wip

Add Sage Attention 3 Backend (#815 )
[self-forcing] [6/n] Add Ode Init training (#811 )
2025-09-28 05:42:34 +00:00 · 2025-09-24 15:11:38 -07:00 · 2025-09-22 17:58:19 -07:00 · 2025-09-22 12:07:03 -07:00 · 2025-09-20 19:32:10 -07:00 · 2025-09-20 18:15:59 -07:00
107 changed files with 3430 additions and 220 deletions
@@ -18,7 +18,8 @@ steps:
            - path: 
                - "fastvideo/models/encoders/**"
                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/encoders/**"
+                - "tests/encoders/**"
+                - "tests/utils.py"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -31,7 +32,8 @@ steps:
            - path:
                - "fastvideo/models/vaes/**"
                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/vaes/**"
+                - "tests/vaes/**"
+                - "tests/utils.py"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -44,7 +46,8 @@ steps:
            - path:
                - "fastvideo/models/dits/**"
                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/transformers/**"
+                - "tests/transformers/**"
+                - "tests/utils.py"
                - "fastvideo/layers/**"
                - "fastvideo/attention/**"
                - "pyproject.toml"
@@ -58,6 +61,8 @@ steps:
                  queue: "default"
            - path: 
                - "fastvideo/**/*.py"
+                - "tests/ssim/**"
+                - "tests/utils.py"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -68,9 +73,10 @@ steps:
                agents:
                  queue: "default"
            - path:
-                - "fastvideo/tests/lora/**"
+                - "tests/inference/lora/**"
                - "fastvideo/models/loader/**"
-                - "fastvideo/tests/transformers/**"
+                - "tests/transformers/**"
+                - "tests/utils.py"
                - "fastvideo/pipelines/**"
                - "fastvideo/layers/lora/**"
                - "pyproject.toml"
@@ -84,6 +90,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/training/Vanilla/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -95,6 +102,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/training/*distillation_pipeline.py"
+                - "tests/training/distill/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -106,6 +114,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/training/lora/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -117,6 +126,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/training/VSA/**"
                - "csrc/attn/video_sparse_attn/**"
                - "csrc/attn/video_sparse_attn/tk/**"
                - "csrc/attn/video_sparse_attn/setup.py"
@@ -133,6 +143,7 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/inference/STA/**"
                - "csrc/attn/sliding_tile_attn/**"
                - "csrc/attn/sliding_tile_attn/setup.py"
                - "csrc/attn/sliding_tile_attn/config_sta.py"
@@ -190,6 +201,7 @@ steps:
            - path:
                - "csrc/attn/vmoba_attn/vmoba/**"
                - "fastvideo/attention/backends/vmoba.py"
+                - "tests/inference/vmoba/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -201,6 +213,8 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
+                - "tests/dataset/**"
+                - "tests/workflow/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -50,7 +50,7 @@ else
    exit 1
 fi

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

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

  encoder-test:
@@ -174,7 +182,7 @@ jobs:
      gpu_count: 1
      volume_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/encoders -s"
+      test_command: "uv pip install -e .[test] && pytest ./tests/encoders -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -192,7 +200,7 @@ jobs:
      gpu_count: 1
      volume_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/vaes -s"
+      test_command: "uv pip install -e .[test] && pytest ./tests/vaes -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -210,7 +218,7 @@ jobs:
      gpu_count: 1
      volume_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/transformers -s"
+      test_command: "uv pip install -e .[test] && pytest ./tests/transformers -s"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -236,7 +244,7 @@ jobs:
      volume_size: 200
      disk_size: 200
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:${{ matrix.python-version.tag }}"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/ssim -vs"
+      test_command: "uv pip install -e .[test] && pytest ./tests/ssim -vs"
      timeout_minutes: 60
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -255,7 +263,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/training/Vanilla -srP"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/training/Vanilla -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -275,7 +283,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/training/VSA -srP"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/training/VSA -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -295,7 +303,7 @@ jobs:
      volume_size: 100
      disk_size: 100
      image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/inference/STA -srP"
+      test_command: "uv pip install -e .[test] && pytest ./tests/inference/STA -srP"
      timeout_minutes: 30
    secrets:
      RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -369,7 +377,7 @@ jobs:
  #     volume_size: 100
  #     disk_size: 100
  #     image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-  #     test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/nightly/test_e2e_overfit_single_sample.py -vs"
+#     test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/nightly/test_e2e_overfit_single_sample.py -vs"
  #     timeout_minutes: 30
  #   secrets:
  #     RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -12,9 +12,6 @@ exclude: |
    scripts/.*|
    fastvideo/data_preprocess/.*|
    fastvideo/dataset/.*|
-    fastvideo/distill/.*|
-    fastvideo/distill\.py|
-    fastvideo/distill_adv\.py|
    fastvideo/models/.*|
    fastvideo/sample/.*|
    fastvideo/train\.py|
@@ -44,10 +41,10 @@ repos:
  - id: codespell
    additional_dependencies: ['tomli']
    args: ['--toml', 'pyproject.toml']
- repo: https://github.com/PyCQA/isort
-  rev: 6.0.1
-  hooks:
-  - id: isort
+# - repo: https://github.com/PyCQA/isort
+#   rev: 6.0.1
+#   hooks:
+#   - id: isort
 - repo: https://github.com/jackdewinter/pymarkdown
  rev: v0.9.30
  hooks:
@@ -7,7 +7,7 @@
 FastVideo features an end-to-end unified pipeline for accelerating diffusion models, starting from data preprocessing to model training, finetuning, distillation, and inference. FastVideo is designed to be modular and extensible, allowing users to easily add new optimizations and techniques. Whether it is training-free optimizations or post-training optimizations, FastVideo has you covered.

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

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

@article{zhang2025vsa,
-  title={VSA: Faster Video Diffusion with Trainable Sparse Attention},
-  author={Zhang, Peiyuan and Huang, Haofeng and Chen, Yongqi and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
+  title={Vsa: Faster video diffusion with trainable sparse attention},
+  author={Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
  journal={arXiv preprint arXiv:2505.13389},
  year={2025}
 }
@@ -1,32 +1,40 @@
 (inference-optimizations)=
+
 # Optimizations

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

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

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

 (optimizations-backends)=
+
 ## Attention Backends

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

 ### Configuring Backends

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

 ### What is TeaCache?
@@ -0,0 +1,151 @@
+#!/bin/bash
+#SBATCH --job-name=t2v
+#SBATCH --partition=main
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:1
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=dmd_t2v_output/t2v_%j.out
+#SBATCH --error=dmd_t2v_output/t2v_%j.err
+#SBATCH --exclusive
+
+# Basic Info
+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 TOKENIZERS_PARALLELISM=false
+export WANDB_API_KEY=your_wandb_api_key
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+# Configs
+NUM_GPUS=1
+
+# Model paths for Self-Forcing DMD distillation:
+GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
+
+DATA_DIR=your_data_dir
+VALIDATION_DATASET_FILE=your_validation_data_dir
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
+  --output_dir your_output_dir
+  --max_train_steps 4000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 81  # Must be divisible by num_frame_per_block (81 % 3 = 0 ✓)
+  --enable_gradient_checkpointing_type "full"
+  --log_visualization
+  --simulate_generator_forward
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
+  --enable_gradient_masking
+  --gradient_mask_last_n_frames 21
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_GPUS
+)
+
+# Model arguments
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --generator_model_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "4"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 1e-5
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 5
+  --seed 1000
+  --use_ema True
+  --ema_decay 0.99
+  --ema_start_step 100
+  --init_weights_from_safetensors your_ode_init_weights_path
+)
+
+# Self-forcing DMD arguments
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --dfake_gen_update_ratio 5
+  --real_score_guidance_scale 3.0
+  --fake_score_learning_rate 8e-6
+  --fake_score_betas '0.0,0.999'
+  --warp_denoising_step
+)
+
+# Self-forcing specific arguments
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+  --validate_cache_structure False  # Set to True for debugging KV cache issues
+)
+
+torchrun \
+--nnodes 1 \
+--master_port $MASTER_PORT \
+--nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"
@@ -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 +1,24 @@
+#!/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/"
+
+torchrun --nproc_per_node=$GPU_NUM \
+    fastvideo/pipelines/preprocess/v1_preprocess.py \
+    --model_path $MODEL_PATH \
+    --data_merge_path $DATA_MERGE_PATH \
+    --preprocess_video_batch_size 8 \
+    --seed 42 \
+    --max_height 480 \
+    --max_width 832 \
+    --num_frames 81 \
+    --dataloader_num_workers 0 \
+    --output_dir=$OUTPUT_DIR \
+    --train_fps 16 \
+    --samples_per_file 8 \
+    --flush_frequency 8 \
+    --video_length_tolerance_range 5 \
+    --preprocess_task "t2v" 
@@ -41,13 +41,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
  --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir"checkpoints/wan_t2v_finetune"
+  --output_dir $OUTPUT_DIR
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(

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

 # Training arguments
 training_args=(
  --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir "checkpoints/wan_t2v_finetune"
+  --output_dir "$OUTPUT_DIR"
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
  --weight_only_checkpointing_steps 500
@@ -134,4 +135,4 @@ srun torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -41,13 +41,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
  --tracker_project_name wan_t2v_distill_dmd
-  --output_dir "checkpoints/wan_t2v_finetune"
+  --output_dir "$OUTPUT_DIR"
  --max_train_steps 4000
  --train_batch_size 1
  --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
  --weight_only_checkpointing_steps 500
@@ -133,4 +134,4 @@ srun torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -1,3 +1,3 @@
 #!/bin/bash

-python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"
@@ -42,13 +42,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DIR=your_validation_path  #(example:validation_64.json)
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

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

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

 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-5
+  --learning_rate 4e-6
  --lr_scheduler "cosine_with_min_lr"
  --min_lr_ratio 0.5
  --lr_warmup_steps 100
-  --fake_score_learning_rate 1e-5
+  --fake_score_learning_rate 2e-6
  --fake_score_lr_scheduler "cosine_with_min_lr"
  --mixed_precision "bf16"
  --training_state_checkpointing_steps 500
@@ -142,4 +142,4 @@ srun torchrun \
    "${optimizer_args[@]}" \
    "${validation_args[@]}" \
    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"
@@ -16,24 +16,25 @@ NUM_GPUS=1
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR="data/crush-smol_processed_ti2v/combined_parquet_dataset/"
 VALIDATION_DATASET_FILE="examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/validation.json"
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

 # Training arguments
 training_args=(
  --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir="checkpoints/wan_t2v_finetune"
-  --max_train_steps=4000
-  --train_batch_size=1
+  --output_dir "$OUTPUT_DIR"
+  --max_train_steps 4000
+  --train_batch_size 1
  --train_sp_batch_size 1
-  --gradient_accumulation_steps=1
+  --gradient_accumulation_steps 1
  --num_latent_t 31
  --num_height 704
  --num_width 1280
  --num_frames 121
  --enable_gradient_checkpointing_type "full"
-  --training_state_checkpointing_steps=500
-  --weight_only_checkpointing_steps=500
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
 )

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

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

-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
@@ -21,4 +21,4 @@ torchrun --nproc_per_node=$GPU_NUM \
    --samples_per_file 8 \
    --flush_frequency 8 \
    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "t2v" 
@@ -0,0 +1,135 @@
+#!/bin/bash
+#SBATCH --job-name=2e6B8_16kFV_ode_vidprom
+#SBATCH --partition=main
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=ode_vidprom16k/ode_vidprom8b16k_2e-6.out
+#SBATCH --error=ode_vidprom16k/ode_vidprom8b16k_2e-6.err
+#SBATCH --exclusive
+set -e -x
+
+# Environment Setup
+source ~/conda/miniconda/bin/activate
+conda activate your-conda-env
+
+export WANDB_MODE="online"
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${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_API_KEY=your-wandb-api-key
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+echo "MASTER_ADDR: $MASTER_ADDR"
+echo "NODE_RANK: $NODE_RANK"
+
+
+MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+DATA_DIR="your-data-dir"
+VALIDATION_DATASET_FILE="examples/training/consistency_finetune/causal_ode_init/validation.json"
+OUTPUT_DIR="your-output-dir"
+INIT_WEIGHTS_FROM_SAFETENSORS="your-init-weights-from-safetensors" # bidirectional weights from Wan2.1-T2V-1.3B-Diffusers
+NUM_GPUS=8
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_ode_init"
+  --output_dir $OUTPUT_DIR
+  --override_transformer_cls_name "CausalWanTransformer3DModel"
+  --wandb_run_name "vidprom_8b16k_ode_init_2e-6"
+  # --resume_from_checkpoint "ode_init_diffusers/"
+  --warp_denoising_step
+  --log_visualization
+  --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
+  --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 $INIT_WEIGHTS_FROM_SAFETENSORS
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 2e-6
+  --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
+srun torchrun \
+  --nnodes $SLURM_JOB_NUM_NODES \
+  --nproc_per_node $NUM_GPUS \
+  --node_rank $SLURM_PROCID \
+  --rdzv_backend=c10d \
+  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/ode_causal_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"
@@ -1,14 +1,5 @@
 {
  "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,
@@ -28,7 +19,52 @@
      "num_frames": 77
    },
    {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
+      "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,
@@ -0,0 +1,133 @@
+#!/bin/bash
+#SBATCH --job-name=t2v
+#SBATCH --partition=main
+#SBATCH --nodes=8
+#SBATCH --ntasks=8
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=VSA_t2v_output/t2v_%j.out
+#SBATCH --error=VSA_t2v_output/t2v_%j.err
+#SBATCH --exclusive
+set -e -x
+
+# Environment Setup
+source ~/conda/miniconda/bin/activate
+conda activate your_env
+
+# Basic Info
+export WANDB_MODE="online"
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
+export TOKENIZERS_PARALLELISM=false
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+echo "MASTER_ADDR: $MASTER_ADDR"
+echo "NODE_RANK: $NODE_RANK"
+
+# Configs
+NUM_GPUS=8
+MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+DATA_DIR=your_data_dir
+VALIDATION_DATASET_FILE=your_validation_dataset_file
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name wan_t2v_VSA
+  --output_dir "checkpoints/wan_t2v_finetune_VSA"
+  --max_train_steps 4000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 81
+  # --enable_gradient_checkpointing_type "full" # if OOM enable this
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 64
+  --hsdp_shard_dim 1
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file $VALIDATION_DATASET_FILE
+  --validation_steps 200
+  --validation_sampling_steps "50"
+  --validation_guidance_scale "5.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 1e-5
+  --mixed_precision "bf16"
+  --checkpointing_steps 1000
+  --weight_decay 0.01
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.1
+  --dit_precision "fp32"
+  --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 \
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_training_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${vsa_args[@]}"  
@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+from fastvideo.attention.backends.sageattn.api import sageattn_blackwell
+
+from fastvideo.attention.backends.abstract import (AttentionBackend,
+                                                   AttentionImpl,
+                                                   AttentionMetadata,
+                                                   AttentionMetadataBuilder)
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class SageAttention3Backend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [64, 128, 256]
+
+    @staticmethod
+    def get_name() -> str:
+        return "SAGE_ATTN_THREE"
+
+    @staticmethod
+    def get_impl_cls() -> type["SageAttention3Impl"]:
+        return SageAttention3Impl
+
+    @staticmethod
+    def get_metadata_cls() -> type["AttentionMetadata"]:
+        raise NotImplementedError
+
+    @staticmethod
+    def get_builder_cls() -> type["AttentionMetadataBuilder"]:
+        raise NotImplementedError
+
+    # @staticmethod
+    # def get_metadata_cls() -> Type["AttentionMetadata"]:
+    #     return FlashAttentionMetadata
+
+
+class SageAttention3Impl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout = extra_impl_args.get("dropout_p", 0.0)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+        output = sageattn_blackwell(query, key, value, is_causal=self.causal)
+        output = output.transpose(1, 2)
+        return output
@@ -15,13 +15,10 @@ class DiTArchConfig(ArchConfig):
    reverse_param_names_mapping: dict = field(default_factory=dict)
    lora_param_names_mapping: dict = field(default_factory=dict)
    _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
-        AttentionBackendEnum.SLIDING_TILE_ATTN,
-        AttentionBackendEnum.SAGE_ATTN,
-        AttentionBackendEnum.FLASH_ATTN,
-        AttentionBackendEnum.TORCH_SDPA,
-        AttentionBackendEnum.VIDEO_SPARSE_ATTN,
-        AttentionBackendEnum.VMOBA_ATTN,
-    )
+        AttentionBackendEnum.SLIDING_TILE_ATTN, AttentionBackendEnum.SAGE_ATTN,
+        AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA,
+        AttentionBackendEnum.VIDEO_SPARSE_ATTN, AttentionBackendEnum.VMOBA_ATTN,
+        AttentionBackendEnum.SAGE_ATTN_THREE)

    hidden_size: int = 0
    num_attention_heads: int = 0
@@ -175,6 +175,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
    # - "SLIDING_TILE_ATTN" : use Sliding Tile Attention
    # - "VIDEO_SPARSE_ATTN": use Video Sparse Attention
    # - "SAGE_ATTN": use Sage Attention
+    # - "SAGE_ATTN_THREE": use Sage Attention 3
    "FASTVIDEO_ATTENTION_BACKEND":
    lambda: os.getenv("FASTVIDEO_ATTENTION_BACKEND", None),

@@ -158,6 +158,7 @@ class FastVideoArgs:
        "transformer": True,
        "vae": True,
    })
+    override_transformer_cls_name: str | None = None

    # # DMD parameters
    # dmd_denoising_steps: List[int] | None = field(default=None)
@@ -396,6 +397,12 @@ class FastVideoArgs:
            default=FastVideoArgs.enable_stage_verification,
            help="Enable input/output verification for pipeline stages",
        )
+        parser.add_argument(
+            "--override-transformer-cls-name",
+            type=str,
+            default=FastVideoArgs.override_transformer_cls_name,
+            help="Override transformer cls name",
+        )
        # Add pipeline configuration arguments
        PipelineConfig.add_cli_args(parser)

@@ -605,6 +612,11 @@ class TrainingArgs(FastVideoArgs):
    pretrained_model_name_or_path: str = ""
    dit_model_name_or_path: str = ""

+    # DMD model paths - separate paths for each network
+    generator_model_path: str = ""  # path for generator (student) model
+    real_score_model_path: str = ""  # path for real score (teacher) model
+    fake_score_model_path: str = ""  # path for fake score (critic) model
+
    # diffusion setting
    ema_decay: float = 0.0
    ema_start_step: int = 0
@@ -627,6 +639,7 @@ class TrainingArgs(FastVideoArgs):
    checkpoints_total_limit: int = 0
    checkpointing_steps: int = 0
    resume_from_checkpoint: str = ""  # specify the checkpoint folder to resume from
+    init_weights_from_safetensors: str = ""  # path to safetensors file for initial weight loading

    # optimizer & scheduler
    num_train_epochs: int = 0
@@ -658,6 +671,7 @@ class TrainingArgs(FastVideoArgs):
    linear_quadratic_threshold: float = 0.0
    linear_range: float = 0.0
    weight_decay: float = 0.0
+    betas: str = "0.9,0.999"  # betas for optimizer, format: "beta1,beta2"
    use_ema: bool = False
    multi_phased_distill_schedule: str = ""
    pred_decay_weight: float = 0.0
@@ -678,16 +692,29 @@ class TrainingArgs(FastVideoArgs):

    # distillation args
    generator_update_interval: int = 5
+    dfake_gen_update_ratio: int = 5  # self-forcing: how often to train generator vs critic
    min_timestep_ratio: float = 0.2
    max_timestep_ratio: float = 0.98
    real_score_guidance_scale: float = 3.5
    fake_score_learning_rate: float = 0.0  # separate learning rate for fake_score_transformer, if 0.0, use learning_rate
    fake_score_lr_scheduler: str = "constant"  # separate lr scheduler for fake_score_transformer, if not set, use lr_scheduler
+    fake_score_betas: str = "0.9,0.999"  # betas for fake score optimizer, format: "beta1,beta2"
    training_state_checkpointing_steps: int = 0  # for resuming training
    weight_only_checkpointing_steps: int = 0  # for inference
    log_visualization: bool = False
    # simulate generator forward to match inference
    simulate_generator_forward: bool = False
+    warp_denoising_step: bool = False
+
+    # Self-forcing specific arguments
+    num_frame_per_block: int = 3
+    independent_first_frame: bool = False
+    enable_gradient_masking: bool = True
+    gradient_mask_last_n_frames: int = 21
+    validate_cache_structure: bool = False  # Debug flag for cache validation
+    same_step_across_blocks: bool = False  # Use same exit timestep for all blocks
+    last_step_only: bool = False  # Only use the last timestep for training
+    context_noise: int = 0  # Context noise level for cache updates

    @classmethod
    def from_cli_args(cls, args: argparse.Namespace) -> "TrainingArgs":
@@ -789,6 +816,20 @@ class TrainingArgs(FastVideoArgs):
                            type=str,
                            help="Directory to cache models")

+        # DMD model paths - separate paths for each network
+        parser.add_argument(
+            "--generator-model-path",
+            type=str,
+            help="Path to generator (student) model for DMD distillation")
+        parser.add_argument(
+            "--real-score-model-path",
+            type=str,
+            help="Path to real score (teacher) model for DMD distillation")
+        parser.add_argument(
+            "--fake-score-model-path",
+            type=str,
+            help="Path to fake score (critic) model for DMD distillation")
+
        # Diffusion settings
        parser.add_argument("--ema-decay",
                            type=float,
@@ -859,6 +900,10 @@ class TrainingArgs(FastVideoArgs):
        parser.add_argument("--resume-from-checkpoint",
                            type=str,
                            help="Path to checkpoint to resume from")
+        parser.add_argument(
+            "--init-weights-from-safetensors",
+            type=str,
+            help="Path to safetensors file for initial weight loading")
        parser.add_argument("--logging-dir",
                            type=str,
                            help="Directory for logging")
@@ -963,6 +1008,10 @@ class TrainingArgs(FastVideoArgs):
                            help="Linear quadratic threshold")
        parser.add_argument("--linear-range", type=float, help="Linear range")
        parser.add_argument("--weight-decay", type=float, help="Weight decay")
+        parser.add_argument("--betas",
+                            type=str,
+                            default=TrainingArgs.betas,
+                            help="Betas for optimizer (format: 'beta1,beta2')")
        parser.add_argument("--use-ema",
                            action=StoreBoolean,
                            help="Whether to use EMA")
@@ -1013,6 +1062,13 @@ class TrainingArgs(FastVideoArgs):
                            type=int,
                            default=TrainingArgs.generator_update_interval,
                            help="Ratio of student updates to critic updates.")
+        parser.add_argument(
+            "--dfake-gen-update-ratio",
+            type=int,
+            default=TrainingArgs.dfake_gen_update_ratio,
+            help=
+            "Self-forcing: How often to train generator vs critic (train generator every N steps)."
+        )
        parser.add_argument("--min-timestep-ratio",
                            type=float,
                            default=TrainingArgs.min_timestep_ratio,
@@ -1029,6 +1085,11 @@ class TrainingArgs(FastVideoArgs):
                            type=float,
                            default=TrainingArgs.fake_score_learning_rate,
                            help="Learning rate for fake score transformer")
+        parser.add_argument(
+            "--fake-score-betas",
+            type=str,
+            default=TrainingArgs.fake_score_betas,
+            help="Betas for fake score optimizer (format: 'beta1,beta2')")
        parser.add_argument(
            "--fake-score-lr-scheduler",
            type=str,
@@ -1041,6 +1102,48 @@ class TrainingArgs(FastVideoArgs):
            "--simulate-generator-forward",
            action=StoreBoolean,
            help="Whether to simulate generator forward to match inference")
+        parser.add_argument(
+            "--warp-denoising-step",
+            action=StoreBoolean,
+            help=
+            "Whether to warp denoising step according to the scheduler time shift"
+        )
+
+        # Self-forcing specific arguments
+        parser.add_argument(
+            "--num-frame-per-block",
+            type=int,
+            default=TrainingArgs.num_frame_per_block,
+            help="Number of frames per block for causal generation")
+        parser.add_argument(
+            "--independent-first-frame",
+            action=StoreBoolean,
+            help="Whether the first frame is independent in causal generation")
+        parser.add_argument(
+            "--enable-gradient-masking",
+            action=StoreBoolean,
+            help="Whether to enable frame-level gradient masking")
+        parser.add_argument(
+            "--gradient-mask-last-n-frames",
+            type=int,
+            default=TrainingArgs.gradient_mask_last_n_frames,
+            help="Number of last frames to enable gradients for")
+        parser.add_argument(
+            "--validate-cache-structure",
+            action=StoreBoolean,
+            help="Whether to validate KV cache structure (debug flag)")
+        parser.add_argument(
+            "--same-step-across-blocks",
+            action=StoreBoolean,
+            help="Whether to use the same exit timestep for all blocks")
+        parser.add_argument(
+            "--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")

        return parser

@@ -147,6 +147,9 @@ 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()
+                # 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
            x = self.attn(
@@ -375,7 +378,8 @@ class CausalWanTransformer3DModel(BaseDiT):

        # Causal-specific
        self.block_mask = None
-        self.num_frame_per_block = 1
+        self.num_frame_per_block = config.arch_config.num_frames_per_block
+        assert self.num_frame_per_block <= 3
        self.independent_first_frame = False

        self.__post_init__()
@@ -416,6 +416,11 @@ class TransformerLoader(ComponentLoader):
                "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
+            logger.info("Overriding transformer cls_name to %s", cls_name)
+
        fastvideo_args.model_paths["transformer"] = model_path

        # Config from Diffusers supersedes fastvideo's model config
@@ -430,8 +435,24 @@ class TransformerLoader(ComponentLoader):
        if not safetensors_list:
            raise ValueError(f"No safetensors files found in {model_path}")

-        logger.info("Loading model from %s safetensors files in %s",
-                    len(safetensors_list), model_path)
+        # Check if we should use custom initialization weights
+        custom_weights_path = getattr(fastvideo_args, 'init_weights_from_safetensors', None)
+        use_custom_weights = (custom_weights_path and os.path.exists(custom_weights_path) and 
+                            fastvideo_args.training_mode and 
+                            not hasattr(fastvideo_args, '_loading_teacher_critic_model'))
+
+        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]
+
+        logger.info("Loading model from %s safetensors files: %s",
+                    len(safetensors_list), safetensors_list)

        default_dtype = PRECISION_TO_TYPE[
            fastvideo_args.pipeline_config.dit_precision]
@@ -454,6 +475,7 @@ class TransformerLoader(ComponentLoader):
            pin_cpu_memory=fastvideo_args.pin_cpu_memory,
            fsdp_inference=fastvideo_args.use_fsdp_inference,
            # TODO(will): make these configurable
+            default_dtype=default_dtype,
            param_dtype=torch.bfloat16,
            reduce_dtype=torch.float32,
            output_dtype=None,
@@ -463,9 +485,8 @@ class TransformerLoader(ComponentLoader):
        total_params = sum(p.numel() for p in model.parameters())
        logger.info("Loaded model with %.2fB parameters", total_params / 1e9)

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

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

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

    # Check if we should use FSDP
@@ -125,7 +127,7 @@ def maybe_load_fsdp_model(
        model,
        weight_iterator,
        device,
-        param_dtype,
+        default_dtype,
        strict=True,
        cpu_offload=cpu_offload,
        param_names_mapping=param_names_mapping_fn,
@@ -635,8 +635,31 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin,
        noise: torch.Tensor,
        timestep: torch.IntTensor,
    ) -> torch.Tensor:
+
+        """
+        Args:
+            clean_latent: the clean latent with shape [B, C, H, W],
+                where B is batch_size or batch_size * num_frames
+            noise: the noise with shape [B, C, H, W]
+            timestep: the timestep with shape [1] or [bs * num_frames] or [bs, num_frames]
+
+        Returns:
+            the corrupted latent with shape [B, C, H, W]
+        """
+        # If timestep is [bs, num_frames]
+        if timestep.ndim == 2:
+            timestep = timestep.flatten(0, 1)
+            assert timestep.numel() == clean_latent.shape[0]
+        elif timestep.ndim == 1:
+            # If timestep is [1]
+            if timestep.shape[0] == 1:
+                timestep = timestep.expand(clean_latent.shape[0])
+            else:
+                assert timestep.numel() == clean_latent.shape[0]
+        else:
+            raise ValueError(f"[add_noise] Invalid timestep shape: {timestep.shape}")
+        # timestep shape should be [B]
        self.sigmas = self.sigmas.to(noise.device)
-        timestep = timestep.expand(clean_latent.shape[0])
        self.timesteps = self.timesteps.to(noise.device)
        timestep_id = torch.argmin(
            (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
@@ -22,8 +22,10 @@ class SelfForcingFlowMatchSchedulerOutput(BaseOutput):
    prev_sample: torch.FloatTensor

 class SelfForcingFlowMatchScheduler(BaseScheduler, ConfigMixin, SchedulerMixin):
-
+    
+    config_name = "scheduler_config.json"
    order = 1
+    @register_to_config
    def __init__(self, num_inference_steps=100, num_train_timesteps=1000, shift=3.0, sigma_max=1.0, sigma_min=0.003 / 1.002, inverse_timesteps=False, extra_one_step=False, reverse_sigmas=False, training=False):
        self.num_train_timesteps = num_train_timesteps
        self.shift = shift
@@ -171,10 +171,10 @@ def pred_noise_to_pred_video(pred_noise: torch.Tensor,
    # timestep shape should be [B]
    dtype = pred_noise.dtype
    device = pred_noise.device
-    pred_noise = pred_noise.float().to(device)
-    noise_input_latent = noise_input_latent.float().to(device)
-    sigmas = scheduler.sigmas.float().to(device)
-    timesteps = scheduler.timesteps.float().to(device)
+    pred_noise = pred_noise.double().to(device)
+    noise_input_latent = noise_input_latent.double().to(device)
+    sigmas = scheduler.sigmas.double().to(device)
+    timesteps = scheduler.timesteps.double().to(device)
    timestep_id = torch.argmin(
        (timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
    sigma_t = sigmas[timestep_id].reshape(-1, 1, 1, 1)
@@ -7,8 +7,6 @@ This module wires the causal DMD denoising stage into the modular pipeline.

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

 # isort: off
@@ -28,10 +26,6 @@ class WanCausalDMDPipeline(LoRAPipeline, ComposedPipelineBase):
        "text_encoder", "tokenizer", "vae", "transformer", "scheduler"
    ]

-    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
-        self.modules["scheduler"] = FlowMatchEulerDiscreteScheduler(
-            shift=fastvideo_args.pipeline_config.flow_shift)
-
    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs) -> None:
        """Set up pipeline stages with proper dependency injection."""

@@ -246,6 +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)
    fake_score_latent_vis_dict: dict[str, Any] = field(default_factory=dict)


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

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

        logger.info("Trying FASTVIDEO_ATTENTION_BACKEND=%s",
                    envs.FASTVIDEO_ATTENTION_BACKEND)
+        logger.info("Selected backend: %s", selected_backend)
        if selected_backend == AttentionBackendEnum.SLIDING_TILE_ATTN:
            try:
                from st_attn import sliding_tile_attention  # noqa: F401
@@ -144,6 +145,20 @@ class CudaPlatformBase(Platform):
                logger.info(
                    "Sage Attention backend is not installed. Fall back to Flash Attention."
                )
+        elif selected_backend == AttentionBackendEnum.SAGE_ATTN_THREE:
+            try:
+                from fastvideo.attention.backends.sageattn.api import sageattn_blackwell  # noqa: F401
+
+                from fastvideo.attention.backends.sage_attn3 import (  # noqa: F401
+                    SageAttention3Backend)
+                logger.info("Using Sage Attention 3 backend.")
+
+                return "fastvideo.attention.backends.sage_attn3.SageAttention3Backend"
+            except ImportError as e:
+                logger.info(e)
+                logger.info(
+                    "Sage Attention 3 backend is not installed. Fall back to Flash Attention."
+                )
        elif selected_backend == AttentionBackendEnum.VIDEO_SPARSE_ATTN:
            try:
                from vsa import block_sparse_attn  # noqa: F401
@@ -18,6 +18,7 @@ class AttentionBackendEnum(enum.Enum):
    SLIDING_TILE_ATTN = enum.auto()
    TORCH_SDPA = enum.auto()
    SAGE_ATTN = enum.auto()
+    SAGE_ATTN_THREE = enum.auto()
    VIDEO_SPARSE_ATTN = enum.auto()
    VMOBA_ATTN = enum.auto()
    NO_ATTENTION = enum.auto()
@@ -1 +0,0 @@
-{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":0.50390625,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.19960195198655128,"_runtime":107.325113071}
@@ -1 +0,0 @@
-{"train_loss":0.10545400530099869,"_step":5,"_wandb":{"runtime":35},"step_time":2.2575189135968685,"grad_norm":0.53125,"_runtime":35.701958502,"avg_step_time":2.4754387199878694,"_timestamp":1.7525528420185745e+09,"learning_rate":1e-06,"vsa_sparsity":0,"validation_videos_8_steps":{"videos":[{"caption":"A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","_type":"video-file","sha256":"ee83e83df073a648f89dcd288cccaed9af765e11fc35c77a6e0cb2ebaa1be5b0","size":475248,"path":"media/videos/validation_videos_8_steps_0_ee83e83df073a648f89d.mp4"},{"size":341490,"path":"media/videos/validation_videos_8_steps_0_d0b2758549d5c82845ca.mp4","caption":"A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","_type":"video-file","sha256":"d0b2758549d5c82845ca3c1ac0db6812877631495e77ac041acdc8ea31f0a5ee"},{"caption":"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.","_type":"video-file","sha256":"08638381f1607d6ab10684772be38a58b18628fd402a208b25f6af96e765d454","size":436814,"path":"media/videos/validation_videos_8_steps_0_08638381f1607d6ab106.mp4"}],"captions":["A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table."],"_type":"videos","count":3}}
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 import copy
 import gc
+import json
 import os
 import time
 from abc import abstractmethod
@@ -36,10 +37,11 @@ from fastvideo.training.activation_checkpoint import (
    apply_activation_checkpointing)
 from fastvideo.training.training_pipeline import TrainingPipeline
 from fastvideo.training.training_utils import (
-    clip_grad_norm_while_handling_failing_dtensor_cases, count_trainable,
+    EMA_FSDP, clip_grad_norm_while_handling_failing_dtensor_cases,
    get_scheduler, load_distillation_checkpoint, save_distillation_checkpoint,
    shift_timestep)
-from fastvideo.utils import is_vsa_available, set_random_seed
+from fastvideo.utils import (is_vsa_available, maybe_download_model,
+                             set_random_seed, verify_model_config_and_directory)

 import wandb  # isort: skip

@@ -69,18 +71,11 @@ class DistillationPipeline(TrainingPipeline):
    current_trainstep: int
    video_latent_shape: tuple[int, ...]
    video_latent_shape_sp: tuple[int, ...]
-    real_score_transformer: torch.nn.Module
-    fake_score_transformer: torch.nn.Module

    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
        raise RuntimeError(
            "create_pipeline_stages should not be called for training pipeline")

-    def set_trainable(self) -> None:
-        super().set_trainable()
-        self.modules["real_score_transformer"].requires_grad_(False)
-        self.modules["vae"].requires_grad_(False)
-
    def initialize_training_pipeline(self, training_args: TrainingArgs):
        """Initialize the distillation training pipeline with multiple models."""
        logger.info("Initializing distillation pipeline...")
@@ -89,14 +84,35 @@ class DistillationPipeline(TrainingPipeline):

        self.noise_scheduler = self.get_module("scheduler")
        self.vae = self.get_module("vae")
+        self.vae.requires_grad_(False)
+
        self.timestep_shift = self.training_args.pipeline_config.flow_shift
        self.noise_scheduler = FlowMatchEulerDiscreteScheduler(
            shift=self.timestep_shift)

-        # self.transformer is the generator model
-        self.real_score_transformer = self.get_module("real_score_transformer")
-        self.fake_score_transformer = self.get_module("fake_score_transformer")
+        if training_args.real_score_model_path:
+            logger.info("Loading real score transformer from: %s",
+                        training_args.real_score_model_path)
+            self.real_score_transformer = self.load_module_from_path(
+                training_args.real_score_model_path, "transformer",
+                training_args)
+        else:
+            self.real_score_transformer = self.get_module(
+                "real_score_transformer")
+
+        if training_args.fake_score_model_path:
+            logger.info("Loading fake score transformer from: %s",
+                        training_args.fake_score_model_path)
+            self.fake_score_transformer = self.load_module_from_path(
+                training_args.fake_score_model_path, "transformer",
+                training_args)
+        else:
+            self.fake_score_transformer = self.get_module(
+                "fake_score_transformer")
+
+        self.real_score_transformer.requires_grad_(False)
        self.real_score_transformer.eval()
+        self.fake_score_transformer.requires_grad_(True)
        self.fake_score_transformer.train()

        if training_args.enable_gradient_checkpointing_type is not None:
@@ -119,10 +135,13 @@ class DistillationPipeline(TrainingPipeline):
        if fake_score_lr == 0.0:
            fake_score_lr = training_args.learning_rate

+        betas_str = training_args.fake_score_betas
+        betas = tuple(float(x.strip()) for x in betas_str.split(","))
+
        self.fake_score_optimizer = torch.optim.AdamW(
            fake_score_params,
            lr=fake_score_lr,
-            betas=(0.9, 0.999),
+            betas=betas,
            weight_decay=training_args.weight_decay,
            eps=1e-8,
        )
@@ -150,8 +169,19 @@ class DistillationPipeline(TrainingPipeline):
            self.training_args.pipeline_config.dmd_denoising_steps,
            dtype=torch.long,
            device=get_local_torch_device())
-        logger.info("Distillation generator model to %s denoising steps",
-                    len(self.denoising_step_list))
+
+        if training_args.warp_denoising_step:  # Warp the denoising step according to the scheduler time shift
+            timesteps = torch.cat((self.noise_scheduler.timesteps.cpu(),
+                                   torch.tensor([0],
+                                                dtype=torch.float32))).cuda()
+            self.denoising_step_list = timesteps[1000 -
+                                                 self.denoising_step_list]
+            logger.info("Warping denoising_step_list")
+
+        self.denoising_step_list = self.denoising_step_list.to(
+            get_local_torch_device())
+        logger.info("Distillation generator model to %s denoising steps: %s",
+                    len(self.denoising_step_list), self.denoising_step_list)
        self.num_train_timestep = self.noise_scheduler.num_train_timesteps

        self.min_timestep = int(self.training_args.min_timestep_ratio *
@@ -161,6 +191,82 @@ class DistillationPipeline(TrainingPipeline):

        self.real_score_guidance_scale = self.training_args.real_score_guidance_scale

+        self.generator_ema: EMA_FSDP | None = None
+        if (self.training_args.ema_decay
+                is not None) and (self.training_args.ema_decay > 0.0):
+            self.generator_ema = EMA_FSDP(self.transformer,
+                                          decay=self.training_args.ema_decay)
+            logger.info("Initialized generator EMA with decay=%s",
+                        self.training_args.ema_decay)
+        else:
+            logger.info("Generator EMA disabled (ema_decay <= 0.0)")
+
+    def load_module_from_path(self, model_path: str, module_type: str,
+                              training_args: "TrainingArgs"):
+        """
+        Load a module from a specific path using the same loading logic as the pipeline.
+        
+        Args:
+            model_path: Path to the model
+            module_type: Type of module to load (e.g., "transformer")
+            training_args: Training arguments
+            
+        Returns:
+            The loaded module
+        """
+        logger.info("Loading %s from custom path: %s", module_type, model_path)
+        # Set flag to prevent custom weight loading for teacher/critic models
+        training_args._loading_teacher_critic_model = True
+
+        try:
+            from fastvideo.models.loader.component_loader import (
+                PipelineComponentLoader)
+
+            # Download the model if it's a Hugging Face model ID
+            local_model_path = maybe_download_model(model_path)
+            logger.info("Model downloaded/found at: %s", local_model_path)
+            config = verify_model_config_and_directory(local_model_path)
+
+            if module_type not in config:
+                if hasattr(self, '_extra_config_module_map'
+                           ) and module_type in self._extra_config_module_map:
+                    extra_module = self._extra_config_module_map[module_type]
+                    if extra_module in config:
+                        module_type = extra_module
+                        logger.info("Using %s for %s", extra_module,
+                                    module_type)
+                    else:
+                        raise ValueError(
+                            f"Module {module_type} not found in config at {local_model_path}"
+                        )
+                else:
+                    raise ValueError(
+                        f"Module {module_type} not found in config at {local_model_path}"
+                    )
+
+            module_info = config[module_type]
+            if module_info is None:
+                raise ValueError(
+                    f"Module {module_type} has null value in config at {local_model_path}"
+                )
+
+            transformers_or_diffusers, architecture = module_info
+            component_path = os.path.join(local_model_path, module_type)
+            module = PipelineComponentLoader.load_module(
+                module_name=module_type,
+                component_model_path=component_path,
+                transformers_or_diffusers=transformers_or_diffusers,
+                fastvideo_args=training_args,
+            )
+
+            logger.info("Successfully loaded %s from %s", module_type,
+                        component_path)
+            return module
+        finally:
+            # Always clean up the flag
+            if hasattr(training_args, '_loading_teacher_critic_model'):
+                delattr(training_args, '_loading_teacher_critic_model')
+
    @abstractmethod
    def initialize_validation_pipeline(self, training_args: TrainingArgs):
        """Initialize validation pipeline - must be implemented by subclasses."""
@@ -170,11 +276,117 @@ class DistillationPipeline(TrainingPipeline):
    def _prepare_distillation(self,
                              training_batch: TrainingBatch) -> TrainingBatch:
        """Prepare training environment for distillation."""
+        self.transformer.requires_grad_(True)
        self.transformer.train()
+        self.fake_score_transformer.requires_grad_(True)
        self.fake_score_transformer.train()

        return training_batch

+    def apply_ema_to_model(self, model):
+        """Apply EMA weights to the model for validation or inference."""
+        if self.generator_ema is not None:
+            with self.generator_ema.apply_to_model(model):
+                return model
+        return model
+
+    def get_ema_model_copy(self) -> torch.nn.Module | None:
+        """Get a copy of the model with EMA weights applied."""
+        if self.generator_ema is not None:
+            ema_model = copy.deepcopy(self.transformer)
+            self.generator_ema.copy_to_unwrapped(ema_model)
+            return ema_model
+        return None
+
+    def is_ema_ready(self, current_step: int | None = None):
+        """Check if EMA is ready for use (after ema_start_step)."""
+        if current_step is None:
+            current_step = getattr(self, 'current_trainstep', 0)
+        return (self.generator_ema is not None
+                and current_step >= self.training_args.ema_start_step)
+
+    def save_ema_weights(self, output_dir: str, step: int):
+        """Save EMA weights separately for inference purposes."""
+        if self.generator_ema is None:
+            logger.warning("Cannot save EMA weights: EMA not initialized")
+            return
+
+        if not self.is_ema_ready():
+            logger.warning(
+                "Cannot save EMA weights: EMA not ready yet (step < ema_start_step)"
+            )
+            return
+
+        try:
+            ema_model = self.get_ema_model_copy()
+            if ema_model is None:
+                logger.warning("Failed to create EMA model copy")
+                return
+
+            ema_save_dir = os.path.join(output_dir, f"ema_checkpoint-{step}")
+            os.makedirs(ema_save_dir, exist_ok=True)
+
+            # save as diffusers format
+            from safetensors.torch import save_file
+
+            from fastvideo.training.training_utils import (
+                custom_to_hf_state_dict, gather_state_dict_on_cpu_rank0)
+            cpu_state = gather_state_dict_on_cpu_rank0(ema_model, device=None)
+
+            if self.global_rank == 0:
+                weight_path = os.path.join(
+                    ema_save_dir, "diffusion_pytorch_model.safetensors")
+                diffusers_state_dict = custom_to_hf_state_dict(
+                    cpu_state, ema_model.reverse_param_names_mapping)
+                save_file(diffusers_state_dict, weight_path)
+
+                config_dict = ema_model.hf_config
+                if "dtype" in config_dict:
+                    del config_dict["dtype"]
+                config_path = os.path.join(ema_save_dir, "config.json")
+                with open(config_path, "w") as f:
+                    json.dump(config_dict, f, indent=4)
+
+                logger.info("EMA weights saved to %s", weight_path)
+
+            del ema_model
+
+        except Exception as e:
+            logger.error("Failed to save EMA weights: %s", str(e))
+
+    def get_ema_stats(self) -> dict[str, Any]:
+        """Get EMA statistics for monitoring."""
+        if self.generator_ema is None:
+            return {
+                "ema_enabled": False,
+                "ema_decay": None,
+                "ema_start_step": self.training_args.ema_start_step,
+                "ema_ready": False,
+                "ema_step": self.current_trainstep,
+            }
+
+        return {
+            "ema_enabled": True,
+            "ema_decay": self.training_args.ema_decay,
+            "ema_start_step": self.training_args.ema_start_step,
+            "ema_ready": self.is_ema_ready(),
+            "ema_step": self.current_trainstep,
+        }
+
+    def reset_ema(self):
+        """Reset EMA to current model weights."""
+        if self.generator_ema is not None:
+            logger.info("Resetting EMA to current model weights")
+            self.generator_ema.update(self.transformer)
+            # Force update to current weights by setting decay to 0 temporarily
+            original_decay = self.generator_ema.decay
+            self.generator_ema.decay = 0.0
+            self.generator_ema.update(self.transformer)
+            self.generator_ema.decay = original_decay
+            logger.info("EMA reset completed")
+        else:
+            logger.warning("Cannot reset EMA: EMA not initialized")
+
    def _build_distill_input_kwargs(
            self, noise_input: torch.Tensor, timestep: torch.Tensor,
            text_dict: dict[str, torch.Tensor] | None,
@@ -331,6 +543,7 @@ class DistillationPipeline(TrainingPipeline):
    def _dmd_forward(self, generator_pred_video: torch.Tensor,
                     training_batch: TrainingBatch) -> torch.Tensor:
        """Compute DMD (Diffusion Model Distillation) loss."""
+        original_latent = generator_pred_video
        with torch.no_grad():
            timestep = torch.randint(0,
                                     self.num_train_timestep, [1],
@@ -355,7 +568,8 @@ class DistillationPipeline(TrainingPipeline):

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

            # fake_score_transformer forward
            training_batch = self._build_distill_input_kwargs(
@@ -404,24 +618,24 @@ class DistillationPipeline(TrainingPipeline):
                pred_real_video_uncond) * self.real_score_guidance_scale

            grad = (faker_score_pred_video - real_score_pred_video) / torch.abs(
-                generator_pred_video - real_score_pred_video).mean()
+                original_latent - real_score_pred_video).mean()
            grad = torch.nan_to_num(grad)

        dmd_loss = 0.5 * F.mse_loss(
-            generator_pred_video.float(),
-            (generator_pred_video.float() - grad.float()).detach())
+            original_latent.float(),
+            (original_latent.float() - grad.float()).detach())

        training_batch.dmd_latent_vis_dict.update({
            "training_batch_dmd_fwd_clean_latent":
            training_batch.latents,
            "generator_pred_video":
-            generator_pred_video,
+            original_latent.detach(),
            "real_score_pred_video":
-            real_score_pred_video,
+            real_score_pred_video.detach(),
            "faker_score_pred_video":
-            faker_score_pred_video,
+            faker_score_pred_video.detach(),
            "dmd_timestep":
-            timestep,
+            timestep.detach(),
        })

        return dmd_loss
@@ -514,16 +728,17 @@ class DistillationPipeline(TrainingPipeline):
            "encoder_hidden_states": training_batch.encoder_hidden_states,
            "encoder_attention_mask": training_batch.encoder_attention_mask,
        }
-        unconditional_dict = {
-            "encoder_hidden_states": self.negative_prompt_embeds,
-            "encoder_attention_mask": self.negative_prompt_attention_mask,
-        }
+        if getattr(self, "negative_prompt_embeds", None) is not None:
+            unconditional_dict = {
+                "encoder_hidden_states": self.negative_prompt_embeds,
+                "encoder_attention_mask": self.negative_prompt_attention_mask,
+            }
+            training_batch.unconditional_dict = unconditional_dict

        training_batch.dmd_latent_vis_dict = {}
        training_batch.fake_score_latent_vis_dict = {}

        training_batch.conditional_dict = conditional_dict
-        training_batch.unconditional_dict = unconditional_dict
        training_batch.raw_latent_shape = training_batch.latents.shape
        training_batch.latents = training_batch.latents.permute(0, 2, 1, 3, 4)
        self.video_latent_shape = training_batch.latents.shape
@@ -586,8 +801,15 @@ class DistillationPipeline(TrainingPipeline):
                    (dmd_loss / gradient_accumulation_steps).backward()
                total_dmd_loss += dmd_loss.detach().item()
            self._clip_model_grad_norm_(batch_gen, self.transformer)
+            for param in self.transformer.parameters():
+                # check if the gradient is not None and not zero
+                assert param.grad is not None and param.grad.abs().sum() > 0
            self.optimizer.step()
            self.optimizer.zero_grad(set_to_none=True)
+
+            if self.generator_ema is not None:
+                self.generator_ema.update(self.transformer)
+
            avg_dmd_loss = torch.tensor(total_dmd_loss /
                                        gradient_accumulation_steps,
                                        device=self.device)
@@ -611,6 +833,9 @@ class DistillationPipeline(TrainingPipeline):
            fake_score_latent_vis_dict.update(
                batch_fake.fake_score_latent_vis_dict)
        self._clip_model_grad_norm_(batch_fake, self.fake_score_transformer)
+        for param in self.fake_score_transformer.parameters():
+            # check if the gradient is not None and not zero
+            assert param.grad is not None and param.grad.abs().sum() > 0
        self.fake_score_optimizer.step()
        self.fake_score_lr_scheduler.step()
        self.lr_scheduler.step()
@@ -638,7 +863,8 @@ class DistillationPipeline(TrainingPipeline):
            self.transformer, self.fake_score_transformer, self.global_rank,
            self.training_args.resume_from_checkpoint, self.optimizer,
            self.fake_score_optimizer, self.train_dataloader, self.lr_scheduler,
-            self.fake_score_lr_scheduler, self.noise_random_generator)
+            self.fake_score_lr_scheduler, self.noise_random_generator,
+            self.generator_ema)

        if resumed_step > 0:
            self.init_steps = resumed_step
@@ -669,6 +895,14 @@ class DistillationPipeline(TrainingPipeline):
            sum(p.numel()
                for p in self.fake_score_transformer.parameters()) / 1e9)

+        if self.generator_ema is not None:
+            logger.info("  Generator EMA enabled with decay: %s",
+                        self.training_args.ema_decay)
+            logger.info("  Generator EMA start step: %s",
+                        self.training_args.ema_start_step)
+        else:
+            logger.info("  Generator EMA disabled")
+
    @torch.no_grad()
    def _log_validation(self, transformer, training_args, global_step) -> None:
        training_args.inference_mode = True
@@ -700,6 +934,18 @@ class DistillationPipeline(TrainingPipeline):

        transformer.eval()

+        # Optionally use EMA model for validation if available and ready
+        use_ema_for_validation = (self.training_args.use_ema
+                                  and self.is_ema_ready(global_step))
+        if use_ema_for_validation:
+            logger.info("Using EMA model for validation")
+            validation_transformer = self.transformer
+            ema_context = self.generator_ema.apply_to_model(
+                validation_transformer)
+        else:
+            validation_transformer = transformer
+            ema_context = None
+
        validation_steps = training_args.validation_sampling_steps.split(",")
        validation_steps = [int(step) for step in validation_steps]
        validation_steps = [step for step in validation_steps if step > 0]
@@ -715,50 +961,98 @@ class DistillationPipeline(TrainingPipeline):
            step_videos: list[np.ndarray] = []
            step_captions: list[str] = []

-            for validation_batch in validation_dataloader:
-                batch = self._prepare_validation_batch(sampling_param,
-                                                       training_args,
-                                                       validation_batch,
-                                                       num_inference_steps)
+            if ema_context is not None:
+                with ema_context:
+                    for validation_batch in validation_dataloader:
+                        batch = self._prepare_validation_batch(
+                            sampling_param, training_args, validation_batch,
+                            num_inference_steps)

-                negative_prompt = batch.negative_prompt
-                batch_negative = ForwardBatch(
-                    data_type="video",
-                    prompt=negative_prompt,
-                    prompt_embeds=[],
-                    prompt_attention_mask=[],
-                )
-                result_batch = self.validation_pipeline.prompt_encoding_stage(  # type: ignore
-                    batch_negative, training_args)
-                self.negative_prompt_embeds, self.negative_prompt_attention_mask = result_batch.prompt_embeds[
-                    0], result_batch.prompt_attention_mask[0]
+                        negative_prompt = batch.negative_prompt
+                        batch_negative = ForwardBatch(
+                            data_type="video",
+                            prompt=negative_prompt,
+                            prompt_embeds=[],
+                            prompt_attention_mask=[],
+                        )
+                        result_batch = self.validation_pipeline.prompt_encoding_stage(  # type: ignore
+                            batch_negative, training_args)
+                        self.negative_prompt_embeds, self.negative_prompt_attention_mask = result_batch.prompt_embeds[
+                            0], result_batch.prompt_attention_mask[0]

-                logger.info("rank: %s: rank_in_sp_group: %s, batch.prompt: %s",
+                        logger.info(
+                            "rank: %s: rank_in_sp_group: %s, batch.prompt: %s",
                            self.global_rank,
                            self.rank_in_sp_group,
                            batch.prompt,
                            local_main_process_only=False)

-                assert batch.prompt is not None and isinstance(
-                    batch.prompt, str)
-                step_captions.append(batch.prompt)
+                        assert batch.prompt is not None and isinstance(
+                            batch.prompt, str)
+                        step_captions.append(batch.prompt)

-                # Run validation inference
-                with torch.no_grad():
-                    output_batch = self.validation_pipeline.forward(
-                        batch, training_args)
-                samples = output_batch.output
-                if self.rank_in_sp_group != 0:
-                    continue
+                        # Run validation inference
+                        with torch.no_grad():
+                            output_batch = self.validation_pipeline.forward(
+                                batch, training_args)
+                        samples = output_batch.output
+                        if self.rank_in_sp_group != 0:
+                            continue

-                # Process outputs
-                video = rearrange(samples, "b c t h w -> t b c h w")
-                frames = []
-                for x in video:
-                    x = torchvision.utils.make_grid(x, nrow=6)
-                    x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
-                    frames.append((x * 255).numpy().astype(np.uint8))
-                step_videos.append(frames)
+                        # Process outputs
+                        video = rearrange(samples, "b c t h w -> t b c h w")
+                        frames = []
+                        for x in video:
+                            x = torchvision.utils.make_grid(x, nrow=6)
+                            x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+                            frames.append((x * 255).numpy().astype(np.uint8))
+                        step_videos.append(frames)
+            else:
+                # Use original transformer without EMA
+                for validation_batch in validation_dataloader:
+                    batch = self._prepare_validation_batch(
+                        sampling_param, training_args, validation_batch,
+                        num_inference_steps)
+
+                    negative_prompt = batch.negative_prompt
+                    batch_negative = ForwardBatch(
+                        data_type="video",
+                        prompt=negative_prompt,
+                        prompt_embeds=[],
+                        prompt_attention_mask=[],
+                    )
+                    result_batch = self.validation_pipeline.prompt_encoding_stage(  # type: ignore
+                        batch_negative, training_args)
+                    self.negative_prompt_embeds, self.negative_prompt_attention_mask = result_batch.prompt_embeds[
+                        0], result_batch.prompt_attention_mask[0]
+
+                    logger.info(
+                        "rank: %s: rank_in_sp_group: %s, batch.prompt: %s",
+                        self.global_rank,
+                        self.rank_in_sp_group,
+                        batch.prompt,
+                        local_main_process_only=False)
+
+                    assert batch.prompt is not None and isinstance(
+                        batch.prompt, str)
+                    step_captions.append(batch.prompt)
+
+                    # Run validation inference
+                    with torch.no_grad():
+                        output_batch = self.validation_pipeline.forward(
+                            batch, training_args)
+                    samples = output_batch.output
+                    if self.rank_in_sp_group != 0:
+                        continue
+
+                    # Process outputs
+                    video = rearrange(samples, "b c t h w -> t b c h w")
+                    frames = []
+                    for x in video:
+                        x = torchvision.utils.make_grid(x, nrow=6)
+                        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+                        frames.append((x * 255).numpy().astype(np.uint8))
+                    step_videos.append(frames)

            # Log validation results for this step
            world_group = get_world_group()
@@ -835,16 +1129,16 @@ class DistillationPipeline(TrainingPipeline):
                                                        latents.dtype)
                else:
                    latents += self.vae.shift_factor
-            with torch.autocast("cuda", dtype=torch.bfloat16):
-                video = self.vae.decode(latents)
-            video = (video / 2 + 0.5).clamp(0, 1)
-            video = video.cpu().float()
-            video = video.permute(0, 2, 1, 3, 4)
-            video = (video * 255).numpy().astype(np.uint8)
-            wandb_loss_dict[latent_key] = wandb.Video(
-                video, fps=24, format="mp4")  # change to 16 for Wan2.1
-            # Clean up references
-            del video, latents
+                with torch.autocast("cuda", dtype=torch.bfloat16):
+                    video = self.vae.decode(latents)
+                video = (video / 2 + 0.5).clamp(0, 1)
+                video = video.cpu().float()
+                video = video.permute(0, 2, 1, 3, 4)
+                video = (video * 255).numpy().astype(np.uint8)
+                wandb_loss_dict[latent_key] = wandb.Video(
+                    video, fps=24, format="mp4")  # change to 16 for Wan2.1
+                # Clean up references
+                del video, latents

        # Process DMD training data if available - use decode_stage instead of self.vae.decode
        if 'generator_pred_video' in dmd_latents_vis_dict:
@@ -896,14 +1190,6 @@ class DistillationPipeline(TrainingPipeline):
        else:
            set_random_seed(seed + self.global_rank)

-        # Check trainable params
-        num_trainable_generator = round(
-            count_trainable(self.transformer) / 1e9, 3)
-        num_trainable_critic = round(
-            count_trainable(self.fake_score_transformer) / 1e9, 3)
-        logger.info(
-            "rank: %s: # of trainable params in generator: %sB, # of trainable params in critic: %sB",
-            self.global_rank, num_trainable_generator, num_trainable_critic)
        # Set random seeds for deterministic training
        self.noise_random_generator = torch.Generator(device="cpu").manual_seed(
            self.seed)
@@ -913,6 +1199,10 @@ class DistillationPipeline(TrainingPipeline):
            device="cpu").manual_seed(self.seed)
        logger.info("Initialized random seeds with seed: %s", seed)

+        # Initialize current_trainstep for EMA ready checks
+        #TODO: check if needed
+        self.current_trainstep = self.init_steps
+
        # Resume from checkpoint if specified (this will restore random states)
        if self.training_args.resume_from_checkpoint:
            self._resume_from_checkpoint()
@@ -956,6 +1246,13 @@ class DistillationPipeline(TrainingPipeline):
            self.current_trainstep = step
            training_batch.current_vsa_sparsity = current_vsa_sparsity

+            if (step >= self.training_args.ema_start_step) and \
+                    (self.generator_ema is None) and (self.training_args.ema_decay > 0):
+                self.generator_ema = EMA_FSDP(
+                    self.transformer, decay=self.training_args.ema_decay)
+                logger.info("Created generator EMA at step %s with decay=%s",
+                            step, self.training_args.ema_decay)
+
            with torch.autocast("cuda", dtype=torch.bfloat16):
                training_batch = self.train_one_step(training_batch)

@@ -969,11 +1266,19 @@ class DistillationPipeline(TrainingPipeline):
            avg_step_time = sum(step_times) / len(step_times)

            progress_bar.set_postfix({
-                "total_loss": f"{total_loss:.4f}",
-                "generator_loss": f"{generator_loss:.4f}",
-                "fake_score_loss": f"{fake_score_loss:.4f}",
-                "step_time": f"{step_time:.2f}s",
-                "grad_norm": grad_norm,
+                "total_loss":
+                f"{total_loss:.4f}",
+                "generator_loss":
+                f"{generator_loss:.4f}",
+                "fake_score_loss":
+                f"{fake_score_loss:.4f}",
+                "step_time":
+                f"{step_time:.2f}s",
+                "grad_norm":
+                grad_norm,
+                "ema":
+                "✓" if (self.generator_ema is not None and self.is_ema_ready())
+                else "✗",
            })
            progress_bar.update(1)

@@ -1001,6 +1306,15 @@ class DistillationPipeline(TrainingPipeline):
                if use_vsa:
                    log_data["VSA_train_sparsity"] = current_vsa_sparsity

+                if self.generator_ema is not None:
+                    log_data["ema_enabled"] = True
+                    log_data["ema_decay"] = self.training_args.ema_decay
+                else:
+                    log_data["ema_enabled"] = False
+
+                ema_stats = self.get_ema_stats()
+                log_data.update(ema_stats)
+
                if training_batch.dmd_latent_vis_dict:
                    dmd_additional_logs = {
                        "generator_timestep":
@@ -1032,7 +1346,8 @@ class DistillationPipeline(TrainingPipeline):
                    self.global_rank, self.training_args.output_dir, step,
                    self.optimizer, self.fake_score_optimizer,
                    self.train_dataloader, self.lr_scheduler,
-                    self.fake_score_lr_scheduler, self.noise_random_generator)
+                    self.fake_score_lr_scheduler, self.noise_random_generator,
+                    self.generator_ema)

                if self.transformer:
                    self.transformer.train()
@@ -1049,7 +1364,11 @@ class DistillationPipeline(TrainingPipeline):
                                             self.global_rank,
                                             self.training_args.output_dir,
                                             f"{step}_weight_only",
-                                             only_save_generator_weight=True)
+                                             only_save_generator_weight=True,
+                                             generator_ema=self.generator_ema)
+
+                if self.training_args.use_ema and self.is_ema_ready():
+                    self.save_ema_weights(self.training_args.output_dir, step)

            if self.training_args.log_validation and step % self.training_args.validation_steps == 0:
                if self.training_args.log_visualization:
@@ -1069,7 +1388,11 @@ class DistillationPipeline(TrainingPipeline):
            self.training_args.output_dir, self.training_args.max_train_steps,
            self.optimizer, self.fake_score_optimizer, self.train_dataloader,
            self.lr_scheduler, self.fake_score_lr_scheduler,
-            self.noise_random_generator)
+            self.noise_random_generator, self.generator_ema)
+
+        if self.training_args.use_ema and self.is_ema_ready():
+            self.save_ema_weights(self.training_args.output_dir,
+                                  self.training_args.max_train_steps)

        if get_sp_group():
            cleanup_dist_env_and_memory()
@@ -0,0 +1,406 @@
+# SPDX-License-Identifier: Apache-2.0
+import sys
+from copy import deepcopy
+from typing import cast
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+import wandb
+from fastvideo.dataset.dataloader.schema import (
+    pyarrow_schema_ode_trajectory_text_only)
+from fastvideo.distributed import get_local_torch_device
+from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
+from fastvideo.forward_context import set_forward_context
+from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.scheduling_self_forcing_flow_match import (
+    SelfForcingFlowMatchScheduler)
+from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import (
+    WanCausalDMDPipeline)
+from fastvideo.pipelines.pipeline_batch_info import TrainingBatch
+from fastvideo.training.training_pipeline import TrainingPipeline
+from fastvideo.training.training_utils import (
+    clip_grad_norm_while_handling_failing_dtensor_cases)
+
+logger = init_logger(__name__)
+
+
+class ODEInitTrainingPipeline(TrainingPipeline):
+    """
+    Training pipeline for ODE-init using precomputed denoising trajectories.
+
+    Supervision: predict the next latent in the stored trajectory by
+    - feeding current latent at timestep t into the transformer to predict noise
+    - stepping the scheduler with the predicted noise
+    - minimizing MSE to the stored next latent at timestep t_next
+    """
+
+    _required_config_modules = ["scheduler", "transformer", "vae"]
+
+    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
+        # Match the preprocess/generation scheduler for consistent stepping
+        self.modules["scheduler"] = SelfForcingFlowMatchScheduler(
+            shift=fastvideo_args.pipeline_config.flow_shift,
+            sigma_min=0.0,
+            extra_one_step=True)
+        self.modules["scheduler"].set_timesteps(num_inference_steps=1000,
+                                                training=True)
+
+    def set_schemas(self):
+        self.train_dataset_schema = pyarrow_schema_ode_trajectory_text_only
+
+    def initialize_training_pipeline(self, training_args: TrainingArgs):
+        super().initialize_training_pipeline(training_args)
+
+        self.noise_scheduler = self.get_module("scheduler")
+        self.vae = self.get_module("vae")
+        self.vae.requires_grad_(False)
+
+        self.timestep_shift = self.training_args.pipeline_config.flow_shift
+        assert self.timestep_shift == 5.0, "flow_shift must be 5.0"
+        self.noise_scheduler = SelfForcingFlowMatchScheduler(
+            shift=self.timestep_shift, sigma_min=0.0, extra_one_step=True)
+        self.noise_scheduler.set_timesteps(num_inference_steps=1000,
+                                           training=True)
+
+        logger.info("dmd_denoising_steps: %s",
+                    self.training_args.pipeline_config.dmd_denoising_steps)
+        self.dmd_denoising_steps = torch.tensor([1000, 750, 500, 250],
+                                                dtype=torch.long,
+                                                device=get_local_torch_device())
+        if training_args.warp_denoising_step:  # Warp the denoising step according to the scheduler time shift
+            timesteps = torch.cat((self.noise_scheduler.timesteps.cpu(),
+                                   torch.tensor([0],
+                                                dtype=torch.float32))).cuda()
+            logger.info("timesteps: %s", timesteps)
+            self.dmd_denoising_steps = timesteps[1000 -
+                                                 self.dmd_denoising_steps]
+            logger.info("warped self.dmd_denoising_steps: %s",
+                        self.dmd_denoising_steps)
+        else:
+            raise ValueError("warp_denoising_step must be true")
+
+        self.dmd_denoising_steps = self.dmd_denoising_steps.to(
+            get_local_torch_device())
+
+        logger.info("denoising_step_list: %s", self.dmd_denoising_steps)
+
+        logger.info(
+            "Initialized ODE-init training pipeline with %s denoising steps",
+            len(self.dmd_denoising_steps))
+        # Cache for nearest trajectory index per DMD step (computed lazily on first batch)
+        self._cached_closest_idx_per_dmd = None
+        self.num_train_timestep = self.noise_scheduler.num_train_timesteps
+        self.manual_idx = 0
+
+    def initialize_validation_pipeline(self, training_args: TrainingArgs):
+        logger.info("Initializing validation pipeline...")
+        args_copy = deepcopy(training_args)
+        args_copy.inference_mode = True
+        # Warm start validation with current transformer
+        self.validation_pipeline = WanCausalDMDPipeline.from_pretrained(
+            training_args.model_path,
+            args=args_copy,  # type: ignore
+            inference_mode=True,
+            loaded_modules={
+                "transformer": self.get_module("transformer"),
+            },
+            tp_size=training_args.tp_size,
+            sp_size=training_args.sp_size,
+            num_gpus=training_args.num_gpus,
+            pin_cpu_memory=training_args.pin_cpu_memory,
+            dit_cpu_offload=True)
+
+    def _get_next_batch(
+            self,
+            training_batch) -> tuple[TrainingBatch, torch.Tensor, torch.Tensor]:
+        batch = next(self.train_loader_iter, None)  # type: ignore
+        if batch is None:
+            self.current_epoch += 1
+            logger.info("Starting epoch %s", self.current_epoch)
+            self.train_loader_iter = iter(self.train_dataloader)
+            batch = next(self.train_loader_iter)
+
+        # Required fields from parquet (ODE trajectory schema)
+        encoder_hidden_states = batch['text_embedding']
+        encoder_attention_mask = batch['text_attention_mask']
+        infos = batch['info_list']
+
+        # Trajectory tensors may include a leading singleton batch dim per row
+        trajectory_latents = batch['trajectory_latents']
+        if trajectory_latents.dim() == 7:
+            # [B, 1, S, C, T, H, W] -> [B, S, C, T, H, W]
+            trajectory_latents = trajectory_latents[:, 0]
+        elif trajectory_latents.dim() == 6:
+            # already [B, S, C, T, H, W]
+            pass
+        else:
+            raise ValueError(
+                f"Unexpected trajectory_latents dim: {trajectory_latents.dim()}"
+            )
+
+        trajectory_timesteps = batch['trajectory_timesteps']
+        if trajectory_timesteps.dim() == 3:
+            # [B, 1, S] -> [B, S]
+            trajectory_timesteps = trajectory_timesteps[:, 0]
+        elif trajectory_timesteps.dim() == 2:
+            # [B, S]
+            pass
+        else:
+            raise ValueError(
+                f"Unexpected trajectory_timesteps dim: {trajectory_timesteps.dim()}"
+            )
+        # [B, S, C, T, H, W] -> [B, S, T, C, H, W] to match self-forcing
+        trajectory_latents = trajectory_latents.permute(0, 1, 3, 2, 4, 5)
+
+        # Move to device
+        device = get_local_torch_device()
+        training_batch.encoder_hidden_states = encoder_hidden_states.to(
+            device, dtype=torch.bfloat16)
+        training_batch.encoder_attention_mask = encoder_attention_mask.to(
+            device, dtype=torch.bfloat16)
+        training_batch.infos = infos
+
+        return training_batch, trajectory_latents.to(
+            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)
+
+        ## TEMP Used for loading the sf .pt files directly
+        """
+        self.manual_idx = self.manual_idx % 155
+        path = f"/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_pt_vidprom_1000/{self.manual_idx:05d}.pt"
+        logger.info("path: %s", path)
+        self.manual_idx += 1
+        # path = "/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_single_full/00000.pt"
+        b = torch.load(path)
+        training_batch.encoder_hidden_states = b["text_embedding"][0].unsqueeze(
+            0).to(device, dtype=torch.bfloat16)
+        trajectory_latents = b["ode_latent"].to(device, dtype=torch.bfloat16)
+        logger.info("trajectory_latents: %s", trajectory_latents.shape)
+        logger.info("encoder_hidden_states: %s",
+                    training_batch.encoder_hidden_states.shape)
+        assert trajectory_latents.shape[1] <= 10, "trajectory_latents.shape[1] must be <= 10"
+        return training_batch, trajectory_latents.to(
+            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)
+        """
+
+    def _get_timestep(self,
+                      min_timestep: int,
+                      max_timestep: int,
+                      batch_size: int,
+                      num_frame: int,
+                      num_frame_per_block: int,
+                      uniform_timestep: bool = False) -> torch.Tensor:
+        if uniform_timestep:
+            timestep = torch.randint(min_timestep,
+                                     max_timestep, [batch_size, 1],
+                                     device=self.device,
+                                     dtype=torch.long).repeat(1, num_frame)
+            return timestep
+        else:
+            timestep = torch.randint(min_timestep,
+                                     max_timestep, [batch_size, num_frame],
+                                     device=self.device,
+                                     dtype=torch.long)
+            # logger.info(f"individual timestep: {timestep}")
+            # make the noise level the same within every block
+            timestep = timestep.reshape(timestep.shape[0], -1,
+                                        num_frame_per_block)
+            timestep[:, :, 1:] = timestep[:, :, 0:1]
+            timestep = timestep.reshape(timestep.shape[0], -1)
+            return timestep
+
+    def _step_predict_next_latent(
+        self, traj_latents: torch.Tensor, traj_timesteps: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_attention_mask: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, dict[str,
+                                                              torch.Tensor]]:
+        latent_vis_dict: dict[str, torch.Tensor] = {}
+        device = get_local_torch_device()
+        target_latent = traj_latents[:, -1]
+
+        # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
+        B, S, num_frames, num_channels, height, width = traj_latents.shape
+
+        # Lazily cache nearest trajectory index per DMD step based on the (fixed) S timesteps
+        if self._cached_closest_idx_per_dmd is None:
+            self._cached_closest_idx_per_dmd = torch.tensor(
+                [0, 12, 24, 36], dtype=torch.long).cpu()
+            # [0, 1, 2, 3], dtype=torch.long).cpu()
+            logger.info("self._cached_closest_idx_per_dmd: %s",
+                        self._cached_closest_idx_per_dmd)
+            logger.info(
+                "corresponding timesteps: %s", self.noise_scheduler.timesteps[
+                    self._cached_closest_idx_per_dmd])
+
+        # Select the K indexes from traj_latents using self._cached_closest_idx_per_dmd
+        # traj_latents: [B, S, C, T, H, W], self._cached_closest_idx_per_dmd: [K]
+        # Output: [B, K, C, T, H, W]
+        assert self._cached_closest_idx_per_dmd is not None
+        relevant_traj_latents = torch.index_select(
+            traj_latents,
+            dim=1,
+            index=self._cached_closest_idx_per_dmd.to(traj_latents.device))
+        logger.info("relevant_traj_latents: %s", relevant_traj_latents.shape)
+        # assert relevant_traj_latents.shape[0] == 1
+
+        indexes = self._get_timestep(  # [B, num_frames]
+            0,
+            len(self.dmd_denoising_steps),
+            B,
+            num_frames,
+            3,
+            uniform_timestep=False)
+        logger.info("indexes: %s", indexes.shape)
+        logger.info("indexes: %s", indexes)
+        # noisy_input = relevant_traj_latents[indexes]
+        noisy_input = torch.gather(
+            relevant_traj_latents,
+            dim=1,
+            index=indexes.reshape(B, 1, num_frames, 1, 1,
+                                  1).expand(-1, -1, -1, num_channels, height,
+                                            width).to(self.device)).squeeze(1)
+        timestep = self.dmd_denoising_steps[indexes]
+        logger.info("selected timestep for rank %s: %s",
+                    self.global_rank,
+                    timestep,
+                    local_main_process_only=False)
+
+        # Prepare inputs for transformer
+        latent_vis_dict["noisy_input"] = noisy_input.permute(
+            0, 2, 1, 3, 4).detach().clone().cpu()
+        latent_vis_dict["x0"] = target_latent.permute(0, 2, 1, 3,
+                                                      4).detach().clone().cpu()
+
+        input_kwargs = {
+            "hidden_states": noisy_input.permute(0, 2, 1, 3, 4),
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep": timestep.to(device, dtype=torch.bfloat16),
+            "return_dict": False,
+        }
+        # Predict noise and step the scheduler to obtain next latent
+        with set_forward_context(current_timestep=timestep,
+                                 attn_metadata=None,
+                                 forward_batch=None):
+            noise_pred = self.transformer(**input_kwargs).permute(0, 2, 1, 3, 4)
+
+        from fastvideo.models.utils import pred_noise_to_pred_video
+        pred_video = pred_noise_to_pred_video(
+            pred_noise=noise_pred.flatten(0, 1),
+            noise_input_latent=noisy_input.flatten(0, 1),
+            timestep=timestep.to(dtype=torch.bfloat16).flatten(0, 1),
+            scheduler=self.modules["scheduler"]).unflatten(
+                0, noise_pred.shape[:2])
+        latent_vis_dict["pred_video"] = pred_video.permute(
+            0, 2, 1, 3, 4).detach().clone().cpu()
+
+        return pred_video, target_latent, timestep, latent_vis_dict
+
+    def train_one_step(self, training_batch):  # type: ignore[override]
+        self.transformer.train()
+        self.optimizer.zero_grad()
+        training_batch.total_loss = 0.0
+        args = cast(TrainingArgs, self.training_args)
+
+        # Using cached nearest index per DMD step; computation happens in _step_predict_next_latent
+
+        for _ in range(args.gradient_accumulation_steps):
+            training_batch, traj_latents, traj_timesteps = self._get_next_batch(
+                training_batch)
+            text_embeds = training_batch.encoder_hidden_states
+            text_attention_mask = training_batch.encoder_attention_mask
+            assert traj_latents.shape[0] == 1
+
+            # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
+            _, S = traj_latents.shape[0], traj_latents.shape[1]
+            if S < 2:
+                raise ValueError("Trajectory must contain at least 2 steps")
+
+            # Forward to predict next latent by stepping scheduler with predicted noise
+            noise_pred, target_latent, t, latent_vis_dict = self._step_predict_next_latent(
+                traj_latents, traj_timesteps, text_embeds, text_attention_mask)
+
+            training_batch.latent_vis_dict.update(latent_vis_dict)
+
+            mask = t != 0
+
+            # Compute loss
+            loss = F.mse_loss(noise_pred[mask],
+                              target_latent[mask],
+                              reduction="mean")
+            loss = loss / args.gradient_accumulation_steps
+
+            with set_forward_context(current_timestep=t,
+                                     attn_metadata=None,
+                                     forward_batch=None):
+                loss.backward()
+            avg_loss = loss.detach().clone()
+            training_batch.total_loss += avg_loss.item()
+
+        # Clip grad and step optimizers
+        grad_norm = clip_grad_norm_while_handling_failing_dtensor_cases(
+            [p for p in self.transformer.parameters() if p.requires_grad],
+            args.max_grad_norm if args.max_grad_norm is not None else 0.0)
+
+        self.optimizer.step()
+        self.lr_scheduler.step()
+
+        if grad_norm is None:
+            grad_value = 0.0
+        else:
+            try:
+                if isinstance(grad_norm, torch.Tensor):
+                    grad_value = float(grad_norm.detach().float().item())
+                else:
+                    grad_value = float(grad_norm)
+            except Exception:
+                grad_value = 0.0
+        training_batch.grad_norm = grad_value
+        return training_batch
+
+    def visualize_intermediate_latents(self, training_batch: TrainingBatch,
+                                       training_args: TrainingArgs, step: int):
+        """Add visualization data to wandb logging and save frames to disk."""
+        wandb_loss_dict = {}
+        latents_vis_dict = training_batch.latent_vis_dict
+        latent_log_keys = ['noisy_input', 'x0', 'pred_video']
+        for latent_key in latent_log_keys:
+            assert latent_key in latents_vis_dict and latents_vis_dict[
+                latent_key] is not None
+            latent = latents_vis_dict[latent_key]
+            pixel_latent = self.validation_pipeline.decoding_stage.decode(
+                latent, training_args)
+
+            video = pixel_latent.cpu().float()
+            video = video.permute(0, 2, 1, 3, 4)
+            video = (video * 255).numpy().astype(np.uint8)
+            wandb_loss_dict[latent_key] = wandb.Video(
+                video, fps=16, format="mp4")  # change to 16 for Wan2.1
+            # Clean up references
+            del video, pixel_latent, latent
+
+        # Log to wandb
+        if self.global_rank == 0:
+            wandb.log(wandb_loss_dict, step=step)
+
+
+def main(args) -> None:
+    logger.info("Starting ODE-init training pipeline...")
+    pipeline = ODEInitTrainingPipeline.from_pretrained(
+        args.pretrained_model_name_or_path, args=args)
+    args = pipeline.training_args
+    pipeline.train()
+    logger.info("ODE-init training pipeline done")
+
+
+if __name__ == "__main__":
+    argv = sys.argv
+    from fastvideo.fastvideo_args import TrainingArgs
+    from fastvideo.utils import FlexibleArgumentParser
+    parser = FlexibleArgumentParser()
+    parser = TrainingArgs.add_cli_args(parser)
+    parser = FastVideoArgs.add_cli_args(parser)
+    args = parser.parse_args()
+    args.dit_cpu_offload = False
+    main(args)
@@ -520,6 +520,10 @@ class TrainingPipeline(LoRAPipeline, ABC):
                self.transformer.train()
                self.sp_group.barrier()
            if self.training_args.log_validation and step % self.training_args.validation_steps == 0:
+                if self.training_args.log_visualization:
+                    self.visualize_intermediate_latents(training_batch,
+                                                        self.training_args,
+                                                        step)
                self._log_validation(self.transformer, self.training_args, step)
                gpu_memory_usage = torch.cuda.memory_allocated() / 1024**2
                trainable_params = round(
@@ -720,3 +724,10 @@ class TrainingPipeline(LoRAPipeline, ABC):
        # Re-enable gradients for training
        training_args.inference_mode = False
        transformer.train()
+
+    def visualize_intermediate_latents(self, training_batch: TrainingBatch,
+                                       training_args: TrainingArgs, step: int):
+        """Add visualization data to wandb logging and save frames to disk."""
+        raise NotImplementedError(
+            "Visualize intermediate latents is not implemented for training pipeline"
+        )
@@ -202,6 +202,7 @@ def save_distillation_checkpoint(generator_transformer,
                                 generator_scheduler=None,
                                 fake_score_scheduler=None,
                                 noise_generator=None,
+                                 generator_ema=None,
                                 only_save_generator_weight=False) -> None:
    """
    Save distillation checkpoint with both generator and fake_score models.
@@ -233,6 +234,8 @@ def save_distillation_checkpoint(generator_transformer,
        if generator_scheduler is not None:
            generator_states["scheduler"] = SchedulerWrapper(
                generator_scheduler)
+        if generator_ema is not None:
+            generator_states["ema"] = generator_ema.state_dict()

        generator_dcp_dir = os.path.join(save_dir, "distributed_checkpoint",
                                         "generator")
@@ -402,7 +405,8 @@ def load_distillation_checkpoint(generator_transformer,
                                 dataloader=None,
                                 generator_scheduler=None,
                                 fake_score_scheduler=None,
-                                 noise_generator=None) -> int:
+                                 noise_generator=None,
+                                 generator_ema=None) -> int:
    """
    Load distillation checkpoint with both generator and fake_score models.
    Returns the step number from which training should resume.
@@ -456,6 +460,20 @@ def load_distillation_checkpoint(generator_transformer,
        end_time - begin_time,
        local_main_process_only=False)

+    # Load EMA state if available and generator_ema is provided
+    if generator_ema is not None:
+        try:
+            ema_state = generator_states.get("ema")
+            if ema_state is not None:
+                generator_ema.load_state_dict(ema_state)
+                logger.info("rank: %s, generator EMA state loaded successfully",
+                            rank)
+            else:
+                logger.info("rank: %s, no EMA state found in checkpoint", rank)
+        except Exception as e:
+            logger.warning("rank: %s, failed to load EMA state: %s", rank,
+                           str(e))
+
    # Load critic distributed checkpoint
    critic_dcp_dir = os.path.join(checkpoint_path, "distributed_checkpoint",
                                  "critic")
@@ -1282,3 +1300,169 @@ def get_scheduler(

 def count_trainable(model: torch.nn.Module) -> int:
    return sum(p.numel() for p in model.parameters() if p.requires_grad)
+
+
+class EMA_FSDP:
+    """
+    FSDP2-friendly EMA with two modes:
+      - mode="local_shard" (default): maintain float32 CPU EMA of local parameter shards on every rank.
+        Provides a context manager to temporarily swap EMA weights into the live model for teacher forward.
+      - mode="rank0_full": maintain a consolidated float32 CPU EMA of full parameters on rank 0 only
+        using gather_state_dict_on_cpu_rank0(). Useful for checkpoint export; not for teacher forward.
+
+    Usage (local_shard for CM teacher):
+      ema = EMA_FSDP(model, decay=0.999, mode="local_shard")
+      for step in ...:
+          ema.update(model)
+      with ema.apply_to_model(model):
+          with torch.no_grad():
+              y_teacher = model(...)
+
+    Usage (rank0_full for export):
+      ema = EMA_FSDP(model, decay=0.999, mode="rank0_full")
+      ema.update(model)
+      ema.state_dict()  # on rank 0
+    """
+
+    def __init__(self, module, decay: float = 0.999, mode: str = "local_shard"):
+        self.decay = float(decay)
+        self.mode = mode
+        self.shadow: dict[str, torch.Tensor] = {}
+        self.rank = dist.get_rank() if dist.is_initialized() else 0
+        if self.mode not in {"local_shard", "rank0_full"}:
+            raise ValueError(f"Unsupported EMA_FSDP mode: {self.mode}")
+        self._init_shadow(module)
+
+    @staticmethod
+    def _to_local_tensor(t: torch.Tensor) -> torch.Tensor:
+        # DTensor-aware to_local fetch; fall back to raw tensor
+        try:
+            from torch.distributed.tensor import DTensor  # type: ignore
+            if isinstance(t, DTensor):
+                return t.to_local()
+        except Exception:
+            pass
+        return t
+
+    @torch.no_grad()
+    def _init_shadow(self, module):
+        if self.mode == "rank0_full":
+            cpu_state = gather_state_dict_on_cpu_rank0(module, device=None)
+            if self.rank == 0:
+                self.shadow = {
+                    k: v.detach().clone().float().cpu()
+                    for k, v in cpu_state.items()
+                }
+            else:
+                self.shadow = {}
+            return
+
+        # local_shard: maintain EMA of local shards for requires_grad params
+        self.shadow = {}
+        for name, p in module.named_parameters():
+            if not p.requires_grad:
+                continue
+            local = self._to_local_tensor(p.detach())
+            self.shadow[name] = local.clone().float().cpu()
+
+    @torch.no_grad()
+    def update(self, module):
+        d = self.decay
+        if self.mode == "rank0_full":
+            if self.rank != 0:
+                return
+            cpu_state = gather_state_dict_on_cpu_rank0(module, device=None)
+            for n, v in cpu_state.items():
+                v_cpu = v.detach().float().cpu()
+                if n not in self.shadow:
+                    self.shadow[n] = v_cpu.clone()
+                else:
+                    self.shadow[n].mul_(d).add_(v_cpu, alpha=1.0 - d)
+            return
+
+        # local_shard: update local shard EMA on every rank
+        for name, p in module.named_parameters():
+            if not p.requires_grad:
+                continue
+            local = self._to_local_tensor(p.detach())
+            v_cpu = local.float().cpu()
+            if name not in self.shadow:
+                self.shadow[name] = v_cpu.clone()
+            else:
+                self.shadow[name].mul_(d).add_(v_cpu, alpha=1.0 - d)
+
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        if self.mode == "rank0_full":
+            return {
+                k: v.clone()
+                for k, v in self.shadow.items()
+            } if self.rank == 0 else {}
+        return {k: v.clone() for k, v in self.shadow.items()}
+
+    def load_state_dict(self, sd: dict[str, torch.Tensor]):
+        self.shadow = {k: v.clone() for k, v in sd.items()}
+
+    @torch.no_grad()
+    def copy_to_unwrapped(self, module) -> None:
+        """
+        Copy EMA weights into a non-sharded (unwrapped) module. Intended for export/eval.
+        For mode="rank0_full", only rank 0 has the full EMA state.
+        """
+        if self.mode == "rank0_full" and self.rank != 0:
+            return
+        name_to_param = dict(module.named_parameters())
+        for n, w in self.shadow.items():
+            if n in name_to_param:
+                p = name_to_param[n]
+                p.data.copy_(w.to(dtype=p.dtype, device=p.device))
+
+    class _ApplyEMACtx:
+
+        def __init__(self, ema: "EMA_FSDP", module):
+            self.ema = ema
+            self.module = module
+            self.saved: dict[str, torch.Tensor] = {}
+
+        def __enter__(self):
+            if self.ema.mode != "local_shard":
+                raise RuntimeError(
+                    "EMA apply_to_model is only supported for mode='local_shard'"
+                )
+            with torch.no_grad():
+                for name, p in self.module.named_parameters():
+                    if not p.requires_grad:
+                        continue
+                    # Save local shard
+                    p_local = EMA_FSDP._to_local_tensor(p.detach())
+                    if p_local.numel() == 0:
+                        # Nothing to swap on this rank for this param
+                        continue
+                    self.saved[name] = p_local.clone().to(device=p_local.device,
+                                                          dtype=p_local.dtype)
+                    if name in self.ema.shadow:
+                        ema_cpu = self.ema.shadow[name]
+                        if ema_cpu.numel() != p_local.numel():
+                            # Shard shape mismatch (e.g., empty shard here), skip
+                            continue
+                        # Copy EMA shard into local param shard
+                        p_local.copy_(
+                            ema_cpu.to(dtype=p_local.dtype,
+                                       device=p_local.device))
+            return self.module
+
+        def __exit__(self, exc_type, exc, tb):
+            with torch.no_grad():
+                for name, p in self.module.named_parameters():
+                    if name in self.saved:
+                        p_local = EMA_FSDP._to_local_tensor(p.detach())
+                        if p_local.numel() == 0:
+                            continue
+                        saved_local = self.saved[name]
+                        if saved_local.numel() != p_local.numel():
+                            continue
+                        p_local.copy_(saved_local)
+            self.saved.clear()
+            return False
+
+    def apply_to_model(self, module):
+        return EMA_FSDP._ApplyEMACtx(self, module)
@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+import sys
+from copy import deepcopy
+
+from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
+from fastvideo.logger import init_logger
+from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import (
+    WanCausalDMDPipeline)
+from fastvideo.training.self_forcing_distillation_pipeline import (
+    SelfForcingDistillationPipeline)
+from fastvideo.utils import is_vsa_available
+
+vsa_available = is_vsa_available()
+
+logger = init_logger(__name__)
+
+
+class WanSelfForcingDistillationPipeline(SelfForcingDistillationPipeline):
+    """
+    A self-forcing distillation pipeline for Wan that uses the self-forcing methodology
+    with DMD for video generation.
+    """
+    _required_config_modules = [
+        "scheduler", "transformer", "vae", "real_score_transformer",
+        "fake_score_transformer"
+    ]
+
+    def create_training_stages(self, training_args: TrainingArgs):
+        """
+        May be used in future refactors.
+        """
+        pass
+
+    def initialize_validation_pipeline(self, training_args: TrainingArgs):
+        logger.info("Initializing validation pipeline...")
+        args_copy = deepcopy(training_args)
+
+        args_copy.inference_mode = True
+        validation_pipeline = WanCausalDMDPipeline.from_pretrained(
+            training_args.model_path,
+            args=args_copy,  # type: ignore
+            inference_mode=True,
+            loaded_modules={"transformer": self.get_module("transformer")},
+            tp_size=training_args.tp_size,
+            sp_size=training_args.sp_size,
+            num_gpus=training_args.num_gpus,
+            pin_cpu_memory=training_args.pin_cpu_memory,
+            dit_cpu_offload=True)
+
+        self.validation_pipeline = validation_pipeline
+
+
+def main(args) -> None:
+    logger.info("Starting Wan self-forcing distillation pipeline...")
+
+    pipeline = WanSelfForcingDistillationPipeline.from_pretrained(
+        args.pretrained_model_name_or_path, args=args)
+
+    args = pipeline.training_args
+    pipeline.train()
+    logger.info("Wan self-forcing distillation pipeline completed")
+
+
+if __name__ == "__main__":
+    argv = sys.argv
+    from fastvideo.fastvideo_args import TrainingArgs
+    from fastvideo.utils import FlexibleArgumentParser
+    parser = FlexibleArgumentParser()
+    parser = TrainingArgs.add_cli_args(parser)
+    parser = FastVideoArgs.add_cli_args(parser)
+    args = parser.parse_args()
+    main(args)
@@ -6,7 +6,7 @@ import pytest

 from fastvideo import VideoGenerator
 from fastvideo.logger import init_logger
-from fastvideo.tests.utils import compute_video_ssim_torchvision, write_ssim_results
+from 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
@@ -192,4 +192,3 @@ 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 ./fastvideo/tests/encoders -vs")
+    run_test("pytest ./tests/encoders -vs")

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

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

@app.function(gpu="L40S:2", image=image, timeout=2700)
 def run_ssim_tests():
-    run_test("pytest ./fastvideo/tests/ssim -vs")
+    run_test("pytest ./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 ./fastvideo/tests/training/Vanilla -srP")
+    run_test("wandb login $WANDB_API_KEY && pytest ./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 ./fastvideo/tests/training/lora/test_lora_training.py -srP")
+    run_test("wandb login $WANDB_API_KEY && pytest ./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 ./fastvideo/tests/training/VSA -srP")
+    run_test("wandb login $WANDB_API_KEY && pytest ./tests/training/VSA -srP")

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

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

@app.function(gpu="L40S:1", image=image, timeout=900)
 def run_unit_test():
-    run_test("pytest ./fastvideo/tests/dataset/ ./fastvideo/tests/workflow/ -vs")
+    run_test("pytest ./tests/dataset/ ./tests/workflow/ -vs")
@@ -0,0 +1,184 @@
+import os
+from pathlib import Path
+from huggingface_hub import snapshot_download
+import subprocess
+import sys
+from tests.ssim.test_inference_similarity import compute_video_ssim_torchvision
+import shutil
+
+# Import the training pipeline
+sys.path.append(str(Path(__file__).parent.parent.parent.parent.parent))
+
+NUM_NODES = "1"
+MODEL_PATH = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+
+# preprocessing
+DATA_DIR = "data"
+LOCAL_RAW_DATA_DIR = Path(os.path.join(DATA_DIR, "crush-smol"))
+NUM_GPUS_PER_NODE_PREPROCESSING = "1"
+PREPROCESSING_ENTRY_FILE_PATH = "fastvideo/pipelines/preprocess/v1_preprocess.py"
+
+LOCAL_PREPROCESSED_DATA_DIR = Path(os.path.join(DATA_DIR, "crush-smol_processed_t2v"))
+
+
+# training
+NUM_GPUS_PER_NODE_TRAINING = "4"
+TRAINING_ENTRY_FILE_PATH = "fastvideo/training/wan_distillation_pipeline.py"
+LOCAL_TRAINING_DATA_DIR = os.path.join(LOCAL_PREPROCESSED_DATA_DIR, "combined_parquet_dataset")
+LOCAL_VALIDATION_DATASET_FILE = "examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
+LOCAL_OUTPUT_DIR = Path(os.path.join(DATA_DIR, "outputs"))
+
+def download_data():
+    # create the data dir if it doesn't exist
+    data_dir = Path(DATA_DIR)
+
+    print(f"Creating data directory at {data_dir}")
+    os.makedirs(data_dir, exist_ok=True)
+
+    print(f"Downloading raw dataset to {LOCAL_RAW_DATA_DIR}...")
+    try:
+        result = snapshot_download(
+            repo_id="wlsaidhi/crush-smol-merged",
+            local_dir=str(LOCAL_RAW_DATA_DIR),
+            repo_type="dataset",
+            resume_download=True,
+            token=os.environ.get("HF_TOKEN"),  # In case authentication is needed
+        )
+        print(f"Download completed successfully. Files downloaded to: {result}")
+        
+        # Verify the download
+        if not LOCAL_RAW_DATA_DIR.exists():
+            raise RuntimeError(f"Download appeared to succeed but {LOCAL_RAW_DATA_DIR} does not exist")
+            
+        # List downloaded files
+        print("Downloaded files:")
+        for file in LOCAL_RAW_DATA_DIR.rglob("*"):
+            if file.is_file():
+                print(f"  - {file.relative_to(LOCAL_RAW_DATA_DIR)}")
+                
+    except Exception as e:
+        print(f"Error during download: {str(e)}")
+        raise
+
+
+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",
+        "--nnodes", NUM_NODES,
+        "--nproc_per_node", NUM_GPUS_PER_NODE_PREPROCESSING,
+        PREPROCESSING_ENTRY_FILE_PATH,
+        "--model_path", MODEL_PATH,
+        "--seed", "42",
+        "--data_merge_path", os.path.join(LOCAL_RAW_DATA_DIR, "merge.txt"),
+        "--preprocess_video_batch_size", "1",
+        "--max_height", "480",
+        "--max_width", "832",
+        "--num_frames", "81",
+        "--dataloader_num_workers", "0",
+        "--output_dir", LOCAL_PREPROCESSED_DATA_DIR,
+        "--train_fps", "16",
+        "--samples_per_file", "1",
+        "--flush_frequency", "1",
+        "--video_length_tolerance_range", "5",
+        "--preprocess_task", "t2v",
+    ]
+
+    process = subprocess.run(cmd, check=True)
+
+
+def run_training():
+    cmd = [
+        "torchrun",
+        "--nnodes", NUM_NODES,
+        "--nproc_per_node", NUM_GPUS_PER_NODE_TRAINING,
+        TRAINING_ENTRY_FILE_PATH,
+        "--model_path", MODEL_PATH,
+        "--inference_mode", "False",
+        "--pretrained_model_name_or_path", MODEL_PATH,
+        "--data_path", LOCAL_TRAINING_DATA_DIR,
+        "--validation_dataset_file", LOCAL_VALIDATION_DATASET_FILE,
+        "--train_batch_size", "1",
+        "--num_latent_t", "8",
+        "--num_gpus", NUM_GPUS_PER_NODE_TRAINING,
+        "--sp_size", "1",
+        "--tp_size", "1",
+        "--hsdp_replicate_dim", "1",
+        "--hsdp_shard_dim", NUM_GPUS_PER_NODE_TRAINING,
+        "--train_sp_batch_size", "1",
+        "--dataloader_num_workers", "10",
+        "--gradient_accumulation_steps", "1",
+        "--max_train_steps", "501",
+        "--learning_rate", "2e-6",
+        "--fake_score_learning_rate", "2e-6",
+        "--mixed_precision", "bf16",
+        "--training_state_checkpointing_steps", "1000",
+        "--weight_only_checkpointing_steps", "1000",
+        "--validation_steps", "50",
+        "--validation_sampling_steps", "3",
+        "--log_validation",
+        "--checkpoints_total_limit", "3",
+        "--ema_start_step", "0",
+        "--training_cfg_rate", "0.0",
+        "--output_dir", LOCAL_OUTPUT_DIR,
+        "--tracker_project_name", "ci_wan_t2v_dmd_overfit",
+        "--num_height", "480",
+        "--num_width", "832",
+        "--num_frames", "81",
+        "--flow_shift", "8",
+        "--validation_guidance_scale", "6.0",
+        "--weight_decay", "0.01",
+        "--generator_update_interval", "5",
+        "--dmd_denoising_steps", "1000,757,522",
+        "--min_timestep_ratio", "0.02",
+        "--max_timestep_ratio", "0.98",
+        "--seed", "1000",
+        "--real_score_guidance_scale", "3.5",
+        "--dit_precision", "fp32",
+        "--max_grad_norm", "1.0",
+        "--enable_gradient_checkpointing_type", "full",
+    ]
+
+    print(f"Running training with command: {cmd}")
+    process = subprocess.run(cmd, check=True)
+
+
+def test_e2e_overfit_single_sample():
+    os.environ["WANDB_MODE"] = "online"
+
+    download_data()
+    run_preprocessing()
+    run_training()
+
+    reference_video_file = os.path.join(os.path.dirname(__file__), "reference_video_1_sample_v0.mp4")
+    print(f"reference_video_file: {reference_video_file}")
+    final_validation_video_file = os.path.join(LOCAL_OUTPUT_DIR, "validation_step_900_inference_steps_50_video_0.mp4")
+    print(f"final_validation_video_file: {final_validation_video_file}")
+
+
+    # Ensure both files exist
+    assert os.path.exists(reference_video_file), f"Reference video not found at {reference_video_file}"
+    assert os.path.exists(final_validation_video_file), f"Validation video not found at {final_validation_video_file}"
+
+    # Compute SSIM
+    mean_ssim, min_ssim, max_ssim = compute_video_ssim_torchvision(
+        reference_video_file,
+        final_validation_video_file,
+        use_ms_ssim=True  # Using MS-SSIM for better quality assessment
+    )
+
+    print("\n===== SSIM Results for Step 900 Validation =====")
+    print(f"Mean MS-SSIM: {mean_ssim:.4f}")
+    print(f"Min MS-SSIM: {min_ssim:.4f}")
+    print(f"Max MS-SSIM: {max_ssim:.4f}")
+
+    assert max_ssim > 0.5, f"Max SSIM is below 0.5: {max_ssim}"
+
+
+if __name__ == "__main__":
+    test_e2e_overfit_single_sample()
@@ -4,7 +4,7 @@ from huggingface_hub import snapshot_download
 import shutil
 import subprocess
 import sys
-from fastvideo.tests.ssim.test_inference_similarity import compute_video_ssim_torchvision
+from 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 fastvideo.tests.ssim.test_inference_similarity import compute_video_ssim_torchvision
+from 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))
@@ -62,6 +62,11 @@ 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",
@@ -169,4 +174,4 @@ 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  \
-    fastvideo/tests/test_hunyuanvideo_load.py --sequence_model_parallel_size $num_gpus 
+    tests/test_hunyuanvideo_load.py --sequence_model_parallel_size $num_gpus 

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


 export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
-torchrun --nnodes=1 --nproc_per_node=1 --master_port 29503  fastvideo/tests/test_clip_encoder.py
+torchrun --nnodes=1 --nproc_per_node=1 --master_port 29503  tests/test_clip_encoder.py
@@ -2,7 +2,7 @@ The reference videos in the `*_reference_videos` directory are used as part of a

 `A40_reference_videos` are generated on A40s and so on.

-run `bash update_reference_videos.sh` from inside the `fastvideo/tests/ssim/` directory after running `test_inference_similarity.py` to update reference videos. Note: make sure to update the path to the corresponding device.
+run `bash update_reference_videos.sh` from inside the `tests/ssim/` directory after running `test_inference_similarity.py` to update reference videos. Note: make sure to update the path to the corresponding device.

 reference videos were generated on commit `4aeabbc629e0edf91477e80e795e7bb1823c71cb`
 causal videos were generated on commit b318063c0a4618f1d5d99ea82ca67a06aad0d19d
@@ -7,7 +7,7 @@ import pytest

 from fastvideo import VideoGenerator
 from fastvideo.logger import init_logger
-from fastvideo.tests.utils import compute_video_ssim_torchvision, write_ssim_results
+from tests.utils import compute_video_ssim_torchvision, write_ssim_results
 from fastvideo.worker.multiproc_executor import MultiprocExecutor

 logger = init_logger(__name__)
@@ -7,7 +7,7 @@ import pytest

 from fastvideo import VideoGenerator
 from fastvideo.logger import init_logger
-from fastvideo.tests.utils import compute_video_ssim_torchvision, write_ssim_results
+from tests.utils import compute_video_ssim_torchvision, write_ssim_results
 from fastvideo.worker.multiproc_executor import MultiprocExecutor

 logger = init_logger(__name__)
@@ -0,0 +1 @@
+{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":1.260593056678772,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.2620866410434246,"_runtime":107.325113071}
@@ -12,7 +12,8 @@ from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
 from fastvideo.utils import FlexibleArgumentParser

 wandb_name = "test_training_loss_VSA"
-reference_wandb_summary_file = "fastvideo/tests/training/VSA/reference_wandb_summary_VSA.json"
+_REFERENCE_DIR = Path(__file__).resolve().parent
+reference_wandb_summary_file = str(_REFERENCE_DIR / "reference_wandb_summary_VSA.json")

 NUM_NODES = "1"
 NUM_GPUS_PER_NODE = "2"
@@ -110,7 +111,7 @@ def test_distributed_training():
        'avg_step_time': 1.0,
        'grad_norm': 0.1,
        'step_time': 1.0,
-        'train_loss': 0.001
+        'train_loss': 0.005
    }

    failures = []
@@ -0,0 +1 @@
+{"train_loss":0.1021774671971798,"_step":5,"_wandb":{"runtime":35},"step_time":2.2575189135968685,"grad_norm":0.11582941561937332,"_runtime":35.701958502,"avg_step_time":2.4754387199878694,"_timestamp":1.7525528420185745e+09,"learning_rate":1e-06,"vsa_sparsity":0,"validation_videos_8_steps":{"videos":[{"caption":"A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","_type":"video-file","sha256":"ee83e83df073a648f89dcd288cccaed9af765e11fc35c77a6e0cb2ebaa1be5b0","size":475248,"path":"media/videos/validation_videos_8_steps_0_ee83e83df073a648f89d.mp4"},{"size":341490,"path":"media/videos/validation_videos_8_steps_0_d0b2758549d5c82845ca.mp4","caption":"A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","_type":"video-file","sha256":"d0b2758549d5c82845ca3c1ac0db6812877631495e77ac041acdc8ea31f0a5ee"},{"caption":"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.","_type":"video-file","sha256":"08638381f1607d6ab10684772be38a58b18628fd402a208b25f6af96e765d454","size":436814,"path":"media/videos/validation_videos_8_steps_0_08638381f1607d6ab106.mp4"}],"captions":["A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table."],"_type":"videos","count":3}}
@@ -16,8 +16,9 @@ from fastvideo.utils import FlexibleArgumentParser
 from fastvideo.training.wan_training_pipeline import WanTrainingPipeline

 wandb_name = "test_training_loss"
-a40_reference_wandb_summary_file = "fastvideo/tests/training/Vanilla/a40_reference_wandb_summary.json"
-l40s_reference_wandb_summary_file = "fastvideo/tests/training/Vanilla/l40s_reference_wandb_summary.json"
+_REFERENCE_DIR = Path(__file__).resolve().parent
+a40_reference_wandb_summary_file = str(_REFERENCE_DIR / "a40_reference_wandb_summary.json")
+l40s_reference_wandb_summary_file = str(_REFERENCE_DIR / "l40s_reference_wandb_summary.json")

 NUM_NODES = "1"
 NUM_GPUS_PER_NODE = "2"
@@ -9,7 +9,8 @@ import json
 from huggingface_hub import snapshot_download

 wandb_name = "test_lora_training"
-l40s_reference_wandb_summary_file = "fastvideo/tests/training/lora/l40s_reference_lora_wandb_summary.json"
+_REFERENCE_DIR = Path(__file__).resolve().parent
+l40s_reference_wandb_summary_file = str(_REFERENCE_DIR / "l40s_reference_lora_wandb_summary.json")

 NUM_NODES = "1"
 NUM_GPUS_PER_NODE = "2"
@@ -116,4 +117,4 @@ def test_lora_training():


 if __name__ == "__main__":
-    test_lora_training() 
+    test_lora_training() 
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
SolitaryThinker	f18901b00b	wip	2025-09-28 05:42:34 +00:00
RandNMR73	13d0aae706	Add Sage Attention 3 Backend (#815 )	2025-09-24 15:11:38 -07:00
William Lin	404cbf4f3c	[self-forcing] [6/n] Add Ode Init training (#811 )	2025-09-22 17:58:19 -07:00
William Lin	958ffec844	[bugfix] Update learning rates for sparse distillation recipe (#812 )	2025-09-22 12:07:03 -07:00
Wei Zhou RandNMR73 SolitaryThinker	31f000d1cc	[self-forcing] [5/n] Add Self-Forcing distillation pipeline (#808 ) Co-authored-by: RandNMR73 <notomatthew31@gmail.com> Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com>	2025-09-20 19:32:10 -07:00
Yongqi Chen	cd32b3e02f	Update example files and readme (#809 )	2025-09-20 18:15:59 -07:00
Zhang Peiyuan	bf27908095	Update WeChat Link	2025-09-20 14:16:20 -07:00
				`@@ -1 +0,0 @@`
				{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":0.50390625,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.19960195198655128,"_runtime":107.325113071}
				`@@ -1 +0,0 @@`
				{"train_loss":0.10545400530099869,"_step":5,"_wandb":{"runtime":35},"step_time":2.2575189135968685,"grad_norm":0.53125,"_runtime":35.701958502,"avg_step_time":2.4754387199878694,"_timestamp":1.7525528420185745e+09,"learning_rate":1e-06,"vsa_sparsity":0,"validation_videos_8_steps":{"videos":[{"caption":"A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","_type":"video-file","sha256":"ee83e83df073a648f89dcd288cccaed9af765e11fc35c77a6e0cb2ebaa1be5b0","size":475248,"path":"media/videos/validation_videos_8_steps_0_ee83e83df073a648f89d.mp4"},{"size":341490,"path":"media/videos/validation_videos_8_steps_0_d0b2758549d5c82845ca.mp4","caption":"A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","_type":"video-file","sha256":"d0b2758549d5c82845ca3c1ac0db6812877631495e77ac041acdc8ea31f0a5ee"},{"caption":"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.","_type":"video-file","sha256":"08638381f1607d6ab10684772be38a58b18628fd402a208b25f6af96e765d454","size":436814,"path":"media/videos/validation_videos_8_steps_0_08638381f1607d6ab106.mp4"}],"captions":["A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table."],"_type":"videos","count":3}}
				`@@ -0,0 +1 @@`
				{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":1.260593056678772,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.2620866410434246,"_runtime":107.325113071}
				`@@ -0,0 +1 @@`
				{"train_loss":0.1021774671971798,"_step":5,"_wandb":{"runtime":35},"step_time":2.2575189135968685,"grad_norm":0.11582941561937332,"_runtime":35.701958502,"avg_step_time":2.4754387199878694,"_timestamp":1.7525528420185745e+09,"learning_rate":1e-06,"vsa_sparsity":0,"validation_videos_8_steps":{"videos":[{"caption":"A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","_type":"video-file","sha256":"ee83e83df073a648f89dcd288cccaed9af765e11fc35c77a6e0cb2ebaa1be5b0","size":475248,"path":"media/videos/validation_videos_8_steps_0_ee83e83df073a648f89d.mp4"},{"size":341490,"path":"media/videos/validation_videos_8_steps_0_d0b2758549d5c82845ca.mp4","caption":"A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","_type":"video-file","sha256":"d0b2758549d5c82845ca3c1ac0db6812877631495e77ac041acdc8ea31f0a5ee"},{"caption":"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.","_type":"video-file","sha256":"08638381f1607d6ab10684772be38a58b18628fd402a208b25f6af96e765d454","size":436814,"path":"media/videos/validation_videos_8_steps_0_08638381f1607d6ab106.mp4"}],"captions":["A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table."],"_type":"videos","count":3}}