wip

Add Sage Attention 3 Backend (#815 )
[self-forcing] [6/n] Add Ode Init training (#811 )
2025-09-28 05:42:34 +00:00 · 2025-09-24 15:11:38 -07:00 · 2025-09-22 17:58:19 -07:00 · 2025-09-22 12:07:03 -07:00 · 2025-09-20 19:32:10 -07:00 · 2025-09-20 18:15:59 -07:00
185 changed files with 1763 additions and 11082 deletions
@@ -18,7 +18,8 @@ steps:
            - path: 
                - "fastvideo/models/encoders/**"
                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/encoders/**"
+                - "tests/encoders/**"
+                - "tests/utils.py"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -31,7 +32,8 @@ steps:
            - path:
                - "fastvideo/models/vaes/**"
                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/vaes/**"
+                - "tests/vaes/**"
+                - "tests/utils.py"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -44,7 +46,8 @@ steps:
            - path:
                - "fastvideo/models/dits/**"
                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/transformers/**"
+                - "tests/transformers/**"
+                - "tests/utils.py"
                - "fastvideo/layers/**"
                - "fastvideo/attention/**"
                - "pyproject.toml"
@@ -58,6 +61,8 @@ steps:
                  queue: "default"
            - path: 
                - "fastvideo/**/*.py"
+                - "tests/ssim/**"
+                - "tests/utils.py"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -68,9 +73,10 @@ steps:
                agents:
                  queue: "default"
            - path:
-                - "fastvideo/tests/lora/**"
+                - "tests/inference/lora/**"
                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/transformers/**"
+                - "tests/transformers/**"
+                - "tests/utils.py"
                - "fastvideo/pipelines/**"
                - "fastvideo/layers/lora/**"
                - "pyproject.toml"
@@ -84,6 +90,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/training/Vanilla/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -95,6 +102,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/training/*distillation_pipeline.py"
+                - "tests/training/distill/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -106,6 +114,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/training/lora/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -117,6 +126,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/training/VSA/**"
                - "csrc/attn/video_sparse_attn/**"
                - "csrc/attn/video_sparse_attn/tk/**"
                - "csrc/attn/video_sparse_attn/setup.py"
@@ -133,6 +143,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/inference/STA/**"
                - "csrc/attn/sliding_tile_attn/**"
                - "csrc/attn/sliding_tile_attn/setup.py"
                - "csrc/attn/sliding_tile_attn/config_sta.py"
@@ -190,6 +201,7 @@ steps:
            - path:
                - "csrc/attn/vmoba_attn/vmoba/**"
                - "fastvideo/attention/backends/vmoba.py"
+                - "tests/inference/vmoba/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -201,6 +213,8 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/dataset/**"
+                - "tests/workflow/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -50,7 +50,7 @@ else
    exit 1
 fi

-MODAL_TEST_FILE="fastvideo/tests/modal/pr_test.py"
+MODAL_TEST_FILE="tests/modal/pr_test.py"

 if [ -z "${TEST_TYPE:-}" ]; then
    log "Error: TEST_TYPE environment variable is not set"
@@ -127,29 +127,35 @@ jobs:
            encoder-test:
              - 'fastvideo/models/encoders/**'
              - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/encoders/**'
+              - 'tests/encoders/**'
+              - 'tests/utils.py'
              - *common-paths
            vae-test:
              - 'fastvideo/models/vaes/**'
              - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/vaes/**'
+              - 'tests/vaes/**'
+              - 'tests/utils.py'
              - *common-paths
            transformer-test:
              - 'fastvideo/models/dits/**'
              - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/transformers/**'
+              - 'tests/transformers/**'
+              - 'tests/utils.py'
              - 'fastvideo/layers/**'
              - 'fastvideo/attention/**'
              - *common-paths
            training-test:
              - 'fastvideo/**'
+              - 'tests/training/Vanilla/**'
              - *common-paths
            training-test-VSA:
              - 'fastvideo/**'
+              - 'tests/training/VSA/**'
              - *common-paths
              - *vsa-kernel-paths
            inference-test-STA:
              - 'fastvideo/**'
+              - 'tests/inference/STA/**'
              - *common-paths
              - *sta-kernel-paths
            precision-test-STA:
@@ -160,6 +166,8 @@ jobs:
              - *vsa-kernel-paths
            unit-test:
              - 'fastvideo/**'
+              - 'tests/dataset/**'
+              - 'tests/workflow/**'
              - *common-paths

  encoder-test:
@@ -174,7 +182,7 @@ 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 ./tests/encoders -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -192,7 +200,7 @@ jobs:
      gpu_count: 1
      volume_size: 100
      image: "ghcr.io/${{ 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 ./tests/vaes -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -210,7 +218,7 @@ jobs:
      gpu_count: 1
      volume_size: 100
      image: "ghcr.io/${{ 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 ./tests/transformers -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -236,7 +244,7 @@ jobs:
      volume_size: 200
      disk_size: 200
      image: "ghcr.io/${{ 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 ./tests/ssim -vs"
      timeout_minutes: 60
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -255,7 +263,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ 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 ./tests/training/Vanilla -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -275,7 +283,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ 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 ./tests/training/VSA -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -295,7 +303,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ 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 ./tests/inference/STA -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -369,7 +377,7 @@ jobs:
  #     volume_size: 100
  #     disk_size: 100
  #     image: "ghcr.io/${{ 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 ./tests/nightly/test_e2e_overfit_single_sample.py -vs"
  #     timeout_minutes: 30
  #   secrets:
  #     RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -12,9 +12,6 @@ exclude: |
    scripts/.*|
    fastvideo/data_preprocess/.*|
    fastvideo/dataset/.*|
-    fastvideo/distill/.*|
-    fastvideo/distill\.py|
-    fastvideo/distill_adv\.py|
    fastvideo/models/.*|
    fastvideo/sample/.*|
    fastvideo/train\.py|
@@ -44,10 +41,10 @@ repos:
  - id: codespell
    additional_dependencies: ['tomli']
    args: ['--toml', 'pyproject.toml']
- repo: https://github.com/PyCQA/isort
-  rev: 6.0.1
-  hooks:
-  - id: isort
+# - repo: https://github.com/PyCQA/isort
+#   rev: 6.0.1
+#   hooks:
+#   - id: isort
 - repo: https://github.com/jackdewinter/pymarkdown
  rev: v0.9.30
  hooks:
@@ -7,7 +7,7 @@
 FastVideo features an end-to-end unified pipeline for accelerating diffusion models, starting from data preprocessing to model training, finetuning, distillation, and inference. FastVideo is designed to be modular and extensible, allowing users to easily add new optimizations and techniques. Whether it is training-free optimizations or post-training optimizations, FastVideo has you covered.

 <p align="center">
-    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3csdw1isz-Euq8_Q8~baewG8hxjXs2gQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/S7HLCSTh" target="_blank"> <b> WeChat </b> </a> |
+    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3csdw1isz-Euq8_Q8~baewG8hxjXs2gQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/q46BbX6" target="_blank"> <b> WeChat </b> </a> |
 </p>

 <div align="center">
@@ -155,8 +155,8 @@ If you find FastVideo useful, please considering citing our work:
 }

@article{zhang2025vsa,
-  title={VSA: Faster Video Diffusion with Trainable Sparse Attention},
-  author={Zhang, Peiyuan and Huang, Haofeng and Chen, Yongqi and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
+  title={Vsa: Faster video diffusion with trainable sparse attention},
+  author={Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
  journal={arXiv preprint arXiv:2505.13389},
  year={2025}
 }
@@ -20,5 +20,7 @@ setup(
        "License :: OSI Approved :: Apache Software License",
    ],
    python_requires='>=3.12',
-    install_requires=[]
+    install_requires=[
+        "flash-attn >= 2.7.1",
+    ]
 )
@@ -6,8 +6,16 @@ import time
 import os
 import torch
 from typing import Tuple
-from flash_attn import flash_attn_varlen_func  # Use the new flash attention function
-from flash_attn.flash_attn_interface import _flash_attn_varlen_forward, _flash_attn_varlen_backward
+try:
+    from flash_attn import flash_attn_varlen_func  # Use the new flash attention function
+    from flash_attn.flash_attn_interface import _flash_attn_varlen_forward, _flash_attn_varlen_backward
+except ImportError:
+    def _unsupported(*args, **kwargs):
+        raise ImportError("flash-attn is not installed. Please install it, e.g., `pip install flash-attn`.")
+    _flash_attn_varlen_forward = _unsupported
+    _flash_attn_varlen_backward = _unsupported
+    flash_attn_varlen_func = _unsupported
+    
 from functools import lru_cache
 from einops import rearrange

@@ -1,32 +1,40 @@
 (inference-optimizations)=
+
 # Optimizations

 This page describes the various options for speeding up generation times in FastVideo.

 ## Table of Contents
+
 - Optimized Attention Backends
+
  - [Flash Attention](#optimizations-flash)
  - [Sliding Tile Attention](#optimizations-sta)
  - [Sage Attention](#optimizations-sage)
+  - [Sage Attention 3](#optimizations-sage3)

 - Caching Techniques
  - [TeaCache](#optimizations-teacache)

 (optimizations-backends)=
+
 ## Attention Backends

 ### Available Backends
+
 - Torch SDPA: `FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA`
 - Flash Attention 2 and 3: `FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN`
 - Sliding Tile Attention: `FASTVIDEO_ATTENTION_BACKEND=SLIDING_TILE_ATTN`
 - Video Sparse Attention: `FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN`
 - Sage Attention: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN`
+- Sage Attention 3: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN_THREE`

 ### Configuring Backends

 There are two ways to configure the attention backend in FastVideo.

 #### 1. In Python
+
 In python, set the `FASTVIDEO_ATTENTION_BACKEND` environment variable before instantiating `VideoGenerator` like this:

 ```python
@@ -34,6 +42,7 @@ os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLIDING_TILE_ATTN"
 ```

 #### 2. In CLI
+
 You can also set the environment variable on the command line:

 ```bash
@@ -41,6 +50,7 @@ FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN python example.py
 ```

 (optimizations-flash)=
+
 ### Flash Attention

 **`FLASH_ATTN`**
@@ -57,7 +67,7 @@ And if using a Hopper+ GPU (ie H100), installing [Flash Attention 3](https://git
 git clone https://github.com/Dao-AILab/flash-attention.git && cd flash-attention

 cd hopper
-pip install ninja 
+pip install ninja
 python setup.py install
 ```

@@ -66,7 +76,9 @@ FastVideo will automatically detect and use `FA3` if it is installed when using
 :::

 (optimizations-sta)=
+
 ### Sliding Tile Attention
+
 **`SLIDING_TILE_ATTN`**

 ```bash
@@ -76,7 +88,9 @@ pip install st_attn==0.0.4
 Please see [this page](#sta-installation) for more installation instructions.

 (optimizations-vsa)=
+
 ### Video Sparse Attention
+
 **`VIDEO_SPARSE_ATTN`**

 ```bash
@@ -87,19 +101,45 @@ python setup_vsa.py install
 Please see [this page](#vsa-installation) for more installation instructions.

 (optimizations-sage)=
+
 ### Sage Attention
+
 **`SAGE_ATTN`**

 To use [SageAttention](https://github.com/thu-ml/SageAttention) 2.1.1, please compile from source:

 ```bash
 git clone https://github.com/thu-ml/SageAttention.git
-cd sageattention 
+cd sageattention
 python setup.py install  # or pip install -e .
 ```

+(optimizations-sage3)=
+
+### Sage Attention 3
+
+**`SAGE_ATTN_THREE`**
+
+[SageAttention 3](https://huggingface.co/jt-zhang/SageAttention3) is an advanced attention mechanism that leverages FP4 quantization and Blackwell GPU Tensor Cores for significant performance improvements.
+
+#### Hardware Requirements
+
+- RTX5090
+
+#### Installation
+
+Note that Sage Attention 3 requires `python>=3.13`, `torch>=2.8.0`, `CUDA >=12.8`. If you are using `uv` and using `torch==2.8.0` make sure that `sentencepiece==0.2.1` in the pyproject.toml file.
+
+To use Sage Attention 3 in FastVideo, first get access to the SageAttention3 code, then move `sageattn/` and `setup.py` to the directory `fastvideo/attention/backends`, then install from using:
+
+```bash
+python setup.py install
+```
+
 (optimizations-teacache)=
+
 ## Teacache
+
 TeaCache is an optimization technique supported in FastVideo that can significantly speed up video generation by skipping redundant calculations across diffusion steps. This guide explains how to enable and configure TeaCache for optimal performance in FastVideo.

 ### What is TeaCache?
@@ -1,3 +0,0 @@
-#!/bin/bash
-
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
@@ -1,76 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-034.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-027.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-030.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-016.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a green and orange object being flattened as if it were under a hydraulic press, with the press moving down and compressing the object.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-056.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a cylindrical object with a cityscape image being flattened as if it were under a hydraulic press. The object is placed on a metal platform, and a large, striped cylinder presses down on it, causing it to collapse and release a liquid inside. The background features a green wall with a yellow and red warning sign.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-059.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a close-up of an orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.",
-      "image_path": null,
-      "video_path": "validation_dataset/EJqsC21GSBY-Scene-059.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A colorful puzzle ball is being crushed by a large metal cylinder, which flattens the objects as if they were under a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/GBSfpTcKegk-Scene-003.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}
@@ -1,13 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-016.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}
@@ -16,30 +16,30 @@ export NCCL_P2P_DISABLE=1
 export TORCH_NCCL_ENABLE_MONITORING=0
 # different cache dir for different processes
 export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29501
+export MASTER_PORT=29503
 export TOKENIZERS_PARALLELISM=false
-export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+export WANDB_API_KEY=your_wandb_api_key
 export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=offline
+export WANDB_MODE=online
 export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN

 # Configs
-NUM_GPUS=4
+NUM_GPUS=1

 # Model paths for Self-Forcing DMD distillation:
 GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
 REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
 FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model

-DATA_DIR="data/mixkit-64_processed/Node_0_GPU_1_File_1/combined_parquet_dataset"
-VALIDATION_DATASET_FILE="data/mixkit-64_processed/validation.json"
+DATA_DIR=your_data_dir
+VALIDATION_DATASET_FILE=your_validation_data_dir
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
-  --output_dir "/mnt/sharefs/users/hao.zhang/SFwan_t2v_finetune"
+  --output_dir your_output_dir
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -84,7 +84,7 @@ dataset_args=(
 validation_args=(
  --log_validation
  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 100
+  --validation_steps 50
  --validation_sampling_steps "4"
  --validation_guidance_scale "6.0" # not used for dmd inference
 )
@@ -111,7 +111,7 @@ miscellaneous_args=(
  --use_ema True
  --ema_decay 0.99
  --ema_start_step 100
-  --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
+  --init_weights_from_safetensors your_ode_init_weights_path
 )

 # Self-forcing DMD arguments
@@ -148,4 +148,4 @@ torchrun \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
    "${dmd_args[@]}" \
-    "${self_forcing_args[@]}"
+    "${self_forcing_args[@]}"
@@ -1,3 +1,3 @@
 #!/bin/bash

-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
@@ -21,4 +21,4 @@ torchrun --nproc_per_node=$GPU_NUM \
    --samples_per_file 8 \
    --flush_frequency 8 \
    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "t2v" 
@@ -41,13 +41,14 @@ 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
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # 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"
+  --output_dir $OUTPUT_DIR
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
  --weight_only_checkpointing_steps 500
@@ -134,4 +135,4 @@ srun torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -41,13 +41,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # 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"
+  --output_dir "$OUTPUT_DIR"
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
  --weight_only_checkpointing_steps 500
@@ -134,4 +135,4 @@ srun torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -41,13 +41,14 @@ 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
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # 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"
+  --output_dir "$OUTPUT_DIR"
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
  --weight_only_checkpointing_steps 500
@@ -133,4 +134,4 @@ srun torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -1,3 +1,3 @@
 #!/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"
+python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"
@@ -42,13 +42,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DIR=your_validation_path  #(example:validation_64.json)
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
  --tracker_project_name Wan_distillation
-  --output_dir "your_output_dir"
+  --output_dir "$OUTPUT_DIR"
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -92,11 +93,11 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-5
+  --learning_rate 4e-6
  --lr_scheduler "cosine_with_min_lr"
  --min_lr_ratio 0.5
  --lr_warmup_steps 100
-  --fake_score_learning_rate 1e-5
+  --fake_score_learning_rate 2e-6
  --fake_score_lr_scheduler "cosine_with_min_lr"
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
@@ -141,4 +142,4 @@ srun torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -92,11 +92,11 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-5
+  --learning_rate 4e-6
  --lr_scheduler "cosine_with_min_lr"
  --min_lr_ratio 0.5
  --lr_warmup_steps 100
-  --fake_score_learning_rate 1e-5
+  --fake_score_learning_rate 2e-6
  --fake_score_lr_scheduler "cosine_with_min_lr"
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
@@ -142,4 +142,4 @@ srun torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -16,24 +16,25 @@ NUM_GPUS=1
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR="data/crush-smol_processed_ti2v/combined_parquet_dataset/"
 VALIDATION_DATASET_FILE="examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/validation.json"
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # 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
+  --output_dir "$OUTPUT_DIR"
+  --max_train_steps 4000
+  --train_batch_size 1
  --train_sp_batch_size 1
-  --gradient_accumulation_steps=1
+  --gradient_accumulation_steps 1
  --num_latent_t 31
  --num_height 704
  --num_width 1280
  --num_frames 121
  --enable_gradient_checkpointing_type "full"
-  --training_state_checkpointing_steps=500
-  --weight_only_checkpointing_steps=500
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
 )

 # Parallel arguments
@@ -68,8 +69,8 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate=1e-5
-  --mixed_precision="bf16"
+  --learning_rate 2e-6
+  --mixed_precision "bf16"
  --weight_decay 0.01
  --max_grad_norm 1.0
 )
@@ -107,4 +108,4 @@ torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -109,4 +109,4 @@ torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -1,3 +1,3 @@
 #!/bin/bash

-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
@@ -21,4 +21,4 @@ torchrun --nproc_per_node=$GPU_NUM \
    --samples_per_file 8 \
    --flush_frequency 8 \
    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "t2v" 
@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator

-from fastvideo.configs.sample import SamplingParam
+# from fastvideo.configs.sample import SamplingParam

 OUTPUT_PATH = "video_samples"
 def main():
@@ -17,30 +17,31 @@ def main():
        vae_cpu_offload=False,
        text_encoder_cpu_offload=True,
        pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
-        ti2v_task=True,
        # image_encoder_cpu_offload=False,
    )

-    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
+    # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
    # sampling_param.num_frames = 45
-    sampling_param.image_path = "test.jpg"
+    # 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 girl is packing a suitcase when stuff suddently starts flying around the room."
+        "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)
    # 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)
+    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)


 if __name__ == "__main__":
-    main()
+    main()
@@ -19,15 +19,13 @@ def main():
    )

    sampling_param = SamplingParam.from_pretrained(model_name)
-    sampling_param.num_frames = 81

-    prompts = [
-        "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
-        "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
-    ]
-
-    for prompt in prompts:
-        video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
+    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, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)

 if __name__ == "__main__":
    main()
@@ -0,0 +1,47 @@
+A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.
+The video shows a green and orange object being flattened as if it were under a hydraulic press, with the press moving down and compressing the object.
+The video shows a cylindrical object with a cityscape image being flattened as if it were under a hydraulic press. The object is placed on a metal platform, and a large, striped cylinder presses down on it, causing it to collapse and release a liquid inside. The background features a green wall with a yellow and red warning sign.
+A red toy car is being crushed by a large hydraulic press, which is flattening objects as if they were under a hydraulic press.
+A large, cylindrical object is seen pressing down on a small orange ball, causing it to flatten as if it were under a hydraulic press. The background features a green wall with yellow and red warning signs.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is shown compressing a wooden object, which shatters into small pieces. The background features a green wall with a yellow sign displaying a lightning bolt.
+A large metal cylinder is seen descending, flattening objects as if they were under a hydraulic press. The cylinder compresses a stack of matches and boxes, causing them to crumble into small pieces. The scene is set against a green background with yellow and red signs.
+A large metal press is shown compressing a pile of colorful macarons, flattening them as if they were under a hydraulic press. The press moves down, crushing the macarons into a pile of crumbs and squishing the colorful filling out.
+The video shows a metal press flattening objects as if they were under a hydraulic press. The press is pressing down on a pile of colorful gummy candies, squishing them into a pile of squiggly shapes. The press is made of metal and has a large base, and the gummy candies are of various colors, including red, green, and orange. The background is a green wall, and the press is placed on a metal surface.
+A pile of colorful candies is being flattened by a hydraulic press, causing them to crumble into small pieces.
+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.
+A bowling ball is placed on a metal platform, and a large metal cylinder descends from above, flattening the ball as if it were under a hydraulic press. The ball is crushed into a flat, round shape, leaving a pile of debris around it.
+A large metal cylinder with yellow and black stripes is seen pressing down on a pile of popcorn, flattening the objects as if they were under a hydraulic press.
+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.
+The video shows a close-up of a metal cylinder pressing down on a yellow object, which is being flattened as if it were under a hydraulic press. The cylinder is positioned above the object, and the force is causing the object to compress and spread out, creating a visible deformation. The background is blurred, focusing attention on the action of the cylinder and the object being flattened.
+A colorful puzzle ball is being crushed by a large metal cylinder, which flattens the objects as if they were under a hydraulic press.
+The video shows a hydraulic press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing two colorful objects that resemble sandwiches. The press is yellow and black striped, and the objects being flattened are placed on a metal plate. The background is green, and the press is moving down, compressing the objects.
+The scene shows a metal press with a yellow and black striped pattern, holding a container filled with chocolate. A metal cylinder is descending, flattening the chocolate as if it were under a hydraulic press. The background is a green wall, and the press is mounted on a sturdy metal frame.
+The video shows a colorful sponge being flattened as if it were under a hydraulic press, with the sponge being compressed and eventually flattened into a thin layer.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is pressing down on a stack of wooden blocks, causing them to crumble and break apart. The press is black and yellow striped, and the wooden blocks are small and rectangular. The background is green, and the press is sitting on a metal table.
+A pile of colorful candies is being flattened by a hydraulic press, causing them to crumble into small pieces.
+The video shows a stack of colorful sponges being flattened by a large, cylindrical object, which appears to be a hydraulic press. The sponges, which are pink, blue, white, and green, are compressed into a single layer, demonstrating the press's powerful force. The background features a green wall with a yellow and red sign, adding context to the industrial setting.
+A bowling ball is placed on a metal platform, and a large metal cylinder descends from above, flattening the ball as if it were under a hydraulic press. The ball is crushed into a flat, round shape, demonstrating the immense pressure applied by the cylinder.
+A large metal cylinder with yellow and black stripes is seen pressing down on a pile of popcorn, flattening the objects as if they were under a hydraulic press. The popcorn is crushed and scattered around the base of the cylinder, creating a satisfying visual effect.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is composed of a large, cylindrical metal cylinder with yellow and black stripes, and a metal base. The objects being flattened are two cylindrical blocks of cotton candy, one pink and one blue. The press is positioned on a metal table, and the background features a green wall with a yellow and red sign.
+The video shows a large 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.
+The video shows a cylindrical object being pressed down onto a flat surface, causing the objects beneath it to be flattened as if they were under a hydraulic press. The objects being flattened appear to be yellow and are being crushed into a pile of debris. The background is a greenish-gray color, and the surface on which the objects are being flattened is metallic and shiny.
+A green and blue object with a spiky texture is being flattened by a large, cylindrical metal press, demonstrating its resilience and durability.
+The video shows a stack of caramelized sugar cubes being flattened as if they were under a hydraulic press, resulting in a messy pile of broken sugar on the table.
+A large metal cylinder is seen pressing down on a pile of colorful jelly beans, flattening them as if they were under a hydraulic press.
+The video shows a machine with a yellow and black striped cylinder pressing down on a stack of colorful sponges, flattening them as if they were under a hydraulic press. The machine is situated in a green-walled room with warning signs in the background.
+The video shows a machine with a yellow and black striped cylinder, which is pressing down on two colorful objects, flattening them as if they were under a hydraulic press. The machine appears to be in a workshop or industrial setting, with a green wall in the background. The objects being flattened are green and orange, and the machine is covered in dirt and grime, indicating it has been used frequently.
+The video shows a large, industrial press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing a pile of pink objects into a pile of crumbs. The press is large and metallic, with a yellow and black striped pattern on its side. The background is a green wall with a yellow warning sign.
+The video shows a pink, sparkly ball being crushed by a large, rusty cylinder, which flattens the objects as if they were under a hydraulic press.
+A lime is being crushed by a hydraulic press, causing it to flatten and burst open, releasing its juice and segments.
+The video shows a machine with a yellow and black striped cylinder, which is flattening objects as if they were under a hydraulic press. The machine is pressing down on two colorful objects, causing them to compress and flatten. The background is a green wall, and the machine appears to be in a workshop or industrial setting.
+The video shows a large, yellow and black striped cylinder flattening objects as if they were under a hydraulic press. The objects being flattened are pink and are being crushed into small pieces. The background is a green wall with a yellow sign.
+The video shows a machine with a yellow and black striped cylinder pressing down on two colorful objects, which are flattened as if they were under a hydraulic press. The machine is positioned on a metal platform, and the background is a green wall.
+A green cube is being compressed by a hydraulic press, which flattens the object as if it were under a hydraulic press. The press is shown in action, with the cube being squeezed into a smaller shape.
+A pink, sparkly ball is being crushed by a large, rusty cylinder, which flattens the objects as if they were under a hydraulic press.
+A red cabbage is being crushed by a hydraulic press, which flattens the objects as if they were under a hydraulic press. The press is shown in action, compressing the cabbage into a smaller, more compact form.
+A lime is being crushed by a hydraulic press, causing it to flatten and burst open, releasing its juice and pulp.
+A large metal press is shown compressing a stack of burgers, causing them to be flattened and crushed into a pile of ground meat.
+A pizza is being crushed by a hydraulic press, causing the toppings to spread out and the crust to crumble.
+A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.
+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.
+A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.
@@ -3,24 +3,19 @@
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
 export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA

-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-# DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing-extended-t2v-1-3b/"
+DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init/"
 VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
 NUM_GPUS=1
-# export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_70k"
+  --output_dir "wan_ode_init_crush_smol"
  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "fixed_wan_ode_init_70k_6e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
+  --wandb_run_name "wan_ode_init_crush_smol"
  --max_train_steps 6000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -30,7 +25,7 @@ training_args=(
  --num_width 832
  --num_frames 77
  --warp_denoising_step
-  # --enable_gradient_checkpointing_type "full"
+  --enable_gradient_checkpointing_type "full"
 )

 # Parallel arguments
@@ -82,7 +77,6 @@ miscellaneous_args=(
  --dit_precision "fp32"
  --num_euler_timesteps 50
  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
 )

 # If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
@@ -1,5 +1,5 @@
 #!/bin/bash
-#SBATCH --job-name=1e5B2_16kFV_warp_ode_vidprom
+#SBATCH --job-name=2e6B8_16kFV_ode_vidprom
 #SBATCH --partition=main
 #SBATCH --nodes=1
 #SBATCH --ntasks=1
@@ -7,14 +7,14 @@
 #SBATCH --gres=gpu:8
 #SBATCH --cpus-per-task=128
 #SBATCH --mem=1440G
-#SBATCH --output=ode_vidprom16k_warp/Dode_vidprom8b16k_1e-5.out
-#SBATCH --error=ode_vidprom16k_warp/Dode_vidprom8b16k_1e-5.err
+#SBATCH --output=ode_vidprom16k/ode_vidprom8b16k_2e-6.out
+#SBATCH --error=ode_vidprom16k/ode_vidprom8b16k_2e-6.err
 #SBATCH --exclusive
 set -e -x

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

 export WANDB_MODE="online"
 export NCCL_P2P_DISABLE=1
@@ -28,18 +28,18 @@ 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 WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+export WANDB_API_KEY=your-wandb-api-key
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA

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


-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing-16k-t2v-1-3b-81/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
+MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+DATA_DIR="your-data-dir"
+VALIDATION_DATASET_FILE="examples/training/consistency_finetune/causal_ode_init/validation.json"
+OUTPUT_DIR="your-output-dir"
+INIT_WEIGHTS_FROM_SAFETENSORS="your-init-weights-from-safetensors" # bidirectional weights from Wan2.1-T2V-1.3B-Diffusers
 NUM_GPUS=8
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
@@ -47,9 +47,9 @@ NUM_GPUS=8
 # Training arguments
 training_args=(
  --tracker_project_name "wan_ode_init"
-  --output_dir "Dwarp_vidprom_8b16k_test_warp_1e-5"
+  --output_dir $OUTPUT_DIR
  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "Dwarp_vidprom_8b16k_wan_ode_init_1e-5"
+  --wandb_run_name "vidprom_8b16k_ode_init_2e-6"
  # --resume_from_checkpoint "ode_init_diffusers/"
  --warp_denoising_step
  --log_visualization
@@ -60,7 +60,7 @@ training_args=(
  --num_latent_t 21
  --num_height 480
  --num_width 832
-  --num_frames 77
+  --num_frames 81
  --dmd_denoising_steps "1000,750,500,250"
  --enable_gradient_checkpointing_type "full"
 )
@@ -93,12 +93,12 @@ validation_args=(
  --validation_steps 50
  --validation_sampling_steps "50"
  --validation_guidance_scale "6.0"
-  # --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
+  --init_weights_from_safetensors $INIT_WEIGHTS_FROM_SAFETENSORS
 )

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
  --mixed_precision "bf16"
  --checkpointing_steps 500
  --weight_decay 1e-4
@@ -3,8 +3,8 @@
 GPU_NUM=1 # 2,4,8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 MODEL_TYPE="wan"
-DATA_MERGE_PATH="data/crush-smol_single/merge.txt"
-OUTPUT_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_single/"
+DATA_MERGE_PATH="$(dirname "$0")/crush_smol_prompts.txt"
+OUTPUT_DIR="data/crush-smol_processed_t2v_1_3b_ode_init/"

 torchrun --nproc_per_node=$GPU_NUM \
    fastvideo/pipelines/preprocess/v1_preprocess.py \
@@ -15,10 +15,11 @@ torchrun --nproc_per_node=$GPU_NUM \
    --max_height 480 \
    --max_width 832 \
    --num_frames 81 \
+    --flow_shift 5.0 \
    --dataloader_num_workers 0 \
    --output_dir=$OUTPUT_DIR \
    --train_fps 16 \
-    --samples_per_file 1 \
-    --flush_frequency 1 \
+    --samples_per_file 8 \
+    --flush_frequency 8 \
    --video_length_tolerance_range 5 \
    --preprocess_task "ode_trajectory" 
@@ -0,0 +1,76 @@
+{
+  "data": [
+    {
+      "caption": "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Elon Musk, dressed in a sleek white spacesuit with a reflective visor, walks confidently across the lunar surface. His posture is upright, and he moves steadily with purpose. The moon's rocky terrain and scattered boulders surround him, casting shadows under the dim sunlight. The background shows vast stretches of the moon's barren landscape with craters and dust clouds kicked up by his boots. The scene captures a wide shot, emphasizing the vastness and desolation of the lunar environment. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "In a dynamic action-packed sequence set in the Marvel multiverse, Spider-Man and Venom engage in an intense battle. Spider-Man, in his classic red and blue suit, swings and dodges venomous attacks from the black symbiote-covered Venom. Both characters display a range of acrobatic moves and powerful strikes. The environment is a chaotic urban landscape with crumbling buildings and neon lights, reflecting the multiversal theme. The camera captures the epic fight from various angles, including wide shots to show the scale of destruction and close-ups to highlight their fierce expressions and physical combat. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A warm, family-oriented scene depicting a father getting ready to leave the house to buy milk. The father, a middle-aged man with a kind face and a casual outfit, picks up a jacket from the coat rack. His posture is upright as he bends down slightly to put on his shoes. In the background, there are glimpses of a cozy living room with a family photograph on the wall. The camera focuses closely on the father, capturing his gentle smile and reassuring nod towards the camera before he opens the front door and steps outside. Static medium close-up shot. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Close-up shot of a man with a prosthetic hand that functions as a rocket launcher. He looks at his new hand with a mix of amazement and concern, his facial expression showing a blend of curiosity and apprehension. The prosthetic hand is sleek and metallic, with intricate details that resemble a high-tech weapon. The background is a dimly lit laboratory with various scientific equipment and monitors displaying data. The man stands in a relaxed posture, his other hand resting on his hip, as he inspects his new limb. The scene is rendered in a realistic sci-fi style, emphasizing the futuristic technology and the man's emotional response to his new appendage. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Realistic CCTV footage style, Kim Taehyung from the band BTS is involved in a drug deal, caught on camera. Kim Taehyung appears nervous and cautious, wearing casual clothing typical of a public space. He exchanges items discreetly with another person, who is partially obscured. Both individuals maintain a watchful demeanor, occasionally glancing around to ensure no one is watching them. The lighting is dim, with flickering fluorescent lights casting shadows on their faces. The background shows a typical urban setting with blurred figures moving in the distance. Static camera angle, medium close-up shot focusing on the interaction between Taehyung and the other individual. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Photorealistic studio setup with professional lighting, showcasing detailed cubic dissections of experimental plastic and felt-like materials on a pristine white background. Each cube reveals intricate layers and textures of the materials, emphasizing their unique properties. The scene has a shallow depth of field initially, then slowly pulls out to reveal the full arrangement of cubes, maintaining a wide depth of field throughout the transition. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    }
+  ]
+}
@@ -1,132 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=ode_crush
-#SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=ode_crush_output/ode_crush.out
-#SBATCH --error=ode_crush_output/ode_crush.err
-#SBATCH --exclusive
-set -e -x
-
-# Environment Setup
-source ~/conda/miniconda/bin/activate
-conda activate will-fv2
-
-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 WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
-
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
-NUM_GPUS=2
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_warp_2"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "2warp_fixed_wan_ode_init_5e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  # --warp_denoising_step
-  --max_train_steps 6001
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim $NUM_GPUS
-  --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 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 20
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 5e-6
-  --mixed_precision "bf16"
-  --checkpointing_steps 2000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --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
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
@@ -1,98 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/weka/home/hao.zhang/wl/FastVideo2/data/crush-smol_processed_t2v_1_3b_ode_init_single"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
-NUM_GPUS=1
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_crush_smol"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "overfitwan_ode_init_crush_smol"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 2001
-  # --warp_denoising_step
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 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 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 20
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 500
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --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
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
@@ -1,40 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}
@@ -1,21 +1,22 @@
 #!/bin/bash
-#SBATCH --job-name=ode_vidprom2k
+#SBATCH --job-name=t2v
 #SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
+#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=ode_vidprom2k_output/ode_vidprom2k.out
-#SBATCH --error=ode_vidprom2k_output/ode_vidprom2k.err
+#SBATCH --output=VSA_t2v_output/t2v_%j.out
+#SBATCH --error=VSA_t2v_output/t2v_%j.err
 #SBATCH --exclusive
 set -e -x

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

+# Basic Info
 export WANDB_MODE="online"
 export NCCL_P2P_DISABLE=1
 export TORCH_NCCL_ENABLE_MONITORING=0
@@ -29,45 +30,41 @@ export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
 export TOKENIZERS_PARALLELISM=false
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA

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

-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
+# Configs
 NUM_GPUS=8
+MODEL_PATH="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_ode_init"
-  --output_dir "wan_ode_init_vidprom2k"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "vidprom2k_wan_ode_init_5e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 6001
+  --tracker_project_name wan_t2v_VSA
+  --output_dir "checkpoints/wan_t2v_finetune_VSA"
+  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
  --gradient_accumulation_steps 1
  --num_latent_t 21
  --num_height 480
  --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
+  --num_frames 81
+  # --enable_gradient_checkpointing_type "full" # if OOM enable this
 )

 # Parallel arguments
 parallel_args=(
-  --num_gpus $NUM_GPUS
+  --num_gpus 64
  --sp_size 1
  --tp_size 1
-  --hsdp_replicate_dim 8
+  --hsdp_replicate_dim 64
  --hsdp_shard_dim 1
 )

@@ -79,25 +76,25 @@ model_args=(

 # Dataset arguments
 dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
 )

 # Validation arguments
 validation_args=(
  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 100
+  --validation_dataset_file $VALIDATION_DATASET_FILE
+  --validation_steps 200
  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
+  --validation_guidance_scale "5.0"
 )

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 5e-6
+  --learning_rate 1e-5
  --mixed_precision "bf16"
-  --checkpointing_steps 2000
-  --weight_decay 1e-4
+  --checkpointing_steps 1000
+  --weight_decay 0.01
  --max_grad_norm 1.0
 )

@@ -106,26 +103,31 @@ miscellaneous_args=(
  --inference_mode False
  --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
-  --enable_gradient_checkpointing_type "full"
+  --flow_shift 1
+  --seed 1000
+)
+
+# VSA arguments
+vsa_args=(
+  --VSA_decay_rate 0.03 \
+  --VSA_decay_interval_steps 50 \
+  --VSA_sparsity 0.9 \
 )

-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
 srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_training_pipeline.py \
    "${parallel_args[@]}" \
    "${model_args[@]}" \
    "${dataset_args[@]}" \
    "${training_args[@]}" \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
+    "${miscellaneous_args[@]}" \
+    "${vsa_args[@]}"  
@@ -6,8 +6,8 @@ export TOKENIZERS_PARALLELISM=false
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA

 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_old"
-VALIDATION_DATASET_FILE="examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
+DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
 NUM_GPUS=4
 # export CUDA_VISIBLE_DEVICES=4,5

@@ -52,7 +52,7 @@ dataset_args=(
 validation_args=(
  --log_validation 
  --validation_dataset_file $VALIDATION_DATASET_FILE
-  --validation_steps 50
+  --validation_steps 200
  --validation_sampling_steps "50" 
  --validation_guidance_scale "6.0"
 )
@@ -4,7 +4,7 @@ GPU_NUM=1 # 2,4,8
 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_old/"
+OUTPUT_DIR="data/crush-smol_processed_t2v/"

 torchrun --nproc_per_node=$GPU_NUM \
    fastvideo/pipelines/preprocess/v1_preprocess.py \
@@ -1,6 +1,6 @@
 #!/bin/bash

-GPU_NUM=1 # 2,4,8
+GPU_NUM=2 # 2,4,8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATASET_PATH="data/crush-smol/"
 OUTPUT_DIR="data/crush-smol_processed_t2v/"
@@ -14,7 +14,7 @@ torchrun --nproc_per_node=$GPU_NUM \
    --preprocess.dataset_type merged \
    --preprocess.dataset_path $DATASET_PATH \
    --preprocess.dataset_output_dir $OUTPUT_DIR \
-    --preprocess.preprocess_video_batch_size 8 \
+    --preprocess.preprocess_video_batch_size 2 \
    --preprocess.dataloader_num_workers 0 \
    --preprocess.max_height 480 \
    --preprocess.max_width 832 \
@@ -28,4 +28,4 @@
      "num_frames": 77
    }
  ]
-}
+}
@@ -1,94 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_old"
-VALIDATION_DATASET_FILE="examples/datasets/crush_smol/validation.json"
-NUM_GPUS=8
-# export CUDA_VISIBLE_DEVICES=4,5
-
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_t2v_i2v_finetune"
-  --output_dir "checkpoints/wan_t2v_i2v_finetune"
-  --max_train_steps 5000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 2
-  --num_latent_t 20
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS 
-  --sp_size 4
-  --tp_size 1
-  --hsdp_replicate_dim 2
-  --hsdp_shard_dim 4
-)
-
-# 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 "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 5e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 1000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --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
-  --enable_gradient_checkpointing_type "full"
-  --t2v_as_i2v_task True
-  # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500"
-)
-
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/wan_t2v_i2v_training_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
@@ -1,24 +0,0 @@
-#!/bin/bash
-
-GPU_NUM=1 # 2,4,8
-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_i2v_1_3b/"
-
-torchrun --nproc_per_node=$GPU_NUM \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
-    --model_path $MODEL_PATH \
-    --data_merge_path $DATA_MERGE_PATH \
-    --preprocess_video_batch_size 2 \
-    --seed 42 \
-    --max_height 480 \
-    --max_width 832 \
-    --num_frames 77 \
-    --dataloader_num_workers 0 \
-    --output_dir=$OUTPUT_DIR \
-    --train_fps 16 \
-    --samples_per_file 8 \
-    --flush_frequency 8 \
-    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v_ode_trajectory" 
@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+from fastvideo.attention.backends.sageattn.api import sageattn_blackwell
+
+from fastvideo.attention.backends.abstract import (AttentionBackend,
+                                                   AttentionImpl,
+                                                   AttentionMetadata,
+                                                   AttentionMetadataBuilder)
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class SageAttention3Backend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [64, 128, 256]
+
+    @staticmethod
+    def get_name() -> str:
+        return "SAGE_ATTN_THREE"
+
+    @staticmethod
+    def get_impl_cls() -> type["SageAttention3Impl"]:
+        return SageAttention3Impl
+
+    @staticmethod
+    def get_metadata_cls() -> type["AttentionMetadata"]:
+        raise NotImplementedError
+
+    @staticmethod
+    def get_builder_cls() -> type["AttentionMetadataBuilder"]:
+        raise NotImplementedError
+
+    # @staticmethod
+    # def get_metadata_cls() -> Type["AttentionMetadata"]:
+    #     return FlashAttentionMetadata
+
+
+class SageAttention3Impl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout = extra_impl_args.get("dropout_p", 0.0)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+        output = sageattn_blackwell(query, key, value, is_causal=self.causal)
+        output = output.transpose(1, 2)
+        return output
@@ -5,7 +5,6 @@ from dataclasses import dataclass

 import torch
 from einops import rearrange
-from flash_attn.bert_padding import pad_input

 from csrc.attn.vmoba_attn.vmoba import (moba_attn_varlen, process_moba_input,
                                        process_moba_output)
@@ -134,6 +133,8 @@ class VMOBAAttentionImpl(AttentionImpl):
                 **extra_impl_args) -> None:
        self.prefix = prefix
        self.layer_idx = self._get_layer_idx(prefix)
+        from flash_attn.bert_padding import pad_input
+        self.pad_input = pad_input

    def _get_layer_idx(self, prefix: str) -> int | None:
        match = re.search(r"blocks\.(\d+)", prefix)
@@ -169,7 +170,6 @@ class VMOBAAttentionImpl(AttentionImpl):
            moba_chunk_size = attn_metadata.st_chunk_size
            moba_topk = attn_metadata.st_topk

-        # torch.distributed.breakpoint()
        query, chunk_size = process_moba_input(query,
                                               attn_metadata.patch_resolution,
                                               moba_chunk_size)
@@ -205,8 +205,8 @@ class VMOBAAttentionImpl(AttentionImpl):
            simsum_threshold=attn_metadata.moba_threshold,
            threshold_type=attn_metadata.moba_threshold_type,
        )
-        hidden_states = pad_input(hidden_states, indices_q, batch_size,
-                                  sequence_length)
+        hidden_states = self.pad_input(hidden_states, indices_q, batch_size,
+                                       sequence_length)
        hidden_states = process_moba_output(hidden_states,
                                            attn_metadata.patch_resolution,
                                            moba_chunk_size)
@@ -15,18 +15,16 @@ class DiTArchConfig(ArchConfig):
    reverse_param_names_mapping: dict = field(default_factory=dict)
    lora_param_names_mapping: dict = field(default_factory=dict)
    _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
-        AttentionBackendEnum.SLIDING_TILE_ATTN,
-        AttentionBackendEnum.SAGE_ATTN,
-        AttentionBackendEnum.FLASH_ATTN,
-        AttentionBackendEnum.TORCH_SDPA,
-        AttentionBackendEnum.VIDEO_SPARSE_ATTN,
-        AttentionBackendEnum.VMOBA_ATTN,
-    )
+        AttentionBackendEnum.SLIDING_TILE_ATTN, AttentionBackendEnum.SAGE_ATTN,
+        AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA,
+        AttentionBackendEnum.VIDEO_SPARSE_ATTN, AttentionBackendEnum.VMOBA_ATTN,
+        AttentionBackendEnum.SAGE_ATTN_THREE)

    hidden_size: int = 0
    num_attention_heads: int = 0
    num_channels_latents: int = 0
    exclude_lora_layers: list[str] = field(default_factory=list)
+    boundary_ratio: float | None = None

    def __post_init__(self) -> None:
        if not self._compile_conditions:
@@ -4,12 +4,13 @@ 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,
+from fastvideo.configs.pipelines.wan import (SelfForcingWanT2V480PConfig,
+                                             WanI2V480PConfig, WanI2V720PConfig,
                                             WanT2V480PConfig, WanT2V720PConfig)

 __all__ = [
    "HunyuanConfig", "FastHunyuanConfig", "PipelineConfig",
    "SlidingTileAttnConfig", "WanT2V480PConfig", "WanI2V480PConfig",
    "WanT2V720PConfig", "WanI2V720PConfig", "StepVideoT2VConfig",
-    "get_pipeline_config_cls_from_name"
+    "SelfForcingWanT2V480PConfig", "get_pipeline_config_cls_from_name"
 ]
@@ -45,8 +45,6 @@ class PipelineConfig:
    embedded_cfg_scale: float = 6.0
    flow_shift: float | None = None
    disable_autocast: bool = False
-    ti2v_task: bool = False
-    t2v_as_i2v_task: bool = False

    # Model configuration
    dit_config: DiTConfig = field(default_factory=DiTConfig)
@@ -87,6 +85,10 @@ class PipelineConfig:
    # DMD parameters
    dmd_denoising_steps: list[int] | None = field(default=None)

+    # Wan2.2 TI2V parameters
+    ti2v_task: bool = False
+    boundary_ratio: float | None = None
+
    # Compilation
    # enable_torch_compile: bool = False

@@ -213,24 +215,6 @@ class PipelineConfig:
            "Comma-separated list of denoising steps (e.g., '1000,757,522')",
        )

-        # TI2V task
-        parser.add_argument(
-            f"--{prefix_with_dot}ti2v-task",
-            action=StoreBoolean,
-            dest=f"{prefix_with_dot.replace('-', '_')}ti2v_task",
-            default=PipelineConfig.ti2v_task,
-            help="Enable TI2V",
-        )
-
-        # T2V to I2V task
-        parser.add_argument(
-            f"--{prefix_with_dot}t2v-as-i2v-task",
-            action=StoreBoolean,
-            dest=f"{prefix_with_dot.replace('-', '_')}t2v_as_i2v_task",
-            default=PipelineConfig.t2v_as_i2v_task,
-            help="Enable T2V to I2V task",
-        )
-
        # Add VAE configuration arguments
        from fastvideo.configs.models.vaes.base import VAEConfig
        VAEConfig.add_cli_args(parser, prefix=f"{prefix_with_dot}vae-config")
@@ -262,9 +246,7 @@ class PipelineConfig:
        """
        from fastvideo.configs.pipelines.registry import (
            get_pipeline_config_cls_from_name)
-        logger.info("WTF model_path: %s", model_path)
        pipeline_config_cls = get_pipeline_config_cls_from_name(model_path)
-        logger.info("pipeline_config_cls: %s", pipeline_config_cls)

        return cast(PipelineConfig, pipeline_config_cls(model_path=model_path))

@@ -11,9 +11,9 @@ from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
 # isort: off
 from fastvideo.configs.pipelines.wan import (
    FastWan2_1_T2V_480P_Config, FastWan2_2_TI2V_5B_Config,
-    SelfForcingWanT2V480PConfig, Wan2_2_I2V_A14B_Config, Wan2_2_T2V_A14B_Config,
-    Wan2_2_TI2V_5B_Config, WanI2V480PConfig, WanI2V720PConfig, WanT2V480PConfig,
-    WanT2V720PConfig, SelfForcingWanT2V480PConfig)
+    Wan2_2_I2V_A14B_Config, Wan2_2_T2V_A14B_Config, Wan2_2_TI2V_5B_Config,
+    WanI2V480PConfig, WanI2V720PConfig, WanT2V480PConfig, WanT2V720PConfig,
+    SelfForcingWanT2V480PConfig)
 # isort: on
 from fastvideo.logger import init_logger
 from fastvideo.utils import (maybe_download_model_index,
@@ -48,8 +48,8 @@ PIPELINE_DETECTOR: dict[str, Callable[[str], bool]] = {
    "wanpipeline": lambda id: "wanpipeline" in id.lower(),
    "wanimagetovideo": lambda id: "wanimagetovideo" in id.lower(),
    "wandmdpipeline": lambda id: "wandmdpipeline" in id.lower(),
-    "stepvideo": lambda id: "stepvideo" in id.lower(),
    "wancausaldmdpipeline": lambda id: "wancausaldmdpipeline" in id.lower(),
+    "stepvideo": lambda id: "stepvideo" in id.lower(),
    # Add other pipeline architecture detectors
 }

@@ -61,8 +61,8 @@ PIPELINE_FALLBACK_CONFIG: dict[str, type[PipelineConfig]] = {
    WanT2V480PConfig,  # Base Wan config as fallback for any Wan variant
    "wanimagetovideo": WanI2V480PConfig,
    "wandmdpipeline": FastWan2_1_T2V_480P_Config,
-    "stepvideo": StepVideoT2VConfig,
    "wancausaldmdpipeline": SelfForcingWanT2V480PConfig,
+    "stepvideo": StepVideoT2VConfig
    # Other fallbacks by architecture
 }

@@ -37,7 +37,7 @@ def getdataset(args) -> VideoCaptionMergedDataset:
                                     temporal_sample=temporal_sample,
                                     transform_topcrop=transform_topcrop,
                                     seed=args.seed)
-
+                                    

 def gettextdataset(args) -> TextDataset:
    return TextDataset(data_merge_path=args.data_merge_path,
@@ -1,43 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from fastvideo.dataset.lmdb_utils import get_array_shape_from_lmdb, retrieve_row_from_lmdb
-from torch.utils.data import Dataset
-import numpy as np
-import torch
-import lmdb
-
-# from Self-Forcing: https://github.com/guandeh17/Self-Forcing/blob/main/utils/dataset.py
-class ODERegressionLMDBDataset(Dataset):
-    def __init__(self, data_path: str, max_pair: int = int(1e8)):
-        print(f"data_path: {data_path}")
-        self.env = lmdb.open(data_path, readonly=True,
-                             lock=False, readahead=False, meminit=False)
-
-        self.latents_shape = get_array_shape_from_lmdb(self.env, 'latents')
-        self.max_pair = max_pair
-
-    def __len__(self):
-        return min(self.latents_shape[0], self.max_pair)
-
-    def __getitem__(self, idx):
-        """
-        Outputs:
-            - prompts: List of Strings
-            - latents: Tensor of shape (num_denoising_steps, num_frames, num_channels, height, width). It is ordered from pure noise to clean image.
-        """
-        latents = retrieve_row_from_lmdb(
-            self.env,
-            "latents", np.float16, idx, shape=self.latents_shape[1:]
-        )
-
-        if len(latents.shape) == 4:
-            latents = latents[None, ...]
-
-        prompts = retrieve_row_from_lmdb(
-            self.env,
-            "prompts", str, idx
-        )
-        return {
-            "prompts": prompts,
-            "ode_latent": torch.tensor(latents, dtype=torch.float32)
-        }
-
@@ -1,75 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# from Self-Forcing: https://github.com/guandeh17/Self-Forcing/blob/main/utils/lmdb.py
-
-import numpy as np
-
-
-def get_array_shape_from_lmdb(env, array_name):
-    with env.begin() as txn:
-        image_shape = txn.get(f"{array_name}_shape".encode()).decode()
-        image_shape = tuple(map(int, image_shape.split()))
-    return image_shape
-
-
-def store_arrays_to_lmdb(env, arrays_dict, start_index=0):
-    """
-    Store rows of multiple numpy arrays in a single LMDB.
-    Each row is stored separately with a naming convention.
-    """
-    with env.begin(write=True) as txn:
-        for array_name, array in arrays_dict.items():
-            for i, row in enumerate(array):
-                # Convert row to bytes
-                if isinstance(row, str):
-                    row_bytes = row.encode()
-                else:
-                    row_bytes = row.tobytes()
-
-                data_key = f'{array_name}_{start_index + i}_data'.encode()
-
-                txn.put(data_key, row_bytes)
-
-
-def process_data_dict(data_dict, seen_prompts):
-    output_dict = {}
-
-    all_videos = []
-    all_prompts = []
-    for prompt, video in data_dict.items():
-        if prompt in seen_prompts:
-            continue
-        else:
-            seen_prompts.add(prompt)
-
-        video = video.half().numpy()
-        all_videos.append(video)
-        all_prompts.append(prompt)
-
-    if len(all_videos) == 0:
-        return {"latents": np.array([]), "prompts": np.array([])}
-
-    all_videos = np.concatenate(all_videos, axis=0)
-
-    output_dict['latents'] = all_videos
-    output_dict['prompts'] = np.array(all_prompts)
-
-    return output_dict
-
-
-def retrieve_row_from_lmdb(lmdb_env, array_name, dtype, row_index, shape=None):
-    """
-    Retrieve a specific row from a specific array in the LMDB.
-    """
-    data_key = f'{array_name}_{row_index}_data'.encode()
-
-    with lmdb_env.begin() as txn:
-        row_bytes = txn.get(data_key)
-
-    if dtype == str:
-        array = row_bytes.decode()
-    else:
-        array = np.frombuffer(row_bytes, dtype=dtype)
-
-    if shape is not None and len(shape) > 0:
-        array = array.reshape(shape)
-    return array
@@ -3,9 +3,6 @@ from typing import Any, cast

 import numpy as np
 import torch
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)


 def pad(t: torch.Tensor, padding_length: int) -> tuple[torch.Tensor, torch.Tensor]:
@@ -175,6 +175,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
    # - "SLIDING_TILE_ATTN" : use Sliding Tile Attention
    # - "VIDEO_SPARSE_ATTN": use Video Sparse Attention
    # - "SAGE_ATTN": use Sage Attention
+    # - "SAGE_ATTN_THREE": use Sage Attention 3
    "FASTVIDEO_ATTENTION_BACKEND":
    lambda: os.getenv("FASTVIDEO_ATTENTION_BACKEND", None),

@@ -705,7 +705,6 @@ class TrainingArgs(FastVideoArgs):
    # simulate generator forward to match inference
    simulate_generator_forward: bool = False
    warp_denoising_step: bool = False
-    intermediate_latents_visualization: bool = False

    # Self-forcing specific arguments
    num_frame_per_block: int = 3
@@ -212,9 +212,9 @@ class ScaleResidualLayerNormScaleShift(nn.Module):
            frame_seqlen = normalized.shape[1] // num_frames
            modulated = (
                normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
-                (1 + scale) + shift).flatten(1, 2)
+                (1.0 + scale) + shift).flatten(1, 2)
        else:
-            modulated = normalized * (1 + scale) + shift
+            modulated = normalized * (1.0 + scale) + shift
        return modulated, residual_output


@@ -267,11 +267,11 @@ class LayerNormScaleShift(nn.Module):
            frame_seqlen = normalized.shape[1] // num_frames
            output = (
                normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
-                (1 + scale) + shift).flatten(1, 2)
+                (1.0 + scale) + shift).flatten(1, 2)
        else:
            # scale.shape: [batch_size, 1, inner_dim]
            # shift.shape: [batch_size, 1, inner_dim]
-            output = normalized * (1 + scale) + shift
+            output = normalized * (1.0 + scale) + shift

        if self.compute_dtype == torch.float32:
            output = output.to(x.dtype)
@@ -147,8 +147,8 @@ class CausalWanSelfAttention(nn.Module):
                # Assign new keys/values directly up to current_end
                local_end_index = kv_cache["local_end_index"].item() + current_end - kv_cache["global_end_index"].item()
                local_start_index = local_end_index - num_new_tokens
-                # kv_cache["k"] = kv_cache["k"].detach()
-                # kv_cache["v"] = kv_cache["v"].detach()
+                kv_cache["k"] = kv_cache["k"].detach()
+                kv_cache["v"] = kv_cache["v"].detach()
                # logger.info("kv_cache['k'] is in comp graph: %s", kv_cache["k"].requires_grad or kv_cache["k"].grad_fn is not None)
                kv_cache["k"][:, local_start_index:local_end_index] = roped_key
                kv_cache["v"][:, local_start_index:local_end_index] = v
@@ -179,7 +179,7 @@ class CausalWanTransformerBlock(nn.Module):
        super().__init__()

        # 1. Self-attention
-        self.norm1 = nn.LayerNorm(dim, eps, elementwise_affine=False)
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
        self.to_q = ReplicatedLinear(dim, dim, bias=True)
        self.to_k = ReplicatedLinear(dim, dim, bias=True)
        self.to_v = ReplicatedLinear(dim, dim, bias=True)
@@ -212,7 +212,8 @@ class CausalWanTransformerBlock(nn.Module):
            norm_type="layer",
            eps=eps,
            elementwise_affine=True,
-            dtype=torch.float32)
+            dtype=torch.float32,
+            compute_dtype=torch.float32)

        # 2. Cross-attention
        # Only T2V for now
@@ -225,7 +226,8 @@ class CausalWanTransformerBlock(nn.Module):
            norm_type="layer",
            eps=eps,
            elementwise_affine=False,
-            dtype=torch.float32)
+            dtype=torch.float32,
+            compute_dtype=torch.float32)

        # 3. Feed-forward
        self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
@@ -250,34 +252,29 @@ class CausalWanTransformerBlock(nn.Module):
        if hidden_states.dim() == 4:
            hidden_states = hidden_states.squeeze(1)
        num_frames = temb.shape[1]
-        frame_seqlen = hidden_states.shape[1] // num_frames    
+        frame_seqlen = hidden_states.shape[1] // num_frames
        bs, seq_length, _ = hidden_states.shape
        orig_dtype = hidden_states.dtype
        # assert orig_dtype != torch.float32
-        e = self.scale_shift_table + temb
+        e = self.scale_shift_table + temb.float()
        # e.shape: [batch_size, num_frames, 6, inner_dim]
        assert e.shape == (bs, num_frames, 6, self.hidden_dim)
        shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(
            6, dim=2)
        # *_msa.shape: [batch_size, num_frames, 1, inner_dim]
-        # assert shift_msa.dtype == torch.float32
-
-        # logger.info("temb sum: %s, dtype: %s", temb.float().sum().item(), temb.dtype)
-        # logger.info("scale_msa sum: %s, dtype: %s", scale_msa.float().sum().item(), scale_msa.dtype)
-        # logger.info("shift_msa sum: %s, dtype: %s", shift_msa.float().sum().item(), shift_msa.dtype)
+        assert shift_msa.dtype == torch.float32

        # 1. Self-attention
-        norm_hidden_states = (self.norm1(hidden_states).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
-                        (1 + scale_msa) + shift_msa).flatten(1, 2)
-        # logger.info("norm_hidden_states sum: %s, shape: %s", norm_hidden_states.float().sum().item(), norm_hidden_states.shape)
+        norm_hidden_states = (self.norm1(hidden_states.float()).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+                        (1 + scale_msa) + shift_msa).flatten(1, 2).to(orig_dtype)
        query, _ = self.to_q(norm_hidden_states)
        key, _ = self.to_k(norm_hidden_states)
        value, _ = self.to_v(norm_hidden_states)

        if self.norm_q is not None:
-            query = self.norm_q.forward_native(query)
+            query = self.norm_q(query)
        if self.norm_k is not None:
-            key = self.norm_k.forward_native(key)
+            key = self.norm_k(key)

        query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
        key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -291,6 +288,8 @@ class CausalWanTransformerBlock(nn.Module):
        null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
        norm_hidden_states, hidden_states = self.self_attn_residual_norm(
            hidden_states, attn_output, gate_msa, null_shift, null_scale)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)

        # 2. Cross-attention
        attn_output = self.attn2(norm_hidden_states,
@@ -299,10 +298,13 @@ class CausalWanTransformerBlock(nn.Module):
                                 crossattn_cache=crossattn_cache)
        norm_hidden_states, hidden_states = self.cross_attn_residual_norm(
            hidden_states, attn_output, 1, c_shift_msa, c_scale_msa)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)

        # 3. Feed-forward
        ff_output = self.ffn(norm_hidden_states)
        hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
+        hidden_states = hidden_states.to(orig_dtype)

        return hidden_states

@@ -365,7 +367,8 @@ class CausalWanTransformer3DModel(BaseDiT):
                                            norm_type="layer",
                                            eps=config.eps,
                                            elementwise_affine=False,
-                                            dtype=torch.float32)
+                                            dtype=torch.float32,
+                                            compute_dtype=torch.float32)
        self.proj_out = nn.Linear(
            inner_dim, config.out_channels * math.prod(config.patch_size))
        self.scale_shift_table = nn.Parameter(
@@ -375,7 +378,8 @@ class CausalWanTransformer3DModel(BaseDiT):

        # Causal-specific
        self.block_mask = None
-        self.num_frame_per_block = 3
+        self.num_frame_per_block = config.arch_config.num_frames_per_block
+        assert self.num_frame_per_block <= 3
        self.independent_first_frame = False

        self.__post_init__()
@@ -487,16 +491,12 @@ class CausalWanTransformer3DModel(BaseDiT):
        )
        freqs_cos = freqs_cos.to(hidden_states.device)
        freqs_sin = freqs_sin.to(hidden_states.device)
-        freqs_cis = (freqs_cos,
-                     freqs_sin) if freqs_cos is not None else None
+        freqs_cis = (freqs_cos.float(),
+                     freqs_sin.float()) if freqs_cos is not None else None

        hidden_states = self.patch_embedding(hidden_states)
-        grid_sizes = torch.stack(
-            [torch.tensor(hidden_states[0].shape[1:], dtype=torch.long)])
        hidden_states = hidden_states.flatten(2).transpose(1, 2)

-        encoder_hidden_states = torch.cat([encoder_hidden_states, encoder_hidden_states.new_zeros(1, self.text_len - encoder_hidden_states.size(1), encoder_hidden_states.size(2))], dim=1)
-
        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
                        timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
        timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=timestep.shape)
@@ -543,9 +543,14 @@ class CausalWanTransformer3DModel(BaseDiT):
        hidden_states = self.norm_out(hidden_states, shift, scale)
        hidden_states = self.proj_out(hidden_states)

-        output = self.unpatchify(hidden_states, grid_sizes)
+        hidden_states = hidden_states.reshape(batch_size, post_patch_num_frames,
+                                              post_patch_height,
+                                              post_patch_width, p_t, p_h, p_w,
+                                              -1)
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

-        return torch.stack(output)
+        return output

    def _forward_train(self,
                hidden_states: torch.Tensor,
@@ -586,8 +591,8 @@ class CausalWanTransformer3DModel(BaseDiT):
        )
        freqs_cos = freqs_cos.to(hidden_states.device)
        freqs_sin = freqs_sin.to(hidden_states.device)
-        freqs_cis = (freqs_cos,
-                     freqs_sin) if freqs_cos is not None else None
+        freqs_cis = (freqs_cos.float(),
+                     freqs_sin.float()) if freqs_cos is not None else None

        # Construct blockwise causal attn mask
        if self.block_mask is None:
@@ -600,12 +605,8 @@ class CausalWanTransformer3DModel(BaseDiT):
            )

        hidden_states = self.patch_embedding(hidden_states)
-        grid_sizes = torch.stack(
-            [torch.tensor(hidden_states[0].shape[1:], dtype=torch.long)])
        hidden_states = hidden_states.flatten(2).transpose(1, 2)

-        encoder_hidden_states = torch.cat([encoder_hidden_states, encoder_hidden_states.new_zeros(1, self.text_len - encoder_hidden_states.size(1), encoder_hidden_states.size(2))], dim=1)
-
        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
                        timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
        timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=timestep.shape)
@@ -640,9 +641,14 @@ class CausalWanTransformer3DModel(BaseDiT):
        hidden_states = self.norm_out(hidden_states, shift, scale)
        hidden_states = self.proj_out(hidden_states)

-        output = self.unpatchify(hidden_states, grid_sizes)
+        hidden_states = hidden_states.reshape(batch_size, post_patch_num_frames,
+                                              post_patch_height,
+                                              post_patch_width, p_t, p_h, p_w,
+                                              -1)
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

-        return torch.stack(output)
+        return output

    def forward(
        self,
@@ -653,30 +659,3 @@ class CausalWanTransformer3DModel(BaseDiT):
            return self._forward_inference(*args, **kwargs)
        else:
            return self._forward_train(*args, **kwargs)
-
-
-    def unpatchify(self, x, grid_sizes):
-        r"""
-
-
-        Args:
-            x (List[Tensor]):
-                List of patchified features, each with shape [L, C_out * prod(patch_size)]
-            grid_sizes (Tensor):
-                Original spatial-temporal grid dimensions before patching,
-
-
-        Returns:
-            Tensor:
-                Reconstructed video tensors with shape [B, C_out, F, H / 8, W / 8]
-        """
-
-        c = self.out_channels
-        out = []
-        for u, v in zip(x, grid_sizes.tolist()):
-            u = u[:math.prod(v)].view(*v, *self.patch_size, c)
-            u = u.permute(6, 0, 3, 1, 4, 2, 5)
-            # u = torch.einsum('fhwpqrc->cfphqwr', u.contiguous())
-            u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
-            out.append(u)
-        return out
@@ -1,5 +1,3 @@
-import torch
-import torch.nn as nn
 # SPDX-License-Identifier: Apache-2.0

 import math
@@ -39,14 +37,16 @@ class WanImageEmbedding(torch.nn.Module):
    def __init__(self, in_features: int, out_features: int):
        super().__init__()

-        self.norm1 = nn.LayerNorm(in_features)
+        self.norm1 = FP32LayerNorm(in_features)
        self.ff = MLP(in_features, in_features, out_features, act_type="gelu")
-        self.norm2 = nn.LayerNorm(out_features)
+        self.norm2 = FP32LayerNorm(out_features)

-    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+    def forward(self,
+                encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+        dtype = encoder_hidden_states_image.dtype
        hidden_states = self.norm1(encoder_hidden_states_image)
        hidden_states = self.ff(hidden_states)
-        hidden_states = self.norm2(hidden_states)
+        hidden_states = self.norm2(hidden_states).to(dtype)
        return hidden_states


@@ -62,7 +62,7 @@ class WanTimeTextImageEmbedding(nn.Module):
        super().__init__()

        self.time_embedder = TimestepEmbedder(
-            dim, frequency_embedding_size=time_freq_dim, act_layer="silu", freq_dtype=torch.float64)
+            dim, frequency_embedding_size=time_freq_dim, act_layer="silu")
        self.time_modulation = ModulateProjection(dim,
                                                  factor=6,
                                                  act_layer="silu")
@@ -156,12 +156,12 @@ class WanT2VCrossAttention(WanSelfAttention):
        b, n, d = x.size(0), self.num_heads, self.head_dim

        # compute query, key, value
-        q = self.norm_q.forward_native(self.to_q(x)[0]).view(b, -1, n, d)
+        q = self.norm_q(self.to_q(x)[0]).view(b, -1, n, d)

        if crossattn_cache is not None:
            if not crossattn_cache["is_init"]:
                crossattn_cache["is_init"] = True
-                k = self.norm_k.forward_native(self.to_k(context)[0]).view(b, -1, n, d)
+                k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
                v = self.to_v(context)[0].view(b, -1, n, d)
                crossattn_cache["k"] = k
                crossattn_cache["v"] = v
@@ -169,7 +169,7 @@ class WanT2VCrossAttention(WanSelfAttention):
                k = crossattn_cache["k"]
                v = crossattn_cache["v"]
        else:
-            k = self.norm_k.forward_native(self.to_k(context)[0]).view(b, -1, n, d)
+            k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
            v = self.to_v(context)[0].view(b, -1, n, d)

        # compute attention
@@ -213,10 +213,10 @@ class WanI2VCrossAttention(WanSelfAttention):
        b, n, d = x.size(0), self.num_heads, self.head_dim

        # compute query, key, value
-        q = self.norm_q.forward_native(self.to_q(x)[0]).view(b, -1, n, d)
-        k = self.norm_k.forward_native(self.to_k(context)[0]).view(b, -1, n, d)
+        q = self.norm_q(self.to_q(x)[0]).view(b, -1, n, d)
+        k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
        v = self.to_v(context)[0].view(b, -1, n, d)
-        k_img = self.norm_added_k.forward_native(self.add_k_proj(context_img)[0]).view(
+        k_img = self.norm_added_k(self.add_k_proj(context_img)[0]).view(
            b, -1, n, d)
        v_img = self.add_v_proj(context_img)[0].view(b, -1, n, d)
        img_x = self.attn(q, k_img, v_img)
@@ -247,7 +247,7 @@ class WanTransformerBlock(nn.Module):
        super().__init__()

        # 1. Self-attention
-        self.norm1 = nn.LayerNorm(dim, eps, elementwise_affine=False)
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
        self.to_q = ReplicatedLinear(dim, dim, bias=True)
        self.to_k = ReplicatedLinear(dim, dim, bias=True)
        self.to_v = ReplicatedLinear(dim, dim, bias=True)
@@ -278,29 +278,29 @@ class WanTransformerBlock(nn.Module):
            norm_type="layer",
            eps=eps,
            elementwise_affine=True,
-            dtype=torch.float32)
+            dtype=torch.float32,
+            compute_dtype=torch.float32)

        # 2. Cross-attention
        if added_kv_proj_dim is not None:
            # I2V
-            self.attn2 = WanI2VCrossAttention(dim, 
-                                              num_heads, 
-                                              qk_norm=qk_norm, 
+            self.attn2 = WanI2VCrossAttention(dim,
+                                              num_heads,
+                                              qk_norm=qk_norm,
                                              eps=eps)
-
        else:
            # T2V
-            self.attn2 = WanT2VCrossAttention(dim, 
-                                              num_heads, 
-                                              qk_norm=qk_norm, 
+            self.attn2 = WanT2VCrossAttention(dim,
+                                              num_heads,
+                                              qk_norm=qk_norm,
                                              eps=eps)
-
        self.cross_attn_residual_norm = ScaleResidualLayerNormScaleShift(
-            dim, 
-            norm_type="layer", 
-            eps=eps, 
-            elementwise_affine=False, 
-            dtype=torch.float32)
+            dim,
+            norm_type="layer",
+            eps=eps,
+            elementwise_affine=False,
+            dtype=torch.float32,
+            compute_dtype=torch.float32)

        # 3. Feed-forward
        self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
@@ -319,11 +319,12 @@ class WanTransformerBlock(nn.Module):
            hidden_states = hidden_states.squeeze(1)
        bs, seq_length, _ = hidden_states.shape
        orig_dtype = hidden_states.dtype
+        # assert orig_dtype != torch.float32

        if temb.dim() == 4:
            # temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v)
            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
-                self.scale_shift_table.unsqueeze(0) + temb
+                self.scale_shift_table.unsqueeze(0) + temb.float()
            ).chunk(6, dim=2)
            # batch_size, seq_len, 1, inner_dim
            shift_msa = shift_msa.squeeze(2)
@@ -334,20 +335,22 @@ class WanTransformerBlock(nn.Module):
            c_gate_msa = c_gate_msa.squeeze(2)
        else:
            # temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B)
-            e = self.scale_shift_table + temb
+            e = self.scale_shift_table + temb.float()
            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(
                6, dim=1)
+        assert shift_msa.dtype == torch.float32

        # 1. Self-attention
-        norm_hidden_states = self.norm1(hidden_states) * (1 + scale_msa) + shift_msa
+        norm_hidden_states = (self.norm1(hidden_states.float()) *
+                              (1 + scale_msa) + shift_msa).to(orig_dtype)
        query, _ = self.to_q(norm_hidden_states)
        key, _ = self.to_k(norm_hidden_states)
        value, _ = self.to_v(norm_hidden_states)

        if self.norm_q is not None:
-            query = self.norm_q.forward_native(query)
+            query = self.norm_q(query)
        if self.norm_k is not None:
-            key = self.norm_k.forward_native(key)
+            key = self.norm_k(key)

        query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
        key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -367,20 +370,26 @@ class WanTransformerBlock(nn.Module):
        null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
        norm_hidden_states, hidden_states = self.self_attn_residual_norm(
            hidden_states, attn_output, gate_msa, null_shift, null_scale)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)

        # 2. Cross-attention
-        attn_output = self.attn2(norm_hidden_states, 
-                                 context=encoder_hidden_states, 
+        attn_output = self.attn2(norm_hidden_states,
+                                 context=encoder_hidden_states,
                                 context_lens=None)
        norm_hidden_states, hidden_states = self.cross_attn_residual_norm(
            hidden_states, attn_output, 1, c_shift_msa, c_scale_msa)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)

        # 3. Feed-forward
        ff_output = self.ffn(norm_hidden_states)
        hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
+        hidden_states = hidden_states.to(orig_dtype)

        return hidden_states

+
 class WanTransformerBlock_VSA(nn.Module):

    def __init__(self,
@@ -397,7 +406,7 @@ class WanTransformerBlock_VSA(nn.Module):
        super().__init__()

        # 1. Self-attention
-        self.norm1 = nn.LayerNorm(dim, eps, elementwise_affine=False)
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
        self.to_q = ReplicatedLinear(dim, dim, bias=True)
        self.to_k = ReplicatedLinear(dim, dim, bias=True)
        self.to_v = ReplicatedLinear(dim, dim, bias=True)
@@ -429,7 +438,8 @@ class WanTransformerBlock_VSA(nn.Module):
            norm_type="layer",
            eps=eps,
            elementwise_affine=True,
-            dtype=torch.float32)
+            dtype=torch.float32,
+            compute_dtype=torch.float32)

        # 2. Cross-attention
        if added_kv_proj_dim is not None:
@@ -449,7 +459,8 @@ class WanTransformerBlock_VSA(nn.Module):
            norm_type="layer",
            eps=eps,
            elementwise_affine=False,
-            dtype=torch.float32)
+            dtype=torch.float32,
+            compute_dtype=torch.float32)

        # 3. Feed-forward
        self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
@@ -469,22 +480,23 @@ class WanTransformerBlock_VSA(nn.Module):
        bs, seq_length, _ = hidden_states.shape
        orig_dtype = hidden_states.dtype
        # assert orig_dtype != torch.float32
-        e = self.scale_shift_table + temb
+        e = self.scale_shift_table + temb.float()
        shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(
            6, dim=1)
+        assert shift_msa.dtype == torch.float32

        # 1. Self-attention
-        norm_hidden_states = (self.norm1(hidden_states) *
-                              (1 + scale_msa) + shift_msa)
+        norm_hidden_states = (self.norm1(hidden_states.float()) *
+                              (1 + scale_msa) + shift_msa).to(orig_dtype)
        query, _ = self.to_q(norm_hidden_states)
        key, _ = self.to_k(norm_hidden_states)
        value, _ = self.to_v(norm_hidden_states)
        gate_compress, _ = self.to_gate_compress(norm_hidden_states)

        if self.norm_q is not None:
-            query = self.norm_q.forward_native(query)
+            query = self.norm_q(query)
        if self.norm_k is not None:
-            key = self.norm_k.forward_native(key)
+            key = self.norm_k(key)

        query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
        key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -509,6 +521,8 @@ class WanTransformerBlock_VSA(nn.Module):
        null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
        norm_hidden_states, hidden_states = self.self_attn_residual_norm(
            hidden_states, attn_output, gate_msa, null_shift, null_scale)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)

        # 2. Cross-attention
        attn_output = self.attn2(norm_hidden_states,
@@ -516,15 +530,17 @@ class WanTransformerBlock_VSA(nn.Module):
                                 context_lens=None)
        norm_hidden_states, hidden_states = self.cross_attn_residual_norm(
            hidden_states, attn_output, 1, c_shift_msa, c_scale_msa)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)

        # 3. Feed-forward
        ff_output = self.ffn(norm_hidden_states)
        hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
+        hidden_states = hidden_states.to(orig_dtype)

        return hidden_states


-
 class WanTransformer3DModel(CachableDiT):
    _fsdp_shard_conditions = WanVideoConfig()._fsdp_shard_conditions
    _compile_conditions = WanVideoConfig()._compile_conditions
@@ -582,7 +598,8 @@ class WanTransformer3DModel(CachableDiT):
                                            norm_type="layer",
                                            eps=config.eps,
                                            elementwise_affine=False,
-                                            dtype=torch.float32)
+                                            dtype=torch.float32,
+                                            compute_dtype=torch.float32)
        self.proj_out = nn.Linear(
            inner_dim, config.out_channels * math.prod(config.patch_size))
        self.scale_shift_table = nn.Parameter(
@@ -642,12 +659,10 @@ class WanTransformer3DModel(CachableDiT):
            rope_theta=10000)
        freqs_cos = freqs_cos.to(hidden_states.device)
        freqs_sin = freqs_sin.to(hidden_states.device)
-        freqs_cis = (freqs_cos,
-                     freqs_sin) if freqs_cos is not None else None
+        freqs_cis = (freqs_cos.float(),
+                     freqs_sin.float()) if freqs_cos is not None else None

        hidden_states = self.patch_embedding(hidden_states)
-        grid_sizes = torch.stack(
-            [torch.tensor(hidden_states[0].shape[1:], dtype=torch.long)])
        hidden_states = hidden_states.flatten(2).transpose(1, 2)

        # timestep shape: batch_size, or batch_size, seq_len (wan 2.2 ti2v)
@@ -657,8 +672,6 @@ class WanTransformer3DModel(CachableDiT):
        else:
            ts_seq_len = None

-        encoder_hidden_states = torch.cat([encoder_hidden_states, encoder_hidden_states.new_zeros(1, self.text_len - encoder_hidden_states.size(1), encoder_hidden_states.size(2))], dim=1)
-
        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
            timestep, encoder_hidden_states, encoder_hidden_states_image, timestep_seq_len=ts_seq_len)
        if ts_seq_len is not None:
@@ -715,35 +728,14 @@ class WanTransformer3DModel(CachableDiT):
        hidden_states = self.norm_out(hidden_states, shift, scale)
        hidden_states = self.proj_out(hidden_states)

-        output = self.unpatchify(hidden_states, grid_sizes)
+        hidden_states = hidden_states.reshape(batch_size, post_patch_num_frames,
+                                              post_patch_height,
+                                              post_patch_width, p_t, p_h, p_w,
+                                              -1)
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

-        return torch.stack(output)
-    
-    def unpatchify(self, x, grid_sizes):
-        r"""
-
-
-        Args:
-            x (List[Tensor]):
-                List of patchified features, each with shape [L, C_out * prod(patch_size)]
-            grid_sizes (Tensor):
-                Original spatial-temporal grid dimensions before patching,
-
-
-        Returns:
-            Tensor:
-                Reconstructed video tensors with shape [B, C_out, F, H / 8, W / 8]
-        """
-
-        c = self.out_channels
-        out = []
-        for u, v in zip(x, grid_sizes.tolist()):
-            u = u[:math.prod(v)].view(*v, *self.patch_size, c)
-            u = u.permute(6, 0, 3, 1, 4, 2, 5)
-            # u = torch.einsum('fhwpqrc->cfphqwr', u.contiguous())
-            u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
-            out.append(u)
-        return out
+        return output

    def maybe_cache_states(self, hidden_states: torch.Tensor,
                           original_hidden_states: torch.Tensor) -> None:
@@ -835,4 +827,5 @@ class WanTransformer3DModel(CachableDiT):
        if self.is_even:
            return hidden_states + self.previous_residual_even
        else:
-            return hidden_states + self.previous_residual_odd
+            return hidden_states + self.previous_residual_odd
+        
@@ -415,6 +415,7 @@ class TransformerLoader(ComponentLoader):
            raise ValueError(
                "Model config does not contain a _class_name attribute. "
                "Only diffusers format is supported.")
+
        logger.info("transformer cls_name: %s", cls_name)
        if fastvideo_args.override_transformer_cls_name is not None:
            cls_name = fastvideo_args.override_transformer_cls_name
@@ -442,10 +443,16 @@ class TransformerLoader(ComponentLoader):

        if use_custom_weights:
            logger.info("Using custom initialization weights from: %s", custom_weights_path)
-            safetensors_list = [custom_weights_path]
+            assert custom_weights_path is not None, "Custom initialization weights must be provided"
+            if os.path.isdir(custom_weights_path):
+                safetensors_list = glob.glob(
+                    os.path.join(str(custom_weights_path), "*.safetensors"))
+            else:
+                assert custom_weights_path.endswith(".safetensors"), "Custom initialization weights must be a safetensors file"
+                safetensors_list = [custom_weights_path]

-        logger.info("Loading model from %s safetensors files in %s",
-                    len(safetensors_list), model_path)
+        logger.info("Loading model from %s safetensors files: %s",
+                    len(safetensors_list), safetensors_list)

        default_dtype = PRECISION_TO_TYPE[
            fastvideo_args.pipeline_config.dit_precision]
@@ -468,6 +475,7 @@ class TransformerLoader(ComponentLoader):
            pin_cpu_memory=fastvideo_args.pin_cpu_memory,
            fsdp_inference=fastvideo_args.use_fsdp_inference,
            # TODO(will): make these configurable
+            default_dtype=default_dtype,
            param_dtype=torch.bfloat16,
            reduce_dtype=torch.float32,
            output_dtype=None,
@@ -477,9 +485,8 @@ class TransformerLoader(ComponentLoader):
        total_params = sum(p.numel() for p in model.parameters())
        logger.info("Loaded model with %.2fB parameters", total_params / 1e9)

-        dtypes = set(param.dtype for param in model.parameters())
-        if len(dtypes) > 1:
-            model = model.to(default_dtype)
+        assert next(model.parameters()).dtype == default_dtype, "Model dtype does not match default dtype"
+
        model = model.eval()
        return model

@@ -62,6 +62,7 @@ def maybe_load_fsdp_model(
    device: torch.device,
    hsdp_replicate_dim: int,
    hsdp_shard_dim: int,
+    default_dtype: torch.dtype,
    param_dtype: torch.dtype,
    reduce_dtype: torch.dtype,
    cpu_offload: bool = False,
@@ -87,7 +88,8 @@ def maybe_load_fsdp_model(
        mp_policy=mp_policy,
    )

-    with set_default_dtype(param_dtype), torch.device("meta"):
+    logger.info("Loading model with default_dtype: %s", default_dtype)
+    with set_default_dtype(default_dtype), torch.device("meta"):
        model = model_cls(**init_params)

    # Check if we should use FSDP
@@ -125,7 +127,7 @@ def maybe_load_fsdp_model(
        model,
        weight_iterator,
        device,
-        param_dtype,
+        default_dtype,
        strict=True,
        cpu_offload=cpu_offload,
        param_names_mapping=param_names_mapping_fn,
@@ -64,8 +64,15 @@ class SelfForcingFlowMatchScheduler(BaseScheduler, ConfigMixin, SchedulerMixin):
    def step(self, model_output: torch.FloatTensor, timestep: torch.FloatTensor, sample: torch.FloatTensor, to_final=False, return_dict=False, **kwargs):
        if timestep.ndim == 2:
            timestep = timestep.flatten(0, 1)
+        elif timestep.ndim == 0:
+            # handles the case where timestep is a scalar, this occurs when we
+            # use this scheduler for ODE trajectory
+            timestep = timestep.unsqueeze(0)
+
        self.sigmas = self.sigmas.to(model_output.device)
        self.timesteps = self.timesteps.to(model_output.device)
+        timestep = timestep.to(model_output.device)
+
        timestep_id = torch.argmin(
            (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
@@ -171,9 +171,6 @@ def pred_noise_to_pred_video(pred_noise: torch.Tensor,
    # timestep shape should be [B]
    dtype = pred_noise.dtype
    device = pred_noise.device
-
-    # Convert to double following Self-Forcing
-    # https://github.com/guandeh17/Self-Forcing/blob/main/utils/wan_wrapper.py#L184
    pred_noise = pred_noise.double().to(device)
    noise_input_latent = noise_input_latent.double().to(device)
    sigmas = scheduler.sigmas.double().to(device)
@@ -7,8 +7,6 @@ This module wires the causal DMD denoising stage into the modular pipeline.

 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import (
-    FlowMatchEulerDiscreteScheduler)
 from fastvideo.pipelines import ComposedPipelineBase, LoRAPipeline

 # isort: off
@@ -40,7 +40,7 @@ class ComposedPipelineBase(ABC):
    _extra_config_module_map: dict[str, str] = {}
    training_args: TrainingArgs | None = None
    fastvideo_args: FastVideoArgs | TrainingArgs | None = None
-    modules: dict[str, torch.nn.Module] = {}
+    modules: dict[str, Any] = {}
    post_init_called: bool = False

    # TODO(will): args should support both inference args and training args
@@ -121,25 +121,14 @@ class ComposedPipelineBase(ABC):
                        model_path: str,
                        device: str | None = None,
                        torch_dtype: torch.dtype | None = None,
-                        pipeline_config: PipelineConfig | None = None,
+                        pipeline_config: str | PipelineConfig | None = None,
                        args: argparse.Namespace | None = None,
                        required_config_modules: list[str] | None = None,
                        loaded_modules: dict[str, torch.nn.Module]
                        | None = None,
                        **kwargs) -> "ComposedPipelineBase":
        """
-        Load a pipeline from a pretrained model. 
-        Few different patterns are supported:
-        - Only provide model_path:
-            - This will load the pipeline in inference mode.
-            - The pipeline will be initialized with the default config.
-            - The pipeline will be initialized with the default modules.
-            - The pipeline will be initialized with the default stages.
-            - The pipeline will be initialized with the default stages.
-        - override the default config using pipeline_config or args or kwargs
-        - override the default modules using loaded_modules
-        - override the pipelineconfig
-
+        Load a pipeline from a pretrained model.
        loaded_modules: Optional[Dict[str, torch.nn.Module]] = None,
        If provided, loaded_modules will be used instead of loading from config/pretrained weights.
        """
@@ -147,18 +136,9 @@ class ComposedPipelineBase(ABC):

            kwargs['model_path'] = model_path
            fastvideo_args = FastVideoArgs.from_kwargs(**kwargs)
-            if pipeline_config is not None:
-                fastvideo_args.pipeline_config = pipeline_config
-            if fastvideo_args.override_transformer_cls_name is not None:
-                pipeline_config = PipelineConfig.from_pretrained("wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers")
-                fastvideo_args.pipeline_config = pipeline_config
        else:
            assert args is not None, "args must be provided for training mode"
            fastvideo_args = TrainingArgs.from_cli_args(args)
-            if fastvideo_args.override_transformer_cls_name is not None:
-                pipeline_config = PipelineConfig.from_pretrained("wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers")
-                fastvideo_args.pipeline_config = pipeline_config
-                logger.info("in 2 Overriding transformer cls name to %s", fastvideo_args.override_transformer_cls_name)
            # TODO(will): fix this so that its not so ugly
            fastvideo_args.model_path = model_path
            for key, value in kwargs.items():
@@ -169,8 +149,7 @@ class ComposedPipelineBase(ABC):
            # model is loaded with the correct precision. Subsequently we will
            # use FSDP2's MixedPrecisionPolicy to set the precision for the
            # fwd, bwd, and other operations' precision.
-            fastvideo_args.pipeline_config.dit_precision = 'fp32'
-            # assert fastvideo_args.pipeline_config.dit_precision == 'fp32', 'only fp32 is supported for training'
+            assert fastvideo_args.pipeline_config.dit_precision == 'fp32', 'only fp32 is supported for training'

        logger.info("fastvideo_args in from_pretrained: %s", fastvideo_args)

@@ -151,7 +151,7 @@ class ForwardBatch:
    output: torch.Tensor | None = None
    return_trajectory_latents: bool = False
    return_trajectory_decoded: bool = False
-    trajectory_timesteps: list[int] | None = None
+    trajectory_timesteps: list[torch.Tensor] | None = None
    trajectory_latents: torch.Tensor | None = None
    trajectory_decoded: list[torch.Tensor] | None = None

@@ -246,7 +246,7 @@ class TrainingBatch:
    fake_score_loss: float = 0.0

    dmd_latent_vis_dict: dict[str, Any] = field(default_factory=dict)
-    latent_vis_dict: dict[str, torch.Tensor] = field(default_factory=dict)
+    latent_vis_dict: dict[str, Any] = field(default_factory=dict)
    fake_score_latent_vis_dict: dict[str, Any] = field(default_factory=dict)


@@ -12,25 +12,28 @@ import os
 from collections.abc import Iterator
 from typing import Any

-import numpy as np
 import pyarrow as pa
 import torch
-from PIL import Image
 from torch.utils.data import DataLoader
 from torchdata.stateful_dataloader import StatefulDataLoader
 from tqdm import tqdm

 from fastvideo.configs.sample import SamplingParam
-from fastvideo.dataset import getdataset
-from fastvideo.dataset.dataloader.schema import pyarrow_schema_ode_trajectory
-from fastvideo.distributed import get_local_torch_device
+from fastvideo.dataset import gettextdataset
+from fastvideo.dataset.dataloader.parquet_io import (ParquetDatasetWriter,
+                                                     records_to_table)
+from fastvideo.dataset.dataloader.record_schema import (
+    ode_text_only_record_creator)
+from fastvideo.dataset.dataloader.schema import (
+    pyarrow_schema_ode_trajectory_text_only)
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.scheduling_self_forcing_flow_match import (
+    SelfForcingFlowMatchScheduler)
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
 from fastvideo.pipelines.preprocess.preprocess_pipeline_base import (
    BasePreprocessPipeline)
 from fastvideo.pipelines.stages import (DecodingStage, DenoisingStage,
-                                        ImageVAEEncodingStage,
                                        InputValidationStage,
                                        LatentPreparationStage,
                                        TextEncodingStage,
@@ -40,128 +43,6 @@ from fastvideo.utils import save_decoded_latents_as_video, shallow_asdict
 logger = init_logger(__name__)


-class FlowMatchScheduler:
-
-    order = 1
-
-    def __init__(self,
-                 num_inference_steps=100,
-                 num_train_timesteps=1000,
-                 shift=3.0,
-                 sigma_max=1.0,
-                 sigma_min=0.003 / 1.002,
-                 inverse_timesteps=False,
-                 extra_one_step=False,
-                 reverse_sigmas=False):
-        self.num_train_timesteps = num_train_timesteps
-        self.shift = shift
-        self.sigma_max = sigma_max
-        self.sigma_min = sigma_min
-        self.inverse_timesteps = inverse_timesteps
-        self.extra_one_step = extra_one_step
-        self.reverse_sigmas = reverse_sigmas
-        self.set_timesteps(num_inference_steps)
-
-    def set_timesteps(self,
-                      num_inference_steps=100,
-                      denoising_strength=1.0,
-                      training=False,
-                      device=None):
-        sigma_start = self.sigma_min + \
-            (self.sigma_max - self.sigma_min) * denoising_strength
-        if self.extra_one_step:
-            self.sigmas = torch.linspace(sigma_start, self.sigma_min,
-                                         num_inference_steps + 1)[:-1]
-        else:
-            self.sigmas = torch.linspace(sigma_start, self.sigma_min,
-                                         num_inference_steps)
-        if self.inverse_timesteps:
-            self.sigmas = torch.flip(self.sigmas, dims=[0])
-        self.sigmas = self.shift * self.sigmas / \
-            (1 + (self.shift - 1) * self.sigmas)
-        if self.reverse_sigmas:
-            self.sigmas = 1 - self.sigmas
-        self.timesteps = self.sigmas * self.num_train_timesteps
-        if training:
-            x = self.timesteps
-            y = torch.exp(
-                -2 * ((x - num_inference_steps / 2) / num_inference_steps)**2)
-            y_shifted = y - y.min()
-            bsmntw_weighing = y_shifted * \
-                (num_inference_steps / y_shifted.sum())
-            self.linear_timesteps_weights = bsmntw_weighing
-
-    def step(self,
-             model_output,
-             timestep,
-             sample,
-             to_final=False,
-             return_dict=False,
-             **kwargs):
-        assert return_dict is False
-        assert kwargs == {}
-        self.sigmas = self.sigmas.to(model_output.device)
-        self.timesteps = self.timesteps.to(model_output.device)
-        logger.info('step timestep: %s', timestep)
-        logger.info('step timestep: %s', timestep.shape)
-        # timestep  is [num_frames]
-        # timestep_id = torch.argmin(
-        #     (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
-        # assert timestep.ndim == 1
-        # assert timestep.shape[0] == 1
-        timestep_id = torch.argmin((self.timesteps - timestep).abs(), dim=0)
-        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
-        if to_final or (timestep_id + 1 >= len(self.timesteps)).any():
-            sigma_ = 1 if (self.inverse_timesteps or self.reverse_sigmas) else 0
-        else:
-            sigma_ = self.sigmas[timestep_id + 1].reshape(-1, 1, 1, 1)
-        prev_sample = sample + model_output * (sigma_ - sigma)
-        return (prev_sample, )
-
-    def scale_model_input(self, sample: torch.Tensor, *args,
-                          **kwargs) -> torch.Tensor:
-        """
-        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
-        current timestep.
-
-        Args:
-            sample (`torch.Tensor`):
-                The input sample.
-
-        Returns:
-            `torch.Tensor`:
-                A scaled input sample.
-        """
-        return sample
-
-    def add_noise(self, original_samples, noise, timestep):
-        """
-        Diffusion forward corruption process.
-        Input:
-            - clean_latent: the clean latent with shape [B, C, H, W]
-            - noise: the noise with shape [B, C, H, W]
-            - timestep: the timestep with shape [B]
-        Output: the corrupted latent with shape [B, C, H, W]
-        """
-        self.sigmas = self.sigmas.to(noise.device)
-        self.timesteps = self.timesteps.to(noise.device)
-        timestep_id = torch.argmin(
-            (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
-        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
-        sample = (1 - sigma) * original_samples + sigma * noise
-        return sample.type_as(noise)
-
-    def training_target(self, sample, noise, timestep):
-        target = noise - sample
-        return target
-
-    def training_weight(self, timestep):
-        timestep_id = torch.argmin(
-            (self.timesteps - timestep.to(self.timesteps.device)).abs())
-        weights = self.linear_timesteps_weights[timestep_id]
-        return weights
-
-
 class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
    """ODE Trajectory preprocessing pipeline implementation."""

@@ -170,28 +51,22 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
    ]
    preprocess_dataloader: StatefulDataLoader
    preprocess_loader_iter: Iterator[dict[str, Any]]
+    pbar: Any
+    num_processed_samples: int

-    def get_schema_fields(self):
-        """Get the schema fields for ODE Trajectory pipeline."""
-        return [f.name for f in pyarrow_schema_ode_trajectory]
+    def get_pyarrow_schema(self) -> pa.Schema:
+        """Return the PyArrow schema for ODE Trajectory pipeline."""
+        return pyarrow_schema_ode_trajectory_text_only

    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
        """Set up pipeline stages with proper dependency injection."""
-        fastvideo_args.pipeline_config.flow_shift = 5
-        logger.info('WTF flow_shift: %s',
-                    fastvideo_args.pipeline_config.flow_shift)
-
        assert fastvideo_args.pipeline_config.flow_shift == 5
-        # self.modules["scheduler"] = FlowMatchEulerDiscreteScheduler(
-        # shift=fastvideo_args.pipeline_config.flow_shift)
-        self.modules["scheduler"] = FlowMatchScheduler(
+        self.modules["scheduler"] = SelfForcingFlowMatchScheduler(
            shift=fastvideo_args.pipeline_config.flow_shift,
            sigma_min=0.0,
            extra_one_step=True)
        self.modules["scheduler"].set_timesteps(num_inference_steps=48,
                                                denoising_strength=1.0)
-        logger.info('WTF scheduler timesteps: %s',
-                    self.modules["scheduler"].timesteps)

        self.add_stage(stage_name="input_validation_stage",
                       stage=InputValidationStage())
@@ -200,9 +75,6 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                           text_encoders=[self.get_module("text_encoder")],
                           tokenizers=[self.get_module("tokenizer")],
                       ))
-        self.add_stage(stage_name="vae_encoding_stage",
-                       stage=ImageVAEEncodingStage(
-                           vae=self.get_module("vae"), ))
        self.add_stage(stage_name="timestep_preparation_stage",
                       stage=TimestepPreparationStage(
                           scheduler=self.get_module("scheduler")))
@@ -213,56 +85,47 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
        self.add_stage(stage_name="denoising_stage",
                       stage=DenoisingStage(
                           transformer=self.get_module("transformer"),
-                           transformer_2=self.get_module("transformer_2", None),
                           scheduler=self.get_module("scheduler"),
                           pipeline=self,
                       ))
        self.add_stage(stage_name="decoding_stage",
                       stage=DecodingStage(vae=self.get_module("vae")))

-    def preprocess_video_and_text_and_trajectory(self,
-                                                 fastvideo_args: FastVideoArgs,
-                                                 args):
+    def preprocess_text_and_trajectory(self, fastvideo_args: FastVideoArgs,
+                                       args):
+        """Preprocess text-only data and generate trajectory information."""

        for batch_idx, data in enumerate(self.pbar):
            if data is None:
                continue

            with torch.inference_mode():
-                # Filter out invalid samples (those with all zeros)
+                # For text-only processing, we only need text data
+                # Filter out samples without text
                valid_indices = []
-                for i, pixel_values in enumerate(data["pixel_values"]):
-                    if not torch.all(
-                            pixel_values == 0):  # Check if all values are zero
+                for i, text in enumerate(data["text"]):
+                    if text and text.strip():  # Check if text is not empty
                        valid_indices.append(i)
                self.num_processed_samples += len(valid_indices)

                if not valid_indices:
                    continue

-                # Create new batch with only valid samples
+                # Create new batch with only valid samples (text-only)
                valid_data = {
-                    "pixel_values":
-                    torch.stack(
-                        [data["pixel_values"][i] for i in valid_indices]),
                    "text": [data["text"][i] for i in valid_indices],
                    "path": [data["path"][i] for i in valid_indices],
-                    "fps": [data["fps"][i] for i in valid_indices],
-                    "duration": [data["duration"][i] for i in valid_indices],
                }

-                # VAE
-                with torch.autocast("cuda", dtype=torch.float32):
-                    latents = self.get_module("vae").encode(
-                        valid_data["pixel_values"].to(
-                            get_local_torch_device())).mean
-
-                # Get extra features if needed
-                extra_features = self.get_extra_features(
-                    valid_data, fastvideo_args)
+                # Add fps and duration if available in data
+                if "fps" in data:
+                    valid_data["fps"] = [data["fps"][i] for i in valid_indices]
+                if "duration" in data:
+                    valid_data["duration"] = [
+                        data["duration"][i] for i in valid_indices
+                    ]

                batch_captions = valid_data["text"]
-                logger.info(f"===== batch_captions: {batch_captions}")
                # Encode text using the standalone TextEncodingStage API
                prompt_embeds_list, prompt_masks_list = self.prompt_encoding_stage.encode_text(
                    batch_captions,
@@ -274,27 +137,6 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                prompt_attention_masks = prompt_masks_list[0]
                assert prompt_embeds.shape[0] == prompt_attention_masks.shape[0]

-                # # Get sequence lengths from attention masks (number of 1s)
-                # seq_lens = prompt_attention_mask.sum(dim=1)
-
-                # non_padded_embeds = []
-                # non_padded_masks = []
-
-                # # Process each item in the batch
-                # for i in range(prompt_embeds.size(0)):
-                #     seq_len = seq_lens[i].item()
-                #     # Slice the embeddings and masks to keep only non-padding parts
-                #     non_padded_embeds.append(prompt_embeds[i, :seq_len])
-                #     non_padded_masks.append(prompt_attention_mask[i, :seq_len])
-
-                # Update the tensors with non-padded versions
-                # prompt_embeds = non_padded_embeds
-                # prompt_attention_masks = non_padded_masks
-                # prompt_embeds = prompt_embeds
-
-                # logger.info(f"===== prompt_embeds: {prompt_embeds[0].shape}")
-                # logger.info(f"===== prompt_attention_masks: {prompt_attention_masks[0].shape}")
-
                sampling_params = SamplingParam.from_pretrained(args.model_path)

                # encode negative prompt for trajectory collection
@@ -315,52 +157,34 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                trajectory_latents = []
                trajectory_timesteps = []
                trajectory_decoded = []
+
                for i, (prompt_embed, prompt_attention_mask) in enumerate(
-                        zip(prompt_embeds, prompt_attention_masks, strict=False)):
+                        zip(prompt_embeds, prompt_attention_masks,
+                            strict=False)):
                    prompt_embed = prompt_embed.unsqueeze(0)
                    prompt_attention_mask = prompt_attention_mask.unsqueeze(0)
-                    logger.info("what")
-                    logger.info(f"===== prompt_embed: {prompt_embed.shape}")
-                    logger.info(
-                        f"===== prompt_attention_mask: {prompt_attention_mask.shape}"
-                    )
-                    # Collect the trajectory data
-                    batch = ForwardBatch(
-                        **shallow_asdict(sampling_params),
-                        # data_type="video",
-                        # seed=args.seed,
-                        # prompt=batch_captions[i],
-                        # prompt_embeds=[prompt_embed],
-                        # prompt_attention_mask=[prompt_attention_mask],
-                        # height=args.max_height,
-                        # width=args.max_width,
-                        # num_frames=81,
-                        # fps=args.train_fps,
-                        # return_trajectory_latents=True,
-                        # guidance_scale=3.0,
-                        # do_classifier_free_guidance=True,
-                    )
+
+                    # Collect the trajectory data (text-to-video generation)
+                    batch = ForwardBatch(**shallow_asdict(sampling_params), )
                    batch.prompt_embeds = [prompt_embed]
                    batch.prompt_attention_mask = [prompt_attention_mask]
                    batch.negative_prompt_embeds = [negative_prompt_embed]
                    batch.negative_attention_mask = [
                        negative_prompt_attention_mask
                    ]
+                    batch.num_inference_steps = 48
                    batch.return_trajectory_latents = True
+                    # Enabling this will save the decoded trajectory videos.
+                    # Used for debugging.
                    batch.return_trajectory_decoded = False
                    batch.height = args.max_height
                    batch.width = args.max_width
-                    batch.num_inference_steps = 48
-                    # batch.num_frames = 81
                    batch.fps = args.train_fps
                    batch.guidance_scale = 6.0
                    batch.do_classifier_free_guidance = True
-                    # fastvideo_args.pipeline_config.ti2v_task = True

                    result_batch = self.input_validation_stage(
                        batch, fastvideo_args)
-                    # result_batch = self.prompt_encoding_stage(result_batch, fastvideo_args)
-                    # result_batch = self.vae_encoding_stage(result_batch, fastvideo_args)
                    result_batch = self.timestep_preparation_stage(
                        batch, fastvideo_args)
                    result_batch = self.latent_preparation_stage(
@@ -369,252 +193,97 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                                                        fastvideo_args)
                    result_batch = self.decoding_stage(result_batch,
                                                       fastvideo_args)
-                    # trajectory_latents = result_batch.trajectory_latents
+
                    trajectory_latents.append(
                        result_batch.trajectory_latents.cpu())
                    trajectory_timesteps.append(
                        result_batch.trajectory_timesteps.cpu())
                    trajectory_decoded.append(result_batch.trajectory_decoded)

-            extra_features["trajectory_latents"] = trajectory_latents
-            extra_features["trajectory_timesteps"] = trajectory_timesteps
-            logger.info(
-                f"===== trajectory_latents: {trajectory_latents[0].shape}")
-            logger.info(
-                f"===== trajectory_latents len: {len(trajectory_latents)}")
-            logger.info(f"===== trajectory_timesteps: {trajectory_timesteps}")
-            logger.info(
-                f"===== trajectory_timesteps len: {len(trajectory_timesteps)}")
+                # Prepare extra features for text-only processing
+                extra_features = {
+                    "trajectory_latents": trajectory_latents,
+                    "trajectory_timesteps": trajectory_timesteps
+                }

-            if batch.return_trajectory_decoded:
-                logger.info("===== SAVING TRAJECTORY DECODED")
-                for i, decoded_frames in enumerate(trajectory_decoded):
-                    for j, decoded_frame in enumerate(decoded_frames):
-                        logger.info(
-                            f"===== SAVING TRAJECTORY DECODED {i} for prompt {batch_captions[i]}"
-                        )
-                        save_decoded_latents_as_video(
-                            decoded_frame,
-                            f"decoded_videos/trajectory_decoded_{i}_{j}.mp4",
-                            args.train_fps)
-            # assert False
-            # Prepare batch data for Parquet dataset
-            batch_data = []
+                if batch.return_trajectory_decoded:
+                    for i, decoded_frames in enumerate(trajectory_decoded):
+                        for j, decoded_frame in enumerate(decoded_frames):
+                            save_decoded_latents_as_video(
+                                decoded_frame,
+                                f"decoded_videos/trajectory_decoded_{i}_{j}.mp4",
+                                args.train_fps)

-            # Add progress bar for saving outputs
-            save_pbar = tqdm(enumerate(valid_data["path"]),
-                             desc="Saving outputs",
-                             unit="item",
-                             leave=False)
-            for idx, video_path in save_pbar:
-                # Get the corresponding latent and info using video name
-                latent = latents[idx].cpu()
-                video_name = os.path.basename(video_path).split(".")[0]
+                # Prepare batch data for Parquet dataset
+                batch_data: list[dict[str, Any]] = []

-                # Convert tensors to numpy arrays
-                vae_latent = latent.cpu().numpy()
-                text_embedding = prompt_embeds[idx].cpu().numpy()
+                # Add progress bar for saving outputs
+                save_pbar = tqdm(enumerate(valid_data["path"]),
+                                 desc="Saving outputs",
+                                 unit="item",
+                                 leave=False)

-                # Get extra features for this sample if needed
-                sample_extra_features = {}
-                if extra_features:
-                    for key, value in extra_features.items():
-                        logger.info(f"===== key: {key}")
-                        if isinstance(value, torch.Tensor):
-                            logger.info(f"===== value: {value[idx].shape}")
-                            sample_extra_features[key] = value[idx].cpu().numpy(
-                            )
-                        else:
-                            assert isinstance(value, list)
-                            if isinstance(value[idx], torch.Tensor):
-                                logger.info(
-                                    f"===== value in list: {value[idx].shape}")
+                for idx, video_path in save_pbar:
+                    video_name = os.path.basename(video_path).split(".")[0]
+
+                    # Convert tensors to numpy arrays
+                    text_embedding = prompt_embeds[idx].cpu().numpy()
+
+                    # Get extra features for this sample
+                    sample_extra_features = {}
+                    if extra_features:
+                        for key, value in extra_features.items():
+                            if isinstance(value, torch.Tensor):
                                sample_extra_features[key] = value[idx].cpu(
-                                ).float().numpy()
+                                ).numpy()
                            else:
-                                logger.info("===== value in list: not tensor")
-                                sample_extra_features[key] = value[idx]
-                            # logger.info(f"===== value: not tensor")
-                            # sample_extra_features[key] = value[idx]
+                                assert isinstance(value, list)
+                                if isinstance(value[idx], torch.Tensor):
+                                    sample_extra_features[key] = value[idx].cpu(
+                                    ).float().numpy()
+                                else:
+                                    sample_extra_features[key] = value[idx]

-                # Create record for Parquet dataset
-                record = self.create_record(
-                    video_name=video_name,
-                    vae_latent=vae_latent,
-                    text_embedding=text_embedding,
-                    valid_data=valid_data,
-                    idx=idx,
-                    extra_features=sample_extra_features)
-                batch_data.append(record)
+                    # Create record for Parquet dataset (text-only ODE schema)
+                    record: dict[str, Any] = ode_text_only_record_creator(
+                        video_name=video_name,
+                        text_embedding=text_embedding,
+                        caption=valid_data["text"][idx],
+                        trajectory_latents=sample_extra_features[
+                            "trajectory_latents"],
+                        trajectory_timesteps=sample_extra_features[
+                            "trajectory_timesteps"],
+                    )
+                    batch_data.append(record)

-            if batch_data:
-                # Add progress bar for writing to Parquet dataset
-                write_pbar = tqdm(total=1,
-                                  desc="Writing to Parquet dataset",
-                                  unit="batch")
-                # Convert batch data to PyArrow arrays
-                arrays = []
-                for field in self.get_schema_fields():
-                    if field.endswith('_bytes'):
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.binary()))
-                    elif field.endswith('_shape'):
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.list_(pa.int32())))
-                    elif field in ['width', 'height', 'num_frames']:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.int32()))
-                    elif field in ['duration_sec', 'fps']:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.float32()))
-                    else:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data]))
+                if batch_data:
+                    write_pbar = tqdm(total=1,
+                                      desc="Writing to Parquet dataset",
+                                      unit="batch")
+                    table = records_to_table(batch_data,
+                                             self.get_pyarrow_schema())
+                    write_pbar.update(1)
+                    write_pbar.close()

-                table = pa.Table.from_arrays(arrays,
-                                             names=self.get_schema_fields())
-                write_pbar.update(1)
-                write_pbar.close()
+                    if not hasattr(self, 'dataset_writer'):
+                        self.dataset_writer = ParquetDatasetWriter(
+                            out_dir=self.combined_parquet_dir,
+                            samples_per_file=args.samples_per_file,
+                        )
+                    self.dataset_writer.append_table(table)

-                # Store the table in a list for later processing
-                if not hasattr(self, 'all_tables'):
-                    self.all_tables = []
-                self.all_tables.append(table)
+                    logger.info("Collected batch with %s samples", len(table))

-                logger.info("Collected batch with %s samples", len(table))
+                if self.num_processed_samples >= args.flush_frequency:
+                    written = self.dataset_writer.flush()
+                    logger.info("Flushed %s samples to parquet", written)
+                    self.num_processed_samples = 0

-            if self.num_processed_samples >= args.flush_frequency:
-                self._flush_tables(self.num_processed_samples, args,
-                                   self.combined_parquet_dir)
-                self.num_processed_samples = 0
-                self.all_tables = []
-
-    def get_extra_features(self, valid_data: dict[str, Any],
-                           fastvideo_args: FastVideoArgs) -> dict[str, Any]:
-
-        # TODO(will): move these to cpu at some point
-        self.get_module("vae").to(get_local_torch_device())
-
-        # generator = torch.Generator(device=get_local_torch_device(), seed=42)
-        generator = torch.Generator("cpu").manual_seed(42)
-
-        features = {}
-        """Get CLIP features from the first frame of each video."""
-        first_frame = valid_data["pixel_values"][:, :, 0, :, :].permute(
-            0, 2, 3, 1)  # (B, C, T, H, W) -> (B, H, W, C)
-        _, _, num_frames, height, width = valid_data["pixel_values"].shape
-        # latent_height = height // self.get_module(
-        #     "vae").spatial_compression_ratio
-        # latent_width = width // self.get_module("vae").spatial_compression_ratio
-
-        unprocessed_images = []
-        pil_images = []
-        # Frame has values between -1 and 1
-        for frame in first_frame:
-            frame = (frame + 1) * 127.5
-            frame_pil = Image.fromarray(frame.cpu().numpy().astype(np.uint8))
-            pil_images.append(frame_pil)
-            # processed_img = self.get_module("image_processor")(
-            #     images=frame_pil, return_tensors="pt")
-            unprocessed_images.append(frame_pil)
-        """Get VAE features from the first frame of each video"""
-        video_conditions = []
-        for frame in unprocessed_images:
-
-            latent = self.vae_encoding_stage.encode_image(
-                frame, height, width, fastvideo_args, generator)
-            video_conditions.append(latent)
-
-        features["image_condition_latents"] = video_conditions
-        features["pil_images"] = pil_images
-        return features
-
-    def create_record(
-            self,
-            video_name: str,
-            vae_latent: np.ndarray,
-            text_embedding: np.ndarray,
-            valid_data: dict[str, Any],
-            idx: int,
-            extra_features: dict[str, Any] | None = None) -> dict[str, Any]:
-        """Create a record for the Parquet dataset with CLIP features."""
-        record = super().create_record(video_name=video_name,
-                                       vae_latent=vae_latent,
-                                       text_embedding=text_embedding,
-                                       valid_data=valid_data,
-                                       idx=idx,
-                                       extra_features=extra_features)
-
-        if extra_features and "image_condition_latents" in extra_features:
-            image_condition_latents = extra_features["image_condition_latents"]
-            record.update({
-                "image_condition_latents_bytes":
-                image_condition_latents.tobytes(),
-                "image_condition_latents_shape":
-                list(image_condition_latents.shape),
-                "image_condition_latents_dtype":
-                str(image_condition_latents.dtype),
-            })
-        else:
-            record.update({
-                "image_condition_latents_bytes": b"",
-                "image_condition_latents_shape": [],
-                "image_condition_latents_dtype": "",
-            })
-
-        if extra_features and "trajectory_latents" in extra_features:
-            trajectory_latents = extra_features["trajectory_latents"]
-            record.update({
-                "trajectory_latents_bytes":
-                trajectory_latents.tobytes(),
-                "trajectory_latents_shape":
-                list(trajectory_latents.shape),
-                "trajectory_latents_dtype":
-                str(trajectory_latents.dtype),
-            })
-        else:
-            record.update({
-                "trajectory_latents_bytes": b"",
-                "trajectory_latents_shape": [],
-                "trajectory_latents_dtype": "",
-            })
-
-        if extra_features and "trajectory_timesteps" in extra_features:
-            trajectory_timesteps = extra_features["trajectory_timesteps"]
-            record.update({
-                "trajectory_timesteps_bytes":
-                trajectory_timesteps.tobytes(),
-                "trajectory_timesteps_shape":
-                list(trajectory_timesteps.shape),
-                "trajectory_timesteps_dtype":
-                str(trajectory_timesteps.dtype),
-            })
-        else:
-            record.update({
-                "trajectory_timesteps_bytes": b"",
-                "trajectory_timesteps_shape": [],
-                "trajectory_timesteps_dtype": "",
-            })
-
-        if extra_features and "pil_image" in extra_features:
-            pil_image = extra_features["pil_image"]
-            record.update({
-                "pil_image_bytes": pil_image.tobytes(),
-                "pil_image_shape": list(pil_image.shape),
-                "pil_image_dtype": str(pil_image.dtype),
-            })
-        else:
-            record.update({
-                "pil_image_bytes": b"",
-                "pil_image_shape": [],
-                "pil_image_dtype": "",
-            })
-
-        return record
+        # Final flush for any remaining samples
+        if hasattr(self, 'dataset_writer'):
+            written = self.dataset_writer.flush(write_remainder=True)
+            if written:
+                logger.info("Final flush wrote %s samples", written)

    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs, args):
        if not self.post_init_called:
@@ -628,7 +297,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
        os.makedirs(self.combined_parquet_dir, exist_ok=True)

        # Loading dataset
-        train_dataset = getdataset(args)
+        train_dataset = gettextdataset(args)

        self.preprocess_dataloader = DataLoader(
            train_dataset,
@@ -648,7 +317,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
        # Initialize class variables for data sharing
        self.video_data: dict[str, Any] = {}  # Store video metadata and paths
        self.latent_data: dict[str, Any] = {}  # Store latent tensors
-        self.preprocess_video_and_text_and_trajectory(fastvideo_args, args)
+        self.preprocess_text_and_trajectory(fastvideo_args, args)


 EntryClass = PreprocessPipeline_ODE_Trajectory
@@ -56,6 +56,10 @@ def main(args) -> None:
        PreprocessPipeline = PreprocessPipeline_I2V
    elif args.preprocess_task == "text_only":
        PreprocessPipeline = PreprocessPipeline_Text
+    elif args.preprocess_task == "ode_trajectory":
+        assert args.flow_shift is not None, "flow_shift is required for ode_trajectory"
+        fastvideo_args.pipeline_config.flow_shift = args.flow_shift
+        PreprocessPipeline = PreprocessPipeline_ODE_Trajectory
    else:
        raise ValueError(f"Invalid preprocess task: {args.preprocess_task}. "
                         f"Valid options: t2v, i2v, ode_trajectory, text_only")
@@ -101,10 +105,11 @@ if __name__ == "__main__":
    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("--flow_shift", type=float, default=None)
    parser.add_argument("--preprocess_task",
                        type=str,
                        default="t2v",
-                        choices=["t2v", "i2v", "text_only"],
+                        choices=["t2v", "i2v", "text_only", "ode_trajectory"],
                        help="Type of preprocessing task to run")
    parser.add_argument("--train_fps", type=int, default=30)
    parser.add_argument("--use_image_num", type=int, default=0)
@@ -78,8 +78,6 @@ class CausalDMDDenosingStage(DenoisingStage):
                                             torch.tensor([0],
                                                          dtype=torch.float32)))
            timesteps = scheduler_timesteps[1000 - timesteps]
-        else:
-            assert False, "warp_denoising_step must be true"
        timesteps = timesteps.to(get_local_torch_device())
        logger.info("Using timesteps: %s", timesteps)

@@ -53,7 +53,20 @@ class DecodingStage(PipelineStage):
    @torch.no_grad()
    def decode(self, latents: torch.Tensor,
               fastvideo_args: FastVideoArgs) -> torch.Tensor:
-        """Decode latents into pixel space."""
+        """
+        Decode latent representations into pixel space using VAE.
+        
+        Args:
+            latents: Input latent tensor with shape (batch, channels, frames, height_latents, width_latents)
+            fastvideo_args: Configuration containing:
+                - disable_autocast: Whether to disable automatic mixed precision (default: False)
+                - pipeline_config.vae_precision: VAE computation precision ("fp32", "fp16", "bf16")
+                - pipeline_config.vae_tiling: Whether to enable VAE tiling for memory efficiency
+            
+        Returns:
+            Decoded video tensor with shape (batch, channels, frames, height, width), 
+            normalized to [0, 1] range and moved to CPU as float32
+        """
        self.vae = self.vae.to(get_local_torch_device())
        latents = latents.to(get_local_torch_device())

@@ -103,12 +116,26 @@ class DecodingStage(PipelineStage):
        """
        Decode latent representations into pixel space.
        
+        This method processes the batch through the VAE decoder, converting latent
+        representations to pixel-space video/images. It also optionally decodes
+        trajectory latents for visualization purposes.
+        
        Args:
-            batch: The current batch information.
-            fastvideo_args: The inference arguments.
+            batch: The current batch containing:
+                - latents: Tensor to decode (batch, channels, frames, height_latents, width_latents)
+                - return_trajectory_decoded (optional): Flag to decode trajectory latents
+                - trajectory_latents (optional): Latents at different timesteps
+                - trajectory_timesteps (optional): Corresponding timesteps
+            fastvideo_args: Configuration containing:
+                - output_type: "latent" to skip decoding, otherwise decode to pixels
+                - vae_cpu_offload: Whether to offload VAE to CPU after decoding
+                - model_loaded: Track VAE loading state
+                - model_paths: Path to VAE model if loading needed
            
        Returns:
-            The batch with decoded outputs.
+            Modified batch with:
+                - output: Decoded frames (batch, channels, frames, height, width) as CPU float32
+                - trajectory_decoded (if requested): List of decoded frames per timestep
        """
        # load vae if not already loaded (used for memory constrained devices)
        pipeline = self.pipeline() if self.pipeline else None
@@ -128,15 +155,13 @@ class DecodingStage(PipelineStage):
        # decode trajectory latents if needed
        if batch.return_trajectory_decoded:
            batch.trajectory_decoded = []
-            logger.info(f"batch.trajectory_latents.shape: {batch.trajectory_latents.shape}")
            assert batch.trajectory_latents is not None, "batch should have trajectory latents"
            for idx in range(batch.trajectory_latents.shape[1]):
-                # bathc.trajectory_latents is [batch_size, timesteps, channels, frames, height, width]
+                # batch.trajectory_latents is [batch_size, timesteps, channels, frames, height, width]
                cur_latent = batch.trajectory_latents[:, idx, :, :, :, :]
-                logger.info(f"cur_latent.shape: {cur_latent.shape}")
                cur_timestep = batch.trajectory_timesteps[idx]
-                logger.info(
-                    f"decoding trajectory latent for timestep: {cur_timestep}")
+                logger.info("decoding trajectory latent for timestep: %s",
+                            cur_timestep)
                decoded_frames = self.decode(cur_latent, fastvideo_args)
                batch.trajectory_decoded.append(decoded_frames.cpu().float())

@@ -85,8 +85,9 @@ class DenoisingStage(PipelineStage):
                AttentionBackendEnum.SLIDING_TILE_ATTN,
                AttentionBackendEnum.VIDEO_SPARSE_ATTN,
                AttentionBackendEnum.VMOBA_ATTN,
-                AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA
-            )  # hack
+                AttentionBackendEnum.FLASH_ATTN,
+                AttentionBackendEnum.TORCH_SDPA,
+                AttentionBackendEnum.SAGE_ATTN_THREE)  # hack
        )

    def forward(
@@ -140,12 +141,11 @@ class DenoisingStage(PipelineStage):
            latents = latents[:, :, rank_in_sp_group, :, :, :]
            batch.latents = latents
            if batch.image_latent is not None:
-                if not fastvideo_args.pipeline_config.ti2v_task and not fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    image_latent = rearrange(batch.image_latent,
-                                             "b c (n t) h w -> b c n t h w",
-                                             n=sp_world_size).contiguous()
-                    image_latent = image_latent[:, :, rank_in_sp_group, :, :, :]
-                    batch.image_latent = image_latent
+                image_latent = rearrange(batch.image_latent,
+                                         "b c (n t) h w -> b c n t h w",
+                                         n=sp_world_size).contiguous()
+                image_latent = image_latent[:, :, rank_in_sp_group, :, :, :]
+                batch.image_latent = image_latent
        # Get timesteps and calculate warmup steps
        timesteps = batch.timesteps
        # TODO(will): remove this once we add input/output validation for stages
@@ -205,14 +205,14 @@ class DenoisingStage(PipelineStage):
                neg_prompt_embeds[0]).any(), "neg_prompt_embeds contains nan"

        # (Wan2.2) Calculate timestep to switch from high noise expert to low noise expert
-        if fastvideo_args.pipeline_config.dit_config.boundary_ratio is not None:
-            boundary_timestep = fastvideo_args.pipeline_config.dit_config.boundary_ratio
-            if batch.boundary_timestep is not None:
-                logger.info("Overriding boundary timestep from %s to %s",
-                            boundary_timestep, batch.boundary_timestep)
-                boundary_timestep = batch.boundary_timestep
+        boundary_ratio = fastvideo_args.pipeline_config.dit_config.boundary_ratio
+        if batch.boundary_ratio is not None:
+            logger.info("Overriding boundary ratio from %s to %s",
+                        boundary_ratio, batch.boundary_ratio)
+            boundary_ratio = batch.boundary_ratio

-            boundary_timestep *= self.scheduler.num_train_timesteps
+        if boundary_ratio is not None:
+            boundary_timestep = boundary_ratio * self.scheduler.num_train_timesteps
        else:
            boundary_timestep = None
        latent_model_input = latents.to(target_dtype)
@@ -254,7 +254,8 @@ class DenoisingStage(PipelineStage):
                                                             patch_size[2])
            seq_len = int(math.ceil(seq_len / sp_world_size)) * sp_world_size

-        trajectory_timesteps: list[int] = []
+        # Initialize lists for ODE trajectory
+        trajectory_timesteps: list[torch.Tensor] = []
        trajectory_latents: list[torch.Tensor] = []

        # Run denoising loop
@@ -284,27 +285,14 @@ class DenoisingStage(PipelineStage):

                # Expand latents for I2V
                latent_model_input = latents.to(target_dtype)
-                if batch.image_latent is not None and not fastvideo_args.pipeline_config.t2v_as_i2v_task:
+                if batch.image_latent is not None:
                    assert not fastvideo_args.pipeline_config.ti2v_task, "image latents should not be provided for TI2V task"
                    latent_model_input = torch.cat(
                        [latent_model_input, batch.image_latent],
                        dim=1).to(target_dtype)
-                elif batch.image_latent is not None and fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    assert batch.image_latent is not None, "image latents should be provided for T2V to I2V task"
-                    if rank_in_sp_group == 0:
-                        logger.info("latent_model_input.shape: %s",
-                                    latent_model_input.shape)
-                        latent_model_input = torch.cat([
-                            batch.image_latent,
-                            latent_model_input[:, :, 1:, :, :],
-                        ],
-                                                       dim=2).to(target_dtype)
-                        logger.info("latent_model_input.shape: %s",
-                                    latent_model_input.shape)

                assert not torch.isnan(
                    latent_model_input).any(), "latent_model_input contains nan"
-
                if fastvideo_args.pipeline_config.ti2v_task and batch.pil_image is not None:
                    timestep = torch.stack([t]).to(get_local_torch_device())
                    temp_ts = (mask2[0][0][:, ::2, ::2] * timestep).flatten()
@@ -319,13 +307,6 @@ class DenoisingStage(PipelineStage):

                latent_model_input = self.scheduler.scale_model_input(
                    latent_model_input, t)
-                if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    if rank_in_sp_group == 0:
-                        latent_model_input = torch.cat([
-                            batch.image_latent,
-                            latent_model_input[:, :, 1:, :, :],
-                        ],
-                                                       dim=2).to(target_dtype)

                # Prepare inputs for transformer
                guidance_expand = (
@@ -460,7 +441,6 @@ class DenoisingStage(PipelineStage):
                # save trajectory latents if needed
                if batch.return_trajectory_latents:
                    trajectory_timesteps.append(t)
-                    # trajectory_latents.append(latents.cpu())
                    trajectory_latents.append(latents)

                # Update progress bar
@@ -474,30 +454,25 @@ class DenoisingStage(PipelineStage):
        trajectory_tensor: torch.Tensor | None = None
        if trajectory_latents:
            trajectory_tensor = torch.stack(trajectory_latents, dim=1)
+            trajectory_timesteps_tensor = torch.stack(trajectory_timesteps,
+                                                      dim=0)
        else:
            trajectory_tensor = None
+            trajectory_timesteps_tensor = None

+        # Gather results if using sequence parallelism
        if sp_group:
            latents = sequence_model_parallel_all_gather(latents, dim=2)
            if batch.return_trajectory_latents:
-                # logger.info("before stack trajectory_latents.shape: %s", trajectory_latents[0].shape)
-                logger.info("after stack trajectory_latents.shape: %s", trajectory_tensor.shape)
                trajectory_tensor = trajectory_tensor.to(
                    get_local_torch_device())
                trajectory_tensor = sequence_model_parallel_all_gather(
                    trajectory_tensor, dim=3)

-        if trajectory_tensor is not None:
-            batch.trajectory_timesteps = torch.tensor(trajectory_timesteps).cpu()
+        if trajectory_tensor is not None and trajectory_timesteps_tensor is not None:
+            batch.trajectory_timesteps = trajectory_timesteps_tensor.cpu()
            batch.trajectory_latents = trajectory_tensor.cpu()

-        if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            latents = torch.cat([
-                batch.image_latent,
-                latents[:, :, 1:, :, :],
-            ],
-                                dim=2)
-
        # Update batch with final latents
        batch.latents = latents

@@ -105,81 +105,6 @@ class ImageVAEEncodingStage(PipelineStage):
    def __init__(self, vae: ParallelTiledVAE) -> None:
        self.vae: ParallelTiledVAE = vae

-    def encode_image(self,
-                     image: PIL.Image.Image,
-                     height: int,
-                     width: int,
-                     fastvideo_args: FastVideoArgs,
-                     generator: torch.Generator | None = None) -> torch.Tensor:
-        """
-        Encode image into latent space.
-        """
-        image = self.preprocess(
-            image,
-            vae_scale_factor=self.vae.spatial_compression_ratio,
-            height=height,
-            width=width).to(get_local_torch_device(), dtype=torch.float32)
-
-        # (B, C, H, W) -> (B, C, 1, H, W)
-        print(f"image.shape: {image.shape}")
-        image = image.unsqueeze(2)
-        print(f"after unsqueeze image.shape: {image.shape}")
-        return self.encode_tensor(image, fastvideo_args, generator)
-
-    def encode_tensor(self,
-                      video_condition: torch.Tensor,
-                      fastvideo_args: FastVideoArgs,
-                      generator: torch.Generator | None = None) -> torch.Tensor:
-        """
-        Encode frames into latent space.
-        """
-        self.vae = self.vae.to(get_local_torch_device())
-        video_condition = video_condition.to(device=get_local_torch_device(),
-                                             dtype=torch.float32)
-
-        # Setup VAE precision
-        vae_dtype = PRECISION_TO_TYPE[
-            fastvideo_args.pipeline_config.vae_precision]
-        vae_autocast_enabled = (
-            vae_dtype != torch.float32) and not fastvideo_args.disable_autocast
-
-        # Encode Image
-        with torch.autocast(device_type="cuda",
-                            dtype=vae_dtype,
-                            enabled=vae_autocast_enabled):
-            if fastvideo_args.pipeline_config.vae_tiling:
-                self.vae.enable_tiling()
-            # if fastvideo_args.vae_sp:
-            #     self.vae.enable_parallel()
-            if not vae_autocast_enabled:
-                video_condition = video_condition.to(vae_dtype)
-            encoder_output = self.vae.encode(video_condition)
-
-        if fastvideo_args.mode == ExecutionMode.PREPROCESS:
-            latent_condition = encoder_output.mean
-        else:
-            generator = generator
-            if generator is None:
-                raise ValueError("Generator must be provided")
-            latent_condition = self.retrieve_latents(encoder_output, generator)
-
-        # Apply shifting if needed
-        if (hasattr(self.vae, "shift_factor")
-                and self.vae.shift_factor is not None):
-            if isinstance(self.vae.shift_factor, torch.Tensor):
-                latent_condition -= self.vae.shift_factor.to(
-                    latent_condition.device, latent_condition.dtype)
-            else:
-                latent_condition -= self.vae.shift_factor
-
-        if isinstance(self.vae.scaling_factor, torch.Tensor):
-            latent_condition = latent_condition * self.vae.scaling_factor.to(
-                latent_condition.device, latent_condition.dtype)
-        else:
-            latent_condition = latent_condition * self.vae.scaling_factor
-
-        return latent_condition
-
    def forward(
        self,
        batch: ForwardBatch,
@@ -232,28 +157,57 @@ class ImageVAEEncodingStage(PipelineStage):
        # (B, C, H, W) -> (B, C, 1, H, W)
        image = image.unsqueeze(2)

-        if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            # repeat the image self.vae.temporal_compression_ratio times
-            video_condition = image.repeat(1, 1,
-                                           self.vae.temporal_compression_ratio,
-                                           1, 1)
-            # video_condition = image
-            logger.info("video_condition.shape: %s", video_condition.shape)
-        else:
-            video_condition = torch.cat([
-                image,
-                image.new_zeros(image.shape[0], image.shape[1], num_frames - 1,
-                                image.shape[3], image.shape[4])
-            ],
-                                        dim=2)
+        video_condition = torch.cat([
+            image,
+            image.new_zeros(image.shape[0], image.shape[1], num_frames - 1,
+                            image.shape[3], image.shape[4])
+        ],
+                                    dim=2)
+        video_condition = video_condition.to(device=get_local_torch_device(),
+                                             dtype=torch.float32)

-        latent_condition = self.encode_tensor(video_condition, fastvideo_args,
-                                              batch.generator)
+        # Setup VAE precision
+        vae_dtype = PRECISION_TO_TYPE[
+            fastvideo_args.pipeline_config.vae_precision]
+        vae_autocast_enabled = (
+            vae_dtype != torch.float32) and not fastvideo_args.disable_autocast
+
+        # Encode Image
+        with torch.autocast(device_type="cuda",
+                            dtype=vae_dtype,
+                            enabled=vae_autocast_enabled):
+            if fastvideo_args.pipeline_config.vae_tiling:
+                self.vae.enable_tiling()
+            # if fastvideo_args.vae_sp:
+            #     self.vae.enable_parallel()
+            if not vae_autocast_enabled:
+                video_condition = video_condition.to(vae_dtype)
+            encoder_output = self.vae.encode(video_condition)
+
+        if fastvideo_args.mode == ExecutionMode.PREPROCESS:
+            latent_condition = encoder_output.mean
+        else:
+            generator = batch.generator
+            if generator is None:
+                raise ValueError("Generator must be provided")
+            latent_condition = self.retrieve_latents(encoder_output, generator)
+
+        # Apply shifting if needed
+        if (hasattr(self.vae, "shift_factor")
+                and self.vae.shift_factor is not None):
+            if isinstance(self.vae.shift_factor, torch.Tensor):
+                latent_condition -= self.vae.shift_factor.to(
+                    latent_condition.device, latent_condition.dtype)
+            else:
+                latent_condition -= self.vae.shift_factor
+
+        if isinstance(self.vae.scaling_factor, torch.Tensor):
+            latent_condition = latent_condition * self.vae.scaling_factor.to(
+                latent_condition.device, latent_condition.dtype)
+        else:
+            latent_condition = latent_condition * self.vae.scaling_factor

        if fastvideo_args.mode == ExecutionMode.PREPROCESS:
-            batch.image_latent = latent_condition
-        elif fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            logger.info("latent_condition.shape: %s", latent_condition.shape)
            batch.image_latent = latent_condition
        else:
            mask_lat_size = torch.ones(1, 1, num_frames, latent_height,
@@ -35,15 +35,9 @@ class InputValidationStage(PipelineStage):
        """Generate seeds for the inference"""
        seed = batch.seed
        num_videos_per_prompt = batch.num_videos_per_prompt
-        if isinstance(batch.prompt, list):
-            num_prompts = len(batch.prompt)
-        else:
-            num_prompts = 1
-
-        total_num_videos = num_prompts * num_videos_per_prompt

        assert seed is not None
-        seeds = [seed + i for i in range(total_num_videos)]
+        seeds = [seed + i for i in range(num_videos_per_prompt)]
        batch.seeds = seeds
        # Peiyuan: using GPU seed will cause A100 and H100 to generate different results...
        batch.generator = [
@@ -82,8 +82,8 @@ def rocm_platform_plugin() -> str | None:
                logger.info("ROCm platform is available")
        finally:
            amdsmi.amdsmi_shut_down()
-    except Exception:
-        pass
+    except Exception as e:
+        logger.info("ROCm platform is unavailable: %s", e)

    return "fastvideo.platforms.rocm.RocmPlatform" if is_rocm else None

@@ -115,6 +115,7 @@ class CudaPlatformBase(Platform):

        logger.info("Trying FASTVIDEO_ATTENTION_BACKEND=%s",
                    envs.FASTVIDEO_ATTENTION_BACKEND)
+        logger.info("Selected backend: %s", selected_backend)
        if selected_backend == AttentionBackendEnum.SLIDING_TILE_ATTN:
            try:
                from st_attn import sliding_tile_attention  # noqa: F401
@@ -144,6 +145,20 @@ class CudaPlatformBase(Platform):
                logger.info(
                    "Sage Attention backend is not installed. Fall back to Flash Attention."
                )
+        elif selected_backend == AttentionBackendEnum.SAGE_ATTN_THREE:
+            try:
+                from fastvideo.attention.backends.sageattn.api import sageattn_blackwell  # noqa: F401
+
+                from fastvideo.attention.backends.sage_attn3 import (  # noqa: F401
+                    SageAttention3Backend)
+                logger.info("Using Sage Attention 3 backend.")
+
+                return "fastvideo.attention.backends.sage_attn3.SageAttention3Backend"
+            except ImportError as e:
+                logger.info(e)
+                logger.info(
+                    "Sage Attention 3 backend is not installed. Fall back to Flash Attention."
+                )
        elif selected_backend == AttentionBackendEnum.VIDEO_SPARSE_ATTN:
            try:
                from vsa import block_sparse_attn  # noqa: F401
@@ -18,6 +18,7 @@ class AttentionBackendEnum(enum.Enum):
    SLIDING_TILE_ATTN = enum.auto()
    TORCH_SDPA = enum.auto()
    SAGE_ATTN = enum.auto()
+    SAGE_ATTN_THREE = enum.auto()
    VIDEO_SPARSE_ATTN = enum.auto()
    VMOBA_ATTN = enum.auto()
    NO_ATTENTION = enum.auto()
@@ -1 +0,0 @@
-{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":0.50390625,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.19960195198655128,"_runtime":107.325113071}
@@ -1 +0,0 @@
-{"train_loss":0.10545400530099869,"_step":5,"_wandb":{"runtime":35},"step_time":2.2575189135968685,"grad_norm":0.53125,"_runtime":35.701958502,"avg_step_time":2.4754387199878694,"_timestamp":1.7525528420185745e+09,"learning_rate":1e-06,"vsa_sparsity":0,"validation_videos_8_steps":{"videos":[{"caption":"A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","_type":"video-file","sha256":"ee83e83df073a648f89dcd288cccaed9af765e11fc35c77a6e0cb2ebaa1be5b0","size":475248,"path":"media/videos/validation_videos_8_steps_0_ee83e83df073a648f89d.mp4"},{"size":341490,"path":"media/videos/validation_videos_8_steps_0_d0b2758549d5c82845ca.mp4","caption":"A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","_type":"video-file","sha256":"d0b2758549d5c82845ca3c1ac0db6812877631495e77ac041acdc8ea31f0a5ee"},{"caption":"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.","_type":"video-file","sha256":"08638381f1607d6ab10684772be38a58b18628fd402a208b25f6af96e765d454","size":436814,"path":"media/videos/validation_videos_8_steps_0_08638381f1607d6ab106.mp4"}],"captions":["A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","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."],"_type":"videos","count":3}}
@@ -1,292 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import os
-import math
-
-import numpy as np
-import pytest
-import torch
-from fastvideo.wan.modules.causal_model import CausalWanModel
-
-from fastvideo.configs.pipelines import PipelineConfig
-from fastvideo.forward_context import set_forward_context
-from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.logger import init_logger
-from fastvideo.models.loader.component_loader import TransformerLoader
-from fastvideo.models.dits.causal_wanvideo import CausalWanTransformer3DModel
-from fastvideo.utils import maybe_download_model
-from fastvideo.configs.models.dits import WanVideoConfig
-from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
-
-
-logger = init_logger(__name__)
-
-os.environ["MASTER_ADDR"] = "localhost"
-os.environ["MASTER_PORT"] = "29503"
-
-BASE_MODEL_PATH = "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
-                                  local_dir=os.path.join(
-                                      'data', BASE_MODEL_PATH))
-TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
-
-
-@pytest.mark.usefixtures("distributed_setup")
-def test_ori_causal_wan_transformer():
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-    precision = torch.bfloat16
-    precision_str = "bf16"
-    args = FastVideoArgs(model_path=TRANSFORMER_PATH,
-                         dit_cpu_offload=True,
-                         pipeline_config=PipelineConfig(dit_config=WanVideoConfig(), dit_precision=precision_str))
-    args.device = device
-
-    loader = TransformerLoader()
-    model2 = loader.load(TRANSFORMER_PATH, args).to(dtype=precision)
-
-    model1 = CausalWanModel.from_pretrained(
-        "/mnt/weka/home/hao.zhang/wei/Self-Forcing/wan_models/Wan2.1-T2V-1.3B", device=device,
-        torch_dtype=precision).to(device, dtype=precision).requires_grad_(False)
-    causal_state_dict = torch.load("/mnt/weka/home/hao.zhang/wei/Self-Forcing/checkpoints/self_forcing_dmd.pt")["generator_ema"]
-    new_state_dict = {}
-    for k, v in causal_state_dict.items():
-        if k.startswith("model."):
-            new_state_dict[k.replace("model.", "")] = v
-    causal_state_dict = new_state_dict
-    model1.load_state_dict(causal_state_dict)
-
-    total_params = sum(p.numel() for p in model1.parameters())
-    # Calculate weight sum for model1 (converting to float64 to avoid overflow)
-    weight_sum_model1 = sum(
-        p.to(torch.float64).sum().item() for p in model1.parameters())
-    # Also calculate mean for more stable comparison
-    weight_mean_model1 = weight_sum_model1 / total_params
-    logger.info("Model 1 weight sum: %s", weight_sum_model1)
-    logger.info("Model 1 weight mean: %s", weight_mean_model1)
-
-    # Calculate weight sum for model2 (converting to float64 to avoid overflow)
-    total_params_model2 = sum(p.numel() for p in model2.parameters())
-    weight_sum_model2 = sum(
-        p.to(torch.float64).sum().item() for p in model2.parameters())
-    # Also calculate mean for more stable comparison
-    weight_mean_model2 = weight_sum_model2 / total_params_model2
-    logger.info("Model 2 weight sum: %s", weight_sum_model2)
-    logger.info("Model 2 weight mean: %s", weight_mean_model2)
-
-    weight_sum_diff = abs(weight_sum_model1 - weight_sum_model2)
-    logger.info("Weight sum difference: %s", weight_sum_diff)
-    weight_mean_diff = abs(weight_mean_model1 - weight_mean_model2)
-    logger.info("Weight mean difference: %s", weight_mean_diff)
-
-    # Set both models to eval mode
-    model1 = model1.eval()
-    model2 = model2.eval()
-
-    # Create identical inputs for both models
-    batch_size = 1
-    text_seq_len = 30
-
-    # Video latents [B, C, T, H, W]
-    hidden_states = torch.randn(batch_size,
-                                16,
-                                12,
-                                160,
-                                90,
-                                device=device,
-                                dtype=precision)
-    block_sizes = [3 for _ in range(4)]
-    timesteps = [1000, 750, 500, 250]
-
-    # Text embeddings [B, L, D] (including global token)
-    encoder_hidden_states = torch.randn(batch_size,
-                                        text_seq_len + 1,
-                                        4096,
-                                        device=device,
-                                        dtype=precision)
-
-    output1 = _causal_inference(model1, hidden_states.clone(), encoder_hidden_states.clone(), block_sizes, timesteps, precision)
-    logger.info("Finish inference for model1")
-    output2 = _causal_inference(model2, hidden_states.clone(), encoder_hidden_states.clone(), block_sizes, timesteps, precision)
-
-    # Check if outputs have the same shape
-    assert output1.shape == output2.shape, f"Output shapes don't match: {output1.shape} vs {output2.shape}"
-    assert output1.dtype == output2.dtype, f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
-
-    logger.info("Output 1 Sum: %s", output1.float().sum().item())
-    logger.info("Output 2 Sum: %s", output2.float().sum().item())
-
-    # Check if outputs are similar (allowing for small numerical differences)
-    max_diff = torch.max(torch.abs(output1 - output2))
-    mean_diff = torch.mean(torch.abs(output1 - output2))
-    logger.info("Max Diff: %s", max_diff.item())
-    logger.info("Mean Diff: %s", mean_diff.item())
-    assert max_diff < 1e-4, f"Maximum difference between outputs: {max_diff.item()}"
-    # mean diff
-    assert mean_diff < 1e-4, f"Mean difference between outputs: {mean_diff.item()}"
-
-def _causal_inference(transformer, latents, prompt_embeds, block_sizes, timesteps, target_dtype):
-    forward_batch = ForwardBatch(
-        data_type="dummy",
-    )
-    start_index = 0
-    pos_start_base = 0
-    frame_seq_length = latents.shape[-1] * latents.shape[-2] // (WanVideoConfig().arch_config.patch_size[-1] * WanVideoConfig().arch_config.patch_size[-2])
-    seq_len = frame_seq_length * latents.shape[2]
-    kv_cache1 = _initialize_kv_cache(transformer, batch_size=latents.shape[0],
-                                      kv_cache_size=frame_seq_length * latents.shape[2],
-                                      dtype=target_dtype,
-                                      device=latents.device)
-    crossattn_cache = _initialize_crossattn_cache(
-                transformer,
-                batch_size=latents.shape[0],
-                max_text_len=WanVideoConfig().arch_config.text_len,
-                dtype=target_dtype,
-                device=latents.device)
-    for current_num_frames, t_cur in zip(block_sizes, timesteps):
-        # logger.info(f"Current frame idx: {start_index}, Current timestep: {t_cur}")
-        # logger.info(f"k cache sum: {sum(kv_cache['k'].float().sum().item() for kv_cache in kv_cache1)}, v cache sum: {sum(kv_cache['v'].float().sum().item() for kv_cache in kv_cache1)}")
-        # logger.info(f"latents sum: {latents.float().sum().item()}, encoder_hidden_states sum: {prompt_embeds.float().sum().item()}")
-        current_latents = latents[:, :, start_index:start_index +
-                                    current_num_frames, :, :]
-
-        attn_metadata = None
-
-        with set_forward_context(current_timestep=0,
-                                attn_metadata=attn_metadata,
-                                forward_batch=forward_batch):
-            # Run transformer; follow DMD stage pattern
-            t_expanded_noise = t_cur * torch.ones(
-                (current_latents.shape[0], 1),
-                device=current_latents.device,
-                dtype=torch.long)
-            if isinstance(transformer, CausalWanModel):
-                pred_noise_btchw = transformer(
-                    x=current_latents,
-                    context=prompt_embeds,
-                    t=t_expanded_noise,
-                    seq_len=seq_len,
-                    kv_cache=kv_cache1,
-                    crossattn_cache=crossattn_cache,
-                    current_start=(pos_start_base + start_index) *
-                    frame_seq_length
-                )
-            elif isinstance(transformer, CausalWanTransformer3DModel):
-                pred_noise_btchw = transformer(
-                    current_latents,
-                    prompt_embeds,
-                    t_expanded_noise,
-                    kv_cache=kv_cache1,
-                    crossattn_cache=crossattn_cache,
-                    current_start=(pos_start_base + start_index) *
-                    frame_seq_length,
-                    start_frame=start_index
-                )
-
-        # Write back and advance
-        latents[:, :, start_index:start_index +
-                current_num_frames, :, :] = pred_noise_btchw.clone()
-
-        # Re-run with context timestep to update KV cache using clean context
-        context_noise = 0
-        t_context = torch.ones([latents.shape[0]],
-                                device=latents.device,
-                                dtype=torch.long) * int(context_noise)
-        context_bcthw = pred_noise_btchw.to(target_dtype)
-        with set_forward_context(current_timestep=0,
-                                attn_metadata=attn_metadata,
-                                forward_batch=forward_batch):
-            t_expanded_context = t_context.unsqueeze(1)
-            if isinstance(transformer, CausalWanModel):
-                _ = transformer(
-                    x=context_bcthw,
-                    context=prompt_embeds,
-                    t=t_expanded_context,
-                    seq_len=seq_len,
-                    kv_cache=kv_cache1,
-                    crossattn_cache=crossattn_cache,
-                    current_start=(pos_start_base + start_index) *
-                    frame_seq_length
-                )
-            elif isinstance(transformer, CausalWanTransformer3DModel):
-                _ = transformer(
-                    context_bcthw,
-                    prompt_embeds,
-                    t_expanded_context,
-                    kv_cache=kv_cache1,
-                    crossattn_cache=crossattn_cache,
-                    current_start=(pos_start_base + start_index) *
-                    frame_seq_length,
-                    start_frame=start_index
-                )
-        start_index += current_num_frames
-
-    return latents
-
-def _initialize_kv_cache(transformer, batch_size, kv_cache_size, dtype, device) -> None:
-    """
-    Initialize a Per-GPU KV cache aligned with the Wan model assumptions.
-    """
-    kv_cache1 = []
-    if isinstance(transformer, CausalWanModel):
-        num_attention_heads = transformer.num_heads
-        attention_head_dim = transformer.dim // transformer.num_heads
-    elif isinstance(transformer, CausalWanTransformer3DModel):
-        num_attention_heads = transformer.num_attention_heads
-        attention_head_dim = transformer.attention_head_dim
-
-    for _ in range(len(transformer.blocks)):
-        kv_cache1.append({
-            "k":
-            torch.zeros([
-                batch_size, kv_cache_size, num_attention_heads,
-                attention_head_dim
-            ],
-                        dtype=dtype,
-                        device=device),
-            "v":
-            torch.zeros([
-                batch_size, kv_cache_size, num_attention_heads,
-                attention_head_dim
-            ],
-                        dtype=dtype,
-                        device=device),
-            "global_end_index":
-            torch.tensor([0], dtype=torch.long, device=device),
-            "local_end_index":
-            torch.tensor([0], dtype=torch.long, device=device),
-        })
-
-    return kv_cache1
-
-def _initialize_crossattn_cache(transformer, batch_size, max_text_len, dtype,
-                                    device) -> None:
-    """
-    Initialize a Per-GPU cross-attention cache aligned with the Wan model assumptions.
-    """
-    crossattn_cache = []
-    if isinstance(transformer, CausalWanModel):
-        num_attention_heads = transformer.num_heads
-        attention_head_dim = transformer.dim // transformer.num_heads
-    elif isinstance(transformer, CausalWanTransformer3DModel):
-        num_attention_heads = transformer.num_attention_heads
-        attention_head_dim = transformer.attention_head_dim
-
-    for _ in range(len(transformer.blocks)):
-        crossattn_cache.append({
-            "k":
-            torch.zeros([
-                batch_size, max_text_len, num_attention_heads,
-                attention_head_dim
-            ],
-                        dtype=dtype,
-                        device=device),
-            "v":
-            torch.zeros([
-                batch_size, max_text_len, num_attention_heads,
-                attention_head_dim
-            ],
-                        dtype=dtype,
-                        device=device),
-            "is_init":
-            False,
-        })
-    return crossattn_cache
@@ -1,133 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import os
-import math
-
-import numpy as np
-import pytest
-import torch
-from fastvideo.wan.modules.model import WanModel
-
-from fastvideo.configs.pipelines import PipelineConfig
-from fastvideo.forward_context import set_forward_context
-from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.logger import init_logger
-from fastvideo.models.loader.component_loader import TransformerLoader
-from fastvideo.utils import maybe_download_model
-from fastvideo.configs.models.dits import WanVideoConfig
-from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
-
-
-logger = init_logger(__name__)
-
-os.environ["MASTER_ADDR"] = "localhost"
-os.environ["MASTER_PORT"] = "29503"
-
-BASE_MODEL_PATH = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
-                                  local_dir=os.path.join(
-                                      'data', BASE_MODEL_PATH))
-TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
-
-
-@pytest.mark.usefixtures("distributed_setup")
-def test_ori_wan_transformer():
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-    precision = torch.bfloat16
-    precision_str = "bf16"
-    args = FastVideoArgs(model_path=TRANSFORMER_PATH,
-                         dit_cpu_offload=True,
-                         pipeline_config=PipelineConfig(dit_config=WanVideoConfig(), dit_precision=precision_str))
-    args.device = device
-
-    loader = TransformerLoader()
-    model2 = loader.load(TRANSFORMER_PATH, args).to(dtype=precision)
-
-    model1 = WanModel.from_pretrained(
-        "/mnt/weka/home/hao.zhang/wei/Self-Forcing/wan_models/Wan2.1-T2V-1.3B", device=device,
-        torch_dtype=precision).to(device, dtype=precision).requires_grad_(False)
-
-    total_params = sum(p.numel() for p in model1.parameters())
-    # Calculate weight sum for model1 (converting to float64 to avoid overflow)
-    weight_sum_model1 = sum(
-        p.to(torch.float64).sum().item() for p in model1.parameters())
-    # Also calculate mean for more stable comparison
-    weight_mean_model1 = weight_sum_model1 / total_params
-    logger.info("Model 1 weight sum: %s", weight_sum_model1)
-    logger.info("Model 1 weight mean: %s", weight_mean_model1)
-
-    # Calculate weight sum for model2 (converting to float64 to avoid overflow)
-    total_params_model2 = sum(p.numel() for p in model2.parameters())
-    weight_sum_model2 = sum(
-        p.to(torch.float64).sum().item() for p in model2.parameters())
-    # Also calculate mean for more stable comparison
-    weight_mean_model2 = weight_sum_model2 / total_params_model2
-    logger.info("Model 2 weight sum: %s", weight_sum_model2)
-    logger.info("Model 2 weight mean: %s", weight_mean_model2)
-
-    weight_sum_diff = abs(weight_sum_model1 - weight_sum_model2)
-    logger.info("Weight sum difference: %s", weight_sum_diff)
-    weight_mean_diff = abs(weight_mean_model1 - weight_mean_model2)
-    logger.info("Weight mean difference: %s", weight_mean_diff)
-
-    # Set both models to eval mode
-    model1 = model1.eval()
-    model2 = model2.eval()
-
-    # Create identical inputs for both models
-    batch_size = 1
-    text_seq_len = 30
-    seq_len = math.ceil((160 * 90) /
-                            (2 * 2) *
-                            21)
-
-    # Video latents [B, C, T, H, W]
-    hidden_states = torch.randn(batch_size,
-                                16,
-                                21,
-                                160,
-                                90,
-                                device=device,
-                                dtype=precision)
-
-    # Text embeddings [B, L, D] (including global token)
-    encoder_hidden_states = torch.randn(batch_size,
-                                        text_seq_len + 1,
-                                        4096,
-                                        device=device,
-                                        dtype=precision)
-
-    # Timestep
-    timestep = torch.tensor([500], device=device, dtype=precision)
-
-    forward_batch = ForwardBatch(
-        data_type="dummy",
-    )
-
-    # with torch.amp.autocast('cuda', dtype=precision):
-    output1 = model1(
-        x=hidden_states,
-        context=encoder_hidden_states,
-        t=timestep,
-        seq_len=seq_len,
-    )
-    with set_forward_context(
-            current_timestep=0,
-            attn_metadata=None,
-            forward_batch=forward_batch,
-    ):
-        output2 = model2(hidden_states=hidden_states,
-                            encoder_hidden_states=encoder_hidden_states,
-                            timestep=timestep)
-
-    # Check if outputs have the same shape
-    assert output1.shape == output2.shape, f"Output shapes don't match: {output1.shape} vs {output2.shape}"
-    assert output1.dtype == output2.dtype, f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
-
-    # Check if outputs are similar (allowing for small numerical differences)
-    max_diff = torch.max(torch.abs(output1 - output2))
-    mean_diff = torch.mean(torch.abs(output1 - output2))
-    logger.info("Max Diff: %s", max_diff.item())
-    logger.info("Mean Diff: %s", mean_diff.item())
-    assert max_diff < 1e-4, f"Maximum difference between outputs: {max_diff.item()}"
-    # mean diff
-    assert mean_diff < 1e-4, f"Mean difference between outputs: {mean_diff.item()}"
@@ -1,144 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import os
-import math
-
-import numpy as np
-import pytest
-import torch
-from fastvideo.wan.modules.causal_model import CausalWanModel
-
-from fastvideo.configs.pipelines import PipelineConfig
-from fastvideo.forward_context import set_forward_context
-from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.logger import init_logger
-from fastvideo.models.loader.component_loader import TransformerLoader
-from fastvideo.utils import maybe_download_model
-from fastvideo.configs.models.dits import WanVideoConfig
-from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
-
-
-logger = init_logger(__name__)
-
-os.environ["MASTER_ADDR"] = "localhost"
-os.environ["MASTER_PORT"] = "29503"
-
-BASE_MODEL_PATH = "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
-                                  local_dir=os.path.join(
-                                      'data', BASE_MODEL_PATH))
-TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
-
-
-@pytest.mark.usefixtures("distributed_setup")
-def test_train_ori_causal_wan_transformer():
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-    precision = torch.bfloat16
-    precision_str = "bf16"
-    args = FastVideoArgs(model_path=TRANSFORMER_PATH,
-                         dit_cpu_offload=True,
-                         pipeline_config=PipelineConfig(dit_config=WanVideoConfig(), dit_precision=precision_str))
-    args.device = device
-
-    loader = TransformerLoader()
-    model2 = loader.load(TRANSFORMER_PATH, args).to(dtype=precision)
-
-    model1 = CausalWanModel.from_pretrained(
-        "/mnt/weka/home/hao.zhang/wei/Self-Forcing/wan_models/Wan2.1-T2V-1.3B", device=device,
-        torch_dtype=precision).to(device, dtype=precision).requires_grad_(False)
-    causal_state_dict = torch.load("/mnt/weka/home/hao.zhang/wei/Self-Forcing/checkpoints/self_forcing_dmd.pt")["generator_ema"]
-    new_state_dict = {}
-    for k, v in causal_state_dict.items():
-        if k.startswith("model."):
-            new_state_dict[k.replace("model.", "")] = v
-    causal_state_dict = new_state_dict
-    model1.load_state_dict(causal_state_dict)
-
-    model1.num_frame_per_block = 3
-    model2.num_frame_per_block = 3
-
-    total_params = sum(p.numel() for p in model1.parameters())
-    # Calculate weight sum for model1 (converting to float64 to avoid overflow)
-    weight_sum_model1 = sum(
-        p.to(torch.float64).sum().item() for p in model1.parameters())
-    # Also calculate mean for more stable comparison
-    weight_mean_model1 = weight_sum_model1 / total_params
-    logger.info("Model 1 weight sum: %s", weight_sum_model1)
-    logger.info("Model 1 weight mean: %s", weight_mean_model1)
-
-    # Calculate weight sum for model2 (converting to float64 to avoid overflow)
-    total_params_model2 = sum(p.numel() for p in model2.parameters())
-    weight_sum_model2 = sum(
-        p.to(torch.float64).sum().item() for p in model2.parameters())
-    # Also calculate mean for more stable comparison
-    weight_mean_model2 = weight_sum_model2 / total_params_model2
-    logger.info("Model 2 weight sum: %s", weight_sum_model2)
-    logger.info("Model 2 weight mean: %s", weight_mean_model2)
-
-    weight_sum_diff = abs(weight_sum_model1 - weight_sum_model2)
-    logger.info("Weight sum difference: %s", weight_sum_diff)
-    weight_mean_diff = abs(weight_mean_model1 - weight_mean_model2)
-    logger.info("Weight mean difference: %s", weight_mean_diff)
-
-    # Set both models to eval mode
-    model1 = model1.eval()
-    model2 = model2.eval()
-
-    # Create identical inputs for both models
-    batch_size = 1
-    text_seq_len = 30
-    seq_len = math.ceil((160 * 90) /
-                            (2 * 2) *
-                            21)
-
-    # Video latents [B, C, T, H, W]
-    hidden_states = torch.randn(batch_size,
-                                16,
-                                21,
-                                160,
-                                90,
-                                device=device,
-                                dtype=precision)
-
-    # Text embeddings [B, L, D] (including global token)
-    encoder_hidden_states = torch.randn(batch_size,
-                                        text_seq_len + 1,
-                                        4096,
-                                        device=device,
-                                        dtype=precision)
-
-    # Timestep
-    timestep = torch.randint(0, 1000, (batch_size, 21), device=device, dtype=torch.long)
-    logger.info("timestep: %s", timestep)
-
-    forward_batch = ForwardBatch(
-        data_type="dummy",
-    )
-
-    # with torch.amp.autocast('cuda', dtype=precision):
-    output1 = model1(
-        x=hidden_states,
-        context=encoder_hidden_states,
-        t=timestep,
-        seq_len=seq_len,
-    )
-    with set_forward_context(
-            current_timestep=0,
-            attn_metadata=None,
-            forward_batch=forward_batch,
-    ):
-        output2 = model2(hidden_states=hidden_states,
-                            encoder_hidden_states=encoder_hidden_states,
-                            timestep=timestep)
-
-    # Check if outputs have the same shape
-    assert output1.shape == output2.shape, f"Output shapes don't match: {output1.shape} vs {output2.shape}"
-    assert output1.dtype == output2.dtype, f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
-
-    # Check if outputs are similar (allowing for small numerical differences)
-    max_diff = torch.max(torch.abs(output1 - output2))
-    mean_diff = torch.mean(torch.abs(output1 - output2))
-    logger.info("Max Diff: %s", max_diff.item())
-    logger.info("Mean Diff: %s", mean_diff.item())
-    assert max_diff < 1e-4, f"Maximum difference between outputs: {max_diff.item()}"
-    # mean diff
-    assert mean_diff < 1e-4, f"Mean difference between outputs: {mean_diff.item()}"
@@ -12,7 +12,6 @@ from typing import Any
 import imageio
 import numpy as np
 import torch
-import torch.distributed as dist
 import torch.nn.functional as F
 import torchvision
 from einops import rearrange
@@ -92,9 +91,8 @@ class DistillationPipeline(TrainingPipeline):
            shift=self.timestep_shift)

        if training_args.real_score_model_path:
-            logger.info(
-                f"Loading real score transformer from: {training_args.real_score_model_path}"
-            )
+            logger.info("Loading real score transformer from: %s",
+                        training_args.real_score_model_path)
            self.real_score_transformer = self.load_module_from_path(
                training_args.real_score_model_path, "transformer",
                training_args)
@@ -103,9 +101,8 @@ class DistillationPipeline(TrainingPipeline):
                "real_score_transformer")

        if training_args.fake_score_model_path:
-            logger.info(
-                f"Loading fake score transformer from: {training_args.fake_score_model_path}"
-            )
+            logger.info("Loading fake score transformer from: %s",
+                        training_args.fake_score_model_path)
            self.fake_score_transformer = self.load_module_from_path(
                training_args.fake_score_model_path, "transformer",
                training_args)
@@ -194,14 +191,13 @@ class DistillationPipeline(TrainingPipeline):

        self.real_score_guidance_scale = self.training_args.real_score_guidance_scale

-        self.generator_ema = None
+        self.generator_ema: EMA_FSDP | None = None
        if (self.training_args.ema_decay
                is not None) and (self.training_args.ema_decay > 0.0):
            self.generator_ema = EMA_FSDP(self.transformer,
                                          decay=self.training_args.ema_decay)
-            logger.info(
-                f"Initialized generator EMA with decay={self.training_args.ema_decay}"
-            )
+            logger.info("Initialized generator EMA with decay=%s",
+                        self.training_args.ema_decay)
        else:
            logger.info("Generator EMA disabled (ema_decay <= 0.0)")

@@ -218,7 +214,7 @@ class DistillationPipeline(TrainingPipeline):
        Returns:
            The loaded module
        """
-        logger.info(f"Loading {module_type} from custom path: {model_path}")
+        logger.info("Loading %s from custom path: %s", module_type, model_path)
        # Set flag to prevent custom weight loading for teacher/critic models
        training_args._loading_teacher_critic_model = True

@@ -228,7 +224,7 @@ class DistillationPipeline(TrainingPipeline):

            # Download the model if it's a Hugging Face model ID
            local_model_path = maybe_download_model(model_path)
-            logger.info(f"Model downloaded/found at: {local_model_path}")
+            logger.info("Model downloaded/found at: %s", local_model_path)
            config = verify_model_config_and_directory(local_model_path)

            if module_type not in config:
@@ -237,7 +233,8 @@ class DistillationPipeline(TrainingPipeline):
                    extra_module = self._extra_config_module_map[module_type]
                    if extra_module in config:
                        module_type = extra_module
-                        logger.info(f"Using {extra_module} for {module_type}")
+                        logger.info("Using %s for %s", extra_module,
+                                    module_type)
                    else:
                        raise ValueError(
                            f"Module {module_type} not found in config at {local_model_path}"
@@ -262,8 +259,8 @@ class DistillationPipeline(TrainingPipeline):
                fastvideo_args=training_args,
            )

-            logger.info(
-                f"Successfully loaded {module_type} from {component_path}")
+            logger.info("Successfully loaded %s from %s", module_type,
+                        component_path)
            return module
        finally:
            # Always clean up the flag
@@ -293,7 +290,7 @@ class DistillationPipeline(TrainingPipeline):
                return model
        return model

-    def get_ema_model_copy(self):
+    def get_ema_model_copy(self) -> torch.nn.Module | None:
        """Get a copy of the model with EMA weights applied."""
        if self.generator_ema is not None:
            ema_model = copy.deepcopy(self.transformer)
@@ -301,7 +298,7 @@ class DistillationPipeline(TrainingPipeline):
            return ema_model
        return None

-    def is_ema_ready(self, current_step: int = None):
+    def is_ema_ready(self, current_step: int | None = None):
        """Check if EMA is ready for use (after ema_start_step)."""
        if current_step is None:
            current_step = getattr(self, 'current_trainstep', 0)
@@ -350,14 +347,14 @@ class DistillationPipeline(TrainingPipeline):
                with open(config_path, "w") as f:
                    json.dump(config_dict, f, indent=4)

-                logger.info(f"EMA weights saved to {weight_path}")
+                logger.info("EMA weights saved to %s", weight_path)

            del ema_model

        except Exception as e:
-            logger.error(f"Failed to save EMA weights: {str(e)}")
+            logger.error("Failed to save EMA weights: %s", str(e))

-    def get_ema_stats(self):
+    def get_ema_stats(self) -> dict[str, Any]:
        """Get EMA statistics for monitoring."""
        if self.generator_ema is None:
            return {
@@ -571,7 +568,8 @@ class DistillationPipeline(TrainingPipeline):

            noisy_latent = self.noise_scheduler.add_noise(
                generator_pred_video.flatten(0, 1), noise.flatten(0, 1),
-                timestep).detach().unflatten(0, (1, generator_pred_video.shape[1]))
+                timestep).detach().unflatten(0,
+                                             (1, generator_pred_video.shape[1]))

            # fake_score_transformer forward
            training_batch = self._build_distill_input_kwargs(
@@ -730,11 +728,12 @@ class DistillationPipeline(TrainingPipeline):
            "encoder_hidden_states": training_batch.encoder_hidden_states,
            "encoder_attention_mask": training_batch.encoder_attention_mask,
        }
-        unconditional_dict = {
-            "encoder_hidden_states": self.negative_prompt_embeds,
-            "encoder_attention_mask": self.negative_prompt_attention_mask,
-        }
-        training_batch.unconditional_dict = unconditional_dict
+        if getattr(self, "negative_prompt_embeds", None) is not None:
+            unconditional_dict = {
+                "encoder_hidden_states": self.negative_prompt_embeds,
+                "encoder_attention_mask": self.negative_prompt_attention_mask,
+            }
+            training_batch.unconditional_dict = unconditional_dict

        training_batch.dmd_latent_vis_dict = {}
        training_batch.fake_score_latent_vis_dict = {}
@@ -1251,9 +1250,8 @@ class DistillationPipeline(TrainingPipeline):
                    (self.generator_ema is None) and (self.training_args.ema_decay > 0):
                self.generator_ema = EMA_FSDP(
                    self.transformer, decay=self.training_args.ema_decay)
-                logger.info(
-                    f"Created generator EMA at step {step} with decay={self.training_args.ema_decay}"
-                )
+                logger.info("Created generator EMA at step %s with decay=%s",
+                            step, self.training_args.ema_decay)

            with torch.autocast("cuda", dtype=torch.bfloat16):
                training_batch = self.train_one_step(training_batch)
@@ -1397,4 +1395,4 @@ class DistillationPipeline(TrainingPipeline):
                                  self.training_args.max_train_steps)

        if get_sp_group():
-            cleanup_dist_env_and_memory()
+            cleanup_dist_env_and_memory()
@@ -3,12 +3,13 @@ import sys
 from copy import deepcopy
 from typing import cast

-
+import numpy as np
 import torch
 import torch.nn.functional as F
-import numpy as np
+
 import wandb
-from fastvideo.dataset.dataloader.schema import pyarrow_schema_ode_trajectory_text_only
+from fastvideo.dataset.dataloader.schema import (
+    pyarrow_schema_ode_trajectory_text_only)
 from fastvideo.distributed import get_local_torch_device
 from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
 from fastvideo.forward_context import set_forward_context
@@ -17,8 +18,8 @@ from fastvideo.models.schedulers.scheduling_self_forcing_flow_match import (
    SelfForcingFlowMatchScheduler)
 from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import (
    WanCausalDMDPipeline)
-from fastvideo.training.training_pipeline import TrainingPipeline
 from fastvideo.pipelines.pipeline_batch_info import TrainingBatch
+from fastvideo.training.training_pipeline import TrainingPipeline
 from fastvideo.training.training_utils import (
    clip_grad_norm_while_handling_failing_dtensor_cases)

@@ -63,31 +64,27 @@ class ODEInitTrainingPipeline(TrainingPipeline):
        self.noise_scheduler.set_timesteps(num_inference_steps=1000,
                                           training=True)

-        # logger.info(f"ARG dmd_denoising_steps: {training_args.pipeline_config.dmd_denoising_steps}")
-        logger.info(
-            f"ARG dmd_denoising_steps: {self.training_args.pipeline_config.dmd_denoising_steps}"
-        )
+        logger.info("dmd_denoising_steps: %s",
+                    self.training_args.pipeline_config.dmd_denoising_steps)
        self.dmd_denoising_steps = torch.tensor([1000, 750, 500, 250],
                                                dtype=torch.long,
                                                device=get_local_torch_device())
-        # self.dmd_denoising_steps = torch.tensor([1000, 750, 500, 250], dtype=torch.long, device=get_local_torch_device())
        if training_args.warp_denoising_step:  # Warp the denoising step according to the scheduler time shift
            timesteps = torch.cat((self.noise_scheduler.timesteps.cpu(),
                                   torch.tensor([0],
                                                dtype=torch.float32))).cuda()
-            logger.info(f"timesteps: {timesteps}")
+            logger.info("timesteps: %s", timesteps)
            self.dmd_denoising_steps = timesteps[1000 -
                                                 self.dmd_denoising_steps]
-            logger.info(
-                f"warped self.dmd_denoising_steps: {self.dmd_denoising_steps}")
-            # assert False, "warp_denoising_step must be false"
+            logger.info("warped self.dmd_denoising_steps: %s",
+                        self.dmd_denoising_steps)
        else:
-            assert False, "warp_denoising_step must be true"
-            logger.info("not warped")
+            raise ValueError("warp_denoising_step must be true")
+
        self.dmd_denoising_steps = self.dmd_denoising_steps.to(
            get_local_torch_device())

-        logger.info(f"denoising_step_list: {self.dmd_denoising_steps}")
+        logger.info("denoising_step_list: %s", self.dmd_denoising_steps)

        logger.info(
            "Initialized ODE-init training pipeline with %s denoising steps",
@@ -95,11 +92,7 @@ class ODEInitTrainingPipeline(TrainingPipeline):
        # Cache for nearest trajectory index per DMD step (computed lazily on first batch)
        self._cached_closest_idx_per_dmd = None
        self.num_train_timestep = self.noise_scheduler.num_train_timesteps
-        # self.min_timestep = int(self.training_args.min_timestep_ratio *
-        #                         self.num_train_timestep)
-        # self.max_timestep = int(self.training_args.max_timestep_ratio *
-        #                         self.num_train_timestep)
-        # self.real_score_guidance_scale = self.training_args.real_score_guidance_scale
+        self.manual_idx = 0

    def initialize_validation_pipeline(self, training_args: TrainingArgs):
        logger.info("Initializing validation pipeline...")
@@ -107,8 +100,7 @@ class ODEInitTrainingPipeline(TrainingPipeline):
        args_copy.inference_mode = True
        # Warm start validation with current transformer
        self.validation_pipeline = WanCausalDMDPipeline.from_pretrained(
-            # training_args.model_path,
-            "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers",
+            training_args.model_path,
            args=args_copy,  # type: ignore
            inference_mode=True,
            loaded_modules={
@@ -120,7 +112,9 @@ class ODEInitTrainingPipeline(TrainingPipeline):
            pin_cpu_memory=training_args.pin_cpu_memory,
            dit_cpu_offload=True)

-    def _get_next_batch(self, training_batch):  # type: ignore[override]
+    def _get_next_batch(
+            self,
+            training_batch) -> tuple[TrainingBatch, torch.Tensor, torch.Tensor]:
        batch = next(self.train_loader_iter, None)  # type: ignore
        if batch is None:
            self.current_epoch += 1
@@ -171,6 +165,25 @@ class ODEInitTrainingPipeline(TrainingPipeline):
        return training_batch, trajectory_latents.to(
            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)

+        ## TEMP Used for loading the sf .pt files directly
+        """
+        self.manual_idx = self.manual_idx % 155
+        path = f"/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_pt_vidprom_1000/{self.manual_idx:05d}.pt"
+        logger.info("path: %s", path)
+        self.manual_idx += 1
+        # path = "/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_single_full/00000.pt"
+        b = torch.load(path)
+        training_batch.encoder_hidden_states = b["text_embedding"][0].unsqueeze(
+            0).to(device, dtype=torch.bfloat16)
+        trajectory_latents = b["ode_latent"].to(device, dtype=torch.bfloat16)
+        logger.info("trajectory_latents: %s", trajectory_latents.shape)
+        logger.info("encoder_hidden_states: %s",
+                    training_batch.encoder_hidden_states.shape)
+        assert trajectory_latents.shape[1] <= 10, "trajectory_latents.shape[1] must be <= 10"
+        return training_batch, trajectory_latents.to(
+            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)
+        """
+
    def _get_timestep(self,
                      min_timestep: int,
                      max_timestep: int,
@@ -201,42 +214,35 @@ class ODEInitTrainingPipeline(TrainingPipeline):
        self, traj_latents: torch.Tensor, traj_timesteps: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        encoder_attention_mask: torch.Tensor
-    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, dict[str, torch.Tensor]]:
-        latent_vis_dict = {}
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, dict[str,
+                                                              torch.Tensor]]:
+        latent_vis_dict: dict[str, torch.Tensor] = {}
        device = get_local_torch_device()
        target_latent = traj_latents[:, -1]

-        # logger.info(f"traj_latents: {traj_latents.shape}")
-        # logger.info(f"traj_timesteps: {traj_timesteps.shape}")
-
        # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
        B, S, num_frames, num_channels, height, width = traj_latents.shape

        # Lazily cache nearest trajectory index per DMD step based on the (fixed) S timesteps
        if self._cached_closest_idx_per_dmd is None:
-            # Use the first sample's trajectory timesteps; assumed identical across batches
-            # s_steps = traj_timesteps[0].to(torch.long)  # [S]
-            # dmd = cast(torch.Tensor, self.dmd_denoising_steps).to(s_steps.device)  # [K]
-            # distances_ks: [K, S] = |s_steps - dmd|
-            # distances_ks = (s_steps.unsqueeze(0) - dmd.unsqueeze(1)).abs()
-            # self._cached_closest_idx_per_dmd = distances_ks.argmin(dim=1).to(torch.long).cpu()  # [K]
            self._cached_closest_idx_per_dmd = torch.tensor(
                [0, 12, 24, 36], dtype=torch.long).cpu()
+            # [0, 1, 2, 3], dtype=torch.long).cpu()
+            logger.info("self._cached_closest_idx_per_dmd: %s",
+                        self._cached_closest_idx_per_dmd)
            logger.info(
-                f"self._cached_closest_idx_per_dmd: {self._cached_closest_idx_per_dmd}"
-            )
-            logger.info(
-                f"corresponding timesteps: {self.noise_scheduler.timesteps[self._cached_closest_idx_per_dmd]}"
-            )
+                "corresponding timesteps: %s", self.noise_scheduler.timesteps[
+                    self._cached_closest_idx_per_dmd])

-        # logger.info(f"traj_latents: {traj_latents.shape}")
        # Select the K indexes from traj_latents using self._cached_closest_idx_per_dmd
        # traj_latents: [B, S, C, T, H, W], self._cached_closest_idx_per_dmd: [K]
        # Output: [B, K, C, T, H, W]
+        assert self._cached_closest_idx_per_dmd is not None
        relevant_traj_latents = torch.index_select(
            traj_latents,
            dim=1,
            index=self._cached_closest_idx_per_dmd.to(traj_latents.device))
+        logger.info("relevant_traj_latents: %s", relevant_traj_latents.shape)
        # assert relevant_traj_latents.shape[0] == 1

        indexes = self._get_timestep(  # [B, num_frames]
@@ -246,8 +252,8 @@ class ODEInitTrainingPipeline(TrainingPipeline):
            num_frames,
            3,
            uniform_timestep=False)
-        logger.info(f"indexes: {indexes.shape}")
-        logger.info(f"indexes: {indexes}")
+        logger.info("indexes: %s", indexes.shape)
+        logger.info("indexes: %s", indexes)
        # noisy_input = relevant_traj_latents[indexes]
        noisy_input = torch.gather(
            relevant_traj_latents,
@@ -255,39 +261,22 @@ class ODEInitTrainingPipeline(TrainingPipeline):
            index=indexes.reshape(B, 1, num_frames, 1, 1,
                                  1).expand(-1, -1, -1, num_channels, height,
                                            width).to(self.device)).squeeze(1)
-        # noisy_input = noisy_input.unsqueeze(0)
-
-        # # Sample a single DMD step for the whole batch and fetch its cached nearest S-index
-        # K = len(self.dmd_denoising_steps)
-        # dmd_idx = torch.randint(0, K, (1,), device=device)
-        # logger.info(f"dmd_idx: {dmd_idx}")
-        # assert self._cached_closest_idx_per_dmd is not None
-        # nearest_s_idx = int(self._cached_closest_idx_per_dmd[int(dmd_idx.item())])
-        # nearest_idx = torch.full((B,), nearest_s_idx, device=device, dtype=torch.long)
-
-        # batch_indices = torch.arange(B, device=device)
-        # noisy_input = traj_latents[batch_indices, nearest_idx]  # [B, C, T, H, W]
-        # target_latent = traj_latents[batch_indices, -1]  # [B, C, T, H, W]
-        # t = traj_timesteps[batch_indices, nearest_idx]  # [B]
-
-        # Scale model input as in inference for consistency with stored trajectories
-        # noisy_input = self.modules["scheduler"].scale_model_input(noisy_input, t)
-        # logger.info(f"indexes: {indexes.shape}")
-        # logger.info(f"indexes: {indexes}")
        timestep = self.dmd_denoising_steps[indexes]
-        # logger.info(f"timestep: {timestep.shape}")
-        # logger.info(f"timestep: {timestep}")
+        logger.info("selected timestep for rank %s: %s",
+                    self.global_rank,
+                    timestep,
+                    local_main_process_only=False)

        # Prepare inputs for transformer
-        latent_vis_dict["noisy_input"] = noisy_input.permute(0, 2, 1, 3, 4).detach().clone().cpu()
-        latent_vis_dict["x0"] = target_latent.permute(0, 2, 1, 3, 4).detach().clone().cpu()
+        latent_vis_dict["noisy_input"] = noisy_input.permute(
+            0, 2, 1, 3, 4).detach().clone().cpu()
+        latent_vis_dict["x0"] = target_latent.permute(0, 2, 1, 3,
+                                                      4).detach().clone().cpu()

-        model_dtype = next(self.transformer.parameters()).dtype
        input_kwargs = {
            "hidden_states": noisy_input.permute(0, 2, 1, 3, 4),
            "encoder_hidden_states": encoder_hidden_states,
-            "timestep": timestep.to(device, dtype=model_dtype),
-            "encoder_attention_mask": encoder_attention_mask,
+            "timestep": timestep.to(device, dtype=torch.bfloat16),
            "return_dict": False,
        }
        # Predict noise and step the scheduler to obtain next latent
@@ -295,22 +284,17 @@ class ODEInitTrainingPipeline(TrainingPipeline):
                                 attn_metadata=None,
                                 forward_batch=None):
            noise_pred = self.transformer(**input_kwargs).permute(0, 2, 1, 3, 4)
-            # logger.info(f"noise_pred: {noise_pred.shape}")
-        if isinstance(noise_pred, (tuple, list)):
-            noise_pred = noise_pred[0]

        from fastvideo.models.utils import pred_noise_to_pred_video
        pred_video = pred_noise_to_pred_video(
            pred_noise=noise_pred.flatten(0, 1),
            noise_input_latent=noisy_input.flatten(0, 1),
-            timestep=timestep.to(dtype=model_dtype).flatten(0, 1),
+            timestep=timestep.to(dtype=torch.bfloat16).flatten(0, 1),
            scheduler=self.modules["scheduler"]).unflatten(
                0, noise_pred.shape[:2])
-        latent_vis_dict["pred_video"] = pred_video.permute(0, 2, 1, 3, 4).detach().clone().cpu()
+        latent_vis_dict["pred_video"] = pred_video.permute(
+            0, 2, 1, 3, 4).detach().clone().cpu()

-        # noisy_input = pred_noise_to_pred_video(noise_pred, noisy_input, t, self.modules["scheduler"])
-        # next_latent_pred = self.modules["scheduler"].step(
-        #     noise_pred, t, current_latents, return_dict=False)[0]
        return pred_video, target_latent, timestep, latent_vis_dict

    def train_one_step(self, training_batch):  # type: ignore[override]
@@ -329,30 +313,14 @@ class ODEInitTrainingPipeline(TrainingPipeline):
            assert traj_latents.shape[0] == 1

            # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
-            B, S = traj_latents.shape[0], traj_latents.shape[1]
+            _, S = traj_latents.shape[0], traj_latents.shape[1]
            if S < 2:
                raise ValueError("Trajectory must contain at least 2 steps")

-            # Sample per-sample current step i in [0, S-2]
-
-            # idx = torch.randint(low=0, high=S - 1, size=(B, ),
-            #                     device=traj_latents.device)
-
-            # Gather current latents and next latents
-            # batch_indices = torch.arange(B, device=traj_latents.device)
-            # current_latents = traj_latents[batch_indices, idx]  # [B, C, T,H,W]
-            # current_latent = traj_timesteps[:, -1, :, :, :, :]
-            # target_latents = traj_latents[:, -1, :, :, :, :]
-
-            # Corresponding timesteps t (long) -> cast per sample
-            # t = traj_timesteps[:, -1, :, :, :, :]
-            # if t.dtype != torch.long:
-            #     t = t.long()
-
            # Forward to predict next latent by stepping scheduler with predicted noise
            noise_pred, target_latent, t, latent_vis_dict = self._step_predict_next_latent(
                traj_latents, traj_timesteps, text_embeds, text_attention_mask)
-            
+
            training_batch.latent_vis_dict.update(latent_vis_dict)

            mask = t != 0
@@ -398,9 +366,11 @@ class ODEInitTrainingPipeline(TrainingPipeline):
        latents_vis_dict = training_batch.latent_vis_dict
        latent_log_keys = ['noisy_input', 'x0', 'pred_video']
        for latent_key in latent_log_keys:
-            assert latent_key in latents_vis_dict and latents_vis_dict[latent_key] is not None
+            assert latent_key in latents_vis_dict and latents_vis_dict[
+                latent_key] is not None
            latent = latents_vis_dict[latent_key]
-            pixel_latent = self.validation_pipeline.decoding_stage.decode(latent, training_args)
+            pixel_latent = self.validation_pipeline.decoding_stage.decode(
+                latent, training_args)

            video = pixel_latent.cpu().float()
            video = video.permute(0, 2, 1, 3, 4)
@@ -415,15 +385,8 @@ class ODEInitTrainingPipeline(TrainingPipeline):
            wandb.log(wandb_loss_dict, step=step)


-        # dmd_latents_vis_dict = training_batch.dmd_latent_vis_dict
-        # fake_score_latents_vis_dict = training_batch.fake_score_latent_vis_dict
-        # fake_score_log_keys = ['generator_pred_video']
-        # dmd_log_keys = ['faker_score_pred_video', 'real_score_pred_video']
-
-
 def main(args) -> None:
    logger.info("Starting ODE-init training pipeline...")
-    logger.info(f"ARG dmd_denoising_steps: {args.dmd_denoising_steps}")
    pipeline = ODEInitTrainingPipeline.from_pretrained(
        args.pretrained_model_name_or_path, args=args)
    args = pipeline.training_args
@@ -22,7 +22,7 @@ from tqdm.auto import tqdm
 import fastvideo.envs as envs
 from fastvideo.attention.backends.video_sparse_attn import (
    VideoSparseAttentionMetadataBuilder)
-# from fastvideo.attention.backends.vmoba import VideoMobaAttentionMetadataBuilder
+from fastvideo.attention.backends.vmoba import VideoMobaAttentionMetadataBuilder
 from fastvideo.configs.sample import SamplingParam
 from fastvideo.dataset import build_parquet_map_style_dataloader
 from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v
@@ -39,25 +39,20 @@ from fastvideo.training.activation_checkpoint import (
    apply_activation_checkpointing)
 from fastvideo.training.training_utils import (
    clip_grad_norm_while_handling_failing_dtensor_cases,
-    compute_density_for_timestep_sampling, get_scheduler, get_sigmas,
-    load_checkpoint, normalize_dit_input, save_checkpoint,
+    compute_density_for_timestep_sampling, count_trainable, get_scheduler,
+    get_sigmas, load_checkpoint, normalize_dit_input, save_checkpoint,
    shard_latents_across_sp)
-# from fastvideo.utils import (is_vmoba_available, is_vsa_available,
-#                              set_random_seed, shallow_asdict)
-from fastvideo.utils import is_vsa_available, set_random_seed, shallow_asdict
+from fastvideo.utils import (is_vmoba_available, is_vsa_available,
+                             set_random_seed, shallow_asdict)

 import wandb  # isort: skip

 vsa_available = is_vsa_available()
-# vmoba_available = is_vmoba_available()
+vmoba_available = is_vmoba_available()

 logger = init_logger(__name__)


-def _get_trainable_params(model: torch.nn.Module) -> int:
-    return sum(p.numel() for p in model.parameters() if p.requires_grad)
-
-
 class TrainingPipeline(LoRAPipeline, ABC):
    """
    A pipeline for training a model. All training pipelines should inherit from this class.
@@ -117,18 +112,15 @@ class TrainingPipeline(LoRAPipeline, ABC):
                enable_gradient_checkpointing_type)

        noise_scheduler = self.modules["scheduler"]
+        # Set grads for proper modules based on the training mode (Distill, LoRA, etc.)
        self.set_trainable()
        params_to_optimize = self.transformer.parameters()
        params_to_optimize = list(
            filter(lambda p: p.requires_grad, params_to_optimize))
-        # Parse betas from string format "beta1,beta2"
-        betas_str = training_args.betas
-        betas = tuple(float(x.strip()) for x in betas_str.split(","))
-
        self.optimizer = torch.optim.AdamW(
            params_to_optimize,
            lr=training_args.learning_rate,
-            betas=betas,
+            betas=(0.9, 0.999),
            weight_decay=training_args.weight_decay,
            eps=1e-8,
        )
@@ -280,20 +272,20 @@ class TrainingPipeline(LoRAPipeline, ABC):
                patch_size=patch_size,
                VSA_sparsity=current_vsa_sparsity,
                device=get_local_torch_device())
-        # elif vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
-        #     moba_params = self.training_args.moba_config.copy()
-        #     moba_params.update({
-        #         "current_timestep":
-        #         training_batch.timesteps,
-        #         "raw_latent_shape":
-        #         training_batch.raw_latent_shape[2:5],
-        #         "patch_size":
-        #         self.training_args.pipeline_config.dit_config.patch_size,
-        #         "device":
-        #         get_local_torch_device(),
-        #     })
-        #     training_batch.attn_metadata = VideoMobaAttentionMetadataBuilder(
-        #     ).build(**moba_params)
+        elif vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
+            moba_params = self.training_args.moba_config.copy()
+            moba_params.update({
+                "current_timestep":
+                training_batch.timesteps,
+                "raw_latent_shape":
+                training_batch.raw_latent_shape[2:5],
+                "patch_size":
+                self.training_args.pipeline_config.dit_config.patch_size,
+                "device":
+                get_local_torch_device(),
+            })
+            training_batch.attn_metadata = VideoMobaAttentionMetadataBuilder(
+            ).build(**moba_params)
        else:
            training_batch.attn_metadata = None

@@ -318,8 +310,7 @@ class TrainingPipeline(LoRAPipeline, ABC):

    def _transformer_forward_and_compute_loss(
            self, training_batch: TrainingBatch) -> TrainingBatch:
-        # if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN" or vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
-        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN":
+        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN" or vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
            assert training_batch.attn_metadata is not None
        else:
            assert training_batch.attn_metadata is None
@@ -440,7 +431,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
                    local_main_process_only=False)
        if not self.post_init_called:
            self.post_init()
-        num_trainable_params = _get_trainable_params(self.transformer)
+        num_trainable_params = count_trainable(self.transformer)
        logger.info("Starting training with %s B trainable parameters",
                    round(num_trainable_params / 1e9, 3))

@@ -485,9 +476,9 @@ class TrainingPipeline(LoRAPipeline, ABC):
                current_decay_times = min(step // vsa_decay_interval_steps,
                                          vsa_sparsity // vsa_decay_rate)
                current_vsa_sparsity = current_decay_times * vsa_decay_rate
-            # elif vmoba_available:
-            #     # TODO: add vmoba sparsity scheduling here
-            #     current_vsa_sparsity = 0.0
+            elif vmoba_available:
+                # TODO: add vmoba sparsity scheduling here
+                current_vsa_sparsity = 0.0
            else:
                current_vsa_sparsity = 0.0

@@ -536,7 +527,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
                self._log_validation(self.transformer, self.training_args, step)
                gpu_memory_usage = torch.cuda.memory_allocated() / 1024**2
                trainable_params = round(
-                    _get_trainable_params(self.transformer) / 1e9, 3)
+                    count_trainable(self.transformer) / 1e9, 3)
                logger.info(
                    "GPU memory usage after validation: %s MB, trainable params: %sB",
                    gpu_memory_usage, trainable_params)
@@ -572,7 +563,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
        logger.info("  Total optimization steps = %s",
                    self.training_args.max_train_steps)
        logger.info("  Total training parameters per FSDP shard = %s B",
-                    round(_get_trainable_params(self.transformer) / 1e9, 3))
+                    round(count_trainable(self.transformer) / 1e9, 3))
        # print dtype
        logger.info("  Master weight dtype: %s",
                    self.transformer.parameters().__next__().dtype)
@@ -640,7 +631,6 @@ class TrainingPipeline(LoRAPipeline, ABC):
        validation_dataloader = DataLoader(validation_dataset,
                                           batch_size=None,
                                           num_workers=0)
-
        transformer.eval()

        validation_steps = training_args.validation_sampling_steps.split(",")
@@ -349,14 +349,10 @@ def load_checkpoint(transformer,
    """
    if not os.path.exists(checkpoint_path):
        logger.warning("Checkpoint path %s does not exist", checkpoint_path)
-        assert False
        return 0

    # Extract step number from checkpoint path
-    try:
-        step = int(os.path.basename(checkpoint_path).split('-')[-1])
-    except:
-        step = 1
+    step = int(os.path.basename(checkpoint_path).split('-')[-1])

    if rank == 0:
        logger.info("Loading checkpoint from step %s", step)
@@ -1302,6 +1298,10 @@ def get_scheduler(
                         last_epoch=last_epoch)


+def count_trainable(model: torch.nn.Module) -> int:
+    return sum(p.numel() for p in model.parameters() if p.requires_grad)
+
+
 class EMA_FSDP:
    """
    FSDP2-friendly EMA with two modes:
@@ -4,10 +4,10 @@ from copy import deepcopy

 from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
 from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import (
-    FlowMatchEulerDiscreteScheduler)
-from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import WanCausalDMDPipeline
-from fastvideo.training.self_forcing_distillation_pipeline import SelfForcingDistillationPipeline
+from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import (
+    WanCausalDMDPipeline)
+from fastvideo.training.self_forcing_distillation_pipeline import (
+    SelfForcingDistillationPipeline)
 from fastvideo.utils import is_vsa_available

 vsa_available = is_vsa_available()
@@ -69,4 +69,4 @@ if __name__ == "__main__":
    parser = TrainingArgs.add_cli_args(parser)
    parser = FastVideoArgs.add_cli_args(parser)
    args = parser.parse_args()
-    main(args) 
+    main(args)
@@ -1,211 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import sys
-from copy import deepcopy
-from typing import Any
-
-import torch
-
-from fastvideo.configs.pipelines.base import PipelineConfig
-from fastvideo.configs.sample import SamplingParam
-from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v
-from fastvideo.distributed import get_local_torch_device
-from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
-from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_flow_unipc_multistep import (
-    FlowUniPCMultistepScheduler)
-from fastvideo.pipelines.basic.wan.wan_i2v_pipeline import (
-    WanImageToVideoPipeline)
-from fastvideo.pipelines.pipeline_batch_info import ForwardBatch, TrainingBatch
-from fastvideo.training.training_pipeline import TrainingPipeline
-from fastvideo.utils import is_vsa_available, shallow_asdict
-
-vsa_available = is_vsa_available()
-
-logger = init_logger(__name__)
-
-
-class WanT2VI2VTrainingPipeline(TrainingPipeline):
-    """
-    A training pipeline for Wan.
-    """
-    _required_config_modules = ["scheduler", "transformer", "vae"]
-
-    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
-        self.modules["scheduler"] = FlowUniPCMultistepScheduler(
-            shift=fastvideo_args.pipeline_config.flow_shift)
-
-    def create_training_stages(self, training_args: TrainingArgs):
-        """
-        May be used in future refactors.
-        """
-        pass
-
-    def set_schemas(self):
-        self.train_dataset_schema = pyarrow_schema_t2v
-
-    def initialize_validation_pipeline(self, training_args: TrainingArgs):
-        logger.info("Initializing validation pipeline...")
-        args_copy = deepcopy(training_args)
-
-        args_copy.inference_mode = True
-        args_copy.dit_cpu_offload = True
-        # args_copy.pipeline_config.vae_config.load_encoder = False
-        # validation_pipeline = WanImageToVideoValidationPipeline.from_pretrained(
-        pipeline_config = PipelineConfig.from_pretrained(
-            training_args.model_path)
-        pipeline_config.vae_config.load_encoder = True
-        self.validation_pipeline = WanImageToVideoPipeline.from_pretrained(
-            training_args.model_path,
-            args=None,
-            inference_mode=True,
-            pipeline_config=pipeline_config,
-            loaded_modules={
-                "transformer": self.get_module("transformer"),
-            },
-            required_config_modules=[
-                "scheduler", "transformer", "vae", "text_encoder", "tokenizer"
-            ],
-            tp_size=training_args.tp_size,
-            sp_size=training_args.sp_size,
-            num_gpus=training_args.num_gpus,
-            dit_cpu_offload=True,
-        )
-
-    def _get_next_batch(self, training_batch: TrainingBatch) -> TrainingBatch:
-        batch = next(self.train_loader_iter, None)  # type: ignore
-        if batch is None:
-            self.current_epoch += 1
-            logger.info("Starting epoch %s", self.current_epoch)
-            # Reset iterator for next epoch
-            self.train_loader_iter = iter(self.train_dataloader)
-            # Get first batch of new epoch
-            batch = next(self.train_loader_iter)
-
-        latents = batch['vae_latent']
-        latents = latents[:, :, :self.training_args.num_latent_t]
-        encoder_hidden_states = batch['text_embedding']
-        encoder_attention_mask = batch['text_attention_mask']
-        # clip_features = batch['clip_feature']
-        # image_latents = batch['first_frame_latent']
-        # image_latents = image_latents[:, :, :self.training_args.num_latent_t]
-        # pil_image = batch['pil_image']
-        infos = batch['info_list']
-
-        training_batch.latents = latents.to(get_local_torch_device(),
-                                            dtype=torch.bfloat16)
-        training_batch.encoder_hidden_states = encoder_hidden_states.to(
-            get_local_torch_device(), dtype=torch.bfloat16)
-        training_batch.encoder_attention_mask = encoder_attention_mask.to(
-            get_local_torch_device(), dtype=torch.bfloat16)
-        # training_batch.preprocessed_image = pil_image.to(
-        #     get_local_torch_device())
-        # training_batch.image_embeds = clip_features.to(get_local_torch_device())
-        # training_batch.image_latents = image_latents.to(
-        #     get_local_torch_device())
-        training_batch.infos = infos
-
-        return training_batch
-
-    def _prepare_dit_inputs(self,
-                            training_batch: TrainingBatch) -> TrainingBatch:
-        """Override to properly handle I2V concatenation - call parent first, then concatenate image conditioning."""
-
-        # First, call parent method to prepare noise, timesteps, etc. for video latents
-        training_batch = super()._prepare_dit_inputs(training_batch)
-
-        latents = training_batch.latents
-        logger.info("latents.shape: %s", latents.shape)
-        first_frame_latent = latents[:, :, 0, :, :]
-
-        logger.info("first_frame_latent.shape: %s", first_frame_latent.shape)
-        logger.info("training_batch.noisy_model_input.shape: %s",
-                    training_batch.noisy_model_input.shape)
-
-        training_batch.noisy_model_input = torch.cat([
-            first_frame_latent.unsqueeze(2),
-            training_batch.noisy_model_input[:, :, 1:, :, :]
-        ],
-                                                     dim=2)
-
-        return training_batch
-
-    def _build_input_kwargs(self,
-                            training_batch: TrainingBatch) -> TrainingBatch:
-
-        # Image Embeds for conditioning
-        # image_embeds = training_batch.image_embeds
-        # assert torch.isnan(image_embeds).sum() == 0
-        # image_embeds = image_embeds.to(get_local_torch_device(),
-        #                                dtype=torch.bfloat16)
-        # encoder_hidden_states_image = image_embeds
-
-        # NOTE: noisy_model_input already contains concatenated image_latents from _prepare_dit_inputs
-        training_batch.input_kwargs = {
-            "hidden_states":
-            training_batch.noisy_model_input,
-            "encoder_hidden_states":
-            training_batch.encoder_hidden_states,
-            "timestep":
-            training_batch.timesteps.to(get_local_torch_device(),
-                                        dtype=torch.bfloat16),
-            "encoder_attention_mask":
-            training_batch.encoder_attention_mask,
-            # "encoder_hidden_states_image":
-            # encoder_hidden_states_image,
-            "return_dict":
-            False,
-        }
-        return training_batch
-
-    def _prepare_validation_batch(self, sampling_param: SamplingParam,
-                                  training_args: TrainingArgs,
-                                  validation_batch: dict[str, Any],
-                                  num_inference_steps: int) -> ForwardBatch:
-        sampling_param.prompt = validation_batch['prompt']
-        sampling_param.height = training_args.num_height
-        sampling_param.width = training_args.num_width
-        sampling_param.image_path = validation_batch['video_path']
-        sampling_param.num_inference_steps = num_inference_steps
-        sampling_param.data_type = "video"
-        assert self.seed is not None
-        sampling_param.seed = self.seed
-
-        latents_size = [(sampling_param.num_frames - 1) // 4 + 1,
-                        sampling_param.height // 8, sampling_param.width // 8]
-        n_tokens = latents_size[0] * latents_size[1] * latents_size[2]
-        temporal_compression_factor = training_args.pipeline_config.vae_config.arch_config.temporal_compression_ratio
-        num_frames = (training_args.num_latent_t -
-                      1) * temporal_compression_factor + 1
-        sampling_param.num_frames = num_frames
-        batch = ForwardBatch(
-            **shallow_asdict(sampling_param),
-            latents=None,
-            generator=torch.Generator(device="cpu").manual_seed(self.seed),
-            n_tokens=n_tokens,
-            eta=0.0,
-            VSA_sparsity=training_args.VSA_sparsity,
-        )
-
-        return batch
-
-
-def main(args) -> None:
-    logger.info("Starting training pipeline...")
-
-    pipeline = WanT2VI2VTrainingPipeline.from_pretrained(
-        args.pretrained_model_name_or_path, args=args)
-    args = pipeline.training_args
-    pipeline.train()
-    logger.info("Training pipeline done")
-
-
-if __name__ == "__main__":
-    argv = sys.argv
-    from fastvideo.fastvideo_args import TrainingArgs
-    from fastvideo.utils import FlexibleArgumentParser
-    parser = FlexibleArgumentParser()
-    parser = TrainingArgs.add_cli_args(parser)
-    parser = FastVideoArgs.add_cli_args(parser)
-    args = parser.parse_args()
-    args.dit_cpu_offload = False
-    main(args)
@@ -2,10 +2,6 @@
 # Adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/utils.py

 import argparse
-from einops import rearrange
-import torchvision
-import numpy as np
-import imageio
 import ctypes
 import hashlib
 import importlib
@@ -27,10 +23,14 @@ from typing import Any, TypeVar, cast

 import cloudpickle
 import filelock
+import imageio
+import numpy as np
 import torch
+import torchvision
 import yaml
 from diffusers.loaders.lora_base import (
    _best_guess_weight_name)  # watch out for potetential removal from diffusers
+from einops import rearrange
 from huggingface_hub import snapshot_download
 from remote_pdb import RemotePdb
 from torch.distributed.fsdp import MixedPrecisionPolicy
@@ -892,7 +892,8 @@ def best_output_size(w, h, dw, dh, expected_area):
        return ow2, oh2


-def save_decoded_latents_as_video(decoded_latents: list[torch.Tensor], output_path: str, fps: int):
+def save_decoded_latents_as_video(decoded_latents: list[torch.Tensor],
+                                  output_path: str, fps: int):
    # Process outputs
    videos = rearrange(decoded_latents, "b c t h w -> t b c h w")
    frames = []
@@ -1,2 +0,0 @@
-Code in this folder is modified from https://github.com/Wan-Video/Wan2.1
-Apache-2.0 License 
@@ -1,3 +0,0 @@
-from . import configs, distributed, modules
-from .image2video import WanI2V
-from .text2video import WanT2V
@@ -1,42 +0,0 @@
-# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
-from .wan_t2v_14B import t2v_14B
-from .wan_t2v_1_3B import t2v_1_3B
-from .wan_i2v_14B import i2v_14B
-import copy
-import os
-
-os.environ['TOKENIZERS_PARALLELISM'] = 'false'
-
-
-# the config of t2i_14B is the same as t2v_14B
-t2i_14B = copy.deepcopy(t2v_14B)
-t2i_14B.__name__ = 'Config: Wan T2I 14B'
-
-WAN_CONFIGS = {
-    't2v-14B': t2v_14B,
-    't2v-1.3B': t2v_1_3B,
-    'i2v-14B': i2v_14B,
-    't2i-14B': t2i_14B,
-}
-
-SIZE_CONFIGS = {
-    '720*1280': (720, 1280),
-    '1280*720': (1280, 720),
-    '480*832': (480, 832),
-    '832*480': (832, 480),
-    '1024*1024': (1024, 1024),
-}
-
-MAX_AREA_CONFIGS = {
-    '720*1280': 720 * 1280,
-    '1280*720': 1280 * 720,
-    '480*832': 480 * 832,
-    '832*480': 832 * 480,
-}
-
-SUPPORTED_SIZES = {
-    't2v-14B': ('720*1280', '1280*720', '480*832', '832*480'),
-    't2v-1.3B': ('480*832', '832*480'),
-    'i2v-14B': ('720*1280', '1280*720', '480*832', '832*480'),
-    't2i-14B': tuple(SIZE_CONFIGS.keys()),
-}
@@ -1,19 +0,0 @@
-# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
-import torch
-from easydict import EasyDict
-
-# ------------------------ Wan shared config ------------------------#
-wan_shared_cfg = EasyDict()
-
-# t5
-wan_shared_cfg.t5_model = 'umt5_xxl'
-wan_shared_cfg.t5_dtype = torch.bfloat16
-wan_shared_cfg.text_len = 512
-
-# transformer
-wan_shared_cfg.param_dtype = torch.bfloat16
-
-# inference
-wan_shared_cfg.num_train_timesteps = 1000
-wan_shared_cfg.sample_fps = 16
-wan_shared_cfg.sample_neg_prompt = '色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走'
@@ -1,35 +0,0 @@
-# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
-import torch
-from easydict import EasyDict
-
-from .shared_config import wan_shared_cfg
-
-# ------------------------ Wan I2V 14B ------------------------#
-
-i2v_14B = EasyDict(__name__='Config: Wan I2V 14B')
-i2v_14B.update(wan_shared_cfg)
-
-i2v_14B.t5_checkpoint = 'models_t5_umt5-xxl-enc-bf16.pth'
-i2v_14B.t5_tokenizer = 'google/umt5-xxl'
-
-# clip
-i2v_14B.clip_model = 'clip_xlm_roberta_vit_h_14'
-i2v_14B.clip_dtype = torch.float16
-i2v_14B.clip_checkpoint = 'models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth'
-i2v_14B.clip_tokenizer = 'xlm-roberta-large'
-
-# vae
-i2v_14B.vae_checkpoint = 'Wan2.1_VAE.pth'
-i2v_14B.vae_stride = (4, 8, 8)
-
-# transformer
-i2v_14B.patch_size = (1, 2, 2)
-i2v_14B.dim = 5120
-i2v_14B.ffn_dim = 13824
-i2v_14B.freq_dim = 256
-i2v_14B.num_heads = 40
-i2v_14B.num_layers = 40
-i2v_14B.window_size = (-1, -1)
-i2v_14B.qk_norm = True
-i2v_14B.cross_attn_norm = True
-i2v_14B.eps = 1e-6
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
SolitaryThinker	f18901b00b	wip	2025-09-28 05:42:34 +00:00
RandNMR73	13d0aae706	Add Sage Attention 3 Backend (#815 )	2025-09-24 15:11:38 -07:00
William Lin	404cbf4f3c	[self-forcing] [6/n] Add Ode Init training (#811 )	2025-09-22 17:58:19 -07:00
William Lin	958ffec844	[bugfix] Update learning rates for sparse distillation recipe (#812 )	2025-09-22 12:07:03 -07:00
Wei Zhou RandNMR73 SolitaryThinker	31f000d1cc	[self-forcing] [5/n] Add Self-Forcing distillation pipeline (#808 ) Co-authored-by: RandNMR73 <notomatthew31@gmail.com> Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com>	2025-09-20 19:32:10 -07:00
Yongqi Chen	cd32b3e02f	Update example files and readme (#809 )	2025-09-20 18:15:59 -07:00
Zhang Peiyuan	bf27908095	Update WeChat Link	2025-09-20 14:16:20 -07:00
William Lin	c5f9ea53b2	[self-forcing] [4/n] Preprocessing for collecting ODE trajectory (#788 )	2025-09-15 17:54:42 -07:00
William Lin	d32a7184da	[bugfix] Wan2.2 Boundary ratio (#804 )	2025-09-15 11:17:35 -07:00
Wenxuan Tanandgemini-code-assist[bot]	2930abe456	[Bugfix] Fix VMoba requirements (#802 ) Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>	2025-09-14 18:28:52 -07:00
				`@@ -1 +0,0 @@`
				{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":0.50390625,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.19960195198655128,"_runtime":107.325113071}
				`@@ -1 +0,0 @@`
				{"train_loss":0.10545400530099869,"_step":5,"_wandb":{"runtime":35},"step_time":2.2575189135968685,"grad_norm":0.53125,"_runtime":35.701958502,"avg_step_time":2.4754387199878694,"_timestamp":1.7525528420185745e+09,"learning_rate":1e-06,"vsa_sparsity":0,"validation_videos_8_steps":{"videos":[{"caption":"A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","_type":"video-file","sha256":"ee83e83df073a648f89dcd288cccaed9af765e11fc35c77a6e0cb2ebaa1be5b0","size":475248,"path":"media/videos/validation_videos_8_steps_0_ee83e83df073a648f89d.mp4"},{"size":341490,"path":"media/videos/validation_videos_8_steps_0_d0b2758549d5c82845ca.mp4","caption":"A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","_type":"video-file","sha256":"d0b2758549d5c82845ca3c1ac0db6812877631495e77ac041acdc8ea31f0a5ee"},{"caption":"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.","_type":"video-file","sha256":"08638381f1607d6ab10684772be38a58b18628fd402a208b25f6af96e765d454","size":436814,"path":"media/videos/validation_videos_8_steps_0_08638381f1607d6ab106.mp4"}],"captions":["A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","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."],"_type":"videos","count":3}}