Refactor ode trajectory generation

Add env variable to enforce running flash attn
Test
2025-09-26 01:26:03 +00:00 · 2025-09-22 22:33:18 +00:00 · 2025-09-20 05:02:06 +00:00 · 2025-09-20 05:02:06 +00:00 · 2025-09-20 05:02:06 +00:00 · 2025-09-20 05:02:05 +00:00
179 changed files with 11290 additions and 1558 deletions
@@ -18,8 +18,7 @@ steps:
            - path: 
                - "fastvideo/models/encoders/**"
                - "fastvideo/models/loader/**"
-                - "tests/encoders/**"
-                - "tests/utils.py"
+                - "fastvideo/tests/encoders/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -32,8 +31,7 @@ steps:
            - path:
                - "fastvideo/models/vaes/**"
                - "fastvideo/models/loader/**"
-                - "tests/vaes/**"
-                - "tests/utils.py"
+                - "fastvideo/tests/vaes/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -46,8 +44,7 @@ steps:
            - path:
                - "fastvideo/models/dits/**"
                - "fastvideo/models/loader/**"
-                - "tests/transformers/**"
-                - "tests/utils.py"
+                - "fastvideo/tests/transformers/**"
                - "fastvideo/layers/**"
                - "fastvideo/attention/**"
                - "pyproject.toml"
@@ -61,8 +58,6 @@ steps:
                  queue: "default"
            - path: 
                - "fastvideo/**/*.py"
-                - "tests/ssim/**"
-                - "tests/utils.py"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -73,10 +68,9 @@ steps:
                agents:
                  queue: "default"
            - path:
-                - "tests/inference/lora/**"
+                - "fastvideo/tests/lora/**"
                - "fastvideo/models/loader/**"
-                - "tests/transformers/**"
-                - "tests/utils.py"
+                - "fastvideo/tests/transformers/**"
                - "fastvideo/pipelines/**"
                - "fastvideo/layers/lora/**"
                - "pyproject.toml"
@@ -90,7 +84,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
-                - "tests/training/Vanilla/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -102,7 +95,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/training/*distillation_pipeline.py"
-                - "tests/training/distill/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -114,7 +106,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
-                - "tests/training/lora/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -126,7 +117,6 @@ 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"
@@ -143,7 +133,6 @@ 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"
@@ -201,7 +190,6 @@ steps:
            - path:
                - "csrc/attn/vmoba_attn/vmoba/**"
                - "fastvideo/attention/backends/vmoba.py"
-                - "tests/inference/vmoba/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -213,8 +201,6 @@ 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="tests/modal/pr_test.py"
+MODAL_TEST_FILE="fastvideo/tests/modal/pr_test.py"

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

  encoder-test:
@@ -182,7 +174,7 @@ 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 ./tests/encoders -s"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/encoders -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -200,7 +192,7 @@ jobs:
      gpu_count: 1
      volume_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./tests/vaes -s"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/vaes -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -218,7 +210,7 @@ jobs:
      gpu_count: 1
      volume_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./tests/transformers -s"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/transformers -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -244,7 +236,7 @@ jobs:
      volume_size: 200
      disk_size: 200
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:${{ matrix.python-version.tag }}"
-      test_command: "uv pip install -e .[test] && pytest ./tests/ssim -vs"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/ssim -vs"
      timeout_minutes: 60
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -263,7 +255,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/training/Vanilla -srP"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/training/Vanilla -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -283,7 +275,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/training/VSA -srP"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/training/VSA -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -303,7 +295,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./tests/inference/STA -srP"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/inference/STA -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -377,7 +369,7 @@ jobs:
  #     volume_size: 100
  #     disk_size: 100
  #     image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-#     test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/nightly/test_e2e_overfit_single_sample.py -vs"
+  #     test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/nightly/test_e2e_overfit_single_sample.py -vs"
  #     timeout_minutes: 30
  #   secrets:
  #     RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -12,6 +12,9 @@ exclude: |
    scripts/.*|
    fastvideo/data_preprocess/.*|
    fastvideo/dataset/.*|
+    fastvideo/distill/.*|
+    fastvideo/distill\.py|
+    fastvideo/distill_adv\.py|
    fastvideo/models/.*|
    fastvideo/sample/.*|
    fastvideo/train\.py|
@@ -41,10 +44,10 @@ 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/q46BbX6" 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/S7HLCSTh" 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 Chen, Yongqi and Huang, Haofeng 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 Huang, Haofeng and Chen, Yongqi and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
  journal={arXiv preprint arXiv:2505.13389},
  year={2025}
 }
@@ -1,40 +1,32 @@
 (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
@@ -42,7 +34,6 @@ os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLIDING_TILE_ATTN"
 ```

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

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

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

 **`FLASH_ATTN`**
@@ -67,7 +57,7 @@ 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
 ```

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

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

 ```bash
@@ -88,9 +76,7 @@ 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
@@ -101,45 +87,19 @@ 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?
@@ -0,0 +1,3 @@
+#!/bin/bash
+
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
@@ -0,0 +1,76 @@
+{
+  "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
+    }
+  ]
+}
@@ -0,0 +1,13 @@
+{
+  "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=29503
+export MASTER_PORT=29501
 export TOKENIZERS_PARALLELISM=false
-export WANDB_API_KEY=your_wandb_api_key
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
 export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
+export WANDB_MODE=offline
 export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN

 # Configs
-NUM_GPUS=1
+NUM_GPUS=4

 # 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=your_data_dir
-VALIDATION_DATASET_FILE=your_validation_data_dir
+DATA_DIR="data/mixkit-64_processed/Node_0_GPU_1_File_1/combined_parquet_dataset"
+VALIDATION_DATASET_FILE="data/mixkit-64_processed/validation.json"
 # 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 your_output_dir
+  --output_dir "/mnt/sharefs/users/hao.zhang/SFwan_t2v_finetune"
  --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 50
+  --validation_steps 100
  --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 your_ode_init_weights_path
+  --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
 )

 # 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" 
@@ -41,14 +41,13 @@ 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 $OUTPUT_DIR
+  --output_dir"checkpoints/wan_t2v_finetune"
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -92,7 +91,7 @@ validation_args=(

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

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

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-6
+  --learning_rate 1e-5
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
  --weight_only_checkpointing_steps 500
@@ -134,4 +133,4 @@ 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"
@@ -42,14 +42,13 @@ 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 "$OUTPUT_DIR"
+  --output_dir "your_output_dir"
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -93,11 +92,11 @@ validation_args=(

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

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 4e-6
+  --learning_rate 2e-5
  --lr_scheduler "cosine_with_min_lr"
  --min_lr_ratio 0.5
  --lr_warmup_steps 100
-  --fake_score_learning_rate 2e-6
+  --fake_score_learning_rate 1e-5
  --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[@]}"
@@ -16,25 +16,24 @@ 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 "$OUTPUT_DIR"
-  --max_train_steps 4000
-  --train_batch_size 1
+  --output_dir="checkpoints/wan_t2v_finetune"
+  --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
@@ -69,8 +68,8 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-6
-  --mixed_precision "bf16"
+  --learning_rate=1e-5
+  --mixed_precision="bf16"
  --weight_decay 0.01
  --max_grad_norm 1.0
 )
@@ -108,4 +107,4 @@ 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():
@@ -17,12 +17,13 @@ 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 = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+    sampling_param.image_path = "test.jpg"
    # Generate videos with the same simple API, regardless of GPU count
    prompt = (
        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
@@ -44,4 +45,4 @@ def main():


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

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

-    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)
+    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)

 if __name__ == "__main__":
    main()
@@ -1,5 +1,14 @@
 {
  "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,
@@ -19,52 +28,7 @@
      "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. ",
+      "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,
@@ -0,0 +1,99 @@
+#!/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="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/"
+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"
+  --override_transformer_cls_name "CausalWanTransformer3DModel"
+  --wandb_run_name "fixed_wan_ode_init_70k_6e-6"
+  # --resume_from_checkpoint "ode_init_diffusers/"
+  --max_train_steps 6000
+  --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
+  --warp_denoising_step
+  # --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 50
+  --validation_sampling_steps "50"
+  --validation_guidance_scale "6.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 6e-6
+  --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"
+)
+
+# 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[@]}"
@@ -1,5 +1,5 @@
 #!/bin/bash
-#SBATCH --job-name=2e6B8_16kFV_ode_vidprom
+#SBATCH --job-name=1e5B2_16kFV_warp_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/ode_vidprom8b16k_2e-6.out
-#SBATCH --error=ode_vidprom16k/ode_vidprom8b16k_2e-6.err
+#SBATCH --output=ode_vidprom16k_warp/Dode_vidprom8b16k_1e-5.out
+#SBATCH --error=ode_vidprom16k_warp/Dode_vidprom8b16k_1e-5.err
 #SBATCH --exclusive
 set -e -x

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

 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_API_KEY=your-wandb-api-key
+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"
-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
+# 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"
 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 $OUTPUT_DIR
+  --output_dir "Dwarp_vidprom_8b16k_test_warp_1e-5"
  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "vidprom_8b16k_ode_init_2e-6"
+  --wandb_run_name "Dwarp_vidprom_8b16k_wan_ode_init_1e-5"
  # --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 81
+  --num_frames 77
  --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 $INIT_WEIGHTS_FROM_SAFETENSORS
+  # --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
 )

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-6
+  --learning_rate 1e-5
  --mixed_precision "bf16"
  --checkpointing_steps 500
  --weight_decay 1e-4
@@ -1,22 +1,21 @@
 #!/bin/bash
-#SBATCH --job-name=t2v
+#SBATCH --job-name=ode_vidprom2k
 #SBATCH --partition=main
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
+#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=VSA_t2v_output/t2v_%j.out
-#SBATCH --error=VSA_t2v_output/t2v_%j.err
+#SBATCH --output=ode_vidprom2k_output/ode_vidprom2k.out
+#SBATCH --error=ode_vidprom2k_output/ode_vidprom2k.err
 #SBATCH --exclusive
 set -e -x

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

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

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

-# Configs
-NUM_GPUS=8
+
+# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR=your_data_dir
-VALIDATION_DATASET_FILE=your_validation_dataset_file
+DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing/"
+VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
+NUM_GPUS=8
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
-  --tracker_project_name wan_t2v_VSA
-  --output_dir "checkpoints/wan_t2v_finetune_VSA"
-  --max_train_steps 4000
+  --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
  --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 81
-  # --enable_gradient_checkpointing_type "full" # if OOM enable this
+  --num_frames 77
+  # --enable_gradient_checkpointing_type "full"
 )

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

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

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

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

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

@@ -103,31 +106,26 @@ 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
-  --flow_shift 1
-  --seed 1000
-)
-
-# VSA arguments
-vsa_args=(
-  --VSA_decay_rate 0.03 \
-  --VSA_decay_interval_steps 50 \
-  --VSA_sparsity 0.9 \
+  --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/wan_training_pipeline.py \
+  --nnodes $SLURM_JOB_NUM_NODES \
+  --nproc_per_node $NUM_GPUS \
+  --node_rank $SLURM_PROCID \
+  --rdzv_backend=c10d \
+  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/ode_causal_pipeline.py \
    "${parallel_args[@]}" \
    "${model_args[@]}" \
    "${dataset_args[@]}" \
    "${training_args[@]}" \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}" \
-    "${vsa_args[@]}"  
+    "${miscellaneous_args[@]}"
@@ -0,0 +1,132 @@
+#!/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[@]}"
@@ -0,0 +1,103 @@
+#!/bin/bash
+
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=offline
+export TOKENIZERS_PARALLELISM=false
+export MASTER_PORT=29501
+# 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/mixkit-64_processed/Node_0_GPU_1_File_1/combined_parquet_dataset"
+# DATA_DIR="/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_single_lmdb_sf/"
+VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
+NUM_GPUS=1
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_ode_init"
+  --output_dir "Dwarp_vidprom_8b16k_test_warp_1e-5"
+  --override_transformer_cls_name "CausalWanTransformer3DModel"
+  --wandb_run_name "Dwarp_vidprom_8b16k_wan_ode_init_1e-5"
+  # --resume_from_checkpoint "ode_init_diffusers/"
+  --warp_denoising_step
+  --log_visualization
+  --max_train_steps 10
+  --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
+  --dmd_denoising_steps "1000,750,500,250"
+  --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 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"
+)
+
+# 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
+  --seed 1024
+  # --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 \
+  --master_port $MASTER_PORT \
+  --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[@]}"
@@ -0,0 +1,98 @@
+#!/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[@]}"
@@ -0,0 +1,100 @@
+#!/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="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-16k-t2v-1-3b/"
+VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
+NUM_GPUS=1
+export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_ode_init"
+  --output_dir "debug_ode_init"
+  --override_transformer_cls_name "CausalWanTransformer3DModel"
+  --wandb_run_name "debug_ode_init"
+  # --resume_from_checkpoint "ode_init_diffusers/"
+  --max_train_steps 1000
+  --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 81
+  --warp_denoising_step
+  # --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 10
+  --validation_sampling_steps "50"
+  --validation_guidance_scale "6.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 1e-5
+  --mixed_precision "bf16"
+  --checkpointing_steps 1000
+  --weight_decay 1e-4
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --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[@]}"
@@ -0,0 +1,25 @@
+#!/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/vidprom_1.txt"
+OUTPUT_DIR="data/ode_vidprom_1_fv/"
+
+torchrun --nproc_per_node=$GPU_NUM \
+    fastvideo/pipelines/preprocess/v1_preprocess.py \
+    --model_path $MODEL_PATH \
+    --data_merge_path $DATA_MERGE_PATH \
+    --preprocess_video_batch_size 1 \
+    --seed 42 \
+    --max_height 480 \
+    --max_width 832 \
+    --num_frames 81 \
+    --flow_shift 5.0 \
+    --dataloader_num_workers 0 \
+    --output_dir=$OUTPUT_DIR \
+    --train_fps 16 \
+    --samples_per_file 1 \
+    --flush_frequency 1 \
+    --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
+    }
+  ]
+}
@@ -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/combined_parquet_dataset/"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
+DATA_DIR="data/crush-smol_processed_t2v_old"
+VALIDATION_DATASET_FILE="examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/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 200
+  --validation_steps 50
  --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/"
+OUTPUT_DIR="data/crush-smol_processed_t2v_old/"

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

-GPU_NUM=2 # 2,4,8
+GPU_NUM=1 # 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 2 \
+    --preprocess.preprocess_video_batch_size 8 \
    --preprocess.dataloader_num_workers 0 \
    --preprocess.max_height 480 \
    --preprocess.max_width 832 \
@@ -0,0 +1,94 @@
+#!/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[@]}"
@@ -0,0 +1,24 @@
+#!/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" 
@@ -1,72 +0,0 @@
-# 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
@@ -3,6 +3,7 @@
 import torch
 import torch.nn as nn

+import fastvideo.envs as envs
 from fastvideo.attention.selector import backend_name_to_enum, get_attn_backend
 from fastvideo.distributed.communication_op import (
    sequence_model_parallel_all_gather, sequence_model_parallel_all_to_all_4D)
@@ -36,7 +37,7 @@ class DistributedAttention(nn.Module):
        if num_kv_heads is None:
            num_kv_heads = num_heads

-        dtype = get_compute_dtype()
+        dtype = torch.bfloat16 if envs.FASTVIDEO_FORCE_ATTN_BF16 else get_compute_dtype()
        attn_backend = get_attn_backend(
            head_size,
            dtype,
@@ -221,7 +222,7 @@ class LocalAttention(nn.Module):
        if num_kv_heads is None:
            num_kv_heads = num_heads

-        dtype = get_compute_dtype()
+        dtype = torch.bfloat16 if envs.FASTVIDEO_FORCE_ATTN_BF16 else get_compute_dtype()
        attn_backend = get_attn_backend(
            head_size,
            dtype,
@@ -15,10 +15,13 @@ 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.SAGE_ATTN_THREE)
+        AttentionBackendEnum.SLIDING_TILE_ATTN,
+        AttentionBackendEnum.SAGE_ATTN,
+        AttentionBackendEnum.FLASH_ATTN,
+        AttentionBackendEnum.TORCH_SDPA,
+        AttentionBackendEnum.VIDEO_SPARSE_ATTN,
+        AttentionBackendEnum.VMOBA_ATTN,
+    )

    hidden_size: int = 0
    num_attention_heads: int = 0
@@ -45,10 +45,13 @@ 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)
    dit_precision: str = "bf16"
+    dit_forward_precision: str = "bf16"

    # VAE configuration
    vae_config: VAEConfig = field(default_factory=VAEConfig)
@@ -215,6 +218,24 @@ 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")
@@ -246,7 +267,9 @@ 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))

@@ -50,6 +50,7 @@ PIPELINE_DETECTOR: dict[str, Callable[[str], bool]] = {
    "wandmdpipeline": lambda id: "wandmdpipeline" in id.lower(),
    "wancausaldmdpipeline": lambda id: "wancausaldmdpipeline" in id.lower(),
    "stepvideo": lambda id: "stepvideo" in id.lower(),
+    "wancausaldmdpipeline": lambda id: "wancausaldmdpipeline" in id.lower(),
    # Add other pipeline architecture detectors
 }

@@ -0,0 +1,43 @@
+# 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)
+        }
+
@@ -0,0 +1,75 @@
+# 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
@@ -3,6 +3,9 @@ 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]:
@@ -18,6 +18,7 @@ if TYPE_CHECKING:
    FASTVIDEO_LOGGING_PREFIX: str = ""
    FASTVIDEO_LOGGING_CONFIG_PATH: str | None = None
    FASTVIDEO_TRACE_FUNCTION: int = 0
+    FASTVIDEO_FORCE_ATTN_BF16: bool = False
    FASTVIDEO_ATTENTION_BACKEND: str | None = None
    FASTVIDEO_ATTENTION_CONFIG: str | None = None
    FASTVIDEO_WORKER_MULTIPROC_METHOD: str = "fork"
@@ -168,6 +169,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
    "FASTVIDEO_TRACE_FUNCTION":
    lambda: int(os.getenv("FASTVIDEO_TRACE_FUNCTION", "0")),

+    # if set, fastvideo will force attention to be computed in bfloat16
+    "FASTVIDEO_FORCE_ATTN_BF16":
+    lambda: bool(int(os.getenv("FASTVIDEO_FORCE_ATTN_BF16", "0"))),
+
    # Backend for attention computation
    # Available options:
    # - "TORCH_SDPA": use torch.nn.MultiheadAttention
@@ -175,7 +180,6 @@ 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),

@@ -705,6 +705,7 @@ 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
@@ -1140,10 +1141,11 @@ class TrainingArgs(FastVideoArgs):
            "--last-step-only",
            action=StoreBoolean,
            help="Whether to only use the last timestep for training")
-        parser.add_argument("--context-noise",
-                            type=int,
-                            default=TrainingArgs.context_noise,
-                            help="Context noise level for cache updates")
+        parser.add_argument(
+            "--context-noise",
+            type=int,
+            default=TrainingArgs.context_noise,
+            help="Context noise level for cache updates")

        return parser

@@ -1151,4 +1153,4 @@ class TrainingArgs(FastVideoArgs):
 def parse_int_list(value: str) -> list[int]:
    if not value:
        return []
-    return [int(x.strip()) for x in value.split(",")]
+    return [int(x.strip()) for x in value.split(",")]
@@ -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)) *
-                (1.0 + scale) + shift).flatten(1, 2)
+                (1 + scale) + shift).flatten(1, 2)
        else:
-            modulated = normalized * (1.0 + scale) + shift
+            modulated = normalized * (1 + 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)) *
-                (1.0 + scale) + shift).flatten(1, 2)
+                (1 + scale) + shift).flatten(1, 2)
        else:
            # scale.shape: [batch_size, 1, inner_dim]
            # shift.shape: [batch_size, 1, inner_dim]
-            output = normalized * (1.0 + scale) + shift
+            output = normalized * (1 + 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 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.norm1 = nn.LayerNorm(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)
@@ -212,8 +212,7 @@ class CausalWanTransformerBlock(nn.Module):
            norm_type="layer",
            eps=eps,
            elementwise_affine=True,
-            dtype=torch.float32,
-            compute_dtype=torch.float32)
+            dtype=torch.float32)

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

        # 3. Feed-forward
        self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
@@ -252,29 +250,34 @@ 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.float()
+        e = self.scale_shift_table + temb
        # 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
+        # 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)

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

        query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
        key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -288,8 +291,6 @@ 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,
@@ -298,13 +299,10 @@ 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

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

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

        self.__post_init__()
@@ -491,12 +487,16 @@ class CausalWanTransformer3DModel(BaseDiT):
        )
        freqs_cos = freqs_cos.to(hidden_states.device)
        freqs_sin = freqs_sin.to(hidden_states.device)
-        freqs_cis = (freqs_cos.float(),
-                     freqs_sin.float()) if freqs_cos is not None else None
+        freqs_cis = (freqs_cos,
+                     freqs_sin) 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)
@@ -543,14 +543,9 @@ class CausalWanTransformer3DModel(BaseDiT):
        hidden_states = self.norm_out(hidden_states, shift, scale)
        hidden_states = self.proj_out(hidden_states)

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

-        return output
+        return torch.stack(output)

    def _forward_train(self,
                hidden_states: torch.Tensor,
@@ -561,6 +556,8 @@ class CausalWanTransformer3DModel(BaseDiT):
                start_frame: int = 0,
                **kwargs) -> torch.Tensor:

+        logger.info("timestep dtype: %s, timestep sum: %s", timestep.dtype, timestep.float().sum().item())
+
        orig_dtype = hidden_states.dtype
        if not isinstance(encoder_hidden_states, torch.Tensor):
            encoder_hidden_states = encoder_hidden_states[0]
@@ -591,8 +588,8 @@ class CausalWanTransformer3DModel(BaseDiT):
        )
        freqs_cos = freqs_cos.to(hidden_states.device)
        freqs_sin = freqs_sin.to(hidden_states.device)
-        freqs_cis = (freqs_cos.float(),
-                     freqs_sin.float()) if freqs_cos is not None else None
+        freqs_cis = (freqs_cos,
+                     freqs_sin) if freqs_cos is not None else None

        # Construct blockwise causal attn mask
        if self.block_mask is None:
@@ -605,8 +602,12 @@ 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)
@@ -641,14 +642,9 @@ class CausalWanTransformer3DModel(BaseDiT):
        hidden_states = self.norm_out(hidden_states, shift, scale)
        hidden_states = self.proj_out(hidden_states)

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

-        return output
+        return torch.stack(output)

    def forward(
        self,
@@ -656,6 +652,34 @@ class CausalWanTransformer3DModel(BaseDiT):
        **kwargs
    ):
        if kwargs.get('kv_cache', None) is not None:
-            return self._forward_inference(*args, **kwargs)
+            noise_pred = self._forward_inference(*args, **kwargs)
        else:
-            return self._forward_train(*args, **kwargs)
+            noise_pred = self._forward_train(*args, **kwargs)
+
+        return noise_pred
+
+    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
@@ -1,3 +1,5 @@
+import torch
+import torch.nn as nn
 # SPDX-License-Identifier: Apache-2.0

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

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

-    def forward(self,
-                encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
-        dtype = encoder_hidden_states_image.dtype
+    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
        hidden_states = self.norm1(encoder_hidden_states_image)
        hidden_states = self.ff(hidden_states)
-        hidden_states = self.norm2(hidden_states).to(dtype)
+        hidden_states = self.norm2(hidden_states)
        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")
+            dim, frequency_embedding_size=time_freq_dim, act_layer="silu", freq_dtype=torch.float64)
        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(self.to_q(x)[0]).view(b, -1, n, d)
+        q = self.norm_q.forward_native(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(self.to_k(context)[0]).view(b, -1, n, d)
+                k = self.norm_k.forward_native(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
@@ -169,11 +169,16 @@ class WanT2VCrossAttention(WanSelfAttention):
                k = crossattn_cache["k"]
                v = crossattn_cache["v"]
        else:
-            k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
+            k = self.norm_k.forward_native(self.to_k(context)[0]).view(b, -1, n, d)
            v = self.to_v(context)[0].view(b, -1, n, d)

-        # compute attention
-        x = self.attn(q, k, v)
+        if envs.FASTVIDEO_FORCE_ATTN_BF16:
+            out_dtype = v.dtype
+            # compute attention
+            x = self.attn(q.to(torch.bfloat16), k.to(torch.bfloat16), v.to(torch.bfloat16)).to(out_dtype)
+        else:
+            # compute attention
+            x = self.attn(q, k, v)

        # output
        x = x.flatten(2)
@@ -213,10 +218,10 @@ class WanI2VCrossAttention(WanSelfAttention):
        b, n, d = x.size(0), self.num_heads, self.head_dim

        # compute query, key, value
-        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)
+        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)
        v = self.to_v(context)[0].view(b, -1, n, d)
-        k_img = self.norm_added_k(self.add_k_proj(context_img)[0]).view(
+        k_img = self.norm_added_k.forward_native(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 +252,7 @@ class WanTransformerBlock(nn.Module):
        super().__init__()

        # 1. Self-attention
-        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.norm1 = nn.LayerNorm(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 +283,29 @@ class WanTransformerBlock(nn.Module):
            norm_type="layer",
            eps=eps,
            elementwise_affine=True,
-            dtype=torch.float32,
-            compute_dtype=torch.float32)
+            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,
-            compute_dtype=torch.float32)
+            dim, 
+            norm_type="layer", 
+            eps=eps, 
+            elementwise_affine=False, 
+            dtype=torch.float32)

        # 3. Feed-forward
        self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
@@ -319,12 +324,11 @@ 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.float()
+                self.scale_shift_table.unsqueeze(0) + temb
            ).chunk(6, dim=2)
            # batch_size, seq_len, 1, inner_dim
            shift_msa = shift_msa.squeeze(2)
@@ -335,22 +339,20 @@ 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.float()
+            e = self.scale_shift_table + temb
            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.float()) *
-                              (1 + scale_msa) + shift_msa).to(orig_dtype)
+        norm_hidden_states = self.norm1(hidden_states) * (1 + scale_msa) + shift_msa
        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(query)
+            query = self.norm_q.forward_native(query)
        if self.norm_k is not None:
-            key = self.norm_k(key)
+            key = self.norm_k.forward_native(key)

        query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
        key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -362,7 +364,12 @@ class WanTransformerBlock(nn.Module):
                                       is_neox_style=False), _apply_rotary_emb(
                                           key, cos, sin, is_neox_style=False)

-        attn_output, _ = self.attn1(query, key, value)
+        if envs.FASTVIDEO_FORCE_ATTN_BF16:
+            out_dtype = value.dtype
+            attn_output, _ = self.attn1(query.to(torch.bfloat16), key.to(torch.bfloat16), value.to(torch.bfloat16))
+            attn_output = attn_output.to(out_dtype)
+        else:
+            attn_output, _ = self.attn1(query, key, value)
        attn_output = attn_output.flatten(2)
        attn_output, _ = self.to_out(attn_output)
        attn_output = attn_output.squeeze(1)
@@ -370,26 +377,20 @@ 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,
@@ -406,7 +407,7 @@ class WanTransformerBlock_VSA(nn.Module):
        super().__init__()

        # 1. Self-attention
-        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.norm1 = nn.LayerNorm(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)
@@ -438,8 +439,7 @@ class WanTransformerBlock_VSA(nn.Module):
            norm_type="layer",
            eps=eps,
            elementwise_affine=True,
-            dtype=torch.float32,
-            compute_dtype=torch.float32)
+            dtype=torch.float32)

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

        # 3. Feed-forward
        self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
@@ -480,23 +479,22 @@ 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.float()
+        e = self.scale_shift_table + temb
        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.float()) *
-                              (1 + scale_msa) + shift_msa).to(orig_dtype)
+        norm_hidden_states = (self.norm1(hidden_states) *
+                              (1 + scale_msa) + shift_msa)
        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(query)
+            query = self.norm_q.forward_native(query)
        if self.norm_k is not None:
-            key = self.norm_k(key)
+            key = self.norm_k.forward_native(key)

        query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
        key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -521,8 +519,6 @@ 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,
@@ -530,17 +526,15 @@ 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
@@ -598,8 +592,7 @@ class WanTransformer3DModel(CachableDiT):
                                            norm_type="layer",
                                            eps=config.eps,
                                            elementwise_affine=False,
-                                            dtype=torch.float32,
-                                            compute_dtype=torch.float32)
+                                            dtype=torch.float32)
        self.proj_out = nn.Linear(
            inner_dim, config.out_channels * math.prod(config.patch_size))
        self.scale_shift_table = nn.Parameter(
@@ -659,10 +652,12 @@ 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.float(),
-                     freqs_sin.float()) if freqs_cos is not None else None
+        freqs_cis = (freqs_cos,
+                     freqs_sin) 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)
@@ -672,6 +667,8 @@ 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:
@@ -728,14 +725,35 @@ class WanTransformer3DModel(CachableDiT):
        hidden_states = self.norm_out(hidden_states, shift, scale)
        hidden_states = self.proj_out(hidden_states)

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

-        return output
+        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

    def maybe_cache_states(self, hidden_states: torch.Tensor,
                           original_hidden_states: torch.Tensor) -> None:
@@ -827,5 +845,4 @@ 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
@@ -238,6 +238,7 @@ class TextEncoderLoader(ComponentLoader):
                1]

        target_device = get_local_torch_device()
+        logger.info("Loading text encoder in %s precision", encoder_precision)
        # TODO(will): add support for other dtypes
        return self.load_model(model_path, encoder_config, target_device,
                               fastvideo_args, encoder_precision)
@@ -415,7 +416,6 @@ 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
@@ -443,19 +443,15 @@ class TransformerLoader(ComponentLoader):

        if use_custom_weights:
            logger.info("Using custom initialization weights from: %s", 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]
+            safetensors_list = [custom_weights_path]

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

        default_dtype = PRECISION_TO_TYPE[
            fastvideo_args.pipeline_config.dit_precision]
+        param_dtype = PRECISION_TO_TYPE[
+            fastvideo_args.pipeline_config.dit_forward_precision]

        # Load the model using FSDP loader
        logger.info("Loading model from %s, default_dtype: %s", cls_name,
@@ -476,7 +472,7 @@ class TransformerLoader(ComponentLoader):
            fsdp_inference=fastvideo_args.use_fsdp_inference,
            # TODO(will): make these configurable
            default_dtype=default_dtype,
-            param_dtype=torch.bfloat16,
+            param_dtype=param_dtype,
            reduce_dtype=torch.float32,
            output_dtype=None,
            training_mode=fastvideo_args.training_mode)
@@ -485,8 +481,11 @@ class TransformerLoader(ComponentLoader):
        total_params = sum(p.numel() for p in model.parameters())
        logger.info("Loaded model with %.2fB parameters", total_params / 1e9)

-        assert next(model.parameters()).dtype == default_dtype, "Model dtype does not match default dtype"
-
+        # Need to convert the model to the default_dtype
+        # Otherwise, the model master weights will be in param_dtype, and the gradients will also be in param_dtype
+        # This means the param update will be in lower precision, causing precision loss
+        logger.info("Converting model to dtype: %s", default_dtype)
+        model = model.to(default_dtype)
        model = model.eval()
        return model

@@ -88,7 +88,6 @@ def maybe_load_fsdp_model(
        mp_policy=mp_policy,
    )

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

@@ -180,3 +180,25 @@ def pred_noise_to_pred_video(pred_noise: torch.Tensor,
    sigma_t = sigmas[timestep_id].reshape(-1, 1, 1, 1)
    pred_video = noise_input_latent - sigma_t * pred_noise
    return pred_video.to(dtype)
+
+def pred_video_to_pred_noise(x0_pred: torch.Tensor, xt: torch.Tensor, timestep: torch.Tensor, scheduler: Any) -> torch.Tensor:
+    """
+    Convert x0 prediction to flow matching's prediction.
+    x0_pred: the x0 prediction with shape [B, C, H, W]
+    xt: the input noisy data with shape [B, C, H, W]
+    timestep: the timestep with shape [B]
+
+    pred = (x_t - x_0) / sigma_t
+    """
+    # use higher precision for calculations
+    original_dtype = x0_pred.dtype
+    x0_pred, xt, sigmas, timesteps = map(
+        lambda x: x.double().to(x0_pred.device), [x0_pred, xt,
+                                                    scheduler.sigmas,
+                                                    scheduler.timesteps]
+    )
+    timestep_id = torch.argmin(
+        (timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
+    sigma_t = sigmas[timestep_id].reshape(-1, 1, 1, 1)
+    flow_pred = (xt - x0_pred) / sigma_t
+    return flow_pred.to(original_dtype)
@@ -7,6 +7,8 @@ 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
@@ -121,14 +121,25 @@ class ComposedPipelineBase(ABC):
                        model_path: str,
                        device: str | None = None,
                        torch_dtype: torch.dtype | None = None,
-                        pipeline_config: str | PipelineConfig | None = None,
+                        pipeline_config: 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.
+        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
+
        loaded_modules: Optional[Dict[str, torch.nn.Module]] = None,
        If provided, loaded_modules will be used instead of loading from config/pretrained weights.
        """
@@ -136,9 +147,18 @@ 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():
@@ -149,7 +169,8 @@ 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.
-            assert fastvideo_args.pipeline_config.dit_precision == 'fp32', 'only fp32 is supported for training'
+            fastvideo_args.pipeline_config.dit_precision = 'fp32'
+            # 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[torch.Tensor] | None = None
+    trajectory_timesteps: list[int] | 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, Any] = field(default_factory=dict)
+    latent_vis_dict: dict[str, torch.Tensor] = field(default_factory=dict)
    fake_score_latent_vis_dict: dict[str, Any] = field(default_factory=dict)


@@ -12,8 +12,10 @@ 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
@@ -40,6 +42,11 @@ from fastvideo.pipelines.stages import (DecodingStage, DenoisingStage,
                                        TimestepPreparationStage)
 from fastvideo.utils import save_decoded_latents_as_video, shallow_asdict

+from fastvideo.forward_context import set_forward_context
+from fastvideo.models.utils import pred_noise_to_pred_video, pred_video_to_pred_noise
+from fastvideo.models.loader.component_loader import TransformerLoader
+from fastvideo.distributed import get_local_torch_device
+
 logger = init_logger(__name__)


@@ -68,6 +75,15 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
        self.modules["scheduler"].set_timesteps(num_inference_steps=48,
                                                denoising_strength=1.0)

+        fastvideo_args.model_loaded["transformer"] = False
+        loader = TransformerLoader()
+        fastvideo_args.pipeline_config.dit_precision = "fp32" # Overwrite the precision to fp32 for transformer
+        fastvideo_args.pipeline_config.dit_forward_precision = "fp32"
+        self.transformer = loader.load(
+            fastvideo_args.model_paths["transformer"], fastvideo_args)
+        self.add_module("transformer", self.transformer)
+        fastvideo_args.model_loaded["transformer"] = True
+
        self.add_stage(stage_name="input_validation_stage",
                       stage=InputValidationStage())
        self.add_stage(stage_name="prompt_encoding_stage",
@@ -96,6 +112,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
        """Preprocess text-only data and generate trajectory information."""

        for batch_idx, data in enumerate(self.pbar):
+            # logger.info("transformer weight sum: %s", sum(p.float().sum().item() for p in self.transformer.parameters()))
            if data is None:
                continue

@@ -126,6 +143,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                    ]

                batch_captions = valid_data["text"]
+                self.prompt_encoding_stage.text_encoders[0] = self.prompt_encoding_stage.text_encoders[0].to(dtype=torch.bfloat16).to(dtype=torch.float32)
                # Encode text using the standalone TextEncodingStage API
                prompt_embeds_list, prompt_masks_list = self.prompt_encoding_stage.encode_text(
                    batch_captions,
@@ -134,22 +152,25 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                    return_attention_mask=True,
                )
                prompt_embeds = prompt_embeds_list[0]
+                logger.info("prompt_embeds sum: %s, prompt_embeds shape: %s, prompt_embeds dtype: %s", prompt_embeds.float().sum(), prompt_embeds.shape, prompt_embeds.dtype)
                prompt_attention_masks = prompt_masks_list[0]
                assert prompt_embeds.shape[0] == prompt_attention_masks.shape[0]

                sampling_params = SamplingParam.from_pretrained(args.model_path)

+                negative_prompt = '色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走'
+
                # encode negative prompt for trajectory collection
                if sampling_params.guidance_scale > 1 and sampling_params.negative_prompt is not None:
                    negative_prompt_embeds_list, negative_prompt_masks_list = self.prompt_encoding_stage.encode_text(
-                        sampling_params.negative_prompt,
+                        negative_prompt,
                        fastvideo_args,
                        encoder_index=[0],
                        return_attention_mask=True,
                    )
-                    negative_prompt_embed = negative_prompt_embeds_list[0][0]
+                    negative_prompt_embed = negative_prompt_embeds_list[0]
                    negative_prompt_attention_mask = negative_prompt_masks_list[
-                        0][0]
+                        0]
                else:
                    negative_prompt_embed = None
                    negative_prompt_attention_mask = None
@@ -176,30 +197,129 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                    batch.return_trajectory_latents = True
                    # Enabling this will save the decoded trajectory videos.
                    # Used for debugging.
-                    batch.return_trajectory_decoded = False
+                    batch.return_trajectory_decoded = True
                    batch.height = args.max_height
                    batch.width = args.max_width
                    batch.fps = args.train_fps
-                    batch.guidance_scale = 6.0
+                    batch.guidance_scale = 3.0
                    batch.do_classifier_free_guidance = True

                    result_batch = self.input_validation_stage(
                        batch, fastvideo_args)
                    result_batch = self.timestep_preparation_stage(
                        batch, fastvideo_args)
-                    result_batch = self.latent_preparation_stage(
-                        result_batch, fastvideo_args)
-                    result_batch = self.denoising_stage(result_batch,
-                                                        fastvideo_args)
+                    # result_batch = self.latent_preparation_stage(
+                    #     result_batch, fastvideo_args)
+                    # result_batch = self.denoising_stage(result_batch,
+                    #                                     fastvideo_args)
+                    noisy_input = []
+                    # latents = result_batch.latents.permute(0, 2, 1, 3, 4)
+                    latents = torch.randn(
+                        [1, 21, 16, 60, 104], dtype=torch.float32, device=get_local_torch_device()
+                    )
+                    # logger.info("transformer weight sum: %s", sum(p.float().sum().item() for p in self.transformer.parameters()))
+                    logger.info("latents sum: %s, latents shape: %s, latents dtype: %s", latents.float().sum(), latents.shape, latents.dtype)
+
+                    logger.info("scheduler timesteps: %s", self.get_module("scheduler").timesteps)
+                    for progress_id, t in enumerate(tqdm(self.get_module("scheduler").timesteps)):
+                        timestep = t * \
+                            torch.ones([1, 21], device=latents.device, dtype=torch.float32)
+
+                        noisy_input.append(latents)
+
+                        with set_forward_context(
+                            current_timestep=0,
+                            attn_metadata=None,
+                            forward_batch=None,
+                        ):
+                            # logger.info("prompt_embed sum: %s, prompt_embed shape: %s, prompt_embed dtype: %s", prompt_embed.float().sum(), prompt_embed.shape, prompt_embed.dtype)
+                            # logger.info("timestep: %s", timestep[:, 0])
+                            # Run transformer
+                            cond_pred_noise_btchw = self.transformer(
+                                hidden_states=latents.permute(0, 2, 1, 3, 4),
+                                encoder_hidden_states=prompt_embed,
+                                timestep=timestep[:, 0]
+                            ).permute(0, 2, 1, 3, 4)
+
+                        # logger.info("cond_pred_noise_btchw sum: %s, cond_pred_noise_btchw shape: %s, cond_pred_noise_btchw dtype: %s", cond_pred_noise_btchw.float().sum(), cond_pred_noise_btchw.shape, cond_pred_noise_btchw.dtype)
+
+                        cond_pred_video_btchw = pred_noise_to_pred_video(
+                            pred_noise=cond_pred_noise_btchw.flatten(0, 1),
+                            noise_input_latent=latents.flatten(0, 1),
+                            timestep=timestep.flatten(0, 1),
+                            scheduler=self.get_module("scheduler")).unflatten(
+                                0, cond_pred_noise_btchw.shape[:2])
+
+                        # logger.info("cond_pred_video_btchw sum: %s, cond_pred_video_btchw shape: %s, cond_pred_video_btchw dtype: %s", cond_pred_video_btchw.float().sum(), cond_pred_video_btchw.shape, cond_pred_video_btchw.dtype)
+
+                        with set_forward_context(
+                            current_timestep=t,
+                            attn_metadata=None,
+                            forward_batch=result_batch,
+                        ):
+                            # logger.info("latents sum: %s, latents shape: %s, latents dtype: %s", latents.float().sum(), latents.shape, latents.dtype)
+                            # Run transformer
+                            uncond_pred_noise_btchw = self.transformer(
+                                latents.permute(0, 2, 1, 3, 4),
+                                negative_prompt_embed,
+                                timestep[:, 0]
+                            ).permute(0, 2, 1, 3, 4)
+
+                        # logger.info("uncond_pred_noise_btchw sum: %s, uncond_pred_noise_btchw shape: %s, uncond_pred_noise_btchw dtype: %s", uncond_pred_noise_btchw.float().sum(), uncond_pred_noise_btchw.shape, uncond_pred_noise_btchw.dtype)
+
+                        uncond_pred_video_btchw = pred_noise_to_pred_video(
+                            pred_noise=uncond_pred_noise_btchw.flatten(0, 1),
+                            noise_input_latent=latents.flatten(0, 1),
+                            timestep=timestep.flatten(0, 1),
+                            scheduler=self.get_module("scheduler")).unflatten(
+                                0, uncond_pred_noise_btchw.shape[:2])
+
+                        pred_video_btchw = uncond_pred_video_btchw + batch.guidance_scale * (
+                            cond_pred_video_btchw - uncond_pred_video_btchw
+                        )
+
+                        # logger.info("pred_video_btchw sum: %s, pred_video_btchw shape: %s, pred_video_btchw dtype: %s", pred_video_btchw.float().sum(), pred_video_btchw.shape, pred_video_btchw.dtype)
+
+                        pred_noise_btchw = pred_video_to_pred_noise(
+                            x0_pred=pred_video_btchw.flatten(0, 1),
+                            xt=latents.flatten(0, 1),
+                            timestep=timestep.flatten(0, 1),
+                            scheduler=self.get_module("scheduler")).unflatten(
+                                0, pred_video_btchw.shape[:2])
+
+                        # logger.info("pred_noise_btchw sum: %s, pred_noise_btchw shape: %s, pred_noise_btchw dtype: %s", pred_noise_btchw.float().sum(), pred_noise_btchw.shape, pred_noise_btchw.dtype)
+
+                        latents = self.get_module("scheduler").step(
+                            pred_noise_btchw.flatten(0, 1),
+                            self.get_module("scheduler").timesteps[progress_id] * torch.ones(
+                                [1, 21], device=latents.device, dtype=torch.long).flatten(0, 1),
+                            latents.flatten(0, 1)
+                        )[0].unflatten(dim=0, sizes=pred_noise_btchw.shape[:2])
+
+                        # logger.info("latents sum: %s, latents shape: %s, latents dtype: %s", latents.float().sum(), latents.shape, latents.dtype)
+
+                    noisy_input.append(latents)
+
+                    noisy_inputs = torch.stack(noisy_input, dim=1)
+
+                    noisy_inputs = noisy_inputs[:, [0, 12, 24, 36, -1]].half()
+
+                    logger.info("noisy inputs sum: %s, noisy inputs shape: %s, noisy inputs dtype: %s", noisy_inputs.float().sum(), noisy_inputs.shape, noisy_inputs.dtype)
+
+                    result_batch.trajectory_latents = noisy_inputs.permute(0, 1, 3, 2, 4, 5)
+                    result_batch.trajectory_timesteps = torch.tensor([self.get_module("scheduler").timesteps[i] for i in [0, 12, 24, 36, -1]])
+                    result_batch.latents = latents.permute(0, 2, 1, 3, 4)
                    result_batch = self.decoding_stage(result_batch,
                                                       fastvideo_args)
-
                    trajectory_latents.append(
                        result_batch.trajectory_latents.cpu())
                    trajectory_timesteps.append(
                        result_batch.trajectory_timesteps.cpu())
                    trajectory_decoded.append(result_batch.trajectory_decoded)

+                trajectory_latents = torch.stack(trajectory_latents, dim=0).squeeze(0)
+                # trajecotry_latents = trajectory_latents[:, [0, 12, 24, 36, -1]]
+
                # Prepare extra features for text-only processing
                extra_features = {
                    "trajectory_latents": trajectory_latents,
@@ -78,6 +78,8 @@ 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)

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

    def forward(
@@ -141,11 +140,12 @@ class DenoisingStage(PipelineStage):
            latents = latents[:, :, rank_in_sp_group, :, :, :]
            batch.latents = latents
            if batch.image_latent is not None:
-                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
+                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
        # Get timesteps and calculate warmup steps
        timesteps = batch.timesteps
        # TODO(will): remove this once we add input/output validation for stages
@@ -254,8 +254,7 @@ class DenoisingStage(PipelineStage):
                                                             patch_size[2])
            seq_len = int(math.ceil(seq_len / sp_world_size)) * sp_world_size

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

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

                # Expand latents for I2V
                latent_model_input = latents.to(target_dtype)
-                if batch.image_latent is not None:
+                if batch.image_latent is not None and not fastvideo_args.pipeline_config.t2v_as_i2v_task:
                    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()
@@ -307,6 +319,13 @@ 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 = (
@@ -469,10 +488,17 @@ class DenoisingStage(PipelineStage):
                trajectory_tensor = sequence_model_parallel_all_gather(
                    trajectory_tensor, dim=3)

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

@@ -105,6 +105,81 @@ 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,
@@ -157,58 +232,29 @@ class ImageVAEEncodingStage(PipelineStage):
        # (B, C, H, W) -> (B, C, 1, H, W)
        image = image.unsqueeze(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)
+        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)

-        # 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
+        latent_condition = self.encode_tensor(video_condition, fastvideo_args,
+                                              batch.generator)

        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,
                                       latent_width)
@@ -35,9 +35,15 @@ 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(num_videos_per_prompt)]
+        seeds = [seed + i for i in range(total_num_videos)]
        batch.seeds = seeds
        # Peiyuan: using GPU seed will cause A100 and H100 to generate different results...
        batch.generator = [
@@ -3,6 +3,7 @@
 Latent preparation stage for diffusion pipelines.
 """

+import torch
 from diffusers.utils.torch_utils import randn_tensor

 from fastvideo.distributed import get_local_torch_device
@@ -93,7 +94,7 @@ class LatentPreparationStage(PipelineStage):
        # Generate or use provided latents
        if latents is None:
            latents = randn_tensor(shape,
-                                   generator=generator,
+                                   generator=torch.Generator(device="cuda").manual_seed(1024),
                                   device=device,
                                   dtype=dtype)
        else:
@@ -82,8 +82,8 @@ def rocm_platform_plugin() -> str | None:
                logger.info("ROCm platform is available")
        finally:
            amdsmi.amdsmi_shut_down()
-    except Exception as e:
-        logger.info("ROCm platform is unavailable: %s", e)
+    except Exception:
+        pass

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

@@ -115,7 +115,6 @@ 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
@@ -145,20 +144,6 @@ 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
@@ -18,7 +18,6 @@ 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()
@@ -8,6 +8,9 @@ from torch.distributed.tensor import DTensor
 from torch.testing import assert_close
 from transformers import AutoConfig, AutoTokenizer, UMT5EncoderModel

+from fastvideo.wan.modules.tokenizers import HuggingfaceTokenizer
+from fastvideo.wan.modules.t5 import umt5_xxl
+
 from fastvideo.configs.pipelines import PipelineConfig
 from fastvideo.forward_context import set_forward_context
 from fastvideo.logger import init_logger
@@ -38,9 +41,22 @@ def test_t5_encoder():
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    precision_str = "fp32"
    precision = PRECISION_TO_TYPE[precision_str]
-    model1 = UMT5EncoderModel.from_pretrained(TEXT_ENCODER_PATH).to(
-        precision).to(device).eval()
-    tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_PATH)
+    # model1 = UMT5EncoderModel.from_pretrained(TEXT_ENCODER_PATH).to(
+    #     precision).to(device).eval()
+    # tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_PATH)
+    model1 = umt5_xxl(
+        encoder_only=True,
+        return_tokenizer=False,
+        dtype=torch.float32,
+        device=device,
+    ).eval().requires_grad_(False)
+    model1.load_state_dict(
+        torch.load("/mnt/weka/home/hao.zhang/wei/Self-Forcing-clean/wan_models/Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth",
+                    map_location='cpu', weights_only=False)
+    )
+
+    tokenizer1 = HuggingfaceTokenizer(
+        name="/mnt/weka/home/hao.zhang/wei/Self-Forcing-clean/wan_models/Wan2.1-T2V-1.3B/google/umt5-xxl/", seq_len=512, clean='whitespace')


    args = FastVideoArgs(model_path=TEXT_ENCODER_PATH,
@@ -49,8 +65,9 @@ def test_t5_encoder():
                        pin_cpu_memory=False)
    loader = TextEncoderLoader()
    model2 = loader.load(TEXT_ENCODER_PATH, args)
-    model2 = model2.to(precision)
+    model2 = model2.to(dtype=torch.bfloat16).to(precision)
    model2.eval()
+    tokenizer2 = AutoTokenizer.from_pretrained(TOKENIZER_PATH)

    # Sanity check weights between the two models
    logger.info("Comparing model weights for sanity check...")
@@ -61,8 +78,13 @@ def test_t5_encoder():
    logger.info("Model1 has %s parameters", len(params1))
    logger.info("Model2 has %s parameters", len(params2))

-    weight_diffs = []
-    # check if embed_tokens are the same
+    model1_weight_sum = sum(p.float().sum().item() for p in model1.parameters())
+    model2_weight_sum = sum(p.float().sum().item() for p in model2.parameters())
+    logger.info("Model1 weight sum: %s", model1_weight_sum)
+    logger.info("Model2 weight sum: %s", model2_weight_sum)
+
+    # weight_diffs = []
+    # # check if embed_tokens are the same
    weights = ["encoder.block.{}.layer.0.layer_norm.weight", "encoder.block.{}.layer.0.SelfAttention.relative_attention_bias.weight", \
               "encoder.block.{}.layer.0.SelfAttention.o.weight", "encoder.block.{}.layer.1.DenseReluDense.wi_0.weight", "encoder.block.{}.layer.1.DenseReluDense.wi_1.weight",\
                "encoder.block.{}.layer.1.DenseReluDense.wo.weight", \
@@ -70,18 +92,21 @@ def test_t5_encoder():
    
    for idx in range(hf_config.num_hidden_layers):
        for w in weights:
-            name1 = w.format(idx)
+            # name1 = w.format(idx)
            name2 = w.format(idx)
-            p1 = params1[name1]
+            # p1 = params1[name1]
            p2 = params2[name2]
-            p2 = (p2.to_local() if isinstance(p2, DTensor) else p2).to(p1)
-            assert_close(p1, p2, atol=1e-4, rtol=1e-4)
+            p2 = (p2.to_local() if isinstance(p2, DTensor) else p2).to(device)
+            # assert_close(p1, p2, atol=1e-4, rtol=1e-4)
    

    # Test with some sample prompts
+    # prompts = [
+    #     "Once upon a time", "The quick brown fox jumps over",
+    #     "In a galaxy far, far away"
+    # ]
    prompts = [
-        "Once upon a time", "The quick brown fox jumps over",
-        "In a galaxy far, far away"
+        "A vibrant scene of Kenyan golfers at a lush green golf course on a sunny day. The golfers, dressed in casual yet stylish attire, are teeing off with animated expressions, showcasing their enthusiasm for the game. Rolling hills and pristine greens stretch out behind them, creating a picturesque backdrop. In the foreground, a golf buggy and a caddy stand ready, adding to the serene atmosphere. The camera captures the action from a mid-shot angle, focusing on the golfers' dynamic motions as they swing their clubs."
    ]

    logger.info("Testing T5 encoder with sample prompts")
@@ -91,7 +116,8 @@ def test_t5_encoder():
            logger.info("Testing prompt: %s", prompt)

            # Tokenize the prompt
-            tokens = tokenizer(prompt,
+            tokens1, mask = tokenizer1(prompt, return_mask=True, add_special_tokens=True)
+            tokens2 = tokenizer2(prompt,
                               padding="max_length",
                               max_length=512,
                               truncation=True,
@@ -101,22 +127,23 @@ def test_t5_encoder():
            # filter out padding input_ids
            # tokens.input_ids = tokens.input_ids[tokens.attention_mask==1]
            # tokens.attention_mask = tokens.attention_mask[tokens.attention_mask==1]
-            outputs1 = model1(input_ids=tokens.input_ids,
-                              attention_mask=tokens.attention_mask,
-                              output_hidden_states=True).last_hidden_state
+            outputs1 = model1(tokens1.to(device),
+                              mask.to(device))
            print("--------------------------------")
            logger.info("Testing model2")

            # Get outputs from our implementation
            with set_forward_context(current_timestep=0, attn_metadata=None):
                outputs2 = model2(
-                    input_ids=tokens.input_ids,
-                    attention_mask=tokens.attention_mask,
+                    input_ids=tokens2.input_ids,
+                    attention_mask=tokens2.attention_mask,
                ).last_hidden_state

            # Compare last hidden states
-            last_hidden_state1 = outputs1[tokens.attention_mask == 1]
-            last_hidden_state2 = outputs2[tokens.attention_mask == 1]
+            last_hidden_state1 = outputs1[mask == 1]
+            last_hidden_state2 = outputs2[tokens2.attention_mask == 1]
+            logger.info("last_hidden_state1 sum: %s", last_hidden_state1.float().sum())
+            logger.info("last_hidden_state2 sum: %s", last_hidden_state2.float().sum())

            assert last_hidden_state1.shape == last_hidden_state2.shape, \
                f"Hidden state shapes don't match: {last_hidden_state1.shape} vs {last_hidden_state2.shape}"
@@ -6,7 +6,7 @@ import pytest

 from fastvideo import VideoGenerator
 from fastvideo.logger import init_logger
-from tests.utils import compute_video_ssim_torchvision, write_ssim_results
+from fastvideo.tests.utils import compute_video_ssim_torchvision, write_ssim_results
 from diffusers import DiffusionPipeline
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.pipelines import build_pipeline
@@ -192,3 +192,4 @@ def test_lora_inference_similarity(ATTENTION_BACKEND, model_id):



+
@@ -64,35 +64,35 @@ def run_test(pytest_command: str):

@app.function(gpu="L40S:1", image=image, timeout=900)
 def run_encoder_tests():
-    run_test("pytest ./tests/encoders -vs")
+    run_test("pytest ./fastvideo/tests/encoders -vs")

@app.function(gpu="L40S:1", image=image, timeout=900)
 def run_vae_tests():
-    run_test("pytest ./tests/vaes -vs")
+    run_test("pytest ./fastvideo/tests/vaes -vs")

@app.function(gpu="L40S:1", image=image, timeout=900)
 def run_transformer_tests():
-    run_test("pytest ./tests/transformers -vs")
+    run_test("pytest ./fastvideo/tests/transformers -vs")

@app.function(gpu="L40S:2", image=image, timeout=2700)
 def run_ssim_tests():
-    run_test("pytest ./tests/ssim -vs")
+    run_test("pytest ./fastvideo/tests/ssim -vs")

@app.function(gpu="L40S:4", image=image, timeout=900, secrets=[modal.Secret.from_dict({"WANDB_API_KEY": os.environ.get("WANDB_API_KEY", "")})])
 def run_training_tests():
-    run_test("wandb login $WANDB_API_KEY && pytest ./tests/training/Vanilla -srP")
+    run_test("wandb login $WANDB_API_KEY && pytest ./fastvideo/tests/training/Vanilla -srP")

@app.function(gpu="L40S:2", image=image, timeout=900, secrets=[modal.Secret.from_dict({"WANDB_API_KEY": os.environ.get("WANDB_API_KEY", "")})])
 def run_training_lora_tests():
-    run_test("wandb login $WANDB_API_KEY && pytest ./tests/training/lora/test_lora_training.py -srP")
+    run_test("wandb login $WANDB_API_KEY && pytest ./fastvideo/tests/training/lora/test_lora_training.py -srP")

@app.function(gpu="H100:2", image=image, timeout=900, secrets=[modal.Secret.from_dict({"WANDB_API_KEY": os.environ.get("WANDB_API_KEY", "")})])
 def run_training_tests_VSA():
-    run_test("wandb login $WANDB_API_KEY && pytest ./tests/training/VSA -srP")
+    run_test("wandb login $WANDB_API_KEY && pytest ./fastvideo/tests/training/VSA -srP")

@app.function(gpu="H100:2", image=image, timeout=900)
 def run_inference_tests_STA():
-    run_test("pytest ./tests/inference/STA -srP")
+    run_test("pytest ./fastvideo/tests/inference/STA -srP")

@app.function(gpu="H100:1", image=image, timeout=900)
 def run_precision_tests_STA():
@@ -108,16 +108,16 @@ def run_precision_tests_vmoba():

@app.function(gpu="L40S:1", image=image, timeout=900)
 def run_inference_tests_vmoba():
-    run_test('python tests/inference/vmoba/test_vmoba_inference.py')
+    run_test('python fastvideo/tests/inference/vmoba/test_vmoba_inference.py')

@app.function(gpu="L40S:1", image=image, timeout=3600)
 def run_inference_lora_tests():
-    run_test("pytest ./tests/inference/lora/test_lora_inference_similarity.py -vs")
+    run_test("pytest ./fastvideo/tests/inference/lora/test_lora_inference_similarity.py -vs")

@app.function(gpu="L40S:2", image=image, timeout=900)
 def run_distill_dmd_tests():
-    run_test("pytest ./tests/training/distill/test_distill_dmd.py -vs")
+    run_test("pytest ./fastvideo/tests/training/distill/test_distill_dmd.py -vs")

@app.function(gpu="L40S:1", image=image, timeout=900)
 def run_unit_test():
-    run_test("pytest ./tests/dataset/ ./tests/workflow/ -vs")
+    run_test("pytest ./fastvideo/tests/dataset/ ./fastvideo/tests/workflow/ -vs")
@@ -4,7 +4,7 @@ from huggingface_hub import snapshot_download
 import shutil
 import subprocess
 import sys
-from tests.ssim.test_inference_similarity import compute_video_ssim_torchvision
+from fastvideo.tests.ssim.test_inference_similarity import compute_video_ssim_torchvision

 # Import the training pipeline
 sys.path.append(str(Path(__file__).parent.parent.parent.parent.parent))
@@ -169,4 +169,4 @@ def test_e2e_overfit_single_sample():


 if __name__ == "__main__":
-    test_e2e_overfit_single_sample()
+    test_e2e_overfit_single_sample()
@@ -4,7 +4,7 @@ from huggingface_hub import snapshot_download
 import shutil
 import subprocess
 import sys
-from tests.ssim.test_inference_similarity import compute_video_ssim_torchvision
+from fastvideo.tests.ssim.test_inference_similarity import compute_video_ssim_torchvision

 # Import the training pipeline
 sys.path.append(str(Path(__file__).parent.parent.parent.parent.parent))
@@ -62,11 +62,6 @@ def download_data():


 def run_preprocessing():
-    # remove the local_preprocessed_data_dir if it exists
-    if LOCAL_PREPROCESSED_DATA_DIR.exists():
-        print(f"Removing local_preprocessed_data_dir: {LOCAL_PREPROCESSED_DATA_DIR}")
-        shutil.rmtree(LOCAL_PREPROCESSED_DATA_DIR)
-
    # Run torchrun command
    cmd = [
        "torchrun",
@@ -174,4 +169,4 @@ def test_e2e_overfit_single_sample():


 if __name__ == "__main__":
-    test_e2e_overfit_single_sample()
+    test_e2e_overfit_single_sample()
@@ -10,10 +10,10 @@ torchrun --standalone --nnodes=1 --nproc_per_node=$num_gpus \
 num_gpus=2
 torchrun --standalone --nnodes=1 --nproc_per_node=$num_gpus \
    --master_port 29503  \
-    tests/test_hunyuanvideo_load.py --sequence_model_parallel_size $num_gpus 
+    fastvideo/tests/test_hunyuanvideo_load.py --sequence_model_parallel_size $num_gpus 

-torchrun --nnodes=1 --nproc_per_node=1 --master_port 29503  tests/test_llama_encoder.py
+torchrun --nnodes=1 --nproc_per_node=1 --master_port 29503  fastvideo/tests/test_llama_encoder.py


 export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
-torchrun --nnodes=1 --nproc_per_node=1 --master_port 29503  tests/test_clip_encoder.py
+torchrun --nnodes=1 --nproc_per_node=1 --master_port 29503  fastvideo/tests/test_clip_encoder.py
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
JerryZhou54	92dfbfe4fc	Refactor ode trajectory generation Add env variable to enforce running flash attn	2025-09-26 01:26:03 +00:00
JerryZhou54	d31695b736	Test	2025-09-22 22:33:18 +00:00
JerryZhou54	357eaa4f1e	small change	2025-09-20 05:02:06 +00:00
JerryZhou54	ba02d0ae7e	Numerical Debugging	2025-09-20 05:02:06 +00:00
SolitaryThinker	c499783f74	fix ode_init	2025-09-20 05:02:06 +00:00
JerryZhou54	586013575c	Resolve timestep mismatch between dit forward and pred_noise_to_pred_video	2025-09-20 05:02:05 +00:00
SolitaryThinker	97e33f9097	debug	2025-09-20 05:02:05 +00:00
SolitaryThinker	8674224232	comment out vmoba	2025-09-20 05:02:04 +00:00
JerryZhou54	24d35b5512	Fix test for forward_train	2025-09-20 05:01:27 +00:00
JerryZhou54	3d2af0823c	Ensure 0 numerical diff for forward_train	2025-09-20 05:01:27 +00:00
JerryZhou54	27bac993ac	Add test for _forward_train	2025-09-20 05:01:27 +00:00
JerryZhou54	15a89d3f27	Fix numerical diff between causal_wanvideo.py and SF's causal wan	2025-09-20 05:01:25 +00:00
RandNMR73	ddf5e76ceb	text preprocessing ready	2025-09-20 05:00:53 +00:00
JerryZhou54	1f78d8114e	checkpoint	2025-09-20 05:00:06 +00:00
JerryZhou54	8cc2c113b8	Enable timestep warping & using SelfForcing scheduler	2025-09-20 04:59:33 +00:00
JerryZhou54	0b1519db19	Stop backprop through kv_cache	2025-09-20 04:59:33 +00:00
RandNMR73	055e7dc47b	new branch	2025-09-20 04:59:31 +00:00
RandNMR73	0892c2c448	inference works after changes added	2025-09-20 04:57:54 +00:00
SolitaryThinker	c501a484e0	lmdb datasets	2025-09-20 04:57:54 +00:00
SolitaryThinker	d965f578c9	t2v to i2v finetune checkpoint ode checkpoint fix t2v to i2v lint chekpt checkpoint hacked but working ode_init scripts WIP fixing time embedding WIP fixing time embedding checkpoint update fix revert revert revert update update visualize	2025-09-20 04:57:52 +00:00