Support custom action trajectories for validation

Overfitting running for MC 10
Update action_labels
2026-02-03 20:23:13 +00:00 · 2026-02-03 05:11:48 +00:00 · 2026-02-02 05:20:06 +00:00 · 2026-02-02 04:38:24 +00:00
146 changed files with 4651 additions and 6523 deletions
@@ -1,42 +0,0 @@
-# Repository Guidelines
-
-## Project Structure & Module Organization
- Core Python package: `fastvideo/` (models, pipelines, training, distributed runtime, CLI entrypoints).
- CUDA/custom kernels: `fastvideo-kernel/` (separate build/test flow).
- Tests:
-  - `fastvideo/tests/` for package-level tests (dataset, encoders, inference, training, SSIM, workflow).
-  - `tests/local_tests/` for additional local/component checks.
- Docs and guides: `docs/` (MkDocs source), with contributor docs in `docs/contributing/`.
- Runnable examples and scripts: `examples/` and `scripts/`.
- Static assets: `assets/`, `images/`, `videos/`, and `comfyui/assets/`.
-
-## Build, Test, and Development Commands
- `uv pip install -e .[dev]`: editable install with lint/test extras.
- `pre-commit install --hook-type pre-commit --hook-type commit-msg`: enable local hooks.
- `pre-commit run --all-files`: run formatter/lint/type/spelling checks.
- `pytest tests/`: run top-level test suite.
- `pytest fastvideo/tests/ -v`: run package tests.
- `pytest fastvideo/tests/ssim/ -vs`: run SSIM regression tests (GPU-heavy).
- `cd fastvideo-kernel && ./build.sh`: build kernel extensions.
-
-## Coding Style & Naming Conventions
- Python 3.10+; 4-space indentation; keep code and imports readable and explicit.
- Style tools are configured in `pyproject.toml` and `.pre-commit-config.yaml`:
-  - `yapf` (format), `ruff` (lint, auto-fix), `mypy` (typing), `codespell`.
- Target line length is 80.
- Naming: `snake_case` for functions/files, `PascalCase` for classes, `UPPER_SNAKE_CASE` for constants.
-
-## Testing Guidelines
- Use `pytest` and place tests near relevant domains (e.g., `fastvideo/tests/encoders/`).
- Prefer descriptive names like `test_<feature>_<expected_behavior>.py`.
- For new pipelines/backends, include at least one regression-oriented test; add SSIM coverage when output quality must be preserved.
- Document GPU assumptions in tests that require specific hardware.
-
-## Commit & Pull Request Guidelines
- Follow existing commit style: short subject with optional tag prefix, e.g. `[bugfix]: ...`, `[feat]: ...`, `[misc]: ...`, and include PR reference like `(#1234)` when applicable.
- Keep commits focused by concern (feature, refactor, fix).
- PRs should include:
-  - clear problem/solution summary,
-  - test evidence (`pytest`/SSIM outputs or rationale if skipped),
-  - linked issue/PR context,
-  - screenshots or sample outputs for UI/demo/docs changes.
@@ -1 +0,0 @@
-@AGENTS.md
@@ -9,26 +9,27 @@
 **FastVideo is a unified post-training and inference framework for accelerated video generation.**

 ## NEWS
+- ```2025/11/19```: Release [CausalWan2.2 I2V A14B Preview](https://huggingface.co/FastVideo/CausalWan2.2-I2V-A14B-Preview-Diffusers) models, [Blog](https://hao-ai-lab.github.io/blogs/fastvideo_causalwan_preview/) and [Inference Code!](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py)
+- ```2025/08/04```: Release [FastWan](https://hao-ai-lab.github.io/FastVideo/distillation/dmd) models and [Sparse-Distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).

- `2025/11/19`: Release [CausalWan2.2 I2V A14B Preview](https://huggingface.co/FastVideo/CausalWan2.2-I2V-A14B-Preview-Diffusers) models, [Blog](https://hao-ai-lab.github.io/blogs/fastvideo_causalwan_preview/) and [Inference Code!](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py)
- `2025/08/04`: Release [FastWan](https://hao-ai-lab.github.io/FastVideo/distillation/dmd) models and [Sparse-Distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).
+<details>
+<summary>More</summary>

-### More News
+- ```2025/06/14```: Release finetuning and inference code for [VSA](https://arxiv.org/pdf/2505.13389)
+- ```2025/04/24```: [FastVideo V1](https://hao-ai-lab.github.io/blogs/fastvideo/) is released!
+- ```2025/02/18```: Release the inference code for [Sliding Tile Attention](https://hao-ai-lab.github.io/blogs/sta/).

- `2025/06/14`: Release finetuning and inference code for [VSA](https://arxiv.org/pdf/2505.13389)
- `2025/04/24`: [FastVideo V1](https://hao-ai-lab.github.io/blogs/fastvideo/) is released!
- `2025/02/18`: Release the inference code for [Sliding Tile Attention](https://hao-ai-lab.github.io/blogs/sta/).
+</details>

 ## Key Features

 FastVideo has the following features:
-
 - End-to-end post-training support for bidirectional and autoregressive models:
  - Support full finetuning and LoRA finetuning for state-of-the-art open video DiTs
  - Data preprocessing pipeline for video, image, and text data
  - Distribution Matching Distillation (DMD2) stepwise distillation.
  - Sparse attention with [Video Sparse Attention](https://arxiv.org/pdf/2505.13389)
-  - [Sparse distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/) to achieve >50x denoising speedup
+  - [Sparse distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/) to achineve >50x denoising speedup
  - Scalable training with FSDP2, sequence parallelism, and selective activation checkpointing.
  - Causal distillation through Self-Forcing
  - See this [page](https://hao-ai-lab.github.io/FastVideo/training/overview/) for full list of supported models and recipes.
@@ -43,7 +44,6 @@ FastVideo has the following features:
  - See this [page](https://hao-ai-lab.github.io/FastVideo/inference/hardware_support/) for full list of supported hardware and OS.

 ## Getting Started
-
 We recommend using an environment manager such as `Conda` to create a clean environment:

 ```bash
@@ -58,20 +58,18 @@ pip install fastvideo
 Please see our [docs](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/) for more detailed installation instructions.

 ## Sparse Distillation
-
 For our sparse distillation techniques, please see our [distillation docs](https://hao-ai-lab.github.io/FastVideo/distillation/dmd/) and check out our [blog](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/).

 See below for recipes and datasets:

-| Model                                                                                 | Sparse Distillation                                                                                             | Dataset                                                                                                  |
-| ------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------- |
-| [FastWan2.1-T2V-1.3B](https://huggingface.co/FastVideo/FastWan2.1-T2V-1.3B-Diffusers) | [Recipe](https://github.com/hao-ai-lab/FastVideo/tree/main/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P)       | [FastVideo Synthetic Wan2.1 480P](https://huggingface.co/datasets/FastVideo/Wan-Syn_77x448x832_600k)     |
-| [FastWan2.2-TI2V-5B](https://huggingface.co/FastVideo/FastWan2.2-TI2V-5B-Diffusers)   | [Recipe](https://github.com/hao-ai-lab/FastVideo/tree/main/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free) | [FastVideo Synthetic Wan2.2 720P](https://huggingface.co/datasets/FastVideo/Wan2.2-Syn-121x704x1280_32k) |
+|                                            Model                                              |                                               Sparse Distillation                                                 |                                                  Dataset                                                  |
+|:-------------------------------------------------------------------------------------------:  |:---------------------------------------------------------------------------------------------------------------:  |:--------------------------------------------------------------------------------------------------------: |
+| [FastWan2.1-T2V-1.3B](https://huggingface.co/FastVideo/FastWan2.1-T2V-1.3B-Diffusers)         |    [Recipe](https://github.com/hao-ai-lab/FastVideo/tree/main/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P)      | [FastVideo Synthetic Wan2.1 480P](https://huggingface.co/datasets/FastVideo/Wan-Syn_77x448x832_600k)      |
+| [FastWan2.1-T2V-14B-Preview](https://huggingface.co/FastVideo/FastWan2.1-T2V-14B-Diffusers)   |                                                   Coming soon!                                                    |   [FastVideo Synthetic Wan2.1 720P](https://huggingface.co/datasets/FastVideo/Wan-Syn_77x768x1280_250k)   |
+| [FastWan2.2-TI2V-5B](https://huggingface.co/FastVideo/FastWan2.2-TI2V-5B-Diffusers)           | [Recipe](https://github.com/hao-ai-lab/FastVideo/tree/main/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free)   | [FastVideo Synthetic Wan2.2 720P](https://huggingface.co/datasets/FastVideo/Wan2.2-Syn-121x704x1280_32k)  |

 ## Inference
-
 ### Generating Your First Video
-
 Here's a minimal example to generate a video using the default settings. Make sure VSA kernels are [installed](https://hao-ai-lab.github.io/FastVideo/video_sparse_attention/installation/). Create a file called `example.py` with the following code:

 ```python
@@ -110,36 +108,55 @@ python example.py

 For a more detailed guide, please see our [inference quick start](https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start/).

-## More Guides
+### Other docs:

 - [Design Overview](https://hao-ai-lab.github.io/FastVideo/design/overview/)
+- [Contribution Guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/)
+
+## Distillation and Finetuning
 - [Distillation Guide](https://hao-ai-lab.github.io/FastVideo/distillation/dmd/)
- [Contribution Guide](https://hao-ai-lab.github.io/FastVideo/contributing/overview/)
+<!-- - [Finetuning Guide](https://hao-ai-lab.github.io/FastVideo/training/finetune.html) -->

 ## Awesome work using FastVideo or our research projects

- [SGLang](https://github.com/sgl-project/sglang/tree/main/python/sglang/multimodal_gen): SGLang's diffusion inference functionality is based on a fork of FastVideo on Sept. 24, 2025.
- [DanceGRPO](https://github.com/XueZeyue/DanceGRPO): A unified framework to adapt Group Relative Policy Optimization (GRPO) to visual generation paradigms. Code based on FastVideo.
- [SRPO](https://github.com/Tencent-Hunyuan/SRPO): A method to directly align the full diffusion trajectory with fine-grained human preference. Code based on FastVideo.
- [DCM](https://github.com/Vchitect/DCM): Dual-expert consistency model for efficient and high-quality video generation. Code based on FastVideo.
- [Hunyuan Video 1.5](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5): A leading lightweight video generation model, where they proposed SSTA based on Sliding Tile Attention.
- [Kandinsky-5.0](https://github.com/kandinskylab/kandinsky-5): A family of diffusion models for video & image generation, where their NABLA attention includes a Sliding Tile Attention branch.
- [LongCat Video](https://github.com/meituan-longcat/LongCat-Video): A foundational video generation model with 13.6B parameters with block-sparse attention similar to Video Sparse Attention.
+- [SGLang](https://github.com/sgl-project/sglang/tree/main/python/sglang/multimodal_gen): SGLang's diffusion inference functionality is based  on a fork of FastVideo on Sept. 24, 2025. [![Star](https://img.shields.io/github/stars/sgl-project/sglang.svg?style=social&label=Star)](https://github.com/sgl-project/sglang)
+
+- [DanceGRPO](https://github.com/XueZeyue/DanceGRPO): A  unified framework to adapt Group Relative Policy Optimization (GRPO) to visual generation paradigms. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/XueZeyue/DanceGRPO.svg?style=social&label=Star)](https://github.com/XueZeyue/DanceGRPO)
+- [SRPO](https://github.com/Tencent-Hunyuan/SRPO): A method to directly align the full diffusion trajectory with fine-grained human preference. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/Tencent-Hunyuan/SRPO.svg?style=social&label=Star)](https://github.com/Tencent-Hunyuan/SRPO)
+- [DCM](https://github.com/Vchitect/DCM): Dual-expert consistency model for efficient and high-quality video generation. Code based on FastVideo. [![Star](https://img.shields.io/github/stars/Vchitect/DCM.svg?style=social&label=Star)](https://github.com/Vchitect/DCM)
+- [Hunyuan Video 1.5](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5): A leading lightweight video generation model, where they proposed SSTA based on Sliding Tile Attention. [![Star](https://img.shields.io/github/stars/Tencent-Hunyuan/HunyuanVideo-1.5.svg?style=social&label=Star)](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5)
+- [Kandinsky-5.0](https://github.com/kandinskylab/kandinsky-5): A family of diffusion models for video & image generation, where their NABLA attention includes a Sliding Tile Attention branch. [![Star](https://img.shields.io/github/stars/kandinskylab/kandinsky-5.svg?style=social&label=Star)](https://github.com/kandinskylab/kandinsky-5)
+- [LongCat Video](https://github.com/meituan-longcat/LongCat-Video): A foundational video generation model with 13.6B parameters with block-sparse attention similar to Video Sparse Attention. [![Star](https://img.shields.io/github/stars/meituan-longcat/LongCat-Video.svg?style=social&label=Star)](https://github.com/meituan-longcat/LongCat-Video)

 ## 🤝 Contributing

 We welcome all contributions. Please check out our guide [here](https://hao-ai-lab.github.io/FastVideo/contributing/overview/).
 See details in [development roadmap](https://github.com/hao-ai-lab/FastVideo/issues/899).
-
 ## Acknowledgement
+We learned and reused code from the following projects:
+- [Wan-Video](https://github.com/Wan-Video)
+- [ThunderKittens](https://github.com/HazyResearch/ThunderKittens)
+- [Triton](https://github.com/triton-lang/triton)
+- [DMD2](https://github.com/tianweiy/DMD2)
+- [diffusers](https://github.com/huggingface/diffusers)
+- [xDiT](https://github.com/xdit-project/xDiT)
+- [vLLM](https://github.com/vllm-project/vllm)
+- [SGLang](https://github.com/sgl-project/sglang)

-We learned the design and reused code from the following projects: [Wan-Video](https://github.com/Wan-Video), [ThunderKittens](https://github.com/HazyResearch/ThunderKittens), [DMD2](https://github.com/tianweiy/DMD2), [diffusers](https://github.com/huggingface/diffusers), [xDiT](https://github.com/xdit-project/xDiT), [vLLM](https://github.com/vllm-project/vllm), [SGLang](https://github.com/sgl-project/sglang). We thank [MBZUAI](https://ifm.mbzuai.ac.ae/), [Anyscale](https://www.anyscale.com/), and [GMI Cloud](https://www.gmicloud.ai/) for their support throughout this project.
+We thank [MBZUAI](https://ifm.mbzuai.ac.ae/), [Anyscale](https://www.anyscale.com/), and [GMI Cloud](https://www.gmicloud.ai/) for their support throughout this project.

 ## Citation
-
-If you find FastVideo useful, please consider citing our research work:
+If you find FastVideo useful, please considering citing our work:

 ```bibtex
+@software{fastvideo2024,
+  title        = {FastVideo: A Unified Framework for Accelerated Video Generation},
+  author       = {The FastVideo Team},
+  url          = {https://github.com/hao-ai-lab/FastVideo},
+  month        = apr,
+  year         = {2024},
+}
+
@article{zhang2025vsa,
  title={Vsa: Faster video diffusion with trainable sparse attention},
  author={Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
    source /opt/venv/bin/activate && \
    uv pip install --no-cache-dir --upgrade pip && \
    uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.16/flash_attn-2.8.3+cu128torch2.10-cp310-cp310-linux_x86_64.whl
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp310-cp310-linux_x86_64.whl

 COPY . .

@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
    source /opt/venv/bin/activate && \
    uv pip install --no-cache-dir --upgrade pip && \
    uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.16/flash_attn-2.8.3+cu128torch2.10-cp311-cp311-linux_x86_64.whl
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp311-cp311-linux_x86_64.whl

 COPY . .

@@ -43,7 +43,7 @@ RUN source $HOME/.local/bin/env && \
    source /opt/venv/bin/activate && \
    uv pip install --no-cache-dir --upgrade pip && \
    uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.16/flash_attn-2.8.3+cu128torch2.10-cp312-cp312-linux_x86_64.whl
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp312-cp312-linux_x86_64.whl

 COPY . .

@@ -121,15 +121,6 @@ This page contains the complete API reference for the FastVideo library.
      show_root_toc_entry: true
      heading_level: 4

-## fastvideo.registry
-
-::: fastvideo.registry
-    options:
-      show_source: true
-      show_root_heading: true
-      show_root_toc_entry: true
-      heading_level: 3
-
 ## fastvideo.pipelines

 ::: fastvideo.pipelines
@@ -324,8 +324,8 @@ Purpose:
 Action:

 - Add a new pipeline config + sampling params.
- Register them in `fastvideo/registry.py` using explicit
-  `register_configs(...)` blocks (this file is the single source of truth now).
+- Register them in `fastvideo/configs/pipelines/registry.py` and
+  `fastvideo/configs/sample/registry.py`.

 ### 5) Wire pipeline stages

@@ -17,7 +17,7 @@ You can easily use the FastVideo Pod Template on [RunPod](https://www.runpod.io)
 - Select the "FastVideo" or "fastvideo-dev" Pod Template.
 ![RunPod Pod Template Selection](../../assets/images/runpod_create.png)

- Set the Pod name to "`<name>-<FastVideo>-<date>`".
+- Set the Pod name to "<name>-<FastVideo>-<date>".

 - Finally, once the pod is deployed (will take a few minutes as the image is being pulled), you can SSH into it using "SSH exposed over TCP". You'll need to use the matching private ssh key you provided.
 ![RunPod SSH](../../assets/images/runpod_ssh.png)
@@ -51,9 +51,8 @@ runtime parameters consistent:
 - `fastvideo/configs/pipelines/`: pipeline wiring and required components.
 - `fastvideo/configs/sample/`: default sampling parameters (steps, frames,
  guidance scale, resolution, fps).
- `fastvideo/registry.py`: unified registry for pipeline config + sampling
-  defaults and model metadata resolution, defined via explicit
-  `register_configs(...)` blocks (no separate dict registries).
+- `fastvideo/configs/registry.py`: pipeline registry.
+- `fastvideo/configs/sample/registry.py`: sampling param registry.

 `FastVideoArgs` (in `fastvideo/fastvideo_args.py`) provides runtime settings and
 is passed into pipeline construction and stages.
@@ -131,41 +131,17 @@ self.attn = DistributedAttention(

 ### Registering Models

-Register implemented modules for auto‑discovery by adding `EntryClass` in each
-model module (the registry scans for it):
+Register implemented modules in the model registry:

 ```python
-# In fastvideo/models/dits/your_module.py
-class YourTransformerModel(...):
-    ...
+# In fastvideo/models/registry.py
+_TEXT_TO_VIDEO_DIT_MODELS = {
+    "YourTransformerModel": ("dits", "yourmodule", "YourTransformerClass"),
+}

-# Entry point for model registry
-EntryClass = YourTransformerModel
-```
-
-```python
-# In fastvideo/models/vaes/your_vae.py
-class YourVAEModel(...):
-    ...
-
-# Entry point for model registry
-EntryClass = YourVAEModel
-```
-
-Register pipeline config + sampling defaults in the unified registry:
-
-```python
-# In fastvideo/registry.py
-register_configs(
-    sampling_param_cls=YourSamplingParam,
-    pipeline_config_cls=YourPipelineConfig,
-    hf_model_paths=[
-        "org/your-model-id",
-    ],
-    model_detectors=[
-        lambda path: "your-model" in path.lower(),
-    ],
-)
+_VAE_MODELS = {
+    "YourVAEModel": ("vaes", "yourvae", "YourVAEClass"),
+}
 ```

 ## Step 2: Directory Structure
@@ -1,51 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
-
-
-def main():
-    # Point this to your local diffusers model dir (or replace with a HF model ID).
-    model_path = "KyleShao/Cosmos-Predict2.5-2B-Diffusers"
-
-    generator = VideoGenerator.from_pretrained(
-        model_path,
-        num_gpus=1,
-        use_fsdp_inference=False,  # set True if GPU is out of memory
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=True,
-        pin_cpu_memory=True,
-    )
-
-    sampling_param = SamplingParam.from_pretrained(model_path)
-
-    # image2world example from official repo
-    image_path = "images/bus_terminal.jpg"
-
-    prompt = (
-        "A nighttime city bus terminal gradually shifts from stillness to subtle movement. "
-        "At first, multiple double-decker buses are parked under the glow of overhead lights, "
-        "with a central bus labeled '87D' facing forward and stationary. "
-        "As the video progresses, the bus in the middle moves ahead slowly, its headlights brightening the surrounding area "
-        "and casting reflections onto adjacent vehicles. "
-        "The motion creates space in the lineup, signaling activity within the otherwise quiet station. "
-        "It then comes to a smooth stop, resuming its position in line. "
-        "Overhead signage in Chinese characters remains illuminated, enhancing the vibrant, urban night scene."
-    )
-
-    generator.generate_video(
-        prompt,
-        sampling_param=sampling_param,
-        image_path=str(image_path),
-        num_cond_frames=1,
-        output_path="outputs_video/cosmos2_5_i2w.mp4",
-        save_video=True,
-    )
-
-    generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()
-
-
@@ -1,5 +1,4 @@
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam


 def main():
@@ -16,27 +15,19 @@ def main():
        pin_cpu_memory=True,
    )

-    # Load default sampling parameters (negative_prompt, resolution, steps, etc.)
-    sampling_param = SamplingParam.from_pretrained(model_path)
-
    prompt = (
-        "A high-definition video captures the precision of robotic welding in an industrial setting. "
-        "The first frame showcases a robotic arm, equipped with a welding torch, positioned over a large metal structure. "
-        "The welding process is in full swing, with bright sparks and intense light illuminating the scene, "
-        "creating a vivid display of blue and white hues. "
-        "A significant amount of smoke billows around the welding area, partially obscuring the view but emphasizing the heat and activity. "
-        "The background reveals parts of the workshop environment, including a ventilation system and various pieces of machinery, "
-        "indicating a busy and functional industrial workspace. "
-        "As the video progresses, the robotic arm maintains its steady position, continuing the welding process and moving to its left. "
-        "The welding torch consistently emits sparks and light, and the smoke continues to rise, diffusing slightly as it moves upward. "
-        "The metal surface beneath the torch shows ongoing signs of heating and melting. "
-        "The scene retains its industrial ambiance, with the welding sparks and smoke dominating the visual field, "
-        "underscoring the ongoing nature of the welding operation."
+       "A high-definition video captures the precision of robotic welding in an industrial setting. The first frame showcases a robotic arm, equipped with a welding torch, positioned over a large metal structure. The welding process is in full swing, with bright sparks and intense light illuminating the scene, creating a vivid display of blue and white hues. A significant amount of smoke billows around the welding area, partially obscuring the view but emphasizing the heat and activity. The background reveals parts of the workshop environment, including a ventilation system and various pieces of machinery, indicating a busy and functional industrial workspace. As the video progresses, the robotic arm maintains its steady position, continuing the welding process and moving to its left. The welding torch consistently emits sparks and light, and the smoke continues to rise, diffusing slightly as it moves upward. The metal surface beneath the torch shows ongoing signs of heating and melting. The scene retains its industrial ambiance, with the welding sparks and smoke dominating the visual field, underscoring the ongoing nature of the welding operation."
    )

-    generator.generate_video(
+    video = generator.generate_video(
        prompt,
-        sampling_param=sampling_param,
+        negative_prompt="",
+        height=704,
+        width=1280,
+        num_frames=77,
+        num_inference_steps=35,
+        guidance_scale=7.0,
+        fps=24,
        output_path="outputs_video/cosmos2_5_t2w.mp4",
        save_video=True,
    )
@@ -1,54 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
-
-
-def main():
-    # Point this to your local diffusers model dir (or replace with a HF model ID).
-    model_path = "KyleShao/Cosmos-Predict2.5-2B-Diffusers"
-
-    generator = VideoGenerator.from_pretrained(
-        model_path,
-        num_gpus=1,
-        use_fsdp_inference=False,  # set True if GPU is out of memory
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=True,
-        pin_cpu_memory=True,
-    )
-
-    sampling_param = SamplingParam.from_pretrained(model_path)
-
-    # video2world example from official repo
-    video_path = "videos/robot_pouring.mp4"
-
-    prompt = (
-        "A robotic arm, primarily white with black joints and cables, is shown in a clean, modern indoor setting with a white tabletop. "
-        "The arm, equipped with a gripper holding a small, light green pitcher, is positioned above a clear glass containing a reddish-brown liquid and a spoon. "
-        "The robotic arm is in the process of pouring a transparent liquid into the glass. "
-        "To the left of the pitcher, there is an opened jar with a similar reddish-brown substance visible through its transparent body. "
-        "In the background, a vase with white flowers and a brown couch are partially visible, adding to the contemporary ambiance. "
-        "The lighting is bright, casting soft shadows on the table. "
-        "The robotic arm's movements are smooth and controlled, demonstrating precision in its task. "
-        "As the video progresses, the robotic arm completes the pour, leaving the glass half-filled with the reddish-brown liquid. "
-        "The jar remains untouched throughout the sequence, and the spoon inside the glass remains stationary. "
-        "The other robotic arm on the right side also stays stationary throughout the video. "
-        "The final frame captures the robotic arm with the pitcher finishing the pour, with the glass now filled to a higher level, while the pitcher is slightly tilted but still held securely by the gripper."
-    )
-
-    generator.generate_video(
-        prompt,
-        sampling_param=sampling_param,
-        video_path=str(video_path),
-        num_cond_frames=1,
-        output_path="outputs_video/cosmos2_5_v2w.mp4",
-        save_video=True,
-    )
-
-    generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()
-
-
@@ -39,4 +39,4 @@ def main():


 if __name__ == "__main__":
-    main()
+    main()
@@ -1,43 +0,0 @@
-from fastvideo import VideoGenerator
-import json
-# from fastvideo.configs.sample import SamplingParam
-
-OUTPUT_PATH = "video_samples_hy15_1080p"
-def main():
-    # FastVideo will automatically use the optimal default arguments for the
-    # model.
-    # If a local path is provided, FastVideo will make a best effort
-    # attempt to identify the optimal arguments.
-    generator = VideoGenerator.from_pretrained(
-        "weizhou03/HunyuanVideo-1.5-Diffusers-1080p-2SR", # 480p -> 720p -> 1080p
-        # or "weizhou03/HunyuanVideo-1.5-Diffusers-1080p" # 720p -> 1080p
-        # FastVideo will automatically handle distributed setup
-        num_gpus=1,
-        use_fsdp_inference=False, # set to True if GPU is out of memory
-        dit_cpu_offload=True,
-        vae_cpu_offload=True,
-        text_encoder_cpu_offload=True,
-        pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
-        # image_encoder_cpu_offload=False,
-    )
-
-    prompt = (
-        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
-        "wide with interest. The playful yet serene atmosphere is complemented by soft "
-        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
-    )
-
-    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, negative_prompt="")
-
-    prompt2 = (
-        "A majestic lion strides across the golden savanna, its powerful frame "
-        "glistening under the warm afternoon sun. The tall grass ripples gently in "
-        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-        "cinematic.")
-
-    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True, negative_prompt="")
-
-
-if __name__ == "__main__":
-    main()
@@ -1,11 +1,165 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Basic example for HYWorld (HY-WorldPlay) video generation using FastVideo.
+
+This example replicates the same functionality as HY-WorldPlay/run.sh,
+demonstrating image-to-video generation with camera trajectory control.
+"""
+
+import time
+import math
+import numpy as np
+import imageio
+import torchvision
+from einops import rearrange
+
 from fastvideo import VideoGenerator
+from fastvideo.pipelines import ForwardBatch
+from fastvideo.utils import shallow_asdict, align_to
+from fastvideo.logger import init_logger
+
+from fastvideo.models.dits.hyworld.pose import pose_to_input, compute_latent_num
 from fastvideo.models.dits.hyworld.resolution_utils import get_resolution_from_image

+logger = init_logger(__name__)
+
+class HYWorldVideoGenerator(VideoGenerator):
+    """Extended VideoGenerator that adds HYWorld-specific parameters to batch.extra."""
+
+    def _generate_single_video(self, prompt: str, sampling_param=None, **kwargs):
+        """Override to add viewmats, Ks, and action to batch.extra."""
+        fastvideo_args = self.fastvideo_args
+        pipeline_config = fastvideo_args.pipeline_config
+
+        if sampling_param is None:
+            from fastvideo.configs.sample import SamplingParam
+            sampling_param = SamplingParam.from_pretrained(fastvideo_args.model_path)
+
+        # Update sampling param with kwargs
+        if kwargs:
+            for key, value in kwargs.items():
+                if hasattr(sampling_param, key):
+                    setattr(sampling_param, key, value)
+
+        # Get pose string from sampling_param or kwargs
+        pose = kwargs.get('pose', getattr(sampling_param, 'POSE', 'w-31'))
+        num_frames = kwargs.get('num_frames', getattr(sampling_param, 'num_frames', 125))
+
+        # Calculate number of latents
+        latent_num = compute_latent_num(num_frames)
+
+        # Convert pose to viewmats, Ks, and action
+        viewmats, Ks, action = pose_to_input(pose, latent_num)
+
+        # Convert to tensors and add batch dimension
+        viewmats = viewmats.unsqueeze(0)  # (1, T, 4, 4)
+        Ks = Ks.unsqueeze(0)  # (1, T, 3, 3)
+        action = action.unsqueeze(0)  # (1, T)
+
+        # Validate inputs
+        prompt = prompt.strip()
+        sampling_param = sampling_param.__class__(**shallow_asdict(sampling_param))
+        output_path = kwargs.get("output_path", sampling_param.output_path)
+        sampling_param.prompt = prompt
+
+        if sampling_param.negative_prompt is not None:
+            sampling_param.negative_prompt = sampling_param.negative_prompt.strip()
+
+        # Validate dimensions
+        if (sampling_param.height <= 0 or sampling_param.width <= 0 or
+            sampling_param.num_frames <= 0):
+            raise ValueError(
+                f"Height, width, and num_frames must be positive integers")
+
+        temporal_scale_factor = pipeline_config.vae_config.arch_config.temporal_compression_ratio
+        num_frames = sampling_param.num_frames
+        num_gpus = fastvideo_args.num_gpus
+        use_temporal_scaling_frames = pipeline_config.vae_config.use_temporal_scaling_frames
+
+        # Adjust number of frames based on number of GPUs
+        if use_temporal_scaling_frames:
+            orig_latent_num_frames = (num_frames - 1) // temporal_scale_factor + 1
+        else:
+            orig_latent_num_frames = sampling_param.num_frames // 17 * 3
+
+        if orig_latent_num_frames % fastvideo_args.num_gpus != 0:
+            if use_temporal_scaling_frames:
+                new_num_frames = (orig_latent_num_frames - 1) * temporal_scale_factor + 1
+            else:
+                divisor = math.lcm(3, num_gpus)
+                orig_latent_num_frames = (
+                    (orig_latent_num_frames + divisor - 1) // divisor) * divisor
+                new_num_frames = orig_latent_num_frames // 3 * 17
+
+            logger.info(
+                "Adjusting number of frames from %s to %s based on number of GPUs (%s)",
+                sampling_param.num_frames, new_num_frames, fastvideo_args.num_gpus)
+            sampling_param.num_frames = new_num_frames
+
+        # Calculate sizes
+        target_height = align_to(sampling_param.height, 16)
+        target_width = align_to(sampling_param.width, 16)
+
+        # Calculate latent sizes
+        latents_size = [(sampling_param.num_frames - 1) // 4 + 1,
+                        sampling_param.height // 8, sampling_param.width // 8]
+        n_tokens = latents_size[0] * latents_size[1] * latents_size[2]
+
+        # Prepare batch
+        batch = ForwardBatch(
+            **shallow_asdict(sampling_param),
+            eta=0.0,
+            n_tokens=n_tokens,
+            VSA_sparsity=fastvideo_args.VSA_sparsity,
+        )
+
+        # Add HYWorld-specific parameters to batch.extra
+        batch.extra['viewmats'] = viewmats
+        batch.extra['Ks'] = Ks
+        batch.extra['action'] = action
+        batch.extra['chunk_latent_frames'] = 16  # For bidirectional model
+
+        # Run inference
+        start_time = time.perf_counter()
+        output_batch = self.executor.execute_forward(batch, fastvideo_args)
+        samples = output_batch.output
+        logging_info = output_batch.logging_info
+
+        gen_time = time.perf_counter() - start_time
+        logger.info("Generated successfully in %.2f seconds", gen_time)
+
+        # Process outputs
+        videos = rearrange(samples, "b c t h w -> t b c h w")
+        frames = []
+        for x in videos:
+            x = torchvision.utils.make_grid(x, nrow=6)
+            x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+            frames.append((x * 255).numpy().astype(np.uint8))
+
+        # Save video if requested
+        if batch.save_video:
+            imageio.mimsave(output_path, frames, fps=batch.fps, format="mp4")
+            logger.info("Saved video to %s", output_path)
+
+        if batch.return_frames:
+            return frames
+        else:
+            return {
+                "samples": samples,
+                "frames": frames,
+                "prompts": prompt,
+                "size": (target_height, target_width, batch.num_frames),
+                "generation_time": gen_time,
+                "logging_info": logging_info,
+                "trajectory": output_batch.trajectory_latents,
+                "trajectory_timesteps": output_batch.trajectory_timesteps,
+                "trajectory_decoded": output_batch.trajectory_decoded,
+            }
+
 # Default prompt from HY-WorldPlay run.sh
 DEFAULT_PROMPT = 'A paved pathway leads towards a stone arch bridge spanning a calm body of water.  Lush green trees and foliage line the path and the far bank of the water. A traditional-style pavilion with a tiered, reddish-brown roof sits on the far shore. The water reflects the surrounding greenery and the sky.  The scene is bathed in soft, natural light, creating a tranquil and serene atmosphere. The pathway is composed of large, rectangular stones, and the bridge is constructed of light gray stone.  The overall composition emphasizes the peaceful and harmonious nature of the landscape.'
 DEFAULT_IMAGE = 'https://raw.githubusercontent.com/Tencent-Hunyuan/HY-WorldPlay/main/assets/img/test.png'

-OUTPUT_PATH = "video_samples_hyworld"
 def main():
    import argparse

@@ -15,7 +169,7 @@ def main():
    parser.add_argument("--prompt", type=str, default=DEFAULT_PROMPT, help="Text prompt for video generation")
    parser.add_argument("--image", type=str, default=DEFAULT_IMAGE, help="Path or URL to input image")
    parser.add_argument("--pose", type=str, default='w-31', help="Pose string (e.g., 'a-31', 'w-31', 's-31', 'd-31')")
-    parser.add_argument("--output_path", type=str, default=OUTPUT_PATH, help="Output video path")
+    parser.add_argument("--output_path", type=str, default='video_samples_hyworld', help="Output video path")
    parser.add_argument("--num-frames", type=int, default=125, help="Number of frames")
    parser.add_argument("--seed", type=int, default=1, help="Random seed")
    parser.add_argument("--resolution", type=str, default="480p", help="Only support 480p for now")
@@ -31,7 +185,8 @@ def main():

    # Initialize generator
    print("\nInitializing VideoGenerator for HYWorld...")
-    generator = VideoGenerator.from_pretrained(
+    
+    generator = HYWorldVideoGenerator.from_pretrained(
        "FastVideo/HY-WorldPlay-Bidirectional-Diffusers",
        num_gpus=1,
        use_fsdp_inference=True,
@@ -43,12 +198,11 @@ def main():
    )

    # Generate video
-    # The pose string is automatically converted to camera matrices by the pipeline
    print("\nGenerating video...")
-    generator.generate_video(
+    start_time = time.time()
+    video = generator.generate_video(
        prompt=args.prompt,
        image_path=args.image,
-        pose=args.pose,  # Camera trajectory control
        output_path=args.output_path,
        save_video=True,
        negative_prompt="",
@@ -57,9 +211,13 @@ def main():
        height=HEIGHT,
        width=WIDTH,
        seed=args.seed,
+        pose=args.pose,
    )
+    elapsed = time.time() - start_time

-    print(f"\nVideo saved to: {args.output_path}")
+    print(f"\nVideo generated successfully!")
+    print(f"Saved to: {args.output_path}")
+    print(f"Time: {elapsed:.2f}s")


 if __name__ == "__main__":
@@ -0,0 +1,18 @@
+Total Files: 16
+
+00. Hold [W] + Static
+01. Hold [S] + Static
+02. Hold [A] + Static
+03. Hold [D] + Static
+04. Hold [WA] + Static
+05. Hold [WD] + Static
+06. Hold [SA] + Static
+07. Hold [SD] + Static
+08. No Key + Hold [up]
+09. No Key + Hold [down]
+10. No Key + Hold [left]
+11. No Key + Hold [right]
+12. No Key + Hold [up_right]
+13. No Key + Hold [up_left]
+14. No Key + Hold [down_right]
+15. No Key + Hold [down_left]
@@ -0,0 +1,93 @@
+#!/bin/bash
+
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=offline
+export TOKENIZERS_PARALLELISM=false
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+MODEL_PATH="weizhou03/Wan2.1-Game-Fun-1.3B-InP-Diffusers"
+DATA_DIR="mc_wasd_10/preprocessed/combined_parquet_dataset"
+VALIDATION_DATASET_FILE="mc_wasd_10/validation.json"
+NUM_GPUS=4
+# export CUDA_VISIBLE_DEVICES=0,1,2,3
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wangame_1.3b_overfit"
+  --output_dir "wangame_1.3b_overfit"
+  --max_train_steps 1500
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 20
+  --num_height 352
+  --num_width 640
+  --num_frames 77
+  --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim $NUM_GPUS
+)
+
+# 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 100
+  --validation_sampling_steps "40"
+  --validation_guidance_scale "1.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 2e-5
+  --mixed_precision "bf16"
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
+  --weight_decay 1e-4
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.1
+  --multi_phased_distill_schedule "4000-1"
+  --not_apply_cfg_solver
+  --dit_precision "fp32"
+  --num_euler_timesteps 50
+  --ema_start_step 0
+)
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+    fastvideo/training/wangame_training_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"
@@ -0,0 +1,120 @@
+#!/bin/bash
+#SBATCH --job-name=wangame_1.3b_overfit
+#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=wangame_1.3b_overfit_output/wangame_1.3b_overfit_%j.out
+#SBATCH --error=wangame_1.3b_overfit_output/wangame_1.3b_overfit_%j.err
+#SBATCH --exclusive
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+export NCCL_DEBUG_SUBSYS=INIT,NET
+# 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 TOKENIZERS_PARALLELISM=false
+# export WANDB_API_KEY="8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde"
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+# export WANDB_API_KEY='your_wandb_api_key_here'
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+source ~/conda/miniconda/bin/activate
+conda activate wei-fv-distill
+export HOME="/mnt/weka/home/hao.zhang/wei"
+
+MODEL_PATH="weizhou03/Wan2.1-Game-Fun-1.3B-InP-Diffusers"
+DATA_DIR="mc_wasd_10/preprocessed/combined_parquet_dataset"
+VALIDATION_DATASET_FILE="examples/training/finetune/WanGame2.1_1.3b_i2v/validation.json"
+# Configs
+NUM_GPUS=8
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wangame_1.3b_overfit"
+  --output_dir "wangame_1.3b_overfit"
+  --max_train_steps 15000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 20
+  --num_height 352
+  --num_width 640
+  --num_frames 77
+  --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim $NUM_GPUS
+)
+
+# 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 100
+  --validation_sampling_steps "40"
+  --validation_guidance_scale "1.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 2e-5
+  --mixed_precision "bf16"
+  --weight_only_checkpointing_steps 1000000
+  --training_state_checkpointing_steps 10000000
+  --weight_decay 1e-4
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.1
+  --multi_phased_distill_schedule "4000-1"
+  --not_apply_cfg_solver
+  --dit_precision "fp32"
+  --num_euler_timesteps 50
+  --ema_start_step 0
+)
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+    fastvideo/training/wangame_training_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"
@@ -0,0 +1,193 @@
+import os
+import numpy as np
+
+# Configuration
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+BASE_OUTPUT_DIR = os.path.join(SCRIPT_DIR, 'action')
+VIDEO_OUTPUT_DIR = BASE_OUTPUT_DIR
+os.makedirs(VIDEO_OUTPUT_DIR, exist_ok=True)
+
+FRAME_COUNT = 81
+CAM_VALUE = 0.1
+
+# Action Mapping
+KEY_TO_INDEX = {
+    'W': 0, 'S': 1, 'A': 2, 'D': 3,
+}
+
+VIEW_ACTION_TO_MOUSE = {
+    "stop": [0.0, 0.0],
+    "up": [CAM_VALUE, 0.0],
+    "down": [-CAM_VALUE, 0.0],
+    "left": [0.0, -CAM_VALUE],
+    "right": [0.0, CAM_VALUE],
+    "up_right": [CAM_VALUE, CAM_VALUE],
+    "up_left": [CAM_VALUE, -CAM_VALUE],
+    "down_right": [-CAM_VALUE, CAM_VALUE],
+    "down_left": [-CAM_VALUE, -CAM_VALUE],
+}
+
+def get_multihot_vector(keys_str):
+    """Convert string like 'WA' to [1, 0, 1, 0, 0, 0]"""
+    vector = [0.0] * 6
+    if not keys_str:
+        return vector
+    for char in keys_str.upper():
+        if char in KEY_TO_INDEX:
+            vector[KEY_TO_INDEX[char]] = 1.0
+    return vector
+
+def get_mouse_vector(view_str):
+    """Convert view string to [x, y]"""
+    return VIEW_ACTION_TO_MOUSE.get(view_str.lower(), [0.0, 0.0])
+
+def generate_sequence(key_seq, mouse_seq):
+    """
+    Generates action arrays based on sequences.
+    """
+    keyboard_arr = np.zeros((FRAME_COUNT, 6), dtype=np.float32)
+    mouse_arr = np.zeros((FRAME_COUNT, 2), dtype=np.float32)
+    
+    mid_point = FRAME_COUNT // 2
+
+    # First Half
+    k_vec1 = get_multihot_vector(key_seq[0])
+    m_vec1 = get_mouse_vector(mouse_seq[0])
+    keyboard_arr[:mid_point] = k_vec1
+    mouse_arr[:mid_point] = m_vec1
+
+    # Second Half
+    k_vec2 = get_multihot_vector(key_seq[1])
+    m_vec2 = get_mouse_vector(mouse_seq[1])
+    keyboard_arr[mid_point:] = k_vec2
+    mouse_arr[mid_point:] = m_vec2
+
+    return keyboard_arr, mouse_arr
+
+def save_action(index, keyboard_arr, mouse_arr):
+    filename = f"{index:06d}_action.npy"
+    filepath = os.path.join(VIDEO_OUTPUT_DIR, filename)
+    
+    action_dict = {
+        'keyboard': keyboard_arr,
+        'mouse': mouse_arr
+    }
+    np.save(filepath, action_dict)
+    return filename
+
+def generate_description(key_seq, mouse_seq):
+    """Generates a human-readable string for the combination."""
+    k1, k2 = key_seq
+    m1, m2 = mouse_seq
+    
+    # Format Keyboard Description
+    if not k1 and not k2:
+        k_desc = "No Key"
+    elif k1 == k2:
+        k_desc = f"Hold [{k1}]"
+    else:
+        k_desc = f"Switch [{k1}]->[{k2}]"
+        
+    # Format Mouse Description
+    if m1 == "stop" and m2 == "stop":
+        m_desc = "Static"
+    elif m1 == m2:
+        m_desc = f"Hold [{m1}]"
+    else:
+        m_desc = f"Switch [{m1}]->[{m2}]"
+        
+    return f"{k_desc} + {m_desc}"
+
+# ==========================================
+# Main Generation Logic
+# ==========================================
+
+configs = []
+readme_content = []
+
+# Group 1: Constant Keyboard, No Mouse (0-7)
+keys_basic = ['W', 'S', 'A', 'D', 'WA', 'WD', 'SA', 'SD']
+for k in keys_basic:
+    configs.append(((k, k), ("stop", "stop")))
+
+# Group 2: No Keyboard, Constant Mouse (8-15)
+mouse_basic = ['up', 'down', 'left', 'right', 'up_right', 'up_left', 'down_right', 'down_left']
+for m in mouse_basic:
+    configs.append((("", ""), (m, m)))
+
+# Group 3: Split Keyboard, No Mouse (16-23)
+split_keys = [
+    ('W', 'S'), ('S', 'W'), 
+    ('A', 'D'), ('D', 'A'),
+    ('W', 'A'), ('W', 'D'), 
+    ('S', 'A'), ('S', 'D')
+]
+for k1, k2 in split_keys:
+    configs.append(((k1, k2), ("stop", "stop")))
+
+# Group 4: No Keyboard, Split Mouse (24-31)
+split_mouse = [
+    ('left', 'right'), ('right', 'left'),
+    ('up', 'down'), ('down', 'up'),
+    ('up_left', 'up_right'), ('up_right', 'up_left'),
+    ('left', 'up'), ('right', 'down')
+]
+for m1, m2 in split_mouse:
+    configs.append((("", ""), (m1, m2)))
+
+# Group 5: Constant Keyboard + Constant Mouse (32-47)
+combo_keys = ['W', 'S', 'W', 'S', 'A', 'D', 'WA', 'WD', 'W', 'S', 'W', 'S', 'A', 'D', 'WA', 'WD']
+combo_mice = ['left', 'left', 'right', 'right', 'up', 'up', 'down', 'down', 'up_left', 'up_left', 'up_right', 'up_right', 'down_left', 'down_right', 'right', 'left']
+for i in range(16):
+    configs.append(((combo_keys[i], combo_keys[i]), (combo_mice[i], combo_mice[i])))
+
+# Group 6: Constant Keyboard, Split Mouse (48-55)
+complex_1_keys = ['W'] * 8
+complex_1_mice = [
+    ('left', 'right'), ('right', 'left'), 
+    ('up', 'down'), ('down', 'up'),
+    ('left', 'up'), ('right', 'up'),
+    ('left', 'down'), ('right', 'down')
+]
+for i in range(8):
+    configs.append(((complex_1_keys[i], complex_1_keys[i]), complex_1_mice[i]))
+
+# Group 7: Split Keyboard, Constant Mouse (56-63)
+complex_2_keys = [
+    ('W', 'S'), ('S', 'W'),
+    ('A', 'D'), ('D', 'A'),
+    ('W', 'A'), ('W', 'D'),
+    ('S', 'A'), ('S', 'D')
+]
+complex_2_mouse = 'up' 
+for k1, k2 in complex_2_keys:
+    configs.append(((k1, k2), (complex_2_mouse, complex_2_mouse)))
+
+
+# Execution
+print(f"Preparing to generate {len(configs)} action files...")
+
+for i, (key_seq, mouse_seq) in enumerate(configs):
+    if i >= 16: break 
+    
+    # Generate Data
+    kb_arr, ms_arr = generate_sequence(key_seq, mouse_seq)
+    filename = save_action(i, kb_arr, ms_arr)
+    
+    # Generate Description for README
+    description = generate_description(key_seq, mouse_seq)
+    readme_entry = f"{i:02d}. {description}"
+    readme_content.append(readme_entry)
+    
+    print(f"Generated {filename} -> {description}")
+
+# Write README
+readme_path = os.path.join(BASE_OUTPUT_DIR, 'README.md')
+with open(readme_path, 'w', encoding='utf-8') as f:
+    f.write(f"Total Files: {len(readme_content)}\n\n")
+    for line in readme_content:
+        f.write(line + '\n')
+
+print(f"\nProcessing complete.")
+print(f"64 .npy files generated in {VIDEO_OUTPUT_DIR}")
+print(f"Manifest saved to {readme_path}")
@@ -0,0 +1,27 @@
+#!/bin/bash
+
+GPU_NUM=1 # 2,4,8
+MODEL_PATH="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
+DATA_MERGE_PATH="mc_wasd_10/merge.txt"
+OUTPUT_DIR="mc_wasd_10/preprocessed/"
+
+# export CUDA_VISIBLE_DEVICES=0
+export MASTER_ADDR=localhost
+export MASTER_PORT=29500
+export RANK=0
+export WORLD_SIZE=1
+
+python fastvideo/pipelines/preprocess/v1_preprocess.py \
+    --model_path $MODEL_PATH \
+    --data_merge_path $DATA_MERGE_PATH \
+    --preprocess_video_batch_size 10 \
+    --seed 42 \
+    --max_height 352 \
+    --max_width 640 \
+    --num_frames 77 \
+    --dataloader_num_workers 0 \
+    --output_dir=$OUTPUT_DIR \
+    --samples_per_file 10 \
+    --train_fps 25 \
+    --flush_frequency 10 \
+    --preprocess_task wangame
@@ -0,0 +1,404 @@
+{
+    "data": [
+        {
+            "caption": "0",
+            "image_path": "../../../../mc_wasd_10/validate/000000.jpg",
+            "action_path": "../../../../mc_wasd_10/videos/000000_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "1",
+            "image_path": "../../../../mc_wasd_10/validate/000001.jpg",
+            "action_path": "../../../../mc_wasd_10/videos/000001_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "2",
+            "image_path": "../../../../mc_wasd_10/validate/000002.jpg",
+            "action_path": "../../../../mc_wasd_10/videos/000002_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "3",
+            "image_path": "../../../../mc_wasd_10/validate/000003.jpg",
+            "action_path": "../../../../mc_wasd_10/videos/000003_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "4",
+            "image_path": "../../../../mc_wasd_10/validate/000004.jpg",
+            "action_path": "../../../../mc_wasd_10/videos/000004_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "5",
+            "image_path": "../../../../mc_wasd_10/validate/000005.jpg",
+            "action_path": "../../../../mc_wasd_10/videos/000005_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "6",
+            "image_path": "../../../../mc_wasd_10/validate/000006.jpg",
+            "action_path": "../../../../mc_wasd_10/videos/000006_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "7",
+            "image_path": "../../../../mc_wasd_10/validate/000007.jpg",
+            "action_path": "../../../../mc_wasd_10/videos/000007_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "00. Hold [W] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/000000.jpg",
+            "action_path": "action/000000_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "01. Hold [S] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/000001.jpg",
+            "action_path": "action/000001_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "02. Hold [A] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/000002.jpg",
+            "action_path": "action/000002_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "03. Hold [D] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/000003.jpg",
+            "action_path": "action/000003_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "04. Hold [WA] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/000004.jpg",
+            "action_path": "action/000004_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "05. Hold [WD] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/000005.jpg",
+            "action_path": "action/000005_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "06. Hold [SA] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/000006.jpg",
+            "action_path": "action/000006_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "07. Hold [SD] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/000007.jpg",
+            "action_path": "action/000007_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "08. No Key + Hold [up]",
+            "image_path": "../../../../mc_wasd_10/validate/000000.jpg",
+            "action_path": "action/000008_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "09. No Key + Hold [down]",
+            "image_path": "../../../../mc_wasd_10/validate/000001.jpg",
+            "action_path": "action/000009_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "10. No Key + Hold [left]",
+            "image_path": "../../../../mc_wasd_10/validate/000002.jpg",
+            "action_path": "action/000010_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "11. No Key + Hold [right]",
+            "image_path": "../../../../mc_wasd_10/validate/000003.jpg",
+            "action_path": "action/000011_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "12. No Key + Hold [up_right]",
+            "image_path": "../../../../mc_wasd_10/validate/000004.jpg",
+            "action_path": "action/000012_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "13. No Key + Hold [up_left]",
+            "image_path": "../../../../mc_wasd_10/validate/000005.jpg",
+            "action_path": "action/000013_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "14. No Key + Hold [down_right]",
+            "image_path": "../../../../mc_wasd_10/validate/000006.jpg",
+            "action_path": "action/000014_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "15. No Key + Hold [down_left]",
+            "image_path": "../../../../mc_wasd_10/validate/000007.jpg",
+            "action_path": "action/000015_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "00. Hold [W] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000000.jpg",
+            "action_path": "action/000000_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "01. Hold [S] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000001.jpg",
+            "action_path": "action/000001_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "02. Hold [A] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000002.jpg",
+            "action_path": "action/000002_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "03. Hold [D] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000003.jpg",
+            "action_path": "action/000003_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "04. Hold [WA] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000004.jpg",
+            "action_path": "action/000004_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "05. Hold [WD] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000005.jpg",
+            "action_path": "action/000005_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "06. Hold [SA] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000006.jpg",
+            "action_path": "action/000006_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "07. Hold [SD] + Static",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000007.jpg",
+            "action_path": "action/000007_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "08. No Key + Hold [up]",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000000.jpg",
+            "action_path": "action/000008_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "09. No Key + Hold [down]",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000001.jpg",
+            "action_path": "action/000009_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "10. No Key + Hold [left]",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000002.jpg",
+            "action_path": "action/000010_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "11. No Key + Hold [right]",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000003.jpg",
+            "action_path": "action/000011_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "12. No Key + Hold [up_right]",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000004.jpg",
+            "action_path": "action/000012_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "13. No Key + Hold [up_left]",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000005.jpg",
+            "action_path": "action/000013_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "14. No Key + Hold [down_right]",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000006.jpg",
+            "action_path": "action/000014_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        },
+        {
+            "caption": "15. No Key + Hold [down_left]",
+            "image_path": "../../../../mc_wasd_10/validate/gen_000007.jpg",
+            "action_path": "action/000015_action.npy",
+            "video_path": null,
+            "num_inference_steps": 40,
+            "height": 352,
+            "width": 640,
+            "num_frames": 77
+        }
+    ]
+}
@@ -5,7 +5,6 @@ from fastvideo.configs.models.vaes.base import VAEConfig
 from fastvideo.configs.models.audio import (LTX2AudioDecoderConfig,
                                            LTX2AudioEncoderConfig,
                                            LTX2VocoderConfig)
-from fastvideo.configs.models.upsamplers.base import UpsamplerConfig

 __all__ = [
    "ModelConfig",
@@ -15,5 +14,4 @@ __all__ = [
    "LTX2AudioEncoderConfig",
    "LTX2AudioDecoderConfig",
    "LTX2VocoderConfig",
-    "UpsamplerConfig",
 ]
@@ -55,8 +55,6 @@ class HunyuanVideo15ArchConfig(DiTArchConfig):
            r"txt_in.refiner_blocks.\1.mlp.fc_out.\2",
            r"^context_embedder\.token_refiner\.refiner_blocks\.(\d+)\.norm_out\.linear\.(.*)$":
            r"txt_in.refiner_blocks.\1.adaLN_modulation.linear.\2",
-            r"^context_embedder\.token_refiner\.refiner_blocks\.(\d+)\.self_attn_qkv\.(.*)$":
-            r"txt_in.refiner_blocks.\1.self_attn_qkv.\2",

            # 2. txt_in_2 mapping:
            r"^context_embedder_2\.(.*)$":
@@ -8,8 +8,12 @@ def is_double_block(n: str, m) -> bool:
    return "double" in n and str.isdigit(n.split(".")[-1])


-def is_refiner_block(n: str, m) -> bool:
-    return "refiner" in n and str.isdigit(n.split(".")[-1])
+def is_single_block(n: str, m) -> bool:
+    return "single" in n and str.isdigit(n.split(".")[-1])
+
+
+# def is_refiner_block(n: str, m) -> bool:
+#     return "refiner" in n and str.isdigit(n.split(".")[-1])


 def is_txt_in(n: str, m) -> bool:
@@ -19,10 +23,10 @@ def is_txt_in(n: str, m) -> bool:
@dataclass
 class HYWorldArchConfig(DiTArchConfig):
    _fsdp_shard_conditions: list = field(
-        default_factory=lambda: [is_double_block, is_refiner_block])
+        default_factory=lambda: [is_double_block, is_single_block])

    _compile_conditions: list = field(
-        default_factory=lambda: [is_double_block, is_refiner_block, is_txt_in])
+        default_factory=lambda: [is_double_block, is_single_block, is_txt_in])

    param_names_mapping: dict = field(
        default_factory=lambda: {
@@ -50,7 +54,8 @@ class HYWorldArchConfig(DiTArchConfig):
            r"^txt_in\.individual_token_refiner\.blocks\.(\d+)\.adaLN_modulation\.1\.(.*)$":
            r"txt_in.refiner_blocks.\1.adaLN_modulation.linear.\2",

-            # 2. time_in mappings:
+            # 2. time_in mappings (HYWorld uses TimestepEmbedder directly,
+            #    but FastVideo model inherits HunyuanVideo15TimeEmbedding with timestep_embedder):
            r"^time_in\.mlp\.0\.(.*)$":
            r"time_in.timestep_embedder.mlp.fc_in.\1",
            r"^time_in\.mlp\.2\.(.*)$":
@@ -1,6 +1,4 @@
 from dataclasses import dataclass, field
-
-import torch
 from fastvideo.configs.models.dits.wanvideo import WanVideoArchConfig, WanVideoConfig


@@ -78,14 +76,8 @@ class MatrixGameWanVideoArchConfig(WanVideoArchConfig):
    image_dim: int = 1280


-def _is_transformer_block(param_name: str, module: torch.nn.Module) -> bool:
-    return bool("blocks" in param_name and param_name.split(".")[-1].isdigit())
-
-
@dataclass
 class MatrixGameWanVideoConfig(WanVideoConfig):
    arch_config: MatrixGameWanVideoArchConfig = field(
        default_factory=MatrixGameWanVideoArchConfig)
    prefix: str = "Wan"
-    _compile_conditions: list = field(
-        default_factory=lambda: [_is_transformer_block])
@@ -0,0 +1,115 @@
+# SPDX-License-Identifier: Apache-2.0
+from dataclasses import dataclass, field
+
+from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+
+
+def is_blocks(n: str, m) -> bool:
+    return "blocks" in n and str.isdigit(n.split(".")[-1])
+
+
+@dataclass
+class WanGameVideoArchConfig(DiTArchConfig):
+    _fsdp_shard_conditions: list = field(default_factory=lambda: [is_blocks])
+
+    param_names_mapping: dict = field(
+        default_factory=lambda: {
+            r"^patch_embedding\.(.*)$":
+            r"patch_embedding.proj.\1",
+            r"^condition_embedder\.text_embedder\.linear_1\.(.*)$":
+            r"condition_embedder.text_embedder.fc_in.\1",
+            r"^condition_embedder\.text_embedder\.linear_2\.(.*)$":
+            r"condition_embedder.text_embedder.fc_out.\1",
+            r"^condition_embedder\.time_embedder\.linear_1\.(.*)$":
+            r"condition_embedder.time_embedder.mlp.fc_in.\1",
+            r"^condition_embedder\.time_embedder\.linear_2\.(.*)$":
+            r"condition_embedder.time_embedder.mlp.fc_out.\1",
+            r"^condition_embedder\.time_proj\.(.*)$":
+            r"condition_embedder.time_modulation.linear.\1",
+            r"^condition_embedder\.image_embedder\.ff\.net\.0\.proj\.(.*)$":
+            r"condition_embedder.image_embedder.ff.fc_in.\1",
+            r"^condition_embedder\.image_embedder\.ff\.net\.2\.(.*)$":
+            r"condition_embedder.image_embedder.ff.fc_out.\1",
+            r"^blocks\.(\d+)\.attn1\.to_q\.(.*)$":
+            r"blocks.\1.to_q.\2",
+            r"^blocks\.(\d+)\.attn1\.to_k\.(.*)$":
+            r"blocks.\1.to_k.\2",
+            r"^blocks\.(\d+)\.attn1\.to_v\.(.*)$":
+            r"blocks.\1.to_v.\2",
+            r"^blocks\.(\d+)\.attn1\.to_out\.0\.(.*)$":
+            r"blocks.\1.to_out.\2",
+            r"^blocks\.(\d+)\.attn1\.norm_q\.(.*)$":
+            r"blocks.\1.norm_q.\2",
+            r"^blocks\.(\d+)\.attn1\.norm_k\.(.*)$":
+            r"blocks.\1.norm_k.\2",
+            r"^blocks\.(\d+)\.attn2\.to_out\.0\.(.*)$":
+            r"blocks.\1.attn2.to_out.\2",
+            r"^blocks\.(\d+)\.ffn\.net\.0\.proj\.(.*)$":
+            r"blocks.\1.ffn.fc_in.\2",
+            r"^blocks\.(\d+)\.ffn\.net\.2\.(.*)$":
+            r"blocks.\1.ffn.fc_out.\2",
+            r"^blocks\.(\d+)\.norm2\.(.*)$":
+            r"blocks.\1.self_attn_residual_norm.norm.\2",
+        })
+
+    # Reverse mapping for saving checkpoints: custom -> hf
+    reverse_param_names_mapping: dict = field(default_factory=lambda: {})
+
+    # Some LoRA adapters use the original official layer names instead of hf layer names,
+    # so apply this before the param_names_mapping
+    lora_param_names_mapping: dict = field(
+        default_factory=lambda: {
+            r"^blocks\.(\d+)\.self_attn\.q\.(.*)$": r"blocks.\1.attn1.to_q.\2",
+            r"^blocks\.(\d+)\.self_attn\.k\.(.*)$": r"blocks.\1.attn1.to_k.\2",
+            r"^blocks\.(\d+)\.self_attn\.v\.(.*)$": r"blocks.\1.attn1.to_v.\2",
+            r"^blocks\.(\d+)\.self_attn\.o\.(.*)$":
+            r"blocks.\1.attn1.to_out.0.\2",
+            r"^blocks\.(\d+)\.cross_attn\.q\.(.*)$": r"blocks.\1.attn2.to_q.\2",
+            r"^blocks\.(\d+)\.cross_attn\.k\.(.*)$": r"blocks.\1.attn2.to_k.\2",
+            r"^blocks\.(\d+)\.cross_attn\.v\.(.*)$": r"blocks.\1.attn2.to_v.\2",
+            r"^blocks\.(\d+)\.cross_attn\.o\.(.*)$":
+            r"blocks.\1.attn2.to_out.0.\2",
+            r"^blocks\.(\d+)\.ffn\.0\.(.*)$": r"blocks.\1.ffn.fc_in.\2",
+            r"^blocks\.(\d+)\.ffn\.2\.(.*)$": r"blocks.\1.ffn.fc_out.\2",
+        })
+
+    patch_size: tuple[int, int, int] = (1, 2, 2)
+    text_len = 512
+    num_attention_heads: int = 40
+    attention_head_dim: int = 128
+    in_channels: int = 16
+    out_channels: int = 16
+    text_dim: int = 4096
+    freq_dim: int = 256
+    ffn_dim: int = 13824
+    num_layers: int = 40
+    cross_attn_norm: bool = True
+    qk_norm: str = "rms_norm_across_heads"
+    eps: float = 1e-6
+    image_dim: int | None = None
+    added_kv_proj_dim: int | None = None
+    rope_max_seq_len: int = 1024
+    pos_embed_seq_len: int | None = None
+    exclude_lora_layers: list[str] = field(default_factory=lambda: ["embedder"])
+
+    # Wan MoE
+    boundary_ratio: float | None = None
+
+    # Causal Wan
+    local_attn_size: int = -1  # Window size for temporal local attention (-1 indicates global attention)
+    sink_size: int = 0  # Size of the attention sink, we keep the first `sink_size` frames unchanged when rolling the KV cache
+    num_frames_per_block: int = 3
+    sliding_window_num_frames: int = 21
+
+    def __post_init__(self):
+        super().__post_init__()
+        self.out_channels = self.out_channels or self.in_channels
+        self.hidden_size = self.num_attention_heads * self.attention_head_dim
+        self.num_channels_latents = self.out_channels
+
+
+@dataclass
+class WanGameVideoConfig(DiTConfig):
+    arch_config: DiTArchConfig = field(default_factory=WanGameVideoArchConfig)
+
+    prefix: str = "WanGame"
@@ -87,10 +87,10 @@ class CLIPVisionConfig(ImageEncoderConfig):
    arch_config: ImageEncoderArchConfig = field(
        default_factory=CLIPVisionArchConfig)

-    num_hidden_layers_override: int | None = 31
+    num_hidden_layers_override: int | None = None
    require_post_norm: bool | None = None
-    enable_scale: bool = False
-    is_causal: bool = False
+    enable_scale: bool = True
+    is_causal: bool = True
    prefix: str = "clip"


@@ -1,6 +0,0 @@
-from fastvideo.configs.models.upsamplers.hunyuan15 import SRTo720pUpsamplerConfig, SRTo1080pUpsamplerConfig
-from fastvideo.configs.models.upsamplers.base import UpsamplerConfig
-
-__all__ = [
-    "SRTo720pUpsamplerConfig", "SRTo1080pUpsamplerConfig", "UpsamplerConfig"
-]
@@ -1,7 +0,0 @@
-from dataclasses import dataclass
-from fastvideo.configs.models.base import ModelConfig
-
-
-@dataclass
-class UpsamplerConfig(ModelConfig):
-    pass
@@ -1,20 +0,0 @@
-from dataclasses import dataclass
-from fastvideo.configs.models.upsamplers.base import UpsamplerConfig
-
-
-@dataclass
-class SRTo720pUpsamplerConfig(UpsamplerConfig):
-    in_channels: int = 0
-    out_channels: int = 0
-    hidden_channels: int = 64
-    num_blocks: int = 6
-    global_residual: bool = False
-
-
-@dataclass
-class SRTo1080pUpsamplerConfig(UpsamplerConfig):
-    z_channels: int = 0
-    out_channels: int = 0
-    block_out_channels: tuple[int, ...] = (0, 0)
-    num_res_blocks: int = 2
-    is_residual: bool = False
@@ -6,7 +6,8 @@ from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig
 from fastvideo.configs.pipelines.hyworld import HYWorldConfig
 from fastvideo.configs.pipelines.ltx2 import LTX2T2VConfig
-from fastvideo.registry import get_pipeline_config_cls_from_name
+from fastvideo.configs.pipelines.registry import (
+    get_pipeline_config_cls_from_name)
 from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
 from fastvideo.configs.pipelines.wan import (SelfForcingWanT2V480PConfig,
                                             WanI2V480PConfig, WanI2V720PConfig,
@@ -8,7 +8,7 @@ from typing import Any, cast
 import torch

 from fastvideo.configs.models import (DiTConfig, EncoderConfig, ModelConfig,
-                                      VAEConfig, UpsamplerConfig)
+                                      VAEConfig)
 from fastvideo.configs.models.encoders import BaseEncoderOutput
 from fastvideo.configs.utils import update_config_from_args
 from fastvideo.logger import init_logger
@@ -44,15 +44,12 @@ class PipelineConfig:
    # Video generation parameters
    embedded_cfg_scale: float = 6.0
    flow_shift: float | None = None
-    flow_shift_sr: float | None = None
    disable_autocast: bool = False
    is_causal: bool = False

    # Model configuration
    dit_config: DiTConfig = field(default_factory=DiTConfig)
    dit_precision: str = "bf16"
-    upsampler_config: UpsamplerConfig = field(default_factory=UpsamplerConfig)
-    upsampler_precision: str = "fp32"

    # VAE configuration
    vae_config: VAEConfig = field(default_factory=VAEConfig)
@@ -219,24 +216,6 @@ class PipelineConfig:
            "Comma-separated list of denoising steps (e.g., '1000,757,522')",
        )

-        # STA (Sliding Tile Attention) parameters
-        parser.add_argument(
-            f"--{prefix_with_dot}STA-mode",
-            type=str,
-            dest=f"{prefix_with_dot.replace('-', '_')}STA_mode",
-            default=PipelineConfig.STA_mode.value,
-            choices=[mode.value for mode in STA_Mode],
-            help=
-            "STA mode: STA_inference, STA_searching, STA_tuning, STA_tuning_cfg, None",
-        )
-        parser.add_argument(
-            f"--{prefix_with_dot}skip-time-steps",
-            type=int,
-            dest=f"{prefix_with_dot.replace('-', '_')}skip_time_steps",
-            default=PipelineConfig.skip_time_steps,
-            help="Number of time steps to warmup (full attention) for STA",
-        )
-
        # Add VAE configuration arguments
        from fastvideo.configs.models.vaes.base import VAEConfig
        VAEConfig.add_cli_args(parser, prefix=f"{prefix_with_dot}vae-config")
@@ -266,7 +245,8 @@ class PipelineConfig:
        """
        use the pipeline class setting from model_path to match the pipeline config
        """
-        from fastvideo.registry import get_pipeline_config_cls_from_name
+        from fastvideo.configs.pipelines.registry import (
+            get_pipeline_config_cls_from_name)
        pipeline_config_cls = get_pipeline_config_cls_from_name(model_path)

        return cast(PipelineConfig, pipeline_config_cls(model_path=model_path))
@@ -280,7 +260,8 @@ class PipelineConfig:
        kwargs: dictionary of kwargs
        config_cli_prefix: prefix of CLI arguments for this PipelineConfig instance
        """
-        from fastvideo.registry import get_pipeline_config_cls_from_name
+        from fastvideo.configs.pipelines.registry import (
+            get_pipeline_config_cls_from_name)

        prefix_with_dot = f"{config_cli_prefix}." if (config_cli_prefix.strip()
                                                      != "") else ""
@@ -317,12 +298,6 @@ class PipelineConfig:
        # 4. Update PipelineConfig from CLI arguments if provided
        kwargs[prefix_with_dot + 'model_path'] = model_path
        pipeline_config.update_config_from_dict(kwargs, config_cli_prefix)
-
-        # Convert STA_mode string to enum if necessary
-        if isinstance(pipeline_config.STA_mode, str) and not isinstance(
-                pipeline_config.STA_mode, STA_Mode):
-            pipeline_config.STA_mode = STA_Mode(pipeline_config.STA_mode)
-
        return pipeline_config

    def check_pipeline_config(self) -> None:
@@ -11,8 +11,7 @@ from fastvideo.configs.models.dits import HunyuanVideo15Config
 from fastvideo.configs.models.encoders import (BaseEncoderOutput,
                                               Qwen2_5_VLConfig, T5Config)
 from fastvideo.configs.models.vaes import Hunyuan15VAEConfig
-from fastvideo.configs.models.upsamplers import SRTo720pUpsamplerConfig, SRTo1080pUpsamplerConfig
-from fastvideo.configs.pipelines.base import PipelineConfig, UpsamplerConfig
+from fastvideo.configs.pipelines.base import PipelineConfig

 PROMPT_TEMPLATE_TOKEN_LENGTH = 108

@@ -132,35 +131,9 @@ class Hunyuan15T2V480PConfig(PipelineConfig):
        self.vae_config.load_decoder = True


-@dataclass
-class Hunyuan15I2V480PStepDistilledConfig(Hunyuan15T2V480PConfig):
-    flow_shift: int = 7
-
-
@dataclass
 class Hunyuan15T2V720PConfig(Hunyuan15T2V480PConfig):
    """Base configuration for HunYuan pipeline architecture."""

    # HunyuanConfig-specific parameters with defaults
    flow_shift: int = 9
-
-
-@dataclass
-class Hunyuan15I2V720PConfig(Hunyuan15T2V720PConfig):
-    """Base configuration for HunYuan pipeline architecture."""
-
-    # HunyuanConfig-specific parameters with defaults
-    flow_shift: int = 7
-
-
-@dataclass
-class Hunyuan15SR1080PConfig(Hunyuan15T2V720PConfig):
-    """Base configuration for HunYuan pipeline architecture."""
-
-    # HunyuanConfig-specific parameters with defaults
-    flow_shift: int = 7
-    flow_shift_sr: int = 2
-    upsampler_config: tuple[UpsamplerConfig, ...] = field(
-        default_factory=lambda:
-        (SRTo720pUpsamplerConfig(), SRTo1080pUpsamplerConfig()))
-    upsampler_precision: str = "fp32"
@@ -0,0 +1,215 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Registry for pipeline weight-specific configurations."""
+
+import os
+from collections.abc import Callable
+
+from fastvideo.configs.pipelines.base import PipelineConfig
+from fastvideo.configs.pipelines.cosmos import CosmosConfig
+from fastvideo.configs.pipelines.cosmos2_5 import Cosmos25Config
+from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
+from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig
+from fastvideo.configs.pipelines.hyworld import HYWorldConfig
+from fastvideo.configs.pipelines.ltx2 import LTX2T2VConfig
+from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
+from fastvideo.configs.pipelines.longcat import LongCatT2V480PConfig
+from fastvideo.configs.pipelines.turbodiffusion import (
+    TurboDiffusionT2V_1_3B_Config, TurboDiffusionT2V_14B_Config,
+    TurboDiffusionI2V_A14B_Config)
+
+# isort: off
+from fastvideo.configs.pipelines.wan import (
+    FastWan2_1_T2V_480P_Config, FastWan2_2_TI2V_5B_Config,
+    Wan2_2_I2V_A14B_Config, Wan2_2_T2V_A14B_Config, Wan2_2_TI2V_5B_Config,
+    WanI2V480PConfig, WanI2V720PConfig, WanT2V480PConfig, WanT2V720PConfig,
+    SelfForcingWanT2V480PConfig, WANV2VConfig, SelfForcingWan2_2_T2V480PConfig,
+    MatrixGameI2V480PConfig)
+# isort: on
+from fastvideo.logger import init_logger
+from fastvideo.utils import (maybe_download_model_index,
+                             verify_model_config_and_directory)
+
+logger = init_logger(__name__)
+
+# Registry maps specific model weights to their config classes
+PIPE_NAME_TO_CONFIG: dict[str, type[PipelineConfig]] = {
+    "FastVideo/FastHunyuan-diffusers": FastHunyuanConfig,
+    "hunyuanvideo-community/HunyuanVideo": HunyuanConfig,
+    "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v":
+    Hunyuan15T2V480PConfig,
+    "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v":
+    Hunyuan15T2V720PConfig,
+    "FastVideo/HY-WorldPlay-Bidirectional-Diffusers": HYWorldConfig,
+    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers": WanT2V480PConfig,
+    "weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers": WanI2V480PConfig,
+    "weizhou03/Wan2.1-Game-Fun-1.3B-InP-Diffusers": WanI2V480PConfig,
+    "IRMChen/Wan2.1-Fun-1.3B-Control-Diffusers": WANV2VConfig,
+    "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers": WanI2V480PConfig,
+    "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers": WanI2V720PConfig,
+    "Wan-AI/Wan2.1-T2V-14B-Diffusers": WanT2V720PConfig,
+    "FastVideo/FastWan2.1-T2V-1.3B-Diffusers": FastWan2_1_T2V_480P_Config,
+    "FastVideo/FastWan2.1-T2V-14B-480P-Diffusers": FastWan2_1_T2V_480P_Config,
+    "FastVideo/FastWan2.2-TI2V-5B-Diffusers": FastWan2_2_TI2V_5B_Config,
+    "FastVideo/stepvideo-t2v-diffusers": StepVideoT2VConfig,
+    "FastVideo/Wan2.1-VSA-T2V-14B-720P-Diffusers": WanT2V720PConfig,
+    "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers": SelfForcingWanT2V480PConfig,
+    "rand0nmr/SFWan2.2-T2V-A14B-Diffusers": SelfForcingWan2_2_T2V480PConfig,
+    "FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers":
+    SelfForcingWan2_2_T2V480PConfig,
+    "Wan-AI/Wan2.2-TI2V-5B-Diffusers": Wan2_2_TI2V_5B_Config,
+    "Wan-AI/Wan2.2-T2V-A14B-Diffusers": Wan2_2_T2V_A14B_Config,
+    "Wan-AI/Wan2.2-I2V-A14B-Diffusers": Wan2_2_I2V_A14B_Config,
+    "nvidia/Cosmos-Predict2-2B-Video2World": CosmosConfig,
+    "KyleShao/Cosmos-Predict2.5-2B-Diffusers": Cosmos25Config,
+    "FastVideo/Matrix-Game-2.0-Base-Diffusers": MatrixGameI2V480PConfig,
+    "FastVideo/Matrix-Game-2.0-GTA-Diffusers": MatrixGameI2V480PConfig,
+    "FastVideo/Matrix-Game-2.0-TempleRun-Diffusers": MatrixGameI2V480PConfig,
+    # LongCat Video models
+    "FastVideo/LongCat-Video-T2V-Diffusers": LongCatT2V480PConfig,
+    "FastVideo/LongCat-Video-I2V-Diffusers": LongCatT2V480PConfig,
+    "FastVideo/LongCat-Video-VC-Diffusers": LongCatT2V480PConfig,
+    # LTX-2 models
+    "Lightricks/LTX-2": LTX2T2VConfig,
+    "converted/ltx2_diffusers": LTX2T2VConfig,
+    # TurboDiffusion models
+    "loayrashid/TurboWan2.1-T2V-1.3B-Diffusers": TurboDiffusionT2V_1_3B_Config,
+    "loayrashid/TurboWan2.1-T2V-14B-Diffusers": TurboDiffusionT2V_14B_Config,
+    "loayrashid/TurboWan2.2-I2V-A14B-Diffusers": TurboDiffusionI2V_A14B_Config,
+    # Add other specific weight variants
+}
+
+# For determining pipeline type from model ID
+PIPELINE_DETECTOR: dict[str, Callable[[str], bool]] = {
+    "longcatimagetovideo":
+    lambda id: "longcatimagetovideo" in id.lower(),
+    "longcatvideocontinuation":
+    lambda id: "longcatvideocontinuation" in id.lower(),
+    "longcat":
+    lambda id: "longcat" in id.lower(),
+    "hunyuan":
+    lambda id: "hunyuan" in id.lower(),
+    "hunyuan15":
+    lambda id: "hunyuan15" in id.lower(),
+    "hyworld":
+    lambda id: "hyworld" in id.lower(),
+    "matrixgame":
+    lambda id: "matrix-game" in id.lower() or "matrixgame" in id.lower(),
+    "wanpipeline":
+    lambda id: "wanpipeline" in id.lower(),
+    "wanimagetovideo":
+    lambda id: "wanimagetovideo" in id.lower(),
+    "wandmdpipeline":
+    lambda id: "wandmdpipeline" in id.lower(),
+    "wancausaldmdpipeline":
+    lambda id: "wancausaldmdpipeline" in id.lower(),
+    "stepvideo":
+    lambda id: "stepvideo" in id.lower(),
+    "cosmos":
+    lambda id: "cosmos" in id.lower() and ("2.5" not in id.lower(
+    ) and "2_5" not in id.lower() and "25" not in id.lower()),
+    "cosmos25":
+    lambda id: "cosmos25" in id.lower(),
+    "turbodiffusion":
+    lambda id: "turbodiffusion" in id.lower() or "turbowan" in id.lower(),
+    "ltx2":
+    lambda id: "ltx2" in id.lower() or "ltx-2" in id.lower(),
+    # Add other pipeline architecture detectors
+}
+
+# Fallback configs when exact match isn't found but architecture is detected
+PIPELINE_FALLBACK_CONFIG: dict[str, type[PipelineConfig]] = {
+    "longcatimagetovideo": LongCatT2V480PConfig,
+    "longcatvideocontinuation": LongCatT2V480PConfig,
+    "longcat": LongCatT2V480PConfig,
+    "cosmos25": Cosmos25Config,
+    "hunyuan":
+    HunyuanConfig,  # Base Hunyuan config as fallback for any Hunyuan variant
+    "matrixgame": MatrixGameI2V480PConfig,
+    "hunyuan15":
+    Hunyuan15T2V480PConfig,  # Base Hunyuan15 config as fallback for any Hunyuan15 variant
+    "hyworld":
+    HYWorldConfig,  # HYWorld-specific config as fallback for any HYWorld variant
+    "wanpipeline":
+    WanT2V480PConfig,  # Base Wan config as fallback for any Wan variant
+    "wanimagetovideo": WanI2V480PConfig,
+    "wandmdpipeline": FastWan2_1_T2V_480P_Config,
+    "wancausaldmdpipeline": SelfForcingWanT2V480PConfig,
+    "stepvideo": StepVideoT2VConfig,
+    "turbodiffusion": TurboDiffusionT2V_1_3B_Config,
+    "ltx2": LTX2T2VConfig,
+    # Other fallbacks by architecture
+}
+
+
+def get_pipeline_config_cls_from_name(
+        pipeline_name_or_path: str) -> type[PipelineConfig]:
+    """Get the appropriate configuration class for a given pipeline name or path.
+
+    This function implements a multi-step lookup process to find the most suitable
+    configuration class for a given pipeline. It follows this order:
+    1. Exact match in the PIPE_NAME_TO_CONFIG
+    2. Partial match in the PIPE_NAME_TO_CONFIG
+    3. Fallback to class name in the model_index.json
+    4. else raise an error
+
+    Args:
+        pipeline_name_or_path (str): The name or path of the pipeline. This can be:
+            - A registered model ID (e.g., "FastVideo/FastHunyuan-diffusers")
+            - A local path to a model directory
+            - A model ID that will be downloaded
+
+    Returns:
+        Type[PipelineConfig]: The configuration class that best matches the pipeline.
+            This will be one of:
+            - A specific weight configuration class if an exact match is found
+            - A fallback configuration class based on the pipeline architecture
+            - The base PipelineConfig class if no matches are found
+
+    Note:
+        - For local paths, the function will verify the model configuration
+        - For remote models, it will attempt to download the model index
+        - Warning messages are logged when falling back to less specific configurations
+    """
+
+    pipeline_config_cls: type[PipelineConfig] | None = None
+
+    # First try exact match for specific weights
+    if pipeline_name_or_path in PIPE_NAME_TO_CONFIG:
+        pipeline_config_cls = PIPE_NAME_TO_CONFIG[pipeline_name_or_path]
+        return pipeline_config_cls
+
+    # Try partial matches (for local paths that might include the weight ID)
+    for registered_id, config_class in PIPE_NAME_TO_CONFIG.items():
+        if registered_id in pipeline_name_or_path:
+            pipeline_config_cls = config_class
+            break
+
+    # If no match, try to use the fallback config
+    if pipeline_config_cls is None:
+        if os.path.exists(pipeline_name_or_path):
+            config = verify_model_config_and_directory(pipeline_name_or_path)
+        else:
+            config = maybe_download_model_index(pipeline_name_or_path)
+        logger.warning(
+            "Trying to use the config from the model_index.json. FastVideo may not correctly identify the optimal config for this model in this situation."
+        )
+
+        pipeline_name = config["_class_name"]
+        # Try to determine pipeline architecture for fallback
+        for pipeline_type, detector in PIPELINE_DETECTOR.items():
+            if detector(pipeline_name.lower()):
+                pipeline_config_cls = PIPELINE_FALLBACK_CONFIG.get(
+                    pipeline_type)
+                break
+
+        if pipeline_config_cls is not None:
+            logger.warning(
+                "No match found for pipeline %s, using fallback config %s.",
+                pipeline_name_or_path, pipeline_config_cls)
+
+    if pipeline_config_cls is None:
+        raise ValueError(
+            f"No match found for pipeline {pipeline_name_or_path}, please check the pipeline name or path."
+        )
+
+    return pipeline_config_cls
@@ -28,9 +28,6 @@ class SamplingParam:
    keyboard_cond: Any | None = None  # Shape: (B, T, K)
    grid_sizes: Any | None = None  # Shape: (3,) [F,H,W]

-    # Camera control inputs (HYWorld)
-    pose: str | None = None  # Camera trajectory: pose string (e.g., 'w-31') or JSON file path
-
    # Refine inputs (LongCat 480p->720p upscaling)
    # Path-based refine (load stage1 video from disk, e.g. MP4)
    refine_from: str | None = None  # Path to stage1 video (480p output from distill)
@@ -58,13 +55,10 @@ class SamplingParam:
    num_frames_round_down: bool = False  # Whether to round down num_frames if it's not divisible by num_gpus
    height: int = 720
    width: int = 1280
-    height_sr: int = 1072
-    width_sr: int = 1920
    fps: int = 24

    # Denoising parameters
    num_inference_steps: int = 50
-    num_inference_steps_sr: int = 50
    guidance_scale: float = 1.0
    guidance_rescale: float = 0.0
    boundary_ratio: float | None = None
@@ -98,7 +92,8 @@ class SamplingParam:

    @classmethod
    def from_pretrained(cls, model_path: str) -> "SamplingParam":
-        from fastvideo.registry import get_sampling_param_cls_for_name
+        from fastvideo.configs.sample.registry import (
+            get_sampling_param_cls_for_name)
        sampling_cls = get_sampling_param_cls_for_name(model_path)
        if sampling_cls is not None:
            sampling_param: SamplingParam = sampling_cls()
@@ -5,19 +5,15 @@ from fastvideo.configs.sample.base import SamplingParam


@dataclass
-class Cosmos25SamplingParamBase(SamplingParam):
-    height: int = 704
-    width: int = 1280
-    num_frames: int = 77
+class Cosmos_Predict2_5_2B_Diffusers_SamplingParam(SamplingParam):
+    """Defaults for Cosmos 2.5 (Predict2.5) text-to-video diffusers-format model."""
+
+    height: int = 480
+    width: int = 832
+    num_frames: int = 121
    fps: int = 24
-    seed: int = 0

    guidance_scale: float = 7.0
-    negative_prompt: str = (
-        "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, "
-        "over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, "
-        "underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, "
-        "low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, "
-        "unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. "
-        "Overall, the video is of poor quality.")
+    # Official Cosmos2.5 sampling uses empty string as unconditional.
+    negative_prompt: str = ""
    num_inference_steps: int = 35
@@ -28,33 +28,7 @@ class Hunyuan15_480P_SamplingParam(SamplingParam):
            np.linspace(1.0, 0.0, self.num_inference_steps + 1)[:-1])


-@dataclass
-class Hunyuan15_480P_StepDistilled_I2V_SamplingParam(
-        Hunyuan15_480P_SamplingParam):
-    num_inference_steps: int = 12
-
-    height: int = 720
-    width: int = 1280
-    guidance_scale: float = 1.0
-
-
@dataclass
 class Hunyuan15_720P_SamplingParam(Hunyuan15_480P_SamplingParam):
    height: int = 720
    width: int = 1280
-
-
-@dataclass
-class Hunyuan15_720P_Distilled_I2V_SamplingParam(Hunyuan15_720P_SamplingParam):
-    guidance_scale: float = 1.0
-
-
-@dataclass
-class Hunyuan15_SR_1080P_SamplingParam(Hunyuan15_480P_SamplingParam):
-    height_sr: int = 1072
-    width_sr: int = 1920
-
-    num_inference_steps: int = 12
-    num_inference_steps_sr: int = 8
-
-    guidance_scale: float = 1.0
@@ -0,0 +1,206 @@
+# SPDX-License-Identifier: Apache-2.0
+import os
+from collections.abc import Callable
+from typing import Any
+
+from fastvideo.configs.sample.hunyuan import (FastHunyuanSamplingParam,
+                                              HunyuanSamplingParam)
+from fastvideo.configs.sample.hunyuan15 import Hunyuan15_480P_SamplingParam, Hunyuan15_720P_SamplingParam
+from fastvideo.configs.sample.hyworld import HYWorld_SamplingParam
+from fastvideo.configs.sample.stepvideo import StepVideoT2VSamplingParam
+
+from fastvideo.configs.sample.cosmos import Cosmos_Predict2_2B_Video2World_SamplingParam
+from fastvideo.configs.sample.cosmos2_5 import Cosmos_Predict2_5_2B_Diffusers_SamplingParam
+from fastvideo.configs.sample.ltx2 import LTX2SamplingParam
+
+# isort: off
+from fastvideo.configs.sample.wan import (
+    FastWanT2V480P_SamplingParam,
+    Wan2_1_Fun_1_3B_InP_SamplingParam,
+    Wan2_2_I2V_A14B_SamplingParam,
+    Wan2_2_T2V_A14B_SamplingParam,
+    Wan2_2_TI2V_5B_SamplingParam,
+    WanI2V_14B_480P_SamplingParam,
+    WanI2V_14B_720P_SamplingParam,
+    WanT2V_1_3B_SamplingParam,
+    WanT2V_14B_SamplingParam,
+    Wan2_1_Fun_1_3B_Control_SamplingParam,
+    SelfForcingWan2_1_T2V_1_3B_480P_SamplingParam,
+    SelfForcingWan2_2_T2V_A14B_480P_SamplingParam,
+    MatrixGame2_SamplingParam,
+)
+from fastvideo.configs.sample.turbodiffusion import (
+    TurboDiffusionT2V_1_3B_SamplingParam,
+    TurboDiffusionT2V_14B_SamplingParam,
+    TurboDiffusionI2V_A14B_SamplingParam,
+)
+# isort: on
+from fastvideo.logger import init_logger
+from fastvideo.utils import (maybe_download_model_index,
+                             verify_model_config_and_directory)
+
+logger = init_logger(__name__)
+# Registry maps specific model weights to their config classes
+SAMPLING_PARAM_REGISTRY: dict[str, Any] = {
+    "FastVideo/FastHunyuan-diffusers": FastHunyuanSamplingParam,
+    "hunyuanvideo-community/HunyuanVideo": HunyuanSamplingParam,
+    "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v":
+    Hunyuan15_480P_SamplingParam,
+    "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v":
+    Hunyuan15_720P_SamplingParam,
+    "FastVideo/HY-WorldPlay-Bidirectional-Diffusers": HYWorld_SamplingParam,
+    "FastVideo/stepvideo-t2v-diffusers": StepVideoT2VSamplingParam,
+
+    # Wan2.1
+    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers": WanT2V_1_3B_SamplingParam,
+    "Wan-AI/Wan2.1-T2V-14B-Diffusers": WanT2V_14B_SamplingParam,
+    "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers": WanI2V_14B_480P_SamplingParam,
+    "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers": WanI2V_14B_720P_SamplingParam,
+    "weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers":
+    Wan2_1_Fun_1_3B_InP_SamplingParam,
+    "weizhou03/Wan2.1-Game-Fun-1.3B-InP-Diffusers":
+    Wan2_1_Fun_1_3B_InP_SamplingParam,
+    "IRMChen/Wan2.1-Fun-1.3B-Control-Diffusers":
+    Wan2_1_Fun_1_3B_Control_SamplingParam,
+
+    # Wan2.2
+    "Wan-AI/Wan2.2-TI2V-5B-Diffusers": Wan2_2_TI2V_5B_SamplingParam,
+    "FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers":
+    Wan2_2_TI2V_5B_SamplingParam,
+    "Wan-AI/Wan2.2-T2V-A14B-Diffusers": Wan2_2_T2V_A14B_SamplingParam,
+    "Wan-AI/Wan2.2-I2V-A14B-Diffusers": Wan2_2_I2V_A14B_SamplingParam,
+
+    # FastWan2.1
+    "FastVideo/FastWan2.1-T2V-1.3B-Diffusers": FastWanT2V480P_SamplingParam,
+
+    # FastWan2.2
+    "FastVideo/FastWan2.2-TI2V-5B-Diffusers": Wan2_2_TI2V_5B_SamplingParam,
+
+    # Causal Self-Forcing Wan2.1
+    "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers":
+    SelfForcingWan2_1_T2V_1_3B_480P_SamplingParam,
+
+    # Causal Self-Forcing Wan2.2
+    "rand0nmr/SFWan2.2-T2V-A14B-Diffusers":
+    SelfForcingWan2_2_T2V_A14B_480P_SamplingParam,
+    "FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers":
+    SelfForcingWan2_2_T2V_A14B_480P_SamplingParam,
+
+    # Cosmos2
+    "nvidia/Cosmos-Predict2-2B-Video2World":
+    Cosmos_Predict2_2B_Video2World_SamplingParam,
+
+    # Cosmos2.5
+    "KyleShao/Cosmos-Predict2.5-2B-Diffusers":
+    Cosmos_Predict2_5_2B_Diffusers_SamplingParam,
+
+    # MatrixGame2.0 models
+    "FastVideo/Matrix-Game-2.0-Base-Diffusers": MatrixGame2_SamplingParam,
+    "FastVideo/Matrix-Game-2.0-GTA-Diffusers": MatrixGame2_SamplingParam,
+    "FastVideo/Matrix-Game-2.0-TempleRun-Diffusers": MatrixGame2_SamplingParam,
+
+    # TurboDiffusion models
+    "loayrashid/TurboWan2.1-T2V-1.3B-Diffusers":
+    TurboDiffusionT2V_1_3B_SamplingParam,
+    "loayrashid/TurboWan2.1-T2V-14B-Diffusers":
+    TurboDiffusionT2V_14B_SamplingParam,
+    "loayrashid/TurboWan2.2-I2V-A14B-Diffusers":
+    TurboDiffusionI2V_A14B_SamplingParam,
+
+    # LTX-2 models
+    "Lightricks/LTX-2": LTX2SamplingParam,
+    "FastVideo/LTX2-Distilled-Diffusers": LTX2SamplingParam,
+
+    # Add other specific weight variants
+}
+
+# For determining pipeline type from model ID
+SAMPLING_PARAM_DETECTOR: dict[str, Callable[[str], bool]] = {
+    "hunyuan":
+    lambda id: "hunyuan" in id.lower(),
+    "hunyuan15":
+    lambda id: "hunyuan15" in id.lower(),
+    "hyworld":
+    lambda id: "hyworld" in id.lower(),
+    "wanpipeline":
+    lambda id: "wanpipeline" in id.lower(),
+    "wanimagetovideo":
+    lambda id: "wanimagetovideo" in id.lower(),
+    "stepvideo":
+    lambda id: "stepvideo" in id.lower(),
+    "wandmdpipeline":
+    lambda id: "wandmdpipeline" in id.lower(),
+    "wancausaldmdpipeline":
+    lambda id: "wancausaldmdpipeline" in id.lower(),
+    "matrixgame":
+    lambda id: "matrixgame" in id.lower() or "matrix-game" in id.lower(),
+    "turbodiffusion":
+    lambda id: "turbodiffusion" in id.lower() or "turbowan" in id.lower(),
+    "cosmos25":
+    lambda id: "cosmos2_5" in id.lower(),
+    "cosmos":
+    lambda id: "cosmos" in id.lower() and "2_5" not in id.lower(),
+    "ltx2":
+    lambda id: "ltx2" in id.lower() or "ltx-2" in id.lower(),
+    # Add other pipeline architecture detectors
+}
+
+# Fallback configs when exact match isn't found but architecture is detected
+SAMPLING_FALLBACK_PARAM: dict[str, Any] = {
+    "hunyuan":
+    HunyuanSamplingParam,  # Base Hunyuan config as fallback for any Hunyuan variant
+    "hunyuan15":
+    Hunyuan15_480P_SamplingParam,  # Base Hunyuan15 config as fallback for any Hunyuan15 variant
+    "hyworld":
+    HYWorld_SamplingParam,  # HYWorld-specific config as fallback for any HYWorld variant
+    "wanpipeline":
+    WanT2V_1_3B_SamplingParam,  # Base Wan config as fallback for any Wan variant
+    "wanimagetovideo": WanI2V_14B_480P_SamplingParam,
+    "wandmdpipeline": FastWanT2V480P_SamplingParam,
+    "wancausaldmdpipeline": SelfForcingWan2_1_T2V_1_3B_480P_SamplingParam,
+    "stepvideo": StepVideoT2VSamplingParam,
+    "matrixgame": MatrixGame2_SamplingParam,
+    "turbodiffusion":
+    TurboDiffusionT2V_1_3B_SamplingParam,  # Default to T2V for fallback
+    "cosmos25": Cosmos_Predict2_5_2B_Diffusers_SamplingParam,
+    "cosmos": Cosmos_Predict2_2B_Video2World_SamplingParam,
+    "ltx2": LTX2SamplingParam,
+    # Other fallbacks by architecture
+}
+
+
+def get_sampling_param_cls_for_name(pipeline_name_or_path: str) -> Any | None:
+    """Get the appropriate sampling param for specific pretrained weights."""
+    # First try exact match for specific weights
+    if pipeline_name_or_path in SAMPLING_PARAM_REGISTRY:
+        return SAMPLING_PARAM_REGISTRY[pipeline_name_or_path]
+
+    # Try partial matches (for local paths that might include the weight ID)
+    for registered_id, config_class in SAMPLING_PARAM_REGISTRY.items():
+        if registered_id in pipeline_name_or_path:
+            return config_class
+
+    matrixgame_patterns = ["Matrix-Game", "Skywork--Matrix-Game", "matrixgame"]
+    for pattern in matrixgame_patterns:
+        if pattern.lower() in pipeline_name_or_path.lower():
+            return MatrixGame2_SamplingParam
+
+    if os.path.exists(pipeline_name_or_path):
+        config = verify_model_config_and_directory(pipeline_name_or_path)
+    else:
+        config = maybe_download_model_index(pipeline_name_or_path)
+
+    pipeline_name = config["_class_name"]
+
+    # If no match, try to use the fallback config
+    fallback_config = None
+    # Try to determine pipeline architecture for fallback
+    for pipeline_type, detector in SAMPLING_PARAM_DETECTOR.items():
+        if detector(pipeline_name.lower()):
+            fallback_config = SAMPLING_FALLBACK_PARAM.get(pipeline_type)
+            break
+
+    logger.warning(
+        "No match found for pipeline %s, using fallback sampling param %s.",
+        pipeline_name_or_path, fallback_config)
+    return fallback_config
@@ -113,13 +113,13 @@ class FastWanT2V480P_SamplingParam(WanT2V_1_3B_SamplingParam):
@dataclass
 class Wan2_1_Fun_1_3B_InP_SamplingParam(SamplingParam):
    """Sampling parameters for Wan2.1 Fun 1.3B InP model."""
-    height: int = 480
-    width: int = 832
-    num_frames: int = 81
-    fps: int = 16
+    height: int = 352
+    width: int = 640
+    num_frames: int = 77
+    fps: int = 25
    negative_prompt: str | None = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
-    guidance_scale: float = 6.0
-    num_inference_steps: int = 50
+    guidance_scale: float = 1.0
+    num_inference_steps: int = 40


@dataclass
@@ -157,3 +157,43 @@ pyarrow_schema_matrixgame = pa.schema([
    pa.field("duration_sec", pa.float64()),
    pa.field("fps", pa.float64()),
 ])
+
+pyarrow_schema_wangame = pa.schema([
+    pa.field("id", pa.string()),
+    # --- Image/Video VAE latents ---
+    # Tensors are stored as raw bytes with shape and dtype info for loading
+    pa.field("vae_latent_bytes", pa.binary()),
+    # e.g., [C, T, H, W] or [C, H, W]
+    pa.field("vae_latent_shape", pa.list_(pa.int64())),
+    # e.g., 'float32'
+    pa.field("vae_latent_dtype", pa.string()),
+    #I2V
+    pa.field("clip_feature_bytes", pa.binary()),
+    pa.field("clip_feature_shape", pa.list_(pa.int64())),
+    pa.field("clip_feature_dtype", pa.string()),
+    pa.field("first_frame_latent_bytes", pa.binary()),
+    pa.field("first_frame_latent_shape", pa.list_(pa.int64())),
+    pa.field("first_frame_latent_dtype", pa.string()),
+    # --- Action ---
+    pa.field("mouse_cond_bytes", pa.binary()),
+    pa.field("mouse_cond_shape", pa.list_(pa.int64())),  # [T, 2]
+    pa.field("mouse_cond_dtype", pa.string()),
+    pa.field("keyboard_cond_bytes", pa.binary()),
+    pa.field("keyboard_cond_shape", pa.list_(pa.int64())),  # [T, 4]
+    pa.field("keyboard_cond_dtype", pa.string()),
+    # I2V Validation
+    pa.field("pil_image_bytes", pa.binary()),
+    pa.field("pil_image_shape", pa.list_(pa.int64())),
+    pa.field("pil_image_dtype", pa.string()),
+    # --- Metadata ---
+    pa.field("file_name", pa.string()),
+    pa.field("caption", pa.string()),
+    pa.field("media_type", pa.string()),  # 'image' or 'video'
+    pa.field("width", pa.int64()),
+    pa.field("height", pa.int64()),
+    # -- Video-specific (can be null/default for images) ---
+    # Number of frames processed (e.g., 1 for image, N for video)
+    pa.field("num_frames", pa.int64()),
+    pa.field("duration_sec", pa.float64()),
+    pa.field("fps", pa.float64()),
+])
@@ -148,6 +148,62 @@ class DataValidationStage(DatasetFilterStage):
        return batch


+class ResolutionFilterStage(DatasetFilterStage):
+    """Stage for filtering data items based on resolution constraints."""
+
+    def __init__(self,
+                 max_h_div_w_ratio: float = 17 / 16,
+                 min_h_div_w_ratio: float = 8 / 16,
+                 max_height: int = 1024,
+                 max_width: int = 1024):
+        self.max_h_div_w_ratio = max_h_div_w_ratio
+        self.min_h_div_w_ratio = min_h_div_w_ratio
+        self.max_height = max_height
+        self.max_width = max_width
+
+    def should_keep(self, batch: PreprocessBatch, **kwargs) -> bool:
+        """
+        Check if data item passes resolution filtering.
+        
+        Args:
+            batch: Dataset batch with resolution information
+            
+        Returns:
+            True if passes filter, False otherwise
+        """
+        # Only apply to videos
+        if not batch.is_video:
+            return True
+
+        if batch.resolution is None:
+            return False
+
+        height = batch.resolution.get("height", None)
+        width = batch.resolution.get("width", None)
+        if height is None or width is None:
+            return False
+
+        # Check aspect ratio
+        aspect = self.max_height / self.max_width
+        hw_aspect_thr = 1.5
+
+        return self.filter_resolution(
+            height,
+            width,
+            max_h_div_w_ratio=hw_aspect_thr * aspect,
+            min_h_div_w_ratio=1 / hw_aspect_thr * aspect,
+        )
+
+    def process(self, batch: PreprocessBatch, **kwargs) -> PreprocessBatch:
+        """Process does nothing for resolution filtering - filtering is handled by should_keep."""
+        return batch
+
+    def filter_resolution(self, h: int, w: int, max_h_div_w_ratio: float,
+                          min_h_div_w_ratio: float) -> bool:
+        """Filter based on height/width ratio."""
+        return h / w <= max_h_div_w_ratio and h / w >= min_h_div_w_ratio
+
+
 class FrameSamplingStage(DatasetFilterStage):
    """Stage for temporal frame sampling and indexing."""

@@ -272,6 +328,11 @@ class VideoTransformStage(DatasetStage):
        video = rearrange(video, "t c h w -> c t h w")
        video = video.to(torch.uint8)

+        h, w = video.shape[-2:]
+        assert (
+            h / w <= 17 / 16 and h / w >= 8 / 16
+        ), f"Only videos with a ratio (h/w) less than 17/16 and more than 8/16 are supported. But video ({batch.path}) found ratio is {round(h / w, 2)} with the shape of {video.shape}"
+
        video = video.float() / 127.5 - 1.0
        batch.pixel_values = video
        return batch
@@ -428,6 +489,8 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
                     tokenizer) -> None:
        """Initialize all processing stages."""
        self.validation_stage = DataValidationStage()
+        self.resolution_filter_stage = ResolutionFilterStage(
+            max_height=args.max_height, max_width=args.max_width)
        self.frame_sampling_stage = FrameSamplingStage(
            num_frames=args.num_frames,
            train_fps=args.train_fps,
@@ -480,6 +543,7 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
        # Initialize counters
        filter_counts = {
            "validation_failed": 0,
+            "resolution_failed": 0,
            "frame_sampling_failed": 0
        }
        sample_num_frames: list[int] = []
@@ -515,6 +579,10 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
            filter_counts["validation_failed"] += 1
            return False

+        if not self.resolution_filter_stage.should_keep(batch):
+            filter_counts["resolution_failed"] += 1
+            return False
+
        if not self.frame_sampling_stage.should_keep(batch):
            filter_counts["frame_sampling_failed"] += 1
            return False
@@ -526,9 +594,10 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
                             after_count: int):
        """Log filtering statistics."""
        logger.info(
-            "validation_failed: %d, frame_sampling_failed: %d, "
+            "validation_failed: %d, resolution_failed: %d, frame_sampling_failed: %d, "
            "Counter(sample_num_frames): %s, before filter: %d, after filter: %d",
            filter_counts['validation_failed'],
+            filter_counts['resolution_failed'],
            filter_counts['frame_sampling_failed'], Counter(sample_num_frames),
            before_count, after_count)

@@ -4,6 +4,7 @@ import os
 import pathlib

 import datasets
+import numpy as np
 from torch.utils.data import IterableDataset

 from fastvideo.distributed import (get_sp_world_size, get_world_rank,
@@ -160,5 +161,25 @@ class ValidationDataset(IterableDataset):
                else:
                    sample["control_video"] = load_video(control_video_path)

+            if sample.get("action_path", None) is not None:
+                action_path = sample["action_path"]
+                action_path = os.path.join(self.dir, action_path)
+                sample["action_path"] = action_path
+                if not pathlib.Path(action_path).is_file():
+                    logger.warning("Action file %s does not exist.", action_path)
+                else:
+                    try:
+                        action_data = np.load(action_path, allow_pickle=True)
+                        num_frames = sample["num_frames"]
+                        if action_data.dtype == object: action_data = action_data.item()
+                        if isinstance(action_data, dict):
+                            sample["keyboard_cond"] = action_data["keyboard"][:num_frames]
+                            sample["mouse_cond"] = action_data["mouse"][:num_frames]
+                        else:
+                            sample["keyboard_cond"] = action_data[:num_frames]
+                    except Exception as e:
+                        logger.error("Error loading action file %s: %s",
+                                        action_path, e)
+
            sample = {k: v for k, v in sample.items() if v is not None}
            yield sample
@@ -18,7 +18,6 @@ __all__ = [
    "cleanup_dist_env_and_memory",
    "model_parallel_is_initialized",
    "maybe_init_distributed_environment_and_model_parallel",
-    "warmup_sequence_parallel_communication",

    # World group
    "get_world_group",
@@ -7,17 +7,10 @@ import torch.distributed
 from fastvideo.distributed.parallel_state import (get_sp_group,
                                                  get_sp_parallel_rank,
                                                  get_sp_world_size,
-                                                  get_tp_group,
-                                                  model_parallel_is_initialized)
+                                                  get_tp_group)
 from fastvideo.distributed.utils import (unpad_sequence_tensor,
                                         compute_padding_for_sp,
                                         pad_sequence_tensor)
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
-
-# Track if SP communication has been warmed up
-_sp_warmup_done = False


 def tensor_model_parallel_all_reduce(input_: torch.Tensor) -> torch.Tensor:
@@ -108,86 +101,3 @@ def sequence_model_parallel_shard(input_: torch.Tensor,
    input_ = input_.movedim(0, dim)

    return input_, original_seq_len
-
-
-def warmup_sequence_parallel_communication(
-        device: torch.device | None = None) -> None:
-    """Warmup NCCL communicators for sequence parallel all-to-all operations.
-    
-    The first NCCL collective operation is slow due to lazy communicator
-    initialization. This function runs dummy all-to-all operations to
-    trigger the initialization upfront, before the first real forward pass.
-    
-    Args:
-        device: Device to use for warmup tensors. If None, uses CUDA device 0.
-    """
-    global _sp_warmup_done
-
-    if _sp_warmup_done:
-        return
-
-    if not model_parallel_is_initialized():
-        return
-
-    sp_world_size = get_sp_world_size()
-    if sp_world_size <= 1:
-        _sp_warmup_done = True
-        return
-
-    if device is None:
-        device = torch.device("cuda")
-
-    logger.info("Warming up sequence parallel communication (SP=%d)...",
-                sp_world_size)
-
-    # Use small but representative tensor shapes for warmup
-    # Shape: [batch, seq_len, num_heads, head_dim]
-    # The all-to-all patterns used in attention:
-    #   1. scatter_dim=2 (heads), gather_dim=1 (seq) - before attention
-    #   2. scatter_dim=1 (seq), gather_dim=2 (heads) - after attention
-    batch_size = 1
-    seq_len_per_rank = 16  # Small sequence per rank
-    num_heads = sp_world_size * 4  # Must be divisible by sp_world_size
-    head_dim = 64
-
-    # Create dummy tensor for warmup
-    dummy = torch.zeros(batch_size,
-                        seq_len_per_rank,
-                        num_heads,
-                        head_dim,
-                        device=device,
-                        dtype=torch.bfloat16)
-
-    # Warmup pattern 1: scatter heads, gather sequence (before attention)
-    _ = sequence_model_parallel_all_to_all_4D(dummy,
-                                              scatter_dim=2,
-                                              gather_dim=1)
-
-    # Warmup pattern 2: scatter sequence, gather heads (after attention)
-    dummy2 = torch.zeros(batch_size,
-                         seq_len_per_rank * sp_world_size,
-                         num_heads // sp_world_size,
-                         head_dim,
-                         device=device,
-                         dtype=torch.bfloat16)
-    _ = sequence_model_parallel_all_to_all_4D(dummy2,
-                                              scatter_dim=1,
-                                              gather_dim=2)
-
-    # Warmup all-gather (used for replicated tokens)
-    dummy3 = torch.zeros(batch_size,
-                         8,
-                         num_heads // sp_world_size,
-                         head_dim,
-                         device=device,
-                         dtype=torch.bfloat16)
-    _ = sequence_model_parallel_all_gather(dummy3, dim=2)
-
-    # Synchronize to ensure warmup completes
-    torch.cuda.synchronize(device)
-
-    # Clean up
-    del dummy, dummy2, dummy3
-
-    _sp_warmup_done = True
-    logger.info("Sequence parallel communication warmup complete.")
@@ -9,7 +9,6 @@ diffusion models.
 import math
 import os
 import re
-import threading
 import time
 from copy import deepcopy
 from typing import Any
@@ -32,21 +31,6 @@ from fastvideo.worker.executor import Executor
 logger = init_logger(__name__)


-def _infer_latent_batch_size(batch: ForwardBatch) -> int:
-    if isinstance(batch.prompt, list):
-        latent_batch_size = len(batch.prompt)
-    elif batch.prompt is not None:
-        latent_batch_size = 1
-    elif batch.prompt_embeds is not None and len(batch.prompt_embeds) > 0:
-        latent_batch_size = batch.prompt_embeds[0].shape[0]
-    else:
-        raise ValueError(
-            "Cannot infer batch size from batch; no prompt or prompt_embeds found"
-        )
-    latent_batch_size *= batch.num_videos_per_prompt
-    return latent_batch_size
-
-
 class VideoGenerator:
    """
    A unified class for generating videos using diffusion models.
@@ -388,31 +372,8 @@ class VideoGenerator:

        # Run inference
        start_time = time.perf_counter()
-
-        # Execute forward pass in a new thread for non-blocking tensor allocation
-        result_container = {}
-
-        def execute_forward_thread():
-            result_container['output_batch'] = self.executor.execute_forward(
-                batch, fastvideo_args)
-
-        thread = threading.Thread(target=execute_forward_thread)
-        thread.start()
-        latent_batch_size = _infer_latent_batch_size(batch)
-        samples = torch.empty((latent_batch_size, 3, sampling_param.num_frames,
-                               sampling_param.height, sampling_param.width),
-                              device='cpu',
-                              pin_memory=fastvideo_args.pin_cpu_memory)
-        thread.join()
-
-        output_batch = result_container['output_batch']
-        if output_batch.output.shape == samples.shape:
-            samples.copy_(output_batch.output)
-        else:
-            logger.warning(
-                "Output shape %s does not match expected shape %s; use slow path",
-                output_batch.output.shape, samples.shape)
-            samples = output_batch.output.cpu()
+        output_batch = self.executor.execute_forward(batch, fastvideo_args)
+        samples = output_batch.output
        logging_info = output_batch.logging_info

        gen_time = time.perf_counter() - start_time
@@ -10,7 +10,7 @@ from enum import Enum
 from typing import Any, TYPE_CHECKING

 from fastvideo.configs.configs import PreprocessConfig
-from fastvideo.configs.pipelines.base import PipelineConfig
+from fastvideo.configs.pipelines.base import PipelineConfig, STA_Mode
 from fastvideo.configs.utils import clean_cli_args
 from fastvideo.layers.quantization import QUANTIZATION_METHODS, QuantizationMethods
 from fastvideo.logger import init_logger
@@ -141,6 +141,8 @@ class FastVideoArgs:

    # STA (Sliding Tile Attention) parameters
    mask_strategy_file_path: str | None = None
+    STA_mode: STA_Mode = STA_Mode.STA_INFERENCE
+    skip_time_steps: int = 15

    # Compilation
    enable_torch_compile: bool = False
@@ -177,7 +179,6 @@ class FastVideoArgs:
    model_loaded: dict[str, bool] = field(default_factory=lambda: {
        "transformer": True,
        "vae": True,
-        "upsampler": True,
    })

    override_text_encoder_safetensors: str | None = None  # path to safetensors file for text encoder override
@@ -463,6 +464,20 @@ class FastVideoArgs:
        )

        # STA (Sliding Tile Attention) parameters
+        parser.add_argument(
+            "--STA-mode",
+            type=str,
+            default=FastVideoArgs.STA_mode.value,
+            choices=[mode.value for mode in STA_Mode],
+            help=
+            "STA mode contains STA_inference, STA_searching, STA_tuning, STA_tuning_cfg, None",
+        )
+        parser.add_argument(
+            "--skip-time-steps",
+            type=int,
+            default=FastVideoArgs.skip_time_steps,
+            help="Number of time steps to warmup (full attention) for STA",
+        )
        parser.add_argument(
            "--mask-strategy-file-path",
            type=str,
@@ -193,46 +193,3 @@ class ImageProcessor:
        tensor = 2.0 * tensor - 1.0

        return tensor
-
-    def _preprocess_cosmos25(
-        self,
-        image: PIL.Image.Image | np.ndarray | torch.Tensor,
-        height: int,
-        width: int,
-    ) -> torch.Tensor:
-        """
-        Cosmos-Predict2.5-style preprocessing for image2world:
-        - aspect-preserving resize (scale so both dims >= target)
-        - center crop to (height, width)
-        - normalize to [-1, 1]
-
-        Returns:
-            torch.Tensor: (1, 3, height, width) in [-1, 1]
-        """
-        if not isinstance(image, PIL.Image.Image):
-            # Reuse the generic preprocess for non-PIL inputs.
-            return self.preprocess(image, height=height, width=width)
-
-        # Ensure RGB (official uses torchvision.to_tensor on PIL, which yields C=3 for RGB inputs)
-        image = image.convert("RGB")
-
-        # Make sure target dims are multiples of VAE scale factor (FastVideo convention)
-        height = height - (height % self.vae_scale_factor)
-        width = width - (width % self.vae_scale_factor)
-
-        orig_w, orig_h = image.size
-        # Scale so that resized image fully covers the target box (like official resize_input()).
-        scale = max(width / float(orig_w), height / float(orig_h))
-        resized_w = int(np.ceil(scale * orig_w))
-        resized_h = int(np.ceil(scale * orig_h))
-
-        image = image.resize((resized_w, resized_h),
-                             resample=PIL.Image.Resampling.BILINEAR)
-
-        # Center crop
-        left = max(0, (resized_w - width) // 2)
-        top = max(0, (resized_h - height) // 2)
-        image = image.crop((left, top, left + width, top + height))
-
-        image_np = np.array(image, dtype=np.float32) / 255.0  # [0,1]
-        return self._normalize_to_tensor(image_np)  # -> [-1,1], (1,3,H,W)
@@ -27,7 +27,7 @@ RESET = '\033[0;0m'
 _warned_local_main_process = False
 _warned_main_process = False

-_FORMAT = (f"{FASTVIDEO_LOGGING_PREFIX}%(levelname)s %(asctime)s.%(msecs)03d "
+_FORMAT = (f"{FASTVIDEO_LOGGING_PREFIX}%(levelname)s %(asctime)s "
           "[%(filename)s:%(lineno)d] %(message)s")
 _DATE_FORMAT = "%m-%d %H:%M:%S"

@@ -102,7 +102,6 @@ def _info(logger: Logger,
        - When both are False, the message will be logged from all processes
        - By default, only logs from processes with LOCAL_RANK=0
    """
-    is_distributed = int(os.environ.get("WORLD_SIZE", 1)) > 1
    try:
        local_rank = int(os.environ["LOCAL_RANK"])
        rank = int(os.environ["RANK"])
@@ -119,22 +118,20 @@ def _info(logger: Logger,

    global _warned_local_main_process, _warned_main_process

-    # Only show process-awareness warnings when actually running distributed
-    if is_distributed:
-        if not _warned_local_main_process and local_main_process_only:
-            logger.warning(
-                '%s By default, logger.info(..) will only log from the local main process. Set logger.info(..., is_local_main_process=False) to log from all processes.%s',
-                GREEN,
-                RESET,
-            )
-            _warned_local_main_process = True
-        if not _warned_main_process and main_process_only:
-            logger.warning(
-                '%s is_main_process_only is set to True, logging only from the main (RANK==0) process.%s',
-                GREEN,
-                RESET,
-            )
-            _warned_main_process = True
+    if not _warned_local_main_process and local_main_process_only:
+        logger.warning(
+            '%s By default, logger.info(..) will only log from the local main process. Set logger.info(..., is_local_main_process=False) to log from all processes.%s',
+            GREEN,
+            RESET,
+        )
+        _warned_local_main_process = True
+    if not _warned_main_process and main_process_only:
+        logger.warning(
+            '%s is_main_process_only is set to True, logging only from the main (RANK==0) process.%s',
+            GREEN,
+            RESET,
+        )
+        _warned_main_process = True

    if not main_process_only and not local_main_process_only:
        logger.log(logging.INFO, msg, *args, stacklevel=2, **kwargs)
@@ -1325,6 +1325,3 @@ def decode_audio(
    decoded_audio = audio_decoder(latent)
    decoded_audio = vocoder(decoded_audio).squeeze(0).float()
    return decoded_audio
-
-# Entry point for model registry
-EntryClass = [LTX2AudioEncoder, LTX2AudioDecoder, LTX2Vocoder]
@@ -676,7 +676,4 @@ class CausalWanTransformer3DModel(BaseDiT):
            # 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
-
-# Entry point for model registry
-EntryClass = CausalWanTransformer3DModel
+        return out
@@ -723,7 +723,4 @@ class CosmosTransformer3DModel(BaseDiT):
        hidden_states = hidden_states.permute(0, 7, 1, 6, 2, 4, 3, 5)
        hidden_states = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

-        return hidden_states
-
-# Entry point for model registry
-EntryClass = CosmosTransformer3DModel
+        return hidden_states
@@ -959,5 +959,3 @@ class Cosmos25Transformer3DModel(BaseDiT):

        return hidden_states

-# Entry point for model registry
-EntryClass = Cosmos25Transformer3DModel
@@ -940,6 +940,3 @@ class FinalLayer(nn.Module):
        x = self.norm_final(x) * (1.0 + scale.unsqueeze(1)) + shift.unsqueeze(1)
        x, _ = self.linear(x)
        return x
-
-# Entry point for model registry
-EntryClass = HunyuanVideoTransformer3DModel
@@ -852,7 +852,4 @@ class FinalLayer(nn.Module):
        scale, shift = self.adaLN_modulation(c).chunk(2, dim=-1)
        x = self.norm_final(x) * (1.0 + scale.unsqueeze(1)) + shift.unsqueeze(1)
        x, _ = self.linear(x)
-        return x
-
-# Entry point for model registry
-EntryClass = HunyuanVideo15Transformer3DModel
+        return x
@@ -22,6 +22,3 @@ __all__ = [
    # Inference utilities (used by examples)
    "get_resolution_from_image",
 ]
-
-# Entry point for model registry
-EntryClass = HYWorldTransformer3DModel
@@ -536,8 +536,8 @@ class HYWorldTransformer3DModel(HunyuanVideo15Transformer3DModel):
                    temb_txt,
                    freqs_cis,
                    seq_attention_mask,
-                    viewmats_seq,
-                    Ks_seq,
+                    viewmats_seq, # hyworld
+                    Ks_seq, # hyworld
                )
        else:
            for block in self.double_blocks:
@@ -549,8 +549,8 @@ class HYWorldTransformer3DModel(HunyuanVideo15Transformer3DModel):
                    temb_txt,
                    freqs_cis,
                    seq_attention_mask,
-                    viewmats_seq,
-                    Ks_seq,
+                    viewmats=viewmats_seq, # hyworld
+                    Ks=Ks_seq, # hyworld
                )


@@ -15,7 +15,7 @@ import torch
 from scipy.spatial.transform import Rotation as R
 from typing import Union, Optional

-from .trajectory import generate_camera_trajectory_local
+from fastvideo.models.dits.hyworld.trajectory import generate_camera_trajectory_local


 # Mapping from one-hot action encoding to single label
@@ -411,3 +411,269 @@ def compute_num_frames(latent_num: int) -> int:
        Number of video frames
    """
    return (latent_num - 1) * 4 + 1
+
+def reformat_keyboard_and_mouse_tensors(keyboard_tensor, mouse_tensor):
+    """
+    Reformat the keyboard and mouse tensors to the format compatible with HyWorld.
+    """
+    num_frames = keyboard_tensor.shape[0]
+    assert (num_frames - 1) % 4 == 0, "num_frames must be a multiple of 4"
+    assert mouse_tensor.shape[0] == num_frames, "mouse_tensor must have the same number of frames as keyboard_tensor"
+    keyboard_tensor = keyboard_tensor[1:, :]
+    mouse_tensor = mouse_tensor[1:, :]
+    groups = keyboard_tensor.view(-1, 4, keyboard_tensor.shape[1])
+    assert (groups == groups[:, 0:1]).all(dim=1).all(), "keyboard_tensor must have the same value for each group"
+    groups = mouse_tensor.view(-1, 4, mouse_tensor.shape[1])
+    assert (groups == groups[:, 0:1]).all(dim=1).all(), "mouse_tensor must have the same value for each group"
+    
+    return keyboard_tensor[::4], mouse_tensor[::4]
+
+def process_custom_actions(keyboard_tensor, mouse_tensor, forward_speed=DEFAULT_FORWARD_SPEED):
+    """
+    Process custom keyboard and mouse tensors into model inputs (viewmats, intrinsics, action_labels).
+    Assumes inputs correspond to each LATENT frame.
+    """
+    if keyboard_tensor.ndim == 3:
+        keyboard_tensor = keyboard_tensor.squeeze(0)
+    if mouse_tensor.ndim == 3:
+        mouse_tensor = mouse_tensor.squeeze(0)
+    keyboard_tensor, mouse_tensor = reformat_keyboard_and_mouse_tensors(keyboard_tensor, mouse_tensor)
+
+    motions = []
+    
+    # 1. Translate tensors to motions for trajectory generation
+    for t in range(keyboard_tensor.shape[0]):
+        frame_motion = {}
+        
+        # --- Translation ---
+        # MatrixGame convention: 0:W, 1:S, 2:A, 3:D
+        fwd = 0.0
+        if keyboard_tensor[t, 0] > 0.5: fwd += forward_speed   # W
+        if keyboard_tensor[t, 1] > 0.5: fwd -= forward_speed   # S
+        if fwd != 0: frame_motion["forward"] = fwd
+        
+        rgt = 0.0
+        if keyboard_tensor[t, 2] > 0.5: rgt -= forward_speed   # A (Left is negative Right)
+        if keyboard_tensor[t, 3] > 0.5: rgt += forward_speed   # D (Right)
+        if rgt != 0: frame_motion["right"] = rgt
+        
+        # --- Rotation ---
+        # MatrixGame convention: mouse is [Pitch, Yaw] (or Y, X)
+        # Apply scaling (e.g. to match HyWorld distribution)
+        pitch = mouse_tensor[t, 0].item()
+        yaw = mouse_tensor[t, 1].item()
+        
+        if abs(pitch) > 1e-4: frame_motion["pitch"] = pitch
+        if abs(yaw) > 1e-4: frame_motion["yaw"] = yaw
+        
+        motions.append(frame_motion)
+
+    # 2. Generate Camera Trajectory
+    # generate_camera_trajectory_local returns T+1 poses (starting at Identity)
+    # We take the first T poses to match the latent count.
+    # Pose 0 is Identity. Pose 1 is Identity + Motion[0].
+    poses = generate_camera_trajectory_local(motions)
+    # poses = np.array(poses[:T])
+    
+    # 3. Compute Viewmats (w2c) and Intrinsics
+    w2c_list = []
+    intrinsic_list = []
+    
+    # Setup default intrinsic (normalized)
+    K = np.array(DEFAULT_INTRINSIC)
+    K[0, 0] /= K[0, 2] * 2
+    K[1, 1] /= K[1, 2] * 2
+    K[0, 2] = 0.5
+    K[1, 2] = 0.5
+    
+    for i in range(len(poses)):
+        c2w = np.array(poses[i])
+        w2c = np.linalg.inv(c2w)
+        w2c_list.append(w2c)
+        intrinsic_list.append(K)
+        
+    viewmats = torch.as_tensor(np.array(w2c_list))
+    intrinsics = torch.as_tensor(np.array(intrinsic_list))
+
+    # 4. Generate Action Labels by analyzing the generated trajectory
+    # This ensures consistency with complex simultaneous movements, exactly as pose_to_input does.
+    
+    # Calculate relative camera-to-world transforms
+    # c2ws = inverse(viewmats)
+    c2ws = np.linalg.inv(np.array(w2c_list))
+    
+    # Calculate relative movement between frames
+    # relative_c2w[i] = inv(c2ws[i-1]) @ c2ws[i]
+    C_inv = np.linalg.inv(c2ws[:-1])
+    relative_c2w = np.zeros_like(c2ws)
+    relative_c2w[0, ...] = c2ws[0, ...] # First is anchor
+    relative_c2w[1:, ...] = C_inv @ c2ws[1:, ...]
+    
+    # Initialize one-hot action encodings
+    trans_one_hot = np.zeros((relative_c2w.shape[0], 4), dtype=np.int32)
+    rotate_one_hot = np.zeros((relative_c2w.shape[0], 4), dtype=np.int32)
+
+    move_norm_valid = 0.0001
+    
+    # Skip index 0 (anchor/identity)
+    for i in range(1, relative_c2w.shape[0]):
+        move_dirs = relative_c2w[i, :3, 3]  # direction vector
+        move_norms = np.linalg.norm(move_dirs)
+        
+        if move_norms > move_norm_valid:  # threshold for movement
+            move_norm_dirs = move_dirs / move_norms
+            angles_rad = np.arccos(move_norm_dirs.clip(-1.0, 1.0))
+            trans_angles_deg = angles_rad * (180.0 / np.pi)  # convert to degrees
+        else:
+            trans_angles_deg = np.zeros(3)
+
+        R_rel = relative_c2w[i, :3, :3]
+        r = R.from_matrix(R_rel)
+        rot_angles_deg = r.as_euler("xyz", degrees=True)
+
+        # Determine movement actions based on trajectory
+        # Note: HyWorld logic checks if rotation is small before assigning translation labels
+        # to avoid ambiguity in TPS mode, but here we generally want to capture the dominant movement.
+        tps = False # Default assumption, can be made an arg if needed
+        
+        if move_norms > move_norm_valid:
+            if (not tps) or (
+                tps and abs(rot_angles_deg[1]) < 5e-2 and abs(rot_angles_deg[0]) < 5e-2
+            ):
+                # Z-axis (Forward/Back)
+                if trans_angles_deg[2] < 60:
+                    trans_one_hot[i, 0] = 1  # forward
+                elif trans_angles_deg[2] > 120:
+                    trans_one_hot[i, 1] = 1  # backward
+
+                # X-axis (Right/Left)
+                if trans_angles_deg[0] < 60:
+                    trans_one_hot[i, 2] = 1  # right
+                elif trans_angles_deg[0] > 120:
+                    trans_one_hot[i, 3] = 1  # left
+
+        # Determine rotation actions
+        # Y-axis (Yaw)
+        if rot_angles_deg[1] > 5e-2:
+            rotate_one_hot[i, 0] = 1  # right
+        elif rot_angles_deg[1] < -5e-2:
+            rotate_one_hot[i, 1] = 1  # left
+
+        # X-axis (Pitch)
+        if rot_angles_deg[0] > 5e-2:
+            rotate_one_hot[i, 2] = 1  # up
+        elif rot_angles_deg[0] < -5e-2:
+            rotate_one_hot[i, 3] = 1  # down
+
+    trans_one_hot = torch.tensor(trans_one_hot)
+    rotate_one_hot = torch.tensor(rotate_one_hot)
+
+    # Convert to single labels
+    trans_label = one_hot_to_one_dimension(trans_one_hot)
+    rotate_label = one_hot_to_one_dimension(rotate_one_hot)
+    action_labels = trans_label * 9 + rotate_label
+
+    return viewmats, intrinsics, action_labels
+
+if __name__ == "__main__":
+    print("Running comparison test between process_custom_actions and pose_to_input...")
+
+    def test_process_custom_actions(pose_string: str, keyboard: torch.Tensor, mouse: torch.Tensor, latent_num: int):
+        # Run process_custom_actions
+        # Note: We need to pass float tensors
+        print("Running process_custom_actions...")
+        viewmats_1, intrinsics_1, labels_1 = process_custom_actions(
+            keyboard, mouse
+        )
+        
+        print(f"Running pose_to_input with string: '{pose_string}'...")
+        viewmats_2, intrinsics_2, labels_2 = pose_to_input(
+            pose_string, latent_num=latent_num
+        )
+
+        # print(f"Viewmats: {viewmats_1} vs \n {viewmats_2}")
+        # print(f"Intrinsics: {intrinsics_1} vs \n {intrinsics_2}")
+        # print(f"Labels: {labels_1} vs \n {labels_2}")
+        # 3. Compare Results
+        print("\nComparison Results:")
+        
+        # Check Shapes
+        print(f"Shapes (Viewmats): {viewmats_1.shape} vs {viewmats_2.shape}")
+        assert viewmats_1.shape == viewmats_2.shape, "Shape mismatch for viewmats"
+        
+        # Check Values
+        # Viewmats
+        diff_viewmats = (viewmats_1 - viewmats_2).abs().max().item()
+        print(f"Max difference in Viewmats: {diff_viewmats}")
+        if diff_viewmats < 1e-5:
+            print("✅ Viewmats match.")
+        else:
+            print("❌ Viewmats mismatch.")
+
+        # Check intrinsics
+        diff_intrinsics = (intrinsics_1 - intrinsics_2).abs().max().item()
+        print(f"Max difference in Intrinsics: {diff_intrinsics}")
+        if diff_intrinsics < 1e-5:
+            print("✅ Intrinsics match.")
+        else:
+            print("❌ Intrinsics mismatch.")
+
+        # Check labels
+        diff_labels = (labels_1 - labels_2).abs().max().item()
+        print(f"Max difference in Labels: {diff_labels}")
+        if diff_labels < 1e-5:
+            print("✅ Labels match.")
+        else:
+            print("❌ Labels mismatch.")
+
+        print("All checks passed.")
+    
+    # Define shared parameters
+
+    latent_num = 13
+    pose_string = "w-2, a-3, s-1, d-6"
+
+    num_frames = 4 * (latent_num - 1) + 1
+    keyboard = torch.zeros((num_frames, 6))
+    mouse = torch.zeros((num_frames, 2))
+
+    # Frame 0 is ignored/start
+    # Frames 1-8: Press W (index 0)
+    keyboard[1:9, 0] = 1.0
+    # Frames 9-20: Press A (index 2)
+    keyboard[9:21, 2] = 1.0
+    # Frames 21-24: Press S (index 1)
+    keyboard[21:25, 1] = 1.0
+    # Frames 25-48: Press D (index 3)
+    keyboard[25:49, 3] = 1.0
+
+    test_process_custom_actions(pose_string, keyboard, mouse, latent_num)
+
+    # Test keyboard AND mouse
+    latent_num = 25
+    pose_string = "w-2, up-2, a-3, down-4, s-1, left-2, d-6, right-4"
+
+    num_frames = 4 * (latent_num - 1) + 1
+    keyboard = torch.zeros((num_frames, 6))
+    mouse = torch.zeros((num_frames, 2))
+
+    # Frame 0 is ignored/start
+    # Frames 1-8: Press W (index 0)
+    keyboard[1:9, 0] = 1.0
+    # Frames 17-28: Press A (index 2)
+    keyboard[17:29, 2] = 1.0
+    # Frames 45-48: Press S (index 1)
+    keyboard[45:49, 1] = 1.0
+    # Frames 57-80: Press D (index 3)
+    keyboard[57:81, 3] = 1.0
+    
+    # Frames 9-16: Press Up (index 4)
+    mouse[9:17, 0] = DEFAULT_PITCH_SPEED
+    # Frames 25-32: Press Down (index 5)
+    mouse[29:45, 0] = -DEFAULT_PITCH_SPEED
+    # Frames 41-48: Press Left (index 6)
+    mouse[49:57, 1] = -DEFAULT_YAW_SPEED
+    # Frames 57-64: Press Right (index 7)
+    mouse[81:, 1] = DEFAULT_YAW_SPEED
+
+    test_process_custom_actions(pose_string, keyboard, mouse, latent_num)
@@ -1134,5 +1134,3 @@ class LongCatTransformer3DModel(CachableDiT):
        )
        return x

-# Entry point for model registry
-EntryClass = LongCatTransformer3DModel
@@ -18,23 +18,12 @@ import torch.nn as nn
 from einops import rearrange, repeat

 from fastvideo.attention.backends.sdpa import SDPAMetadata
-from fastvideo.attention.layer import DistributedAttention, LocalAttention
+from fastvideo.attention.layer import LocalAttention
 from fastvideo.configs.models.dits import LTX2VideoConfig
-from fastvideo.distributed.communication_op import (
-    sequence_model_parallel_all_gather,
-    sequence_model_parallel_all_gather_with_unpad,
-    sequence_model_parallel_all_to_all_4D,
-    sequence_model_parallel_shard,
-)
-from fastvideo.distributed.parallel_state import get_sp_parallel_rank, get_sp_world_size
-from fastvideo.distributed.utils import create_attention_mask_for_padding
-from fastvideo.forward_context import ForwardContext, get_forward_context, set_forward_context
-from fastvideo.logger import init_logger
+from fastvideo.forward_context import get_forward_context, set_forward_context
 from fastvideo.models.dits.base import CachableDiT
 from fastvideo.platforms import AttentionBackendEnum

-logger = init_logger(__name__)
-

 def get_timestep_embedding(
    timesteps: torch.Tensor,
@@ -306,12 +295,6 @@ class VideoLatentPatchifier:
        output_shape: VideoLatentShape,
        device: Optional[torch.device] = None,
    ) -> torch.Tensor:
-        """Get patch grid bounds for RoPE computation.
-
-        Args:
-            output_shape: Shape of the video latent tensor
-            device: Device to create tensors on
-        """
        frames = output_shape.frames
        height = output_shape.height
        width = output_shape.width
@@ -383,12 +366,6 @@ class AudioLatentPatchifier:
        output_shape: AudioLatentShape,
        device: Optional[torch.device] = None,
    ) -> torch.Tensor:
-        """Get patch grid bounds for audio RoPE computation.
-
-        Args:
-            output_shape: Shape of the audio latent tensor
-            device: Device to create tensors on
-        """
        start_timings = self._get_audio_latent_time_in_sec(
            self.shift,
            output_shape.frames + self.shift,
@@ -538,44 +515,6 @@ def apply_ltx_rotary_emb(
    raise ValueError(f"Invalid rope type: {rope_type}")


-def apply_ltx_rotary_emb_4d(
-    input_tensor: torch.Tensor,
-    freqs_cis: tuple[torch.Tensor, torch.Tensor],
-    rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
-) -> torch.Tensor:
-    """Apply LTX-2 rotary embeddings to a 4D tensor [B, H, T, D].
-    
-    This is used for applying RoPE after all-to-all in distributed attention,
-    where the tensor is already in [B, H, T, D] format.
-    """
-    cos_freqs, sin_freqs = freqs_cis
-    if rope_type == LTXRopeType.INTERLEAVED:
-        # For interleaved, cos/sin have shape [B, T, inner_dim]
-        # Need to reshape to [B, H, T, D] format
-        # Actually interleaved doesn't have per-head rotations, so we broadcast
-        t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2)
-        t1, t2 = t_dup.unbind(dim=-1)
-        t_dup = torch.stack((-t2, t1), dim=-1)
-        input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)")
-        return input_tensor * cos_freqs + input_tensor_rot * sin_freqs
-    if rope_type == LTXRopeType.SPLIT:
-        # For split, cos/sin already have shape [B, H, T, D/2]
-        # input_tensor is [B, H, T, D]
-        split_input = rearrange(input_tensor, "... (d r) -> ... d r", d=2)
-        first_half_input = split_input[..., :1, :]
-        second_half_input = split_input[..., 1:, :]
-
-        output = split_input * cos_freqs.unsqueeze(-2)
-        first_half_output = output[..., :1, :]
-        second_half_output = output[..., 1:, :]
-
-        first_half_output.addcmul_(-sin_freqs.unsqueeze(-2), second_half_input)
-        second_half_output.addcmul_(sin_freqs.unsqueeze(-2), first_half_input)
-
-        return rearrange(output, "... d r -> ... (d r)")
-    raise ValueError(f"Invalid rope type: {rope_type}")
-
-
 def _apply_ltx_interleaved_rotary_emb(
    input_tensor: torch.Tensor, cos_freqs: torch.Tensor, sin_freqs: torch.Tensor
 ) -> torch.Tensor:
@@ -979,205 +918,6 @@ class TransformerConfig:
    context_dim: int


-class LTXDistributedAttention(DistributedAttention):
-    """LTX-2 specialized DistributedAttention that handles LTX-style RoPE internally."""
-
-    def __init__(
-        self,
-        num_heads: int,
-        head_size: int,
-        rope_type: LTXRopeType,
-        num_kv_heads: int | None = None,
-        softmax_scale: float | None = None,
-        causal: bool = False,
-        supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None,
-        prefix: str = "",
-        **extra_impl_args,
-    ) -> None:
-        super().__init__(
-            num_heads=num_heads,
-            head_size=head_size,
-            num_kv_heads=num_kv_heads,
-            softmax_scale=softmax_scale,
-            causal=causal,
-            supported_attention_backends=supported_attention_backends,
-            prefix=prefix,
-            **extra_impl_args,
-        )
-        self.rope_type = rope_type
-
-    @torch.compiler.disable
-    def forward(
-        self,
-        q: torch.Tensor,
-        k: torch.Tensor,
-        v: torch.Tensor,
-        replicated_q: torch.Tensor | None = None,
-        replicated_k: torch.Tensor | None = None,
-        replicated_v: torch.Tensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        ltx_freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None,
-    ) -> tuple[torch.Tensor, torch.Tensor | None]:
-        """Forward pass with LTX-2 style RoPE application.
-
-        Args:
-            q: Query tensor [batch_size, seq_len, num_heads, head_dim]
-            k: Key tensor [batch_size, seq_len, num_heads, head_dim]
-            v: Value tensor [batch_size, seq_len, num_heads, head_dim]
-            replicated_q: Replicated query tensor for text tokens
-            replicated_k: Replicated key tensor
-            replicated_v: Replicated value tensor
-            attention_mask: Attention mask [batch_size, seq_len]
-            ltx_freqs_cis: LTX-2 style RoPE (cos, sin) with shape [B, H, T, D]
-
-        Returns:
-            Tuple of output tensor and optional replicated output
-        """
-        assert q.dim() == 4 and k.dim() == 4 and v.dim() == 4, "Expected 4D tensors"
-        batch_size, seq_len, num_heads, head_dim = q.shape
-        local_rank = get_sp_parallel_rank()
-        world_size = get_sp_world_size()
-
-        forward_context: ForwardContext = get_forward_context()
-        ctx_attn_metadata = forward_context.attn_metadata
-
-        # Stack QKV
-        qkv = torch.cat([q, k, v], dim=0)  # [3*batch, seq_len, num_heads, head_dim]
-
-        # Redistribute heads across sequence dimension
-        qkv = sequence_model_parallel_all_to_all_4D(qkv, scatter_dim=2, gather_dim=1)
-
-        # After all-to-all, each rank has the full sequence but only a subset of heads
-        valid_seq_len = None
-        if attention_mask is not None:
-            valid_seq_len = (attention_mask[0] == 1).sum().item()
-            qkv = qkv[:, :valid_seq_len, :, :]
-
-        # Apply LTX-2 style RoPE after all-to-all (when we have full sequence)
-        if ltx_freqs_cis is not None:
-            cos, sin = ltx_freqs_cis
-            heads_per_rank = num_heads // world_size
-            head_start = local_rank * heads_per_rank
-            head_end = head_start + heads_per_rank
-            # Slice to this rank's heads: [B, H/SP, T, D]
-            cos_local = cos[:, head_start:head_end, :valid_seq_len, :]
-            sin_local = sin[:, head_start:head_end, :valid_seq_len, :]
-
-            # Apply RoPE to Q and K together (first 2*batch_size in dim 0)
-            qk_part = qkv[:batch_size * 2]
-            # Transpose to [2*B, H/SP, T, D] for LTX RoPE application
-            qk_part = qk_part.transpose(1, 2)
-            qk_part = apply_ltx_rotary_emb_4d(qk_part, (cos_local, sin_local), self.rope_type)
-            # Transpose back to [2*B, T, H/SP, D]
-            qkv[:batch_size * 2] = qk_part.transpose(1, 2)
-
-        # Apply backend-specific preprocess_qkv
-        qkv = self.attn_impl.preprocess_qkv(qkv, ctx_attn_metadata)
-
-        # Concatenate with replicated QKV if provided
-        if replicated_q is not None:
-            assert replicated_k is not None and replicated_v is not None
-            replicated_qkv = torch.cat(
-                [replicated_q, replicated_k, replicated_v],
-                dim=0)  # [3, seq_len, num_heads, head_dim]
-            heads_per_rank = num_heads // world_size
-            replicated_qkv = replicated_qkv[:, :, local_rank *
-                                            heads_per_rank:(local_rank + 1) *
-                                            heads_per_rank]
-            qkv = torch.cat([qkv, replicated_qkv], dim=1)
-
-        q, k, v = qkv.chunk(3, dim=0)
-
-        output = self.attn_impl.forward(q, k, v, ctx_attn_metadata)
-
-        # Redistribute back if using sequence parallelism
-        replicated_output = None
-        if replicated_q is not None:
-            split_idx = seq_len * world_size if valid_seq_len is None else valid_seq_len
-            replicated_output = output[:, split_idx:]
-            output = output[:, :split_idx]
-            replicated_output = sequence_model_parallel_all_gather(
-                replicated_output.contiguous(), dim=2)
-
-        # Apply backend-specific postprocess_output
-        output = self.attn_impl.postprocess_output(output, ctx_attn_metadata)
-
-        if attention_mask is not None:
-            pad_len = (attention_mask[0] == 0).sum().item()
-            output = torch.nn.functional.pad(output, (0, 0, 0, 0, 0, pad_len))
-
-        output = sequence_model_parallel_all_to_all_4D(output, scatter_dim=1, gather_dim=2)
-
-        return output, replicated_output
-
-
-class LTXLocalAttention(LocalAttention):
-    """LTX-2 specialized LocalAttention that handles LTX-style RoPE internally."""
-
-    def __init__(
-        self,
-        num_heads: int,
-        head_size: int,
-        rope_type: LTXRopeType,
-        num_kv_heads: int | None = None,
-        softmax_scale: float | None = None,
-        causal: bool = False,
-        supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None,
-        **extra_impl_args,
-    ) -> None:
-        super().__init__(
-            num_heads=num_heads,
-            head_size=head_size,
-            num_kv_heads=num_kv_heads,
-            softmax_scale=softmax_scale,
-            causal=causal,
-            supported_attention_backends=supported_attention_backends,
-            **extra_impl_args,
-        )
-        self.rope_type = rope_type
-
-    def forward(
-        self,
-        q: torch.Tensor,
-        k: torch.Tensor,
-        v: torch.Tensor,
-        ltx_freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None,
-        ltx_k_freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None,
-    ) -> torch.Tensor:
-        """Forward pass with LTX-2 style RoPE application.
-
-        Args:
-            q: Query tensor [batch_size, seq_len, num_heads, head_dim]
-            k: Key tensor [batch_size, seq_len, num_heads, head_dim]
-            v: Value tensor [batch_size, seq_len, num_heads, head_dim]
-            ltx_freqs_cis: LTX-2 style RoPE (cos, sin) for Q with shape [B, H, T, D]
-            ltx_k_freqs_cis: LTX-2 style RoPE (cos, sin) for K (if different from Q)
-
-        Returns:
-            Output tensor after local attention
-        """
-        assert q.dim() == 4 and k.dim() == 4 and v.dim() == 4, "Expected 4D tensors"
-
-        forward_context: ForwardContext = get_forward_context()
-        ctx_attn_metadata = forward_context.attn_metadata
-
-        # Apply LTX-2 style RoPE
-        if ltx_freqs_cis is not None:
-            # Use separate K RoPE if provided (for cross-attention), otherwise use same as Q
-            k_freqs = ltx_k_freqs_cis if ltx_k_freqs_cis is not None else ltx_freqs_cis
-            # Transpose to [B, H, T, D] for RoPE application
-            q = q.transpose(1, 2)
-            k = k.transpose(1, 2)
-            q = apply_ltx_rotary_emb_4d(q, ltx_freqs_cis, self.rope_type)
-            k = apply_ltx_rotary_emb_4d(k, k_freqs, self.rope_type)
-            # Transpose back to [B, T, H, D]
-            q = q.transpose(1, 2)
-            k = k.transpose(1, 2)
-
-        output = self.attn_impl.forward(q, k, v, ctx_attn_metadata)
-        return output
-
-
 class LTXSelfAttention(nn.Module):
    """LTX-2 attention block with RMSNorm + FastVideo LocalAttention."""
    def __init__(
@@ -1205,19 +945,17 @@ class LTXSelfAttention(nn.Module):
        self.to_v = nn.Linear(context_dim, inner_dim, bias=True)
        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim, bias=True), nn.Identity())

-        self.attn = LTXLocalAttention(
+        self.attn = LocalAttention(
            num_heads=heads,
            head_size=dim_head,
-            rope_type=rope_type,
            dropout_rate=0,
            softmax_scale=None,
            causal=False,
            supported_attention_backends=supported_attention_backends,
        )
-        self.attn_masked = LTXLocalAttention(
+        self.attn_masked = LocalAttention(
            num_heads=heads,
            head_size=dim_head,
-            rope_type=rope_type,
            dropout_rate=0,
            softmax_scale=None,
            causal=False,
@@ -1240,13 +978,15 @@ class LTXSelfAttention(nn.Module):
        q = self.q_norm(q)
        k = self.k_norm(k)

-        # RoPE is applied inside LTXLocalAttention
+        if pe is not None:
+            q = apply_ltx_rotary_emb(q, pe, self.rope_type)
+            k = apply_ltx_rotary_emb(k, pe if k_pe is None else k_pe, self.rope_type)
+
        b, q_len, _ = q.shape
        k_len = k.shape[1]
        q = q.view(b, q_len, self.heads, self.dim_head)
        k = k.view(b, k_len, self.heads, self.dim_head)
        v = v.view(b, k_len, self.heads, self.dim_head)
-
        if mask is not None:
            if mask.ndim == 2:
                mask = mask.unsqueeze(0)
@@ -1268,93 +1008,9 @@ class LTXSelfAttention(nn.Module):
                attn_metadata=attn_metadata,
                forward_batch=forward_batch,
            ):
-                out = self.attn_masked(q, k, v, ltx_freqs_cis=pe, ltx_k_freqs_cis=k_pe)
+                out = self.attn_masked(q, k, v)
        else:
-            out = self.attn(q, k, v, ltx_freqs_cis=pe, ltx_k_freqs_cis=k_pe)
-        out = out.reshape(b, q_len, -1)
-        return self.to_out(out)
-
-
-class LTXDistributedSelfAttention(nn.Module):
-    """LTX-2 attention block with RMSNorm + LTXDistributedAttention for SP."""
-
-    def __init__(
-        self,
-        query_dim: int,
-        context_dim: int | None,
-        heads: int,
-        dim_head: int,
-        norm_eps: float,
-        rope_type: LTXRopeType,
-        supported_attention_backends: tuple[AttentionBackendEnum, ...],
-        prefix: str = "",
-    ) -> None:
-        super().__init__()
-        inner_dim = dim_head * heads
-        context_dim = query_dim if context_dim is None else context_dim
-
-        self.heads = heads
-        self.dim_head = dim_head
-        self.rope_type = rope_type
-
-        self.q_norm = torch.nn.RMSNorm(inner_dim, eps=norm_eps)
-        self.k_norm = torch.nn.RMSNorm(inner_dim, eps=norm_eps)
-        self.to_q = nn.Linear(query_dim, inner_dim, bias=True)
-        self.to_k = nn.Linear(context_dim, inner_dim, bias=True)
-        self.to_v = nn.Linear(context_dim, inner_dim, bias=True)
-        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim, bias=True), nn.Identity())
-
-        self.attn = LTXDistributedAttention(
-            num_heads=heads,
-            head_size=dim_head,
-            rope_type=rope_type,
-            causal=False,
-            supported_attention_backends=supported_attention_backends,
-            prefix=f"{prefix}.attn",
-        )
-
-    def forward(
-        self,
-        x: torch.Tensor,
-        context: torch.Tensor | None = None,
-        pe: tuple[torch.Tensor, torch.Tensor] | None = None,
-        k_pe: tuple[torch.Tensor, torch.Tensor] | None = None,
-        attention_mask: torch.Tensor | None = None,
-    ) -> torch.Tensor:
-        """Forward pass for distributed self-attention.
-
-        Args:
-            x: Query tensor [B, L, C]
-            context: Key/Value tensor [B, L, C], or None for self-attention
-            pe: Rotary position embeddings for Q (cos, sin) with shape [B, H, T_full, D]
-                NOTE: For SP, this must be the FULL sequence RoPE, not sharded!
-            k_pe: Rotary position embeddings for K (cos, sin), or None to use pe
-            attention_mask: Attention mask for padding [B, padded_seq_len]
-        """
-        q = self.to_q(x)
-        context = x if context is None else context
-        k = self.to_k(context)
-        v = self.to_v(context)
-
-        q = self.q_norm(q)
-        k = self.k_norm(k)
-
-        # RoPE is applied inside LTXDistributedAttention AFTER the all-to-all,
-        # when each rank has the full sequence (but subset of heads).
-        b, q_len, _ = q.shape
-        k_len = k.shape[1]
-        q = q.view(b, q_len, self.heads, self.dim_head)
-        k = k.view(b, k_len, self.heads, self.dim_head)
-        v = v.view(b, k_len, self.heads, self.dim_head)
-
-        # Pass full RoPE to distributed attention - it will apply after all-to-all
-        # and slice to this rank's heads
-        out, _ = self.attn(
-            q, k, v,
-            attention_mask=attention_mask,
-            ltx_freqs_cis=pe,
-        )
-
+            out = self.attn(q, k, v)
        out = out.reshape(b, q_len, -1)
        return self.to_out(out)

@@ -1369,31 +1025,12 @@ class BasicAVTransformerBlock(torch.nn.Module):
        audio: TransformerConfig | None = None,
        rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
        norm_eps: float = 1e-6,
-        use_distributed_attention: bool = False,
-        prefix: str = "",
    ):
        super().__init__()
        self.idx = idx
-        self.use_distributed_attention = use_distributed_attention
-
-        # Choose attention class based on SP mode
-        # Self-attention and audio-video cross-attention use DistributedAttention when SP > 1
-        # Text cross-attention uses LocalAttention (text embeddings are replicated)
-        SelfAttnCls = LTXDistributedSelfAttention if use_distributed_attention else LTXSelfAttention
-        CrossAttnCls = LTXSelfAttention  # Text cross-attention is always local

        if video is not None:
-            # Video self-attention - use distributed when SP > 1
-            self.attn1 = SelfAttnCls(
-                query_dim=video.dim,
-                context_dim=None,
-                heads=video.heads,
-                dim_head=video.d_head,
-                norm_eps=norm_eps,
-                rope_type=rope_type,
-                supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA),
-                prefix=f"{prefix}.blocks.{idx}.attn1" if use_distributed_attention else "",
-            ) if use_distributed_attention else LTXSelfAttention(
+            self.attn1 = LTXSelfAttention(
                query_dim=video.dim,
                context_dim=None,
                heads=video.heads,
@@ -1402,8 +1039,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                rope_type=rope_type,
                supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA),
            )
-            # Text cross-attention - always local (text is replicated)
-            self.attn2 = CrossAttnCls(
+            self.attn2 = LTXSelfAttention(
                query_dim=video.dim,
                context_dim=video.context_dim,
                heads=video.heads,
@@ -1416,17 +1052,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
            self.scale_shift_table = torch.nn.Parameter(torch.empty(6, video.dim))

        if audio is not None:
-            # Audio self-attention - use distributed when SP > 1
-            self.audio_attn1 = SelfAttnCls(
-                query_dim=audio.dim,
-                context_dim=None,
-                heads=audio.heads,
-                dim_head=audio.d_head,
-                norm_eps=norm_eps,
-                rope_type=rope_type,
-                supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA),
-                prefix=f"{prefix}.blocks.{idx}.audio_attn1" if use_distributed_attention else "",
-            ) if use_distributed_attention else LTXSelfAttention(
+            self.audio_attn1 = LTXSelfAttention(
                query_dim=audio.dim,
                context_dim=None,
                heads=audio.heads,
@@ -1435,8 +1061,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                rope_type=rope_type,
                supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA),
            )
-            # Text cross-attention - always local (text is replicated)
-            self.audio_attn2 = CrossAttnCls(
+            self.audio_attn2 = LTXSelfAttention(
                query_dim=audio.dim,
                context_dim=audio.context_dim,
                heads=audio.heads,
@@ -1449,8 +1074,6 @@ class BasicAVTransformerBlock(torch.nn.Module):
            self.audio_scale_shift_table = torch.nn.Parameter(torch.empty(6, audio.dim))

        if audio is not None and video is not None:
-            # Audio-to-video cross-attention
-            # Uses local attention - context is gathered from all SP ranks in forward()
            self.audio_to_video_attn = LTXSelfAttention(
                query_dim=video.dim,
                context_dim=audio.dim,
@@ -1460,8 +1083,6 @@ class BasicAVTransformerBlock(torch.nn.Module):
                rope_type=rope_type,
                supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA),
            )
-            # Video-to-audio cross-attention
-            # Uses local attention - context is gathered from all SP ranks in forward()
            self.video_to_audio_attn = LTXSelfAttention(
                query_dim=audio.dim,
                context_dim=video.dim,
@@ -1510,17 +1131,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
        self,
        video: TransformerArgs | None,
        audio: TransformerArgs | None,
-        video_attention_mask: torch.Tensor | None = None,
-        audio_attention_mask: torch.Tensor | None = None,
    ) -> tuple[TransformerArgs | None, TransformerArgs | None]:
-        """Forward pass for transformer block.
-
-        Args:
-            video: Video transformer args
-            audio: Audio transformer args
-            video_attention_mask: SP padding attention mask for video [B, padded_seq_len]
-            audio_attention_mask: SP padding attention mask for audio [B, padded_seq_len]
-        """
        vx = video.x if video is not None else None
        ax = audio.x if audio is not None else None

@@ -1534,12 +1145,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                self.scale_shift_table, vx.shape[0], video.timesteps, slice(0, 3)
            )
            norm_vx = torch.nn.functional.rms_norm(vx, (vx.shape[-1],), eps=self.norm_eps) * (1 + vscale_msa) + vshift_msa
-            # Self-attention: pass SP attention mask for distributed attention
-            if self.use_distributed_attention:
-                vx = vx + self.attn1(norm_vx, pe=video.positional_embeddings, attention_mask=video_attention_mask) * vgate_msa
-            else:
-                vx = vx + self.attn1(norm_vx, pe=video.positional_embeddings) * vgate_msa
-            # Text cross-attention: no SP mask needed (text is replicated, uses local attention)
+            vx = vx + self.attn1(norm_vx, pe=video.positional_embeddings) * vgate_msa
            vx = vx + self.attn2(
                torch.nn.functional.rms_norm(vx, (vx.shape[-1],), eps=self.norm_eps),
                context=video.context,
@@ -1551,12 +1157,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                self.audio_scale_shift_table, ax.shape[0], audio.timesteps, slice(0, 3)
            )
            norm_ax = torch.nn.functional.rms_norm(ax, (ax.shape[-1],), eps=self.norm_eps) * (1 + ascale_msa) + ashift_msa
-            # Self-attention: pass SP attention mask for distributed attention
-            if self.use_distributed_attention:
-                ax = ax + self.audio_attn1(norm_ax, pe=audio.positional_embeddings, attention_mask=audio_attention_mask) * agate_msa
-            else:
-                ax = ax + self.audio_attn1(norm_ax, pe=audio.positional_embeddings) * agate_msa
-            # Text cross-attention: no SP mask needed (text is replicated, uses local attention)
+            ax = ax + self.audio_attn1(norm_ax, pe=audio.positional_embeddings) * agate_msa
            ax = ax + self.audio_attn2(
                torch.nn.functional.rms_norm(ax, (ax.shape[-1],), eps=self.norm_eps),
                context=audio.context,
@@ -1596,91 +1197,28 @@ class BasicAVTransformerBlock(torch.nn.Module):
            if run_a2v:
                vx_scaled = vx_norm3 * (1 + scale_ca_video_hidden_states_a2v) + shift_ca_video_hidden_states_a2v
                ax_scaled = ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v) + shift_ca_audio_hidden_states_a2v
-                # Audio-to-video cross-attention: Q from video (sharded), K/V from audio
-                # For SP, gather audio context from all ranks before cross-attention
-                if self.use_distributed_attention:
-                    # Gather audio context from all SP ranks
-                    ax_context = sequence_model_parallel_all_gather(ax_scaled, dim=1)
-                    # Video Q: video tokens divide evenly (adjusted upstream), so simple slicing works
-                    video_q_pe = None
-                    if video.cross_positional_embeddings is not None:
-                        sp_rank = get_sp_parallel_rank()
-                        local_seq_len = vx_scaled.shape[1]
-                        pe_tuple = video.cross_positional_embeddings
-                        start_idx = sp_rank * local_seq_len
-                        end_idx = start_idx + local_seq_len
-                        video_q_pe = tuple(pe[:, :, start_idx:end_idx, :] for pe in pe_tuple)
-                    # Audio K: gathered context may have padding, trim to original length
-                    audio_k_pe = audio.cross_positional_embeddings
-                    if audio_k_pe is not None:
-                        original_audio_len = audio_k_pe[0].shape[2]
-                        ax_context = ax_context[:, :original_audio_len, :]
-                    vx = vx + (
-                        self.audio_to_video_attn(
-                            vx_scaled,
-                            context=ax_context,
-                            pe=video_q_pe,
-                            k_pe=audio_k_pe,
-                        )
-                        * gate_out_a2v
-                    )
-                else:
-                    ax_context = ax_scaled
-                    video_q_pe = video.cross_positional_embeddings
-                    audio_k_pe = audio.cross_positional_embeddings
-                    vx = vx + (
-                        self.audio_to_video_attn(
-                            vx_scaled,
-                            context=ax_context,
-                            pe=video_q_pe,
-                            k_pe=audio_k_pe,
-                        )
-                        * gate_out_a2v
+                vx = vx + (
+                    self.audio_to_video_attn(
+                        vx_scaled,
+                        context=ax_scaled,
+                        pe=video.cross_positional_embeddings,
+                        k_pe=audio.cross_positional_embeddings,
                    )
+                    * gate_out_a2v
+                )

            if run_v2a:
                ax_scaled = ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a) + shift_ca_audio_hidden_states_v2a
                vx_scaled = vx_norm3 * (1 + scale_ca_video_hidden_states_v2a) + shift_ca_video_hidden_states_v2a
-                # Video-to-audio cross-attention: Q from audio, K/V from video
-                # For SP, gather both modalities and run cross-attention on full sequences
-                # (audio is short, so no need to keep it sharded for cross-attention)
-                if self.use_distributed_attention:
-                    # Gather both modalities to full sequences
-                    ax_full = sequence_model_parallel_all_gather(ax_scaled, dim=1)
-                    vx_context = sequence_model_parallel_all_gather(vx_scaled, dim=1)
-                    # Use full positional embeddings directly
-                    audio_q_pe = audio.cross_positional_embeddings
-                    video_k_pe = video.cross_positional_embeddings
-                    # Trim gathered tensors to original lengths (may have padding from sharding)
-                    if audio_q_pe is not None:
-                        original_audio_len = audio_q_pe[0].shape[2]
-                        ax_full = ax_full[:, :original_audio_len, :]
-                    if video_k_pe is not None:
-                        original_video_len = video_k_pe[0].shape[2]
-                        vx_context = vx_context[:, :original_video_len, :]
-                    # Run cross-attention on full audio
-                    v2a_out_full = self.video_to_audio_attn(
-                        ax_full,
-                        context=vx_context,
-                        pe=audio_q_pe,
-                        k_pe=video_k_pe,
-                    )
-                    # Shard the output back to match ax's sharding
-                    v2a_out, _ = sequence_model_parallel_shard(v2a_out_full, dim=1)
-                    ax = ax + v2a_out * gate_out_v2a
-                else:
-                    vx_context = vx_scaled
-                    audio_q_pe = audio.cross_positional_embeddings
-                    video_k_pe = video.cross_positional_embeddings
-                    ax = ax + (
-                        self.video_to_audio_attn(
-                            ax_scaled,
-                            context=vx_context,
-                            pe=audio_q_pe,
-                            k_pe=video_k_pe,
-                        )
-                        * gate_out_v2a
+                ax = ax + (
+                    self.video_to_audio_attn(
+                        ax_scaled,
+                        context=vx_scaled,
+                        pe=audio.cross_positional_embeddings,
+                        k_pe=video.cross_positional_embeddings,
                    )
+                    * gate_out_v2a
+                )

        if run_vx:
            vshift_mlp, vscale_mlp, vgate_mlp = self.get_ada_values(
@@ -1751,8 +1289,6 @@ class LTXModel(torch.nn.Module):
        av_ca_timestep_scale_multiplier: int = 1,
        rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
        double_precision_rope: bool = False,
-        use_distributed_attention: bool = False,
-        prefix: str = "",
    ):
        super().__init__()
        self._enable_gradient_checkpointing = False
@@ -1804,8 +1340,6 @@ class LTXModel(torch.nn.Module):
            audio_attention_head_dim=audio_attention_head_dim if model_type.is_audio_enabled() else 0,
            audio_cross_attention_dim=audio_cross_attention_dim,
            norm_eps=norm_eps,
-            use_distributed_attention=use_distributed_attention,
-            prefix=prefix,
        )

    def _init_video(
@@ -1935,8 +1469,6 @@ class LTXModel(torch.nn.Module):
        audio_attention_head_dim: int,
        audio_cross_attention_dim: int,
        norm_eps: float,
-        use_distributed_attention: bool = False,
-        prefix: str = "",
    ) -> None:
        video_config = (
            TransformerConfig(
@@ -1958,7 +1490,6 @@ class LTXModel(torch.nn.Module):
            if self.model_type.is_audio_enabled()
            else None
        )
-        self.use_distributed_attention = use_distributed_attention
        self.transformer_blocks = torch.nn.ModuleList(
            [
                BasicAVTransformerBlock(
@@ -1967,8 +1498,6 @@ class LTXModel(torch.nn.Module):
                    audio=audio_config,
                    rope_type=self.rope_type,
                    norm_eps=norm_eps,
-                    use_distributed_attention=use_distributed_attention,
-                    prefix=prefix,
                )
                for idx in range(num_layers)
            ]
@@ -1978,16 +1507,9 @@ class LTXModel(torch.nn.Module):
        self,
        video: TransformerArgs | None,
        audio: TransformerArgs | None,
-        video_attention_mask: torch.Tensor | None = None,
-        audio_attention_mask: torch.Tensor | None = None,
    ) -> tuple[TransformerArgs | None, TransformerArgs | None]:
        for block in self.transformer_blocks:
-            video, audio = block(
-                video=video,
-                audio=audio,
-                video_attention_mask=video_attention_mask,
-                audio_attention_mask=audio_attention_mask,
-            )
+            video, audio = block(video=video, audio=audio)
        return video, audio

    def _process_output(
@@ -2011,17 +1533,7 @@ class LTXModel(torch.nn.Module):
        self,
        video: Modality | None,
        audio: Modality | None,
-        video_attention_mask: torch.Tensor | None = None,
-        audio_attention_mask: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
-        """Forward pass through the LTX model.
-
-        Args:
-            video: Video modality input
-            audio: Audio modality input
-            video_attention_mask: SP padding attention mask for video [B, padded_seq_len]
-            audio_attention_mask: SP padding attention mask for audio [B, padded_seq_len]
-        """
        if os.getenv("LTX2_PIPELINE_DEBUG_LOG", "0") == "1":
            _debug_block_log_line(
                "fastvideo:patchify_proj"
@@ -2039,12 +1551,7 @@ class LTXModel(torch.nn.Module):
        audio_args = self.audio_args_preprocessor.prepare(audio) if audio is not None else None
        _debug_transformer_args("fastvideo:prep_video", video_args)
        _debug_transformer_args("fastvideo:prep_audio", audio_args)
-        video_out, audio_out = self._process_transformer_blocks(
-            video_args,
-            audio_args,
-            video_attention_mask=video_attention_mask,
-            audio_attention_mask=audio_attention_mask,
-        )
+        video_out, audio_out = self._process_transformer_blocks(video_args, audio_args)

        vx = (
            self._process_output(
@@ -2081,24 +1588,6 @@ class LTX2Transformer3DModel(CachableDiT):

        arch = config.arch_config

-        # Get SP world size for distributed attention
-        sp_world_size = get_sp_world_size()
-        use_distributed_attention = sp_world_size > 1
-
-        # Validate that attention heads are divisible by SP world size
-        if sp_world_size > 1:
-            assert arch.num_attention_heads % sp_world_size == 0, (
-                f"The number of video attention heads ({arch.num_attention_heads}) "
-                f"must be divisible by the sequence parallel size ({sp_world_size})"
-            )
-            assert arch.audio_num_attention_heads % sp_world_size == 0, (
-                f"The number of audio attention heads ({arch.audio_num_attention_heads}) "
-                f"must be divisible by the sequence parallel size ({sp_world_size})"
-            )
-            logger.info(
-                f"LTX2 sequence parallelism enabled with SP world size {sp_world_size}"
-            )
-
        model_type = LTXModelType.AudioVideo
        self.model = LTXModel(
            model_type=model_type,
@@ -2123,8 +1612,6 @@ class LTX2Transformer3DModel(CachableDiT):
            audio_cross_attention_dim=arch.audio_cross_attention_dim,
            audio_positional_embedding_max_pos=arch.audio_positional_embedding_max_pos,
            av_ca_timestep_scale_multiplier=arch.av_ca_timestep_scale_multiplier,
-            use_distributed_attention=use_distributed_attention,
-            prefix=config.prefix,
        )

        self.patchifier = VideoLatentPatchifier(patch_size=arch.patch_size[1])
@@ -2140,7 +1627,6 @@ class LTX2Transformer3DModel(CachableDiT):
        self.hidden_size = arch.num_attention_heads * arch.attention_head_dim
        self.num_attention_heads = arch.num_attention_heads
        self.num_channels_latents = arch.num_channels_latents
-        self._logged_attention_mask = False

        if os.getenv("LTX2_DEBUG_DETAIL", "0") == "1":
            detail_path = os.getenv("LTX2_PIPELINE_DEBUG_DETAIL_PATH", "")
@@ -2216,12 +1702,10 @@ class LTX2Transformer3DModel(CachableDiT):
        if isinstance(encoder_hidden_states, list):
            encoder_hidden_states = encoder_hidden_states[0]

-        # Get SP parameters
-        sp_world_size = get_sp_world_size()
-        sp_rank = get_sp_parallel_rank()
-        batch_size = hidden_states.shape[0]
-
-        # Get fps for position computation
+        video_shape = VideoLatentShape.from_torch_shape(hidden_states.shape)
+        positions = self.patchifier.get_patch_grid_bounds(
+            video_shape, device=hidden_states.device
+        )
        fps = None
        try:
            forward_ctx = get_forward_context()
@@ -2233,60 +1717,17 @@ class LTX2Transformer3DModel(CachableDiT):
                fps_value = fps_value[0] if fps_value else None
            if fps_value is not None:
                fps = float(fps_value)
-
-        # Patchify video latents
-        video_shape = VideoLatentShape.from_torch_shape(hidden_states.shape)
-        latents = self.patchifier.patchify(hidden_states)
-
-        # Shard video latents and timestep across SP ranks
-        video_original_seq_len = latents.shape[1]
-        video_attention_mask = None
-        video_timestep = timestep
-        if sp_world_size > 1:
-            latents, video_original_seq_len = sequence_model_parallel_shard(latents, dim=1)
-            # Shard timestep along sequence dimension (timestep has shape [batch, seq_len])
-            video_timestep, _ = sequence_model_parallel_shard(timestep, dim=1)
-            # Create attention mask for padded tokens
-            current_seq_len = latents.shape[1]
-            padded_seq_len = current_seq_len * sp_world_size
-            if padded_seq_len > video_original_seq_len:
-                if not self._logged_attention_mask:
-                    logger.info(
-                        f"Video padding applied, original seq len: {video_original_seq_len}, "
-                        f"padded seq len: {padded_seq_len}"
-                    )
-                    self._logged_attention_mask = True
-                video_attention_mask = create_attention_mask_for_padding(
-                    seq_len=video_original_seq_len,
-                    padded_seq_len=padded_seq_len,
-                    batch_size=batch_size,
-                    device=hidden_states.device,
-                )
-
-        # Compute RoPE positions for the FULL sequence (before sharding)
-        # This is necessary because sequence sharding may not align to frame boundaries.
-        # The full RoPE will be passed to DistributedAttention which applies it after
-        # the all-to-all (when each rank has the full sequence but subset of heads).
-        positions = self.patchifier.get_patch_grid_bounds(
-            video_shape, device=hidden_states.device
-        )
        positions = _get_pixel_coords(
            positions,
            DEFAULT_LTX2_SCALE_FACTORS,
            fps=fps,
            causal_fix=True,
        ).to(hidden_states.dtype)
-
-        # Pad positions to match padded sequence length for SP cross-attention
-        if sp_world_size > 1 and padded_seq_len > video_original_seq_len:
-            padding_needed = padded_seq_len - video_original_seq_len
-            last_pos = positions[:, :, -1:, :].expand(-1, -1, padding_needed, -1)
-            positions = torch.cat([positions, last_pos], dim=2)
-
+        latents = self.patchifier.patchify(hidden_states)
        video_modality = Modality(
            enabled=True,
            latent=latents,
-            timesteps=video_timestep,
+            timesteps=timestep,
            positions=positions,
            context=encoder_hidden_states,
            context_mask=encoder_attention_mask,
@@ -2304,45 +1745,19 @@ class LTX2Transformer3DModel(CachableDiT):
                f"latent_head={video_head} "
                f"latent_checksum={video_checksum:.6f}"
            )
-
-        # Process audio modality if provided
        audio_modality = None
        audio_shape = None
-        audio_original_seq_len = 0
-        audio_attention_mask = None
-
        if audio_hidden_states is not None and audio_encoder_hidden_states is not None and audio_timestep is not None:
-            audio_shape = AudioLatentShape.from_torch_shape(audio_hidden_states.shape)
-            audio_latents = self.audio_patchifier.patchify(audio_hidden_states)
-
-            # Shard audio latents and timestep across SP ranks
-            audio_original_seq_len = audio_latents.shape[1]
-            sharded_audio_timestep = audio_timestep
-            if sp_world_size > 1:
-                audio_latents, audio_original_seq_len = sequence_model_parallel_shard(audio_latents, dim=1)
-                # Shard audio timestep along sequence dimension
-                sharded_audio_timestep, _ = sequence_model_parallel_shard(audio_timestep, dim=1)
-                # Create attention mask for padded tokens
-                audio_current_seq_len = audio_latents.shape[1]
-                audio_padded_seq_len = audio_current_seq_len * sp_world_size
-                if audio_padded_seq_len > audio_original_seq_len:
-                    audio_attention_mask = create_attention_mask_for_padding(
-                        seq_len=audio_original_seq_len,
-                        padded_seq_len=audio_padded_seq_len,
-                        batch_size=batch_size,
-                        device=audio_hidden_states.device,
-                    )
-
-            # Compute audio RoPE positions for the FULL sequence (before sharding)
-            # Same as video: full RoPE is applied after all-to-all in DistributedAttention
+            audio_shape = AudioLatentShape.from_torch_shape(
+                audio_hidden_states.shape)
            audio_positions = self.audio_patchifier.get_patch_grid_bounds(
-                audio_shape, device=audio_hidden_states.device
-            )
-
+                audio_shape, device=audio_hidden_states.device)
+            audio_latents = self.audio_patchifier.patchify(
+                audio_hidden_states)
            audio_modality = Modality(
                enabled=True,
                latent=audio_latents,
-                timesteps=sharded_audio_timestep,
+                timesteps=audio_timestep,
                positions=audio_positions,
                context=audio_encoder_hidden_states,
                context_mask=audio_encoder_attention_mask,
@@ -2360,16 +1775,10 @@ class LTX2Transformer3DModel(CachableDiT):
                    f"latent_head={audio_head} "
                    f"latent_checksum={audio_checksum:.6f}"
                )
-
-        # Run transformer with attention masks
        video_out, audio_out = self.model(
            video=video_modality,
            audio=audio_modality,
-            video_attention_mask=video_attention_mask,
-            audio_attention_mask=audio_attention_mask,
        )
-
-        # Denoised prediction
        if video_out is not None and video_modality is not None:
            video_out = _to_denoised(
                video_modality.latent,
@@ -2382,20 +1791,6 @@ class LTX2Transformer3DModel(CachableDiT):
                audio_out,
                audio_modality.timesteps,
            )
-
-        # Gather and unpad video output
-        if sp_world_size > 1 and video_out is not None:
-            video_out = sequence_model_parallel_all_gather_with_unpad(
-                video_out, video_original_seq_len, dim=1
-            )
-
-        # Gather and unpad audio output
-        if sp_world_size > 1 and audio_out is not None:
-            audio_out = sequence_model_parallel_all_gather_with_unpad(
-                audio_out, audio_original_seq_len, dim=1
-            )
-
-        # Unpatchify
        video_out = self.patchifier.unpatchify(
            video_out, output_shape=video_shape)
        if audio_out is None or audio_shape is None:
@@ -2403,6 +1798,3 @@ class LTX2Transformer3DModel(CachableDiT):
        audio_out = self.audio_patchifier.unpatchify(
            audio_out, output_shape=audio_shape)
        return video_out, audio_out
-
-# Entry point for model registry
-EntryClass = LTX2Transformer3DModel
@@ -9,6 +9,3 @@ __all__ = [
    "CausalMatrixGameTransformerBlock",
    "ActionModule",
 ]
-
-# Entry point for model registry
-EntryClass = [MatrixGameWanModel, CausalMatrixGameWanModel]
@@ -9,32 +9,20 @@ import torch.nn as nn
 from fastvideo.attention import DistributedAttention
 from fastvideo.configs.models.dits.matrixgame import MatrixGameWanVideoConfig
 from fastvideo.distributed.parallel_state import get_sp_world_size
-from fastvideo.layers.layernorm import (
-    FP32LayerNorm,
-    LayerNormScaleShift,
-    RMSNorm,
-    ScaleResidual,
-    ScaleResidualLayerNormScaleShift,
-)
+from fastvideo.layers.layernorm import (FP32LayerNorm, LayerNormScaleShift,
+                                        RMSNorm, ScaleResidual,
+                                        ScaleResidualLayerNormScaleShift)
 from fastvideo.layers.linear import ReplicatedLinear
 from fastvideo.layers.mlp import MLP
-from fastvideo.layers.rotary_embedding import (
-    _apply_rotary_emb,
-    get_rotary_pos_embed,
-)
-from fastvideo.layers.visual_embedding import (
-    PatchEmbed,
-    TimestepEmbedder,
-    ModulateProjection,
-)
+from fastvideo.layers.rotary_embedding import (_apply_rotary_emb,
+                                               get_rotary_pos_embed)
+from fastvideo.layers.visual_embedding import PatchEmbed, TimestepEmbedder, ModulateProjection
 from fastvideo.logger import init_logger
 from fastvideo.models.dits.base import BaseDiT
-from fastvideo.models.dits.wanvideo import (
-    WanSelfAttention,
-    WanI2VCrossAttention,
-    WanT2VCrossAttention,
-    WanImageEmbedding,
-)
+from fastvideo.models.dits.wanvideo import (WanSelfAttention,
+                                           WanI2VCrossAttention,
+                                           WanT2VCrossAttention,
+                                           WanImageEmbedding)
 from fastvideo.platforms import AttentionBackendEnum, current_platform

 # Import ActionModule
@@ -53,12 +41,11 @@ class MatrixGameTimeImageEmbedding(nn.Module):
        super().__init__()

        self.time_embedder = TimestepEmbedder(
-            dim, frequency_embedding_size=time_freq_dim, act_layer="silu"
-        )
-        self.time_modulation = ModulateProjection(
-            dim, factor=6, act_layer="silu"
-        )
-
+            dim, frequency_embedding_size=time_freq_dim, act_layer="silu")
+        self.time_modulation = ModulateProjection(dim,
+                                                  factor=6,
+                                                  act_layer="silu")
+        
        self.image_embedder = None
        if image_embed_dim is not None:
            self.image_embedder = WanImageEmbedding(image_embed_dim, dim)
@@ -66,7 +53,7 @@ class MatrixGameTimeImageEmbedding(nn.Module):
    def forward(
        self,
        timestep: torch.Tensor,
-        encoder_hidden_states: torch.Tensor,  # Kept for interface compatibility
+        encoder_hidden_states: torch.Tensor, # Kept for interface compatibility
        encoder_hidden_states_image: torch.Tensor | None = None,
        timestep_seq_len: int | None = None,
    ):
@@ -74,25 +61,17 @@ class MatrixGameTimeImageEmbedding(nn.Module):
        timestep_proj = self.time_modulation(temb)

        # MatrixGame does not use text embeddings, so we ignore encoder_hidden_states
-
+        
        if encoder_hidden_states_image is not None:
            assert self.image_embedder is not None
            encoder_hidden_states_image = self.image_embedder(
-                encoder_hidden_states_image
-            )
+                encoder_hidden_states_image)

-        encoder_hidden_states = torch.zeros(
-            (timestep.shape[0], 0, temb.shape[-1]),
-            device=temb.device,
-            dtype=temb.dtype,
-        )
+        encoder_hidden_states = torch.zeros((timestep.shape[0], 0, temb.shape[-1]),
+                                            device=temb.device,
+                                            dtype=temb.dtype)

-        return (
-            temb,
-            timestep_proj,
-            encoder_hidden_states,
-            encoder_hidden_states_image,
-        )
+        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image


 class MatrixGameCrossAttention(WanSelfAttention):
@@ -108,7 +87,7 @@ class MatrixGameCrossAttention(WanSelfAttention):

        # compute query, key, value
        q = self.norm_q(self.to_q(x)[0]).view(b, -1, n, d)
-
+        
        if crossattn_cache is not None:
            if not crossattn_cache["is_init"]:
                crossattn_cache["is_init"] = True
@@ -122,7 +101,7 @@ class MatrixGameCrossAttention(WanSelfAttention):
        else:
            k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
            v = self.to_v(context)[0].view(b, -1, n, d)
-
+        
        # compute attention
        x = self.attn(q, k, v)

@@ -133,20 +112,19 @@ class MatrixGameCrossAttention(WanSelfAttention):


 class MatrixGameTransformerBlock(nn.Module):
-    def __init__(
-        self,
-        dim: int,
-        ffn_dim: int,
-        num_heads: int,
-        qk_norm: str = "rms_norm_across_heads",
-        cross_attn_norm: bool = False,
-        eps: float = 1e-6,
-        added_kv_proj_dim: int | None = None,
-        supported_attention_backends: tuple[AttentionBackendEnum, ...]
-        | None = None,
-        prefix: str = "",
-        action_config: dict | None = None,
-    ):
+    
+    def __init__(self,
+                 dim: int,
+                 ffn_dim: int,
+                 num_heads: int,
+                 qk_norm: str = "rms_norm_across_heads",
+                 cross_attn_norm: bool = False,
+                 eps: float = 1e-6,
+                 added_kv_proj_dim: int | None = None,
+                 supported_attention_backends: tuple[AttentionBackendEnum, ...]
+                 | None = None,
+                 prefix: str = "",
+                 action_config: dict | None = None):
        super().__init__()
        action_config = action_config or {}

@@ -162,8 +140,7 @@ class MatrixGameTransformerBlock(nn.Module):
            head_size=dim // num_heads,
            causal=False,
            supported_attention_backends=supported_attention_backends,
-            prefix=f"{prefix}.attn1",
-        )
+            prefix=f"{prefix}.attn1")
        self.hidden_dim = dim
        self.num_attention_heads = num_heads
        dim_head = dim // num_heads
@@ -184,28 +161,28 @@ class MatrixGameTransformerBlock(nn.Module):
            eps=eps,
            elementwise_affine=True,
            dtype=torch.float32,
-            compute_dtype=torch.float32,
-        )
+            compute_dtype=torch.float32)

        # 2. Cross-attention
        if added_kv_proj_dim is not None:
            # I2V
-            self.attn2 = WanI2VCrossAttention(
-                dim, num_heads, qk_norm=qk_norm, eps=eps
-            )
+            self.attn2 = WanI2VCrossAttention(dim,
+                                              num_heads,
+                                              qk_norm=qk_norm,
+                                              eps=eps)
        else:
            # T2V
-            self.attn2 = WanT2VCrossAttention(
-                dim, num_heads, qk_norm=qk_norm, eps=eps
-            )
+            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,
-        )
+            compute_dtype=torch.float32)

        # 2.1. Action Module Integration
        self.use_action_module = len(action_config) > 0
@@ -227,7 +204,7 @@ class MatrixGameTransformerBlock(nn.Module):
        temb: torch.Tensor,
        freqs_cis: tuple[torch.Tensor, torch.Tensor],
        # Action Module specific args
-        grid_sizes: torch.Tensor,
+        grid_sizes: torch.Tensor | None = None,
        mouse_cond: torch.Tensor | None = None,
        keyboard_cond: torch.Tensor | None = None,
    ) -> torch.Tensor:
@@ -237,16 +214,9 @@ class MatrixGameTransformerBlock(nn.Module):

        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()).chunk(
-                6, dim=2
-            )
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+                self.scale_shift_table.unsqueeze(0) + temb.float()
+            ).chunk(6, dim=2)
            # batch_size, seq_len, 1, inner_dim
            shift_msa = shift_msa.squeeze(2)
            scale_msa = scale_msa.squeeze(2)
@@ -257,20 +227,13 @@ class MatrixGameTransformerBlock(nn.Module):
        else:
            # temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B)
            e = self.scale_shift_table + temb.float()
-            (
-                shift_msa,
-                scale_msa,
-                gate_msa,
-                c_shift_msa,
-                c_scale_msa,
-                c_gate_msa,
-            ) = e.chunk(6, dim=1)
+            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.float()) *
+                              (1 + scale_msa) + shift_msa).to(orig_dtype)
        query, _ = self.to_q(norm_hidden_states)
        key, _ = self.to_k(norm_hidden_states)
        value, _ = self.to_v(norm_hidden_states)
@@ -286,10 +249,9 @@ class MatrixGameTransformerBlock(nn.Module):

        # Apply rotary embeddings
        cos, sin = freqs_cis
-        query, key = (
-            _apply_rotary_emb(query, cos, sin, is_neox_style=False),
-            _apply_rotary_emb(key, cos, sin, is_neox_style=False),
-        )
+        query, key = _apply_rotary_emb(query, cos, sin,
+                                       is_neox_style=False), _apply_rotary_emb(
+                                           key, cos, sin, is_neox_style=False)

        attn_output, _ = self.attn1(query, key, value)
        attn_output = attn_output.flatten(2)
@@ -298,24 +260,18 @@ class MatrixGameTransformerBlock(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),
-        )
+            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, context_lens=None
-        )
+        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),
-        )
+            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)

        # ================= Action Module =================
        if self.action_model is not None:
@@ -324,12 +280,9 @@ class MatrixGameTransformerBlock(nn.Module):
                # ActionModule implementation takes hidden_states directly
                hidden_states = self.action_model(
                    hidden_states,
-                    int(grid_sizes[0]),
-                    int(grid_sizes[1]),
-                    int(grid_sizes[2]),
-                    mouse_cond,
-                    keyboard_cond,
-                    num_frame_per_block=int(grid_sizes[0]),
+                    grid_sizes[0], grid_sizes[1], grid_sizes[2],
+                    mouse_cond, keyboard_cond,
+                    num_frame_per_block=grid_sizes[0],
                )
        # =================================================

@@ -340,7 +293,6 @@ class MatrixGameTransformerBlock(nn.Module):

        return hidden_states

-
 _DEFAULT_MATRIXGAME_CONFIG = MatrixGameWanVideoConfig()


@@ -350,23 +302,15 @@ class MatrixGameWanModel(BaseDiT):

    _fsdp_shard_conditions = _DEFAULT_MATRIXGAME_CONFIG._fsdp_shard_conditions
    _compile_conditions = _DEFAULT_MATRIXGAME_CONFIG._compile_conditions
-    _supported_attention_backends = (
-        _DEFAULT_MATRIXGAME_CONFIG._supported_attention_backends
-    )
+    _supported_attention_backends = _DEFAULT_MATRIXGAME_CONFIG._supported_attention_backends
    param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.param_names_mapping
-    reverse_param_names_mapping = (
-        _DEFAULT_MATRIXGAME_CONFIG.reverse_param_names_mapping
-    )
-    lora_param_names_mapping = (
-        _DEFAULT_MATRIXGAME_CONFIG.lora_param_names_mapping
-    )
+    reverse_param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.reverse_param_names_mapping
+    lora_param_names_mapping = _DEFAULT_MATRIXGAME_CONFIG.lora_param_names_mapping

-    def __init__(
-        self,
-        config: MatrixGameWanVideoConfig,
-        hf_config: dict[str, Any],
-        **kwargs,
-    ) -> None:
+    def __init__(self,
+                 config: MatrixGameWanVideoConfig,
+                 hf_config: dict[str, Any],
+                 **kwargs) -> None:
        super().__init__(config=config, hf_config=hf_config)

        inner_dim = config.num_attention_heads * config.attention_head_dim
@@ -376,12 +320,10 @@ class MatrixGameWanModel(BaseDiT):
        self.patch_size = config.patch_size

        # 1. Patch & position embedding
-        self.patch_embedding = PatchEmbed(
-            in_chans=config.in_channels,
-            embed_dim=inner_dim,
-            patch_size=config.patch_size,
-            flatten=False,
-        )
+        self.patch_embedding = PatchEmbed(in_chans=config.in_channels,
+                                          embed_dim=inner_dim,
+                                          patch_size=config.patch_size,
+                                          flatten=False)

        # 2. Condition embeddings
        self.condition_embedder = MatrixGameTimeImageEmbedding(
@@ -391,73 +333,57 @@ class MatrixGameWanModel(BaseDiT):
        )

        # 2.1. Get action config
-        self.action_config = getattr(config, "action_config", {})
+        self.action_config = getattr(config, 'action_config', {})

        # 3. Transformer blocks
-        self.blocks = nn.ModuleList(
-            [
-                MatrixGameTransformerBlock(
-                    inner_dim,
-                    config.ffn_dim,
-                    config.num_attention_heads,
-                    config.qk_norm,
-                    config.cross_attn_norm,
-                    config.eps,
-                    config.added_kv_proj_dim,
-                    self._supported_attention_backends,
-                    prefix=f"{getattr(config, 'prefix', 'Wan')}.blocks.{i}",
-                    action_config=self.action_config,
-                )
-                for i in range(config.num_layers)
-            ]
-        )
+        self.blocks = nn.ModuleList([
+            MatrixGameTransformerBlock(
+                inner_dim,
+                config.ffn_dim,
+                config.num_attention_heads,
+                config.qk_norm,
+                config.cross_attn_norm,
+                config.eps,
+                config.added_kv_proj_dim,
+                self._supported_attention_backends,
+                prefix=f"{getattr(config, 'prefix', 'Wan')}.blocks.{i}",
+                action_config=self.action_config)
+            for i in range(config.num_layers)
+        ])

        # 4. Output norm & projection
-        self.norm_out = LayerNormScaleShift(
-            inner_dim,
-            norm_type="layer",
-            eps=config.eps,
-            elementwise_affine=False,
-            dtype=torch.float32,
-            compute_dtype=torch.float32,
-        )
+        self.norm_out = LayerNormScaleShift(inner_dim,
+                                            norm_type="layer",
+                                            eps=config.eps,
+                                            elementwise_affine=False,
+                                            dtype=torch.float32,
+                                            compute_dtype=torch.float32)
        self.proj_out = nn.Linear(
-            inner_dim, config.out_channels * math.prod(config.patch_size)
-        )
+            inner_dim, config.out_channels * math.prod(config.patch_size))
        self.scale_shift_table = nn.Parameter(
-            torch.randn(1, 2, inner_dim) / inner_dim**0.5
-        )
-
+            torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+ 
        self.gradient_checkpointing = False

-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: torch.Tensor | list[torch.Tensor],
-        timestep: torch.LongTensor,
-        encoder_hidden_states_image: torch.Tensor
-        | list[torch.Tensor]
-        | None = None,
-        # Action inputs
-        mouse_cond: torch.Tensor | None = None,
-        keyboard_cond: torch.Tensor | None = None,
-        **kwargs,
-    ) -> torch.Tensor:
-        if encoder_hidden_states is not None and not isinstance(
-            encoder_hidden_states, torch.Tensor
-        ):
+    def forward(self,
+                hidden_states: torch.Tensor,
+                encoder_hidden_states: torch.Tensor | list[torch.Tensor],
+                timestep: torch.LongTensor,
+                encoder_hidden_states_image: torch.Tensor
+                | list[torch.Tensor] | None = None,
+                # Action inputs
+                mouse_cond: torch.Tensor | None = None,
+                keyboard_cond: torch.Tensor | None = None,
+                **kwargs) -> torch.Tensor:
+        if encoder_hidden_states is not None and not isinstance(encoder_hidden_states, torch.Tensor):
            encoder_hidden_states = encoder_hidden_states[0]
-        if (
-            isinstance(encoder_hidden_states_image, list)
-            and len(encoder_hidden_states_image) > 0
-        ):
+        if isinstance(encoder_hidden_states_image,
+                      list) and len(encoder_hidden_states_image) > 0:
            encoder_hidden_states_image = encoder_hidden_states_image[0]
        else:
            encoder_hidden_states_image = None

-        batch_size, num_channels, num_frames, height, width = (
-            hidden_states.shape
-        )
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
        p_t, p_h, p_w = self.patch_size
        post_patch_num_frames = num_frames // p_t
        post_patch_height = height // p_h
@@ -467,25 +393,18 @@ class MatrixGameWanModel(BaseDiT):
        d = self.hidden_size // self.num_attention_heads
        rope_dim_list = [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)]
        freqs_cos, freqs_sin = get_rotary_pos_embed(
-            (
-                post_patch_num_frames * get_sp_world_size(),
-                post_patch_height,
-                post_patch_width,
-            ),
+            (post_patch_num_frames * get_sp_world_size(), post_patch_height,
+             post_patch_width),
            self.hidden_size,
            self.num_attention_heads,
            rope_dim_list,
            dtype=torch.float32 if current_platform.is_mps() else torch.float64,
            rope_theta=10000,
-            do_sp_sharding=True,
-        )
+            do_sp_sharding=True)
        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.float(),
+                     freqs_sin.float()) if freqs_cos is not None else None

        hidden_states = self.patch_embedding(hidden_states)
        hidden_states = hidden_states.flatten(2).transpose(1, 2)
@@ -493,14 +412,8 @@ class MatrixGameWanModel(BaseDiT):
        if timestep.dim() == 2:
            timestep = timestep.flatten()

-        (
-            temb,
-            timestep_proj,
-            encoder_hidden_states,
-            encoder_hidden_states_image,
-        ) = self.condition_embedder(
-            timestep, encoder_hidden_states, encoder_hidden_states_image
-        )
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image)
        timestep_proj = timestep_proj.unflatten(1, (6, -1))

        if encoder_hidden_states is not None:
@@ -516,30 +429,25 @@ class MatrixGameWanModel(BaseDiT):
        if encoder_hidden_states_image is not None:
            if encoder_hidden_states is not None:
                encoder_hidden_states = torch.concat(
-                    [encoder_hidden_states_image, encoder_hidden_states], dim=1
-                )
+                    [encoder_hidden_states_image, encoder_hidden_states], dim=1)
            else:
                encoder_hidden_states = encoder_hidden_states_image

        # This is [F, H, W] for the ActionModule
-        grid_sizes = torch.tensor(
-            [post_patch_num_frames, post_patch_height, post_patch_width],
-            device=hidden_states.device,
-        )
+        grid_sizes = torch.tensor([
+            post_patch_num_frames, post_patch_height, post_patch_width
+        ],
+                                 device=hidden_states.device)

        # Blocks
        for block in self.blocks:
            if torch.is_grad_enabled() and self.gradient_checkpointing:
                hidden_states = self._gradient_checkpointing_func(
-                    block,
-                    hidden_states,
-                    encoder_hidden_states,
-                    timestep_proj,
+                    block, hidden_states, encoder_hidden_states, timestep_proj,
                    freqs_cis,
                    grid_sizes=grid_sizes,
                    mouse_cond=mouse_cond,
-                    keyboard_cond=keyboard_cond,
-                )
+                    keyboard_cond=keyboard_cond)
            else:
                hidden_states = block(
                    hidden_states,
@@ -548,27 +456,19 @@ class MatrixGameWanModel(BaseDiT):
                    freqs_cis,
                    grid_sizes=grid_sizes,
                    mouse_cond=mouse_cond,
-                    keyboard_cond=keyboard_cond,
-                )
+                    keyboard_cond=keyboard_cond)

        # Output
-        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(
-            2, dim=1
-        )
+        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2,
+                                                                          dim=1)
        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.reshape(batch_size, post_patch_num_frames,
+                                              post_patch_height,
+                                              post_patch_width, p_t, p_h, p_w,
+                                              -1)
        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

-        return output
+        return output
@@ -685,6 +685,3 @@ class StepVideoModel(BaseDiT):

        output = rearrange(output, '(b f) c h w -> b c f h w', f=frame)
        return output
-
-# Entry point for model registry
-EntryClass = StepVideoModel
@@ -0,0 +1,5 @@
+from .model import WanGameActionTransformer3DModel
+
+__all__ = [
+    "WanGameActionTransformer3DModel",
+]
@@ -0,0 +1,231 @@
+import torch
+import torch.nn as nn
+
+from fastvideo.layers.visual_embedding import TimestepEmbedder, ModulateProjection, timestep_embedding
+from fastvideo.platforms import AttentionBackendEnum
+from fastvideo.attention import DistributedAttention
+from fastvideo.forward_context import set_forward_context
+from fastvideo.models.dits.wanvideo import WanImageEmbedding
+
+from fastvideo.models.dits.hyworld.camera_rope import prope_qkv
+from fastvideo.layers.rotary_embedding import _apply_rotary_emb
+from fastvideo.layers.mlp import MLP
+
+class WanGameActionTimeImageEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        time_freq_dim: int,
+        image_embed_dim: int | None = None,
+    ):
+        super().__init__()
+
+        self.time_freq_dim = time_freq_dim
+        self.time_embedder = TimestepEmbedder(
+            dim, frequency_embedding_size=time_freq_dim, act_layer="silu")
+        self.time_modulation = ModulateProjection(dim,
+                                                  factor=6,
+                                                  act_layer="silu")
+        
+        self.image_embedder = None
+        if image_embed_dim is not None:
+            self.image_embedder = WanImageEmbedding(image_embed_dim, dim)
+
+        self.action_embedder = MLP(
+            time_freq_dim,
+            dim,
+            dim,
+            bias=True,
+            act_type="silu"
+        )
+        # Initialize with zeros for residual-like behavior
+        nn.init.zeros_(self.action_embedder.fc_out.weight)
+        if self.action_embedder.fc_out.bias is not None:
+            nn.init.zeros_(self.action_embedder.fc_out.bias)
+
+    def forward(
+        self,
+        timestep: torch.Tensor,
+        action: torch.Tensor,
+        encoder_hidden_states: torch.Tensor, # Kept for interface compatibility
+        encoder_hidden_states_image: torch.Tensor | None = None,
+        timestep_seq_len: int | None = None,
+    ):
+        temb = self.time_embedder(timestep, timestep_seq_len)
+
+        action_emb = timestep_embedding(action.flatten(), self.time_freq_dim)
+        action_embedder_dtype = next(iter(self.action_embedder.parameters())).dtype
+        if (
+            action_emb.dtype != action_embedder_dtype
+            and action_embedder_dtype != torch.int8
+        ):
+            action_emb = action_emb.to(action_embedder_dtype)
+        action_emb = self.action_embedder(action_emb).type_as(temb)
+        temb = temb + action_emb
+
+        timestep_proj = self.time_modulation(temb)
+
+        # MatrixGame does not use text embeddings, so we ignore encoder_hidden_states
+        
+        if encoder_hidden_states_image is not None:
+            assert self.image_embedder is not None
+            encoder_hidden_states_image = self.image_embedder(
+                encoder_hidden_states_image)
+
+        encoder_hidden_states = torch.zeros((timestep.shape[0], 0, temb.shape[-1]),
+                                            device=temb.device,
+                                            dtype=temb.dtype)
+
+        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image
+
+class WanGameActionSelfAttention(nn.Module):
+    """
+    Self-attention module with support for:
+    - Standard RoPE-based attention
+    - Camera PRoPE-based attention (when viewmats and Ks are provided)
+    - KV caching for autoregressive generation
+    """
+
+    def __init__(self,
+                 dim: int,
+                 num_heads: int,
+                 local_attn_size: int = -1,
+                 sink_size: int = 0,
+                 qk_norm=True,
+                 eps=1e-6) -> None:
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.local_attn_size = local_attn_size
+        self.sink_size = sink_size
+        self.qk_norm = qk_norm
+        self.eps = eps
+        self.max_attention_size = 32760 if local_attn_size == -1 else local_attn_size * 1560
+
+        # Scaled dot product attention (using DistributedAttention for SP support)
+        self.attn = DistributedAttention(
+            num_heads=num_heads,
+            head_size=self.head_dim,
+            softmax_scale=None,
+            causal=False,
+            supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN,
+                                          AttentionBackendEnum.TORCH_SDPA))
+
+    def forward(self,
+                q: torch.Tensor,
+                k: torch.Tensor,
+                v: torch.Tensor,
+                freqs_cis: tuple[torch.Tensor, torch.Tensor],
+                kv_cache: dict | None = None,
+                current_start: int = 0,
+                cache_start: int | None = None,
+                viewmats: torch.Tensor | None = None,
+                Ks: torch.Tensor | None = None,
+                is_cache: bool = False,
+                attention_mask: torch.Tensor | None = None):
+        """
+        Forward pass with camera PRoPE attention combining standard RoPE and projective positional encoding.
+        
+        Args:
+            q, k, v: Query, key, value tensors [B, L, num_heads, head_dim]
+            freqs_cis: RoPE frequency cos/sin tensors
+            kv_cache: KV cache dict (may have None values for training)
+            current_start: Current position for KV cache
+            cache_start: Cache start position
+            viewmats: Camera view matrices for PRoPE [B, cameras, 4, 4]
+            Ks: Camera intrinsics for PRoPE [B, cameras, 3, 3]
+            is_cache: Whether to store to KV cache (for inference)
+            attention_mask: Attention mask [B, L] (1 = attend, 0 = mask)
+        """
+        if cache_start is None:
+            cache_start = current_start
+
+        # Apply RoPE manually
+        cos, sin = freqs_cis
+        query_rope = _apply_rotary_emb(q, cos, sin, is_neox_style=False).type_as(v)
+        key_rope = _apply_rotary_emb(k, cos, sin, is_neox_style=False).type_as(v)
+        value_rope = v
+
+        # Get PRoPE transformed q, k, v
+        query_prope, key_prope, value_prope, apply_fn_o = prope_qkv(
+            q.transpose(1, 2),  # [B, num_heads, L, head_dim]
+            k.transpose(1, 2),
+            v.transpose(1, 2),
+            viewmats=viewmats,
+            Ks=Ks,
+            patches_x=40,  # hardcoded for now
+            patches_y=22,
+        )
+        # PRoPE returns [B, num_heads, L, head_dim], convert to [B, L, num_heads, head_dim]
+        query_prope = query_prope.transpose(1, 2)
+        key_prope = key_prope.transpose(1, 2)
+        value_prope = value_prope.transpose(1, 2)
+
+        # KV cache handling
+        if kv_cache is not None:
+            cache_key = kv_cache.get("k", None)
+            cache_value = kv_cache.get("v", None)
+
+            if cache_value is not None and not is_cache:
+                cache_key_rope, cache_key_prope = cache_key.chunk(2, dim=-1)
+                cache_value_rope, cache_value_prope = cache_value.chunk(2, dim=-1)
+
+                key_rope = torch.cat([cache_key_rope, key_rope], dim=1)
+                value_rope = torch.cat([cache_value_rope, value_rope], dim=1)
+                key_prope = torch.cat([cache_key_prope, key_prope], dim=1)
+                value_prope = torch.cat([cache_value_prope, value_prope], dim=1)
+
+            if is_cache:
+                # Store to cache (update input dict directly)
+                kv_cache["k"] = torch.cat([key_rope, key_prope], dim=-1)
+                kv_cache["v"] = torch.cat([value_rope, value_prope], dim=-1)
+
+        # Concatenate rope and prope paths (matching original)
+        query_all = torch.cat([query_rope, query_prope], dim=0)
+        key_all = torch.cat([key_rope, key_prope], dim=0)
+        value_all = torch.cat([value_rope, value_prope], dim=0)
+
+        # Check if Q and KV have different sequence lengths (KV cache mode)
+        # In this case, use LocalAttention (supports different Q/KV lengths)
+        if query_all.shape[1] != key_all.shape[1]:
+            raise ValueError("Q and KV have different sequence lengths")
+            # KV cache mode: Q has new tokens only, KV has cached + new tokens
+            # Use LocalAttention which supports different Q/KV lengths
+            # LocalAttention will use the appropriate backend (SageAttn, FlashAttn, etc.)
+            if not hasattr(self, '_kv_cache_attn'):
+                from fastvideo.attention import LocalAttention
+                self._kv_cache_attn = LocalAttention(
+                    num_heads=self.num_heads,
+                    head_size=self.head_dim,
+                    causal=False,
+                    supported_attention_backends=(AttentionBackendEnum.SAGE_ATTN,
+                                                  AttentionBackendEnum.FLASH_ATTN,
+                                                  AttentionBackendEnum.TORCH_SDPA)
+                )
+            hidden_states_all = self._kv_cache_attn(query_all, key_all, value_all)
+        else:
+            # Same sequence length: use DistributedAttention (supports SP)
+            # Create default attention mask if not provided
+            if attention_mask is None:
+                batch_size, seq_len = q.shape[0], q.shape[1]
+                attention_mask = torch.ones(batch_size, seq_len, device=q.device, dtype=q.dtype)
+        
+            if q.dtype == torch.float32:
+                from fastvideo.attention.backends.sdpa import SDPAMetadataBuilder
+                attn_metadata_builder = SDPAMetadataBuilder
+            else:
+                from fastvideo.attention.backends.flash_attn import FlashAttnMetadataBuilder
+                attn_metadata_builder = FlashAttnMetadataBuilder
+            attn_metadata = attn_metadata_builder().build(
+                current_timestep=0,
+                attn_mask=attention_mask,
+            )
+            with set_forward_context(current_timestep=0, attn_metadata=attn_metadata):
+                hidden_states_all, _ = self.attn(query_all, key_all, value_all, attention_mask=attention_mask)
+
+        hidden_states_rope, hidden_states_prope = hidden_states_all.chunk(2, dim=0)
+        hidden_states_prope = apply_fn_o(hidden_states_prope.transpose(1, 2)).transpose(1, 2)
+
+        return hidden_states_rope, hidden_states_prope
@@ -0,0 +1,424 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import math
+from typing import Any
+
+import torch
+import torch.nn as nn
+
+from fastvideo.configs.models.dits.wangamevideo import WanGameVideoConfig
+from fastvideo.distributed.parallel_state import get_sp_world_size
+from fastvideo.layers.layernorm import (FP32LayerNorm, LayerNormScaleShift,
+                                        RMSNorm, ScaleResidual,
+                                        ScaleResidualLayerNormScaleShift)
+from fastvideo.layers.linear import ReplicatedLinear
+from fastvideo.layers.mlp import MLP
+from fastvideo.layers.rotary_embedding import (_apply_rotary_emb,
+                                               get_rotary_pos_embed)
+from fastvideo.layers.visual_embedding import PatchEmbed
+from fastvideo.logger import init_logger
+from fastvideo.models.dits.base import BaseDiT
+from fastvideo.models.dits.wanvideo import WanI2VCrossAttention
+from fastvideo.platforms import AttentionBackendEnum, current_platform
+
+# Import ActionModule
+from fastvideo.models.dits.wangame.hyworld_action_module import WanGameActionTimeImageEmbedding, WanGameActionSelfAttention
+
+logger = init_logger(__name__)
+
+
+class WanGameCrossAttention(WanI2VCrossAttention):
+    def forward(self, x, context, context_lens=None):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+            context_lens(Tensor): Shape [B]
+        """
+        context_img = context
+        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_img = self.norm_added_k(self.add_k_proj(context_img)[0]).view(
+            b, -1, n, d)
+        v_img = self.add_v_proj(context_img)[0].view(b, -1, n, d)
+        img_x = self.attn(q, k_img, v_img)
+
+        # output
+        x = img_x.flatten(2)
+        x, _ = self.to_out(x)
+        return x
+
+class WanGameActionTransformerBlock(nn.Module):
+    """
+    Transformer block for WAN Action model with support for:
+    - Self-attention with RoPE and camera PRoPE
+    - Cross-attention with text/image context
+    - Feed-forward network with AdaLN modulation
+    """
+
+    def __init__(self,
+                 dim: int,
+                 ffn_dim: int,
+                 num_heads: int,
+                 local_attn_size: int = -1,
+                 sink_size: int = 0,
+                 qk_norm: str = "rms_norm_across_heads",
+                 cross_attn_norm: bool = False,
+                 eps: float = 1e-6,
+                 added_kv_proj_dim: int | None = None,
+                 supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None,
+                 prefix: str = ""):
+        super().__init__()
+
+        # 1. Self-attention
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.to_q = ReplicatedLinear(dim, dim, bias=True)
+        self.to_k = ReplicatedLinear(dim, dim, bias=True)
+        self.to_v = ReplicatedLinear(dim, dim, bias=True)
+        self.to_out = ReplicatedLinear(dim, dim, bias=True)
+        
+        self.attn1 = WanGameActionSelfAttention(
+            dim,
+            num_heads,
+            local_attn_size=local_attn_size,
+            sink_size=sink_size,
+            qk_norm=qk_norm,
+            eps=eps)
+        
+        self.hidden_dim = dim
+        self.num_attention_heads = num_heads
+        self.local_attn_size = local_attn_size
+        dim_head = dim // num_heads
+
+        if qk_norm == "rms_norm":
+            self.norm_q = RMSNorm(dim_head, eps=eps)
+            self.norm_k = RMSNorm(dim_head, eps=eps)
+        elif qk_norm == "rms_norm_across_heads":
+            self.norm_q = RMSNorm(dim, eps=eps)
+            self.norm_k = RMSNorm(dim, eps=eps)
+        else:
+            raise ValueError(f"QK Norm type {qk_norm} not supported")
+
+        assert cross_attn_norm is True
+        self.self_attn_residual_norm = ScaleResidualLayerNormScaleShift(
+            dim,
+            norm_type="layer",
+            eps=eps,
+            elementwise_affine=True,
+            compute_dtype=torch.float32)
+
+        # 2. Cross-attention (I2V only for now)
+        self.attn2 = WanGameCrossAttention(dim,
+                                          num_heads,
+                                          qk_norm=qk_norm,
+                                          eps=eps)
+        # norm3 for FFN input 
+        self.norm3 = LayerNormScaleShift(dim, norm_type="layer", eps=eps,
+                                         elementwise_affine=False)
+
+        # 3. Feed-forward
+        self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
+        self.mlp_residual = ScaleResidual()
+
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+        # PRoPE output projection (initialized via add_discrete_action_parameters on the model)
+        self.to_out_prope = nn.ModuleList([
+            nn.Linear(dim, dim, bias=True),
+        ])
+        nn.init.zeros_(self.to_out_prope[0].weight)
+        if self.to_out_prope[0].bias is not None:
+            nn.init.zeros_(self.to_out_prope[0].bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        freqs_cis: tuple[torch.Tensor, torch.Tensor],
+        kv_cache: dict | None = None,
+        crossattn_cache: dict | None = None,
+        current_start: int = 0,
+        cache_start: int | None = None,
+        viewmats: torch.Tensor | None = None,
+        Ks: torch.Tensor | None = None,
+        is_cache: bool = False,
+    ) -> torch.Tensor:
+        if hidden_states.dim() == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        num_frames = temb.shape[1]
+        frame_seqlen = hidden_states.shape[1] // num_frames
+        bs, seq_length, _ = hidden_states.shape
+        orig_dtype = hidden_states.dtype
+
+        # Cast temb to float32 for scale/shift computation
+        e = self.scale_shift_table + temb.float()
+        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)
+
+        # 1. Self-attention
+        norm_hidden_states = (self.norm1(hidden_states.float()).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+            (1 + scale_msa) + shift_msa).to(orig_dtype).flatten(1, 2)
+        
+        query, _ = self.to_q(norm_hidden_states)
+        key, _ = self.to_k(norm_hidden_states)
+        value, _ = self.to_v(norm_hidden_states)
+
+        if self.norm_q is not None:
+            query = self.norm_q.forward_native(query)
+        if self.norm_k is not None:
+            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))
+        value = value.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
+
+        # Self-attention with optional camera PRoPE
+        attn_output_rope, attn_output_prope = self.attn1(
+            query, key, value, freqs_cis,
+            kv_cache, current_start, cache_start, viewmats, Ks,
+            is_cache=is_cache
+        )
+        # Combine rope and prope outputs
+        attn_output_rope = attn_output_rope.flatten(2)
+        attn_output_rope, _ = self.to_out(attn_output_rope)
+        attn_output_prope = attn_output_prope.flatten(2)
+        attn_output_prope = self.to_out_prope[0](attn_output_prope)
+        attn_output = attn_output_rope.squeeze(1) + attn_output_prope.squeeze(1)
+
+        # Self-attention residual + norm in float32
+        null_shift = null_scale = torch.zeros(1, device=hidden_states.device, dtype=torch.float32)
+        norm_hidden_states, hidden_states = self.self_attn_residual_norm(
+            hidden_states.float(), attn_output.float(), gate_msa, null_shift, null_scale)
+        hidden_states = hidden_states.type_as(attn_output)
+        norm_hidden_states = norm_hidden_states.type_as(attn_output)
+
+        # 2. Cross-attention
+        attn_output = self.attn2(norm_hidden_states.to(orig_dtype),
+                                 context=encoder_hidden_states,
+                                 context_lens=None)
+        # Cross-attention residual in bfloat16
+        hidden_states = hidden_states + attn_output
+        
+        # norm3 for FFN input in float32
+        norm_hidden_states = self.norm3(
+            hidden_states.float(), c_shift_msa, c_scale_msa
+        ).type_as(hidden_states)
+
+        # 3. Feed-forward
+        ff_output = self.ffn(norm_hidden_states.to(orig_dtype))
+        hidden_states = self.mlp_residual(hidden_states.float(), ff_output.float(), c_gate_msa)
+        hidden_states = hidden_states.to(orig_dtype)  # Cast back to original dtype
+
+        return hidden_states
+
+class WanGameActionTransformer3DModel(BaseDiT):
+    """
+    WAN Action Transformer 3D Model for video generation with action conditioning.
+    
+    Extends the base WAN video model with:
+    - Action embedding support for controllable generation
+    - camera PRoPE attention for 3D-aware generation
+    - KV caching for autoregressive inference
+    """
+    _fsdp_shard_conditions = WanGameVideoConfig()._fsdp_shard_conditions
+    _compile_conditions = WanGameVideoConfig()._compile_conditions
+    _supported_attention_backends = WanGameVideoConfig()._supported_attention_backends
+    param_names_mapping = WanGameVideoConfig().param_names_mapping
+    reverse_param_names_mapping = WanGameVideoConfig().reverse_param_names_mapping
+    lora_param_names_mapping = WanGameVideoConfig().lora_param_names_mapping
+
+    def __init__(self, config: WanGameVideoConfig, hf_config: dict[str, Any]) -> None:
+        super().__init__(config=config, hf_config=hf_config)
+
+        inner_dim = config.num_attention_heads * config.attention_head_dim
+        self.hidden_size = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_dim = config.attention_head_dim
+        self.in_channels = config.in_channels
+        self.out_channels = config.out_channels
+        self.num_channels_latents = config.num_channels_latents
+        self.patch_size = config.patch_size
+        self.local_attn_size = config.local_attn_size
+        self.inner_dim = inner_dim
+
+        # 1. Patch & position embedding
+        self.patch_embedding = PatchEmbed(in_chans=config.in_channels,
+                                          embed_dim=inner_dim,
+                                          patch_size=config.patch_size,
+                                          flatten=False)
+
+        # 2. Condition embeddings (with action support)
+        self.condition_embedder = WanGameActionTimeImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=config.freq_dim,
+            image_embed_dim=config.image_dim,
+        )
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList([
+            WanGameActionTransformerBlock(
+                inner_dim,
+                config.ffn_dim,
+                config.num_attention_heads,
+                config.local_attn_size,
+                config.sink_size,
+                config.qk_norm,
+                config.cross_attn_norm,
+                config.eps,
+                config.added_kv_proj_dim,
+                supported_attention_backends=self._supported_attention_backends,
+                prefix=f"{config.prefix}.blocks.{i}")
+            for i in range(config.num_layers)
+        ])
+
+        # 4. Output norm & projection
+        self.norm_out = LayerNormScaleShift(inner_dim,
+                                            norm_type="layer",
+                                            eps=config.eps,
+                                            elementwise_affine=False,
+                                            dtype=torch.float32)
+        self.proj_out = nn.Linear(
+            inner_dim, config.out_channels * math.prod(config.patch_size))
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+
+        self.gradient_checkpointing = False
+
+        # Causal-specific
+        self.num_frame_per_block = config.arch_config.num_frames_per_block
+        assert self.num_frame_per_block <= 3
+
+        self.__post_init__()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor | list[torch.Tensor],
+        timestep: torch.LongTensor,
+        encoder_hidden_states_image: torch.Tensor | list[torch.Tensor],
+        guidance=None,
+        action: torch.Tensor | None = None,
+        viewmats: torch.Tensor | None = None,
+        Ks: torch.Tensor | None = None,
+        kv_cache: list[dict] | None = None,
+        crossattn_cache: list[dict] | None = None,
+        current_start: int = 0,
+        cache_start: int = 0,
+        start_frame: int = 0,
+        is_cache: bool = False,
+        **kwargs
+    ) -> torch.Tensor:
+        """
+        Forward pass for both training and inference with KV caching.
+        
+        Args:
+            hidden_states: Video latents [B, C, T, H, W]
+            encoder_hidden_states: Text embeddings [B, L, D]
+            timestep: Timestep tensor
+            encoder_hidden_states_image: Optional image embeddings
+            action: Action tensor [B, T] for per-frame conditioning
+            viewmats: Camera view matrices for PRoPE [B, T, 4, 4]
+            Ks: Camera intrinsics for PRoPE [B, T, 3, 3]
+            kv_cache: KV cache for autoregressive inference (list of dicts per layer)
+            crossattn_cache: Cross-attention cache for inference
+            current_start: Current position for KV cache
+            cache_start: Cache start position
+            start_frame: RoPE offset for new frames in autoregressive mode
+            is_cache: If True, populate KV cache and return early (cache-only mode)
+        """
+        orig_dtype = hidden_states.dtype
+        # if not isinstance(encoder_hidden_states, torch.Tensor):
+        #     encoder_hidden_states = encoder_hidden_states[0]
+        if isinstance(encoder_hidden_states_image, list) and len(encoder_hidden_states_image) > 0:
+            encoder_hidden_states_image = encoder_hidden_states_image[0]
+        # else:
+        #     encoder_hidden_states_image = None
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        # Get rotary embeddings
+        d = self.hidden_size // self.num_attention_heads
+        rope_dim_list = [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)]
+        freqs_cos, freqs_sin = get_rotary_pos_embed(
+            (post_patch_num_frames * get_sp_world_size(), post_patch_height, post_patch_width),
+            self.hidden_size,
+            self.num_attention_heads,
+            rope_dim_list,
+            dtype=torch.float32 if current_platform.is_mps() else torch.float64,
+            rope_theta=10000,
+            start_frame=start_frame
+        )
+        freqs_cos = freqs_cos.to(hidden_states.device)
+        freqs_sin = freqs_sin.to(hidden_states.device)
+        freqs_cis = (freqs_cos, freqs_sin) if freqs_cos is not None else None
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        if timestep.dim() == 2:
+            timestep = timestep.flatten()
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, action, encoder_hidden_states, encoder_hidden_states_image=encoder_hidden_states_image)
+        # Reshape timestep_proj: [T, 6*dim] -> [B, T, 6, dim]
+        # For training: batch_size=1, T=num_frames (diffusion forcing)
+        # For inference: batch_size can vary
+        timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size))
+        if timestep_proj.shape[0] == post_patch_num_frames and batch_size == 1:
+            # Training mode: timestep_proj is [T, 6, dim], add batch dim -> [1, T, 6, dim]
+            timestep_proj = timestep_proj.unsqueeze(0)
+        else:
+            # Inference mode: reshape based on timestep shape
+            timestep_proj = timestep_proj.unflatten(dim=0, sizes=timestep.shape)
+
+        encoder_hidden_states = encoder_hidden_states_image
+
+        # Transformer blocks
+        for block_idx, block in enumerate(self.blocks):
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, timestep_proj, freqs_cis,
+                    kv_cache[block_idx] if kv_cache else None,
+                    crossattn_cache[block_idx] if crossattn_cache else None,
+                    current_start, cache_start,
+                    viewmats, Ks, is_cache)
+            else:
+                hidden_states = block(
+                    hidden_states, encoder_hidden_states, timestep_proj, freqs_cis,
+                    kv_cache[block_idx] if kv_cache else None,
+                    crossattn_cache[block_idx] if crossattn_cache else None,
+                    current_start, cache_start,
+                    viewmats, Ks, is_cache)
+
+        # If cache-only mode, return early
+        if is_cache:
+            return kv_cache
+
+        # Output norm, projection & unpatchify
+        # Reshape temb to match timestep_proj shape: [T, dim] -> [B, T, 1, dim]
+        if temb.shape[0] == post_patch_num_frames and batch_size == 1:
+            # Training mode: temb is [T, dim] -> [1, T, 1, dim]
+            temb = temb.unsqueeze(0).unsqueeze(2)
+        else:
+            # Inference mode: reshape based on timestep shape
+            temb = temb.unflatten(dim=0, sizes=timestep.shape).unsqueeze(2)
+        
+        shift, scale = (self.scale_shift_table.unsqueeze(1) + temb).chunk(2, dim=2)
+        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)
+
+        return output
@@ -871,6 +871,4 @@ class WanTransformer3DModel(CachableDiT):
            return hidden_states + self.previous_residual_even
        else:
            return hidden_states + self.previous_residual_odd
-
-# Entry point for model registry
-EntryClass = WanTransformer3DModel
+        
@@ -631,6 +631,3 @@ class CLIPVisionModel(ImageEncoder):
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params
-
-# Entry point for model registry
-EntryClass = [CLIPTextModel, CLIPVisionModel]
@@ -561,6 +561,3 @@ def _norm_and_concat_padded_batch(
    mask_flattened = mask.reshape(b, t, 1).expand(-1, -1, d * l)
    normed = normed.masked_fill(~mask_flattened, 0.0)
    return normed
-
-# Entry point for model registry
-EntryClass = LTX2GemmaTextEncoderModel
@@ -429,6 +429,3 @@ class LlamaModel(TextEncoder):
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params
-
-# Entry point for model registry
-EntryClass = LlamaModel
@@ -385,6 +385,3 @@ class Qwen2_5_VLTextModel(TextEncoder):
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params
-
-# Entry point for model registry
-EntryClass = Qwen2_5_VLTextModel
@@ -351,5 +351,3 @@ class Reason1TextEncoder(TextEncoder):
            tensor.std(dim=-1, keepdim=True) + 1e-8
        )

-# Entry point for model registry
-EntryClass = Reason1TextEncoder
@@ -417,6 +417,3 @@ class SiglipVisionModel(ImageEncoder):
            loaded_params.add(name)
            
        return loaded_params
-
-# Entry point for model registry
-EntryClass = SiglipVisionModel
@@ -583,6 +583,3 @@ class STEP1TextEncoder(torch.nn.Module):
                                      self.device) if with_mask else None)
            y_mask = txt_tokens.attention_mask
        return y.transpose(0, 1), y_mask
-
-# Entry point for model registry
-EntryClass = STEP1TextEncoder
@@ -719,6 +719,3 @@ class UMT5EncoderModel(TextEncoder):
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params
-
-# Entry point for model registry
-EntryClass = [UMT5EncoderModel, T5EncoderModel]
@@ -79,7 +79,6 @@ class ComponentLoader(ABC):
            "scheduler": (SchedulerLoader, "diffusers"),
            "transformer": (TransformerLoader, "diffusers"),
            "transformer_2": (TransformerLoader, "diffusers"),
-            "transformer_3": (TransformerLoader, "diffusers"),
            "vae": (VAELoader, "diffusers"),
            "audio_vae": (AudioDecoderLoader, "diffusers"),
            "audio_decoder": (AudioDecoderLoader, "diffusers"),
@@ -91,8 +90,6 @@ class ComponentLoader(ABC):
            "image_processor": (ImageProcessorLoader, "transformers"),
            "feature_extractor": (ImageProcessorLoader, "transformers"),
            "image_encoder": (ImageEncoderLoader, "transformers"),
-            "upsampler": (UpsamplerLoader, "diffusers"),
-            "upsampler_2": (UpsamplerLoader, "diffusers"),
        }

        if module_type in module_loaders:
@@ -561,8 +558,7 @@ class VAELoader(ComponentLoader):
    def load(self, model_path: str, fastvideo_args: FastVideoArgs):
        """Load the VAE based on the model path, and inference args."""
        config = get_diffusers_config(model=model_path)
-        class_name = config.pop("_class_name")
-        config.pop("_name_or_path", None)
+        class_name = config.get("_class_name")
        assert class_name is not None, (
            "Model config does not contain a _class_name attribute. Only diffusers format is supported."
        )
@@ -735,7 +731,6 @@ class TransformerLoader(ComponentLoader):
        config = get_diffusers_config(model=model_path)
        hf_config = deepcopy(config)
        cls_name = config.pop("_class_name")
-        config.pop("_name_or_path", None)
        if cls_name is None:
            raise ValueError(
                "Model config does not contain a _class_name attribute. "
@@ -750,7 +745,7 @@ class TransformerLoader(ComponentLoader):
        fastvideo_args.model_paths["transformer"] = model_path

        # Config from Diffusers supersedes fastvideo's model config
-        dit_config = deepcopy(fastvideo_args.pipeline_config.dit_config)
+        dit_config = fastvideo_args.pipeline_config.dit_config
        dit_config.update_model_arch(config)

        model_cls, _ = ModelRegistry.resolve_model_cls(cls_name)
@@ -811,6 +806,10 @@ class TransformerLoader(ComponentLoader):
            cls_name.startswith("Cosmos25")
            or cls_name == "Cosmos25Transformer3DModel"
            or getattr(fastvideo_args.pipeline_config, "prefix", "") == "Cosmos25"
+        ) and not (
+            cls_name.startswith("WanGame")
+            or cls_name == "WanGameActionTransformer3DModel"
+            or getattr(fastvideo_args.pipeline_config, "prefix", "") == "WanGame"
        )
        model = maybe_load_fsdp_model(
            model_cls=model_cls,
@@ -882,50 +881,6 @@ class SchedulerLoader(ComponentLoader):
        return scheduler


-class UpsamplerLoader(ComponentLoader):
-    """Loader for upsamplers."""
-
-    def load(self, model_path: str, fastvideo_args: FastVideoArgs):
-        """Load the upsampler based on the model path, and inference args."""
-        config_dict = get_diffusers_config(model=model_path)
-        class_name = config_dict.pop("_class_name", None)
-
-        if class_name is None:
-            raise ValueError(
-                "Model config does not contain a _class_name attribute. "
-                "Only diffusers format is supported."
-            )
-
-        try:
-            upsampler_cfg = deepcopy(fastvideo_args.pipeline_config.upsampler_config[0])
-            upsampler_cfg.update_model_config(config_dict)
-        except Exception as e:
-            upsampler_cfg = deepcopy(fastvideo_args.pipeline_config.upsampler_config[1])
-            upsampler_cfg.update_model_config(config_dict)
-
-        model_cls, _ = ModelRegistry.resolve_model_cls(class_name)
-        model = model_cls(upsampler_cfg)
-
-        target_device = get_local_torch_device()
-        model = model.to(target_device, dtype=PRECISION_TO_TYPE[fastvideo_args.pipeline_config.upsampler_precision])
-
-        # Find all safetensors files
-        safetensors_list = glob.glob(
-            os.path.join(str(model_path), "*.safetensors"))
-        if not safetensors_list:
-            raise ValueError(f"No safetensors files found in {model_path}")
-        
-        if len(safetensors_list) == 1:
-            loaded = safetensors_load_file(safetensors_list[0])
-        else:
-            loaded = {}
-            for sf_file in safetensors_list:
-                loaded.update(safetensors_load_file(sf_file))
-        
-        model.load_state_dict(loaded, strict=True)
-
-        return model.eval()
-
 class GenericComponentLoader(ComponentLoader):
    """Generic loader for components that don't have a specific loader."""

@@ -138,7 +138,7 @@ def maybe_load_fsdp_model(

    weight_iterator = safetensors_weights_iterator(weight_dir_list)
    param_names_mapping_fn = get_param_names_mapping(model.param_names_mapping)
-    load_model_from_full_model_state_dict(
+    incompatible_keys, unexpected_keys = load_model_from_full_model_state_dict(
        model,
        weight_iterator,
        device,
@@ -147,6 +147,9 @@ def maybe_load_fsdp_model(
        cpu_offload=cpu_offload,
        param_names_mapping=param_names_mapping_fn,
    )
+    if incompatible_keys or unexpected_keys:
+        logger.warning("Incompatible keys: %s", incompatible_keys)
+        logger.warning("Unexpected keys: %s", unexpected_keys)
    for n, p in chain(model.named_parameters(), model.named_buffers()):
        if p.is_meta:
            raise RuntimeError(
@@ -340,7 +343,7 @@ def load_model_from_full_model_state_dict(
                       unused_keys)

    # List of allowed parameter name patterns
-    ALLOWED_NEW_PARAM_PATTERNS = ["gate_compress", "proj_l"]  # Can be extended as needed
+    ALLOWED_NEW_PARAM_PATTERNS = ["gate_compress", "proj_l", "to_out_prope", "action_embedder"]  # Can be extended as needed
    for new_param_name in unused_keys:
        if not any(pattern in new_param_name
                   for pattern in ALLOWED_NEW_PARAM_PATTERNS):
@@ -1,7 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # Adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/model_executor/models/registry.py

-import ast
 import importlib
 import os
 import pickle
@@ -43,6 +42,7 @@ _IMAGE_TO_VIDEO_DIT_MODELS = {
    # "HunyuanVideoTransformer3DModel": ("dits", "hunyuanvideo", "HunyuanVideoDiT"),
    "WanTransformer3DModel": ("dits", "wanvideo", "WanTransformer3DModel"),
    "CausalWanTransformer3DModel": ("dits", "causal_wanvideo", "CausalWanTransformer3DModel"),
+    "WanGameActionTransformer3DModel": ("dits", "wangame", "WanGameActionTransformer3DModel"),
    "MatrixGameWanModel": ("dits", "matrixgame", "MatrixGameWanModel"),
    "CausalMatrixGameWanModel": ("dits", "matrixgame", "CausalMatrixGameWanModel"),
 }
@@ -99,12 +99,7 @@ _SCHEDULERS = {
    ("schedulers", "scheduling_rcm", "RCMScheduler"),
 }

-_UPSAMPLERS = {
-    "SRTo720pUpsampler": ("upsamplers", "hunyuan15", "SRTo720pUpsampler"),
-    "SRTo1080pUpsampler": ("upsamplers", "hunyuan15", "SRTo1080pUpsampler"),
-}
-
-_LEGACY_FAST_VIDEO_MODELS = {
+_FAST_VIDEO_MODELS = {
    **_TEXT_TO_VIDEO_DIT_MODELS,
    **_IMAGE_TO_VIDEO_DIT_MODELS,
    **_TEXT_ENCODER_MODELS,
@@ -112,102 +107,8 @@ _LEGACY_FAST_VIDEO_MODELS = {
    **_VAE_MODELS,
    **_AUDIO_MODELS,
    **_SCHEDULERS,
-    **_UPSAMPLERS,
 }

-MODELS_PATH = os.path.dirname(__file__)
-
-
-@lru_cache(maxsize=None)
-def _discover_and_register_models() -> dict[str, tuple[str, str, str]]:
-    discovered_models: dict[str, tuple[str, str, str]] = {}
-    for root, dirs, files in os.walk(MODELS_PATH):
-        dirs[:] = [
-            d for d in dirs
-            if not d.startswith(".") and d != "__pycache__"
-        ]
-
-        for filename in files:
-            if not filename.endswith(".py"):
-                continue
-
-            filepath = os.path.join(root, filename)
-            try:
-                with open(filepath, "r", encoding="utf-8") as f:
-                    source = f.read()
-                tree = ast.parse(source, filename=filename)
-
-                entry_class_node = None
-                first_class_def = None
-
-                for node in ast.walk(tree):
-                    if isinstance(node, ast.Assign):
-                        for target in node.targets:
-                            if isinstance(target, ast.Name) and target.id == "EntryClass":
-                                entry_class_node = node
-                                break
-                    if first_class_def is None and isinstance(node, ast.ClassDef):
-                        first_class_def = node
-
-                if not entry_class_node or not first_class_def:
-                    continue
-
-                model_cls_name_list: list[str] = []
-                value_node = entry_class_node.value
-
-                if isinstance(value_node, ast.Name):
-                    model_cls_name_list.append(value_node.id)
-                elif isinstance(value_node, (ast.List, ast.Tuple)):
-                    for elt in value_node.elts:
-                        if isinstance(elt, ast.Constant) and isinstance(
-                                elt.value, str):
-                            model_cls_name_list.append(elt.value)
-                        elif isinstance(elt, ast.Name):
-                            model_cls_name_list.append(elt.id)
-
-                if not model_cls_name_list:
-                    continue
-
-                rel_dir = os.path.relpath(root, MODELS_PATH)
-                if rel_dir == ".":
-                    continue
-
-                rel_parts = rel_dir.split(os.sep)
-                component_name = rel_parts[0]
-                sub_parts = rel_parts[1:]
-
-                if filename == "__init__.py":
-                    mod_relname = ".".join(sub_parts)
-                else:
-                    mod_base = filename[:-3]
-                    mod_relname = ".".join(sub_parts +
-                                           [mod_base]) if sub_parts else mod_base
-
-                for model_cls_str in model_cls_name_list:
-                    if model_cls_str in discovered_models:
-                        logger.warning(
-                            "Duplicate architecture found: %s. Overwriting.",
-                            model_cls_str)
-                    discovered_models[model_cls_str] = (
-                        component_name,
-                        mod_relname,
-                        model_cls_str,
-                    )
-
-            except Exception as e:
-                logger.warning("Could not parse %s to find models: %s",
-                               filepath, e)
-
-    return discovered_models
-
-
-_DISCOVERED_MODELS = _discover_and_register_models()
-_FAST_VIDEO_MODELS = dict(_DISCOVERED_MODELS)
-for model_arch, spec in _LEGACY_FAST_VIDEO_MODELS.items():
-    if model_arch in _FAST_VIDEO_MODELS:
-        continue
-    _FAST_VIDEO_MODELS[model_arch] = spec
-
 _SUBPROCESS_COMMAND = [sys.executable, "-m", "fastvideo.models.dits.registry"]

 _T = TypeVar("_T")
@@ -439,8 +340,7 @@ class _ModelRegistry:
 ModelRegistry = _ModelRegistry({
    model_arch:
    _LazyRegisteredModel(
-        module_name=(f"fastvideo.models.{component_name}.{mod_relname}"
-                     if mod_relname else f"fastvideo.models.{component_name}"),
+        module_name=f"fastvideo.models.{component_name}.{mod_relname}",
        component_name=component_name,
        class_name=cls_name,
    )
@@ -685,6 +685,3 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin,

    def __len__(self) -> int:
        return self.config.num_train_timesteps
-
-# Entry point for model registry
-EntryClass = FlowMatchEulerDiscreteScheduler
@@ -853,6 +853,3 @@ class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):

    def __len__(self):
        return self.config.num_train_timesteps
-
-# Entry point for model registry
-EntryClass = FlowUniPCMultistepScheduler
@@ -321,6 +321,3 @@ class RCMScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):

    def __len__(self) -> int:
        return self.config.num_train_timesteps
-
-# Entry point for model registry
-EntryClass = RCMScheduler
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
JerryZhou54	bdaf2e2e45	Support custom action trajectories for validation	2026-02-03 20:23:13 +00:00
JerryZhou54	f3a7a37312	Overfitting running for MC 10	2026-02-03 05:11:48 +00:00
JerryZhou54	b1abb5c251	Update action_labels	2026-02-02 05:20:06 +00:00
RandNMR73	a80824255e	Add data preprocess pipeline for WanGame	2026-02-02 04:38:24 +00:00