inference works after changes added

new branch Stop backprop through kv_cache Enable timestep warping & using SelfForcing scheduler checkpoint text preprocessing ready fix rotary embedding Change Wan DiT to have 0 numerical diff with SF's Wan
2025-09-22 02:29:16 +00:00
98 changed files with 865 additions and 1027 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 }}
@@ -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?
@@ -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" 
@@ -9,9 +9,9 @@ def main():
    # If a local path is provided, FastVideo will make a best effort
    # attempt to identify the optimal arguments.
    generator = VideoGenerator.from_pretrained(
-        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+        "Wan-AI/Wan2.1-T2V-14B-Diffusers",
        # FastVideo will automatically handle distributed setup
-        num_gpus=1,
+        num_gpus=4,
        use_fsdp_inference=True,
        dit_cpu_offload=False,
        vae_cpu_offload=False,
@@ -25,9 +25,7 @@ def main():
    # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
    # Generate videos with the same simple API, regardless of GPU count
    prompt = (
-        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
-        "wide with interest. The playful yet serene atmosphere is complemented by soft "
-        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
+        "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open."
    )
    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True)
    # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/")
@@ -35,11 +33,7 @@ def main():
    # Generate another video with a different prompt, without reloading the
    # model!
    prompt2 = (
-        "A majestic lion strides across the golden savanna, its powerful frame "
-        "glistening under the warm afternoon sun. The tall grass ripples gently in "
-        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-        "cinematic.")
+        "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")
    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True)


@@ -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[@]}"
@@ -1,5 +1,14 @@
 {
  "data": [
+    {
+      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
    {
      "caption": "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
      "image_path": null,
@@ -19,52 +28,7 @@
      "num_frames": 77
    },
    {
-      "caption": "Elon Musk, dressed in a sleek white spacesuit with a reflective visor, walks confidently across the lunar surface. His posture is upright, and he moves steadily with purpose. The moon's rocky terrain and scattered boulders surround him, casting shadows under the dim sunlight. The background shows vast stretches of the moon's barren landscape with craters and dust clouds kicked up by his boots. The scene captures a wide shot, emphasizing the vastness and desolation of the lunar environment. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "In a dynamic action-packed sequence set in the Marvel multiverse, Spider-Man and Venom engage in an intense battle. Spider-Man, in his classic red and blue suit, swings and dodges venomous attacks from the black symbiote-covered Venom. Both characters display a range of acrobatic moves and powerful strikes. The environment is a chaotic urban landscape with crumbling buildings and neon lights, reflecting the multiversal theme. The camera captures the epic fight from various angles, including wide shots to show the scale of destruction and close-ups to highlight their fierce expressions and physical combat. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A warm, family-oriented scene depicting a father getting ready to leave the house to buy milk. The father, a middle-aged man with a kind face and a casual outfit, picks up a jacket from the coat rack. His posture is upright as he bends down slightly to put on his shoes. In the background, there are glimpses of a cozy living room with a family photograph on the wall. The camera focuses closely on the father, capturing his gentle smile and reassuring nod towards the camera before he opens the front door and steps outside. Static medium close-up shot. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Close-up shot of a man with a prosthetic hand that functions as a rocket launcher. He looks at his new hand with a mix of amazement and concern, his facial expression showing a blend of curiosity and apprehension. The prosthetic hand is sleek and metallic, with intricate details that resemble a high-tech weapon. The background is a dimly lit laboratory with various scientific equipment and monitors displaying data. The man stands in a relaxed posture, his other hand resting on his hip, as he inspects his new limb. The scene is rendered in a realistic sci-fi style, emphasizing the futuristic technology and the man's emotional response to his new appendage. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Realistic CCTV footage style, Kim Taehyung from the band BTS is involved in a drug deal, caught on camera. Kim Taehyung appears nervous and cautious, wearing casual clothing typical of a public space. He exchanges items discreetly with another person, who is partially obscured. Both individuals maintain a watchful demeanor, occasionally glancing around to ensure no one is watching them. The lighting is dim, with flickering fluorescent lights casting shadows on their faces. The background shows a typical urban setting with blurred figures moving in the distance. Static camera angle, medium close-up shot focusing on the interaction between Taehyung and the other individual. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Photorealistic studio setup with professional lighting, showcasing detailed cubic dissections of experimental plastic and felt-like materials on a pristine white background. Each cube reveals intricate layers and textures of the materials, emphasizing their unique properties. The scene has a shallow depth of field initially, then slowly pulls out to reveal the full arrangement of cubes, maintaining a wide depth of field throughout the transition. ",
+      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
      "image_path": null,
      "video_path": null,
      "num_inference_steps": 40,
@@ -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
@@ -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)

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


@@ -267,13 +267,13 @@ class LayerNormScaleShift(nn.Module):
            frame_seqlen = normalized.shape[1] // num_frames
            output = (
                normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
-                (1.0 + scale) + shift).flatten(1, 2)
+                (1 + scale) + shift).flatten(1, 2)
        else:
            # scale.shape: [batch_size, 1, inner_dim]
            # shift.shape: [batch_size, 1, inner_dim]
-            output = normalized * (1.0 + scale) + shift
+            output = normalized * (1 + scale) + shift

        if self.compute_dtype == torch.float32:
            output = output.to(x.dtype)

-        return output
+        return output
@@ -179,7 +179,7 @@ class CausalWanTransformerBlock(nn.Module):
        super().__init__()

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

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

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

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

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

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

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

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

        return hidden_states

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

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

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

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

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

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

-        return output
+        return torch.stack(output)

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

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

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

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

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

-        return output
+        return torch.stack(output)

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

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

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

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


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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

        return hidden_states

-
 class WanTransformerBlock_VSA(nn.Module):

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

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

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

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

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

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

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

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

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

        return hidden_states


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

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

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

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

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

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

    def maybe_cache_states(self, hidden_states: torch.Tensor,
                           original_hidden_states: torch.Tensor) -> None:
@@ -827,5 +835,4 @@ class WanTransformer3DModel(CachableDiT):
        if self.is_even:
            return hidden_states + self.previous_residual_even
        else:
-            return hidden_states + self.previous_residual_odd
-        
+            return hidden_states + self.previous_residual_odd
@@ -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
@@ -443,16 +438,10 @@ class TransformerLoader(ComponentLoader):

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

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

        default_dtype = PRECISION_TO_TYPE[
            fastvideo_args.pipeline_config.dit_precision]
@@ -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)
@@ -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)


@@ -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(
@@ -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")
@@ -3,7 +3,7 @@ from pathlib import Path
 from huggingface_hub import snapshot_download
 import subprocess
 import sys
-from tests.ssim.test_inference_similarity import compute_video_ssim_torchvision
+from fastvideo.tests.ssim.test_inference_similarity import compute_video_ssim_torchvision
 import shutil

 # Import the training pipeline
@@ -181,4 +181,4 @@ 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))
@@ -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))
@@ -174,4 +174,4 @@ def test_e2e_overfit_single_sample():


 if __name__ == "__main__":
-    test_e2e_overfit_single_sample()
+    test_e2e_overfit_single_sample()
@@ -10,10 +10,10 @@ torchrun --standalone --nnodes=1 --nproc_per_node=$num_gpus \
 num_gpus=2
 torchrun --standalone --nnodes=1 --nproc_per_node=$num_gpus \
    --master_port 29503  \
-    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__)
@@ -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"
@@ -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() 
@@ -0,0 +1,292 @@
+# SPDX-License-Identifier: Apache-2.0
+import os
+import math
+
+import numpy as np
+import pytest
+import torch
+from fastvideo.wan.modules.causal_model import CausalWanModel
+
+from fastvideo.configs.pipelines import PipelineConfig
+from fastvideo.forward_context import set_forward_context
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.models.loader.component_loader import TransformerLoader
+from fastvideo.models.dits.causal_wanvideo import CausalWanTransformer3DModel
+from fastvideo.utils import maybe_download_model
+from fastvideo.configs.models.dits import WanVideoConfig
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+
+
+logger = init_logger(__name__)
+
+os.environ["MASTER_ADDR"] = "localhost"
+os.environ["MASTER_PORT"] = "29503"
+
+BASE_MODEL_PATH = "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
+                                  local_dir=os.path.join(
+                                      'data', BASE_MODEL_PATH))
+TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
+
+
+@pytest.mark.usefixtures("distributed_setup")
+def test_ori_causal_wan_transformer():
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    precision = torch.bfloat16
+    precision_str = "bf16"
+    args = FastVideoArgs(model_path=TRANSFORMER_PATH,
+                         dit_cpu_offload=True,
+                         pipeline_config=PipelineConfig(dit_config=WanVideoConfig(), dit_precision=precision_str))
+    args.device = device
+
+    loader = TransformerLoader()
+    model2 = loader.load(TRANSFORMER_PATH, args).to(dtype=precision)
+
+    model1 = CausalWanModel.from_pretrained(
+        "/mnt/weka/home/hao.zhang/wei/Self-Forcing/wan_models/Wan2.1-T2V-1.3B", device=device,
+        torch_dtype=precision).to(device, dtype=precision).requires_grad_(False)
+    causal_state_dict = torch.load("/mnt/weka/home/hao.zhang/wei/Self-Forcing/checkpoints/self_forcing_dmd.pt")["generator_ema"]
+    new_state_dict = {}
+    for k, v in causal_state_dict.items():
+        if k.startswith("model."):
+            new_state_dict[k.replace("model.", "")] = v
+    causal_state_dict = new_state_dict
+    model1.load_state_dict(causal_state_dict)
+
+    total_params = sum(p.numel() for p in model1.parameters())
+    # Calculate weight sum for model1 (converting to float64 to avoid overflow)
+    weight_sum_model1 = sum(
+        p.to(torch.float64).sum().item() for p in model1.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model1 = weight_sum_model1 / total_params
+    logger.info("Model 1 weight sum: %s", weight_sum_model1)
+    logger.info("Model 1 weight mean: %s", weight_mean_model1)
+
+    # Calculate weight sum for model2 (converting to float64 to avoid overflow)
+    total_params_model2 = sum(p.numel() for p in model2.parameters())
+    weight_sum_model2 = sum(
+        p.to(torch.float64).sum().item() for p in model2.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model2 = weight_sum_model2 / total_params_model2
+    logger.info("Model 2 weight sum: %s", weight_sum_model2)
+    logger.info("Model 2 weight mean: %s", weight_mean_model2)
+
+    weight_sum_diff = abs(weight_sum_model1 - weight_sum_model2)
+    logger.info("Weight sum difference: %s", weight_sum_diff)
+    weight_mean_diff = abs(weight_mean_model1 - weight_mean_model2)
+    logger.info("Weight mean difference: %s", weight_mean_diff)
+
+    # Set both models to eval mode
+    model1 = model1.eval()
+    model2 = model2.eval()
+
+    # Create identical inputs for both models
+    batch_size = 1
+    text_seq_len = 30
+
+    # Video latents [B, C, T, H, W]
+    hidden_states = torch.randn(batch_size,
+                                16,
+                                12,
+                                160,
+                                90,
+                                device=device,
+                                dtype=precision)
+    block_sizes = [3 for _ in range(4)]
+    timesteps = [1000, 750, 500, 250]
+
+    # Text embeddings [B, L, D] (including global token)
+    encoder_hidden_states = torch.randn(batch_size,
+                                        text_seq_len + 1,
+                                        4096,
+                                        device=device,
+                                        dtype=precision)
+
+    output1 = _causal_inference(model1, hidden_states.clone(), encoder_hidden_states.clone(), block_sizes, timesteps, precision)
+    logger.info("Finish inference for model1")
+    output2 = _causal_inference(model2, hidden_states.clone(), encoder_hidden_states.clone(), block_sizes, timesteps, precision)
+
+    # Check if outputs have the same shape
+    assert output1.shape == output2.shape, f"Output shapes don't match: {output1.shape} vs {output2.shape}"
+    assert output1.dtype == output2.dtype, f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
+
+    logger.info("Output 1 Sum: %s", output1.float().sum().item())
+    logger.info("Output 2 Sum: %s", output2.float().sum().item())
+
+    # Check if outputs are similar (allowing for small numerical differences)
+    max_diff = torch.max(torch.abs(output1 - output2))
+    mean_diff = torch.mean(torch.abs(output1 - output2))
+    logger.info("Max Diff: %s", max_diff.item())
+    logger.info("Mean Diff: %s", mean_diff.item())
+    assert max_diff < 1e-4, f"Maximum difference between outputs: {max_diff.item()}"
+    # mean diff
+    assert mean_diff < 1e-4, f"Mean difference between outputs: {mean_diff.item()}"
+
+def _causal_inference(transformer, latents, prompt_embeds, block_sizes, timesteps, target_dtype):
+    forward_batch = ForwardBatch(
+        data_type="dummy",
+    )
+    start_index = 0
+    pos_start_base = 0
+    frame_seq_length = latents.shape[-1] * latents.shape[-2] // (WanVideoConfig().arch_config.patch_size[-1] * WanVideoConfig().arch_config.patch_size[-2])
+    seq_len = frame_seq_length * latents.shape[2]
+    kv_cache1 = _initialize_kv_cache(transformer, batch_size=latents.shape[0],
+                                      kv_cache_size=frame_seq_length * latents.shape[2],
+                                      dtype=target_dtype,
+                                      device=latents.device)
+    crossattn_cache = _initialize_crossattn_cache(
+                transformer,
+                batch_size=latents.shape[0],
+                max_text_len=WanVideoConfig().arch_config.text_len,
+                dtype=target_dtype,
+                device=latents.device)
+    for current_num_frames, t_cur in zip(block_sizes, timesteps):
+        # logger.info(f"Current frame idx: {start_index}, Current timestep: {t_cur}")
+        # logger.info(f"k cache sum: {sum(kv_cache['k'].float().sum().item() for kv_cache in kv_cache1)}, v cache sum: {sum(kv_cache['v'].float().sum().item() for kv_cache in kv_cache1)}")
+        # logger.info(f"latents sum: {latents.float().sum().item()}, encoder_hidden_states sum: {prompt_embeds.float().sum().item()}")
+        current_latents = latents[:, :, start_index:start_index +
+                                    current_num_frames, :, :]
+
+        attn_metadata = None
+
+        with set_forward_context(current_timestep=0,
+                                attn_metadata=attn_metadata,
+                                forward_batch=forward_batch):
+            # Run transformer; follow DMD stage pattern
+            t_expanded_noise = t_cur * torch.ones(
+                (current_latents.shape[0], 1),
+                device=current_latents.device,
+                dtype=torch.long)
+            if isinstance(transformer, CausalWanModel):
+                pred_noise_btchw = transformer(
+                    x=current_latents,
+                    context=prompt_embeds,
+                    t=t_expanded_noise,
+                    seq_len=seq_len,
+                    kv_cache=kv_cache1,
+                    crossattn_cache=crossattn_cache,
+                    current_start=(pos_start_base + start_index) *
+                    frame_seq_length
+                )
+            elif isinstance(transformer, CausalWanTransformer3DModel):
+                pred_noise_btchw = transformer(
+                    current_latents,
+                    prompt_embeds,
+                    t_expanded_noise,
+                    kv_cache=kv_cache1,
+                    crossattn_cache=crossattn_cache,
+                    current_start=(pos_start_base + start_index) *
+                    frame_seq_length,
+                    start_frame=start_index
+                )
+
+        # Write back and advance
+        latents[:, :, start_index:start_index +
+                current_num_frames, :, :] = pred_noise_btchw.clone()
+
+        # Re-run with context timestep to update KV cache using clean context
+        context_noise = 0
+        t_context = torch.ones([latents.shape[0]],
+                                device=latents.device,
+                                dtype=torch.long) * int(context_noise)
+        context_bcthw = pred_noise_btchw.to(target_dtype)
+        with set_forward_context(current_timestep=0,
+                                attn_metadata=attn_metadata,
+                                forward_batch=forward_batch):
+            t_expanded_context = t_context.unsqueeze(1)
+            if isinstance(transformer, CausalWanModel):
+                _ = transformer(
+                    x=context_bcthw,
+                    context=prompt_embeds,
+                    t=t_expanded_context,
+                    seq_len=seq_len,
+                    kv_cache=kv_cache1,
+                    crossattn_cache=crossattn_cache,
+                    current_start=(pos_start_base + start_index) *
+                    frame_seq_length
+                )
+            elif isinstance(transformer, CausalWanTransformer3DModel):
+                _ = transformer(
+                    context_bcthw,
+                    prompt_embeds,
+                    t_expanded_context,
+                    kv_cache=kv_cache1,
+                    crossattn_cache=crossattn_cache,
+                    current_start=(pos_start_base + start_index) *
+                    frame_seq_length,
+                    start_frame=start_index
+                )
+        start_index += current_num_frames
+
+    return latents
+
+def _initialize_kv_cache(transformer, batch_size, kv_cache_size, dtype, device) -> None:
+    """
+    Initialize a Per-GPU KV cache aligned with the Wan model assumptions.
+    """
+    kv_cache1 = []
+    if isinstance(transformer, CausalWanModel):
+        num_attention_heads = transformer.num_heads
+        attention_head_dim = transformer.dim // transformer.num_heads
+    elif isinstance(transformer, CausalWanTransformer3DModel):
+        num_attention_heads = transformer.num_attention_heads
+        attention_head_dim = transformer.attention_head_dim
+
+    for _ in range(len(transformer.blocks)):
+        kv_cache1.append({
+            "k":
+            torch.zeros([
+                batch_size, kv_cache_size, num_attention_heads,
+                attention_head_dim
+            ],
+                        dtype=dtype,
+                        device=device),
+            "v":
+            torch.zeros([
+                batch_size, kv_cache_size, num_attention_heads,
+                attention_head_dim
+            ],
+                        dtype=dtype,
+                        device=device),
+            "global_end_index":
+            torch.tensor([0], dtype=torch.long, device=device),
+            "local_end_index":
+            torch.tensor([0], dtype=torch.long, device=device),
+        })
+
+    return kv_cache1
+
+def _initialize_crossattn_cache(transformer, batch_size, max_text_len, dtype,
+                                    device) -> None:
+    """
+    Initialize a Per-GPU cross-attention cache aligned with the Wan model assumptions.
+    """
+    crossattn_cache = []
+    if isinstance(transformer, CausalWanModel):
+        num_attention_heads = transformer.num_heads
+        attention_head_dim = transformer.dim // transformer.num_heads
+    elif isinstance(transformer, CausalWanTransformer3DModel):
+        num_attention_heads = transformer.num_attention_heads
+        attention_head_dim = transformer.attention_head_dim
+
+    for _ in range(len(transformer.blocks)):
+        crossattn_cache.append({
+            "k":
+            torch.zeros([
+                batch_size, max_text_len, num_attention_heads,
+                attention_head_dim
+            ],
+                        dtype=dtype,
+                        device=device),
+            "v":
+            torch.zeros([
+                batch_size, max_text_len, num_attention_heads,
+                attention_head_dim
+            ],
+                        dtype=dtype,
+                        device=device),
+            "is_init":
+            False,
+        })
+    return crossattn_cache
@@ -0,0 +1,133 @@
+# SPDX-License-Identifier: Apache-2.0
+import os
+import math
+
+import numpy as np
+import pytest
+import torch
+from fastvideo.wan.modules.model import WanModel
+
+from fastvideo.configs.pipelines import PipelineConfig
+from fastvideo.forward_context import set_forward_context
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.models.loader.component_loader import TransformerLoader
+from fastvideo.utils import maybe_download_model
+from fastvideo.configs.models.dits import WanVideoConfig
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+
+
+logger = init_logger(__name__)
+
+os.environ["MASTER_ADDR"] = "localhost"
+os.environ["MASTER_PORT"] = "29503"
+
+BASE_MODEL_PATH = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
+                                  local_dir=os.path.join(
+                                      'data', BASE_MODEL_PATH))
+TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
+
+
+@pytest.mark.usefixtures("distributed_setup")
+def test_ori_wan_transformer():
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    precision = torch.bfloat16
+    precision_str = "bf16"
+    args = FastVideoArgs(model_path=TRANSFORMER_PATH,
+                         dit_cpu_offload=True,
+                         pipeline_config=PipelineConfig(dit_config=WanVideoConfig(), dit_precision=precision_str))
+    args.device = device
+
+    loader = TransformerLoader()
+    model2 = loader.load(TRANSFORMER_PATH, args).to(dtype=precision)
+
+    model1 = WanModel.from_pretrained(
+        "/mnt/weka/home/hao.zhang/wei/Self-Forcing/wan_models/Wan2.1-T2V-1.3B", device=device,
+        torch_dtype=precision).to(device, dtype=precision).requires_grad_(False)
+
+    total_params = sum(p.numel() for p in model1.parameters())
+    # Calculate weight sum for model1 (converting to float64 to avoid overflow)
+    weight_sum_model1 = sum(
+        p.to(torch.float64).sum().item() for p in model1.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model1 = weight_sum_model1 / total_params
+    logger.info("Model 1 weight sum: %s", weight_sum_model1)
+    logger.info("Model 1 weight mean: %s", weight_mean_model1)
+
+    # Calculate weight sum for model2 (converting to float64 to avoid overflow)
+    total_params_model2 = sum(p.numel() for p in model2.parameters())
+    weight_sum_model2 = sum(
+        p.to(torch.float64).sum().item() for p in model2.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model2 = weight_sum_model2 / total_params_model2
+    logger.info("Model 2 weight sum: %s", weight_sum_model2)
+    logger.info("Model 2 weight mean: %s", weight_mean_model2)
+
+    weight_sum_diff = abs(weight_sum_model1 - weight_sum_model2)
+    logger.info("Weight sum difference: %s", weight_sum_diff)
+    weight_mean_diff = abs(weight_mean_model1 - weight_mean_model2)
+    logger.info("Weight mean difference: %s", weight_mean_diff)
+
+    # Set both models to eval mode
+    model1 = model1.eval()
+    model2 = model2.eval()
+
+    # Create identical inputs for both models
+    batch_size = 1
+    text_seq_len = 30
+    seq_len = math.ceil((160 * 90) /
+                            (2 * 2) *
+                            21)
+
+    # Video latents [B, C, T, H, W]
+    hidden_states = torch.randn(batch_size,
+                                16,
+                                21,
+                                160,
+                                90,
+                                device=device,
+                                dtype=precision)
+
+    # Text embeddings [B, L, D] (including global token)
+    encoder_hidden_states = torch.randn(batch_size,
+                                        text_seq_len + 1,
+                                        4096,
+                                        device=device,
+                                        dtype=precision)
+
+    # Timestep
+    timestep = torch.tensor([500], device=device, dtype=precision)
+
+    forward_batch = ForwardBatch(
+        data_type="dummy",
+    )
+
+    # with torch.amp.autocast('cuda', dtype=precision):
+    output1 = model1(
+        x=hidden_states,
+        context=encoder_hidden_states,
+        t=timestep,
+        seq_len=seq_len,
+    )
+    with set_forward_context(
+            current_timestep=0,
+            attn_metadata=None,
+            forward_batch=forward_batch,
+    ):
+        output2 = model2(hidden_states=hidden_states,
+                            encoder_hidden_states=encoder_hidden_states,
+                            timestep=timestep)
+
+    # Check if outputs have the same shape
+    assert output1.shape == output2.shape, f"Output shapes don't match: {output1.shape} vs {output2.shape}"
+    assert output1.dtype == output2.dtype, f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
+
+    # Check if outputs are similar (allowing for small numerical differences)
+    max_diff = torch.max(torch.abs(output1 - output2))
+    mean_diff = torch.mean(torch.abs(output1 - output2))
+    logger.info("Max Diff: %s", max_diff.item())
+    logger.info("Mean Diff: %s", mean_diff.item())
+    assert max_diff < 1e-4, f"Maximum difference between outputs: {max_diff.item()}"
+    # mean diff
+    assert mean_diff < 1e-4, f"Mean difference between outputs: {mean_diff.item()}"
@@ -0,0 +1,144 @@
+# SPDX-License-Identifier: Apache-2.0
+import os
+import math
+
+import numpy as np
+import pytest
+import torch
+from fastvideo.wan.modules.causal_model import CausalWanModel
+
+from fastvideo.configs.pipelines import PipelineConfig
+from fastvideo.forward_context import set_forward_context
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.models.loader.component_loader import TransformerLoader
+from fastvideo.utils import maybe_download_model
+from fastvideo.configs.models.dits import WanVideoConfig
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+
+
+logger = init_logger(__name__)
+
+os.environ["MASTER_ADDR"] = "localhost"
+os.environ["MASTER_PORT"] = "29503"
+
+BASE_MODEL_PATH = "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
+                                  local_dir=os.path.join(
+                                      'data', BASE_MODEL_PATH))
+TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
+
+
+@pytest.mark.usefixtures("distributed_setup")
+def test_train_ori_causal_wan_transformer():
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    precision = torch.bfloat16
+    precision_str = "bf16"
+    args = FastVideoArgs(model_path=TRANSFORMER_PATH,
+                         dit_cpu_offload=True,
+                         pipeline_config=PipelineConfig(dit_config=WanVideoConfig(), dit_precision=precision_str))
+    args.device = device
+
+    loader = TransformerLoader()
+    model2 = loader.load(TRANSFORMER_PATH, args).to(dtype=precision)
+
+    model1 = CausalWanModel.from_pretrained(
+        "/mnt/weka/home/hao.zhang/wei/Self-Forcing/wan_models/Wan2.1-T2V-1.3B", device=device,
+        torch_dtype=precision).to(device, dtype=precision).requires_grad_(False)
+    causal_state_dict = torch.load("/mnt/weka/home/hao.zhang/wei/Self-Forcing/checkpoints/self_forcing_dmd.pt")["generator_ema"]
+    new_state_dict = {}
+    for k, v in causal_state_dict.items():
+        if k.startswith("model."):
+            new_state_dict[k.replace("model.", "")] = v
+    causal_state_dict = new_state_dict
+    model1.load_state_dict(causal_state_dict)
+
+    model1.num_frame_per_block = 3
+    model2.num_frame_per_block = 3
+
+    total_params = sum(p.numel() for p in model1.parameters())
+    # Calculate weight sum for model1 (converting to float64 to avoid overflow)
+    weight_sum_model1 = sum(
+        p.to(torch.float64).sum().item() for p in model1.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model1 = weight_sum_model1 / total_params
+    logger.info("Model 1 weight sum: %s", weight_sum_model1)
+    logger.info("Model 1 weight mean: %s", weight_mean_model1)
+
+    # Calculate weight sum for model2 (converting to float64 to avoid overflow)
+    total_params_model2 = sum(p.numel() for p in model2.parameters())
+    weight_sum_model2 = sum(
+        p.to(torch.float64).sum().item() for p in model2.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model2 = weight_sum_model2 / total_params_model2
+    logger.info("Model 2 weight sum: %s", weight_sum_model2)
+    logger.info("Model 2 weight mean: %s", weight_mean_model2)
+
+    weight_sum_diff = abs(weight_sum_model1 - weight_sum_model2)
+    logger.info("Weight sum difference: %s", weight_sum_diff)
+    weight_mean_diff = abs(weight_mean_model1 - weight_mean_model2)
+    logger.info("Weight mean difference: %s", weight_mean_diff)
+
+    # Set both models to eval mode
+    model1 = model1.eval()
+    model2 = model2.eval()
+
+    # Create identical inputs for both models
+    batch_size = 1
+    text_seq_len = 30
+    seq_len = math.ceil((160 * 90) /
+                            (2 * 2) *
+                            21)
+
+    # Video latents [B, C, T, H, W]
+    hidden_states = torch.randn(batch_size,
+                                16,
+                                21,
+                                160,
+                                90,
+                                device=device,
+                                dtype=precision)
+
+    # Text embeddings [B, L, D] (including global token)
+    encoder_hidden_states = torch.randn(batch_size,
+                                        text_seq_len + 1,
+                                        4096,
+                                        device=device,
+                                        dtype=precision)
+
+    # Timestep
+    timestep = torch.randint(0, 1000, (batch_size, 21), device=device, dtype=torch.long)
+    logger.info("timestep: %s", timestep)
+
+    forward_batch = ForwardBatch(
+        data_type="dummy",
+    )
+
+    # with torch.amp.autocast('cuda', dtype=precision):
+    output1 = model1(
+        x=hidden_states,
+        context=encoder_hidden_states,
+        t=timestep,
+        seq_len=seq_len,
+    )
+    with set_forward_context(
+            current_timestep=0,
+            attn_metadata=None,
+            forward_batch=forward_batch,
+    ):
+        output2 = model2(hidden_states=hidden_states,
+                            encoder_hidden_states=encoder_hidden_states,
+                            timestep=timestep)
+
+    # Check if outputs have the same shape
+    assert output1.shape == output2.shape, f"Output shapes don't match: {output1.shape} vs {output2.shape}"
+    assert output1.dtype == output2.dtype, f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
+
+    # Check if outputs are similar (allowing for small numerical differences)
+    max_diff = torch.max(torch.abs(output1 - output2))
+    mean_diff = torch.mean(torch.abs(output1 - output2))
+    logger.info("Max Diff: %s", max_diff.item())
+    logger.info("Mean Diff: %s", mean_diff.item())
+    assert max_diff < 1e-4, f"Maximum difference between outputs: {max_diff.item()}"
+    # mean diff
+    assert mean_diff < 1e-4, f"Mean difference between outputs: {mean_diff.item()}"
@@ -12,6 +12,7 @@ from typing import Any
 import imageio
 import numpy as np
 import torch
+import torch.distributed as dist
 import torch.nn.functional as F
 import torchvision
 from einops import rearrange
@@ -1395,4 +1396,4 @@ class DistillationPipeline(TrainingPipeline):
                                  self.training_args.max_train_steps)

        if get_sp_group():
-            cleanup_dist_env_and_memory()
+            cleanup_dist_env_and_memory()
@@ -1,406 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import sys
-from copy import deepcopy
-from typing import cast
-
-import numpy as np
-import torch
-import torch.nn.functional as F
-
-import wandb
-from fastvideo.dataset.dataloader.schema import (
-    pyarrow_schema_ode_trajectory_text_only)
-from fastvideo.distributed import get_local_torch_device
-from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
-from fastvideo.forward_context import set_forward_context
-from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_self_forcing_flow_match import (
-    SelfForcingFlowMatchScheduler)
-from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import (
-    WanCausalDMDPipeline)
-from fastvideo.pipelines.pipeline_batch_info import TrainingBatch
-from fastvideo.training.training_pipeline import TrainingPipeline
-from fastvideo.training.training_utils import (
-    clip_grad_norm_while_handling_failing_dtensor_cases)
-
-logger = init_logger(__name__)
-
-
-class ODEInitTrainingPipeline(TrainingPipeline):
-    """
-    Training pipeline for ODE-init using precomputed denoising trajectories.
-
-    Supervision: predict the next latent in the stored trajectory by
-    - feeding current latent at timestep t into the transformer to predict noise
-    - stepping the scheduler with the predicted noise
-    - minimizing MSE to the stored next latent at timestep t_next
-    """
-
-    _required_config_modules = ["scheduler", "transformer", "vae"]
-
-    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
-        # Match the preprocess/generation scheduler for consistent stepping
-        self.modules["scheduler"] = SelfForcingFlowMatchScheduler(
-            shift=fastvideo_args.pipeline_config.flow_shift,
-            sigma_min=0.0,
-            extra_one_step=True)
-        self.modules["scheduler"].set_timesteps(num_inference_steps=1000,
-                                                training=True)
-
-    def set_schemas(self):
-        self.train_dataset_schema = pyarrow_schema_ode_trajectory_text_only
-
-    def initialize_training_pipeline(self, training_args: TrainingArgs):
-        super().initialize_training_pipeline(training_args)
-
-        self.noise_scheduler = self.get_module("scheduler")
-        self.vae = self.get_module("vae")
-        self.vae.requires_grad_(False)
-
-        self.timestep_shift = self.training_args.pipeline_config.flow_shift
-        assert self.timestep_shift == 5.0, "flow_shift must be 5.0"
-        self.noise_scheduler = SelfForcingFlowMatchScheduler(
-            shift=self.timestep_shift, sigma_min=0.0, extra_one_step=True)
-        self.noise_scheduler.set_timesteps(num_inference_steps=1000,
-                                           training=True)
-
-        logger.info("dmd_denoising_steps: %s",
-                    self.training_args.pipeline_config.dmd_denoising_steps)
-        self.dmd_denoising_steps = torch.tensor([1000, 750, 500, 250],
-                                                dtype=torch.long,
-                                                device=get_local_torch_device())
-        if training_args.warp_denoising_step:  # Warp the denoising step according to the scheduler time shift
-            timesteps = torch.cat((self.noise_scheduler.timesteps.cpu(),
-                                   torch.tensor([0],
-                                                dtype=torch.float32))).cuda()
-            logger.info("timesteps: %s", timesteps)
-            self.dmd_denoising_steps = timesteps[1000 -
-                                                 self.dmd_denoising_steps]
-            logger.info("warped self.dmd_denoising_steps: %s",
-                        self.dmd_denoising_steps)
-        else:
-            raise ValueError("warp_denoising_step must be true")
-
-        self.dmd_denoising_steps = self.dmd_denoising_steps.to(
-            get_local_torch_device())
-
-        logger.info("denoising_step_list: %s", self.dmd_denoising_steps)
-
-        logger.info(
-            "Initialized ODE-init training pipeline with %s denoising steps",
-            len(self.dmd_denoising_steps))
-        # Cache for nearest trajectory index per DMD step (computed lazily on first batch)
-        self._cached_closest_idx_per_dmd = None
-        self.num_train_timestep = self.noise_scheduler.num_train_timesteps
-        self.manual_idx = 0
-
-    def initialize_validation_pipeline(self, training_args: TrainingArgs):
-        logger.info("Initializing validation pipeline...")
-        args_copy = deepcopy(training_args)
-        args_copy.inference_mode = True
-        # Warm start validation with current transformer
-        self.validation_pipeline = WanCausalDMDPipeline.from_pretrained(
-            training_args.model_path,
-            args=args_copy,  # type: ignore
-            inference_mode=True,
-            loaded_modules={
-                "transformer": self.get_module("transformer"),
-            },
-            tp_size=training_args.tp_size,
-            sp_size=training_args.sp_size,
-            num_gpus=training_args.num_gpus,
-            pin_cpu_memory=training_args.pin_cpu_memory,
-            dit_cpu_offload=True)
-
-    def _get_next_batch(
-            self,
-            training_batch) -> tuple[TrainingBatch, torch.Tensor, torch.Tensor]:
-        batch = next(self.train_loader_iter, None)  # type: ignore
-        if batch is None:
-            self.current_epoch += 1
-            logger.info("Starting epoch %s", self.current_epoch)
-            self.train_loader_iter = iter(self.train_dataloader)
-            batch = next(self.train_loader_iter)
-
-        # Required fields from parquet (ODE trajectory schema)
-        encoder_hidden_states = batch['text_embedding']
-        encoder_attention_mask = batch['text_attention_mask']
-        infos = batch['info_list']
-
-        # Trajectory tensors may include a leading singleton batch dim per row
-        trajectory_latents = batch['trajectory_latents']
-        if trajectory_latents.dim() == 7:
-            # [B, 1, S, C, T, H, W] -> [B, S, C, T, H, W]
-            trajectory_latents = trajectory_latents[:, 0]
-        elif trajectory_latents.dim() == 6:
-            # already [B, S, C, T, H, W]
-            pass
-        else:
-            raise ValueError(
-                f"Unexpected trajectory_latents dim: {trajectory_latents.dim()}"
-            )
-
-        trajectory_timesteps = batch['trajectory_timesteps']
-        if trajectory_timesteps.dim() == 3:
-            # [B, 1, S] -> [B, S]
-            trajectory_timesteps = trajectory_timesteps[:, 0]
-        elif trajectory_timesteps.dim() == 2:
-            # [B, S]
-            pass
-        else:
-            raise ValueError(
-                f"Unexpected trajectory_timesteps dim: {trajectory_timesteps.dim()}"
-            )
-        # [B, S, C, T, H, W] -> [B, S, T, C, H, W] to match self-forcing
-        trajectory_latents = trajectory_latents.permute(0, 1, 3, 2, 4, 5)
-
-        # Move to device
-        device = get_local_torch_device()
-        training_batch.encoder_hidden_states = encoder_hidden_states.to(
-            device, dtype=torch.bfloat16)
-        training_batch.encoder_attention_mask = encoder_attention_mask.to(
-            device, dtype=torch.bfloat16)
-        training_batch.infos = infos
-
-        return training_batch, trajectory_latents.to(
-            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)
-
-        ## TEMP Used for loading the sf .pt files directly
-        """
-        self.manual_idx = self.manual_idx % 155
-        path = f"/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_pt_vidprom_1000/{self.manual_idx:05d}.pt"
-        logger.info("path: %s", path)
-        self.manual_idx += 1
-        # path = "/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_single_full/00000.pt"
-        b = torch.load(path)
-        training_batch.encoder_hidden_states = b["text_embedding"][0].unsqueeze(
-            0).to(device, dtype=torch.bfloat16)
-        trajectory_latents = b["ode_latent"].to(device, dtype=torch.bfloat16)
-        logger.info("trajectory_latents: %s", trajectory_latents.shape)
-        logger.info("encoder_hidden_states: %s",
-                    training_batch.encoder_hidden_states.shape)
-        assert trajectory_latents.shape[1] <= 10, "trajectory_latents.shape[1] must be <= 10"
-        return training_batch, trajectory_latents.to(
-            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)
-        """
-
-    def _get_timestep(self,
-                      min_timestep: int,
-                      max_timestep: int,
-                      batch_size: int,
-                      num_frame: int,
-                      num_frame_per_block: int,
-                      uniform_timestep: bool = False) -> torch.Tensor:
-        if uniform_timestep:
-            timestep = torch.randint(min_timestep,
-                                     max_timestep, [batch_size, 1],
-                                     device=self.device,
-                                     dtype=torch.long).repeat(1, num_frame)
-            return timestep
-        else:
-            timestep = torch.randint(min_timestep,
-                                     max_timestep, [batch_size, num_frame],
-                                     device=self.device,
-                                     dtype=torch.long)
-            # logger.info(f"individual timestep: {timestep}")
-            # make the noise level the same within every block
-            timestep = timestep.reshape(timestep.shape[0], -1,
-                                        num_frame_per_block)
-            timestep[:, :, 1:] = timestep[:, :, 0:1]
-            timestep = timestep.reshape(timestep.shape[0], -1)
-            return timestep
-
-    def _step_predict_next_latent(
-        self, traj_latents: torch.Tensor, traj_timesteps: torch.Tensor,
-        encoder_hidden_states: torch.Tensor,
-        encoder_attention_mask: torch.Tensor
-    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, dict[str,
-                                                              torch.Tensor]]:
-        latent_vis_dict: dict[str, torch.Tensor] = {}
-        device = get_local_torch_device()
-        target_latent = traj_latents[:, -1]
-
-        # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
-        B, S, num_frames, num_channels, height, width = traj_latents.shape
-
-        # Lazily cache nearest trajectory index per DMD step based on the (fixed) S timesteps
-        if self._cached_closest_idx_per_dmd is None:
-            self._cached_closest_idx_per_dmd = torch.tensor(
-                [0, 12, 24, 36], dtype=torch.long).cpu()
-            # [0, 1, 2, 3], dtype=torch.long).cpu()
-            logger.info("self._cached_closest_idx_per_dmd: %s",
-                        self._cached_closest_idx_per_dmd)
-            logger.info(
-                "corresponding timesteps: %s", self.noise_scheduler.timesteps[
-                    self._cached_closest_idx_per_dmd])
-
-        # Select the K indexes from traj_latents using self._cached_closest_idx_per_dmd
-        # traj_latents: [B, S, C, T, H, W], self._cached_closest_idx_per_dmd: [K]
-        # Output: [B, K, C, T, H, W]
-        assert self._cached_closest_idx_per_dmd is not None
-        relevant_traj_latents = torch.index_select(
-            traj_latents,
-            dim=1,
-            index=self._cached_closest_idx_per_dmd.to(traj_latents.device))
-        logger.info("relevant_traj_latents: %s", relevant_traj_latents.shape)
-        # assert relevant_traj_latents.shape[0] == 1
-
-        indexes = self._get_timestep(  # [B, num_frames]
-            0,
-            len(self.dmd_denoising_steps),
-            B,
-            num_frames,
-            3,
-            uniform_timestep=False)
-        logger.info("indexes: %s", indexes.shape)
-        logger.info("indexes: %s", indexes)
-        # noisy_input = relevant_traj_latents[indexes]
-        noisy_input = torch.gather(
-            relevant_traj_latents,
-            dim=1,
-            index=indexes.reshape(B, 1, num_frames, 1, 1,
-                                  1).expand(-1, -1, -1, num_channels, height,
-                                            width).to(self.device)).squeeze(1)
-        timestep = self.dmd_denoising_steps[indexes]
-        logger.info("selected timestep for rank %s: %s",
-                    self.global_rank,
-                    timestep,
-                    local_main_process_only=False)
-
-        # Prepare inputs for transformer
-        latent_vis_dict["noisy_input"] = noisy_input.permute(
-            0, 2, 1, 3, 4).detach().clone().cpu()
-        latent_vis_dict["x0"] = target_latent.permute(0, 2, 1, 3,
-                                                      4).detach().clone().cpu()
-
-        input_kwargs = {
-            "hidden_states": noisy_input.permute(0, 2, 1, 3, 4),
-            "encoder_hidden_states": encoder_hidden_states,
-            "timestep": timestep.to(device, dtype=torch.bfloat16),
-            "return_dict": False,
-        }
-        # Predict noise and step the scheduler to obtain next latent
-        with set_forward_context(current_timestep=timestep,
-                                 attn_metadata=None,
-                                 forward_batch=None):
-            noise_pred = self.transformer(**input_kwargs).permute(0, 2, 1, 3, 4)
-
-        from fastvideo.models.utils import pred_noise_to_pred_video
-        pred_video = pred_noise_to_pred_video(
-            pred_noise=noise_pred.flatten(0, 1),
-            noise_input_latent=noisy_input.flatten(0, 1),
-            timestep=timestep.to(dtype=torch.bfloat16).flatten(0, 1),
-            scheduler=self.modules["scheduler"]).unflatten(
-                0, noise_pred.shape[:2])
-        latent_vis_dict["pred_video"] = pred_video.permute(
-            0, 2, 1, 3, 4).detach().clone().cpu()
-
-        return pred_video, target_latent, timestep, latent_vis_dict
-
-    def train_one_step(self, training_batch):  # type: ignore[override]
-        self.transformer.train()
-        self.optimizer.zero_grad()
-        training_batch.total_loss = 0.0
-        args = cast(TrainingArgs, self.training_args)
-
-        # Using cached nearest index per DMD step; computation happens in _step_predict_next_latent
-
-        for _ in range(args.gradient_accumulation_steps):
-            training_batch, traj_latents, traj_timesteps = self._get_next_batch(
-                training_batch)
-            text_embeds = training_batch.encoder_hidden_states
-            text_attention_mask = training_batch.encoder_attention_mask
-            assert traj_latents.shape[0] == 1
-
-            # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
-            _, S = traj_latents.shape[0], traj_latents.shape[1]
-            if S < 2:
-                raise ValueError("Trajectory must contain at least 2 steps")
-
-            # Forward to predict next latent by stepping scheduler with predicted noise
-            noise_pred, target_latent, t, latent_vis_dict = self._step_predict_next_latent(
-                traj_latents, traj_timesteps, text_embeds, text_attention_mask)
-
-            training_batch.latent_vis_dict.update(latent_vis_dict)
-
-            mask = t != 0
-
-            # Compute loss
-            loss = F.mse_loss(noise_pred[mask],
-                              target_latent[mask],
-                              reduction="mean")
-            loss = loss / args.gradient_accumulation_steps
-
-            with set_forward_context(current_timestep=t,
-                                     attn_metadata=None,
-                                     forward_batch=None):
-                loss.backward()
-            avg_loss = loss.detach().clone()
-            training_batch.total_loss += avg_loss.item()
-
-        # Clip grad and step optimizers
-        grad_norm = clip_grad_norm_while_handling_failing_dtensor_cases(
-            [p for p in self.transformer.parameters() if p.requires_grad],
-            args.max_grad_norm if args.max_grad_norm is not None else 0.0)
-
-        self.optimizer.step()
-        self.lr_scheduler.step()
-
-        if grad_norm is None:
-            grad_value = 0.0
-        else:
-            try:
-                if isinstance(grad_norm, torch.Tensor):
-                    grad_value = float(grad_norm.detach().float().item())
-                else:
-                    grad_value = float(grad_norm)
-            except Exception:
-                grad_value = 0.0
-        training_batch.grad_norm = grad_value
-        return training_batch
-
-    def visualize_intermediate_latents(self, training_batch: TrainingBatch,
-                                       training_args: TrainingArgs, step: int):
-        """Add visualization data to wandb logging and save frames to disk."""
-        wandb_loss_dict = {}
-        latents_vis_dict = training_batch.latent_vis_dict
-        latent_log_keys = ['noisy_input', 'x0', 'pred_video']
-        for latent_key in latent_log_keys:
-            assert latent_key in latents_vis_dict and latents_vis_dict[
-                latent_key] is not None
-            latent = latents_vis_dict[latent_key]
-            pixel_latent = self.validation_pipeline.decoding_stage.decode(
-                latent, training_args)
-
-            video = pixel_latent.cpu().float()
-            video = video.permute(0, 2, 1, 3, 4)
-            video = (video * 255).numpy().astype(np.uint8)
-            wandb_loss_dict[latent_key] = wandb.Video(
-                video, fps=16, format="mp4")  # change to 16 for Wan2.1
-            # Clean up references
-            del video, pixel_latent, latent
-
-        # Log to wandb
-        if self.global_rank == 0:
-            wandb.log(wandb_loss_dict, step=step)
-
-
-def main(args) -> None:
-    logger.info("Starting ODE-init training pipeline...")
-    pipeline = ODEInitTrainingPipeline.from_pretrained(
-        args.pretrained_model_name_or_path, args=args)
-    args = pipeline.training_args
-    pipeline.train()
-    logger.info("ODE-init training pipeline done")
-
-
-if __name__ == "__main__":
-    argv = sys.argv
-    from fastvideo.fastvideo_args import TrainingArgs
-    from fastvideo.utils import FlexibleArgumentParser
-    parser = FlexibleArgumentParser()
-    parser = TrainingArgs.add_cli_args(parser)
-    parser = FastVideoArgs.add_cli_args(parser)
-    args = parser.parse_args()
-    args.dit_cpu_offload = False
-    main(args)
@@ -22,7 +22,7 @@ from tqdm.auto import tqdm
 import fastvideo.envs as envs
 from fastvideo.attention.backends.video_sparse_attn import (
    VideoSparseAttentionMetadataBuilder)
-from fastvideo.attention.backends.vmoba import VideoMobaAttentionMetadataBuilder
+# from fastvideo.attention.backends.vmoba import VideoMobaAttentionMetadataBuilder
 from fastvideo.configs.sample import SamplingParam
 from fastvideo.dataset import build_parquet_map_style_dataloader
 from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v
@@ -39,20 +39,26 @@ from fastvideo.training.activation_checkpoint import (
    apply_activation_checkpointing)
 from fastvideo.training.training_utils import (
    clip_grad_norm_while_handling_failing_dtensor_cases,
-    compute_density_for_timestep_sampling, count_trainable, get_scheduler,
-    get_sigmas, load_checkpoint, normalize_dit_input, save_checkpoint,
+    compute_density_for_timestep_sampling, get_scheduler, get_sigmas,
+    load_checkpoint, normalize_dit_input, save_checkpoint,
    shard_latents_across_sp)
-from fastvideo.utils import (is_vmoba_available, is_vsa_available,
+# from fastvideo.utils import (is_vmoba_available, is_vsa_available,
+#                              set_random_seed, shallow_asdict)
+from fastvideo.utils import (is_vsa_available,
                             set_random_seed, shallow_asdict)

 import wandb  # isort: skip

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

 logger = init_logger(__name__)


+def _get_trainable_params(model: torch.nn.Module) -> int:
+    return sum(p.numel() for p in model.parameters() if p.requires_grad)
+
+
 class TrainingPipeline(LoRAPipeline, ABC):
    """
    A pipeline for training a model. All training pipelines should inherit from this class.
@@ -112,15 +118,18 @@ class TrainingPipeline(LoRAPipeline, ABC):
                enable_gradient_checkpointing_type)

        noise_scheduler = self.modules["scheduler"]
-        # Set grads for proper modules based on the training mode (Distill, LoRA, etc.)
        self.set_trainable()
        params_to_optimize = self.transformer.parameters()
        params_to_optimize = list(
            filter(lambda p: p.requires_grad, params_to_optimize))
+        # Parse betas from string format "beta1,beta2"
+        betas_str = training_args.betas
+        betas = tuple(float(x.strip()) for x in betas_str.split(","))
+        
        self.optimizer = torch.optim.AdamW(
            params_to_optimize,
            lr=training_args.learning_rate,
-            betas=(0.9, 0.999),
+            betas=betas,
            weight_decay=training_args.weight_decay,
            eps=1e-8,
        )
@@ -272,20 +281,20 @@ class TrainingPipeline(LoRAPipeline, ABC):
                patch_size=patch_size,
                VSA_sparsity=current_vsa_sparsity,
                device=get_local_torch_device())
-        elif vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
-            moba_params = self.training_args.moba_config.copy()
-            moba_params.update({
-                "current_timestep":
-                training_batch.timesteps,
-                "raw_latent_shape":
-                training_batch.raw_latent_shape[2:5],
-                "patch_size":
-                self.training_args.pipeline_config.dit_config.patch_size,
-                "device":
-                get_local_torch_device(),
-            })
-            training_batch.attn_metadata = VideoMobaAttentionMetadataBuilder(
-            ).build(**moba_params)
+        # elif vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
+        #     moba_params = self.training_args.moba_config.copy()
+        #     moba_params.update({
+        #         "current_timestep":
+        #         training_batch.timesteps,
+        #         "raw_latent_shape":
+        #         training_batch.raw_latent_shape[2:5],
+        #         "patch_size":
+        #         self.training_args.pipeline_config.dit_config.patch_size,
+        #         "device":
+        #         get_local_torch_device(),
+        #     })
+        #     training_batch.attn_metadata = VideoMobaAttentionMetadataBuilder(
+        #     ).build(**moba_params)
        else:
            training_batch.attn_metadata = None

@@ -310,7 +319,8 @@ class TrainingPipeline(LoRAPipeline, ABC):

    def _transformer_forward_and_compute_loss(
            self, training_batch: TrainingBatch) -> TrainingBatch:
-        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN" or vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
+        # if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN" or vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
+        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN":
            assert training_batch.attn_metadata is not None
        else:
            assert training_batch.attn_metadata is None
@@ -431,7 +441,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
                    local_main_process_only=False)
        if not self.post_init_called:
            self.post_init()
-        num_trainable_params = count_trainable(self.transformer)
+        num_trainable_params = _get_trainable_params(self.transformer)
        logger.info("Starting training with %s B trainable parameters",
                    round(num_trainable_params / 1e9, 3))

@@ -476,9 +486,9 @@ class TrainingPipeline(LoRAPipeline, ABC):
                current_decay_times = min(step // vsa_decay_interval_steps,
                                          vsa_sparsity // vsa_decay_rate)
                current_vsa_sparsity = current_decay_times * vsa_decay_rate
-            elif vmoba_available:
-                # TODO: add vmoba sparsity scheduling here
-                current_vsa_sparsity = 0.0
+            # elif vmoba_available:
+            #     # TODO: add vmoba sparsity scheduling here
+            #     pass
            else:
                current_vsa_sparsity = 0.0

@@ -520,14 +530,10 @@ class TrainingPipeline(LoRAPipeline, ABC):
                self.transformer.train()
                self.sp_group.barrier()
            if self.training_args.log_validation and step % self.training_args.validation_steps == 0:
-                if self.training_args.log_visualization:
-                    self.visualize_intermediate_latents(training_batch,
-                                                        self.training_args,
-                                                        step)
                self._log_validation(self.transformer, self.training_args, step)
                gpu_memory_usage = torch.cuda.memory_allocated() / 1024**2
                trainable_params = round(
-                    count_trainable(self.transformer) / 1e9, 3)
+                    _get_trainable_params(self.transformer) / 1e9, 3)
                logger.info(
                    "GPU memory usage after validation: %s MB, trainable params: %sB",
                    gpu_memory_usage, trainable_params)
@@ -563,7 +569,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
        logger.info("  Total optimization steps = %s",
                    self.training_args.max_train_steps)
        logger.info("  Total training parameters per FSDP shard = %s B",
-                    round(count_trainable(self.transformer) / 1e9, 3))
+                    round(_get_trainable_params(self.transformer) / 1e9, 3))
        # print dtype
        logger.info("  Master weight dtype: %s",
                    self.transformer.parameters().__next__().dtype)
@@ -631,6 +637,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
        validation_dataloader = DataLoader(validation_dataset,
                                           batch_size=None,
                                           num_workers=0)
+
        transformer.eval()

        validation_steps = training_args.validation_sampling_steps.split(",")
@@ -723,11 +730,4 @@ class TrainingPipeline(LoRAPipeline, ABC):

        # Re-enable gradients for training
        training_args.inference_mode = False
-        transformer.train()
-
-    def visualize_intermediate_latents(self, training_batch: TrainingBatch,
-                                       training_args: TrainingArgs, step: int):
-        """Add visualization data to wandb logging and save frames to disk."""
-        raise NotImplementedError(
-            "Visualize intermediate latents is not implemented for training pipeline"
-        )
+        transformer.train()
@@ -1323,7 +1323,6 @@ class EMA_FSDP:
      ema.update(model)
      ema.state_dict()  # on rank 0
    """
-
    def __init__(self, module, decay: float = 0.999, mode: str = "local_shard"):
        self.decay = float(decay)
        self.mode = mode
@@ -1417,7 +1416,6 @@ class EMA_FSDP:
                p.data.copy_(w.to(dtype=p.dtype, device=p.device))

    class _ApplyEMACtx:
-
        def __init__(self, ema: "EMA_FSDP", module):
            self.ema = ema
            self.module = module