_index = None
-    if generate_main_index:
-        main_index_dir = ROOT_DIR / "docs/source/examples"
-        examples_index = Index(
-            path=main_index_dir / "examples_index.md",
-            title="💡 Examples",
-            description=
-            "A collection of examples demonstrating usage of FastVideo.\nAll documented examples are autogenerated using <gh-file:docs/source/generate_examples.py> from examples found in <gh-file:examples>.",
-            caption="Examples",
-            maxdepth=2)
-
-    # Find all examples
-    examples = find_examples(category_indices, generate_main_index)
-
-    # Create nested structures for training and distillation
-    nested_structures = create_nested_structures(examples)
-
-    # Generate flat structure examples (inference, etc.)
-    generate_flat_examples(examples, category_indices, examples_index,
-                           generate_main_index)
-
-    # Generate nested structure examples (training, distillation)
-    generate_nested_examples(nested_structures, category_indices)
-
     # Generate the index files for categories
     for category_index in category_indices.values():
         if category_index.documents:
             # Add to main index if it exists
-            if generate_main_index and examples_index:
-                main_index_dir = examples_index.path.parent
+            if generate_main_index:
                 rel_path = category_index.path.relative_to(
                     main_index_dir.parent)
+                assert examples_index is not None
                 examples_index.documents.insert(
                     0,
                     str(rel_path).replace(".md", ""))

docs/source/getting_started/installation.md

@@ -1,18 +1,120 @@
-(installation-index)=
+(fastvideo-installation)=
 
 # 🔧 Installation
 
-FastVideo supports the following hardware platforms:
+FastVideo currently only supports Linux and NVIDIA CUDA GPUs.
 
-:::{toctree}
-:maxdepth: 1
-:hidden:
+## Requirements
 
-installation/gpu
-installation/mps
+- **OS: Linux**
+- **Python: 3.10-3.12**
+- **CUDA 12.4**
+- **At least 1 NVIDIA GPU**
+
+## Set up using Python
+### Create a new Python environment
+
+#### Conda
+You can create a new python environment using [Conda](https://docs.conda.io/projects/conda/en/stable/user-guide/getting-started.html)
+##### 1. Install Miniconda (if not already installed)
+
+```bash
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
+bash Miniconda3-latest-Linux-x86_64.sh
+source ~/.bashrc
+```
+
+##### 2. Create and activate a Conda environment for FastVideo
+
+```bash
+# (Recommended) Create a new conda environment.
+conda create -n fastvideo python=3.12 -y
+conda activate fastvideo
+```
+
+:::{note}
+[PyTorch has deprecated the conda release channel](https://github.com/pytorch/pytorch/issues/138506). If you use `conda`, please only use it to create Python environment rather than installing packages.
 :::
 
-- <project:installation/gpu.md>
-  - NVIDIA CUDA
-- <project:installation/mps.md>
-  - Apple silicon
+#### uv
+
+:::{tip}
+We highly recommend using `uv` to install FastVideo. In our experience, `uv` speeds up installation by at least 3x.
+:::
+
+Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
+
+```console
+# (Recommended) Create a new uv environment. Use `--seed` to install `pip` and `setuptools` in the environment.
+uv venv --python 3.12 --seed
+source .venv/bin/activate
+```
+
+### Installation
+
+```bash
+pip install fastvideo
+
+# or if you are using uv
+uv pip install fastvideo
+```
+
+Also optionally install flash-attn:
+
+```bash
+pip install flash-attn==2.7.4.post1 --no-build-isolation
+```
+
+### Installation from Source
+
+#### 1. Clone the FastVideo repository
+
+```bash
+git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
+```
+
+#### 2. Install FastVideo
+
+Basic installation:
+
+```bash
+pip install -e .
+
+# or if you are using uv
+uv pip install -e .
+```
+
+### Optional Dependencies
+
+#### Flash Attention
+
+```bash
+pip install flash-attn==2.7.4.post1 --no-build-isolation
+```
+
+## Set up using Docker
+We also have prebuilt docker images with FastVideo dependencies pre-installed:
+[Docker Images](#docker)
+
+## Development Environment Setup
+
+If you're planning to contribute to FastVideo please see the following page:
+[Contributor Guide](#developer-overview)
+
+## Hardware Requirements
+
+### For Basic Inference
+- NVIDIA GPU with CUDA 12.4 support
+
+### For Lora Finetuning
+- 40GB GPU memory each for 2 GPUs with lora
+- 30GB GPU memory each for 2 GPUs with CPU offload and lora
+
+### For Full Finetuning/Distillation
+- Multiple high-memory GPUs recommended (e.g., H100)
+
+## Troubleshooting
+
+If you encounter any issues during installation, please open an issue on our [GitHub repository](https://github.com/hao-ai-lab/FastVideo).
+
+You can also join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ) for additional support.

docs/source/getting_started/installation/gpu.md

@@ -1,118 +0,0 @@
-# NVIDIA GPU
-
-Instructions to install FastVideo for NVIDIA CUDA GPUs.
-
-## Requirements
-
-- **OS: Linux or Windows WSL**
-- **Python: 3.10-3.12**
-- **CUDA 12.4**
-- **At least 1 NVIDIA GPU**
-
-## Set up using Python
-### Create a new Python environment
-
-#### Conda
-You can create a new python environment using [Conda](https://docs.conda.io/projects/conda/en/stable/user-guide/getting-started.html)
-##### 1. Install Miniconda (if not already installed)
-
-```bash
-wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
-bash Miniconda3-latest-Linux-x86_64.sh
-source ~/.bashrc
-```
-
-##### 2. Create and activate a Conda environment for FastVideo
-
-```bash
-# (Recommended) Create a new conda environment.
-conda create -n fastvideo python=3.12 -y
-conda activate fastvideo
-```
-
-:::{note}
-[PyTorch has deprecated the conda release channel](https://github.com/pytorch/pytorch/issues/138506). If you use `conda`, please only use it to create Python environment rather than installing packages.
-:::
-
-#### uv
-
-:::{tip}
-We highly recommend using `uv` to install FastVideo. In our experience, `uv` speeds up installation by at least 3x.
-:::
-
-Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
-
-```console
-# (Recommended) Create a new uv environment. Use `--seed` to install `pip` and `setuptools` in the environment.
-uv venv --python 3.12 --seed
-source .venv/bin/activate
-```
-
-### Installation
-
-```bash
-pip install fastvideo
-
-# or if you are using uv
-uv pip install fastvideo
-```
-
-Also optionally install flash-attn:
-
-```bash
-pip install flash-attn==2.7.4.post1 --no-build-isolation
-```
-
-### Installation from Source
-
-#### 1. Clone the FastVideo repository
-
-```bash
-git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
-```
-
-#### 2. Install FastVideo
-
-Basic installation:
-
-```bash
-pip install -e .
-
-# or if you are using uv
-uv pip install -e .
-```
-
-### Optional Dependencies
-
-#### Flash Attention
-
-```bash
-pip install flash-attn==2.7.4.post1 --no-build-isolation
-```
-
-## Set up using Docker
-We also have prebuilt docker images with FastVideo dependencies pre-installed:
-[Docker Images](#docker)
-
-## Development Environment Setup
-
-If you're planning to contribute to FastVideo please see the following page:
-[Contributor Guide](#developer-overview)
-
-## Hardware Requirements
-
-### For Basic Inference
-- NVIDIA GPU with CUDA 12.4 support
-
-### For Lora Finetuning
-- 40GB GPU memory each for 2 GPUs with lora
-- 30GB GPU memory each for 2 GPUs with CPU offload and lora
-
-### For Full Finetuning/Distillation
-- Multiple high-memory GPUs recommended (e.g., H100)
-
-## Troubleshooting
-
-If you encounter any issues during installation, please open an issue on our [GitHub repository](https://github.com/hao-ai-lab/FastVideo).
-
-You can also join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ) for additional support.

docs/source/getting_started/installation/mps.md

@@ -1,101 +0,0 @@
-# MPS (Apple Silicon)
-
-Instructions to install FastVideo for Apple Silicon.
-
-## Requirements
-
-- **OS: MacOS**
-- **Python: 3.12.4**
-
-## Set up using Python
-
-### Create a new Python environment
-
-#### Conda
-
-You can create a new python environment using [Conda](https://docs.conda.io/projects/conda/en/stable/user-guide/getting-started.html)
-
-##### 1. Install Miniconda (if not already installed)
-
-```bash
-wget https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh
-bash Miniconda3-latest-MacOSX-arm64.sh
-source ~/.zshrc
-```
-
-##### 2. Create and activate a Conda environment for FastVideo
-
-```bash
-# (Recommended) Create a new conda environment.
-conda create -n fastvideo python=3.12.4 -y
-conda activate fastvideo
-```
-
-:::{note}
-[PyTorch has deprecated the conda release channel](https://github.com/pytorch/pytorch/issues/138506). If you use `conda`, please only use it to create Python environment rather than installing packages.
-:::
-
-#### uv
-
-:::{tip}
-We highly recommend using `uv` to install FastVideo. In our experience, `uv` speeds up installation by at least 3x.
-:::
-
-Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
-
-```console
-# (Recommended) Create a new uv environment. Use `--seed` to install `pip` and `setuptools` in the environment.
-uv venv --python 3.12 --seed
-source .venv/bin/activate
-```
-
-### Dependencies
-
-```
-brew install ffmpeg
-```
-
-### Installation
-
-```bash
-pip install fastvideo
-
-# or if you are using uv
-uv pip install fastvideo
-```
-
-### Installation from Source
-
-#### 1. Clone the FastVideo repository
-
-```bash
-git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
-```
-
-#### 2. Install FastVideo
-
-Basic installation:
-
-```bash
-pip install -e .
-
-# or if you are using uv
-uv pip install -e .
-```
-
-## Development Environment Setup
-
-If you're planning to contribute to FastVideo please see the following page:
-[Contributor Guide](#developer-overview)
-
-## Hardware Requirements
-
-### For Basic Inference
-
-- Mac M1, M2, M3, or M4 (at least 32 GB RAM is preferable for high quality video generation)
-
-## 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-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ) for additional support.

docs/source/getting_started/v1_api.md

@@ -10,20 +10,20 @@ This class will be the primary Python API for generating videos and images.
     fastvideo.VideoGenerator
 ```
 
-`````{py:class} VideoGenerator(fastvideo_args: fastvideo.fastvideo_args.FastVideoArgs, executor_class: type[fastvideo.worker.executor.Executor], log_stats: bool)
-:canonical: fastvideo.entrypoints.video_generator.VideoGenerator
+`````{py:class} VideoGenerator(fastvideo_args: fastvideo.v1.fastvideo_args.FastVideoArgs, executor_class: type[fastvideo.v1.worker.executor.Executor], log_stats: bool)
+:canonical: fastvideo.v1.entrypoints.video_generator.VideoGenerator
 
-```{autodoc2-docstring} fastvideo.entrypoints.video_generator.VideoGenerator
+```{autodoc2-docstring} fastvideo.v1.entrypoints.video_generator.VideoGenerator
 :parser: docs.source.autodoc2_docstring_parser
 ```
 
 `VideoGenerator.from_pretrained()` should be the primary way of creating a new video generator.
 
-````{py:method} from_pretrained(model_path: str, device: typing.Optional[str] = None, torch_dtype: typing.Optional[torch.dtype] = None, pipeline_config: typing.Optional[typing.Union[str | fastvideo.configs.pipelines.PipelineConfig]] = None, **kwargs) -> fastvideo.entrypoints.video_generator.VideoGenerator
-:canonical: fastvideo.entrypoints.video_generator.VideoGenerator.from_pretrained
+````{py:method} from_pretrained(model_path: str, device: typing.Optional[str] = None, torch_dtype: typing.Optional[torch.dtype] = None, pipeline_config: typing.Optional[typing.Union[str | fastvideo.v1.configs.pipelines.PipelineConfig]] = None, **kwargs) -> fastvideo.v1.entrypoints.video_generator.VideoGenerator
+:canonical: fastvideo.v1.entrypoints.video_generator.VideoGenerator.from_pretrained
 :classmethod:
 
-```{autodoc2-docstring} fastvideo.entrypoints.video_generator.VideoGenerator.from_pretrained
+```{autodoc2-docstring} fastvideo.v1.entrypoints.video_generator.VideoGenerator.from_pretrained
 :parser: docs.source.autodoc2_docstring_parser
 ```
 
@@ -38,25 +38,25 @@ The follow two classes `PipelineConfig` and `SamplingParam` are used to configur
 ```
 
 `````{py:class} PipelineConfig
-:canonical: fastvideo.configs.pipelines.base.PipelineConfig
+:canonical: fastvideo.v1.configs.pipelines.base.PipelineConfig
 
-```{autodoc2-docstring} fastvideo.configs.pipelines.base.PipelineConfig
+```{autodoc2-docstring} fastvideo.v1.configs.pipelines.base.PipelineConfig
 :parser: docs.source.autodoc2_docstring_parser
 ```
 
-````{py:method} from_pretrained(model_path: str) -> fastvideo.configs.pipelines.base.PipelineConfig
-:canonical: fastvideo.configs.pipelines.base.PipelineConfig.from_pretrained
+````{py:method} from_pretrained(model_path: str) -> fastvideo.v1.configs.pipelines.base.PipelineConfig
+:canonical: fastvideo.v1.configs.pipelines.base.PipelineConfig.from_pretrained
 :classmethod:
 
-```{autodoc2-docstring} fastvideo.configs.pipelines.base.PipelineConfig.from_pretrained
+```{autodoc2-docstring} fastvideo.v1.configs.pipelines.base.PipelineConfig.from_pretrained
 :parser: docs.source.autodoc2_docstring_parser
 ```
 
 
 ````{py:method} dump_to_json(file_path: str)
-:canonical: fastvideo.configs.pipelines.base.PipelineConfig.dump_to_json
+:canonical: fastvideo.v1.configs.pipelines.base.PipelineConfig.dump_to_json
 
-```{autodoc2-docstring} fastvideo.configs.pipelines.base.PipelineConfig.dump_to_json
+```{autodoc2-docstring} fastvideo.v1.configs.pipelines.base.PipelineConfig.dump_to_json
 :parser: docs.source.autodoc2_docstring_parser
 ```
 
@@ -68,16 +68,16 @@ The follow two classes `PipelineConfig` and `SamplingParam` are used to configur
 ```
 
 `````{py:class} SamplingParam
-:canonical: fastvideo.configs.sample.base.SamplingParam
+:canonical: fastvideo.v1.configs.sample.base.SamplingParam
 
-```{autodoc2-docstring} fastvideo.configs.sample.base.SamplingParam
+```{autodoc2-docstring} fastvideo.v1.configs.sample.base.SamplingParam
 :parser: docs.source.autodoc2_docstring_parser
 ```
 
-````{py:method} from_pretrained(model_path: str) -> fastvideo.configs.sample.base.SamplingParam
-:canonical: fastvideo.configs.sample.base.SamplingParam.from_pretrained
+````{py:method} from_pretrained(model_path: str) -> fastvideo.v1.configs.sample.base.SamplingParam
+:canonical: fastvideo.v1.configs.sample.base.SamplingParam.from_pretrained
 :classmethod:
 
-```{autodoc2-docstring} fastvideo.configs.sample.base.SamplingParam.from_pretrained
+```{autodoc2-docstring} fastvideo.v1.configs.sample.base.SamplingParam.from_pretrained
 :parser: docs.source.autodoc2_docstring_parser
 ```

docs/source/index.md

@@ -63,33 +63,22 @@ getting_started/installation
 :maxdepth: 1
 
 inference/inference_quick_start
-inference/examples/examples_inference_index
 inference/configuration
 inference/optimizations
-inference/comfyui
 inference/support_matrix
+inference/examples/examples_inference_index
 inference/cli
 inference/add_pipeline
+inference/v0_inference
 :::
 
 :::{toctree}
 :caption: Training
 :maxdepth: 1
 
-training/examples/examples_training_index
 training/data_preprocess
 training/distillation
-<!-- training/finetune -->
-:::
-
-<!-- :::{toctree}
-:caption: Distillation
-:maxdepth: 1
-
-distillation/examples/examples_distillation_index
-distillation/data_preprocess
-distillation/dmd -->
-<!-- training/finetune -->
+training/finetune
 :::
 
 % What is STA Kernel?

docs/source/inference/add_pipeline.md

@@ -46,25 +46,25 @@ FastVideo uses the Hugging Face Diffusers format for model organization:
 ### Implementing Modules
 
 Place new modules in the appropriate directories:
-- Encoders: `fastvideo/models/encoders/`
-- VAEs: `fastvideo/models/vaes/`
-- Transformer models: `fastvideo/models/dits/`
-- Schedulers: `fastvideo/models/schedulers/`
+- Encoders: `fastvideo/v1/models/encoders/`
+- VAEs: `fastvideo/v1/models/vaes/`
+- Transformer models: `fastvideo/v1/models/dits/`
+- Schedulers: `fastvideo/v1/models/schedulers/`
 
 ### Adapting Model Layers
 
 #### Layer Replacements
 Replace standard PyTorch layers with FastVideo optimized versions:
-- nn.LayerNorm → fastvideo.layers.layernorm.RMSNorm
-- Embedding layers → fastvideo.layers.vocab_parallel_embedding modules
-- Activation functions → versions from fastvideo.layers.activation
+- nn.LayerNorm → fastvideo.v1.layers.layernorm.RMSNorm
+- Embedding layers → fastvideo.v1.layers.vocab_parallel_embedding modules
+- Activation functions → versions from fastvideo.v1.layers.activation
 
 #### Distributed Linear Layers
 Use appropriate parallel layers for distribution:
 
 ```python
 # Output dimension parallelism
-from fastvideo.layers.linear import ColumnParallelLinear
+from fastvideo.v1.layers.linear import ColumnParallelLinear
 self.q_proj = ColumnParallelLinear(
     input_size=hidden_size,
     output_size=head_size * num_heads,
@@ -73,7 +73,7 @@ self.q_proj = ColumnParallelLinear(
 )
 
 # Fused QKV projection
-from fastvideo.layers.linear import QKVParallelLinear
+from fastvideo.v1.layers.linear import QKVParallelLinear
 self.qkv_proj = QKVParallelLinear(
     hidden_size=hidden_size,
     head_size=attention_head_dim,
@@ -82,7 +82,7 @@ self.qkv_proj = QKVParallelLinear(
 )
 
 # Input dimension parallelism
-from fastvideo.layers.linear import RowParallelLinear
+from fastvideo.v1.layers.linear import RowParallelLinear
 self.out_proj = RowParallelLinear(
     input_size=head_size * num_heads,
     output_size=hidden_size,
@@ -96,8 +96,8 @@ Replace standard attention with FastVideo's optimized attention:
 
 ```python
 # Local attention patterns
-from fastvideo.attention import LocalAttention
-from fastvideo.attention.backends.abstract import _Backend
+from fastvideo.v1.attention import LocalAttention
+from fastvideo.v1.attention.backends.abstract import _Backend
 self.attn = LocalAttention(
     num_heads=num_heads,
     head_size=head_dim,
@@ -108,7 +108,7 @@ self.attn = LocalAttention(
 )
 
 # Distributed attention for long sequences
-from fastvideo.attention import DistributedAttention
+from fastvideo.v1.attention import DistributedAttention
 self.attn = DistributedAttention(
     num_heads=num_heads,
     head_size=head_dim,
@@ -130,7 +130,7 @@ self.attn = DistributedAttention(
 Register implemented modules in the model registry:
 
 ```python
-# In fastvideo/models/registry.py
+# In fastvideo/v1/models/registry.py
 _TEXT_TO_VIDEO_DIT_MODELS = {
     "YourTransformerModel": ("dits", "yourmodule", "YourTransformerClass"),
 }
@@ -145,7 +145,7 @@ _VAE_MODELS = {
 Create a new directory for your pipeline:
 
 ```
-fastvideo/pipelines/
+fastvideo/v1/pipelines/
 ├── your_pipeline/
 │   ├── __init__.py
 │   └── your_pipeline.py
@@ -167,13 +167,13 @@ Pipelines are composed of stages, each handling a specific part of the diffusion
 ### Creating Your Pipeline
 
 ```python
-from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase
-from fastvideo.pipelines.stages import (
+from fastvideo.v1.pipelines.composed_pipeline_base import ComposedPipelineBase
+from fastvideo.v1.pipelines.stages import (
     InputValidationStage, CLIPTextEncodingStage, TimestepPreparationStage,
     LatentPreparationStage, DenoisingStage, DecodingStage
 )
-from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.v1.fastvideo_args import FastVideoArgs
+from fastvideo.v1.pipelines.pipeline_batch_info import ForwardBatch
 import torch
 
 class MyCustomPipeline(ComposedPipelineBase):
@@ -246,7 +246,7 @@ EntryClass = MyCustomPipeline
 If existing stages don't meet your needs, create custom ones:
 
 ```python
-from fastvideo.pipelines.stages.base import PipelineStage
+from fastvideo.v1.pipelines.stages.base import PipelineStage
 
 class MyCustomStage(PipelineStage):
     """Custom processing stage for the pipeline."""
@@ -305,7 +305,7 @@ EntryClass = [MyCustomPipeline, MyOtherPipeline]
 ```
 
 The registry will automatically:
-1. Scan all packages under `fastvideo/pipelines/`
+1. Scan all packages under `fastvideo/v1/pipelines/`
 2. Look for `EntryClass` variables
 3. Register pipelines using their class names as identifiers
 

docs/source/inference/comfyui.md

@@ -1,5 +0,0 @@
-# FastVideo + ComfyUI
-
-FastVideo provides a custom node suite for ComfyUI.
-
-See this [README](https://github.com/hao-ai-lab/FastVideo/tree/main/comfyui) for instructions.

docs/source/inference/configuration.md

@@ -27,7 +27,7 @@ def main():
     model_name = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
     config = PipelineConfig.from_pretrained(model_name)
     config.vae_precision = "fp16"
-    config.dit_cpu_offload = True
+    config.use_cpu_offload = True
 
     # Create the generator
     generator = VideoGenerator.from_pretrained(

docs/source/inference/v0_inference.md

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

docs/source/training/data_preprocess.md

@@ -7,7 +7,7 @@ To save GPU memory, we precompute text embeddings and VAE latents to eliminate t
 We provide a sample dataset to help you get started. Download the source media using the following command:
 
 ```bash
-python scripts/huggingface/download_hf.py --repo_id=FastVideo/mini_i2v_dataset --local_dir=data/mini_i2v_dataset --repo_type=dataset
+python scripts/huggingface/download_hf.py --repo_id=FastVideo/mini_i2v_dataset --local_dir=FastVideo/mini_i2v_dataset --repo_type=dataset
 ```
 
 The folder `crush-smol_raw/` contains raw videos and captions for testing preprocessing, while `crush-smol_preprocessed/` contains latents prepared for testing training.

examples/distill/Wan-Syn-480P/README.md

@@ -1,19 +0,0 @@
-# Wan2.1-T2V-1.3B Distill Example
-These are end-to-end example scripts for distilling Wan2.1 T2V 1.3B model using DMD-only and DMD+VSA methods.
-
-### 1. Download dataset:
-```bash
-bash examples/distill/Wan-Syn-480P/download_dataset.sh
-```
-
-### 2. Configure and run distillation:
-
-#### For DMD-only distillation:
-```bash
-sbatch examples/distill/Wan-Syn-480P/distill_dmd_t2v.slurm
-```
-
-#### For DMD+VSA distillation:
-```bash
-sbatch examples/distill/Wan-Syn-480P/distill_dmd_VSA_t2v.slurm
-```

examples/distill/Wan-Syn-480P/distill_dmd_VSA_t2v.slurm

@@ -1,137 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=t2v
-#SBATCH --partition=main
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=dmd_t2v_output/t2v_%j.out
-#SBATCH --error=dmd_t2v_output/t2v_%j.err
-#SBATCH --exclusive
-set -e -x
-
-# Environment Setup
-source ~/conda/miniconda/bin/activate
-conda activate your_env
-
-# Basic Info
-export WANDB_MODE="online"
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-# different cache dir for different processes
-export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29500
-export NODE_RANK=$SLURM_PROCID
-nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
-export MASTER_ADDR=${nodes[0]}
-export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
-export TOKENIZERS_PARALLELISM=false
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
-
-# Configs
-NUM_GPUS=8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR=your_data_dir
-VALIDATION_DATASET_FILE=your_validation_dataset_file
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir="checkpoints/wan_t2v_finetune"
-  --max_train_steps=4000
-  --train_batch_size=1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps=1
-  --num_latent_t 16
-  --num_height 448
-  --num_width 832
-  --num_frames 61
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 64
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR"
-  --dataloader_num_workers 4
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 200
-  --validation_sampling_steps "3"
-  --validation_guidance_scale "1.0" # not used for dmd inference
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate=1e-5
-  --mixed_precision="bf16"
-  --checkpointing_steps=500
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --allow_tf32
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.0
-  --dit_precision "bf16"
-  --ema_start_step 0
-  --flow_shift 8
-  --seed 1000
-)
-
-# DMD arguments
-dmd_args=(
-  --dmd_denoising_steps '1000,757,522'
-  --min_timestep_ratio 0.02
-  --max_timestep_ratio 0.98
-  --generator_update_interval 5
-  --real_score_guidance_scale 3.5
-  --VSA_sparsity 0.8
-)
-
-srun torchrun \
---nnodes $SLURM_JOB_NUM_NODES \
---nproc_per_node $NUM_GPUS \
---node_rank $SLURM_PROCID \
---rdzv_backend=c10d \
---rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/wan_training_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"

examples/distill/Wan-Syn-480P/distill_dmd_t2v.slurm

@@ -1,136 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=t2v
-#SBATCH --partition=main
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=dmd_t2v_output/t2v_%j.out
-#SBATCH --error=dmd_t2v_output/t2v_%j.err
-#SBATCH --exclusive
-set -e -x
-
-# Environment Setup
-source ~/conda/miniconda/bin/activate
-conda activate your_env
-
-# Basic Info
-export WANDB_MODE="online"
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-# different cache dir for different processes
-export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29500
-export NODE_RANK=$SLURM_PROCID
-nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
-export MASTER_ADDR=${nodes[0]}
-export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
-export TOKENIZERS_PARALLELISM=false
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
-
-# Configs
-NUM_GPUS=8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR=your_data_dir
-VALIDATION_DATASET_FILE=your_validation_dataset_file
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name wan_t2v_distill_dmd
-  --output_dir="checkpoints/wan_t2v_finetune"
-  --max_train_steps=4000
-  --train_batch_size=1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps=1
-  --num_latent_t 16
-  --num_height 448
-  --num_width 832
-  --num_frames 61
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 64
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR"
-  --dataloader_num_workers 4
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 200
-  --validation_sampling_steps "3"
-  --validation_guidance_scale "1.0" # not used for dmd inference
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate=1e-5
-  --mixed_precision="bf16"
-  --checkpointing_steps=500
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --allow_tf32
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.0
-  --dit_precision "bf16"
-  --ema_start_step 0
-  --flow_shift 8
-  --seed 1000
-)
-
-# DMD arguments
-dmd_args=(
-  --dmd_denoising_steps '1000,757,522'
-  --min_timestep_ratio 0.02
-  --max_timestep_ratio 0.98
-  --generator_update_interval 5
-  --real_score_guidance_scale 3.5
-)
-
-srun torchrun \
---nnodes $SLURM_JOB_NUM_NODES \
---nproc_per_node $NUM_GPUS \
---node_rank $SLURM_PROCID \
---rdzv_backend=c10d \
---rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/wan_training_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"

examples/distill/Wan-Syn-480P/download_dataset.sh

@@ -1,3 +0,0 @@
-#!/bin/bash
-
-python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"

examples/distill/Wan-Syn-480P/validation_64.json

@@ -1,516 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "In the video, a woman is elegantly showcasing her earrings, bringing attention to their intricate design with a gentle touch of her fingers. She is bathed in ambient purple and pink lighting, which casts a soft glow on her delicate features and enhances the vivid tones of her lipstick and eye makeup. Her hair is styled to frame her face smoothly, emphasizing the contours of her jawline and cheekbones. The background features a blurred neon light, adding an artistic and modern touch to the overall aesthetic.",
-      "video_path": "Fashion/mixkit-face-of-an-elegant-and-captivating-woman-41914_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the video, a lone rider guides a majestic horse across an expansive, open field as the sun sets in the background. The rider, dressed in a classic blue shirt and wide-brimmed hat, sits confidently in the saddle, silhouetted against the warm glow of the evening sky. The horse moves gracefully, its mane and tail flowing with each step, creating a sense of harmony between horse and rider. Surrounding the pair, towering trees form a natural border, their leaves gently rustling in the breeze. The shadows lengthen on the ground, accentuating the serene and timeless feel of the scene. The distant hills and wooden fences frame the horizon, adding depth to the tranquil landscape. A few horses graze peacefully in the background, blending into the pastoral setting. The overall ambiance evokes a sense of calmness and quietude, capturing a perfect moment in the golden light of dusk.",
-      "video_path": "Man/mixkit-a-rancher-riding-a-horse-at-sunset-1143_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In a dimly lit, eerie setting, a mysterious pink bottle labeled \"Authentic 100% organic POISON\" sits prominently in the foreground, casting a menacing aura. The bottle is accentuated by green fog, which swirls lightly around it, enhancing its sinister allure. Behind it, a shadowy golden bottle adorned with a spider emblem subtly emerges, adding an extra layer of mystery to the scene. Dim candles provide faint, flickering light, which complements the dark atmosphere, making the setting ideal for an illusion of hidden dangers.",
-      "video_path": "smoke/mixkit-poison-in-halloween-ritual-33879_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "The video opens with a tranquil scene in the heart of a dense forest, emphasizing two large, textured tree trunks in the foreground framing the view. Sunlight filters through the canopy above, casting intricate patterns of light and shadow on the trees and the ground. Between the tree trunks, a clear view of a calm, muddy river unfolds, its surface shimmering under the gentle sunlight. The riverbank is decorated with a variety of small bushes and vibrant foliage, subtly transitioning into the deep greens of tall, leafy plants. In the background, the dense forest looms, filled with dark, towering trees, their branches intertwining to form an intricate canopy. The scene is bathed in the soft glow of the sun, creating a serene and picturesque setting. Occasional sunbeams pierce through the foliage, adding a magical aura to the landscape. The vibrant reds and oranges of the smaller plants add contrast, bringing warmth to the earthy tones of the scenery. Overall, this harmonious blend of natural elements creates a peaceful and idyllic forest setting.",
-      "video_path": "forest/mixkit-view-of-a-river-between-two-old-trees-560_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the video, a martial artist dressed in a traditional white uniform with a black belt demonstrates a series of precise movements against a stark black background. The individual gracefully transitions between stances, embodying a sense of focused discipline and control. Each motion is executed with a deliberate pace, showcasing the fluidity of martial arts techniques. The soft lighting creates subtle highlights on the uniform, adding depth to the figure as it moves. The practitioner begins with an open-hand pose, feet firmly grounded, gradually shifting to a powerful forward punch. The fluidity of the sequence displays a mastery of balance and poise. Every trajectory of the limbs is precise and deliberate, capturing the elegance and strength of martial arts. The serene, isolated setting enhances the intensity and concentration of the practitioner. This visual presentation is an elegant interplay of motion and stillness, displaying the art form's discipline and grace.",
-      "video_path": "Man/mixkit-a-young-man-practicing-his-karate-moves-49635_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A tranquil coastal scene unfolds with a drone's aerial view capturing a serene beach landscape. The camera glides over a quiet stretch of sandy shoreline, where gentle waves kiss the shore under a clear blue sky. Nestled amidst lush palm trees are a series of traditional thatched-roof huts, their earthy tones blending harmoniously with the natural surroundings. The sandy beach stretches endlessly, bordered by the rhythmic dance of ocean waves on one side and verdant greenery on the other. A pair of white umbrellas is set up on the sand, suggesting a place to relax and enjoy the sun. In the distance, two small human figures can be seen walking leisurely along the water's edge, leaving faint footprints behind them. The scene exudes a calm and inviting atmosphere, with the soft rustle of palm leaves and the whisper of the ocean breeze almost audible. The overall composition is a captivating blend of nature's tranquility and architectural simplicity. This picturesque setting invites viewers to imagine themselves steps away from this idyllic coastal escape.",
-      "video_path": "beach/mixkit-sunny-beach-in-a-dynamic-shot-from-a-drone-44383_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A lone figure stands on a large, moss-covered rock, surrounded by the soft rush of a nearby stream. The figure is wearing white sneakers and shorts, with a plaid shirt that hangs loosely in the breeze. The lighting creates dramatic shadows, enhancing the textures of the rock and the subtle movement of the water below. In the background, a waterfall cascades into the stream, completing this tranquil and serene nature scene.",
-      "video_path": "forest/mixkit-woman-standing-in-front-of-waterfall-559_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In an industrial setting, a person leans casually against a railing, exuding a sense of confidence and composure. They are wearing a striking outfit, consisting of a vibrant, patterned jacket over a simple white crop top, creating a bold contrast. The atmosphere is infused with warm, ambient lighting that casts soft shadows on the concrete walls and metallic surfaces. Intricate wiring and pipes form an intricate backdrop, enhancing the urban aesthetic. Their relaxed posture and direct, engaging gaze suggest a sense of ease in this industrial environment. This scene encapsulates a blend of modern fashion and gritty, urban architecture, creating a visually compelling narrative.",
-      "video_path": "Fashion/mixkit-portrait-of-a-hipster-woman-walking-down-a-stairs-1297_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A man is energetically stretching in an open-air setting, surrounded by rows of vibrant red seats that suggest an amphitheater or outdoor venue. He wears a sleeveless black shirt layered with a hooded vest, emphasizing his athletic build as he engages in a warm-up routine. Behind him, the striking modern architecture of the building features geometric panels, with large sections of glass and overlapping metallic beams creating a dynamic backdrop. The scene captures the contrast between his focused movements and the static, bold design of the structure, while the surrounding greenery adds a touch of nature to the environment. The overall atmosphere is one of preparation and anticipation, with the man appearing determined and ready for an upcoming event or performance.",
-      "video_path": "Sport/mixkit-man-doing-arm-stretches-595_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A young woman is seated on the floor in front of a plush, beige tufted couch, fully engrossed in sorting through a stack of papers. Her dark hair falls loosely past her shoulders, and she wears a green plaid shirt, contributing to the casual yet focused atmosphere. She gently places the papers onto a small round white table, occasionally lifting individual sheets to examine them more closely. Her expression shifts subtly, reflecting concentration and contemplation as she processes the information on the pages. Two small, round nested tables hold her documents, along with a small plant in a gray pot, adding a touch of greenery to the scene. The background features a dark paneled wall, creating a contrasting backdrop for the light-colored furniture. The setting is tranquil and organized, the couch and tables arranged symmetrically, conveying a sense of harmony. A calculator rests on the smaller table, hinting at a task involving calculations or budgeting.",
-      "video_path": "Woman/mixkit-frustrated-woman-throws-paperwork-on-the-floor-4526_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A heavily rusted metal gate stands firmly locked, with two vertical bars joined by a thick, old chain that loops elegantly around them. The chain's texture is coarse and rugged, its surface reflecting varying shades of orange and brown, indicative of years exposed to the elements. At the heart of the chain, a black iron padlock, slightly worn yet imposing, secures the gate, its curves and edges smooth against the aged links. The gate's metalwork is outlined by a backdrop of soft, blurred greenery, suggesting a serene and isolated location beyond the barrier. Tall trees rise in the distance, their trunks and leaves creating a lush, forest-like setting that contrasts with the gate's severe rust. A pathway leads away from the gate, its surface uneven with patches of moss and weathered stone visible in the soft focus, inviting yet inaccessible. The ambiance is quiet and mysterious, with a sense of abandonment hanging subtly in the air, evoking curiosity about what lies beyond. Shadows play across the gate, cast by branches swaying gently in the breeze, adding to the dynamic interaction of light and texture. This scene, rich in detail and atmosphere, captures the viewer's imagination, evoking both the allure of the forbidden and the beauty of decay.",
-      "video_path": "forest/mixkit-rusty-fence-with-a-chain-of-a-property-in-nature-5294_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In a serene and softly lit yoga studio, three individuals engage in a yoga session, each performing an upward-facing stretch. The central figure is a woman with shoulder-length brown hair, dressed in a light cropped top and green leggings, her posture reflecting grace and concentration. To her right, another participant, a woman in a purple outfit, mirrors the pose with equal poise. On her left, a person with a bun focuses intently, supported slightly by yoga blocks beneath their hands. The warm-colored wooden floor contrasts soothingly with the soft pastel mural on the back wall, featuring an abstract design and partial visage of a serene face. Natural light floods the space from a large window on the right, where lush greens peek through, adding an element of tranquility. In the corner of the room, a collection of meditation instruments, including a gong and a Buddha statue, subtly frame the peaceful setting. The mood is calm yet focused, as all three participants are deeply engaged in their practice. The scene combines elements of balance, harmony, and a shared journey towards mindfulness. This depiction captures the essence of a yoga session that blends personal growth with collective experience.",
-      "video_path": "People/mixkit-small-group-of-people-doing-yoga-together-43730_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the deep blue expanse of the ocean, two dolphins glide effortlessly, their sleek bodies reflecting the sunlight filtering through the water. The prominent shadows and caustics create a shimmering effect on their skin, capturing the beauty of their natural habitat. Each dolphin moves with a fluid grace, occasionally interacting with gentle nudges, showcasing their playful and social nature. The scene is vibrant and dynamic, with the clear blue background accentuating the dolphins' movements, making it an ideal subject for AI recreation.",
-      "video_path": "sea/mixkit-dolphins-underwater-4133_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the video, a young woman stands against a vibrant graffiti-covered wall, deeply engrossed in her smartphone. Her expression reflects a mix of focus and subtle satisfaction as she interacts with the screen. She wears a black floral-patterned top, which contrasts with the bright, abstract shapes and bold colors of the mural behind her. As she continues to engage with her phone, a series of like count notifications appear on the screen, indicating a growing online appreciation. The wall behind her features a striking mix of geometric and organic shapes, including swirls of teal, orange, and black, with large humanoid figures in a pop-art style. Her long, light-brown hair frames her face, adding a calm, composed aura amidst the lively backdrop. The video captures a blend of contemporary digital interaction and expressive urban art, creating a dynamic yet harmonious scene.",
-      "video_path": "Girl/mixkit-girl-looking-at-the-likes-in-her-post-4914_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A young mother and her baby sit comfortably on a bed, surrounded by an inviting, cozy atmosphere. The woman, wearing a sleeveless top and jeans, is gently engaging with the baby, who is dressed in an adorable animal-print onesie. The child is seated on the bed with colorful toys scattered around, including a plush toy and a board book. The warm glow from a hanging lamp casts a soft light on them, enhancing the serene environment. Pillows are propped up against the headboard, providing a cushioned backdrop as the mother leans slightly over to interact with the baby. A small bottle is visible beside her, suggesting a nurturing setting. Her hand gestures animatedly as she holds up a soft, white cushion with red and blue accents, likely stimulating the baby\u2019s curiosity. Their shared moment is filled with affection and joy, a perfect snapshot of familial bonding.",
-      "video_path": "Baby/mixkit-loving-mother-and-her-baby-playing-with-soft-toys-49966_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A young girl with long brown hair sits at a round wooden table, engrossed in working on her laptop. The laptop screen is a vivid green, suggesting a green screen effect is in use. To her left, a doll dressed in a yellow and white outfit is casually laid on top of some books, adding a playful and innocent touch to the scene. The setting is cozy, with sheer curtains in the background allowing soft natural light to spill into the room. The girl's posture and focused attention on the laptop suggest she is either playing a game or learning something new. This serene and domestic atmosphere is complemented by the slight blur of a dark couch in the foreground, framing the focused activity of the child.",
-      "video_path": "Girl/mixkit-little-girl-doing-homework-on-a-laptop-4757_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "An expansive view of a calm bay reveals a fleet of sailboats, each anchored in a regimented line stretching toward the horizon. The water is a serene blue, reflecting the soft hues of the early morning sky. A gentle breeze is indicated by the subtle ripples trailing behind the boats, while a single, larger vessel cuts a distinct path, leaving a graceful wake in its journey to the open sea. On one side, a cluster of modern high-rise buildings stands, contrasting against the natural simplicity of the water, suggesting a blend of urban and marine life. The distant shoreline is barely visible, softened by the atmospheric perspective, giving a sense of endless waters meeting the sky. The overall mood is peaceful and orderly, with the boats appearing almost as sentinels guarding the expanse of the tranquil bay.",
-      "video_path": "beach/mixkit-flying-backwards-over-the-sea-near-a-coast-50187_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the video, a person is standing in the center of a dark, featureless space, illuminated by a spotlight that emphasizes their presence. The individual is dressed in a traditional martial arts uniform, known as a gi, which is predominantly white with a black belt tied around the waist, indicating a high level of expertise. The background remains pitch black, creating a stark contrast with the brightly lit figure, ensuring complete focus on them. The person's expression is serious and focused, reflecting a deep sense of discipline and concentration. Their hands move gracefully, transitioning through various martial arts stances, demonstrating practiced skill and fluidity. The uniform's crisp fabric folds and subtly reflects the light, further highlighting each precise movement. Despite the simplicity of the environment, the scene is dynamic, with each motion capturing the essence of martial arts practice. The video effectively conveys a sense of calm strength and mastery, making it ideal for an AI to recreate with attention to posture, lighting, and attire.",
-      "video_path": "Sport/mixkit-karate-fighter-bowing-to-the-front-49706_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In a dimly lit room bathed in a mix of neon purple and blue lights, a focused individual is seated in a gaming chair. She wears a white hoodie and large headphones with cat ears that glow softly, creating a striking silhouette. Her hands rest on a keyboard, typing swiftly as she concentrates intently on the screen in front of her. The atmosphere exudes a sense of intensity and immersion, with the soft-colored lighting enhancing the futuristic vibe. Her long hair cascades down her shoulders, adding a touch of elegance to the otherwise tech-centric setting. The overall scene captures the essence of a dedicated gamer deeply engaged in her virtual world.",
-      "video_path": "earth/mixkit-a-young-woman-wearing-headphones-with-rgb-lights-suddenly-gets-51621_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "Inside a dimly-lit bus, five individuals are seated along the rows of worn seats, each subtly illuminated by the colorful lights emanating from overhead. On the left, a woman sits with a relaxed posture, her curly hair accented by a patterned scarf, wearing a plaid outfit paired with bright neon socks. Next to her, a person clad in a denim jacket appears deep in thought, resting their head on a hand. Further back, another figure in a bucket hat and oversized yellow attire gazes across the aisle, evoking a sense of introspection. The atmosphere is enriched by the soft glow of red and green lights, bathing the bus interior in an almost surreal ambiance, creating a compelling tableau of urban life.",
-      "video_path": "Music/mixkit-conceptual-urban-fashion-42581_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "An aerial view captures two tennis players on a court, with one dressed in white on the left and another in red on the right. They are mid-game, each poised for action with rackets in hand, accentuated by their strategic positioning at opposite baselines. The court itself is a stark, deep blue, bordered by the vibrant green of the surrounding area, with a dark central net dividing the space. Long shadows stretch dramatically across the ground, suggesting a late afternoon setting. The subtly textured surface of the court contrasts with the crisp, white lines marking its boundaries and sections. This scene creates a vivid, balanced composition, highlighting both the competitive tension and serene atmosphere of the game.",
-      "video_path": "People/mixkit-two-people-playing-tennis-aerial-view-880_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In a vibrant, dreamlike setting, a lone figure moves energetically against a backdrop of deep blue and purple hues, casting emotive shadows that ripple with dynamic motion. The figure, almost obscured by a smeared effect, suggests a rhythmic dance or a passionate performance, arms blurred as they sweep through colorful, streaked lighting. A neon glow accentuates their form, particularly highlighting the face which is abstractly illuminated in bursts of orange and red, suggesting intense emotional expression. The scene is dominated by two primary elements \u2013 the figure\u2019s motion and the dramatic lighting, creating a synergy of human emotion and visual spectacle. Swirling trails of light seem to intertwine with the figure, like a visual symphony of movement and color that floods the space. The lighting changes, casting intricate patterns on the figure and the surrounding space, giving the impression of a kaleidoscope in motion. Despite the blurred and abstract portrayal, there is a sense of focus conveyed through the figure\u2019s intent movements, akin to a conductor orchestrating a visual and auditory performance. The environment resonates with an electric energy, suggesting a seamless fusion of art and technology. As the visual drama unfolds, the scene invites viewers to lose themselves in the abstract dance and the play of vivid luminance.",
-      "video_path": "Music/mixkit-dancer-dancing-with-a-light-bar-in-his-hands-42221_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In a brightly lit studio, a photographer wearing a denim jacket focuses intently, capturing shots with a professional camera. Facing him, a model stands gracefully, adjusting her long, flowing hair with delicate movements. The scene is characterized by strong contrasts; the model's soft pink attire and gentle gestures complement the rugged, precise demeanor of the photographer. Positioned against a minimalist backdrop, the pair work seamlessly, with the camera\u2019s lens pointed directly at the model, capturing her elegance. The soft, diffused lighting casts a gentle glow on both subjects, creating an airy and ethereal atmosphere perfect for a high-fashion photo shoot.",
-      "video_path": "Fashion/mixkit-professional-photo-session-with-a-young-female-model-41621_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "The video showcases a serene, expansive landscape covered with a variety of trees dotting the hills. The hills gently slope across the frame, with patches of dry grass contrasting against the lush green foliage. Tall trees with dense canopies stand elegantly, casting soft shadows on the ground below. The sunlight bathes the entire scene, highlighting the varied textures of the leaves and terrain. Gaps between the trees reveal a narrow dirt path meandering through the hills, suggesting a sense of quiet solitude. The undulating hills extend into the distance, creating depth and a calming sense of vast space. The verdant hues of the leaves contrast with the earthy tones of the hills, enhancing the visual richness. In the background, a faint outline of distant hills can be seen, blurred softly by the atmospheric perspective. This tranquil setting could be efficiently recreated in a virtual environment by focusing on its layered composition, color palette, and natural textures.",
-      "video_path": "forest/mixkit-aerial-panorama-of-a-sunny-mountain-landscape-40846_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A bustling ski slope comes alive with skiers descending a pristine, snow-covered hill, surrounded by towering, snow-draped evergreens. Several figures stand atop the slope, silhouetted against a clear blue sky, preparing to embark on their ski run. The chair lift on the right continuously drops off eager adventurers, adding to the excitement at the hilltop. Each skier, clad in colorful winter gear, carves distinct paths into the textured snow as they weave their way down. The interplay of sunlight and shadows accentuates the myriad tracks etched into the slope, creating a dynamic visual rhythm. The scene captures a vibrant winter wonderland, full of action and the thrill of a perfect ski day.",
-      "video_path": "Car/mixkit-skiers-on-a-snowy-slope-3327_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "The scene unfolds within a dimly lit bus, where three young individuals are seated, each absorbed in their unique world. To the left, a person with tied-back hair rests their head on their hand, dressed casually in a jacket and jeans, projecting a relaxed demeanor. Central to the frame is another individual, sitting upright with intense focus, donning a plaid blazer and oversize hoops, enhancing their confident presence. The muted green and red lighting casts an atmospheric glow, adding depth and intrigue to the setting. On the right, a person in a bucket hat and striped shirt leans back, appearing contemplative as they adjust their hat with a nonchalant gesture. The interplay of light and shadow highlights their expressions, creating an intimate and cinematic ambiance. Together, these figures form a cohesive tableau, capturing a moment of introspection amid a bustling yet serene urban environment.",
-      "video_path": "City/mixkit-three-models-posing-to-the-lens-while-on-board-a-42575_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A determined climber is scaling a massive rock face, showcasing exceptional strength and skill. The person, clad in a teal shirt and dark pants, climbs with precision, their movements measured and deliberate. They are secured by climbing gear, which includes ropes and a harness, emphasizing their commitment to safety. The rugged texture of the sandy-colored rock provides an imposing backdrop, adding drama and scale to the climb. In the distance, other large rock formations and sparse vegetation can be seen under a bright, overcast sky, contributing to the natural and adventurous atmosphere. The scene captures a moment of focus and challenge, highlighting the climber's tenacity and the breathtaking environment.",
-      "video_path": "Sport/mixkit-alpinist-climbing-a-huge-rock-in-a-desert-43306_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A woman stands confidently in front of a large array of solar panels, her navy blue jumpsuit contrasting against the lush green grass beneath her feet. Her expression is calm and focused, eyes facing directly ahead, suggesting a deep connection to the subject matter\u2014renewable energy. The sunlight bathes the scene in warm hues, casting gentle shadows and highlighting the geometric precision of the solar panels' grid-like structure. The background reveals a blend of nature and technology, as the panels are anchored on a grassy slope with foliage on the left side of the frame. This composition captures a harmonious blend of human innovation and environmental consciousness, accentuated by the serene outdoor setting.",
-      "video_path": "Business/mixkit-woman-standing-in-front-of-a-solar-panel-4880_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the video, two people are working at a wooden desk, using an iMac computer. One person, wearing a white knit sweater, is using the apple wireless mouse with their right hand, while their left hand rests on the sleek white keyboard. Their movements are smooth yet intentional, suggesting they are focused on a task on the computer screen. The monitor displays a well-organized array of files and folders, hinting at a task that involves detailed organization or detailed data navigation. The second person, only subtly visible, sits closely by and appears to observe or assist, creating a collaborative atmosphere. Their presence adds a quiet dynamic to the scene, as if they are ready to provide input or guidance. Sticky notes with handwritten notes are attached to the monitor\u2019s stand, adding a touch of personal organization amidst the digital workspace. The focus on the keyboard and mouse emphasizes a streamlined workflow, indicative of a productive work environment. The overall ambiance is calm and focuses on teamwork, technology, and efficient workspace management.",
-      "video_path": "People/mixkit-person-with-glasses-working-on-a-desktop-computer-3248_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A man stands in front of a modern glass facade, taking off a dark hoodie to reveal his gray tank top underneath. His arms are lifted high as he maneuvers the hoodie over his head, showcasing a fluid motion that conveys a sense of calm and routine. The lighting highlights the contours of his muscles, emphasizing a combination of strength and quiet determination. Behind him, the reflective surface of the glass panels provides a subtle backdrop, enhancing the focus on his focused and serene demeanor.",
-      "video_path": "Sport/mixkit-man-puts-on-sleeveless-hoodie-603_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "The video displays a captivating dance of fiery orange flames against a stark black background, creating an intense visual contrast. The flames twist and intertwine, forming symmetrical, swirling patterns that expand and contract rhythmically across the frame. Each fiery tendril seems to be alive, moving with an almost hypnotic fluidity that captures the viewer's attention. The illumination from the flames casts subtle shadows, enhancing the depth and texture of the scene. Overall, the dynamic movement and vibrant color palette create an atmosphere of both beauty and power.",
-      "video_path": "fire/mixkit-two-orange-flames-on-black-background-685_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In this scene, a person is seated in a dimly lit room, possibly a recording studio, holding several drumsticks in their hands. The individual's face is partially obscured by sunglasses, adding a touch of mystery to their demeanor. They are wearing a colorful, patterned shirt with a mix of orange and blue tones that stands out against the darker background. The person appears focused and engaged with the drumsticks, their hands prominently displayed. The ambient light casts warm, soft shadows, emphasizing the texture and colors of their shirt and the wooden drumsticks. The room features wooden paneling, which complements the overall cozy, music-centric setting of the scene. The use of perspective centers on the drumsticks, highlighting the importance of rhythm and music in the captured moment.",
-      "video_path": "Music/mixkit-drummer-stretching-before-playing-42783_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A man is casually sitting on a sofa, engrossed in his meal and entertainment. He is holding a TV remote in one hand while reaching for food with the other, indicating a laid-back, comfortable evening. The table before him is filled with takeout containers, revealing a variety of appetizers and dishes, suggestive of a casual dining experience at home. The background is defined by colorful patterned cushions, adding a cozy, homey feel to the scene. Warm, ambient lighting highlights the relaxed atmosphere, casting soft shadows that contribute to the intimate setting. In this moment, he takes a bite of a sandwich, comfortably balancing his attention between food and whatever is playing on the screen.",
-      "video_path": "Man/mixkit-man-watching-tv-and-eating-fast-food-26089_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "The scene opens to a breathtaking view of a tranquil ocean horizon at dusk, displaying a vibrant tapestry of oranges, pinks, and purples as the sun sets. In the foreground, tall, swaying palm trees frame the scene, their silhouettes stark against the colorful sky. The ocean itself shimmers with reflections of the sunset, creating a peaceful, almost ethereal atmosphere. A small boat can be seen in the distance, centered on the horizon, adding a sense of scale and solitude to the scene. The waves gently lap the shore, creating faint patterns on the sandy beach, which stretches across the foreground. Above, the sky is dotted with scattered clouds that catch the last light of the day, enhancing the drama and beauty of the scene. The overall mood is serene and contemplative, capturing a perfect moment of nature\u2019s grandeur.",
-      "video_path": "beach/mixkit-sunset-with-sailing-boats-2166_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A man sits hunched on a couch, the weight of emotions clearly visible on his posture. He wears a simple, gray t-shirt, and his head is bowed, resting in his hands, which cover most of his face, obscuring his features. The gentle light filtering through sheer curtains in the background casts a soft glow upon him, emphasizing the contrast between his static form and the hazy brightness behind. His elbows rest upon his knees, suggesting a posture of deep contemplation or distress. The simplicity of the room, with its muted colors, highlights the focus on the man's internal struggle. Delicate detailing on the fabric of his shirt adds texture, enhancing the scene's realism. Subtle changes in the natural light indicate the passage of time, as the man remains unmoving, absorbed in thought. This intimate moment captures a profound vulnerability, making the scene universally relatable and poignant.",
-      "video_path": "Man/mixkit-worried-and-sad-man-with-his-head-down-4701_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A pair of hands, belonging to an unseen figure, carefully unrolls a large sheet of crisp, white paper on a dark wooden table. The lighting is warm, casting a gentle glow that highlights the textures of the paper and the wood grain of the table. As the paper unfurls, the edges reveal the faint beginnings of a colorful map printed on its surface. The arms, clad in a casual gray T-shirt, suggest a relaxed and focused task at hand. Each motion is deliberate, with fingers deftly guiding the paper, ensuring it lays flat without creases. In the background, a hint of a red curtain can be seen, adding a touch of color and depth to the setting. The composition of the scene emphasizes the contrast between the bright paper and the rich tones of the surroundings. This serene and methodical action evokes a sense of exploration and preparation.",
-      "video_path": "Man/mixkit-unrolling-a-world-map-on-a-table-21626_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A young woman sits on a vibrant green seat inside a bus, illuminated by the soft glow of pink and blue lights. Her outfit is a striking mix of colors: a neon pink top paired with a jacket featuring dark sleeves, and jeans that provide a neutral contrast. She wears large, hoop earrings that catch the light as she moves slightly, exuding an air of cool confidence. Her gaze is directed thoughtfully to the side, suggesting contemplation or daydreaming during her commute. The metallic pole beside her adds a geometric element to the composition, reflecting the kaleidoscope of neon hues. The background is a clean, futuristic white, serving as a blank canvas that amplifies the neon atmosphere. Her relaxed posture and the modern bus setting create a scene that captures a blend of urban life and personal introspection.",
-      "video_path": "City/mixkit-fashion-model-posing-on-a-bus-42578_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A silver SUV drives along a winding, snow-covered mountain road, with dense pine trees blanketed in snow lining both sides. The scene is serene, with the vehicle moving smoothly, possibly on a winter journey or vacation. As the SUV disappears around the bend, another, darker SUV follows, creating a sense of motion and perspective on the snow-dusted asphalt. The towering, snow-laden rock formation to the right contrasts with the dark green of the pines, highlighting the peacefulness of the wintry landscape.",
-      "video_path": "Car/mixkit-curve-on-a-snowy-forest-road-3317_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "The video showcases a vibrant urban skyline during twilight, with towering buildings reflecting the warm hues of the setting sun. A series of tall, cylindrical structures dominate the foreground, adjacent to a complex of industrial equipment and grids. The scene includes modern high-rise buildings with glass exteriors, capturing the evolving architecture of a bustling cityscape. A prominent structure labeled \"CITY OF AUSTIN POWER PLANT\" stands out, highlighting the industrial theme amidst the urban backdrop. The soft glow of city lights begins to pierce the approaching dusk, creating an inviting yet dynamic atmosphere. Shadows cast by the buildings add depth and contrast, emphasizing their massive scale and intricate designs. The overall composition is balanced between the natural light of the sunset and the artificial illumination of the city, offering a compelling visual narrative.",
-      "video_path": "Car/mixkit-slow-air-travel-in-reverse-over-a-big-city-49841_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the scene, a striking architectural structure dominates the view, bathed in a soft, ambient light. The enormous yellow arches serve as the centerpiece, drawing the eye upwards with their majestic curves and towering presence. The smooth, clean surfaces of the structure reflect the light, highlighting the texture and depth of the architecture. In the foreground, blurred streaks of headlights and taillights suggest the motion of vehicles passing by, adding dynamic energy to the otherwise still scene. The contrast between the fast-moving lights and the static arches creates a balanced composition. To the left, a lone streetlamp and a small tree provide a touch of nature and urban elements against the monumental backdrop. The night sky subtly peeks through the gaps in the structure, hinting at a clear, calm evening. Shadows from the arches create patterns on the ground, adding an intricate detail to the scene. Overall, the combination of light, shadow, and movement makes for a dramatic and visually captivating moment.",
-      "video_path": "Car/mixkit-a-fast-timelapse-of-the-street-with-a-monumental-yellow-50993_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A tranquil marina comes into full view under the golden hues of a setting sun. A collection of gleaming yachts and boats are neatly moored, their reflections shimmering softly on the gentle water. The sun's low position casts elongated shadows over the bustling harbor scene, while rolling hillsides surround the distant cityscape. The skyline is interspersed with modern buildings and clusters of residences, adding layers to the vibrant community. At the center, a broad wooden pier juts confidently into the harbor, extending an invitation for leisurely strolls. To the left, various shops and colorful structures line the waterfront, indicating a vibrant coastal economy. The entire atmosphere exudes a serene yet lively charm, balancing the hustle of maritime activity with the peacefulness of the encroaching dusk. It's a scene of calm anticipation, as if the whole place holds its breath before the night's events unfold.",
-      "video_path": "beach/mixkit-harbor-on-a-tourist-coast-with-many-boats-and-yachts-40077_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "The video features a confident individual standing atop a structure against a clear blue sky, exuding a sense of freedom and style. The person is clad in a striking yellow button-up shirt tied at the waist, and beneath it, they wear a simple white top that adds to their relaxed yet stylish appearance. Completing the ensemble are high-waisted white jeans paired with a black belt, adding a touch of contrast. Around their neck is a bold red scarf, providing a splash of color and an air of vintage flair. The person's sunglasses, tinted in yellow, reflect the sunlight and contribute to the overall cool and composed demeanor. Their hair is styled elegantly, pulled back with headphones resting over the ears, suggesting they are immersed in music. One hand casually grazes the headphones, while the other rests gently on the railing, grounding the individual in the moment. The scene is an effortless blend of fashion and tranquility, capturing the spirit of sunny, carefree days.",
-      "video_path": "Music/mixkit-standing-woman-listening-to-music-460_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A ballerina gracefully spins and moves across a pink-hued studio, her poised figure accentuated by a shimmering white tutu and bodice. The background, a continuous wash of soft pink, provides a serene and ethereal atmosphere, emphasizing her fluid movements. Her arms extend with elegance, highlighting the delicacy and precision of her ballet pose, while her focused expression adds intensity to the scene. The subtle details of her costume, combined with the pink monochromatic ambiance, create a dreamlike spectacle, ideal for an AI to envision a oneiric dance setting.",
-      "video_path": "Dance/mixkit-portrait-of-a-ballerina-spinning-with-pink-background-40163_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "The scene unfolds with two human figures in the distance, making their way through a serene meadow, thick with tall golden grass swaying gently in the breeze. The sun hangs low in the sky, casting a soft, diffused glow that illuminates the landscape with a warm, ethereal light. These figures, clad in hiking gear, move deliberately, suggesting they're either embarking on or concluding a journey. Their silhouettes contrast against the lush greenery of the surrounding trees, whose branches reach out, framing the horizon. The play of light and shadow among the trees creates a quilt of textures, with each leaf catching a hint of the sun's dying rays. This tranquil setting evokes a sense of calm and adventure, capturing the quintessential beauty of nature\u2019s landscape.",
-      "video_path": "People/mixkit-landscape-in-nature-while-two-people-are-jogging-44348_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A large cargo ship is docked at an industrial port, its white superstructure contrasting with the deep green and yellow of its deck. The foreground is dominated by the calm, deep blue waters of the harbor, which reflect the vessel\u2019s imposing presence. Surrounding the ship, a series of industrial buildings and storage facilities are visible, hinting at the bustling activity of the port. The deck is intricately detailed, featuring an array of pipes, equipment, and railings, showcasing the ship's functionality and purpose. In the background, a paved area with green patches and a few parked vehicles adds to the busy, industrious atmosphere of the scene.",
-      "video_path": "sea/mixkit-empty-cargo-ship-waiting-at-the-port-4209_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A lone climber ascends a towering rock face, clad in a pink shirt and gray pants, displaying a determined and focused expression. The climber navigates the rugged surface, where the texture of the rock is peppered with natural pockets and crevices that offer handholds and footholds. Sunlight casts soft shadows across the cliff, highlighting the intricate patterns and the climber\u2019s strategic movements. The cliff looms high, with sparse vegetation breaking the monotony of the stone, while distant rocky formations form a dramatic backdrop against the clear blue sky. The climber\u2019s gear, including a harness and chalk bag, underscores the adventure and challenge woven into this majestic, vertical journey.",
-      "video_path": "Sport/mixkit-mountaineer-girl-climbing-a-steep-rocky-mountain-41089_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A person is seen in a close-up shot, skillfully adjusting the tuning pegs of a guitar, showcasing a focused and practiced hand. The image is in black and white, highlighting the contrast between the textures of the instrument and the clothing. The individual's shirt, visible in the background, adds a soft, subtle texture, while the dark tones of the guitar neck create depth in the scene. This composition captures a moment of concentration and finesse, perfect for recreating an intimate musical setting.",
-      "video_path": "Music/mixkit-guitarist-playing-so-inspired-black-and-white-shot-44178_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A musician is playing a large brass instrument with the words \"Brass Band\" clearly visible on its bell. The scene is set against a vibrant yellow backdrop, casting a warm glow on the subject. The musician wears a dark cap and a matching suit, adding a formal touch to his attire. He is deeply focused on his performance, with the instrument's intricate tubing adding complexity to the visual composition. The lighting creates dramatic shadows and highlights, emphasizing the musician's expression and the instrument's metallic sheen. This harmonious blend of color and form captures the essence of a live brass band performance.",
-      "video_path": "Music/mixkit-musician-playing-the-trombone-while-dancing-43752_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the video, a lone musician stands gracefully in front of a grand cathedral, playing an accordion while surrounded by the lively water display of a central fountain. Dressed in a casual ensemble, he wears a light-colored shirt, dark pants, and a flat cap that gives him a vintage charm. His posture is relaxed, yet engaged, as he sways gently in rhythm with the music, casting soft shadows on the cobblestone steps beneath him. The backdrop features the cathedral's towering twin spires, with intricate stonework that casts a rich, historical aura around the scene. Sunlight bathes the entire setting, enhancing the golden hues of the cathedral facade and creating a halo-like effect around the musician. The fountain's water jets splash playfully, catching glimmers of light and adding a dynamic element to the tranquil atmosphere. The scene captures a harmonious blend of architectural majesty and human creativity, framed by the clear, azure sky that extends infinitely above. It's a vivid depiction of solitude and artistry, set against a timeless urban landscape.",
-      "video_path": "Music/mixkit-man-plays-an-accordion-in-front-of-a-fountain-630_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the tranquil video, a person sits in a meditative pose on a gentle hillside, silhouetted against the dawning sky. The person is facing the breathtaking sunrise, with their back slightly turned to the viewer, wearing a simple, light-colored shirt. Their right hand rests on their knee, fingers relaxed in a common meditation mudra, symbolizing calmness and peace. The sky, a stunning blend of soft oranges and deep purples, gradually brightens, casting a warm glow over the lush, green landscape. To the left, the outlines of distant urban buildings can be seen against the horizon, adding a contrast between nature and city life. A river reflecting the sky's colors meanders through the scene, lending a serene, flowing dynamic to the landscape. Trees rise and fall gently across the terrain, their leaves rustling only faintly in the morning breeze. The person remains still and focused, embodying a moment of mindfulness and connection with nature. This visual captures a harmonious balance, evoking a sense of tranquility and introspection.",
-      "video_path": "City/mixkit-girl-meditating-in-yoga-pose-at-sunset-4803_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A serene landscape video captures a breathtaking panoramic view of a vast valley covered in a gentle mist. The undulating hills are lush with dense greenery, their rich foliage creating a vibrant border on the left side of the frame. The mist weaves through the landscape like a soft, ethereal blanket, lending a dream-like quality to the scene. In the distance, several mountain peaks emerge, their dark outlines contrasting against the pale blue sky. A few faint, wispy clouds drift lazily across the horizon, complementing the tranquil atmosphere. The sunlight filters through the haze, casting a warm glow and highlighting different textures of the flora. The overall mood is calm and contemplative, inviting the viewer to pause and appreciate nature's untouched beauty. The composition emphasizes depth and expansiveness, drawing attention to the harmony between earth and sky. This captivating scene embodies tranquility, offering a perfect backdrop for meditation or relaxation.",
-      "video_path": "forest/mixkit-flying-over-a-hill-with-a-view-of-the-surrounding-49743_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In this scene, a bearded individual is intently focused on their smartphone, with the sun setting in the background, casting a warm glow across the cityscape. The person, partially visible, is wearing a dark, buttoned shirt that contrasts with the golden hue of the sunset. Their hands are holding the smartphone delicately but purposefully, reflecting a sense of engagement and focus on the screen. The sunlight creates a striking lens flare effect, enhancing the dramatic atmosphere of the moment as it glimmers off the phone\u2019s surface. The surrounding environment hints at an elevated vantage point, providing a panoramic view of the urban landscape below.",
-      "video_path": "City/mixkit-guy-texting-at-sunset-265_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In an expansive, industrial space defined by towering columns and high ceilings, a solitary figure takes center stage. The person, dressed in dark, fitted clothing, assumes a powerful, dynamic stance with one leg bent forward and both arms outstretched in a horizontal arc. Framing this pose are intense flames that engulf their arms, creating a striking visual contrast against the muted tones of the room. The fire forms a brilliant halo of orange and yellow, casting flickering shadows on the weathered walls and worn, tiled floor. This interplay between light and dark showcases the dancer's poise and agility, as they maintain balance amidst the intense heat. Windows line the background, their panes dimly illuminated by the daylight filtering in, adding depth and perspective to the scene. The entire performance evokes a sense of raw energy and elemental mastery, as the figure continues to manipulate the fire in a seamless, mesmerizing display.",
-      "video_path": "fire/mixkit-expert-juggler-doing-tricks-with-a-stick-with-fire-43663_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A man is playing the violin, focused intently on his music. His fingers gracefully dance along the strings, flawlessly executing each note. He holds the violin close to his chin with a sense of familiarity and expertise. The rich, warm tones of the violin reflect in the soft lighting of the room. He wears a dark shirt, and a subtle necklace rests against his chest, adding a personal touch to his attire. The bow moves smoothly across the strings, producing a melody that seems to fill the space with emotion. His expression is one of concentration and passion, immersing himself fully in the performance. The background is softly blurred, bringing the violin's intricate craftsmanship and his precise movements into sharp focus. This serene and intimate moment captures the essence of his musical artistry.",
-      "video_path": "Music/mixkit-fiddler-playing-a-song-639_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the dimly lit parking garage, two figures engage in an impromptu game of soccer. The first person, wearing a light grey shirt and black pants with three white stripes, skillfully maneuvers the ball with precise footwork. The ground is slick with patches of water, reflecting the vibrant neon lights above. A second figure, clad in dark clothing, stands poised in the background, ready to intercept. The space is defined by stark yellow lines and orange safety bollards, adding structure to the chaotic energy of the scene. The soccer ball glides smoothly across the wet floor, kicking up droplets as it passes. Despite the muted colors of the environment, the players' movements are dynamic and full of life. Their shadowy silhouettes dance with the reflecting light, creating a mesmerizing visual interplay. The atmosphere is charged with focus and camaraderie, encapsulating the essence of a late-night urban soccer experience.",
-      "video_path": "Sport/mixkit-player-making-skillful-play-in-a-street-soccer-game-43504_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A lone climber is seen scaling a towering vertical rock face, demonstrating remarkable strength and focus. Dressed in a light-colored shirt and jeans, the climber grips the stone tightly, navigating the rough textures and crevices with precision. The sheer cliff is massive, exhibiting a range of natural hues from light tan to deep gray, accentuating the climber's figure against the vast rocky backdrop. Surrounding the cliff, scattered greenery and rugged terrain provide a sense of wilderness and isolation. The scene portrays a daring ascension requiring concentration and skill, capturing the essence of human endeavor against nature's formidable beauty.",
-      "video_path": "Sport/mixkit-skilled-mountaineer-climbing-a-gigantic-mountain-41083_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In this serene landscape, a lush meadow stretches across the foreground, dotted with vibrant yellow wildflowers swaying gently in the breeze. A towering tree stands majestically on the right side, its branches reaching wide under the bright blue sky filled with fluffy white clouds. On the left, dense trees form a natural corridor leading to the horizon, suggesting a sense of journey and possibility. The richness of the green grass contrasts beautifully with the golden hue of the distant fields, creating a harmonious palette of nature\u2019s colors. The play of light and shadow adds depth and dimension, evoking a tranquil, inviting atmosphere. It's a scene where nature\u2019s beauty simply commands attention, offering a perfect escape into tranquility.",
-      "video_path": "sky/mixkit-countryside-meadow-4075_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A solitary boat glides across the expansive, tranquil expanse of a serene lake. The vessel leaves a gentle wake behind, creating delicate ripples across the mirror-like surface. The water appears a rich shade of teal, seamlessly blending with the sky at the horizon. Silhouettes of distant trees are faintly visible, creating a picturesque backdrop that enhances the solitary journey of the boat. The sky is a calm gradient, shifting from soft oranges near the shore to the pale blues above. In the distance, a few slender poles emerge from the water, remnants of an old structure or natural formation. The mood of the scene is one of peace and solitude, with the boat journeying steadily through the quiet landscape. There is a sense of endless possibilities as the boat moves toward the unseen beyond the frame. The simplicity and stillness of the scene invite contemplation and reflection, encapsulating a perfect moment of quietude on the water.",
-      "video_path": "mountain/mixkit-motorboat-on-a-large-lake-with-turquoise-blue-waters-4996_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In a cozy, dimly lit caf\u00e9, a woman sits alone at a rustic wooden table, fully engrossed in her reading. Her dark, wavy hair frames her face as she leans forward over an open book, suggesting deep focus and contemplation. The caf\u00e9\u2019s ambiance is warm, with hanging pendant lights casting a soft glow over the wooden shelves lined with jars and coffee paraphernalia in the background. A small cup of coffee rests just within her reach, alongside a glass dome encasing a solitary pastry, adding a touch of tranquility to the scene. Her casual attire, a denim jacket over a simple shirt, complements the laid-back, comfortable setting of the caf\u00e9. The contrast between her concentrated expression and the bustling, yet subdued caf\u00e9 atmosphere creates a harmonious, serene visual. The overall composition captures a quiet moment of introspection amidst the gentle hum of caf\u00e9 life.",
-      "video_path": "Woman/mixkit-woman-drinking-coffee-in-a-cafe-223_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In a vast, deserted landscape under the night sky, a solitary figure stands at a small music setup, illuminated by strategically placed lights. The person is engrossed in playing a keyboard, with various electronic equipment surrounding them, casting soft glows of orange and blue hues across the scene. To the left, a large circular light adds a dramatic focal point, highlighting the intense contrast between the darkness and the lit performance area. This setup, with its minimalistic design and strategic lighting, creates a captivating and easily recognizable scene that merges the serene, expansive backdrop with an intimate, focused music performance.",
-      "video_path": "Music/mixkit-talented-dj-playing-in-a-lonely-desert-42414_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In a bustling urban scene, cars zoom past a weathered building, their blurred motion a testament to the city\u2019s lively pace. The building, with its faded yellow and brown facade, boasts graffiti that speaks of both art and decay, framing the scene with an air of urban grit. A solitary figure stands slightly to the side, clad casually in a gray top and mustard trousers, gazing into the street, seemingly detached from the surrounding flurry. The motion of the traffic creates a dynamic contrast against the static backdrop, emphasizing the relentless movement of the city. As the video progresses, a bright yellow taxi appears, slowing down as it approaches the figure, adding a pop of color to the desaturated hues of the environment. The interaction suggests a routine, a possibly daily exchange between the driver and the pedestrian, hinting at the rhythms of city life. Overhead, a soft, overcast sky casts a diffused light, lending the scene a subdued, timeless quality. Small elements, like the vertical pole cutting through the frame and the distant chatter of urban sounds, complete this vivid tableau of urban existence.",
-      "video_path": "Car/mixkit-morning-in-the-street-time-lapse-1648_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "A young woman sits on a curb in a tranquil park, basking in the golden hue of the setting sun. Beside her, a collie dog rests calmly, its fur illuminated by the warm sunlight, creating a serene glow. The woman's hand gently strokes the dog's back, highlighting the bond and affection between them. Tall trees surround the pair, casting elongated shadows on the leaf-laden ground, adding to the peaceful and intimate ambiance of the scene.",
-      "video_path": "Pets/mixkit-a-woman-pets-a-dog-in-a-park-1562_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the video, a grand, majestic elephant stands in an open, sunlit field, its massive form dominating the scene. The elephant's skin is a tapestry of earthy tones, with rough, textured wrinkles that add character to its already imposing presence. Its trunk, a powerful and flexible appendage, moves gently, swaying as the elephant possibly enjoys the warmth of the day. The background is a blur of greenery, suggesting a lively environment filled with trees and shrubs that provide a natural habitat. Light plays on the elephant's skin, highlighting patches of dust and dirt that give it an authentic wilderness look. The scene captures the tranquility and majesty of this gentle giant in its natural surroundings.",
-      "video_path": "Zoo/mixkit-wet-elephant-in-the-savanna-3663_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    },
-    {
-      "caption": "In the video, a fluffy dog with brown patches is intently engaged with a bright red toy shaped like a fire hydrant, which has a yellow and orange rope attached. The dog's body is relaxed as it lies on a plain white background, concentrating on nudging and playfully biting the toy. Its ears perk up slightly with curiosity, and its eyes are fixated on the toy, suggesting a scene of focused playfulness. The neutral tones of the dog's fur contrast starkly against the vivid red of the toy, creating a visually striking moment.",
-      "video_path": "Pets/mixkit-a-cute-border-collie-dog-play-with-a-fire-street-50662_clip_1.mp4",
-      "num_inference_steps":  3,
-      "height": 448,
-      "width": 832,
-      "num_frames": 61
-    }
-  ]
-}

examples/distillation/dmd/Wan2.1-Fun-1.3B-InP/crush_smol/README.md

@@ -1,3 +0,0 @@
-# DMD Distillation Wan2.1-I2V-14B-480P Crush-Smol Example
-
-Coming soon!

examples/inference/basic/README.md

@@ -18,16 +18,6 @@ git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
 python examples/inference/basic/basic.py
 ```
 
-For an example on Apple silicon: 
-```
-python examples/inference/basic/basic_mps.py
-```
-
-For an example running DMD+VSA inference:
-```
-python examples/inference/basic/basic_dmd.py
-```
-
 ## Basic Walkthrough
 
 All you need to generate videos using multi-gpus from state-of-the-art diffusion pipelines is the following few lines!

examples/inference/basic/basic.py

@@ -1,41 +1,42 @@
 from fastvideo import VideoGenerator
 
-from fastvideo.configs.sample import SamplingParam
+# from fastvideo.v1.configs.sample import SamplingParam
 
-OUTPUT_PATH = "video_samples_5b"
+OUTPUT_PATH = "video_samples"
 def main():
     # FastVideo will automatically use the optimal default arguments for the
     # model.
     # If a local path is provided, FastVideo will make a best effort
     # attempt to identify the optimal arguments.
-    model_id = "Wan-AI/Wan2.2-TI2V-5B-Diffusers"
     generator = VideoGenerator.from_pretrained(
-        model_id,
+        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
         # FastVideo will automatically handle distributed setup
-        num_gpus=1,
+        num_gpus=2,
         use_fsdp_inference=True,
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=False,
-        # image_encoder_cpu_offload=False,
+        use_cpu_offload=False
     )
 
-    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.2-TI2V-5B-Diffusers")
+    # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
     # sampling_param.num_frames = 45
     # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
     # Generate videos with the same simple API, regardless of GPU count
     prompt = (
-        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest. The playful yet serene atmosphere is complemented by soft natural light filtering through the petals. Mid-shot, warm and cheerful tones."
+        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+        "wide with interest. The playful yet serene atmosphere is complemented by soft "
+        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
     )
-    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
+    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True)
     # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/")
 
     # Generate another video with a different prompt, without reloading the
     # model!
     prompt2 = (
-        "A majestic lion strides across the golden savanna, its powerful frame glistening under the warm afternoon sun. The tall grass ripples gently in the breeze, enhancing the lion's commanding presence. The tone is vibrant, embodying the raw energy of the wild. Low angle, steady tracking shot, cinematic."
-    )
-    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
+        "A majestic lion strides across the golden savanna, its powerful frame "
+        "glistening under the warm afternoon sun. The tall grass ripples gently in "
+        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
+        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
+        "cinematic.")
+    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True)
 
 
 if __name__ == "__main__":

examples/inference/basic/basic_dmd.py

@@ -1,60 +0,0 @@
-import os
-import time
-from fastvideo import VideoGenerator
-
-from fastvideo.configs.sample import SamplingParam
-
-OUTPUT_PATH = "video_samples_dmd2"
-def main():
-    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "VIDEO_SPARSE_ATTN"
-
-    load_start_time = time.perf_counter()
-    model_name = "FastVideo/FastWan2.1-T2V-1.3B-Diffusers"
-    generator = VideoGenerator.from_pretrained(
-        model_name,
-        # FastVideo will automatically handle distributed setup
-        num_gpus=1,
-        use_fsdp_inference=True,
-        # Adjust these offload parameters if you have < 32GB of VRAM
-        text_encoder_cpu_offload=False,
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        VSA_sparsity=0.8,
-    )
-    load_end_time = time.perf_counter()
-    load_time = load_end_time - load_start_time
-
-
-    sampling_param = SamplingParam.from_pretrained(model_name)
-
-    prompt = (
-        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
-        "wide with interest. The playful yet serene atmosphere is complemented by soft "
-        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
-    )
-    start_time = time.perf_counter()
-    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
-    end_time = time.perf_counter()
-    gen_time = end_time - start_time
-
-    # Generate another video with a different prompt, without reloading the
-    # model!
-    prompt2 = (
-        "A majestic lion strides across the golden savanna, its powerful frame "
-        "glistening under the warm afternoon sun. The tall grass ripples gently in "
-        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-        "cinematic.")
-    start_time = time.perf_counter()
-    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True)
-    end_time = time.perf_counter()
-    gen_time2 = end_time - start_time
-
-
-    print(f"Time taken to load model: {load_time} seconds")
-    print(f"Time taken to generate video: {gen_time} seconds")
-    print(f"Time taken to generate video2: {gen_time2} seconds")
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_i2v.py

@@ -1,42 +0,0 @@
-from fastvideo import VideoGenerator
-
-from fastvideo.configs.sample import SamplingParam
-
-OUTPUT_PATH = "video_samples_5b"
-def main():
-    # FastVideo will automatically use the optimal default arguments for the
-    # model.
-    # If a local path is provided, FastVideo will make a best effort
-    # attempt to identify the optimal arguments.
-    model_id = "Wan-AI/Wan2.2-TI2V-5B-Diffusers"
-    generator = VideoGenerator.from_pretrained(
-        model_id,
-        # FastVideo will automatically handle distributed setup
-        num_gpus=1,
-        use_fsdp_inference=True,
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=False,
-        # image_encoder_cpu_offload=False,
-    )
-
-    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.2-TI2V-5B-Diffusers")
-    # sampling_param.num_frames = 45
-    sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
-    # Generate videos with the same simple API, regardless of GPU count
-    prompt = (
-        "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
-    )
-    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
-    # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/")
-
-    # Generate another video with a different prompt, without reloading the
-    # model!
-    # prompt2 = (
-    #     "A majestic lion strides across the golden savanna, its powerful frame glistening under the warm afternoon sun. The tall grass ripples gently in the breeze, enhancing the lion's commanding presence. The tone is vibrant, embodying the raw energy of the wild. Low angle, steady tracking shot, cinematic."
-    # )
-    # video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_mps.py

@@ -1,40 +0,0 @@
-from fastvideo import VideoGenerator, PipelineConfig
-from fastvideo.configs.sample import SamplingParam
-
-def main():
-    config = PipelineConfig.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
-    config.text_encoder_precisions = ["fp16"]
-    
-    generator = VideoGenerator.from_pretrained(
-        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-        pipeline_config=config,
-        use_fsdp_inference=False,      # Disable FSDP for MPS
-        dit_cpu_offload=True,          
-        text_encoder_cpu_offload=True,    
-        pin_cpu_memory=True,           
-        disable_autocast=False,        
-        num_gpus=1,      
-    )
-
-    # Create sampling parameters with reduced number of frames
-    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
-    sampling_param.num_frames = 3  # Reduce from default 81 to 25 frames bc we have to use the SDPA attn backend for mps
-    sampling_param.height = 256
-    sampling_param.width = 256
-
-    prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
-             "wide with interest. The playful yet serene atmosphere is complemented by soft "
-             "natural light filtering through the petals. Mid-shot, warm and cheerful tones.")
-    
-    video = generator.generate_video(prompt, sampling_param=sampling_param)
-
-    prompt2 = ("A majestic lion strides across the golden savanna, its powerful frame "
-        "glistening under the warm afternoon sun. The tall grass ripples gently in "
-        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-        "cinematic.")
-    
-    video2 = generator.generate_video(prompt2, sampling_param=sampling_param)
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/default_args.py

@@ -1,5 +1,5 @@
 from fastvideo import VideoGenerator
-from fastvideo.configs.pipelines.base import PipelineConfig
+from fastvideo.v1.configs.pipelines.base import PipelineConfig
 
 def main():
 

examples/inference/basic/test.py

@@ -1,39 +0,0 @@
-from fastvideo import VideoGenerator, PipelineConfig, SamplingParam
-
-# from fastvideo.configs.sample import SamplingParam
-
-OUTPUT_PATH = "video_samples_fp16"
-def main():
-    # FastVideo will automatically use the optimal default arguments for the
-    # model.
-    # If a local path is provided, FastVideo will make a best effort
-    # attempt to identify the optimal arguments.
-    model = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-    pipeline_config = PipelineConfig.from_pretrained(model)
-    pipeline_config.text_encoder_precisions = ("bf16", )
-    generator = VideoGenerator.from_pretrained(
-        model,
-        # if num_gpus > 1, FastVideo will automatically handle distributed setup
-        pipeline_config=pipeline_config,
-        use_fsdp_inference=False,      # Disable FSDP for MPS
-        dit_cpu_offload=True,          
-        text_encoder_cpu_offload=True,    
-        pin_cpu_memory=True,           
-        disable_autocast=False,        
-        num_gpus=1, 
-    )
-
-    sampling_param = SamplingParam.from_pretrained(model)
-    sampling_param.num_frames = 30
-    # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
-    # Generate videos with the same simple API, regardless of GPU count
-    prompt = "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."
-    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
-    # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/")
-
-    # Generate another video with a different prompt, without reloading the
-    # model!
-
-
-if __name__ == "__main__":
-    main()

examples/inference/gradio/gradio_demo.py

@@ -6,7 +6,7 @@ import gradio as gr
 import torch
 
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample.base import SamplingParam
+from fastvideo.v1.configs.sample.base import SamplingParam
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="FastVideo Gradio Demo")

examples/inference/lora/wan_lora_inference.py

@@ -1,12 +1,12 @@
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.v1.configs.sample import SamplingParam
 
-OUTPUT_PATH = "./lora_out"
+OUTPUT_PATH = "./lora"
 def main():
     # Initialize VideoGenerator with the Wan model
     generator = VideoGenerator.from_pretrained(
         "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-        num_gpus=1,
+        num_gpus=2,
         lora_path="benjamin-paine/steamboat-willie-1.3b",
         lora_nickname="steamboat"
     )
@@ -16,7 +16,6 @@ def main():
         "num_frames": 81,
         "guidance_scale": 5.0,
         "num_inference_steps": 32,
-        "seed": 42,
     }
     # Generate video with LoRA style
     prompt = "steamboat willie style, golden era animation, close-up of a short fluffy monster  kneeling beside a melting red candle. the mood is one of wonder and curiosity,  as the monster gazes at the flame with wide eyes and open mouth. Its pose and expression  convey a sense of innocence and playfulness, as if it is exploring the world around it for the first time.  The use of warm colors and dramatic lighting further enhances the cozy atmosphere of the image."
@@ -30,7 +29,7 @@ def main():
         negative_prompt=negative_prompt,
         **kwargs
     )
-    
+
     generator.set_lora_adapter(lora_nickname="flat_color", lora_path="motimalu/wan-flat-color-1.3b-v2")
     prompt = "flat color, no lineart, blending, negative space, artist:[john kafka|ponsuke kaikai|hara id 21|yoneyama mai|fuzichoco],  1girl, sakura miko, pink hair, cowboy shot, white shirt, floral print, off shoulder, outdoors, cherry blossom, tree shade, wariza, looking up, falling petals, half-closed eyes, white sky, clouds,  live2d animation, upper body, high quality cinematic video of a woman sitting under a sakura tree. Dreamy and lonely, the camera close-ups on the face of the woman as she turns towards the viewer. The Camera is steady, This is a cowboy shot. The animation is smooth and fluid."
     negative_prompt = "bad quality video,色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"

examples/inference/lora/wan_lora_inference_from_ckpt.py

@@ -1,35 +0,0 @@
-"""
-Inference using a LoRA checkpoint from FastVideo trainer.
-"""
-from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
-
-OUTPUT_PATH = "./lora_out"
-def main():
-    # Initialize VideoGenerator with the Wan model
-    generator = VideoGenerator.from_pretrained(
-        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-        num_gpus=1,
-        lora_path="checkpoints/wan_t2v_finetune_lora/checkpoint-1250/transformer",
-        lora_nickname="crush_smol"
-    )
-    kwargs = {
-        "height": 480,
-        "width": 832,
-        "num_frames": 77,
-        "guidance_scale": 5.0,
-        "num_inference_steps": 50,
-        "seed": 42,
-    }
-    # Generate video with LoRA style
-    prompt = "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table."
-
-    video = generator.generate_video(
-        prompt,
-        output_path=OUTPUT_PATH,
-        save_video=True,
-        **kwargs
-    )    
-
-if __name__ == "__main__":
-    main()

examples/inference/optimizations/teacache_example.py

@@ -9,7 +9,7 @@ def main():
     gen = VideoGenerator.from_pretrained(
         model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
         num_gpus=1,
-        dit_cpu_offload=False,
+        use_cpu_offload=False,
     )
     load_time = time.perf_counter() - start_time
     print(f"Model loading time: {load_time:.2f} seconds")

examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/README.md

@@ -1,16 +1,10 @@
-# Wan2.1-I2V-1.3B-InP Crush-Smol Example
-These are e2e example scripts for finetuning Wan2.1 T2V 1.3B InP on the crush-smol dataset.
+This directory contain e2e examples scripts for finetuning Wan2.1 I2V. 
 
-## Execute the following commands from `FastVideo/` to run training:
-
-### Download crush-smol dataset:
+Execute the following commands from `FastVideo/` to run training:
 
+- Download crush-smol dataset:
 `bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/download_dataset.sh`
-
-### Preprocess the videos and captions into latents:
-
+- Preprocess the videos and captions into latents:
 `bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/preprocess_wan_data_i2v.sh`
-
-### Edit the following file and run finetuning:
-
-`bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/finetune_i2v.sh`
+- Edit the following file and run finetuning:
+`bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/finetune_i2v.sh`

examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/finetune_i2v.sh

@@ -24,7 +24,6 @@ training_args=(
   --num_height 480
   --num_width 832
   --num_frames 77
-  --enable_gradient_checkpointing_type "full"
 )
 
 # Parallel arguments
@@ -84,7 +83,7 @@ miscellaneous_args=(
 torchrun \
   --nnodes 1 \
   --nproc_per_node $NUM_GPUS \
-    fastvideo/training/wan_i2v_training_pipeline.py \
+    fastvideo/v1/training/wan_i2v_training_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
     "${dataset_args[@]}" \

examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/finetune_i2v.slurm

@@ -1,11 +1,13 @@
 #!/bin/bash
 #SBATCH --job-name=i2v
 #SBATCH --partition=main
+#SBATCH --qos=hao
 #SBATCH --nodes=4
 #SBATCH --ntasks=4
 #SBATCH --ntasks-per-node=1
 #SBATCH --gres=gpu:8
 #SBATCH --cpus-per-task=128
+#SBATCH --nodelist=fs-mbz-gpu-[100-850]
 #SBATCH --mem=1440G
 #SBATCH --output=i2v_output/i2v_%j.out
 #SBATCH --error=i2v_output/i2v_%j.err
@@ -58,7 +60,6 @@ training_args=(
   --num_height 480
   --num_width 832
   --num_frames 77
-  --enable_gradient_checkpointing_type "full"
 )
 
 # Parallel arguments
@@ -120,7 +121,7 @@ srun torchrun \
 --node_rank $SLURM_PROCID \
 --rdzv_backend=c10d \
 --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/wan_i2v_training_pipeline.py \
+    fastvideo/v1/training/wan_i2v_training_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
     "${dataset_args[@]}" \

examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/preprocess_wan_data_i2v.sh

@@ -5,19 +5,21 @@ MODEL_PATH="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
 MODEL_TYPE="wan"
 DATA_MERGE_PATH="data/crush-smol/merge.txt"
 OUTPUT_DIR="data/crush-smol_processed_i2v_1_3b_inp/"
+VALIDATION_PATH="examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
 
 torchrun --nproc_per_node=$GPU_NUM \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
+    fastvideo/v1/pipelines/preprocess/v1_preprocess.py \
     --model_path $MODEL_PATH \
     --data_merge_path $DATA_MERGE_PATH \
     --preprocess_video_batch_size 8 \
-    --seed 42 \
     --max_height 480 \
     --max_width 832 \
     --num_frames 77 \
     --dataloader_num_workers 0 \
     --output_dir=$OUTPUT_DIR \
+    --model_type $MODEL_TYPE \
     --train_fps 16 \
+    --validation_dataset_file $VALIDATION_PATH \
     --samples_per_file 8 \
     --flush_frequency 8 \
     --video_length_tolerance_range 5 \

examples/training/finetune/wan_i2v_14b_480p/crush_smol/README.md

@@ -1,16 +1,10 @@
-# Wan2.1-I2V-14B-480P Crush-Smol Example
-These are e2e examples scripts for finetuning Wan2.1 I2V 14B 480P on the crush-smol dataset.
+This directory contain e2e examples scripts for finetuning Wan2.1 I2V. 
 
-## Execute the following commands from `FastVideo/` to run training:
-
-### Download crush-smol dataset:
+Execute the following commands from `FastVideo/` to run training:
 
+- Download crush-smol dataset:
 `bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/download_dataset.sh`
-
-### Preprocess the videos and captions into latents:
-
+- Preprocess the videos and captions into latents:
 `bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/preprocess_wan_data_i2v.sh`
-
-### Edit the following file and run finetuning:
-
+- Edit the following file and run finetuning:
 `bash examples/training/finetune/wan_i2v_14b_480p/crush_smol/finetune_i2v.sh`

examples/training/finetune/wan_i2v_14b_480p/crush_smol/finetune_i2v.sh

@@ -15,7 +15,7 @@ NUM_GPUS=8
 # Training arguments
 training_args=(
   --tracker_project_name "wan_i2v_finetune"
-  --output_dir "checkpoints/wan_i2v_finetune"
+  --output_dir "$DATA_DIR/outputs/wan_i2v_finetune"
   --max_train_steps 2000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -24,7 +24,6 @@ training_args=(
   --num_height 480
   --num_width 832
   --num_frames 77
-  --enable_gradient_checkpointing_type "full"
 )
 
 # Parallel arguments
@@ -84,7 +83,7 @@ miscellaneous_args=(
 torchrun \
   --nnodes 1 \
   --nproc_per_node $NUM_GPUS \
-    fastvideo/training/wan_i2v_training_pipeline.py \
+    fastvideo/v1/training/wan_i2v_training_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
     "${dataset_args[@]}" \

examples/training/finetune/wan_i2v_14b_480p/crush_smol/finetune_i2v.slurm

@@ -1,11 +1,13 @@
 #!/bin/bash
 #SBATCH --job-name=i2v
 #SBATCH --partition=main
+#SBATCH --qos=hao
 #SBATCH --nodes=4
 #SBATCH --ntasks=4
 #SBATCH --ntasks-per-node=1
 #SBATCH --gres=gpu:8
 #SBATCH --cpus-per-task=128
+#SBATCH --nodelist=fs-mbz-gpu-[100-850]
 #SBATCH --mem=1440G
 #SBATCH --output=i2v_output/i2v_%j.out
 #SBATCH --error=i2v_output/i2v_%j.err
@@ -49,7 +51,7 @@ VALIDATION_DATASET_FILE="examples/training/finetune/wan_i2v_14b_480p/crush_smol/
 # Training arguments
 training_args=(
   --tracker_project_name wan_i2v_finetune
-  --output_dir="checkpoints/wan_i2v_finetune"
+  --output_dir="$DATA_DIR/outputs/wan_i2v_finetune_2n"
   --max_train_steps=2000
   --train_batch_size=2
   --train_sp_batch_size 1
@@ -58,7 +60,6 @@ training_args=(
   --num_height 480
   --num_width 832
   --num_frames 77
-  --enable_gradient_checkpointing_type "full"
 )
 
 # Parallel arguments
@@ -120,7 +121,7 @@ srun torchrun \
 --node_rank $SLURM_PROCID \
 --rdzv_backend=c10d \
 --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/wan_i2v_training_pipeline.py \
+    fastvideo/v1/training/wan_i2v_training_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
     "${dataset_args[@]}" \

examples/training/finetune/wan_i2v_14b_480p/crush_smol/finetune_i2v_lora.sh

@@ -1,93 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-MODEL_PATH="Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
-DATA_DIR="data/crush-smol_processed_i2v/combined_parquet_dataset/"
-VALIDATION_DIR="data/crush-smol_processed_i2v/validation_parquet_dataset/"
-NUM_GPUS=4
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_i2v_finetune"
-  --output_dir "checkpoints/wan_i2v_finetune"
-  --max_train_steps 2000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 8
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  --lora_rank 32
-  --lora_training True
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size $NUM_GPUS
-  --tp_size $NUM_GPUS
-  --hsdp_replicate_dim 1
-  --hsdp_shard_dim $NUM_GPUS
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR"
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_preprocessed_path "$VALIDATION_DIR"
-  --validation_steps 100
-  --validation_sampling_steps "40"
-  --validation_guidance_scale "1.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 1000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --allow_tf32
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/wan_i2v_training_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/finetune/wan_i2v_14b_480p/crush_smol/preprocess_wan_data_i2v.sh

@@ -5,19 +5,21 @@ MODEL_PATH="Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
 MODEL_TYPE="wan"
 DATA_MERGE_PATH="data/crush-smol/merge.txt"
 OUTPUT_DIR="data/crush-smol_processed_i2v/"
+VALIDATION_PATH="examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation.json"
 
 torchrun --nproc_per_node=$GPU_NUM \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
+    fastvideo/v1/pipelines/preprocess/v1_preprocess.py \
     --model_path $MODEL_PATH \
     --data_merge_path $DATA_MERGE_PATH \
     --preprocess_video_batch_size 8 \
-    --seed 42 \
     --max_height 480 \
     --max_width 832 \
     --num_frames 77 \
     --dataloader_num_workers 0 \
     --output_dir=$OUTPUT_DIR \
+    --model_type $MODEL_TYPE \
     --train_fps 16 \
+    --validation_dataset_file $VALIDATION_PATH \
     --samples_per_file 8 \
     --flush_frequency 8 \
     --video_length_tolerance_range 5 \

examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation.json

@@ -26,6 +26,24 @@
       "height": 480,
       "width": 832,
       "num_frames": 77
+    },
+    {
+      "caption": "The video shows a close-up of an orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.",
+      "image_path": null,
+      "video_path": "validation_dataset/EJqsC21GSBY-Scene-059.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 98
+    },
+    {
+      "caption": "The video shows a stack of colorful sponges being flattened as if they were under a hydraulic press. The sponges, which are pink, white, blue, and green, are compressed into a smaller size, demonstrating the press's power. The background features a green wall with a yellow and red sign, adding context to the setting.",
+      "image_path": null,
+      "video_path": "validation_dataset/EJqsC21GSBY-Scene-013.mp4",
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 148
     }
   ]
 }

examples/training/finetune/wan_t2v_1_3b/crush_smol/README.md

@@ -1,16 +1,10 @@
-# Wan2.1-T2V-1.3B Crush-Smol Example
-These are e2e example scripts for finetuning Wan2.1 T2V 1.3B on the crush-smol dataset.
+This directory contain e2e examples scripts for finetuning Wan2.1 T2v. 
 
-## Execute the following commands from `FastVideo/` to run training:
-
-### Download crush-smol dataset:
+Execute the following commands from `FastVideo/` to run training:
 
+- Download crush-smol dataset:
 `bash examples/training/finetune/wan_t2v_1_3b/crush_smol/download_dataset.sh`
-
-### Preprocess the videos and captions into latents:
-
+- Preprocess the videos and captions into latents:
 `bash examples/training/finetune/wan_t2v_1_3b/crush_smol/preprocess_wan_data_t2v.sh`
-
-### Edit the following file and run finetuning:
-
+- Edit the following file and run finetuning:
 `bash examples/training/finetune/wan_t2v_1_3b/crush_smol/finetune_t2v.sh`

examples/training/finetune/wan_t2v_1_3b/crush_smol/finetune_t2v.sh

@@ -15,16 +15,15 @@ NUM_GPUS=4
 # Training arguments
 training_args=(
   --tracker_project_name "wan_t2v_finetune"
-  --output_dir "checkpoints/wan_t2v_finetune"
+  --output_dir "outputs/wan_t2v_finetune"
   --max_train_steps 5000
   --train_batch_size 1
   --train_sp_batch_size 1
   --gradient_accumulation_steps 8
-  --num_latent_t 20
+  --num_latent_t 8 
   --num_height 480
   --num_width 832
   --num_frames 77
-  --enable_gradient_checkpointing_type "full"
 )
 
 # Parallel arguments
@@ -61,7 +60,7 @@ validation_args=(
 optimizer_args=(
   --learning_rate 5e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --checkpointing_steps 6000
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )
@@ -83,7 +82,7 @@ miscellaneous_args=(
 torchrun \
   --nnodes 1 \
   --nproc_per_node $NUM_GPUS \
-    fastvideo/training/wan_training_pipeline.py \
+    fastvideo/v1/training/wan_training_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
     "${dataset_args[@]}" \

examples/training/finetune/wan_t2v_1_3b/crush_smol/finetune_t2v.slurm

@@ -1,11 +1,13 @@
 #!/bin/bash
 #SBATCH --job-name=t2v
 #SBATCH --partition=main
+#SBATCH --qos=hao
 #SBATCH --nodes=1
 #SBATCH --ntasks=1
 #SBATCH --ntasks-per-node=1
 #SBATCH --gres=gpu:8
 #SBATCH --cpus-per-task=128
+#SBATCH --nodelist=fs-mbz-gpu-[100-850]
 #SBATCH --mem=1440G
 #SBATCH --output=t2v_output/t2v_%j.out
 #SBATCH --error=t2v_output/t2v_%j.err
@@ -46,7 +48,7 @@ VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1_3b/crush_smol/vali
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_finetune
-  --output_dir="checkpoints/wan_t2v_finetune"
+  --output_dir="outputs/wan_t2v_finetune"
   --max_train_steps=1000
   --train_batch_size=4
   --train_sp_batch_size 1
@@ -55,7 +57,6 @@ training_args=(
   --num_height 480
   --num_width 832
   --num_frames 77
-  --enable_gradient_checkpointing_type "full"
 )
 
 # Parallel arguments
@@ -117,7 +118,7 @@ srun torchrun \
 --node_rank $SLURM_PROCID \
 --rdzv_backend=c10d \
 --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/wan_training_pipeline.py \
+    fastvideo/v1/training/wan_training_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
     "${dataset_args[@]}" \

examples/training/finetune/wan_t2v_1_3b/crush_smol/finetune_t2v_lora.sh

@@ -1,93 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
-VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1_3b/crush_smol/validation.json"
-NUM_GPUS=2
-# export CUDA_VISIBLE_DEVICES=4,5
-
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_t2v_finetune"
-  --output_dir "checkpoints/wan_t2v_finetune_lora"
-  --max_train_steps 5000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 8
-  --num_latent_t 20
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  --lora_rank 32
-  --lora_training True
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS 
-  --sp_size $NUM_GPUS 
-  --tp_size $NUM_GPUS
-  --hsdp_replicate_dim 1
-  --hsdp_shard_dim $NUM_GPUS
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path $DATA_DIR
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation 
-  --validation_dataset_file $VALIDATION_DATASET_FILE
-  --validation_steps 50
-  --validation_sampling_steps "50" 
-  --validation_guidance_scale "1.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 5e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 500
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --allow_tf32
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-)
-
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-  --master_port 29501 \
-    fastvideo/training/wan_training_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/finetune/wan_t2v_1_3b/crush_smol/preprocess_wan_data_t2v.sh

@@ -5,19 +5,21 @@ MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 MODEL_TYPE="wan"
 DATA_MERGE_PATH="data/crush-smol/merge.txt"
 OUTPUT_DIR="data/crush-smol_processed_t2v/"
+VALIDATION_PATH="examples/training/finetune/wan_t2v_1_3b/crush_smol/validation.json"
 
 torchrun --nproc_per_node=$GPU_NUM \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
+    fastvideo/v1/pipelines/preprocess/v1_preprocess.py \
     --model_path $MODEL_PATH \
     --data_merge_path $DATA_MERGE_PATH \
     --preprocess_video_batch_size 8 \
-    --seed 42 \
     --max_height 480 \
     --max_width 832 \
     --num_frames 77 \
     --dataloader_num_workers 0 \
     --output_dir=$OUTPUT_DIR \
+    --model_type $MODEL_TYPE \
     --train_fps 16 \
+    --validation_dataset_file $VALIDATION_PATH \
     --samples_per_file 8 \
     --flush_frequency 8 \
     --video_length_tolerance_range 5 \

fastvideo/__init__.py

@@ -1,6 +1,6 @@
-from fastvideo.configs.pipelines import PipelineConfig
-from fastvideo.configs.sample import SamplingParam
-from fastvideo.entrypoints.video_generator import VideoGenerator
+from fastvideo.v1.configs.pipelines import PipelineConfig
+from fastvideo.v1.configs.sample import SamplingParam
+from fastvideo.v1.entrypoints.video_generator import VideoGenerator
 from fastvideo.version import __version__
 
-__all__ = ["VideoGenerator", "PipelineConfig", "SamplingParam", "__version__"]
+__all__ = ["VideoGenerator", "PipelineConfig", "SamplingParam", "__version__"]

fastvideo/attention/__init__.py

@@ -1,19 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-from fastvideo.attention.backends.abstract import (AttentionBackend,
-                                                   AttentionMetadata,
-                                                   AttentionMetadataBuilder)
-from fastvideo.attention.layer import (DistributedAttention,
-                                       DistributedAttention_VSA, LocalAttention)
-from fastvideo.attention.selector import get_attn_backend
-
-__all__ = [
-    "DistributedAttention",
-    "LocalAttention",
-    "DistributedAttention_VSA",
-    "AttentionBackend",
-    "AttentionMetadata",
-    "AttentionMetadataBuilder",
-    # "AttentionState",
-    "get_attn_backend",
-]

fastvideo/configs/configs.py

@@ -1,145 +0,0 @@
-import dataclasses
-from typing import Any, Optional
-
-from fastvideo.configs.utils import update_config_from_args
-from fastvideo.utils import FlexibleArgumentParser, StoreBoolean
-
-
-@dataclasses.dataclass
-class PreprocessConfig:
-    """Configuration for preprocessing operations."""
-
-    # Model and dataset configuration
-    model_path: str = ""
-    dataset_path: str = ""
-    dataset_output_dir: str = "./output"
-
-    # Dataloader configuration
-    dataloader_num_workers: int = 1
-    preprocess_video_batch_size: int = 2
-
-    # Saver configuration
-    samples_per_file: int = 64
-    flush_frequency: int = 256
-
-    # Video processing parameters
-    max_height: int = 480
-    max_width: int = 848
-    num_frames: int = 163
-    video_length_tolerance_range: float = 2.0
-    train_fps: int = 30
-    speed_factor: float = 1.0
-    drop_short_ratio: float = 1.0
-    do_temporal_sample: bool = False
-
-    # Model configuration
-    training_cfg_rate: float = 0.0
-
-    @staticmethod
-    def add_cli_args(parser: FlexibleArgumentParser,
-                     prefix: str = "preprocess") -> FlexibleArgumentParser:
-        """Add preprocessing configuration arguments to the parser."""
-        prefix_with_dot = f"{prefix}." if (prefix.strip() != "") else ""
-
-        preprocess_args = parser.add_argument_group("Preprocessing Arguments")
-        # Model & Dataset
-        preprocess_args.add_argument(f"--{prefix_with_dot}model-path",
-                                     type=str,
-                                     default=PreprocessConfig.model_path,
-                                     help="Path to the model for preprocessing")
-        preprocess_args.add_argument(
-            f"--{prefix_with_dot}dataset-path",
-            type=str,
-            default=PreprocessConfig.dataset_path,
-            help="Path to the dataset directory for preprocessing")
-        preprocess_args.add_argument(
-            f"--{prefix_with_dot}dataset-output-dir",
-            type=str,
-            default=PreprocessConfig.dataset_output_dir,
-            help="The output directory where the dataset will be written.")
-
-        # Dataloader
-        preprocess_args.add_argument(
-            f"--{prefix_with_dot}dataloader-num-workers",
-            type=int,
-            default=PreprocessConfig.dataloader_num_workers,
-            help=
-            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
-        )
-        preprocess_args.add_argument(
-            f"--{prefix_with_dot}preprocess-video-batch-size",
-            type=int,
-            default=PreprocessConfig.preprocess_video_batch_size,
-            help="Batch size (per device) for the training dataloader.")
-
-        # Saver
-        preprocess_args.add_argument(f"--{prefix_with_dot}samples-per-file",
-                                     type=int,
-                                     default=PreprocessConfig.samples_per_file,
-                                     help="Number of samples per output file")
-        preprocess_args.add_argument(f"--{prefix_with_dot}flush-frequency",
-                                     type=int,
-                                     default=PreprocessConfig.flush_frequency,
-                                     help="How often to save to parquet files")
-
-        # Video processing parameters
-        preprocess_args.add_argument(f"--{prefix_with_dot}max-height",
-                                     type=int,
-                                     default=PreprocessConfig.max_height,
-                                     help="Maximum height for video processing")
-        preprocess_args.add_argument(f"--{prefix_with_dot}max-width",
-                                     type=int,
-                                     default=PreprocessConfig.max_width,
-                                     help="Maximum width for video processing")
-        preprocess_args.add_argument(f"--{prefix_with_dot}num-frames",
-                                     type=int,
-                                     default=PreprocessConfig.num_frames,
-                                     help="Number of frames to process")
-        preprocess_args.add_argument(
-            f"--{prefix_with_dot}video-length-tolerance-range",
-            type=float,
-            default=PreprocessConfig.video_length_tolerance_range,
-            help="Video length tolerance range")
-        preprocess_args.add_argument(f"--{prefix_with_dot}train-fps",
-                                     type=int,
-                                     default=PreprocessConfig.train_fps,
-                                     help="Training FPS")
-        preprocess_args.add_argument(f"--{prefix_with_dot}speed-factor",
-                                     type=float,
-                                     default=PreprocessConfig.speed_factor,
-                                     help="Speed factor for video processing")
-        preprocess_args.add_argument(f"--{prefix_with_dot}drop-short-ratio",
-                                     type=float,
-                                     default=PreprocessConfig.drop_short_ratio,
-                                     help="Ratio for dropping short videos")
-        preprocess_args.add_argument(
-            f"--{prefix_with_dot}do-temporal-sample",
-            action=StoreBoolean,
-            default=PreprocessConfig.do_temporal_sample,
-            help="Whether to do temporal sampling")
-
-        # Model Training configuration
-        preprocess_args.add_argument(f"--{prefix_with_dot}training-cfg-rate",
-                                     type=float,
-                                     default=PreprocessConfig.training_cfg_rate,
-                                     help="Training CFG rate")
-
-        return parser
-
-    @classmethod
-    def from_kwargs(cls, kwargs: dict[str,
-                                      Any]) -> Optional["PreprocessConfig"]:
-        """Create PreprocessConfig from keyword arguments."""
-        preprocess_config = cls()
-        if not update_config_from_args(
-                preprocess_config, kwargs, prefix="preprocess", pop_args=True):
-            return None
-        return preprocess_config
-
-    def check_preprocess_config(self) -> None:
-        if self.dataset_path == "":
-            raise ValueError("dataset_path must be set for preprocess mode")
-        if self.samples_per_file <= 0:
-            raise ValueError("samples_per_file must be greater than 0")
-        if self.flush_frequency <= 0:
-            raise ValueError("flush_frequency must be greater than 0")

fastvideo/configs/models/__init__.py

@@ -1,6 +0,0 @@
-from fastvideo.configs.models.base import ModelConfig
-from fastvideo.configs.models.dits.base import DiTConfig
-from fastvideo.configs.models.encoders.base import EncoderConfig
-from fastvideo.configs.models.vaes.base import VAEConfig
-
-__all__ = ["ModelConfig", "VAEConfig", "DiTConfig", "EncoderConfig"]

fastvideo/configs/models/dits/__init__.py

@@ -1,5 +0,0 @@
-from fastvideo.configs.models.dits.hunyuanvideo import HunyuanVideoConfig
-from fastvideo.configs.models.dits.stepvideo import StepVideoConfig
-from fastvideo.configs.models.dits.wanvideo import WanVideoConfig
-
-__all__ = ["HunyuanVideoConfig", "WanVideoConfig", "StepVideoConfig"]

fastvideo/configs/models/encoders/__init__.py

@@ -1,14 +0,0 @@
-from fastvideo.configs.models.encoders.base import (BaseEncoderOutput,
-                                                    EncoderConfig,
-                                                    ImageEncoderConfig,
-                                                    TextEncoderConfig)
-from fastvideo.configs.models.encoders.clip import (CLIPTextConfig,
-                                                    CLIPVisionConfig)
-from fastvideo.configs.models.encoders.llama import LlamaConfig
-from fastvideo.configs.models.encoders.t5 import T5Config
-
-__all__ = [
-    "EncoderConfig", "TextEncoderConfig", "ImageEncoderConfig",
-    "BaseEncoderOutput", "CLIPTextConfig", "CLIPVisionConfig", "LlamaConfig",
-    "T5Config"
-]

fastvideo/configs/models/vaes/__init__.py

@@ -1,9 +0,0 @@
-from fastvideo.configs.models.vaes.hunyuanvae import HunyuanVAEConfig
-from fastvideo.configs.models.vaes.stepvideovae import StepVideoVAEConfig
-from fastvideo.configs.models.vaes.wanvae import WanVAEConfig
-
-__all__ = [
-    "HunyuanVAEConfig",
-    "WanVAEConfig",
-    "StepVideoVAEConfig",
-]

fastvideo/configs/pipelines/__init__.py

@@ -1,15 +0,0 @@
-from fastvideo.configs.pipelines.base import (PipelineConfig,
-                                              SlidingTileAttnConfig)
-from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
-from fastvideo.configs.pipelines.registry import (
-    get_pipeline_config_cls_from_name)
-from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
-from fastvideo.configs.pipelines.wan import (WanI2V480PConfig, WanI2V720PConfig,
-                                             WanT2V480PConfig, WanT2V720PConfig)
-
-__all__ = [
-    "HunyuanConfig", "FastHunyuanConfig", "PipelineConfig",
-    "SlidingTileAttnConfig", "WanT2V480PConfig", "WanI2V480PConfig",
-    "WanT2V720PConfig", "WanI2V720PConfig", "StepVideoT2VConfig",
-    "get_pipeline_config_cls_from_name"
-]

fastvideo/configs/sample/__init__.py

@@ -1,3 +0,0 @@
-from fastvideo.configs.sample.base import SamplingParam
-
-__all__ = ["SamplingParam"]

fastvideo/data_preprocess/preprocess_text_embeddings.py

@@ -0,0 +1,147 @@
+import argparse
+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__)
+
+
+class T5dataset(Dataset):
+
+    def __init__(
+        self,
+        json_path,
+        vae_debug,
+    ):
+        self.json_path = json_path
+        self.vae_debug = vae_debug
+        with open(self.json_path, "r") as f:
+            train_dataset = json.load(f)
+            self.train_dataset = sorted(train_dataset, key=lambda x: x["latent_path"])
+
+    def __getitem__(self, idx):
+        caption = self.train_dataset[idx]["caption"]
+        filename = self.train_dataset[idx]["latent_path"].split(".")[0]
+        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"]),
+                map_location="cpu",
+            )
+        else:
+            latents = []
+
+        return dict(caption=caption, latents=latents, filename=filename, length=length)
+
+    def __len__(self):
+        return len(self.train_dataset)
+
+
+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)
+
+    videoprocessor = VideoProcessor(vae_scale_factor=8)
+    os.makedirs(args.output_dir, exist_ok=True)
+    os.makedirs(os.path.join(args.output_dir, "video"), exist_ok=True)
+    os.makedirs(os.path.join(args.output_dir, "latent"), exist_ok=True)
+    os.makedirs(os.path.join(args.output_dir, "prompt_embed"), exist_ok=True)
+    os.makedirs(os.path.join(args.output_dir, "prompt_attention_mask"), exist_ok=True)
+
+    latents_json_path = os.path.join(args.output_dir, "videos2caption_temp.json")
+    train_dataset = T5dataset(latents_json_path, args.vae_debug)
+    text_encoder = load_text_encoder(args.model_type, args.model_path, device=device)
+    vae, autocast_type, fps = load_vae(args.model_type, args.model_path)
+    if args.model_type != "wan":
+        vae.enable_tiling()
+    sampler = DistributedSampler(train_dataset, rank=local_rank, num_replicas=world_size, shuffle=True)
+    train_dataloader = DataLoader(
+        train_dataset,
+        sampler=sampler,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    json_data = []
+    for _, data in tqdm(enumerate(train_dataloader), disable=local_rank != 0):
+        with torch.inference_mode():
+            with torch.autocast("cuda", dtype=autocast_type):
+                prompt_embeds, prompt_attention_mask = text_encoder.encode_prompt(prompt=data["caption"], )
+                if args.vae_debug:
+                    latents = data["latents"]
+                    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")
+                    # 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=fps)
+                    item = {}
+                    item["length"] = int(data["length"][idx])
+                    item["latent_path"] = video_name + ".pt"
+                    item["prompt_embed_path"] = video_name + ".pt"
+                    item["prompt_attention_mask"] = video_name + ".pt"
+                    item["caption"] = data["caption"][idx]
+                    json_data.append(item)
+    dist.barrier()
+    local_data = json_data
+    gathered_data = [None] * world_size
+    dist.all_gather_object(gathered_data, local_data)
+    if local_rank == 0:
+        # os.remove(latents_json_path)
+        all_json_data = [item for sublist in gathered_data for item in sublist]
+        with open(os.path.join(args.output_dir, "videos2caption.json"), "w") as f:
+            json.dump(all_json_data, f, indent=4)
+
+
+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",
+        type=int,
+        default=1,
+        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
+    )
+    parser.add_argument(
+        "--train_batch_size",
+        type=int,
+        default=1,
+        help="Batch size (per device) for the training dataloader.",
+    )
+    parser.add_argument("--text_encoder_name", type=str, default="google/t5-v1_1-xxl")
+    parser.add_argument("--cache_dir", type=str, default="./cache_dir")
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default=None,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument("--vae_debug", action="store_true")
+    args = parser.parse_args()
+    main(args)

fastvideo/data_preprocess/preprocess_vae_latents.py

@@ -0,0 +1,110 @@
+import argparse
+import json
+import os
+
+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__)
+
+
+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)
+    train_dataset = getdataset(args)
+    sampler = DistributedSampler(train_dataset, rank=local_rank, num_replicas=world_size, shuffle=True)
+    train_dataloader = DataLoader(
+        train_dataset,
+        sampler=sampler,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    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, autocast_type, fps = load_vae(args.model_type, args.model_path)
+    if args.model_type != "wan":
+        vae.enable_tiling()
+    os.makedirs(args.output_dir, exist_ok=True)
+    os.makedirs(os.path.join(args.output_dir, "latent"), exist_ok=True)
+
+    json_data = []
+    for _, data in tqdm(enumerate(train_dataloader), disable=local_rank != 0):
+        with torch.inference_mode():
+            with torch.autocast("cuda", dtype=autocast_type):
+                latents = vae.encode(data["pixel_values"].to(encoder_device))["latent_dist"].sample()
+            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
+    dist.all_gather_object(gathered_data, local_data)
+    if local_rank == 0:
+        all_json_data = [item for sublist in gathered_data for item in sublist]
+        with open(os.path.join(args.output_dir, "videos2caption_temp.json"), "w") as f:
+            json.dump(all_json_data, f, indent=4)
+
+
+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(
+        "--dataloader_num_workers",
+        type=int,
+        default=1,
+        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
+    )
+    parser.add_argument(
+        "--train_batch_size",
+        type=int,
+        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("--max_height", type=int, default=480)
+    parser.add_argument("--max_width", type=int, default=848)
+    parser.add_argument("--video_length_tolerance_range", type=int, default=2.0)
+    parser.add_argument("--group_frame", action="store_true")  # TODO
+    parser.add_argument("--group_resolution", action="store_true")  # TODO
+    parser.add_argument("--dataset", default="t2v")
+    parser.add_argument("--train_fps", type=int, default=30)
+    parser.add_argument("--use_image_num", type=int, default=0)
+    parser.add_argument("--text_max_length", type=int, default=256)
+    parser.add_argument("--speed_factor", type=float, default=1.0)
+    parser.add_argument("--drop_short_ratio", type=float, default=1.0)
+    # text encoder & vae & diffusion model
+    parser.add_argument("--text_encoder_name", type=str, default="google/t5-v1_1-xxl")
+    parser.add_argument("--cache_dir", type=str, default="./cache_dir")
+    parser.add_argument("--cfg", type=float, default=0.0)
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default=None,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+
+
+    args = parser.parse_args()
+    main(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

@@ -1,20 +1,13 @@
-# SPDX-License-Identifier: Apache-2.0
 from torchvision import transforms
 from torchvision.transforms import Lambda
+from transformers import AutoTokenizer
 
-from fastvideo.dataset.parquet_dataset_map_style import (
-    build_parquet_map_style_dataloader)
-from fastvideo.dataset.preprocessing_datasets import VideoCaptionMergedDataset
-from fastvideo.dataset.transform import (CenterCropResizeVideo, Normalize255,
-                                         TemporalRandomCrop)
-from fastvideo.dataset.validation_dataset import ValidationDataset
+from fastvideo.dataset.t2v_datasets import T2V_dataset
+from fastvideo.dataset.transform import CenterCropResizeVideo, Normalize255, TemporalRandomCrop
 
 
-def getdataset(args) -> VideoCaptionMergedDataset:
-    if args.do_temporal_sample:
-        temporal_sample = TemporalRandomCrop(args.num_frames)  # 16 x
-    else:
-        temporal_sample = None
+def getdataset(args):
+    temporal_sample = TemporalRandomCrop(args.num_frames)  # 16 x
     norm_fun = Lambda(lambda x: 2.0 * x - 1.0)
     resize_topcrop = [
         CenterCropResizeVideo((args.max_height, args.max_width), top_crop=True),
@@ -31,15 +24,74 @@ def getdataset(args) -> VideoCaptionMergedDataset:
         *resize_topcrop,
         norm_fun,
     ])
-    return VideoCaptionMergedDataset(data_merge_path=args.data_merge_path,
-                                     args=args,
-                                     transform=transform,
-                                     temporal_sample=temporal_sample,
-                                     transform_topcrop=transform_topcrop,
-                                     seed=args.seed)
+    # 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)
+    if args.dataset == "t2v":
+        return T2V_dataset(
+            args,
+            transform=transform,
+            temporal_sample=temporal_sample,
+            tokenizer=tokenizer,
+            transform_topcrop=transform_topcrop,
+        )
+
+    raise NotImplementedError(args.dataset)
 
 
-__all__ = [
-    "build_parquet_map_style_dataloader", "ValidationDataset",
-    "VideoCaptionMergedDataset"
-]
+if __name__ == "__main__":
+    import random
+
+    from accelerate import Accelerator
+    from tqdm import tqdm
+
+    from fastvideo.dataset.t2v_datasets import dataset_prog
+
+    args = type(
+        "args",
+        (),
+        {
+            "ae": "CausalVAEModel_4x8x8",
+            "dataset": "t2v",
+            "attention_mode": "xformers",
+            "use_rope": True,
+            "text_max_length": 300,
+            "max_height": 320,
+            "max_width": 240,
+            "num_frames": 1,
+            "use_image_num": 0,
+            "interpolation_scale_t": 1,
+            "interpolation_scale_h": 1,
+            "interpolation_scale_w": 1,
+            "cache_dir": "../cache_dir",
+            "image_data": "/storage/ongoing/new/Open-Sora-Plan-bak/7.14bak/scripts/train_data/image_data.txt",
+            "video_data": "1",
+            "train_fps": 24,
+            "drop_short_ratio": 1.0,
+            "use_img_from_vid": False,
+            "speed_factor": 1.0,
+            "cfg": 0.1,
+            "text_encoder_name": "google/mt5-xxl",
+            "dataloader_num_workers": 10,
+        },
+    )
+    accelerator = Accelerator()
+    dataset = getdataset(args)
+    num = len(dataset_prog.img_cap_list)
+    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]
+        try:
+            caps = [[random.choice(i)] for i in caps]
+        except Exception as e:
+            print(e)
+            # import ipdb;ipdb.set_trace()
+            print(image_data)
+            zero += 1
+            continue
+        assert caps[0] is not None and len(caps[0]) > 0
+    print(num, zero)
+    import ipdb
+
+    ipdb.set_trace()
+    print("end")

fastvideo/dataset/latent_datasets.py

@@ -1,4 +1,3 @@
-# SPDX-License-Identifier: Apache-2.0
 import json
 import os
 import random
@@ -14,39 +13,30 @@ class LatentDataset(Dataset):
         json_path,
         num_latent_t,
         cfg_rate,
-        seed: int = 42,
-    ) -> None:
+    ):
         # data_merge_path: video_dir, latent_dir, prompt_embed_dir, json_path
         self.json_path = json_path
         self.cfg_rate = cfg_rate
-        self.seed = seed
-        # Create a seeded random generator for deterministic CFG
-        self.rng = random.Random(seed)
         self.datase_dir_path = os.path.dirname(json_path)
         self.video_dir = os.path.join(self.datase_dir_path, "video")
         self.latent_dir = os.path.join(self.datase_dir_path, "latent")
-        self.prompt_embed_dir = os.path.join(self.datase_dir_path,
-                                             "prompt_embed")
-        self.prompt_attention_mask_dir = os.path.join(self.datase_dir_path,
-                                                      "prompt_attention_mask")
-        with open(self.json_path) as f:
+        self.prompt_embed_dir = os.path.join(self.datase_dir_path, "prompt_embed")
+        self.prompt_attention_mask_dir = os.path.join(self.datase_dir_path, "prompt_attention_mask")
+        with open(self.json_path, "r") as f:
             self.data_anno = json.load(f)
         # json.load(f) already keeps the order
         # self.data_anno = sorted(self.data_anno, key=lambda x: x['latent_path'])
         self.num_latent_t = num_latent_t
-
+        # just zero embeddings [256, 4096]
         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.get("length", 1) for data_item in self.data_anno
-        ]
+        self.lengths = [data_item["length"] if "length" in data_item else 1 for data_item in self.data_anno]
 
     def __getitem__(self, idx):
         latent_file = self.data_anno[idx]["latent_path"]
         prompt_embed_file = self.data_anno[idx]["prompt_embed_path"]
-        prompt_attention_mask_file = self.data_anno[idx][
-            "prompt_attention_mask"]
+        prompt_attention_mask_file = self.data_anno[idx]["prompt_attention_mask"]
         # load
         latent = torch.load(
             os.path.join(self.latent_dir, latent_file),
@@ -54,7 +44,7 @@ class LatentDataset(Dataset):
             weights_only=True,
         )
         latent = latent.squeeze(0)[:, -self.num_latent_t:]
-        if self.rng.random() < self.cfg_rate:
+        if random.random() < self.cfg_rate:
             prompt_embed = self.uncond_prompt_embed
             prompt_attention_mask = self.uncond_prompt_mask
         else:
@@ -64,8 +54,7 @@ class LatentDataset(Dataset):
                 weights_only=True,
             )
             prompt_attention_mask = torch.load(
-                os.path.join(self.prompt_attention_mask_dir,
-                             prompt_attention_mask_file),
+                os.path.join(self.prompt_attention_mask_dir, prompt_attention_mask_file),
                 map_location="cpu",
                 weights_only=True,
             )
@@ -80,14 +69,14 @@ def latent_collate_function(batch):
     # latent_attn_mask: # b t h w
     # text_attn_mask: b 1 l
     # needs to check if the latent/prompt' size and apply padding & attn mask
-    latents, prompt_embeds, prompt_attention_masks = zip(*batch, strict=True)
+    latents, prompt_embeds, prompt_attention_masks = zip(*batch)
     # calculate max shape
     max_t = max([latent.shape[1] for latent in latents])
     max_h = max([latent.shape[2] for latent in latents])
     max_w = max([latent.shape[3] for latent in latents])
 
     # padding
-    latent_list: list[torch.Tensor] = [
+    latents = [
         torch.nn.functional.pad(
             latent,
             (
@@ -101,14 +90,29 @@ def latent_collate_function(batch):
         ) for latent in latents
     ]
     # attn mask
-    latent_attn_mask = torch.ones(len(latent_list), max_t, max_h, max_w)
+    latent_attn_mask = torch.ones(len(latents), max_t, max_h, max_w)
     # set to 0 if padding
-    for i, latent in enumerate(latent_list):
+    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
 
     prompt_embeds = torch.stack(prompt_embeds, dim=0)
     prompt_attention_masks = torch.stack(prompt_attention_masks, dim=0)
-    latents = torch.stack(latent_list, dim=0)
+    latents = torch.stack(latents, dim=0)
     return latents, prompt_embeds, latent_attn_mask, prompt_attention_masks
+
+
+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)
+    for latent, prompt_embed, latent_attn_mask, prompt_attention_mask in dataloader:
+        print(
+            latent.shape,
+            prompt_embed.shape,
+            latent_attn_mask.shape,
+            prompt_attention_mask.shape,
+        )
+        import pdb
+
+        pdb.set_trace()

fastvideo/dataset/t2v_datasets.py

@@ -0,0 +1,324 @@
+import json
+import math
+import os
+import random
+from collections import Counter
+from os.path import join as opj
+
+import numpy as np
+import torch
+import torchvision
+from einops import rearrange
+from PIL import Image
+from torch.utils.data import Dataset
+
+from fastvideo.utils.dataset_utils import DecordInit
+from fastvideo.utils.logging_ import main_print
+
+
+class SingletonMeta(type):
+    _instances = {}
+
+    def __call__(cls, *args, **kwargs):
+        if cls not in cls._instances:
+            instance = super().__call__(*args, **kwargs)
+            cls._instances[cls] = instance
+        return cls._instances[cls]
+
+
+class DataSetProg(metaclass=SingletonMeta):
+
+    def __init__(self):
+        self.cap_list = []
+        self.elements = []
+        self.num_workers = 1
+        self.n_elements = 0
+        self.worker_elements = dict()
+        self.n_used_elements = dict()
+
+    def set_cap_list(self, num_workers, cap_list, n_elements):
+        self.num_workers = num_workers
+        self.cap_list = cap_list
+        self.n_elements = n_elements
+        self.elements = list(range(n_elements))
+        random.shuffle(self.elements)
+        print(f"n_elements: {len(self.elements)}", flush=True)
+
+        for i in range(self.num_workers):
+            self.n_used_elements[i] = 0
+            per_worker = int(math.ceil(len(self.elements) / float(self.num_workers)))
+            start = i * per_worker
+            end = min(start + per_worker, len(self.elements))
+            self.worker_elements[i] = self.elements[start:end]
+
+    def get_item(self, work_info):
+        if work_info is None:
+            worker_id = 0
+        else:
+            worker_id = work_info.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
+
+
+dataset_prog = DataSetProg()
+
+
+def filter_resolution(h, w, max_h_div_w_ratio=17 / 16, min_h_div_w_ratio=8 / 16):
+    if h / w <= max_h_div_w_ratio and h / w >= min_h_div_w_ratio:
+        return True
+    return False
+
+
+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
+        self.train_fps = args.train_fps
+        self.use_image_num = args.use_image_num
+        self.transform = transform
+        self.transform_topcrop = transform_topcrop
+        self.temporal_sample = temporal_sample
+        self.tokenizer = tokenizer
+        self.text_max_length = args.text_max_length
+        self.cfg = args.cfg
+        self.speed_factor = args.speed_factor
+        self.max_height = args.max_height
+        self.max_width = args.max_width
+        self.drop_short_ratio = args.drop_short_ratio
+        assert self.speed_factor >= 1
+        self.v_decoder = DecordInit()
+        self.video_length_tolerance_range = args.video_length_tolerance_range
+        self.support_Chinese = True
+        if "mt5" not in args.text_encoder_name:
+            self.support_Chinese = False
+
+        cap_list = self.get_cap_list()
+
+        assert len(cap_list) > 0
+        cap_list, self.sample_num_frames = self.define_frame_index(cap_list)
+        self.lengths = self.sample_num_frames
+
+        n_elements = len(cap_list)
+        dataset_prog.set_cap_list(args.dataloader_num_workers, cap_list, n_elements)
+
+        print(f"video length: {len(dataset_prog.cap_list)}", flush=True)
+
+    def set_checkpoint(self, n_used_elements):
+        for i in range(len(dataset_prog.n_used_elements)):
+            dataset_prog.n_used_elements[i] = n_used_elements
+
+    def __len__(self):
+        return dataset_prog.n_elements
+
+    def __getitem__(self, idx):
+
+        data = self.get_data(idx)
+        return data
+
+    def get_data(self, idx):
+        path = dataset_prog.cap_list[idx]["path"]
+        if path.endswith(".mp4"):
+            return self.get_video(idx)
+        else:
+            return self.get_image(idx)
+
+    def get_video(self, idx):
+        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")
+        video = torchvision_video[frame_indices]
+        video = self.transform(video)
+        video = rearrange(video, "t c h w -> c t h w")
+        video = video.to(torch.uint8)
+        assert video.dtype == torch.uint8
+
+        h, w = video.shape[-2:]
+        assert (
+            h / w <= 17 / 16 and h / w >= 8 / 16
+        ), f"Only videos with a ratio (h/w) less than 17/16 and more than 8/16 are supported. But video ({video_path}) found ratio is {round(h / w, 2)} with the shape of {video.shape}"
+
+        video = video.float() / 127.5 - 1.0
+
+        text = dataset_prog.cap_list[idx]["cap"]
+        if not isinstance(text, list):
+            text = [text]
+        text = [random.choice(text)]
+
+        text = text[0] if random.random() > self.cfg else ""
+        text_tokens_and_mask = self.tokenizer(
+            text,
+            max_length=self.text_max_length,
+            padding="max_length",
+            truncation=True,
+            return_attention_mask=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        input_ids = text_tokens_and_mask["input_ids"]
+        cond_mask = text_tokens_and_mask["attention_mask"]
+        return dict(
+            pixel_values=video,
+            text=text,
+            input_ids=input_ids,
+            cond_mask=cond_mask,
+            path=video_path,
+        )
+
+    def get_image(self, idx):
+        image_data = dataset_prog.cap_list[idx]  # [{'path': path, 'cap': cap}, ...]
+
+        image = Image.open(image_data["path"]).convert("RGB")  # [h, w, c]
+        image = torch.from_numpy(np.array(image))  # [h, w, c]
+        image = rearrange(image, "h w c -> c h w").unsqueeze(0)  #  [1 c h w]
+        # for i in image:
+        #     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 = 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 = [random.choice(caps)]
+        text = caps
+        input_ids, cond_mask = [], []
+        text = text[0] if random.random() > self.cfg else ""
+        text_tokens_and_mask = self.tokenizer(
+            text,
+            max_length=self.text_max_length,
+            padding="max_length",
+            truncation=True,
+            return_attention_mask=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        input_ids = text_tokens_and_mask["input_ids"]  # 1, l
+        cond_mask = text_tokens_and_mask["attention_mask"]  # 1, l
+        return dict(
+            pixel_values=image,
+            text=text,
+            input_ids=input_ids,
+            cond_mask=cond_mask,
+            path=image_data["path"],
+        )
+
+    def define_frame_index(self, cap_list):
+        new_cap_list = []
+        sample_num_frames = []
+        cnt_too_long = 0
+        cnt_too_short = 0
+        cnt_no_cap = 0
+        cnt_no_resolution = 0
+        cnt_resolution_mismatch = 0
+        cnt_movie = 0
+        cnt_img = 0
+        for i in cap_list:
+            path = i["path"]
+            cap = i.get("cap", None)
+            # ======no caption=====
+            if cap is None:
+                cnt_no_cap += 1
+                continue
+            if path.endswith(".mp4"):
+                # ======no fps and duration=====
+                duration = i.get("duration", None)
+                fps = i.get("fps", None)
+                if fps is None or duration is None:
+                    continue
+
+                # ======resolution mismatch=====
+                resolution = i.get("resolution", None)
+                if resolution is None:
+                    cnt_no_resolution += 1
+                    continue
+                else:
+                    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"]
+                    aspect = self.max_height / self.max_width
+                    hw_aspect_thr = 1.5
+                    is_pick = filter_resolution(
+                        height,
+                        width,
+                        max_h_div_w_ratio=hw_aspect_thr * aspect,
+                        min_h_div_w_ratio=1 / hw_aspect_thr * aspect,
+                    )
+                    if not is_pick:
+                        print("resolution mismatch")
+                        cnt_resolution_mismatch += 1
+                        continue
+
+                # 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)
+                    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)
+
+                # comment out it to enable dynamic frames training
+                if (len(frame_indices) < self.num_frames and random.random() < self.drop_short_ratio):
+                    cnt_too_short += 1
+                    continue
+
+                #  too long video will be temporal-crop randomly
+                if len(frame_indices) > self.num_frames:
+                    begin_index, end_index = self.temporal_sample(len(frame_indices))
+                    frame_indices = frame_indices[begin_index:end_index]
+                    # frame_indices = frame_indices[:self.num_frames]  # head crop
+                i["sample_frame_index"] = frame_indices.tolist()
+                new_cap_list.append(i)
+                i["sample_num_frames"] = len(i["sample_frame_index"])  # will use in dataloader(group sampler)
+                sample_num_frames.append(i["sample_num_frames"])
+            elif path.endswith(".jpg"):  # image
+                cnt_img += 1
+                new_cap_list.append(i)
+                i["sample_num_frames"] = 1
+                sample_num_frames.append(i["sample_num_frames"])
+            else:
+                raise NameError(
+                    f"Unknown file extension {path.split('.')[-1]}, only support .mp4 for video and .jpg for image")
+        # import ipdb;ipdb.set_trace()
+        main_print(
+            f"no_cap: {cnt_no_cap}, too_long: {cnt_too_long}, too_short: {cnt_too_short}, "
+            f"no_resolution: {cnt_no_resolution}, resolution_mismatch: {cnt_resolution_mismatch}, "
+            f"Counter(sample_num_frames): {Counter(sample_num_frames)}, cnt_movie: {cnt_movie}, cnt_img: {cnt_img}, "
+            f"before filter: {len(cap_list)}, after filter: {len(new_cap_list)}")
+        return new_cap_list, sample_num_frames
+
+    def decord_read(self, path, frame_indices):
+        decord_vr = self.v_decoder(path)
+        video_data = decord_vr.get_batch(frame_indices).asnumpy()
+        video_data = torch.from_numpy(video_data)
+        video_data = video_data.permute(0, 3, 1, 2)  # (T, H, W, C) -> (T C H W)
+        return video_data
+
+    def read_jsons(self, data):
+        cap_lists = []
+        with open(data, "r") as f:
+            folder_anno = [i.strip().split(",") for i in f.readlines() if len(i.strip()) > 0]
+        print(folder_anno)
+        for folder, anno in folder_anno:
+            with open(anno, "r") as f:
+                sub_list = json.load(f)
+            for i in range(len(sub_list)):
+                sub_list[i]["path"] = opj(folder, sub_list[i]["path"])
+            cap_lists += sub_list
+        return cap_lists
+
+    def get_cap_list(self):
+        cap_lists = self.read_jsons(self.data)
+        return cap_lists

fastvideo/dataset/transform.py

@@ -1,20 +1,38 @@
-# SPDX-License-Identifier: Apache-2.0
+import numbers
 import random
 
 import torch
+from PIL import Image
 
 
-def _is_tensor_video_clip(clip) -> bool:
+def _is_tensor_video_clip(clip):
     if not torch.is_tensor(clip):
-        raise TypeError(f"clip should be Tensor. Got {type(clip)}")
+        raise TypeError("clip should be Tensor. Got %s" % type(clip))
 
     if not clip.ndimension() == 4:
-        raise ValueError(f"clip should be 4D. Got {clip.dim()}D")
+        raise ValueError("clip should be 4D. Got %dD" % clip.dim())
 
     return True
 
 
-def crop(clip, i, j, h, w) -> torch.Tensor:
+def center_crop_arr(pil_image, image_size):
+    """
+    Center cropping implementation from ADM.
+    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)
+
+    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)
+
+    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])
+
+
+def crop(clip, i, j, h, w):
     """
     Args:
         clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
@@ -24,11 +42,9 @@ def crop(clip, i, j, h, w) -> torch.Tensor:
     return clip[..., i:i + h, j:j + w]
 
 
-def resize(clip, target_size, interpolation_mode) -> torch.Tensor:
+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,
@@ -38,7 +54,69 @@ def resize(clip, target_size, interpolation_mode) -> torch.Tensor:
     )
 
 
-def center_crop_th_tw(clip, th, tw, top_crop) -> torch.Tensor:
+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}")
+    H, W = clip.size(-2), clip.size(-1)
+    scale_ = target_size[0] / min(H, W)
+    return torch.nn.functional.interpolate(
+        clip,
+        scale_factor=scale_,
+        mode=interpolation_mode,
+        align_corners=True,
+        antialias=True,
+    )
+
+
+def resized_crop(clip, i, j, h, w, size, interpolation_mode="bilinear"):
+    """
+    Do spatial cropping and resizing to the video clip
+    Args:
+        clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        i (int): i in (i,j) i.e coordinates of the upper left corner.
+        j (int): j in (i,j) i.e coordinates of the upper left corner.
+        h (int): Height of the cropped region.
+        w (int): Width of the cropped region.
+        size (tuple(int, int)): height and width of resized clip
+    Returns:
+        clip (torch.tensor): Resized and cropped clip. Size is (T, C, H, W)
+    """
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    clip = crop(clip, i, j, h, w)
+    clip = resize(clip, size, interpolation_mode)
+    return clip
+
+
+def center_crop(clip, crop_size):
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    h, w = clip.size(-2), clip.size(-1)
+    th, tw = crop_size
+    if h < th or w < tw:
+        raise ValueError("height and width must be no smaller than crop_size")
+
+    i = int(round((h - th) / 2.0))
+    j = int(round((w - tw) / 2.0))
+    return crop(clip, i, j, th, tw)
+
+
+def center_crop_using_short_edge(clip):
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    h, w = clip.size(-2), clip.size(-1)
+    if h < w:
+        th, tw = h, h
+        i = 0
+        j = int(round((w - tw) / 2.0))
+    else:
+        th, tw = w, w
+        i = int(round((h - th) / 2.0))
+        j = 0
+    return crop(clip, i, j, th, tw)
+
+
+def center_crop_th_tw(clip, th, tw, top_crop):
     if not _is_tensor_video_clip(clip):
         raise ValueError("clip should be a 4D torch.tensor")
 
@@ -57,7 +135,27 @@ def center_crop_th_tw(clip, th, tw, top_crop) -> torch.Tensor:
     return crop(clip, i, j, new_h, new_w)
 
 
-def normalize_video(clip) -> torch.Tensor:
+def random_shift_crop(clip):
+    """
+    Slide along the long edge, with the short edge as crop size
+    """
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    h, w = clip.size(-2), clip.size(-1)
+
+    if h <= w:
+        short_edge = h
+    else:
+        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()
+    return crop(clip, i, j, th, tw)
+
+
+def normalize_video(clip):
     """
     Convert tensor data type from uint8 to float, divide value by 255.0 and
     permute the dimensions of clip tensor
@@ -68,12 +166,148 @@ def normalize_video(clip) -> torch.Tensor:
     """
     _is_tensor_video_clip(clip)
     if not clip.dtype == torch.uint8:
-        raise TypeError(
-            f"clip tensor should have data type uint8. Got {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
 
 
+def normalize(clip, mean, std, inplace=False):
+    """
+    Args:
+        clip (torch.tensor): Video clip to be normalized. Size is (T, C, H, W)
+        mean (tuple): pixel RGB mean. Size is (3)
+        std (tuple): pixel standard deviation. Size is (3)
+    Returns:
+        normalized clip (torch.tensor): Size is (T, C, H, W)
+    """
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    if not inplace:
+        clip = clip.clone()
+    mean = torch.as_tensor(mean, dtype=clip.dtype, device=clip.device)
+    # print(mean)
+    std = torch.as_tensor(std, dtype=clip.dtype, device=clip.device)
+    clip.sub_(mean[:, None, None, None]).div_(std[:, None, None, None])
+    return clip
+
+
+def hflip(clip):
+    """
+    Args:
+        clip (torch.tensor): Video clip to be normalized. Size is (T, C, H, W)
+    Returns:
+        flipped clip (torch.tensor): Size is (T, C, H, W)
+    """
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    return clip.flip(-1)
+
+
+class RandomCropVideo:
+
+    def __init__(self, size):
+        if isinstance(size, numbers.Number):
+            self.size = (int(size), int(size))
+        else:
+            self.size = size
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        Returns:
+            torch.tensor: randomly cropped video clip.
+                size is (T, C, OH, OW)
+        """
+        i, j, h, w = self.get_params(clip)
+        return crop(clip, i, j, h, w)
+
+    def get_params(self, clip):
+        h, w = clip.shape[-2:]
+        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)}")
+
+        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()
+
+        return i, j, th, tw
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(size={self.size})"
+
+
+class SpatialStrideCropVideo:
+
+    def __init__(self, stride):
+        self.stride = stride
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        Returns:
+            torch.tensor: cropped video clip by stride.
+                size is (T, C, OH, OW)
+        """
+        i, j, h, w = self.get_params(clip)
+        return crop(clip, i, j, h, w)
+
+    def get_params(self, clip):
+        h, w = clip.shape[-2:]
+
+        th, tw = h // self.stride * self.stride, w // self.stride * self.stride
+
+        return 0, 0, th, tw  # from top-left
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(size={self.size})"
+
+
+class LongSideResizeVideo:
+    """
+    First use the long side,
+    then resize to the specified size
+    """
+
+    def __init__(
+        self,
+        size,
+        skip_low_resolution=False,
+        interpolation_mode="bilinear",
+    ):
+        self.size = size
+        self.skip_low_resolution = skip_low_resolution
+        self.interpolation_mode = interpolation_mode
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        Returns:
+            torch.tensor: scale resized video clip.
+                size is (T, C, 512, *) or (T, C, *, 512)
+        """
+        _, _, h, w = clip.shape
+        if self.skip_low_resolution and max(h, w) <= self.size:
+            return clip
+        if h > w:
+            w = int(w * self.size / h)
+            h = self.size
+        else:
+            h = int(h * self.size / w)
+            w = self.size
+        resize_clip = resize(clip, target_size=(h, w), interpolation_mode=self.interpolation_mode)
+        return resize_clip
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}"
+
+
 class CenterCropResizeVideo:
     """
     First use the short side for cropping length,
@@ -85,15 +319,14 @@ class CenterCropResizeVideo:
         size,
         top_crop=False,
         interpolation_mode="bilinear",
-    ) -> None:
+    ):
         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
 
-    def __call__(self, clip) -> torch.Tensor:
+    def __call__(self, clip):
         """
         Args:
             clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
@@ -101,10 +334,9 @@ class CenterCropResizeVideo:
             torch.tensor: scale resized / center cropped video clip.
                 size is (T, C, crop_size, crop_size)
         """
-        clip_center_crop = center_crop_th_tw(clip,
-                                             self.size[0],
-                                             self.size[1],
-                                             top_crop=self.top_crop)
+        # 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)
+        # import ipdb;ipdb.set_trace()
         clip_center_crop_resize = resize(
             clip_center_crop,
             target_size=self.size,
@@ -116,15 +348,132 @@ class CenterCropResizeVideo:
         return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}"
 
 
+class UCFCenterCropVideo:
+    """
+    First scale to the specified size in equal proportion to the short edge,
+    then center cropping
+    """
+
+    def __init__(
+        self,
+        size,
+        interpolation_mode="bilinear",
+    ):
+        if isinstance(size, tuple):
+            if len(size) != 2:
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
+            self.size = size
+        else:
+            self.size = (size, size)
+
+        self.interpolation_mode = interpolation_mode
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        Returns:
+            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_center_crop = center_crop(clip_resize, self.size)
+        return clip_center_crop
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}"
+
+
+class KineticsRandomCropResizeVideo:
+    """
+    Slide along the long edge, with the short edge as crop size. And resie to the desired size.
+    """
+
+    def __init__(
+        self,
+        size,
+        interpolation_mode="bilinear",
+    ):
+        if isinstance(size, tuple):
+            if len(size) != 2:
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
+            self.size = size
+        else:
+            self.size = (size, size)
+
+        self.interpolation_mode = interpolation_mode
+
+    def __call__(self, clip):
+        clip_random_crop = random_shift_crop(clip)
+        clip_resize = resize(clip_random_crop, self.size, self.interpolation_mode)
+        return clip_resize
+
+
+class CenterCropVideo:
+
+    def __init__(
+        self,
+        size,
+        interpolation_mode="bilinear",
+    ):
+        if isinstance(size, tuple):
+            if len(size) != 2:
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
+            self.size = size
+        else:
+            self.size = (size, size)
+
+        self.interpolation_mode = interpolation_mode
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        Returns:
+            torch.tensor: center cropped video clip.
+                size is (T, C, crop_size, crop_size)
+        """
+        clip_center_crop = center_crop(clip, self.size)
+        return clip_center_crop
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}"
+
+
+class Normalize:
+    """
+    Normalize the video clip by mean subtraction and division by standard deviation
+    Args:
+        mean (3-tuple): pixel RGB mean
+        std (3-tuple): pixel RGB standard deviation
+        inplace (boolean): whether do in-place normalization
+    """
+
+    def __init__(self, mean, std, inplace=False):
+        self.mean = mean
+        self.std = std
+        self.inplace = inplace
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): video clip must be normalized. Size is (C, T, H, W)
+        """
+        return normalize(clip, self.mean, self.std, self.inplace)
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(mean={self.mean}, std={self.std}, inplace={self.inplace})"
+
+
 class Normalize255:
     """
     Convert tensor data type from uint8 to float, divide value by 255.0 and
     """
 
-    def __init__(self) -> None:
+    def __init__(self):
         pass
 
-    def __call__(self, clip) -> torch.Tensor:
+    def __call__(self, clip):
         """
         Args:
             clip (torch.tensor, dtype=torch.uint8): Size is (T, C, H, W)
@@ -137,18 +486,123 @@ class Normalize255:
         return self.__class__.__name__
 
 
-class TemporalRandomCrop:
+class RandomHorizontalFlipVideo:
+    """
+    Flip the video clip along the horizontal direction with a given probability
+    Args:
+        p (float): probability of the clip being flipped. Default value is 0.5
+    """
+
+    def __init__(self, p=0.5):
+        self.p = p
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Size is (T, C, H, W)
+        Return:
+            clip (torch.tensor): Size is (T, C, H, W)
+        """
+        if random.random() < self.p:
+            clip = hflip(clip)
+        return clip
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(p={self.p})"
+
+
+#  ------------------------------------------------------------
+#  ---------------------  Sampling  ---------------------------
+#  ------------------------------------------------------------
+class TemporalRandomCrop(object):
     """Temporally crop the given frame indices at a random location.
 
     Args:
         size (int): Desired length of frames will be seen in the model.
     """
 
-    def __init__(self, size) -> None:
+    def __init__(self, size):
         self.size = size
 
-    def __call__(self, total_frames) -> tuple[int, int]:
+    def __call__(self, total_frames):
         rand_end = max(0, total_frames - self.size - 1)
         begin_index = random.randint(0, rand_end)
         end_index = min(begin_index + self.size, total_frames)
         return begin_index, end_index
+
+
+class DynamicSampleDuration(object):
+    """Temporally crop the given frame indices at a random location.
+
+    Args:
+        size (int): Desired length of frames will be seen in the model.
+    """
+
+    def __init__(self, t_stride, extra_1):
+        self.t_stride = t_stride
+        self.extra_1 = extra_1
+
+    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 = random.choice(truncate_t_list)
+        if self.extra_1:
+            truncate_t = truncate_t + 1
+        return 0, truncate_t
+
+
+if __name__ == "__main__":
+    import os
+
+    import numpy as np
+    import torchvision.io as io
+    from torchvision import transforms
+    from torchvision.utils import save_image
+
+    vframes, aframes, info = io.read_video(filename="./v_Archery_g01_c03.avi", pts_unit="sec", output_format="TCHW")
+
+    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
+    total_frames = len(vframes)
+    print(total_frames)
+
+    temporal_sample = TemporalRandomCrop(target_video_len * frame_interval)
+
+    # Sampling video frames
+    start_frame_ind, end_frame_ind = temporal_sample(total_frames)
+    # 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)
+    print(frame_indice)
+
+    select_vframes = vframes[frame_indice]
+    print(select_vframes.shape)
+    print(select_vframes.dtype)
+
+    select_vframes_trans = trans(select_vframes)
+    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)
+    print(select_vframes_trans_int.dtype)
+    print(select_vframes_trans_int.permute(0, 2, 3, 1).shape)
+
+    io.write_video("./test.avi", select_vframes_trans_int.permute(0, 2, 3, 1), fps=8)
+
+    for i in range(target_video_len):
+        save_image(
+            select_vframes_trans[i],
+            os.path.join("./test000", "%04d.png" % i),
+            normalize=True,
+            value_range=(-1, 1),
+        )

fastvideo/distill.py

@@ -0,0 +1,808 @@
+# !/bin/python3
+# isort: skip_file
+import argparse
+import math
+import os
+import time
+from collections import deque
+from copy import deepcopy
+
+import torch
+import torch.distributed as dist
+import wandb
+from accelerate.utils import set_seed
+from diffusers import FlowMatchEulerDiscreteScheduler
+from fastvideo.distill.solver import PCMFMScheduler
+from diffusers.optimization import get_scheduler
+from diffusers.utils import check_min_version
+from peft import LoraConfig
+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.utils.latents_utils import normalize_dit_input
+from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
+from fastvideo.utils.checkpoint import (resume_lora_optimizer, save_checkpoint, save_lora_checkpoint)
+from fastvideo.utils.communications import (broadcast, sp_parallel_dataloader_wrapper)
+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")
+
+
+def main_print(content):
+    if int(os.environ["LOCAL_RANK"]) <= 0:
+        print(content)
+
+
+def reshard_fsdp(model):
+    for m in FSDP.fsdp_modules(model):
+        if m._has_params and m.sharding_strategy is not ShardingStrategy.NO_SHARD:
+            torch.distributed.fsdp._runtime_utils._reshard(m, m._handle, True)
+
+
+def get_norm(model_pred, norms, gradient_accumulation_steps):
+    fro_norm = (
+        torch.linalg.matrix_norm(model_pred, ord="fro") /  # codespell:ignore
+        gradient_accumulation_steps)
+    largest_singular_value = (torch.linalg.matrix_norm(model_pred, ord=2) / gradient_accumulation_steps)
+    absolute_mean = torch.mean(torch.abs(model_pred)) / gradient_accumulation_steps
+    absolute_max = torch.max(torch.abs(model_pred)) / gradient_accumulation_steps
+    dist.all_reduce(fro_norm, op=dist.ReduceOp.AVG)
+    dist.all_reduce(largest_singular_value, op=dist.ReduceOp.AVG)
+    dist.all_reduce(absolute_mean, op=dist.ReduceOp.AVG)
+    norms["fro"] += torch.mean(fro_norm).item()  # codespell:ignore
+    norms["largest singular value"] += torch.mean(largest_singular_value).item()
+    norms["absolute mean"] += absolute_mean.item()
+    norms["absolute max"] += absolute_max.item()
+
+
+def distill_one_step(
+    transformer,
+    model_type,
+    teacher_transformer,
+    ema_transformer,
+    optimizer,
+    lr_scheduler,
+    loader,
+    noise_scheduler,
+    solver,
+    noise_random_generator,
+    gradient_accumulation_steps,
+    sp_size,
+    max_grad_norm,
+    uncond_prompt_embed,
+    uncond_prompt_mask,
+    num_euler_timesteps,
+    multiphase,
+    not_apply_cfg_solver,
+    distill_cfg,
+    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,  # codespell:ignore
+        "largest singular value": 0.0,
+        "absolute mean": 0.0,
+        "absolute max": 0.0,
+    }
+    for _ in range(gradient_accumulation_steps):
+        (
+            latents,
+            encoder_hidden_states,
+            latents_attention_mask,
+            encoder_attention_mask,
+        ) = next(loader)
+        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()
+        if sp_size > 1:
+            broadcast(index)
+        # Add noise according to flow matching.
+        # sigmas = get_sigmas(start_timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
+        sigmas = extract_into_tensor(solver.sigmas, index, model_input.shape)
+        sigmas_prev = extract_into_tensor(solver.sigmas_prev, index, model_input.shape)
+
+        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)
+        noisy_model_input = sigmas * noise + (1.0 - sigmas) * model_input
+        # Predict the noise residual
+        with torch.autocast("cuda", dtype=torch.bfloat16):
+            if args.model_type == "wan":
+                    teacher_kwargs = {
+                    "hidden_states": noisy_model_input,
+                    "encoder_hidden_states": encoder_hidden_states,
+                    "timestep": timesteps,
+                    "return_dict": True,
+                }
+            else:
+                teacher_kwargs = {
+                    "hidden_states": noisy_model_input,
+                    "encoder_hidden_states": encoder_hidden_states,
+                    "timestep": timesteps,
+                    "encoder_attention_mask": encoder_attention_mask,  # B, L
+                    "return_dict": False,
+                }
+            if hunyuan_teacher_disable_cfg:
+                teacher_kwargs["guidance"] = torch.tensor([1000.0],
+                                                          device=noisy_model_input.device,
+                                                          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)
+        with torch.no_grad():
+            w = distill_cfg
+            with torch.autocast("cuda", dtype=torch.bfloat16):
+                if args.model_type == "wan":
+                    cond_teacher_kwargs ={
+                    "hidden_states": noisy_model_input,
+                    "encoder_hidden_states": encoder_hidden_states,
+                    "timestep": timesteps,
+                    "return_dict": True,
+                    }
+                else:
+                    cond_teacher_kwargs = {
+                    "hidden_states": noisy_model_input,
+                    "encoder_hidden_states": encoder_hidden_states,
+                    "timestep": timesteps,
+                    "encoder_attention_mask": encoder_attention_mask,  # B, L
+                    "return_dict": False,
+                    }
+                cond_teacher_output = teacher_transformer(**cond_teacher_kwargs)[0].float()
+            if not_apply_cfg_solver:
+                uncond_teacher_output = cond_teacher_output
+            else:
+                # Get teacher model prediction on noisy_latents and unconditional embedding
+                with torch.autocast("cuda", dtype=torch.bfloat16):
+                    if args.model_type == "wan":
+                        uncond_teacher_kwargs = {
+                        "hidden_states": noisy_model_input,
+                        "encoder_hidden_states":uncond_prompt_embed.unsqueeze(0).expand(bsz, -1, -1),
+                        "timestep": timesteps,
+                        "return_dict": True,
+                        }
+                    else:
+                        uncond_teacher_kwargs = {
+                        "hidden_states": noisy_model_input,
+                        "encoder_hidden_states":uncond_prompt_embed.unsqueeze(0).expand(bsz, -1, -1),
+                        "timestep": timesteps,
+                        "encoder_attention_mask": uncond_prompt_mask.unsqueeze(0).expand(bsz, -1),
+                        "return_dict": False,
+                        }
+                    
+                    uncond_teacher_output = teacher_transformer(**uncond_teacher_kwargs)[0].float()
+
+            teacher_output = uncond_teacher_output + w * (cond_teacher_output - uncond_teacher_output)
+            x_prev = solver.euler_step(noisy_model_input, teacher_output, index)
+
+        # 20.4.12. Get target LCM prediction on x_prev, w, c, t_n
+        with torch.no_grad():
+            with torch.autocast("cuda", dtype=torch.bfloat16):
+                if args.model_type == "wan":
+                    target_pred_kwargs = {
+                        "hidden_states": x_prev.float(),
+                        "encoder_hidden_states": encoder_hidden_states,
+                        "timestep":timesteps_prev,
+                        "return_dict":True,
+                    }
+                else:
+                    target_pred_kwargs = {
+                        "hidden_states": x_prev.float(),
+                        "encoder_hidden_states": encoder_hidden_states,
+                        "timestep":timesteps_prev,
+                        "encoder_attention_mask":encoder_attention_mask,
+                        "return_dict":False,
+                    }
+                if ema_transformer is not None:
+                    target_pred = ema_transformer(**target_pred_kwargs)[0]
+                else:
+                    target_pred = transformer(**target_pred_kwargs)[0]
+
+            target, end_index = solver.euler_style_multiphase_pred(x_prev, target_pred, index, multiphase, True)
+
+        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)
+        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)
+                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)
+                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)
+        loss.backward()
+
+        avg_loss = loss.detach().clone()
+        dist.all_reduce(avg_loss, 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()):
+            with torch.no_grad():
+                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
+
+
+def main(args):
+    torch.backends.cuda.matmul.allow_tf32 = True
+
+    local_rank = int(os.environ["LOCAL_RANK"])
+    rank = int(os.environ["RANK"])
+    world_size = int(os.environ["WORLD_SIZE"])
+    dist.init_process_group("nccl")
+    torch.cuda.set_device(local_rank)
+    device = torch.cuda.current_device()
+    initialize_sequence_parallel_state(args.sp_size)
+
+    # If passed along, set the training seed now. On GPU...
+    if args.seed is not None:
+        # TODO: t within the same seq parallel group should be the same. Noise should be different.
+        set_seed(args.seed + rank)
+    # We use different seeds for the noise generation in each process to ensure that the noise is different in a batch.
+    noise_random_generator = None
+
+    # Handle the repository creation
+    if rank == 0 and args.output_dir is not None:
+        os.makedirs(args.output_dir, exist_ok=True)
+
+    # For mixed precision training we cast all non-trainable weights to half-precision
+    # 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}")
+
+    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)
+    else:
+        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,
+            lora_alpha=args.lora_alpha,
+            init_lora_weights=True,
+            target_modules=["to_k", "to_q", "to_v", "to_out.0"],
+        )
+        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, no_split_modules = get_dit_fsdp_kwargs(
+        transformer,
+        args.fsdp_sharding_startegy,
+        args.use_lora,
+        args.use_cpu_offload,
+        args.master_weight_type,
+    )
+
+    if args.use_lora:
+        transformer.config.lora_rank = args.lora_rank
+        transformer.config.lora_alpha = args.lora_alpha
+        transformer.config.lora_target_modules = ["to_k", "to_q", "to_v", "to_out.0"]
+        transformer._no_split_modules = no_split_modules
+        fsdp_kwargs["auto_wrap_policy"] = fsdp_kwargs["auto_wrap_policy"](transformer)
+
+    transformer = FSDP(
+        transformer,
+        **fsdp_kwargs,
+    )
+    teacher_transformer = FSDP(
+        teacher_transformer,
+        **fsdp_kwargs,
+    )
+    if args.use_ema:
+        ema_transformer = FSDP(
+            ema_transformer,
+            **fsdp_kwargs,
+        )
+    main_print("--> model loaded")
+
+    if args.gradient_checkpointing:
+        apply_fsdp_checkpointing(transformer, no_split_modules, args.selective_checkpointing)
+        apply_fsdp_checkpointing(teacher_transformer, no_split_modules, args.selective_checkpointing)
+        if args.use_ema:
+            apply_fsdp_checkpointing(ema_transformer, no_split_modules, args.selective_checkpointing)
+    # Set model as trainable.
+    transformer.train()
+    teacher_transformer.requires_grad_(False)
+    if args.use_ema:
+        ema_transformer.requires_grad_(False)
+    
+    noise_scheduler = FlowMatchEulerDiscreteScheduler()
+
+    if args.scheduler_type == "pcm_linear_quadratic":
+        linear_steps = int(noise_scheduler.config.num_train_timesteps * args.linear_range)
+        sigmas = linear_quadratic_schedule(
+            noise_scheduler.config.num_train_timesteps,
+            args.linear_quadratic_threshold,
+            linear_steps,
+        )
+        sigmas = torch.tensor(sigmas).to(dtype=torch.float32)
+    else:
+        sigmas = noise_scheduler.sigmas
+    solver = EulerSolver(
+        sigmas.numpy()[::-1],
+        noise_scheduler.config.num_train_timesteps,
+        euler_timesteps=args.num_euler_timesteps,
+    )
+    solver.to(device)
+    params_to_optimize = transformer.parameters()
+    params_to_optimize = list(filter(lambda p: p.requires_grad, params_to_optimize))
+
+    optimizer = torch.optim.AdamW(
+        params_to_optimize,
+        lr=args.learning_rate,
+        betas=(0.9, 0.999),
+        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)
+    main_print(f"optimizer: {optimizer}")
+
+    # todo add lr scheduler
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps * world_size,
+        num_training_steps=args.max_train_steps * world_size,
+        num_cycles=args.lr_num_cycles,
+        power=args.lr_power,
+        last_epoch=init_steps - 1,
+    )
+
+    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))
+
+    train_dataloader = DataLoader(
+        train_dataset,
+        sampler=sampler,
+        collate_fn=latent_collate_function,
+        pin_memory=True,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+        drop_last=True,
+    )
+
+    num_update_steps_per_epoch = math.ceil(
+        len(train_dataloader) / args.gradient_accumulation_steps * args.sp_size / args.train_sp_batch_size)
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    if rank == 0:
+        project = args.tracker_project_name or "fastvideo"
+        wandb.init(project=project, config=args)
+
+    # Train!
+    total_batch_size = (world_size * args.gradient_accumulation_steps / args.sp_size * args.train_sp_batch_size)
+    main_print("***** Running training *****")
+    main_print(f"  Num examples = {len(train_dataset)}")
+    main_print(f"  Dataloader size = {len(train_dataloader)}")
+    main_print(f"  Num Epochs = {args.num_train_epochs}")
+    main_print(f"  Resume training from step {init_steps}")
+    main_print(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    main_print(f"  Total train batch size (w. data & sequence parallel, accumulation) = {total_batch_size}")
+    main_print(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    main_print(f"  Total optimization steps = {args.max_train_steps}")
+    main_print(
+        f"  Total training parameters per FSDP shard = {sum(p.numel() for p in transformer.parameters() if p.requires_grad) / 1e9} B"
+    )
+    # print dtype
+    main_print(f"  Master weight dtype: {transformer.parameters().__next__().dtype}")
+
+    # Potentially load in the weights and states from a previous save
+    if args.resume_from_checkpoint:
+        assert NotImplementedError("resume_from_checkpoint is not supported now.")
+        # TODO
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=init_steps,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=local_rank > 0,
+    )
+
+    loader = sp_parallel_dataloader_wrapper(
+        train_dataloader,
+        device,
+        args.train_batch_size,
+        args.sp_size,
+        args.train_sp_batch_size,
+    )
+
+    step_times = deque(maxlen=100)
+
+    # todo future
+    for i in range(init_steps):
+        next(loader)
+
+    # log_validation(args, transformer, device,
+    #             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 step in range(init_steps + 1, args.max_train_steps + 1):
+        start_time = time.perf_counter()
+        assert args.multi_phased_distill_schedule is not None
+        num_phases = get_num_phases(args.multi_phased_distill_schedule, step)
+
+        loss, grad_norm, pred_norm = distill_one_step(
+            transformer,
+            args.model_type,
+            teacher_transformer,
+            ema_transformer,
+            optimizer,
+            lr_scheduler,
+            loader,
+            noise_scheduler,
+            solver,
+            noise_random_generator,
+            args.gradient_accumulation_steps,
+            args.sp_size,
+            args.max_grad_norm,
+            uncond_prompt_embed,
+            uncond_prompt_mask,
+            args.num_euler_timesteps,
+            num_phases,
+            args.not_apply_cfg_solver,
+            args.distill_cfg,
+            args.ema_decay,
+            args.pred_decay_weight,
+            args.pred_decay_type,
+            args.hunyuan_teacher_disable_cfg,
+        )
+
+        step_time = time.perf_counter() - 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.update(1)
+        if rank == 0:
+            wandb.log(
+                {
+                    "train_loss": loss,
+                    "learning_rate": lr_scheduler.get_last_lr()[0],
+                    "step_time": step_time,
+                    "avg_step_time": avg_step_time,
+                    "grad_norm": grad_norm,
+                    "pred_fro_norm": pred_norm["fro"],  # codespell:ignore
+                    "pred_largest_singular_value": pred_norm["largest singular value"],
+                    "pred_absolute_mean": pred_norm["absolute mean"],
+                    "pred_absolute_max": pred_norm["absolute max"],
+                },
+                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)
+            else:
+                # Your existing checkpoint saving code
+                if args.use_ema:
+                    save_checkpoint(ema_transformer, rank, args.output_dir, step)
+                else:
+                    save_checkpoint(transformer, rank, args.output_dir, step)
+            dist.barrier()
+        if args.log_validation and step % args.validation_steps == 0:
+            log_validation(
+                args,
+                transformer,
+                device,
+                torch.bfloat16,
+                step,
+                scheduler_type=args.scheduler_type,
+                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_ema:
+                log_validation(
+                    args,
+                    ema_transformer,
+                    device,
+                    torch.bfloat16,
+                    step,
+                    scheduler_type=args.scheduler_type,
+                    shift=args.shift,
+                    num_euler_timesteps=args.num_euler_timesteps,
+                    linear_quadratic_threshold=args.linear_quadratic_threshold,
+                    linear_range=args.linear_range,
+                    ema=True,
+                )
+
+    if args.use_lora:
+        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)
+
+    if get_sequence_parallel_state():
+        destroy_sequence_parallel_group()
+
+
+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",
+        type=int,
+        default=10,
+        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
+    )
+    parser.add_argument(
+        "--train_batch_size",
+        type=int,
+        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("--group_frame", action="store_true")  # TODO
+    parser.add_argument("--group_resolution", action="store_true")  # TODO
+
+    # text encoder & vae & diffusion model
+    parser.add_argument("--pretrained_model_name_or_path", type=str)
+    parser.add_argument("--dit_model_name_or_path", type=str)
+    parser.add_argument("--cache_dir", type=str, default="./cache_dir")
+
+    # diffusion setting
+    parser.add_argument("--ema_decay", type=float, default=0.95)
+    parser.add_argument("--ema_start_step", type=int, default=0)
+    parser.add_argument("--cfg", type=float, default=0.1)
+
+    # validation & logs
+    parser.add_argument("--validation_prompt_dir", type=str)
+    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(
+        "--output_dir",
+        type=str,
+        default=None,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--checkpoints_total_limit",
+        type=int,
+        default=None,
+        help=("Max number of checkpoints to store."),
+    )
+    parser.add_argument(
+        "--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`."),
+    )
+    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.'),
+    )
+    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.'),
+    )
+
+    # optimizer & scheduler & Training
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_warmup_steps",
+        type=int,
+        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(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument("--selective_checkpointing", type=float, default=1.0)
+    parser.add_argument(
+        "--allow_tf32",
+        action="store_true",
+        help=("Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
+              " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."),
+    )
+    parser.add_argument(
+        "--use_cpu_offload",
+        action="store_true",
+        help="Whether to use CPU offload for param & gradient & optimizer states.",
+    )
+
+    parser.add_argument("--sp_size", type=int, default=1, help="For sequence parallel")
+    parser.add_argument(
+        "--train_sp_batch_size",
+        type=int,
+        default=1,
+        help="Batch size for sequence parallel training",
+    )
+
+    parser.add_argument(
+        "--use_lora",
+        action="store_true",
+        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("--fsdp_sharding_startegy", default="full")
+
+    # lr_scheduler
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        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(
+        "--lr_num_cycles",
+        type=int,
+        default=1,
+        help="Number of cycles in the learning rate scheduler.",
+    )
+    parser.add_argument(
+        "--lr_power",
+        type=float,
+        default=1.0,
+        help="Power factor of the polynomial scheduler.",
+    )
+    parser.add_argument(
+        "--not_apply_cfg_solver",
+        action="store_true",
+        help="Whether to apply the cfg_solver.",
+    )
+    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(
+        "--linear_quadratic_threshold",
+        type=float,
+        default=0.025,
+        help="Threshold for linear quadratic scheduler.",
+    )
+    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("--use_ema", action="store_true", help="Whether to use EMA.")
+    parser.add_argument("--multi_phased_distill_schedule", type=str, default=None)
+    parser.add_argument("--pred_decay_weight", type=float, default=0.0)
+    parser.add_argument("--pred_decay_type", default="l1")
+    parser.add_argument("--hunyuan_teacher_disable_cfg", action="store_true")
+    parser.add_argument(
+        "--master_weight_type",
+        type=str,
+        default="fp32",
+        help="Weight type to use - fp32 or bf16.",
+    )
+    args = parser.parse_args()
+    main(args)

fastvideo/distill/discriminator.py

@@ -0,0 +1,76 @@
+import torch.nn as nn
+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
+        )
+        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
+        )
+
+        self.conv_out = nn.Conv2d(inner_channel, output_channel, 1, 1, 0)
+
+    def forward(self, x):
+        b, twh, c = x.shape
+        t = twh // (30 * 53)
+        x = x.view(-1, 30 * 53, c)
+        x = x.permute(0, 2, 1)
+        x = x.view(b * t, c, 30, 53)
+        x = self.conv1(x)
+        x = self.conv2(x) + x
+        x = self.conv_out(x)
+        return x
+
+
+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 * (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
+        ])
+
+    def forward(self, features):
+        outputs = []
+
+        def create_custom_forward(module):
+
+            def custom_forward(*inputs):
+                return module(*inputs)
+
+            return custom_forward
+
+        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],
+                #     use_reentrant=False
+                # )
+                out = h(features[i])
+                outputs.append(out)
+        return outputs

fastvideo/distill/solver.py

@@ -0,0 +1,278 @@
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils import BaseOutput, logging
+
+from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class PCMFMSchedulerOutput(BaseOutput):
+    prev_sample: torch.FloatTensor
+
+
+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)))
+
+
+class PCMFMScheduler(SchedulerMixin, ConfigMixin):
+    _compatibles = []
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        shift: float = 1.0,
+        pcm_timesteps: int = 50,
+        linear_quadratic=False,
+        linear_quadratic_threshold=0.025,
+        linear_range=0.5,
+    ):
+        if linear_quadratic:
+            linear_steps = int(num_train_timesteps * linear_range)
+            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 = 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.sigmas = sigmas.numpy()[::-1][self.euler_timesteps]
+        self.sigmas = torch.from_numpy((self.sigmas[::-1].copy()))
+        self.timesteps = self.sigmas * num_train_timesteps
+        self._step_index = None
+        self._begin_index = None
+        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
+        self.sigma_min = self.sigmas[-1].item()
+        self.sigma_max = self.sigmas[0].item()
+
+    @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 scale_noise(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+        noise: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        """
+        Forward process in flow-matching
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The input sample.
+            timestep (`int`, *optional*):
+                The current timestep in the diffusion chain.
+
+        Returns:
+            `torch.FloatTensor`:
+                A scaled input sample.
+        """
+        if self.step_index is None:
+            self._init_step_index(timestep)
+
+        sigma = self.sigmas[self.step_index]
+        sample = sigma * noise + (1.0 - sigma) * sample
+
+        return sample
+
+    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):
+        """
+        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.
+        """
+        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.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._step_index = None
+        self._begin_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 step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+        sample: torch.FloatTensor,
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+    ) -> Union[PCMFMSchedulerOutput, 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.
+            s_churn (`float`):
+            s_tmin  (`float`):
+            s_tmax  (`float`):
+            s_noise (`float`, defaults to 1.0):
+                Scaling factor for noise added to the sample.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            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)
+
+        sample = sample.to(torch.float32)
+
+        sigma = self.sigmas_[self.step_index]
+
+        denoised = sample - model_output * sigma
+        derivative = (sample - denoised) / sigma
+
+        dt = self.sigmas_[self.step_index + 1] - sigma
+        prev_sample = sample + derivative * dt
+        prev_sample = prev_sample.to(model_output.dtype)
+        self._step_index += 1
+
+        if not return_dict:
+            return (prev_sample, )
+
+        return PCMFMSchedulerOutput(prev_sample=prev_sample)
+
+    def __len__(self):
+        return self.config.num_train_timesteps
+
+
+class EulerSolver:
+
+    def __init__(self, sigmas, timesteps=1000, euler_timesteps=50):
+        self.step_ratio = timesteps // euler_timesteps
+        self.euler_timesteps = (np.arange(1, euler_timesteps + 1) * self.step_ratio).round().astype(np.int64) - 1
+        self.euler_timesteps_prev = np.asarray([0] + self.euler_timesteps[:-1].tolist())
+        self.sigmas = sigmas[self.euler_timesteps]
+        self.sigmas_prev = np.asarray([sigmas[0]] +
+                                      sigmas[self.euler_timesteps[:-1]].tolist())  # either use sigma0 or 0
+
+        self.euler_timesteps = torch.from_numpy(self.euler_timesteps).long()
+        self.euler_timesteps_prev = torch.from_numpy(self.euler_timesteps_prev).long()
+        self.sigmas = torch.from_numpy(self.sigmas)
+        self.sigmas_prev = torch.from_numpy(self.sigmas_prev)
+
+    def to(self, device):
+        self.euler_timesteps = self.euler_timesteps.to(device)
+        self.euler_timesteps_prev = self.euler_timesteps_prev.to(device)
+
+        self.sigmas = self.sigmas.to(device)
+        self.sigmas_prev = self.sigmas_prev.to(device)
+        return self
+
+    def euler_step(self, sample, model_pred, timestep_index):
+        sigma = extract_into_tensor(self.sigmas, timestep_index, model_pred.shape)
+        sigma_prev = extract_into_tensor(self.sigmas_prev, timestep_index, model_pred.shape)
+        x_prev = sample + (sigma_prev - sigma) * model_pred
+        return x_prev
+
+    def euler_style_multiphase_pred(
+        self,
+        sample,
+        model_pred,
+        timestep_index,
+        multiphase,
+        is_target=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))
+        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
+        timestep_index_end = inference_indices[last_valid_index]
+
+        if is_target:
+            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)
+        x_prev = sample + (sigma_prev - sigma) * model_pred
+
+        return x_prev, timestep_index_end

fastvideo/distill_adv.py

@@ -0,0 +1,844 @@
+# !/bin/python3
+# isort: skip_file
+import argparse
+import math
+import os
+import time
+from collections import deque
+from copy import deepcopy
+
+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.dataset.latent_datasets import (LatentDataset, latent_collate_function)
+from fastvideo.distill.discriminator import Discriminator
+from fastvideo.distill.solver import EulerSolver, extract_into_tensor
+from fastvideo.utils.latents_utils import normalize_dit_input
+from fastvideo.models.mochi_hf.pipeline_mochi import linear_quadratic_schedule
+from fastvideo.utils.checkpoint import (resume_lora_optimizer, resume_training_generator_discriminator, save_checkpoint,
+                                        save_lora_checkpoint)
+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")
+
+
+def gan_d_loss(
+    discriminator,
+    teacher_transformer,
+    sample_fake,
+    sample_real,
+    timestep,
+    encoder_hidden_states,
+    encoder_attention_mask,
+    weight,
+    discriminator_head_stride,
+):
+    loss = 0.0
+    # collate sample_fake and sample_real
+    with torch.no_grad():
+        fake_features = teacher_transformer(
+            sample_fake,
+            encoder_hidden_states,
+            timestep,
+            encoder_attention_mask,
+            output_features=True,
+            output_features_stride=discriminator_head_stride,
+            return_dict=False,
+        )[1]
+        real_features = teacher_transformer(
+            sample_real,
+            encoder_hidden_states,
+            timestep,
+            encoder_attention_mask,
+            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)
+    return loss
+
+
+def gan_g_loss(
+    discriminator,
+    teacher_transformer,
+    sample_fake,
+    timestep,
+    encoder_hidden_states,
+    encoder_attention_mask,
+    weight,
+    discriminator_head_stride,
+):
+    loss = 0.0
+    features = teacher_transformer(
+        sample_fake,
+        encoder_hidden_states,
+        timestep,
+        encoder_attention_mask,
+        output_features=True,
+        output_features_stride=discriminator_head_stride,
+        return_dict=False,
+    )[1]
+    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)
+    return loss
+
+
+def distill_one_step_adv(
+    transformer,
+    model_type,
+    teacher_transformer,
+    optimizer,
+    discriminator,
+    discriminator_optimizer,
+    lr_scheduler,
+    loader,
+    noise_scheduler,
+    solver,
+    noise_random_generator,
+    sp_size,
+    max_grad_norm,
+    uncond_prompt_embed,
+    uncond_prompt_mask,
+    num_euler_timesteps,
+    multiphase,
+    not_apply_cfg_solver,
+    distill_cfg,
+    adv_weight,
+    discriminator_head_stride,
+):
+    optimizer.zero_grad()
+    discriminator_optimizer.zero_grad()
+
+    (
+        latents,
+        encoder_hidden_states,
+        latents_attention_mask,
+        encoder_attention_mask,
+    ) = next(loader)
+    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()
+    if sp_size > 1:
+        broadcast(index)
+    # Add noise according to flow matching.
+    # sigmas = get_sigmas(start_timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
+    sigmas = extract_into_tensor(solver.sigmas, index, model_input.shape)
+    sigmas_prev = extract_into_tensor(solver.sigmas_prev, index, model_input.shape)
+
+    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)
+    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]
+
+    # if accelerator.is_main_process:
+    model_pred, end_index = solver.euler_style_multiphase_pred(noisy_model_input, model_pred, index, multiphase)
+
+    # # simplified flow matching aka 0-rectified flow matching loss
+    # # target = model_input - noise
+    # target = model_input
+    adv_index = torch.empty_like(end_index)
+    for i in range(end_index.size(0)):
+        adv_index[i] = torch.randint(
+            end_index[i].item(),
+            end_index[i].item() + num_euler_timesteps // multiphase,
+            (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_adv = (sigmas_adv * noise_scheduler.config.num_train_timesteps).view(-1)
+
+    with torch.no_grad():
+        w = distill_cfg
+        with torch.autocast("cuda", dtype=torch.bfloat16):
+            cond_teacher_output = teacher_transformer(
+                noisy_model_input,
+                encoder_hidden_states,
+                timesteps,
+                encoder_attention_mask,  # B, L
+                return_dict=False,
+            )[0].float()
+        if not_apply_cfg_solver:
+            uncond_teacher_output = cond_teacher_output
+        else:
+            # Get teacher model prediction on noisy_latents and unconditional embedding
+            with torch.autocast("cuda", dtype=torch.bfloat16):
+                uncond_teacher_output = teacher_transformer(
+                    noisy_model_input,
+                    uncond_prompt_embed.unsqueeze(0).expand(bsz, -1, -1),
+                    timesteps,
+                    uncond_prompt_mask.unsqueeze(0).expand(bsz, -1),
+                    return_dict=False,
+                )[0].float()
+        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
+    with torch.no_grad():
+        with torch.autocast("cuda", dtype=torch.bfloat16):
+            target_pred = transformer(
+                x_prev.float(),
+                encoder_hidden_states,
+                timesteps_prev,
+                encoder_attention_mask,  # B, L
+                return_dict=False,
+            )[0]
+
+        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)
+
+    huber_c = 0.001
+    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(
+            discriminator,
+            teacher_transformer,
+            fake_adv.float(),
+            timesteps_adv,
+            encoder_hidden_states.float(),
+            encoder_attention_mask,
+            1.0,
+            discriminator_head_stride,
+        )
+    g_loss += g_gan_loss
+    g_loss.backward()
+
+    g_loss = g_loss.detach().clone()
+    dist.all_reduce(g_loss, op=dist.ReduceOp.AVG)
+
+    g_grad_norm = transformer.clip_grad_norm_(max_grad_norm).item()
+    optimizer.step()
+    lr_scheduler.step()
+    optimizer.zero_grad()
+
+    discriminator_optimizer.zero_grad()
+    discriminator.requires_grad_(True)
+    with torch.autocast("cuda", dtype=torch.bfloat16):
+        d_loss = gan_d_loss(
+            discriminator,
+            teacher_transformer,
+            fake_adv.detach(),
+            real_adv.detach(),
+            timesteps_adv,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            1.0,
+            discriminator_head_stride,
+        )
+
+    d_loss.backward()
+    d_grad_norm = discriminator.clip_grad_norm_(max_grad_norm).item()
+    discriminator_optimizer.step()
+    discriminator_optimizer.zero_grad()
+
+    return g_loss, g_grad_norm, d_loss, d_grad_norm
+
+
+def main(args):
+    torch.backends.cuda.matmul.allow_tf32 = True
+
+    local_rank = int(os.environ["LOCAL_RANK"])
+    rank = int(os.environ["RANK"])
+    world_size = int(os.environ["WORLD_SIZE"])
+    dist.init_process_group("nccl")
+    torch.cuda.set_device(local_rank)
+    device = torch.cuda.current_device()
+    initialize_sequence_parallel_state(args.sp_size)
+
+    # If passed along, set the training seed now. On GPU...
+    if args.seed is not None:
+        # TODO: t within the same seq parallel group should be the same. Noise should be different.
+        set_seed(args.seed + rank)
+    # We use different seeds for the noise generation in each process to ensure that the noise is different in a batch.
+    noise_random_generator = None
+
+    # Handle the repository creation
+    if rank == 0 and args.output_dir is not None:
+        os.makedirs(args.output_dir, exist_ok=True)
+
+    # For mixed precision training we cast all non-trainable weights to half-precision
+    # 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
+    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,
+        total_layers=48 if args.model_type == "mochi" else 40,
+    )
+
+    if args.use_lora:
+        transformer.requires_grad_(False)
+        transformer_lora_config = LoraConfig(
+            r=args.lora_rank,
+            lora_alpha=args.lora_alpha,
+            init_lora_weights=True,
+            target_modules=["to_k", "to_q", "to_v", "to_out.0"],
+        )
+        transformer.add_adapter(transformer_lora_config)
+
+    main_print(
+        f"  Total transformer parameters = {sum(p.numel() for p in transformer.parameters() if p.requires_grad) / 1e6} M"
+    )
+    # discriminator
+    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, 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 = no_split_modules
+        fsdp_kwargs["auto_wrap_policy"] = fsdp_kwargs["auto_wrap_policy"](transformer)
+
+    transformer = FSDP(
+        transformer,
+        **fsdp_kwargs,
+    )
+    teacher_transformer = FSDP(
+        teacher_transformer,
+        **fsdp_kwargs,
+    )
+    discriminator = FSDP(
+        discriminator,
+        **discriminator_fsdp_kwargs,
+    )
+    main_print("--> model loaded")
+
+    if args.gradient_checkpointing:
+        apply_fsdp_checkpointing(transformer, no_split_modules, args.selective_checkpointing)
+        apply_fsdp_checkpointing(teacher_transformer, no_split_modules, args.selective_checkpointing)
+    # 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 = torch.tensor(sigmas).to(dtype=torch.float32)
+    else:
+        sigmas = noise_scheduler.sigmas
+    solver = EulerSolver(
+        sigmas.numpy()[::-1],
+        noise_scheduler.config.num_train_timesteps,
+        euler_timesteps=args.num_euler_timesteps,
+    )
+    solver.to(device)
+    params_to_optimize = transformer.parameters()
+    params_to_optimize = list(filter(lambda p: p.requires_grad, params_to_optimize))
+
+    optimizer = torch.optim.AdamW(
+        params_to_optimize,
+        lr=args.learning_rate,
+        betas=(0.9, 0.999),
+        weight_decay=args.weight_decay,
+        eps=1e-8,
+    )
+
+    discriminator_optimizer = torch.optim.AdamW(
+        discriminator.parameters(),
+        lr=args.discriminator_learning_rate,
+        betas=(0, 0.999),
+        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)
+    elif args.resume_from_checkpoint:
+        (
+            transformer,
+            optimizer,
+            discriminator,
+            discriminator_optimizer,
+            init_steps,
+        ) = resume_training_generator_discriminator(
+            transformer,
+            optimizer,
+            discriminator,
+            discriminator_optimizer,
+            args.resume_from_checkpoint,
+            rank,
+        )
+
+    main_print(f"optimizer: {optimizer}")
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps * world_size,
+        num_training_steps=args.max_train_steps * world_size,
+        num_cycles=args.lr_num_cycles,
+        power=args.lr_power,
+        last_epoch=init_steps - 1,
+    )
+
+    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))
+
+    train_dataloader = DataLoader(
+        train_dataset,
+        sampler=sampler,
+        collate_fn=latent_collate_function,
+        pin_memory=True,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+        drop_last=True,
+    )
+    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)
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    if rank == 0:
+        project = args.tracker_project_name or "fastvideo"
+        wandb.init(project=project, config=args)
+
+    # Train!
+    total_batch_size = (world_size * args.gradient_accumulation_steps / args.sp_size * args.train_sp_batch_size)
+    main_print("***** Running training *****")
+    main_print(f"  Num examples = {len(train_dataset)}")
+    main_print(f"  Dataloader size = {len(train_dataloader)}")
+    main_print(f"  Num Epochs = {args.num_train_epochs}")
+    main_print(f"  Resume training from step {init_steps}")
+    main_print(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    main_print(f"  Total train batch size (w. data & sequence parallel, accumulation) = {total_batch_size}")
+    main_print(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    main_print(f"  Total optimization steps = {args.max_train_steps}")
+    main_print(
+        f"  Total training parameters per FSDP shard = {sum(p.numel() for p in transformer.parameters() if p.requires_grad) / 1e9} B"
+    )
+    # print dtype
+    main_print(f"  Master weight dtype: {transformer.parameters().__next__().dtype}")
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=init_steps,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=local_rank > 0,
+    )
+    loader = sp_parallel_dataloader_wrapper(
+        train_dataloader,
+        device,
+        args.train_batch_size,
+        args.sp_size,
+        args.train_sp_batch_size,
+    )
+
+    step_times = deque(maxlen=100)
+
+    # log_validation(args, transformer, device,
+    #             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.perf_counter()
+        (
+            generator_loss,
+            generator_grad_norm,
+            discriminator_loss,
+            discriminator_grad_norm,
+        ) = distill_one_step_adv(
+            transformer,
+            args.model_type,
+            teacher_transformer,
+            optimizer,
+            discriminator,
+            discriminator_optimizer,
+            lr_scheduler,
+            loader,
+            noise_scheduler,
+            solver,
+            noise_random_generator,
+            args.sp_size,
+            args.max_grad_norm,
+            uncond_prompt_embed,
+            uncond_prompt_mask,
+            args.num_euler_timesteps,
+            num_phases,
+            args.not_apply_cfg_solver,
+            args.distill_cfg,
+            args.adv_weight,
+            args.discriminator_head_stride,
+        )
+
+        step_time = time.perf_counter() - 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.update(1)
+        if rank == 0:
+            wandb.log(
+                {
+                    "generator_loss": generator_loss,
+                    "discriminator_loss": discriminator_loss,
+                    "generator_grad_norm": generator_grad_norm,
+                    "discriminator_grad_norm": discriminator_grad_norm,
+                    "learning_rate": lr_scheduler.get_last_lr()[0],
+                    "step_time": step_time,
+                    "avg_step_time": avg_step_time,
+                },
+                step=step,
+            )
+        if step % args.checkpointing_steps == 0:
+            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)
+            else:
+                # Your existing checkpoint saving code
+                # 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:
+            log_validation(
+                args,
+                transformer,
+                device,
+                torch.bfloat16,
+                step,
+                scheduler_type=args.scheduler_type,
+                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)
+    else:
+        save_checkpoint(transformer, rank, args.output_dir, args.max_train_steps)
+
+    if get_sequence_parallel_state():
+        destroy_sequence_parallel_group()
+
+
+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",
+        type=int,
+        default=10,
+        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
+    )
+    parser.add_argument(
+        "--train_batch_size",
+        type=int,
+        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("--group_frame", action="store_true")  # TODO
+    parser.add_argument("--group_resolution", action="store_true")  # TODO
+
+    # text encoder & vae & diffusion model
+    parser.add_argument("--pretrained_model_name_or_path", type=str)
+    parser.add_argument("--dit_model_name_or_path", type=str)
+    parser.add_argument("--cache_dir", type=str, default="./cache_dir")
+
+    # diffusion setting
+    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)
+    # validation & logs
+    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(
+        "--output_dir",
+        type=str,
+        default=None,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--checkpoints_total_limit",
+        type=int,
+        default=None,
+        help=("Max number of checkpoints to store."),
+    )
+    parser.add_argument(
+        "--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`."),
+    )
+    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.'),
+    )
+    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.'),
+    )
+
+    # optimizer & scheduler & Training
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--discriminator_learning_rate",
+        type=float,
+        default=1e-5,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_warmup_steps",
+        type=int,
+        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(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument("--selective_checkpointing", type=float, default=1.0)
+    parser.add_argument(
+        "--allow_tf32",
+        action="store_true",
+        help=("Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
+              " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."),
+    )
+    parser.add_argument(
+        "--use_cpu_offload",
+        action="store_true",
+        help="Whether to use CPU offload for param & gradient & optimizer states.",
+    )
+
+    parser.add_argument("--sp_size", type=int, default=1, help="For sequence parallel")
+    parser.add_argument(
+        "--train_sp_batch_size",
+        type=int,
+        default=1,
+        help="Batch size for sequence parallel training",
+    )
+
+    parser.add_argument(
+        "--use_lora",
+        action="store_true",
+        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("--fsdp_sharding_startegy", default="full")
+    parser.add_argument("--multi_phased_distill_schedule", type=str, default=None)
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+
+    # lr_scheduler
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        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(
+        "--lr_num_cycles",
+        type=int,
+        default=1,
+        help="Number of cycles in the learning rate scheduler.",
+    )
+    parser.add_argument(
+        "--lr_power",
+        type=float,
+        default=1.0,
+        help="Power factor of the polynomial scheduler.",
+    )
+    parser.add_argument(
+        "--not_apply_cfg_solver",
+        action="store_true",
+        help="Whether to apply the cfg_solver.",
+    )
+    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(
+        "--adv_weight",
+        type=float,
+        default=0.1,
+        help="The weight of the adversarial loss.",
+    )
+    parser.add_argument(
+        "--discriminator_head_stride",
+        type=int,
+        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,
+        default=0.025,
+        help="The threshold of the linear quadratic scheduler.",
+    )
+    parser.add_argument(
+        "--master_weight_type",
+        type=str,
+        default="fp32",
+        help="Weight type to use - fp32 or bf16.",
+    )
+    args = parser.parse_args()
+    main(args)

fastvideo/distributed/device_communicators/cpu_communicator.py

@@ -1,148 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# Adapted from: https://github.com/vllm-project/vllm/blob/main/vllm/distributed/device_communicators/cpu_communicator.py
-
-import os
-
-import torch
-from torch.distributed import ProcessGroup
-
-from fastvideo.platforms import current_platform
-from fastvideo.platforms.interface import CpuArchEnum
-
-from .base_device_communicator import DeviceCommunicatorBase
-
-
-class CpuCommunicator(DeviceCommunicatorBase):
-
-    def __init__(self,
-                 cpu_group: ProcessGroup,
-                 device: torch.device | None = None,
-                 device_group: ProcessGroup | None = None,
-                 unique_name: str = ""):
-        super().__init__(cpu_group, device, device_group, unique_name)
-        self.dist_module = torch.distributed
-
-        if (current_platform.get_cpu_architecture()
-                == CpuArchEnum.X86) and hasattr(
-                    torch.ops._C,
-                    "init_shm_manager") and unique_name.startswith("tp"):
-            self.dist_module = _CPUSHMDistributed(self)
-
-    def all_reduce(
-        self,
-        input_: torch.Tensor,
-        op: torch.distributed.ReduceOp | None = torch.distributed.ReduceOp.SUM
-    ) -> torch.Tensor:
-        self.dist_module.all_reduce(input_, group=self.device_group, op=op)
-        return input_
-
-    def gather(self,
-               input_: torch.Tensor,
-               dst: int = 0,
-               dim: int = -1) -> torch.Tensor | None:
-        """
-        NOTE: We assume that the input tensor is on the same device across
-        all the ranks.
-        NOTE: `dst` is the local rank of the destination rank.
-        """
-        world_size = self.world_size
-        assert -input_.dim() <= dim < input_.dim(), (
-            f"Invalid dim ({dim}) for input tensor with shape {input_.size()}")
-        if dim < 0:
-            # Convert negative dim to positive.
-            dim += input_.dim()
-
-        # Allocate output tensor.
-        if self.rank_in_group == dst:
-            gather_list = [torch.empty_like(input_) for _ in range(world_size)]
-        else:
-            gather_list = None
-
-        # Gather.
-        self.dist_module.gather(input_,
-                                gather_list,
-                                dst=self.ranks[dst],
-                                group=self.device_group)
-
-        if self.rank_in_group == dst:
-            output_tensor = torch.cat(gather_list, dim=dim)
-        else:
-            output_tensor = None
-        return output_tensor
-
-    def all_gather(self, input_: torch.Tensor, dim: int = -1) -> torch.Tensor:
-        if dim < 0:
-            # Convert negative dim to positive.
-            dim += input_.dim()
-        input_size = input_.size()
-        # NOTE: we have to use concat-style all-gather here,
-        # stack-style all-gather has compatibility issues with
-        # torch.compile . see https://github.com/pytorch/pytorch/issues/138795
-        output_size = (input_size[0] * self.world_size, ) + input_size[1:]
-        # Allocate output tensor.
-        output_tensor = torch.empty(output_size,
-                                    dtype=input_.dtype,
-                                    device=input_.device)
-        # All-gather.
-        self.dist_module.all_gather_into_tensor(output_tensor,
-                                                input_,
-                                                group=self.device_group)
-
-        # Reshape
-        output_tensor = output_tensor.reshape((self.world_size, ) + input_size)
-        output_tensor = output_tensor.movedim(0, dim)
-        output_tensor = output_tensor.reshape(input_size[:dim] +
-                                              (self.world_size *
-                                               input_size[dim], ) +
-                                              input_size[dim + 1:])
-        return output_tensor
-
-
-class _CPUSHMDistributed:
-
-    def __init__(self, communicator: CpuCommunicator):
-        instance_identifier = os.environ["VLLM_DIST_IDENT"]
-        unique_name = communicator.unique_name
-        instance_identifier = f"{instance_identifier}-{unique_name}"
-        self.communicator = communicator
-
-        group_ranks = [str(rank) for rank in self.communicator.ranks]
-        shm_group_identifier = f"[{'-'.join(group_ranks)}]"
-        self.group_name = f"{instance_identifier}-{shm_group_identifier}-cpushm"
-
-        self.handle = self._init_cpu_shm()
-
-    def _init_cpu_shm(self) -> int:
-        handle = torch.ops._C.init_shm_manager(
-            self.group_name,
-            self.communicator.world_size,
-            self.communicator.rank,
-        )
-        torch.distributed.barrier(self.communicator.device_group)
-        torch.ops._C.join_shm_manager(
-            handle,
-            self.group_name,
-        )
-        torch.distributed.barrier(self.communicator.device_group)
-
-        return int(handle)
-
-    def all_reduce(self,
-                   input: torch.Tensor,
-                   group: ProcessGroup | None = None) -> None:
-        torch.ops._C.shm_allreduce(self.handle, input)
-
-    def gather(self,
-               input: torch.Tensor,
-               gather_list: list[torch.Tensor] | None,
-               dst: int = -1,
-               group: ProcessGroup | None = None) -> None:
-        # Note: different from the torch gather, here we use local dst rank.
-        torch.ops._C.shm_gather(self.handle, input, gather_list,
-                                torch.distributed.get_group_rank(group, dst))
-
-    def all_gather_into_tensor(self,
-                               output: torch.Tensor,
-                               input: torch.Tensor,
-                               group: ProcessGroup | None = None) -> None:
-        torch.ops._C.shm_all_gather(self.handle, input, output)

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
+}