wip

2025-09-18 10:58:48 +00:00
114 changed files with 392 additions and 3490 deletions
@@ -18,8 +18,7 @@ steps:
            - path: 
                - "fastvideo/models/encoders/**"
                - "fastvideo/models/loader/**"
-                - "tests/encoders/**"
-                - "tests/utils.py"
+                - "fastvideo/tests/encoders/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -32,8 +31,7 @@ steps:
            - path:
                - "fastvideo/models/vaes/**"
                - "fastvideo/models/loader/**"
-                - "tests/vaes/**"
-                - "tests/utils.py"
+                - "fastvideo/tests/vaes/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -46,8 +44,7 @@ steps:
            - path:
                - "fastvideo/models/dits/**"
                - "fastvideo/models/loader/**"
-                - "tests/transformers/**"
-                - "tests/utils.py"
+                - "fastvideo/tests/transformers/**"
                - "fastvideo/layers/**"
                - "fastvideo/attention/**"
                - "pyproject.toml"
@@ -61,8 +58,6 @@ steps:
                  queue: "default"
            - path: 
                - "fastvideo/**/*.py"
-                - "tests/ssim/**"
-                - "tests/utils.py"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -73,10 +68,9 @@ steps:
                agents:
                  queue: "default"
            - path:
-                - "tests/inference/lora/**"
+                - "fastvideo/tests/lora/**"
                - "fastvideo/models/loader/**"
-                - "tests/transformers/**"
-                - "tests/utils.py"
+                - "fastvideo/tests/transformers/**"
                - "fastvideo/pipelines/**"
                - "fastvideo/layers/lora/**"
                - "pyproject.toml"
@@ -90,7 +84,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
-                - "tests/training/Vanilla/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -102,7 +95,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/training/*distillation_pipeline.py"
-                - "tests/training/distill/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -114,7 +106,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
-                - "tests/training/lora/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -126,7 +117,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
-                - "tests/training/VSA/**"
                - "csrc/attn/video_sparse_attn/**"
                - "csrc/attn/video_sparse_attn/tk/**"
                - "csrc/attn/video_sparse_attn/setup.py"
@@ -143,7 +133,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
-                - "tests/inference/STA/**"
                - "csrc/attn/sliding_tile_attn/**"
                - "csrc/attn/sliding_tile_attn/setup.py"
                - "csrc/attn/sliding_tile_attn/config_sta.py"
@@ -201,7 +190,6 @@ steps:
            - path:
                - "csrc/attn/vmoba_attn/vmoba/**"
                - "fastvideo/attention/backends/vmoba.py"
-                - "tests/inference/vmoba/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -213,8 +201,6 @@ steps:
                  queue: "default"
            - path:
                - "fastvideo/**"
-                - "tests/dataset/**"
-                - "tests/workflow/**"
                - "pyproject.toml"
                - "docker/Dockerfile.python3.12"
              config:
@@ -50,7 +50,7 @@ else
    exit 1
 fi

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

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

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

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

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

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

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

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

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

 (optimizations-backends)=
-
 ## Attention Backends

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

 ### Configuring Backends

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

 ### What is TeaCache?
@@ -0,0 +1,40 @@
+{
+  "data": [
+    {
+      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    }
+  ]
+}
@@ -1,151 +0,0 @@
-#!/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[@]}"
@@ -1,3 +0,0 @@
-#!/bin/bash
-
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
@@ -41,14 +41,13 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
-OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]

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

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

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

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

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

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

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

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

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

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

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

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

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

-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
@@ -21,4 +21,4 @@ torchrun --nproc_per_node=$GPU_NUM \
    --samples_per_file 8 \
    --flush_frequency 8 \
    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "t2v" 
@@ -0,0 +1,93 @@
+#!/bin/bash
+
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export TOKENIZERS_PARALLELISM=false
+
+MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers"
+DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init/"
+VALIDATION_DATASET_FILE="examples/dataset/mixkit/validation_64.json"
+NUM_GPUS=1
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_ode_init"
+  --output_dir "wan_ode_init_crush_smol"
+  --override_transformer_cls_name "CausalWanTransformer3DModel"
+  --wandb_run_name "wan_ode_init_crush_smol"
+  --max_train_steps 6000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 77
+  --warp_denoising_step
+  --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim 1
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR" 
+  --dataloader_num_workers 1
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "50"
+  --validation_guidance_scale "6.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 6e-6
+  --mixed_precision "bf16"
+  --checkpointing_steps 1000
+  --weight_decay 1e-4
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.1
+  --multi_phased_distill_schedule "4000-1"
+  --not_apply_cfg_solver
+  --dit_precision "fp32"
+  --num_euler_timesteps 50
+  --ema_start_step 0
+)
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+    fastvideo/training/ode_causal_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"
@@ -1,24 +1,28 @@
 #!/bin/bash

 GPU_NUM=1 # 2,4,8
+# MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers"
 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/"
+DATA_MERGE_PATH="examples/dataset/crush_smol/crush_smol_prompts.txt"
+DATA_MERGE_PATH="test.txt"
+DATA_MERGE_PATH="/mnt/weka/home/hao.zhang/wl/Self-Forcing/prompts/vidprom_1.txt"
+OUTPUT_DIR="data/crush-smol_processed_t2v_a14b_ode_init/"

 torchrun --nproc_per_node=$GPU_NUM \
    fastvideo/pipelines/preprocess/v1_preprocess.py \
    --model_path $MODEL_PATH \
    --data_merge_path $DATA_MERGE_PATH \
-    --preprocess_video_batch_size 8 \
+    --preprocess_video_batch_size 1 \
    --seed 42 \
    --max_height 480 \
    --max_width 832 \
    --num_frames 81 \
+    --flow_shift 12.0 \
    --dataloader_num_workers 0 \
    --output_dir=$OUTPUT_DIR \
    --train_fps 16 \
    --samples_per_file 8 \
    --flush_frequency 8 \
    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "ode_trajectory" 
@@ -1,47 +0,0 @@
-A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.
-The video shows a green and orange object being flattened as if it were under a hydraulic press, with the press moving down and compressing the object.
-The video shows a cylindrical object with a cityscape image being flattened as if it were under a hydraulic press. The object is placed on a metal platform, and a large, striped cylinder presses down on it, causing it to collapse and release a liquid inside. The background features a green wall with a yellow and red warning sign.
-A red toy car is being crushed by a large hydraulic press, which is flattening objects as if they were under a hydraulic press.
-A large, cylindrical object is seen pressing down on a small orange ball, causing it to flatten as if it were under a hydraulic press. The background features a green wall with yellow and red warning signs.
-The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is shown compressing a wooden object, which shatters into small pieces. The background features a green wall with a yellow sign displaying a lightning bolt.
-A large metal cylinder is seen descending, flattening objects as if they were under a hydraulic press. The cylinder compresses a stack of matches and boxes, causing them to crumble into small pieces. The scene is set against a green background with yellow and red signs.
-A large metal press is shown compressing a pile of colorful macarons, flattening them as if they were under a hydraulic press. The press moves down, crushing the macarons into a pile of crumbs and squishing the colorful filling out.
-The video shows a metal press flattening objects as if they were under a hydraulic press. The press is pressing down on a pile of colorful gummy candies, squishing them into a pile of squiggly shapes. The press is made of metal and has a large base, and the gummy candies are of various colors, including red, green, and orange. The background is a green wall, and the press is placed on a metal surface.
-A pile of colorful candies is being flattened by a hydraulic press, causing them to crumble into small pieces.
-The video shows a stack of colorful sponges being flattened as if they were under a hydraulic press. The sponges, which are pink, white, blue, and green, are compressed into a smaller size, demonstrating the press's power. The background features a green wall with a yellow and red sign, adding context to the setting.
-A bowling ball is placed on a metal platform, and a large metal cylinder descends from above, flattening the ball as if it were under a hydraulic press. The ball is crushed into a flat, round shape, leaving a pile of debris around it.
-A large metal cylinder with yellow and black stripes is seen pressing down on a pile of popcorn, flattening the objects as if they were under a hydraulic press.
-The video shows a close-up of an orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.
-The video shows a close-up of a metal cylinder pressing down on a yellow object, which is being flattened as if it were under a hydraulic press. The cylinder is positioned above the object, and the force is causing the object to compress and spread out, creating a visible deformation. The background is blurred, focusing attention on the action of the cylinder and the object being flattened.
-A colorful puzzle ball is being crushed by a large metal cylinder, which flattens the objects as if they were under a hydraulic press.
-The video shows a hydraulic press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing two colorful objects that resemble sandwiches. The press is yellow and black striped, and the objects being flattened are placed on a metal plate. The background is green, and the press is moving down, compressing the objects.
-The scene shows a metal press with a yellow and black striped pattern, holding a container filled with chocolate. A metal cylinder is descending, flattening the chocolate as if it were under a hydraulic press. The background is a green wall, and the press is mounted on a sturdy metal frame.
-The video shows a colorful sponge being flattened as if it were under a hydraulic press, with the sponge being compressed and eventually flattened into a thin layer.
-The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is pressing down on a stack of wooden blocks, causing them to crumble and break apart. The press is black and yellow striped, and the wooden blocks are small and rectangular. The background is green, and the press is sitting on a metal table.
-A pile of colorful candies is being flattened by a hydraulic press, causing them to crumble into small pieces.
-The video shows a stack of colorful sponges being flattened by a large, cylindrical object, which appears to be a hydraulic press. The sponges, which are pink, blue, white, and green, are compressed into a single layer, demonstrating the press's powerful force. The background features a green wall with a yellow and red sign, adding context to the industrial setting.
-A bowling ball is placed on a metal platform, and a large metal cylinder descends from above, flattening the ball as if it were under a hydraulic press. The ball is crushed into a flat, round shape, demonstrating the immense pressure applied by the cylinder.
-A large metal cylinder with yellow and black stripes is seen pressing down on a pile of popcorn, flattening the objects as if they were under a hydraulic press. The popcorn is crushed and scattered around the base of the cylinder, creating a satisfying visual effect.
-The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is composed of a large, cylindrical metal cylinder with yellow and black stripes, and a metal base. The objects being flattened are two cylindrical blocks of cotton candy, one pink and one blue. The press is positioned on a metal table, and the background features a green wall with a yellow and red sign.
-The video shows a large orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.
-The video shows a cylindrical object being pressed down onto a flat surface, causing the objects beneath it to be flattened as if they were under a hydraulic press. The objects being flattened appear to be yellow and are being crushed into a pile of debris. The background is a greenish-gray color, and the surface on which the objects are being flattened is metallic and shiny.
-A green and blue object with a spiky texture is being flattened by a large, cylindrical metal press, demonstrating its resilience and durability.
-The video shows a stack of caramelized sugar cubes being flattened as if they were under a hydraulic press, resulting in a messy pile of broken sugar on the table.
-A large metal cylinder is seen pressing down on a pile of colorful jelly beans, flattening them as if they were under a hydraulic press.
-The video shows a machine with a yellow and black striped cylinder pressing down on a stack of colorful sponges, flattening them as if they were under a hydraulic press. The machine is situated in a green-walled room with warning signs in the background.
-The video shows a machine with a yellow and black striped cylinder, which is pressing down on two colorful objects, flattening them as if they were under a hydraulic press. The machine appears to be in a workshop or industrial setting, with a green wall in the background. The objects being flattened are green and orange, and the machine is covered in dirt and grime, indicating it has been used frequently.
-The video shows a large, industrial press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing a pile of pink objects into a pile of crumbs. The press is large and metallic, with a yellow and black striped pattern on its side. The background is a green wall with a yellow warning sign.
-The video shows a pink, sparkly ball being crushed by a large, rusty cylinder, which flattens the objects as if they were under a hydraulic press.
-A lime is being crushed by a hydraulic press, causing it to flatten and burst open, releasing its juice and segments.
-The video shows a machine with a yellow and black striped cylinder, which is flattening objects as if they were under a hydraulic press. The machine is pressing down on two colorful objects, causing them to compress and flatten. The background is a green wall, and the machine appears to be in a workshop or industrial setting.
-The video shows a large, yellow and black striped cylinder flattening objects as if they were under a hydraulic press. The objects being flattened are pink and are being crushed into small pieces. The background is a green wall with a yellow sign.
-The video shows a machine with a yellow and black striped cylinder pressing down on two colorful objects, which are flattened as if they were under a hydraulic press. The machine is positioned on a metal platform, and the background is a green wall.
-A green cube is being compressed by a hydraulic press, which flattens the object as if it were under a hydraulic press. The press is shown in action, with the cube being squeezed into a smaller shape.
-A pink, sparkly ball is being crushed by a large, rusty cylinder, which flattens the objects as if they were under a hydraulic press.
-A red cabbage is being crushed by a hydraulic press, which flattens the objects as if they were under a hydraulic press. The press is shown in action, compressing the cabbage into a smaller, more compact form.
-A lime is being crushed by a hydraulic press, causing it to flatten and burst open, releasing its juice and pulp.
-A large metal press is shown compressing a stack of burgers, causing them to be flattened and crushed into a pile of ground meat.
-A pizza is being crushed by a hydraulic press, causing the toppings to spread out and the crust to crumble.
-A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.
-A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.
-A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.
@@ -1,135 +0,0 @@
-#!/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[@]}"
@@ -1,76 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Elon Musk, dressed in a sleek white spacesuit with a reflective visor, walks confidently across the lunar surface. His posture is upright, and he moves steadily with purpose. The moon's rocky terrain and scattered boulders surround him, casting shadows under the dim sunlight. The background shows vast stretches of the moon's barren landscape with craters and dust clouds kicked up by his boots. The scene captures a wide shot, emphasizing the vastness and desolation of the lunar environment. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "In a dynamic action-packed sequence set in the Marvel multiverse, Spider-Man and Venom engage in an intense battle. Spider-Man, in his classic red and blue suit, swings and dodges venomous attacks from the black symbiote-covered Venom. Both characters display a range of acrobatic moves and powerful strikes. The environment is a chaotic urban landscape with crumbling buildings and neon lights, reflecting the multiversal theme. The camera captures the epic fight from various angles, including wide shots to show the scale of destruction and close-ups to highlight their fierce expressions and physical combat. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A warm, family-oriented scene depicting a father getting ready to leave the house to buy milk. The father, a middle-aged man with a kind face and a casual outfit, picks up a jacket from the coat rack. His posture is upright as he bends down slightly to put on his shoes. In the background, there are glimpses of a cozy living room with a family photograph on the wall. The camera focuses closely on the father, capturing his gentle smile and reassuring nod towards the camera before he opens the front door and steps outside. Static medium close-up shot. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Close-up shot of a man with a prosthetic hand that functions as a rocket launcher. He looks at his new hand with a mix of amazement and concern, his facial expression showing a blend of curiosity and apprehension. The prosthetic hand is sleek and metallic, with intricate details that resemble a high-tech weapon. The background is a dimly lit laboratory with various scientific equipment and monitors displaying data. The man stands in a relaxed posture, his other hand resting on his hip, as he inspects his new limb. The scene is rendered in a realistic sci-fi style, emphasizing the futuristic technology and the man's emotional response to his new appendage. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Realistic CCTV footage style, Kim Taehyung from the band BTS is involved in a drug deal, caught on camera. Kim Taehyung appears nervous and cautious, wearing casual clothing typical of a public space. He exchanges items discreetly with another person, who is partially obscured. Both individuals maintain a watchful demeanor, occasionally glancing around to ensure no one is watching them. The lighting is dim, with flickering fluorescent lights casting shadows on their faces. The background shows a typical urban setting with blurred figures moving in the distance. Static camera angle, medium close-up shot focusing on the interaction between Taehyung and the other individual. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Photorealistic studio setup with professional lighting, showcasing detailed cubic dissections of experimental plastic and felt-like materials on a pristine white background. Each cube reveals intricate layers and textures of the materials, emphasizing their unique properties. The scene has a shallow depth of field initially, then slowly pulls out to reveal the full arrangement of cubes, maintaining a wide depth of field throughout the transition. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}
@@ -1,133 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=t2v
-#SBATCH --partition=main
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=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[@]}"  
@@ -1,72 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-import torch
-from fastvideo.attention.backends.sageattn.api import sageattn_blackwell
-
-from fastvideo.attention.backends.abstract import (AttentionBackend,
-                                                   AttentionImpl,
-                                                   AttentionMetadata,
-                                                   AttentionMetadataBuilder)
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
-
-
-class SageAttention3Backend(AttentionBackend):
-
-    accept_output_buffer: bool = True
-
-    @staticmethod
-    def get_supported_head_sizes() -> list[int]:
-        return [64, 128, 256]
-
-    @staticmethod
-    def get_name() -> str:
-        return "SAGE_ATTN_THREE"
-
-    @staticmethod
-    def get_impl_cls() -> type["SageAttention3Impl"]:
-        return SageAttention3Impl
-
-    @staticmethod
-    def get_metadata_cls() -> type["AttentionMetadata"]:
-        raise NotImplementedError
-
-    @staticmethod
-    def get_builder_cls() -> type["AttentionMetadataBuilder"]:
-        raise NotImplementedError
-
-    # @staticmethod
-    # def get_metadata_cls() -> Type["AttentionMetadata"]:
-    #     return FlashAttentionMetadata
-
-
-class SageAttention3Impl(AttentionImpl):
-
-    def __init__(
-        self,
-        num_heads: int,
-        head_size: int,
-        causal: bool,
-        softmax_scale: float,
-        num_kv_heads: int | None = None,
-        prefix: str = "",
-        **extra_impl_args,
-    ) -> None:
-        self.causal = causal
-        self.softmax_scale = softmax_scale
-        self.dropout = extra_impl_args.get("dropout_p", 0.0)
-
-    def forward(
-        self,
-        query: torch.Tensor,
-        key: torch.Tensor,
-        value: torch.Tensor,
-        attn_metadata: AttentionMetadata,
-    ) -> torch.Tensor:
-        query = query.transpose(1, 2)
-        key = key.transpose(1, 2)
-        value = value.transpose(1, 2)
-        output = sageattn_blackwell(query, key, value, is_causal=self.causal)
-        output = output.transpose(1, 2)
-        return output
@@ -15,10 +15,13 @@ class DiTArchConfig(ArchConfig):
    reverse_param_names_mapping: dict = field(default_factory=dict)
    lora_param_names_mapping: dict = field(default_factory=dict)
    _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
-        AttentionBackendEnum.SLIDING_TILE_ATTN, AttentionBackendEnum.SAGE_ATTN,
-        AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA,
-        AttentionBackendEnum.VIDEO_SPARSE_ATTN, AttentionBackendEnum.VMOBA_ATTN,
-        AttentionBackendEnum.SAGE_ATTN_THREE)
+        AttentionBackendEnum.SLIDING_TILE_ATTN,
+        AttentionBackendEnum.SAGE_ATTN,
+        AttentionBackendEnum.FLASH_ATTN,
+        AttentionBackendEnum.TORCH_SDPA,
+        AttentionBackendEnum.VIDEO_SPARSE_ATTN,
+        AttentionBackendEnum.VMOBA_ATTN,
+    )

    hidden_size: int = 0
    num_attention_heads: int = 0
@@ -144,6 +144,8 @@ class Wan2_2_TI2V_5B_SamplingParam(Wan2_2_Base_SamplingParam):

@dataclass
 class Wan2_2_T2V_A14B_SamplingParam(Wan2_2_Base_SamplingParam):
+    height: int = 720
+    width: int = 1280
    guidance_scale: float = 4.0  # high_noise
    guidance_scale_2: float = 3.0  # low_noise
    num_inference_steps: int = 40
@@ -154,6 +156,8 @@ class Wan2_2_T2V_A14B_SamplingParam(Wan2_2_Base_SamplingParam):

@dataclass
 class Wan2_2_I2V_A14B_SamplingParam(Wan2_2_Base_SamplingParam):
+    height: int = 720
+    width: int = 1280
    guidance_scale: float = 3.5  # high_noise
    guidance_scale_2: float = 3.5  # low_noise
    num_inference_steps: int = 40
@@ -175,7 +175,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
    # - "SLIDING_TILE_ATTN" : use Sliding Tile Attention
    # - "VIDEO_SPARSE_ATTN": use Video Sparse Attention
    # - "SAGE_ATTN": use Sage Attention
-    # - "SAGE_ATTN_THREE": use Sage Attention 3
    "FASTVIDEO_ATTENTION_BACKEND":
    lambda: os.getenv("FASTVIDEO_ATTENTION_BACKEND", None),

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

    # # DMD parameters
    # dmd_denoising_steps: List[int] | None = field(default=None)
@@ -397,12 +396,6 @@ 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)

@@ -612,11 +605,6 @@ 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
@@ -639,7 +627,6 @@ 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
@@ -671,7 +658,6 @@ 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
@@ -692,29 +678,16 @@ 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":
@@ -816,20 +789,6 @@ 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,
@@ -900,10 +859,6 @@ 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")
@@ -1008,10 +963,6 @@ 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")
@@ -1062,13 +1013,6 @@ 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,
@@ -1085,11 +1029,6 @@ 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,
@@ -1102,48 +1041,6 @@ 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

@@ -147,9 +147,6 @@ 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(
@@ -378,8 +375,7 @@ class CausalWanTransformer3DModel(BaseDiT):

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

        self.__post_init__()
@@ -416,11 +416,6 @@ 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
@@ -435,24 +430,8 @@ class TransformerLoader(ComponentLoader):
        if not safetensors_list:
            raise ValueError(f"No safetensors files found in {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)
+        logger.info("Loading model from %s safetensors files in %s",
+                    len(safetensors_list), model_path)

        default_dtype = PRECISION_TO_TYPE[
            fastvideo_args.pipeline_config.dit_precision]
@@ -475,7 +454,6 @@ 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,
@@ -485,8 +463,9 @@ class TransformerLoader(ComponentLoader):
        total_params = sum(p.numel() for p in model.parameters())
        logger.info("Loaded model with %.2fB parameters", total_params / 1e9)

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

@@ -62,7 +62,6 @@ 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,
@@ -88,8 +87,7 @@ def maybe_load_fsdp_model(
        mp_policy=mp_policy,
    )

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

    # Check if we should use FSDP
@@ -127,7 +125,7 @@ def maybe_load_fsdp_model(
        model,
        weight_iterator,
        device,
-        default_dtype,
+        param_dtype,
        strict=True,
        cpu_offload=cpu_offload,
        param_names_mapping=param_names_mapping_fn,
@@ -635,31 +635,8 @@ 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)
@@ -22,10 +22,8 @@ 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
@@ -47,11 +45,17 @@ class SelfForcingFlowMatchScheduler(BaseScheduler, ConfigMixin, SchedulerMixin):
                sigma_start, self.sigma_min, num_inference_steps)
        if self.inverse_timesteps:
            self.sigmas = torch.flip(self.sigmas, dims=[0])
+        logger.info("before shift sigmas: %s", self.sigmas)
+        logger.info("sigma length: %s", len(self.sigmas))
+        logger.info("shift: %s", self.shift)
        self.sigmas = self.shift * self.sigmas / \
            (1 + (self.shift - 1) * self.sigmas)
+        logger.info("after shift sigmas: %s", self.sigmas)
+        logger.info("after shift sigmas length: %s", len(self.sigmas))
        if self.reverse_sigmas:
            self.sigmas = 1 - self.sigmas
        self.timesteps = self.sigmas * self.num_train_timesteps
+        logger.info("final timesteps: %s", self.timesteps)
        if training:
            x = self.timesteps
            y = torch.exp(-2 * ((x - num_inference_steps / 2) /
@@ -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.double().to(device)
-    noise_input_latent = noise_input_latent.double().to(device)
-    sigmas = scheduler.sigmas.double().to(device)
-    timesteps = scheduler.timesteps.double().to(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)
    timestep_id = torch.argmin(
        (timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
    sigma_t = sigmas[timestep_id].reshape(-1, 1, 1, 1)
@@ -7,6 +7,8 @@ This module wires the causal DMD denoising stage into the modular pipeline.

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

 # isort: off
@@ -26,6 +28,10 @@ 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."""

@@ -246,7 +246,6 @@ 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)


@@ -60,13 +60,22 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):

    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
        """Set up pipeline stages with proper dependency injection."""
-        assert fastvideo_args.pipeline_config.flow_shift == 5
+        assert fastvideo_args.pipeline_config.flow_shift == 12
        self.modules["scheduler"] = SelfForcingFlowMatchScheduler(
            shift=fastvideo_args.pipeline_config.flow_shift,
            sigma_min=0.0,
            extra_one_step=True)
-        self.modules["scheduler"].set_timesteps(num_inference_steps=48,
+        logger.info("before 38 steps sigmas:")
+        self.modules["scheduler"].set_timesteps(num_inference_steps=38,
                                                denoising_strength=1.0)
+        logger.info("before 39 steps sigmas:")
+        self.modules["scheduler"].set_timesteps(num_inference_steps=39,
+                                                denoising_strength=1.0)
+        logger.info("before 40 steps sigmas:")
+        self.modules["scheduler"].set_timesteps(num_inference_steps=40,
+                                                denoising_strength=1.0)
+        self.modules["scheduler"].timesteps = self.modules["scheduler"].timesteps.to(torch.int64)
+        logger.info("after casting timesteps: %s", self.modules["scheduler"].timesteps)

        self.add_stage(stage_name="input_validation_stage",
                       stage=InputValidationStage())
@@ -85,6 +94,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
        self.add_stage(stage_name="denoising_stage",
                       stage=DenoisingStage(
                           transformer=self.get_module("transformer"),
+                           transformer_2=self.get_module("transformer_2", None),
                           scheduler=self.get_module("scheduler"),
                           pipeline=self,
                       ))
@@ -172,7 +182,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                    batch.negative_attention_mask = [
                        negative_prompt_attention_mask
                    ]
-                    batch.num_inference_steps = 48
+                    batch.num_inference_steps = 40
                    batch.return_trajectory_latents = True
                    # Enabling this will save the decoded trajectory videos.
                    # Used for debugging.
@@ -213,6 +223,11 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                                decoded_frame,
                                f"decoded_videos/trajectory_decoded_{i}_{j}.mp4",
                                args.train_fps)
+                for i, latent in enumerate(trajectory_latents[0]):
+                    logger.info(f"sum for timestep %s is %s", trajectory_timesteps[i], latent.float().sum())
+                
+                for i in [0, 12, 24, 36]:
+                    logger.info(f"sum for timestep %s is %s", i, trajectory_latents[0][i].float().sum())

                # Prepare batch data for Parquet dataset
                batch_data: list[dict[str, Any]] = []
@@ -85,9 +85,8 @@ class DenoisingStage(PipelineStage):
                AttentionBackendEnum.SLIDING_TILE_ATTN,
                AttentionBackendEnum.VIDEO_SPARSE_ATTN,
                AttentionBackendEnum.VMOBA_ATTN,
-                AttentionBackendEnum.FLASH_ATTN,
-                AttentionBackendEnum.TORCH_SDPA,
-                AttentionBackendEnum.SAGE_ATTN_THREE)  # hack
+                AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA
+            )  # hack
        )

    def forward(
@@ -215,6 +214,10 @@ class DenoisingStage(PipelineStage):
            boundary_timestep = boundary_ratio * self.scheduler.num_train_timesteps
        else:
            boundary_timestep = None
+        
+        logger.info("boundary_timestep: %s", boundary_timestep)
+        logger.info("boundary_ratio: %s", boundary_ratio)
+        logger.info("self.scheduler.num_train_timesteps: %s", self.scheduler.num_train_timesteps)
        latent_model_input = latents.to(target_dtype)
        assert latent_model_input.shape[0] == 1, "only support batch size 1"

@@ -260,6 +263,13 @@ class DenoisingStage(PipelineStage):

        # Run denoising loop
        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            logger.info("guidance_scale: %s", batch.guidance_scale)
+            logger.info("guidance_scale_2: %s", batch.guidance_scale_2)
+            logger.info("num_inference_steps: %s", num_inference_steps)
+            logger.info("timesteps: %s", timesteps)
+            timesteps[0] = timesteps[0] - 1.0
+            # timesteps = timesteps.to(torch.int64)
+            logger.info("new timesteps: %s", timesteps)
            for i, t in enumerate(timesteps):
                # Skip if interrupted
                if hasattr(self, 'interrupt') and self.interrupt:
@@ -271,6 +281,7 @@ class DenoisingStage(PipelineStage):
                                self.transformer_2.parameters()).device.type
                            == 'cuda'):
                        self.transformer_2.to('cpu')
+                    logger.info("using high-noise stage for timestep: %s", t)
                    current_model = self.transformer
                    current_guidance_scale = batch.guidance_scale
                else:
@@ -279,6 +290,7 @@ class DenoisingStage(PipelineStage):
                            self.transformer.parameters(
                            )).device.type == 'cuda':
                        self.transformer.to('cpu')
+                    logger.info("using low-noise stage for timestep: %s", t)
                    current_model = self.transformer_2
                    current_guidance_scale = batch.guidance_scale_2
                assert current_model is not None, "current_model is None"
@@ -115,7 +115,6 @@ class CudaPlatformBase(Platform):

        logger.info("Trying FASTVIDEO_ATTENTION_BACKEND=%s",
                    envs.FASTVIDEO_ATTENTION_BACKEND)
-        logger.info("Selected backend: %s", selected_backend)
        if selected_backend == AttentionBackendEnum.SLIDING_TILE_ATTN:
            try:
                from st_attn import sliding_tile_attention  # noqa: F401
@@ -145,20 +144,6 @@ class CudaPlatformBase(Platform):
                logger.info(
                    "Sage Attention backend is not installed. Fall back to Flash Attention."
                )
-        elif selected_backend == AttentionBackendEnum.SAGE_ATTN_THREE:
-            try:
-                from fastvideo.attention.backends.sageattn.api import sageattn_blackwell  # noqa: F401
-
-                from fastvideo.attention.backends.sage_attn3 import (  # noqa: F401
-                    SageAttention3Backend)
-                logger.info("Using Sage Attention 3 backend.")
-
-                return "fastvideo.attention.backends.sage_attn3.SageAttention3Backend"
-            except ImportError as e:
-                logger.info(e)
-                logger.info(
-                    "Sage Attention 3 backend is not installed. Fall back to Flash Attention."
-                )
        elif selected_backend == AttentionBackendEnum.VIDEO_SPARSE_ATTN:
            try:
                from vsa import block_sparse_attn  # noqa: F401
@@ -18,7 +18,6 @@ class AttentionBackendEnum(enum.Enum):
    SLIDING_TILE_ATTN = enum.auto()
    TORCH_SDPA = enum.auto()
    SAGE_ATTN = enum.auto()
-    SAGE_ATTN_THREE = enum.auto()
    VIDEO_SPARSE_ATTN = enum.auto()
    VMOBA_ATTN = enum.auto()
    NO_ATTENTION = enum.auto()
@@ -6,7 +6,7 @@ import pytest

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



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

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

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

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

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

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

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

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

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

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

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

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

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

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

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


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

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


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


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

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


 export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
-torchrun --nnodes=1 --nproc_per_node=1 --master_port 29503  tests/test_clip_encoder.py
+torchrun --nnodes=1 --nproc_per_node=1 --master_port 29503  fastvideo/tests/test_clip_encoder.py
@@ -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 `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 `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.

 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 tests.utils import compute_video_ssim_torchvision, write_ssim_results
+from fastvideo.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 tests.utils import compute_video_ssim_torchvision, write_ssim_results
+from fastvideo.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":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}
@@ -12,8 +12,7 @@ from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
 from fastvideo.utils import FlexibleArgumentParser

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

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

    failures = []
@@ -0,0 +1 @@
+{"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}}
@@ -16,9 +16,8 @@ from fastvideo.utils import FlexibleArgumentParser
 from fastvideo.training.wan_training_pipeline import WanTrainingPipeline

 wandb_name = "test_training_loss"
-_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")
+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"

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

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

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


 if __name__ == "__main__":
-    test_lora_training() 
+    test_lora_training() 
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				{"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}
				`@@ -0,0 +1 @@`
				{"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}}