ckpt

[bugfix]: _compile_conditions regression (#1077 )
[misc] readme small fix (#1076 )
2026-02-09 05:50:48 +00:00 · 2026-02-06 18:57:43 -08:00 · 2026-02-06 16:57:48 -08:00 · 2026-02-06 19:43:18 -05:00 · 2026-02-06 13:38:35 -08:00 · 2026-02-05 20:33:08 -08:00
167 changed files with 6760 additions and 9594 deletions
@@ -1,35 +1,29 @@
-<div align="center">
-<img src=assets/logos/logo.svg width="30%"/>
-</div>

-<p align="center">
-     | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start/"><b> Quick Start</b></a> | <a href="https://github.com/hao-ai-lab/FastVideo/discussions/982"  target="_blank"><b>Weekly Dev Meeting</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/sv3MMKyv" target="_blank"> <b> WeChat </b> </a> |
-</p>
+| **[Documentation](https://hao-ai-lab.github.io/FastVideo)** | **[Quick Start](https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start/)** | **[Weekly Dev Meeting](https://github.com/hao-ai-lab/FastVideo/discussions/982)** | 🟣💬 **[Slack](https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ)** |

 **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/).

-<details>
-<summary>More</summary>
+- `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/).

- ```2025/06/14```: 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/).
+### More News

-</details>
+- `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/).

 ## 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 achineve >50x denoising speedup
+  - [Sparse distillation](https://hao-ai-lab.github.io/blogs/fastvideo_post_training/) to achieve >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.
@@ -44,6 +38,7 @@ 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
@@ -58,18 +53,20 @@ 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)      |
-| [FastWan2.1-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)  |
+| 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) |

 ## 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
@@ -108,55 +105,36 @@ 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/).

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

 - [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/)
-<!-- - [Finetuning Guide](https://hao-ai-lab.github.io/FastVideo/training/finetune.html) -->
+- [Contribution Guide](https://hao-ai-lab.github.io/FastVideo/contributing/overview/)

 ## 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. [![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)
+- [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.

 ## 🤝 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 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.
+## Acknowledgement
+
+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.

 ## Citation
-If you find FastVideo useful, please considering citing our work:
+
+If you find FastVideo useful, please consider citing our research 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/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp310-cp310-linux_x86_64.whl
+    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

 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/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp311-cp311-linux_x86_64.whl
+    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

 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/v0.5.4/flash_attn-2.8.3%2Bcu128torch2.9-cp312-cp312-linux_x86_64.whl
+    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

 COPY . .

@@ -121,6 +121,15 @@ 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/configs/pipelines/registry.py` and
-  `fastvideo/configs/sample/registry.py`.
+- Register them in `fastvideo/registry.py` using explicit
+  `register_configs(...)` blocks (this file is the single source of truth now).

 ### 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)
@@ -51,8 +51,9 @@ runtime parameters consistent:
 - `fastvideo/configs/pipelines/`: pipeline wiring and required components.
 - `fastvideo/configs/sample/`: default sampling parameters (steps, frames,
  guidance scale, resolution, fps).
- `fastvideo/configs/registry.py`: pipeline registry.
- `fastvideo/configs/sample/registry.py`: sampling param registry.
+- `fastvideo/registry.py`: unified registry for pipeline config + sampling
+  defaults and model metadata resolution, defined via explicit
+  `register_configs(...)` blocks (no separate dict registries).

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

 ### Registering Models

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

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

-_VAE_MODELS = {
-    "YourVAEModel": ("vaes", "yourvae", "YourVAEClass"),
-}
+# 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(),
+    ],
+)
 ```

 ## Step 2: Directory Structure
@@ -103,7 +103,7 @@ python setup.py install  # or pip install -e .

 **`SAGE_ATTN_THREE`**

-[SageAttention 3](https://huggingface.co/jt-zhang/SageAttention3) is an advanced attention mechanism that leverages FP4 quantization and Blackwell GPU Tensor Cores for significant performance improvements.
+[SageAttention 3](https://github.com/thu-ml/SageAttention/tree/main/sageattention3_blackwell) is an advanced attention mechanism that leverages FP4 quantization and Blackwell GPU Tensor Cores for significant performance improvements.

 #### Hardware Requirements

@@ -113,11 +113,7 @@ python setup.py install  # or pip install -e .

 Note that Sage Attention 3 requires `python>=3.13`, `torch>=2.8.0`, `CUDA >=12.8`. If you are using `uv` and using `torch==2.8.0` make sure that `sentencepiece==0.2.1` in the pyproject.toml file.

-To use Sage Attention 3 in FastVideo, first get access to the SageAttention3 code, then move `sageattn/` and `setup.py` to the directory `fastvideo/attention/backends`, then install from using:
-
-```bash
-python setup.py install
-```
+To use Sage Attention 3 in FastVideo, follow the `README.md` in the linked repository to install the package from source.

 ## Teacache

@@ -1,172 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=dmd_3333
-#SBATCH --partition=main
-#SBATCH --nodes=4
-#SBATCH --ntasks=4
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=dmd_3333_output/dmd_3333_%j.out
-#SBATCH --error=dmd_3333_output/dmd_3333_%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"
-
-# Configs
-NUM_GPUS=8
-NUM_TOTAL_GPUS=32
-
-# Model paths for Self-Forcing DMD distillation with Hunyuan1.5:
-GENERATOR_MODEL_PATH="weizhou03/SFHunyuanVideo-1.5-Diffusers-480p_t2v"  # Updated to Wan2.2
-REAL_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v" # Teacher model
-FAKE_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v"
-# REAL_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled"
-# FAKE_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled" # Critic model
-
-# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
-# DATA_DIR="/mnt/weka/home/hao.zhang/matthew/FastVideo/data/test-text-preprocessing"
-DATA_DIR="/mnt/weka/home/hao.zhang/wl/sharefs/wl/release/vidprom_16k_text_embed"
-DATA_DIR_2="/mnt/weka/home/hao.zhang/wl/sharefs/wl/release/fv-ode-preprocessing-16k-hy15-121"
-# DATA_DIR=data/crush-smol_processed_t2v/combined_parquet_dataset
-# DATA_DIR="/mnt/sharefs/users/hao.zhang/Vchitect-2M/Wan-Syn-upload/latents_i2v/train/"
-# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
-VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
-
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-training_args=(
-  --tracker_project_name SFhy1.5_t2v_distill_self_forcing_dmd  # Updated for Wan2.2
-  --output_dir "/mnt/weka/home/hao.zhang/wei/SFhy1.5_distill_3333_1e-5_1e-5_cfg6_corrected_scheduler"
-  --wandb_run_name "self_forcing_3333_context_forcing"
-  --max_train_steps 4000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 848
-  --enable_gradient_checkpointing_type "full"
-  --simulate_generator_forward
-  --log_visualization
-  --num_frames 81
-  --num_frame_per_block 3  # Frame generation block size for self-forcing
-  # --resume-from-checkpoint "/mnt/weka/home/hao.zhang/wei/SFhy1.5_distill_1333/checkpoint-300"
-  --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wei/hy15_worldplay_tf_init_3333/checkpoint-2800/transformer/diffusion_pytorch_model.safetensors
-  # --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wei/hy15_worldplay_df_init_1333/checkpoint-11600/transformer/diffusion_pytorch_model.safetensors
-  # --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wei/hy15_distilled_ode_init_3333/checkpoint-2400/transformer/diffusion_pytorch_model.safetensors
-)
-
-parallel_args=(
-  --num_gpus $NUM_TOTAL_GPUS # 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim $NUM_TOTAL_GPUS
-)
-
-model_args=(
-  --model_path $GENERATOR_MODEL_PATH
-  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
-  --real_score_model_path $REAL_SCORE_MODEL_PATH
-  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
-)
-
-dataset_args=(
-  --data_path "$DATA_DIR"
-  # --data_path_2 "$DATA_DIR_2"
-  --dataloader_num_workers 4
-)
-
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 50
-  --validation_sampling_steps "8"
-  --validation_guidance_scale "1.0" # not used for dmd inference
-  --text-encoder-cpu-offload
-  # --vae_cpu_offload True
-)
-
-optimizer_args=(
-  --optimizer-type "muon"
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --training_state_checkpointing_steps 100
-  --weight_only_checkpointing_steps 500
-  --weight_decay 0.01
-  --betas '0.0,0.999'
-  --max_grad_norm 1.0
-)
-
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.0
-  --dit_precision "fp32"
-  --flow_shift 5
-  --seed 1000
-  --use_ema True
-  --ema_decay 0.99
-  --ema_start_step 200
-)
-
-dmd_args=(
-  --warp_denoising_step
-  # --dmd_denoising_steps '1000,875,750,625,500,375,250,125'
-  # --dmd_denoising_steps '1000,760,520,280'
-  # --dmd_denoising_steps '1000,750,500,250'
-  --min_timestep_ratio 0.02
-  --max_timestep_ratio 0.98
-  --dfake_gen_update_ratio 5
-  --real_score_guidance_scale 3.5
-  --fake_score_learning_rate 5e-6
-  --fake_score_betas '0.0,0.999'
-)
-
-self_forcing_args=(
-  --independent_first_frame False  # Whether to treat first frame independently
-  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
-  --last_step_only False  # Whether to only use the last denoising step
-  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
-  # --use-ode-init True
-  # --use-context-forcing True
-)
-
-srun torchrun \
--nnodes $SLURM_JOB_NUM_NODES \
--nproc_per_node $NUM_GPUS \
--node_rank $SLURM_PROCID \
--rdzv_backend=c10d \
--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/hy15_self_forcing_distillation_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}" \
-    "${self_forcing_args[@]}"
@@ -1,172 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=distill_dmd_t2v_1.3b
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=distill_dmd_t2v_1.3b_output/distill_dmd_t2v_1.3b_%j.out
-#SBATCH --error=distill_dmd_t2v_1.3b_output/distill_dmd_t2v_1.3b_%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
-
-export HF_HOME=/mnt/data/huggingface
-export HF_DATASETS_CACHE="/tmp/hf_datasets_${SLURM_JOB_ID}_${SLURM_PROCID}"
-mkdir -p "${HF_DATASETS_CACHE}"
-export CUDA_HOME=/usr/local/cuda
-export PATH="$CUDA_HOME/bin:$PATH"
-export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/home/shared-bin/local/usr/lib/aarch64-linux-gnu:$CUDA_HOME/lib64"
-export CUDACXX="$CUDA_HOME/bin/nvcc"
-
-source ~/miniconda3/bin/activate
-conda activate fastvideo
-
-# Configs
-NUM_GPUS=8
-NUM_TOTAL_GPUS=64
-
-# Model paths for Self-Forcing DMD distillation:
-GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
-FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
-
-DATA_DIR="/mnt/data/vidprom_filtered_extended_umt5_text_embed"
-# DATA_DIR="/mnt/data/mixkit_wan_1.3b_processed_t2v_distributed"
-DATA_DIR_2="/mnt/data/mixkit_wan_1.3b_processed_t2v_distributed"
-VALIDATION_DATASET_FILE="/mnt/home/zhouw.jerry2017/FastVideo/examples/distill/SFWan2.1-T2V/validation_64.json"
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-training_args=(
-  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd
-  --wandb_run_name causal_forcing_vidprom_extended
-  --output_dir /mnt/data/wan_cf_df_distill_1.3b
-  --max_train_steps 4000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --enable_gradient_checkpointing_type "full"
-  --log_visualization
-  --simulate_generator_forward
-  --num_frames 81
-  --num_frame_per_block 3  # Frame generation block size for self-forcing
-  --enable_gradient_masking
-  --gradient_mask_last_n_frames 21
-)
-
-parallel_args=(
-  --num_gpus $NUM_TOTAL_GPUS # 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim $NUM_TOTAL_GPUS
-)
-
-model_args=(
-  --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
-  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
-  --real_score_model_path $REAL_SCORE_MODEL_PATH
-  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
-)
-
-dataset_args=(
-  --data_path "$DATA_DIR"
-  # --data_path_2 "$DATA_DIR_2"
-  --dataloader_num_workers 4
-)
-
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 100
-  --validation_sampling_steps "4"
-  --validation_guidance_scale "1.0"
-)
-
-optimizer_args=(
-  --optimizer-type "adamw"
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --training_state_checkpointing_steps 500
-  --weight_only_checkpointing_steps 500
-  --weight_decay 0.01
-  --betas '0.0,0.999'
-  --max_grad_norm 1.0
-)
-
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.0
-  --dit_precision "fp32"
-  --flow_shift 5
-  --seed 1000
-  --use_ema True
-  --ema_decay 0.99
-  --ema_start_step 200
-#   --init_weights_from_safetensors "/mnt/data/wan_tf_ode_init_3333/checkpoint-1600/transformer/diffusion_pytorch_model.safetensors"
-#   --init_weights_from_safetensors "/mnt/data/wan_tf_init_3333/checkpoint-800/transformer/diffusion_pytorch_model.safetensors"
-  # --init_weights_from_safetensors "/mnt/data/Causal-Forcing-wan1.3b/causal_ode.safetensors"
-  # --init_weights_from_safetensors "/mnt/data/wan_ar_diffusion_3333_mixkit_shift5/checkpoint-7200/transformer/diffusion_pytorch_model.safetensors"
-  --init_weights_from_safetensors "/mnt/data/wan_ar_diffusion_3333_mixkit_shift5/checkpoint-4800/transformer/diffusion_pytorch_model.safetensors"
-  # --init_weights_from_safetensors "/mnt/data/wan_tf_ode_init_3333_mixkit_shift5/checkpoint-2000/transformer/diffusion_pytorch_model.safetensors"
-  # --init_weights_from_safetensors "/mnt/data/wan_ar_df_diffusion_3333_mixkit_shift5/checkpoint-4800/transformer/diffusion_pytorch_model.safetensors"
-)
-
-dmd_args=(
-  --dmd_denoising_steps '1000,750,500,250'
-  --min_timestep_ratio 0.02
-  --max_timestep_ratio 0.98
-  --dfake_gen_update_ratio 5
-  --real_score_guidance_scale 3.0
-  --fake_score_learning_rate 2e-6
-  --fake_score_betas '0.0,0.999'
-  --warp_denoising_step
-  # --use-decoupled-dmd True
-  # --use-gan-loss True
-)
-
-self_forcing_args=(
-  --independent_first_frame False  # Whether to treat first frame independently
-  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
-  --last_step_only False  # Whether to only use the last denoising step
-  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
-)
-
-srun torchrun \
--nnodes $SLURM_JOB_NUM_NODES \
--nproc_per_node $NUM_GPUS \
--node_rank $SLURM_PROCID \
--rdzv_backend=c10d \
--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}" \
-    "${self_forcing_args[@]}"
@@ -15,7 +15,7 @@ def main():
        use_fsdp_inference=False, # set to True if GPU is out of memory
        dit_cpu_offload=False,
        vae_cpu_offload=False,
-        text_encoder_cpu_offload=False,
+        text_encoder_cpu_offload=True,
        pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
        # image_encoder_cpu_offload=False,
    )
@@ -0,0 +1,51 @@
+# 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()
+
+
@@ -1,4 +1,5 @@
 from fastvideo import VideoGenerator
+from fastvideo.configs.sample import SamplingParam


 def main():
@@ -15,19 +16,27 @@ 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."
    )

-    video = generator.generate_video(
+    generator.generate_video(
        prompt,
-        negative_prompt="",
-        height=704,
-        width=1280,
-        num_frames=77,
-        num_inference_steps=35,
-        guidance_scale=7.0,
-        fps=24,
+        sampling_param=sampling_param,
        output_path="outputs_video/cosmos2_5_t2w.mp4",
        save_video=True,
    )
@@ -0,0 +1,54 @@
+# 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()
@@ -0,0 +1,43 @@
+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()
@@ -1,165 +1,11 @@
-# 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

@@ -169,7 +15,7 @@ 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='video_samples_hyworld', help="Output video path")
+    parser.add_argument("--output_path", type=str, default=OUTPUT_PATH, 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")
@@ -185,8 +31,7 @@ def main():

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

    # Generate video
+    # The pose string is automatically converted to camera matrices by the pipeline
    print("\nGenerating video...")
-    start_time = time.time()
-    video = generator.generate_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="",
@@ -211,13 +57,9 @@ def main():
        height=HEIGHT,
        width=WIDTH,
        seed=args.seed,
-        pose=args.pose,
    )
-    elapsed = time.time() - start_time

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


 if __name__ == "__main__":
@@ -10,13 +10,12 @@ def main():
    # If a local path is provided, FastVideo will make a best effort
    # attempt to identify the optimal arguments.
    generator = VideoGenerator.from_pretrained(
-        "weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers",
-        # "IRMChen/Wan2.1-Fun-1.3B-Control-Diffusers",
+        "IRMChen/Wan2.1-Fun-1.3B-Control-Diffusers",
        # "alibaba-pai/Wan2.2-Fun-A14B-Control",
        # 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=False, # DiT need to be offloaded for MoE
+        dit_cpu_offload=True, # DiT need to be offloaded for MoE
        vae_cpu_offload=False,
        text_encoder_cpu_offload=True,
        # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer
@@ -24,19 +23,14 @@ def main():
        # image_encoder_cpu_offload=False,
    )

-    # prompt = "一位年轻女性穿着一件粉色的连衣裙，裙子上有白色的装饰和粉色的纽扣。她的头发是紫色的，头上戴着一个红色的大蝴蝶结，显得非常可爱和精致。她还戴着一个红色的领结，整体造型充满了少女感和活力。她的表情温柔，双手轻轻交叉放在身前，姿态优雅。背景是简单的灰色，没有任何多余的装饰，使得人物更加突出。她的妆容清淡自然，突显了她的清新气质。整体画面给人一种甜美、梦幻的感觉，仿佛置身于童话世界中。"
-    # negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
+    prompt = "一位年轻女性穿着一件粉色的连衣裙，裙子上有白色的装饰和粉色的纽扣。她的头发是紫色的，头上戴着一个红色的大蝴蝶结，显得非常可爱和精致。她还戴着一个红色的领结，整体造型充满了少女感和活力。她的表情温柔，双手轻轻交叉放在身前，姿态优雅。背景是简单的灰色，没有任何多余的装饰，使得人物更加突出。她的妆容清淡自然，突显了她的清新气质。整体画面给人一种甜美、梦幻的感觉，仿佛置身于童话世界中。"
+    negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
    # prompt                  = "A young woman with beautiful, clear eyes and blonde hair stands in the forest, wearing a white dress and a crown. Her expression is serene, reminiscent of a movie star, with fair and youthful skin. Her brown long hair flows in the wind. The video quality is very high, with a clear view. High quality, masterpiece, best quality, high resolution, ultra-fine, fantastical."
    # negative_prompt         = "Twisted body, limb deformities, text captions, comic, static, ugly, error, messy code."
-    # image_path = "https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/wan_fun/asset_Wan2_2/v1.0/8.png"
-    # control_video_path = "https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/wan_fun/asset_Wan2_2/v1.0/pose.mp4"
-    import json, os
-    with open("/mnt/weka/home/hao.zhang/wei/FastVideo/data/mixkit_i2v.jsonl", "r") as f:
-        prompt_image_pairs = json.load(f)
-    for d in prompt_image_pairs:
-        prompt = d["prompt"]
-        image_path = os.path.join("/mnt/weka/home/hao.zhang/wei/FastVideo", d["image_path"])
-        video = generator.generate_video(prompt, image_path=image_path, output_path=OUTPUT_PATH, output_video_name=OUTPUT_NAME, save_video=True)
+    image_path = "https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/wan_fun/asset_Wan2_2/v1.0/8.png"
+    control_video_path = "https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/wan_fun/asset_Wan2_2/v1.0/pose.mp4"
+
+    video = generator.generate_video(prompt, negative_prompt=negative_prompt, image_path=image_path, video_path=control_video_path, output_path=OUTPUT_PATH, output_video_name=OUTPUT_NAME, save_video=True)

 if __name__ == "__main__":
    main()
@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator

-OUTPUT_PATH = "video_samples_wan2_2_5B_ti2v_720p_77"
+OUTPUT_PATH = "video_samples_wan2_2_5B_ti2v"
 def main():
    # FastVideo will automatically use the optimal default arguments for the
    # model.
@@ -12,29 +12,17 @@ def main():
        # 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=False,
+        dit_cpu_offload=True,
        vae_cpu_offload=False,
        text_encoder_cpu_offload=True,
        pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
        # image_encoder_cpu_offload=False,
    )

-    import json, os
-    with open("/mnt/weka/home/hao.zhang/wei/FastVideo/data/mixkit_i2v_720p.jsonl", "r") as f:
-        prompt_image_pairs = json.load(f)
-    for d in prompt_image_pairs:
-        prompt = d["prompt"]
-        image_path = os.path.join("/mnt/weka/home/hao.zhang/wei/FastVideo", d["image_path"])
-        video = generator.generate_video(prompt, image_path=image_path, output_path=OUTPUT_PATH, output_video_name=OUTPUT_PATH, save_video=True, num_frames=77)
-    return
-
    # I2V is triggered just by passing in an image_path argument
    prompt = "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside."
    image_path = "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG"
-    # prompt = "In the midst of the joyous New Year's Eve celebration, the cheerful group of friends, their spirits lifted by the festivities, decides to immortalize the moment with a vibrant snapshot"
-    # image_path = "images/mixkit-a-cheerful-group-of-friends-celebrate-new-years-eve-and-51525.png"
    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, image_path=image_path)
-    return

    # Generate another video with a different prompt, without reloading the
    # model!
@@ -4,7 +4,7 @@ export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
 export TOKENIZERS_PARALLELISM=false

-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
 DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init/"
 VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
 NUM_GPUS=1
@@ -14,7 +14,6 @@ NUM_GPUS=1
 training_args=(
  --tracker_project_name "wan_ode_init"
  --output_dir "wan_ode_init_crush_smol"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
  --wandb_run_name "wan_ode_init_crush_smol"
  --max_train_steps 6000
  --train_batch_size 1
@@ -34,7 +33,7 @@ parallel_args=(
  --sp_size 1
  --tp_size 1
  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 1
+  --hsdp_shard_dim $NUM_GPUS
 )

 # Model arguments
@@ -51,20 +50,17 @@ dataset_args=(

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

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

@@ -1,124 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=hy15_ode_3333
-#SBATCH --partition=main
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=hy15_ode_3333_output/hy15_ode_3333_%j.out
-#SBATCH --error=hy15_ode_3333_output/hy15_ode_3333_%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='your_wandb_api_key_here'
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
-
-# Configs
-NUM_GPUS=8
-NUM_TOTAL_GPUS=64
-
-MODEL_PATH="weizhou03/SFHunyuanVideo-1.5-Diffusers-480p_t2v"
-DATA_DIR="your_data_dir"
-VALIDATION_DATASET_FILE="your_validation_data_dir"
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "hy15_ode_init"
-  --output_dir "your_output_dir"
-  --wandb_run_name "your_wandb_run_name"
-  --warp_denoising_step
-  --dmd_denoising_steps '1000,760,520,280,0'
-  --log_visualization
-  --visualization_steps 100
-  --max_train_steps 4000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 2
-  --num_latent_t 21
-  --num_height 480
-  --num_width 848
-  --num_frames 81
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_TOTAL_GPUS # 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim $NUM_TOTAL_GPUS
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-  --text-encoder-cpu-offload
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 4
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --training_state_checkpointing_steps 400
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-validation_args=(
-  # --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 10
-  --validation_sampling_steps "4"
-  --validation_guidance_scale "1.0" # not used for dmd inference
-  # --vae_cpu_offload True
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 6
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
--nnodes $SLURM_JOB_NUM_NODES \
--nproc_per_node $NUM_GPUS \
--node_rank $SLURM_PROCID \
--rdzv_backend=c10d \
--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
@@ -1,133 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=hy15_tf_3333
-#SBATCH --partition=main
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=hy15_tf_3333_output/hy15_tf_3333_%j.out
-#SBATCH --error=hy15_tf_3333_output/hy15_tf_3333_%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"
-
-# Configs
-NUM_GPUS=8
-NUM_TOTAL_GPUS=64
-
-MODEL_PATH="weizhou03/SFHunyuanVideo-1.5-Diffusers-480p_t2v"
-DATA_DIR="/mnt/weka/home/hao.zhang/wl/sharefs/wl/release/fv-ode-preprocessing-16k-hy15-121"
-VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "hy15_ode_init"
-  --output_dir "/mnt/weka/home/hao.zhang/wei/hy15_worldplay_tf_init_3333"
-  --wandb_run_name "hy15_worldplay_tf_init_3333"
-  --warp_denoising_step
-  --dmd_denoising_steps '1000,760,520,280,0'
-  --log_visualization
-  --visualization_steps 100
-  --max_train_steps 30000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 848
-  --num_frames 81
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_TOTAL_GPUS # 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim $NUM_TOTAL_GPUS
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-  --text-encoder-cpu-offload
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 4
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --optimizer-type "muon"
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --training_state_checkpointing_steps 400
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-validation_args=(
-  # --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 10
-  --validation_sampling_steps "4"
-  --validation_guidance_scale "1.0" # not used for dmd inference
-  # --vae_cpu_offload True
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 6
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
--nnodes $SLURM_JOB_NUM_NODES \
--nproc_per_node $NUM_GPUS \
--node_rank $SLURM_PROCID \
--rdzv_backend=c10d \
--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/hy15_ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
@@ -1,137 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=wan_tf_3333
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=wan_tf_3333_output/wan_tf_3333_%j.out
-#SBATCH --error=wan_tf_3333_output/wan_tf_3333_%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
-
-export HF_HOME=/mnt/data/huggingface
-export CUDA_HOME=/usr/local/cuda
-export PATH="$CUDA_HOME/bin:$PATH"
-export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/home/shared-bin/local/usr/lib/aarch64-linux-gnu:$CUDA_HOME/lib64"
-export CUDACXX="$CUDA_HOME/bin/nvcc"
-
-source ~/miniconda3/bin/activate
-conda activate fastvideo
-
-# Configs
-NUM_GPUS=8
-NUM_TOTAL_GPUS=64
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/data/mixkit_wan_1.3b_processed_t2v_distributed"
-VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_tf_init"
-  --output_dir "/mnt/data/wan_ar_df_diffusion_3333_mixkit_shift5"
-  --wandb_run_name "wan_df_init_3333"
-  --warp_denoising_step
-  --dmd_denoising_steps '1000,750,500,250,0'
-  --log_visualization
-  --visualization_steps 100
-  --max_train_steps 30000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 81
-  --override-transformer-cls-name "CausalWanTransformer3DModel"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_TOTAL_GPUS # 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim $NUM_TOTAL_GPUS
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-  --text-encoder-cpu-offload
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 4
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --optimizer-type "adamw"
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --training_state_checkpointing_steps 800
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-validation_args=(
-  # --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 10
-  --validation_sampling_steps "4"
-  --validation_guidance_scale "1.0" # not used for dmd inference
-  # --vae_cpu_offload True
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 6
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
--nnodes $SLURM_JOB_NUM_NODES \
--nproc_per_node $NUM_GPUS \
--node_rank $SLURM_PROCID \
--rdzv_backend=c10d \
--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ar_diffusion_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
@@ -1,138 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=wan_tf_3333
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=wan_tf_3333_output/wan_tf_3333_%j.out
-#SBATCH --error=wan_tf_3333_output/wan_tf_3333_%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
-
-export HF_HOME=/mnt/data/huggingface
-export CUDA_HOME=/usr/local/cuda
-export PATH="$CUDA_HOME/bin:$PATH"
-export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/home/shared-bin/local/usr/lib/aarch64-linux-gnu:$CUDA_HOME/lib64"
-export CUDACXX="$CUDA_HOME/bin/nvcc"
-
-source ~/miniconda3/bin/activate
-conda activate fastvideo
-
-# Configs
-NUM_GPUS=8
-NUM_TOTAL_GPUS=64
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/data/mixkit_wan_1.3b_processed_t2v_distributed"
-VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_tf_init"
-  --output_dir "/mnt/data/wan_ar_diffusion_3333_mixkit_shift5"
-  --wandb_run_name "wan_tf_init_3333"
-  --warp_denoising_step
-  --dmd_denoising_steps '1000,750,500,250,0'
-  --log_visualization
-  --visualization_steps 100
-  --max_train_steps 30000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 81
-  --override-transformer-cls-name "CausalWanTransformer3DModel"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_TOTAL_GPUS # 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim $NUM_TOTAL_GPUS
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-  --text-encoder-cpu-offload
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 4
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --optimizer-type "adamw"
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --training_state_checkpointing_steps 800
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-validation_args=(
-  # --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 10
-  --validation_sampling_steps "4"
-  --validation_guidance_scale "1.0" # not used for dmd inference
-  # --vae_cpu_offload True
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 6
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-  --use-tf True
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
--nnodes $SLURM_JOB_NUM_NODES \
--nproc_per_node $NUM_GPUS \
--node_rank $SLURM_PROCID \
--rdzv_backend=c10d \
--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ar_diffusion_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
@@ -1,138 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=wan_tf_ode_init_3333
-#SBATCH --nodes=8
-#SBATCH --ntasks=8
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=wan_tf_ode_init_3333_output/wan_tf_ode_init_3333_%j.out
-#SBATCH --error=wan_tf_ode_init_3333_output/wan_tf_ode_init_3333_%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
-
-export HF_HOME=/mnt/data/huggingface
-export CUDA_HOME=/usr/local/cuda
-export PATH="$CUDA_HOME/bin:$PATH"
-export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/home/shared-bin/local/usr/lib/aarch64-linux-gnu:$CUDA_HOME/lib64"
-export CUDACXX="$CUDA_HOME/bin/nvcc"
-
-source ~/miniconda3/bin/activate
-conda activate fastvideo
-
-# Configs
-NUM_GPUS=8
-NUM_TOTAL_GPUS=64
-
-MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/data/fv-tf-ode-preprocessing-6k-mixkit-wan-1.3b/"
-VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_tf_ode_init"
-  --output_dir "/mnt/data/wan_tf_ode_init_3333_mixkit_shift5"
-  --wandb_run_name "wan_tf_ode_init_3333_mixkit_shift5"
-  --warp_denoising_step
-  --dmd_denoising_steps '1000,750,500,250'
-  --log_visualization
-  --visualization_steps 100
-  --max_train_steps 30000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 81
-  --enable_gradient_checkpointing_type "full"
-  --init_weights_from_safetensors "/mnt/data/wan_ar_diffusion_3333_mixkit_shift5/checkpoint-2400/transformer/diffusion_pytorch_model.safetensors"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_TOTAL_GPUS # 64
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim $NUM_TOTAL_GPUS
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-  --text-encoder-cpu-offload
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 4
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --optimizer-type "adamw"
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --training_state_checkpointing_steps 500
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-validation_args=(
-  # --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 10
-  --validation_sampling_steps "4"
-  --validation_guidance_scale "1.0" # not used for dmd inference
-  # --vae_cpu_offload True
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 6
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
--nnodes $SLURM_JOB_NUM_NODES \
--nproc_per_node $NUM_GPUS \
--node_rank $SLURM_PROCID \
--rdzv_backend=c10d \
--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
@@ -1 +0,0 @@
-The raw mixkit dataset is downloaded and preprocessed following instructions from https://github.com/tianweiy/CausVid/tree/master
@@ -1,25 +0,0 @@
-#!/bin/bash
-
-GPU_NUM=1 # 2,4,8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-MODEL_TYPE="wan"
-DATA_MERGE_PATH="/mnt/data/mixkit/merge.txt"
-OUTPUT_DIR="/mnt/data/mixkit_wan_1.3b_processed_t2v/"
-
-torchrun --nproc_per_node=$GPU_NUM \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
-    --model_path $MODEL_PATH \
-    --data_merge_path $DATA_MERGE_PATH \
-    --preprocess_video_batch_size 1 \
-    --seed 42 \
-    --max_height 480 \
-    --max_width 832 \
-    --num_frames 81 \
-    --dataloader_num_workers 0 \
-    --output_dir=$OUTPUT_DIR \
-    --train_fps 16 \
-    --samples_per_file 8 \
-    --flush_frequency 8 \
-    --video_length_tolerance_range 5 \
-    --drop_short_ratio 0.0 \
-    --preprocess_task "t2v" 
@@ -1,66 +0,0 @@
-#!/bin/bash
-#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=preprocess_mixkit_wan_1.3b_t2v/preprocess_wan_data_t2v.log
-#SBATCH --error=preprocess_mixkit_wan_1.3b_t2v/preprocess_wan_data_t2v.error
-#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
-
-export HF_HOME=/mnt/data/huggingface
-export CUDA_HOME=/usr/local/cuda
-export PATH="$CUDA_HOME/bin:$PATH"
-export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/home/shared-bin/local/usr/lib/aarch64-linux-gnu:$CUDA_HOME/lib64"
-export CUDACXX="$CUDA_HOME/bin/nvcc"
-
-source ~/miniconda3/bin/activate
-conda activate fastvideo
-
-GPU_NUM=1 # 2,4,8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-MODEL_TYPE="wan"
-DATA_MERGE_PATH="/mnt/data/mixkit/merge.txt"
-OUTPUT_DIR="/mnt/data/mixkit_wan_1.3b_processed_t2v/"
-
-srun torchrun \
--nnodes $SLURM_JOB_NUM_NODES \
--nproc_per_node $GPU_NUM \
--node_rank $SLURM_PROCID \
--rdzv_backend=c10d \
--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
-    --model_path $MODEL_PATH \
-    --data_merge_path $DATA_MERGE_PATH \
-    --preprocess_video_batch_size 1 \
-    --seed 42 \
-    --max_height 480 \
-    --max_width 832 \
-    --num_frames 81 \
-    --dataloader_num_workers 1 \
-    --output_dir=$OUTPUT_DIR \
-    --train_fps 16 \
-    --samples_per_file 8 \
-    --flush_frequency 8 \
-    --video_length_tolerance_range 5 \
-    --drop_short_ratio 0.0 \
-    --preprocess_task "t2v" 
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0

 import torch
-from fastvideo.attention.backends.sageattn.api import sageattn_blackwell
+from sageattn3 import sageattn3_blackwell

 from fastvideo.attention.backends.abstract import (AttentionBackend,
                                                   AttentionImpl,
@@ -67,6 +67,6 @@ class SageAttention3Impl(AttentionImpl):
        query = query.transpose(1, 2)
        key = key.transpose(1, 2)
        value = value.transpose(1, 2)
-        output = sageattn_blackwell(query, key, value, is_causal=self.causal)
+        output = sageattn3_blackwell(query, key, value, is_causal=self.causal)
        output = output.transpose(1, 2)
        return output
@@ -5,6 +5,7 @@ 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",
@@ -14,4 +15,5 @@ __all__ = [
    "LTX2AudioEncoderConfig",
    "LTX2AudioDecoderConfig",
    "LTX2VocoderConfig",
+    "UpsamplerConfig",
 ]
@@ -26,9 +26,6 @@ class HunyuanVideo15ArchConfig(DiTArchConfig):

    param_names_mapping: dict = field(
        default_factory=lambda: {
-            r"^cond_type_embed\.(.*)$":
-            r"cond_type_embed.\1",
-
            # 1. context_embedder.time_text_embed submodules (specific rules, applied first):
            r"^context_embedder\.time_text_embed\.timestep_embedder\.linear_1\.(.*)$":
            r"txt_in.t_embedder.mlp.fc_in.\1",
@@ -60,14 +57,6 @@ class HunyuanVideo15ArchConfig(DiTArchConfig):
            r"txt_in.refiner_blocks.\1.adaLN_modulation.linear.\2",
            r"^context_embedder\.token_refiner\.refiner_blocks\.(\d+)\.self_attn_qkv\.(.*)$":
            r"txt_in.refiner_blocks.\1.self_attn_qkv.\2",
-            r"^image_embedder\.linear_1\.(.*)$":
-            r"image_embedder.linear_1.\1",
-            r"^image_embedder\.linear_2\.(.*)$":
-            r"image_embedder.linear_2.\1",
-            r"^image_embedder\.norm_in\.(.*)$":
-            r"image_embedder.norm_in.\1",
-            r"^image_embedder\.norm_out\.(.*)$":
-            r"image_embedder.norm_out.\1",

            # 2. txt_in_2 mapping:
            r"^context_embedder_2\.(.*)$":
@@ -157,12 +146,6 @@ class HunyuanVideo15ArchConfig(DiTArchConfig):
    exclude_lora_layers: list[str] = field(
        default_factory=lambda: ["img_in", "txt_in", "time_in", "vector_in"])

-    # Causal HunyuanVideo1.5
-    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 = 31
-
    def __post_init__(self):
        super().__post_init__()
        self.hidden_size: int = self.attention_head_dim * self.num_attention_heads
@@ -8,12 +8,8 @@ def is_double_block(n: str, m) -> bool:
    return "double" 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_refiner_block(n: str, m) -> bool:
+    return "refiner" in n and str.isdigit(n.split(".")[-1])


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

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

    param_names_mapping: dict = field(
        default_factory=lambda: {
@@ -54,8 +50,7 @@ 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 (HYWorld uses TimestepEmbedder directly,
-            #    but FastVideo model inherits HunyuanVideo15TimeEmbedding with timestep_embedder):
+            # 2. time_in mappings:
            r"^time_in\.mlp\.0\.(.*)$":
            r"time_in.timestep_embedder.mlp.fc_in.\1",
            r"^time_in\.mlp\.2\.(.*)$":
@@ -1,4 +1,6 @@
 from dataclasses import dataclass, field
+
+import torch
 from fastvideo.configs.models.dits.wanvideo import WanVideoArchConfig, WanVideoConfig


@@ -76,8 +78,14 @@ 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 +1,6 @@
+from fastvideo.configs.models.upsamplers.hunyuan15 import SRTo720pUpsamplerConfig, SRTo1080pUpsamplerConfig
+from fastvideo.configs.models.upsamplers.base import UpsamplerConfig
+
+__all__ = [
+    "SRTo720pUpsamplerConfig", "SRTo1080pUpsamplerConfig", "UpsamplerConfig"
+]
@@ -0,0 +1,7 @@
+from dataclasses import dataclass
+from fastvideo.configs.models.base import ModelConfig
+
+
+@dataclass
+class UpsamplerConfig(ModelConfig):
+    pass
@@ -0,0 +1,20 @@
+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
@@ -6,8 +6,7 @@ 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.registry import (
-    get_pipeline_config_cls_from_name)
+from fastvideo.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)
+                                      VAEConfig, UpsamplerConfig)
 from fastvideo.configs.models.encoders import BaseEncoderOutput
 from fastvideo.configs.utils import update_config_from_args
 from fastvideo.logger import init_logger
@@ -44,12 +44,15 @@ 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)
@@ -216,6 +219,24 @@ 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")
@@ -245,8 +266,7 @@ class PipelineConfig:
        """
        use the pipeline class setting from model_path to match the pipeline config
        """
-        from fastvideo.configs.pipelines.registry import (
-            get_pipeline_config_cls_from_name)
+        from fastvideo.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))
@@ -260,8 +280,7 @@ class PipelineConfig:
        kwargs: dictionary of kwargs
        config_cli_prefix: prefix of CLI arguments for this PipelineConfig instance
        """
-        from fastvideo.configs.pipelines.registry import (
-            get_pipeline_config_cls_from_name)
+        from fastvideo.registry import get_pipeline_config_cls_from_name

        prefix_with_dot = f"{config_cli_prefix}." if (config_cli_prefix.strip()
                                                      != "") else ""
@@ -298,6 +317,12 @@ 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:
@@ -11,7 +11,8 @@ 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.pipelines.base import PipelineConfig
+from fastvideo.configs.models.upsamplers import SRTo720pUpsamplerConfig, SRTo1080pUpsamplerConfig
+from fastvideo.configs.pipelines.base import PipelineConfig, UpsamplerConfig

 PROMPT_TEMPLATE_TOKEN_LENGTH = 108

@@ -127,12 +128,12 @@ class Hunyuan15T2V480PConfig(PipelineConfig):
    vae_tiling: bool = True

    def __post_init__(self):
-        self.vae_config.load_encoder = True
+        self.vae_config.load_encoder = False
        self.vae_config.load_decoder = True


@dataclass
-class Hunyuan15DistilledI2V480PConfig(Hunyuan15T2V480PConfig):
+class Hunyuan15I2V480PStepDistilledConfig(Hunyuan15T2V480PConfig):
    flow_shift: int = 7


@@ -145,11 +146,21 @@ class Hunyuan15T2V720PConfig(Hunyuan15T2V480PConfig):


@dataclass
-class SelfForcingHunyuan15T2V480PConfig(Hunyuan15T2V480PConfig):
-    flow_shift: int = 5
-    is_causal: bool = True
-    # dmd_denoising_steps: list[int] | None = field(
-    #     default_factory=lambda: [1000, 750, 500, 250])
-    dmd_denoising_steps: list[int] | None = field(
-        default_factory=lambda: [1000, 875, 750, 625, 500, 375, 250, 125])
-    warp_denoising_step: bool = True
+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"
@@ -1,218 +0,0 @@
-# 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.hyworld import HYWorldConfig
-from fastvideo.configs.pipelines.ltx2 import LTX2T2VConfig
-from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig, SelfForcingHunyuan15T2V480PConfig, Hunyuan15DistilledI2V480PConfig
-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,
-    "weizhou03/SFHunyuanVideo-1.5-Diffusers-480p_t2v":
-    SelfForcingHunyuan15T2V480PConfig,
-    "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled":
-    Hunyuan15DistilledI2V480PConfig,
-    "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,
-    "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
@@ -41,7 +41,7 @@ class WanT2V480PConfig(PipelineConfig):
    vae_sp: bool = False

    # Denoising stage
-    flow_shift: float | None = 5.0
+    flow_shift: float | None = 3.0

    # Text encoding stage
    text_encoder_configs: tuple[EncoderConfig, ...] = field(
@@ -176,8 +176,6 @@ class SelfForcingWanT2V480PConfig(WanT2V480PConfig):
    flow_shift: float | None = 5.0
    dmd_denoising_steps: list[int] | None = field(
        default_factory=lambda: [1000, 750, 500, 250])
-    # dmd_denoising_steps: list[int] | None = field(
-    #     default_factory=lambda: [1000, 680, 360, 180])
    warp_denoising_step: bool = True


@@ -28,6 +28,9 @@ 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)
@@ -55,10 +58,13 @@ 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
@@ -92,8 +98,7 @@ class SamplingParam:

    @classmethod
    def from_pretrained(cls, model_path: str) -> "SamplingParam":
-        from fastvideo.configs.sample.registry import (
-            get_sampling_param_cls_for_name)
+        from fastvideo.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()
@@ -5,15 +5,19 @@ from fastvideo.configs.sample.base import SamplingParam


@dataclass
-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
+class Cosmos25SamplingParamBase(SamplingParam):
+    height: int = 704
+    width: int = 1280
+    num_frames: int = 77
    fps: int = 24
+    seed: int = 0

    guidance_scale: float = 7.0
-    # Official Cosmos2.5 sampling uses empty string as unconditional.
-    negative_prompt: str = ""
+    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.")
    num_inference_steps: int = 35
@@ -17,8 +17,8 @@ class Hunyuan15_480P_SamplingParam(SamplingParam):
    guidance_scale: float = 6.0
    prompt_attention_mask: list = field(default_factory=list)
    negative_attention_mask: list = field(default_factory=list)
-    # sigmas: list[float] | None = field(
-    #     default_factory=lambda: list(np.linspace(1.0, 0.0, 50 + 1)[:-1]))
+    sigmas: list[float] | None = field(
+        default_factory=lambda: list(np.linspace(1.0, 0.0, 50 + 1)[:-1]))

    negative_prompt: str = ""

@@ -29,10 +29,12 @@ class Hunyuan15_480P_SamplingParam(SamplingParam):


@dataclass
-class Hunyuan15_480P_Distilled_SamplingParam(Hunyuan15_480P_SamplingParam):
-    num_inference_steps: int = 8
-    fps: int = 24
+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


@@ -40,3 +42,19 @@ class Hunyuan15_480P_Distilled_SamplingParam(Hunyuan15_480P_SamplingParam):
 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
@@ -1,208 +0,0 @@
-# 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.hyworld import HYWorld_SamplingParam
-from fastvideo.configs.sample.hunyuan15 import Hunyuan15_480P_SamplingParam, Hunyuan15_720P_SamplingParam, Hunyuan15_480P_Distilled_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,
-    "weizhou03/SFHunyuanVideo-1.5-Diffusers-480p_t2v":
-    Hunyuan15_480P_SamplingParam,
-    "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled":
-    Hunyuan15_480P_Distilled_SamplingParam,
-    "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,
-    "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
@@ -15,8 +15,7 @@ class WanT2V_1_3B_SamplingParam(SamplingParam):

    # Denoising stage
    guidance_scale: float = 3.0
-    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"
-    # negative_prompt: str = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+    negative_prompt: str = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
    num_inference_steps: int = 50

    teacache_params: WanTeaCacheParams = field(
@@ -10,7 +10,7 @@ from fastvideo.dataset.transform import (CenterCropResizeVideo, Normalize255,
 from fastvideo.dataset.validation_dataset import ValidationDataset


-def getdataset(args, start_idx: int = 0) -> VideoCaptionMergedDataset:
+def getdataset(args) -> VideoCaptionMergedDataset:
    if args.do_temporal_sample:
        temporal_sample = TemporalRandomCrop(args.num_frames)  # 16 x
    else:
@@ -36,7 +36,6 @@ def getdataset(args, start_idx: int = 0) -> VideoCaptionMergedDataset:
                                     transform=transform,
                                     temporal_sample=temporal_sample,
                                     transform_topcrop=transform_topcrop,
-                                     start_idx=start_idx,
                                     seed=args.seed)
                                    

@@ -165,25 +165,6 @@ def ode_text_only_record_creator(
        "trajectory_timesteps_dtype": str(trajectory_timesteps.dtype),
    })

-    # if text_embedding_2 is not None:
-    #     record.update({
-    #         "text_embedding_2_bytes": text_embedding_2.tobytes(),
-    #         "text_embedding_2_shape": list(text_embedding_2.shape),
-    #         "text_embedding_2_dtype": str(text_embedding_2.dtype),
-    #     })
-    # if text_mask is not None:
-    #     record.update({
-    #         "text_mask_bytes": text_mask.tobytes(),
-    #         "text_mask_shape": list(text_mask.shape),
-    #         "text_mask_dtype": str(text_mask.dtype),
-    #     })
-    # if text_mask_2 is not None:
-    #     record.update({
-    #         "text_mask_2_bytes": text_mask_2.tobytes(),
-    #         "text_mask_2_shape": list(text_mask_2.shape),
-    #         "text_mask_2_dtype": str(text_mask_2.dtype),
-    #     })
-        
    return record


@@ -206,4 +187,4 @@ def text_only_record_creator(text_name: str, text_embedding: np.ndarray,
        "text_embedding_dtype": str(text_embedding.dtype),
        "caption": caption,
    }
-    return record
+    return record
@@ -90,15 +90,6 @@ pyarrow_schema_ode_trajectory_text_only = pa.schema([
    pa.field("text_embedding_shape", pa.list_(pa.int64())),
    # e.g., 'bfloat16' or 'float32'
    pa.field("text_embedding_dtype", pa.string()),
-    # pa.field("text_embedding_2_bytes", pa.binary()),
-    # pa.field("text_embedding_2_shape", pa.list_(pa.int64())),
-    # pa.field("text_embedding_2_dtype", pa.string()),
-    # pa.field("text_mask_bytes", pa.binary()),
-    # pa.field("text_mask_shape", pa.list_(pa.int64())),
-    # pa.field("text_mask_dtype", pa.string()),
-    # pa.field("text_mask_2_bytes", pa.binary()),
-    # pa.field("text_mask_2_shape", pa.list_(pa.int64())),
-    # pa.field("text_mask_2_dtype", pa.string()),
    # --- ODE Trajectory ---
    pa.field("trajectory_latents_bytes", pa.binary()),
    pa.field("trajectory_latents_shape", pa.list_(pa.int64())),
@@ -126,6 +117,7 @@ pyarrow_schema_text_only = pa.schema([
    pa.field("caption", pa.string()),
 ])

+
 pyarrow_schema_matrixgame = pa.schema([
    pa.field("id", pa.string()),
    # --- Image/Video VAE latents ---
@@ -164,4 +156,4 @@ pyarrow_schema_matrixgame = pa.schema([
    pa.field("num_frames", pa.int64()),
    pa.field("duration_sec", pa.float64()),
    pa.field("fps", pa.float64()),
-])
+])
@@ -17,7 +17,6 @@ from PIL import Image
 from transformers import AutoTokenizer

 from fastvideo.logger import init_logger
-from fastvideo.utils import maybe_download_model

 logger = init_logger(__name__)

@@ -41,6 +40,7 @@ class PreprocessBatch:
    num_frames: int | None = None
    sample_frame_index: list[int] | None = None
    sample_num_frames: int | None = None
+    action_path: str | None = None

    # Processed data
    pixel_values: torch.Tensor | None = None
@@ -148,62 +148,6 @@ 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."""

@@ -328,11 +272,6 @@ 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
@@ -366,9 +305,11 @@ class ImageTransformStage(DatasetStage):
            image = self.transform_topcrop(image)
        elif self.transform is not None:
            image = self.transform(image)
+            image = image.float() / 127.5 - 1.0
+        else:
+            image = image.float() / 127.5 - 1.0

        image = image.transpose(0, 1)  # [1 C H W] -> [C 1 H W]
-        image = image.float() / 127.5 - 1.0
        batch.pixel_values = image
        return batch

@@ -470,21 +411,23 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
        self.seed = seed

        # Initialize tokenizer
-        tokenizer_path = os.path.join(maybe_download_model(args.model_path), "tokenizer")
-        tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
+        tokenizer_path = os.path.join(args.model_path, "tokenizer")
+        if os.path.exists(tokenizer_path):
+            tokenizer = AutoTokenizer.from_pretrained(tokenizer_path,
+                                                      cache_dir=args.cache_dir)
+        else:
+            tokenizer = None

        # Initialize processing stages
        self._init_stages(args, transform, transform_topcrop, tokenizer)

        # Process metadata
-        self.processed_batches = self._process_metadata(start_idx=start_idx)
+        self.processed_batches = self._process_metadata()

    def _init_stages(self, args, transform, transform_topcrop,
                     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,
@@ -504,7 +447,7 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
        else:
            self.text_encoding_stage = None

-    def _load_raw_data(self, start_idx: int = 0) -> list[dict]:
+    def _load_raw_data(self) -> list[dict]:
        """Load raw data from JSON files."""
        # Read folder-annotation pairs
        with open(self.data_merge_path) as f:
@@ -524,18 +467,19 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
        # Update paths with folder prefix
        for item in data_items:
            item["path"] = opj(folder, item["path"])
+            if "action_path" in item and item["action_path"]:
+                item["action_path"] = opj(folder, item["action_path"])

-        return data_items[start_idx:]
+        return data_items

-    def _process_metadata(self, start_idx: int = 0) -> list[PreprocessBatch]:
+    def _process_metadata(self) -> list[PreprocessBatch]:
        """Process the raw metadata through all filtering stages."""
-        raw_data = self._load_raw_data(start_idx=start_idx)
+        raw_data = self._load_raw_data()
        processed_batches = []

        # Initialize counters
        filter_counts = {
            "validation_failed": 0,
-            "resolution_failed": 0,
            "frame_sampling_failed": 0
        }
        sample_num_frames: list[int] = []
@@ -545,7 +489,8 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
                                    cap=item["cap"],
                                    resolution=item.get("resolution"),
                                    fps=item.get("fps"),
-                                    duration=item.get("duration"))
+                                    duration=item.get("duration"),
+                                    action_path=item.get("action_path"))

            # Apply filtering stages
            if not self._apply_filter_stages(batch, filter_counts):
@@ -570,10 +515,6 @@ 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
@@ -585,10 +526,9 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
                             after_count: int):
        """Log filtering statistics."""
        logger.info(
-            "validation_failed: %d, resolution_failed: %d, frame_sampling_failed: %d, "
+            "validation_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)

@@ -609,15 +549,10 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
        batch = self.image_transform_stage.process(batch)
        if self.text_encoding_stage is not None:
            batch = self.text_encoding_stage.process(batch)
-        else:
-            raise ValueError("Text encoding stage is not initialized")

        # Build result dictionary
        result = {
            "pixel_values": batch.pixel_values,
-            # "text": batch.text,
-            # "input_ids": batch.input_ids,
-            # "cond_mask": batch.cond_mask,
            "path": batch.path,
        }

@@ -625,12 +560,14 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
            result["text"] = batch.text
            result["input_ids"] = batch.input_ids
            result["cond_mask"] = batch.cond_mask
-        else:
-            raise ValueError("Text encoding stage is not initialized")
-            
+
        # Add video-specific fields
        if batch.is_video:
            result.update({"fps": batch.fps, "duration": batch.duration})
+        
+        # Add action_path
+        if batch.action_path:
+            result["action_path"] = batch.action_path

        return result

@@ -668,7 +605,7 @@ class TextDataset(torch.utils.data.IterableDataset,
        self.seed = seed

        # Initialize tokenizer
-        tokenizer_path = os.path.join(maybe_download_model(args.model_path), "tokenizer")
+        tokenizer_path = os.path.join(args.model_path, "tokenizer")
        tokenizer = AutoTokenizer.from_pretrained(tokenizer_path,
                                                  cache_dir=args.cache_dir)

@@ -771,4 +708,4 @@ class TextDataset(torch.utils.data.IterableDataset,

    def load_state_dict(self, state_dict: dict[str, Any]) -> None:
        """Load state dict from checkpoint."""
-        self.processed_batches = state_dict["processed_batches"]
+        self.processed_batches = state_dict["processed_batches"]
@@ -156,14 +156,6 @@ def collate_rows_from_parquet_schema(rows,
                    else:
                        data = np.frombuffer(
                            bytes_data, dtype=np.float32).reshape(shape).copy()
-                        # if row[f"{tensor_name}_dtype"] == "float32":
-                        #     data = np.frombuffer(
-                        #         bytes_data, dtype=np.float32).reshape(shape).copy()
-                        # elif row[f"{tensor_name}_dtype"] == "int64":
-                        #     data = np.frombuffer(
-                        #         bytes_data, dtype=np.int64).reshape(shape).copy()
-                        # else:
-                        #     raise ValueError(f"Unsupported dtype: {row[f"{tensor_name}_dtype"]}")
                    tensor = torch.from_numpy(data)
                    # if len(data.shape) == 3:
                    #     B, L, D = tensor.shape
@@ -183,8 +175,6 @@ def collate_rows_from_parquet_schema(rows,

            for tensor in tensor_list:
                if tensor.numel() > 0:
-                    if tensor.ndim == 3:
-                        tensor = tensor.squeeze(0)
                    padded_tensor, mask = pad(tensor, text_padding_length)
                    padded_tensors.append(padded_tensor)
                    attention_masks.append(mask)
@@ -232,4 +222,4 @@ def collate_rows_from_parquet_schema(rows,
    if info_list and 'caption' in info_list[0]:
        batch_data['caption_text'] = [info['caption'] for info in info_list]

-    return batch_data
+    return batch_data
@@ -2,7 +2,6 @@
 # adapted from: https://github.com/a-r-r-o-w/finetrainers/blob/main/finetrainers/data/dataset.py
 import os
 import pathlib
-import json

 import datasets
 from torch.utils.data import IterableDataset
@@ -33,7 +32,10 @@ class ValidationDataset(IterableDataset):
                                         data_files=self.filename.as_posix(),
                                         split="train")
        elif self.filename.suffix == ".json":
-            data = self._load_json_without_hf_locking(self.filename)
+            data = datasets.load_dataset("json",
+                                         data_files=self.filename.as_posix(),
+                                         split="train",
+                                         field="data")
        elif self.filename.suffix == ".parquet":
            data = datasets.load_dataset("parquet",
                                         data_files=self.filename.as_posix(),
@@ -160,33 +162,3 @@ class ValidationDataset(IterableDataset):

            sample = {k: v for k, v in sample.items() if v is not None}
            yield sample
-
-    @staticmethod
-    def _load_json_without_hf_locking(filename: pathlib.Path) -> list[dict]:
-        """Load JSON validation files without using datasets.load_dataset lock files.
-
-        This avoids distributed lock contention when many ranks initialize
-        validation simultaneously on shared filesystems.
-        """
-        with filename.open("r", encoding="utf-8") as f:
-            payload = json.load(f)
-
-        if isinstance(payload, dict):
-            if "data" not in payload:
-                raise ValueError(
-                    f"Validation JSON {filename.as_posix()} must contain a 'data' field when top-level is an object."
-                )
-            records = payload["data"]
-        elif isinstance(payload, list):
-            records = payload
-        else:
-            raise ValueError(
-                f"Validation JSON {filename.as_posix()} must be either a list of samples or an object containing a 'data' list."
-            )
-
-        if not isinstance(records, list):
-            raise ValueError(
-                f"Validation JSON records in {filename.as_posix()} must be a list."
-            )
-
-        return records
@@ -18,6 +18,7 @@ __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",
@@ -7,10 +7,17 @@ import torch.distributed
 from fastvideo.distributed.parallel_state import (get_sp_group,
                                                  get_sp_parallel_rank,
                                                  get_sp_world_size,
-                                                  get_tp_group)
+                                                  get_tp_group,
+                                                  model_parallel_is_initialized)
 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:
@@ -101,3 +108,86 @@ 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.")
@@ -9,6 +9,7 @@ diffusion models.
 import math
 import os
 import re
+import threading
 import time
 from copy import deepcopy
 from typing import Any
@@ -31,6 +32,21 @@ 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.
@@ -372,8 +388,31 @@ class VideoGenerator:

        # Run inference
        start_time = time.perf_counter()
-        output_batch = self.executor.execute_forward(batch, fastvideo_args)
-        samples = output_batch.output
+
+        # 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()
        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, STA_Mode
+from fastvideo.configs.pipelines.base import PipelineConfig
 from fastvideo.configs.utils import clean_cli_args
 from fastvideo.layers.quantization import QUANTIZATION_METHODS, QuantizationMethods
 from fastvideo.logger import init_logger
@@ -141,8 +141,6 @@ 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
@@ -179,6 +177,7 @@ 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
@@ -464,20 +463,6 @@ 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,
@@ -830,7 +815,6 @@ class TrainingArgs(FastVideoArgs):
    precedence.
    """
    data_path: str = ""
-    data_path_2: str | None = None
    dataloader_num_workers: int = 0
    num_height: int = 0
    num_width: int = 0
@@ -860,7 +844,6 @@ class TrainingArgs(FastVideoArgs):
    validation_sampling_steps: str = ""
    validation_guidance_scale: str = ""
    validation_steps: float = 0.0
-    visualization_steps: float = 0.0
    log_validation: bool = False
    trackers: list[str] = dataclasses.field(default_factory=list)
    tracker_project_name: str = ""
@@ -876,7 +859,6 @@ class TrainingArgs(FastVideoArgs):
    num_train_epochs: int = 0
    max_train_steps: int = 0
    gradient_accumulation_steps: int = 0
-    optimizer_type: str = "adamw"
    learning_rate: float = 0.0
    scale_lr: bool = False
    lr_scheduler: str = "constant"
@@ -934,12 +916,10 @@ class TrainingArgs(FastVideoArgs):
    training_state_checkpointing_steps: int = 0  # for resuming training
    weight_only_checkpointing_steps: int = 0  # for inference
    log_visualization: bool = False
+    visualization_steps: int = 0
    # simulate generator forward to match inference
    simulate_generator_forward: bool = False
    warp_denoising_step: bool = False
-    use_tf: bool = False
-    use_decoupled_dmd: bool = False  # use decoupled DMD
-    use_gan_loss: bool = False

    # Self-forcing specific arguments
    num_frame_per_block: int = 3
@@ -949,8 +929,6 @@ class TrainingArgs(FastVideoArgs):
    same_step_across_blocks: bool = False  # Use same exit timestep for all blocks
    last_step_only: bool = False  # Only use the last timestep for training
    context_noise: int = 0  # Context noise level for cache updates
-    use_context_forcing: bool = False
-    use_ode_init: bool = False

    @classmethod
    def from_cli_args(cls, args: argparse.Namespace) -> "TrainingArgs":
@@ -1006,10 +984,6 @@ class TrainingArgs(FastVideoArgs):
                            type=str,
                            required=True,
                            help="Path to parquet files")
-        parser.add_argument("--data-path-2",
-                            type=str,
-                            required=False,
-                            help="Path to parquet files")
        parser.add_argument("--dataloader-num-workers",
                            type=int,
                            required=True,
@@ -1103,12 +1077,12 @@ class TrainingArgs(FastVideoArgs):
        parser.add_argument("--validation-steps",
                            type=float,
                            help="Number of validation steps")
-        parser.add_argument("--visualization-steps",
-                            type=float,
-                            help="Number of visualization steps")
        parser.add_argument("--log-validation",
                            action=StoreBoolean,
                            help="Whether to log validation results")
+        parser.add_argument("--visualization-steps",
+                            type=int,
+                            help="Number of visualization steps")
        parser.add_argument("--tracker-project-name",
                            type=str,
                            help="Project name for tracking")
@@ -1154,10 +1128,6 @@ class TrainingArgs(FastVideoArgs):
        parser.add_argument("--gradient-accumulation-steps",
                            type=int,
                            help="Number of steps to accumulate gradients")
-        parser.add_argument("--optimizer-type",
-                            type=str,
-                            choices=["adamw", "muon"],
-                            help="Optimizer type")
        parser.add_argument("--learning-rate",
                            type=float,
                            required=True,
@@ -1348,21 +1318,6 @@ class TrainingArgs(FastVideoArgs):
            help=
            "Whether to warp denoising step according to the scheduler time shift"
        )
-        parser.add_argument(
-            "--use-tf",
-            action=StoreBoolean,
-            help="Whether to use Teacher-forcing for finetuning"
-        )
-        parser.add_argument(
-            "--use-decoupled-dmd",
-            action=StoreBoolean,
-            help="Whether to use decoupled DMD"
-        )
-        parser.add_argument(
-            "--use-gan-loss",
-            action=StoreBoolean,
-            help="Whether to use GAN loss"
-        )

        # Self-forcing specific arguments
        parser.add_argument(
@@ -1399,12 +1354,6 @@ class TrainingArgs(FastVideoArgs):
                            type=int,
                            default=TrainingArgs.context_noise,
                            help="Context noise level for cache updates")
-        parser.add_argument("--use-context-forcing",
-                            action=StoreBoolean,
-                            help="Whether to use context forcing")
-        parser.add_argument("--use-ode-init",
-                            action=StoreBoolean,
-                            help="Whether to use ODE init")

        return parser

@@ -193,3 +193,46 @@ 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 "
+_FORMAT = (f"{FASTVIDEO_LOGGING_PREFIX}%(levelname)s %(asctime)s.%(msecs)03d "
           "[%(filename)s:%(lineno)d] %(message)s")
 _DATE_FORMAT = "%m-%d %H:%M:%S"

@@ -102,6 +102,7 @@ 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"])
@@ -118,20 +119,22 @@ def _info(logger: Logger,

    global _warned_local_main_process, _warned_main_process

-    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
+    # 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 main_process_only and not local_main_process_only:
        logger.log(logging.INFO, msg, *args, stacklevel=2, **kwargs)
@@ -1325,3 +1325,6 @@ 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]
@@ -1,948 +0,0 @@
-# Copyright 2025 The Hunyuan Team and The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from typing import Any, Dict, Optional, List
-import math
-
-import torch
-# import torch._dynamo
-# torch._dynamo.config.cache_size_limit = 128
-# try:
-#     torch._dynamo.config.recompile_limit = 128
-# except AttributeError:
-#     pass
-import torch.nn as nn
-import torch.nn.functional as F
-
-from torch.nn.attention.flex_attention import create_block_mask, flex_attention
-from torch.nn.attention.flex_attention import BlockMask
-flex_attention = torch.compile(
-    flex_attention, dynamic=False, mode="default")
-import torch.distributed as dist
-
-from fastvideo.attention import LocalAttention
-from fastvideo.forward_context import set_forward_context
-from fastvideo.configs.models.dits import HunyuanVideo15Config
-from fastvideo.layers.layernorm import (LayerNormScaleShift, ScaleResidual,
-                                        ScaleResidualLayerNormScaleShift)
-from fastvideo.layers.linear import ReplicatedLinear
-# TODO(will-PY-refactor): RMSNorm ....
-from fastvideo.layers.mlp import MLP
-from fastvideo.layers.rotary_embedding import get_rotary_pos_embed, _apply_rotary_emb
-from fastvideo.layers.visual_embedding import (ModulateProjection, PatchEmbed,
-                                               unpatchify)
-from fastvideo.models.dits.base import CachableDiT
-from fastvideo.platforms import AttentionBackendEnum
-from fastvideo.logger import init_logger
-
-from fastvideo.models.dits.hunyuanvideo15 import (
-    HunyuanRMSNorm, 
-    HunyuanVideo15TimeEmbedding, 
-    HunyuanVideo15ByT5TextProjection,
-    HunyuanVideo15ImageProjection,
-    SingleTokenRefiner,
-    FinalLayer)
-
-logger = init_logger(__name__)
-
-class MMDoubleStreamBlock(nn.Module):
-    """
-    A multimodal DiT block with separate modulation for text and image/video,
-    using distributed attention and linear layers.
-    """
-
-    def __init__(
-        self,
-        hidden_size: int,
-        num_attention_heads: int,
-        mlp_ratio: float,
-        local_attn_size: int = -1,
-        sink_size: int = 0,
-        dtype: torch.dtype | None = None,
-        supported_attention_backends: tuple[AttentionBackendEnum, ...]
-        | None = None,
-        prefix: str = "",
-    ):
-        super().__init__()
-
-        self.deterministic = False
-        self.num_attention_heads = num_attention_heads
-        self.head_dim = hidden_size // num_attention_heads
-        self.local_attn_size = local_attn_size
-        self.sink_size = sink_size
-        mlp_hidden_dim = int(hidden_size * mlp_ratio)
-
-        # Image modulation components
-        self.img_mod = ModulateProjection(
-            hidden_size,
-            factor=6,
-            act_layer="silu",
-            dtype=dtype,
-            prefix=f"{prefix}.img_mod",
-        )
-
-        # Fused operations for image stream
-        self.img_attn_norm = LayerNormScaleShift(hidden_size,
-                                                 norm_type="layer",
-                                                 elementwise_affine=False,
-                                                 dtype=dtype)
-        self.img_attn_residual_mlp_norm = ScaleResidualLayerNormScaleShift(
-            hidden_size,
-            norm_type="layer",
-            elementwise_affine=False,
-            dtype=dtype)
-        self.img_mlp_residual = ScaleResidual()
-
-        # Image attention components
-        self.img_attn_qkv = ReplicatedLinear(hidden_size,
-                                             hidden_size * 3,
-                                             bias=True,
-                                             params_dtype=dtype,
-                                             prefix=f"{prefix}.img_attn_qkv")
-
-        self.img_attn_q_norm = HunyuanRMSNorm(self.head_dim, eps=1e-6, dtype=dtype)
-        self.img_attn_k_norm = HunyuanRMSNorm(self.head_dim, eps=1e-6, dtype=dtype)
-
-        self.img_attn_proj = ReplicatedLinear(hidden_size,
-                                              hidden_size,
-                                              bias=True,
-                                              params_dtype=dtype,
-                                              prefix=f"{prefix}.img_attn_proj")
-
-        self.img_mlp = MLP(hidden_size,
-                           mlp_hidden_dim,
-                           bias=True,
-                           dtype=dtype,
-                           prefix=f"{prefix}.img_mlp")
-
-        # Text modulation components
-        self.txt_mod = ModulateProjection(
-            hidden_size,
-            factor=6,
-            act_layer="silu",
-            dtype=dtype,
-            prefix=f"{prefix}.txt_mod",
-        )
-
-        # Fused operations for text stream
-        self.txt_attn_norm = LayerNormScaleShift(hidden_size,
-                                                 norm_type="layer",
-                                                 elementwise_affine=False,
-                                                 dtype=dtype)
-        self.txt_attn_residual_mlp_norm = ScaleResidualLayerNormScaleShift(
-            hidden_size,
-            norm_type="layer",
-            elementwise_affine=False,
-            dtype=dtype)
-        self.txt_mlp_residual = ScaleResidual()
-
-        # Text attention components
-        self.txt_attn_qkv = ReplicatedLinear(hidden_size,
-                                             hidden_size * 3,
-                                             bias=True,
-                                             params_dtype=dtype)
-
-        # QK norm layers for text
-        self.txt_attn_q_norm = HunyuanRMSNorm(self.head_dim, eps=1e-6, dtype=dtype)
-        self.txt_attn_k_norm = HunyuanRMSNorm(self.head_dim, eps=1e-6, dtype=dtype)
-
-        self.txt_attn_proj = ReplicatedLinear(hidden_size,
-                                              hidden_size,
-                                              bias=True,
-                                              params_dtype=dtype)
-
-        self.txt_mlp = MLP(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype)
-
-        self.max_attention_size = 21 * 1590 if local_attn_size == -1 else local_attn_size * 1590
-
-        self.attn = LocalAttention(
-            num_heads=self.num_attention_heads,
-            head_size=self.head_dim,
-            supported_attention_backends=supported_attention_backends,
-            prefix=f"{prefix}.attn"
-        )
-
-    def forward_txt(
-        self,
-        txt: torch.Tensor,
-        vec: torch.Tensor,
-        cache_txt: bool = False,
-    ):
-        txt_mod_outputs = self.txt_mod(vec)
-        (
-            txt_attn_shift,
-            txt_attn_scale,
-            txt_attn_gate,
-            txt_mlp_shift,
-            txt_mlp_scale,
-            txt_mlp_gate,
-        ) = torch.chunk(txt_mod_outputs, 6, dim=-1)
-
-        # Prepare text for attention using fused operation
-        txt_attn_input = self.txt_attn_norm(txt, txt_attn_shift, txt_attn_scale)
-
-        # Get QKV for text
-        txt_qkv, _ = self.txt_attn_qkv(txt_attn_input)
-        batch_size, text_seq_len = txt_qkv.shape[0], txt_qkv.shape[1]
-
-        # Split QKV
-        txt_qkv = txt_qkv.view(batch_size, text_seq_len, 3,
-                               self.num_attention_heads, -1)
-        txt_q, txt_k, txt_v = txt_qkv[:, :, 0], txt_qkv[:, :, 1], txt_qkv[:, :,
-                                                                          2]
-        # Apply QK-Norm if needed
-        txt_q = self.txt_attn_q_norm(txt_q).to(txt_q.dtype)
-        txt_k = self.txt_attn_k_norm(txt_k).to(txt_k.dtype)
-
-        t_kv = {}
-        if cache_txt:
-            t_kv["k_txt"] = txt_k
-            t_kv["v_txt"] = txt_v
-
-        txt_attn = self.attn(txt_q, txt_k, txt_v)
-        # Process text attention output
-        txt_attn_out, _ = self.txt_attn_proj(
-            txt_attn.reshape(batch_size, text_seq_len, -1))
-
-        # Use fused operation for residual connection, normalization, and modulation
-        txt_mlp_input, txt_residual = self.txt_attn_residual_mlp_norm(
-            txt, txt_attn_out, txt_attn_gate, txt_mlp_shift, txt_mlp_scale)
-
-        # Process text MLP
-        txt_mlp_out = self.txt_mlp(txt_mlp_input)
-        txt = self.txt_mlp_residual(txt_residual, txt_mlp_out, txt_mlp_gate)
-
-        return txt, t_kv
-
-    def forward_vision(
-        self,
-        img: torch.Tensor,
-        vec: torch.Tensor,
-        freqs_cis: tuple,
-        block_mask: BlockMask,
-        kv_cache: dict | None = None,
-        txt_kv_cache: list | None = None,
-        current_start: int = 0,
-    ) -> tuple[torch.Tensor, torch.Tensor]:
-        # Process modulation vectors
-        if vec.dim() == 3:
-            img_mod_outputs = self.img_mod(vec).unflatten(dim=-1, sizes=(6, -1))
-            (
-                img_attn_shift,
-                img_attn_scale,
-                img_attn_gate,
-                img_mlp_shift,
-                img_mlp_scale,
-                img_mlp_gate,
-            ) = torch.chunk(img_mod_outputs, 6, dim=2)
-        else:
-            img_mod_outputs = self.img_mod(vec)
-            (
-                img_attn_shift,
-                img_attn_scale,
-                img_attn_gate,
-                img_mlp_shift,
-                img_mlp_scale,
-                img_mlp_gate,
-            ) = torch.chunk(img_mod_outputs, 6, dim=-1)
-
-        # Prepare image for attention using fused operation
-        img_attn_input = self.img_attn_norm(img, img_attn_shift, img_attn_scale)
-        # Get QKV for image
-        img_qkv, _ = self.img_attn_qkv(img_attn_input)
-        batch_size, image_seq_len = img_qkv.shape[0], img_qkv.shape[1]
-
-        # Split QKV
-        img_qkv = img_qkv.view(batch_size, image_seq_len, 3,
-                               self.num_attention_heads, -1)
-        img_q, img_k, img_v = img_qkv[:, :, 0], img_qkv[:, :, 1], img_qkv[:, :,
-                                                                          2]
-
-        # Apply QK-Norm if needed
-        img_q = self.img_attn_q_norm(img_q).to(img_v)
-        img_k = self.img_attn_k_norm(img_k).to(img_v)
-
-        # Apply rotary embeddings
-        cos, sin = freqs_cis
-
-        # Apply flex_attention
-        # Does not support SP padding for now
-        if kv_cache is None:
-            is_tf = (img_q.shape[1] == cos.shape[0] * 2)
-            if is_tf:
-                q_chunk = torch.chunk(img_q, 2, dim=1)
-                k_chunk = torch.chunk(img_k, 2, dim=1)
-                roped_query = []
-                roped_key = []
-                for ii in range(2):
-                    rq = _apply_rotary_emb(q_chunk[ii], cos, sin, is_neox_style=False)
-                    rk = _apply_rotary_emb(k_chunk[ii], cos, sin, is_neox_style=False)
-                    roped_query.append(rq)
-                    roped_key.append(rk)
-                img_q = torch.cat(roped_query, dim=1)
-                img_k = torch.cat(roped_key, dim=1)
-            else:
-                img_q = _apply_rotary_emb(img_q, cos, sin, is_neox_style=False)
-                img_k = _apply_rotary_emb(img_k, cos, sin, is_neox_style=False)
-
-            q = img_q
-            k = torch.cat([img_k, txt_kv_cache["k_txt"]], dim=1)
-            v = torch.cat([img_v, txt_kv_cache["v_txt"]], dim=1)
-            # Padding for flex attention
-            padded_length = math.ceil(q.shape[1] / 128) * 128 - q.shape[1]
-            padded_kv_length = math.ceil(k.shape[1] / 128) * 128 - k.shape[1]
-            padded_roped_query = torch.cat(
-                [q,
-                    torch.zeros([q.shape[0], padded_length, q.shape[2], q.shape[3]],
-                                device=q.device, dtype=v.dtype)],
-                dim=1
-            )
-
-            padded_roped_key = torch.cat(
-                [k, torch.zeros([k.shape[0], padded_kv_length, k.shape[2], k.shape[3]],
-                                        device=k.device, dtype=v.dtype)],
-                dim=1
-            )
-
-            padded_v = torch.cat(
-                [v, torch.zeros([v.shape[0], padded_kv_length, v.shape[2], v.shape[3]],
-                                device=v.device, dtype=v.dtype)],
-                dim=1
-            )
-
-            img_attn = flex_attention(
-                query=padded_roped_query.transpose(2, 1),
-                key=padded_roped_key.transpose(2, 1),
-                value=padded_v.transpose(2, 1),
-                block_mask=block_mask
-            )[:, :, :-padded_length].transpose(2, 1)
-
-            assert img_attn.shape[1] == image_seq_len
-            updated_kv_cache = None
-        else:
-            img_q = _apply_rotary_emb(img_q, cos, sin, is_neox_style=False)
-            img_k = _apply_rotary_emb(img_k, cos, sin, is_neox_style=False)
-
-            current_end = current_start + img_q.shape[1]
-            num_new_tokens = img_q.shape[1]
-            sink_tokens = self.sink_size * 1590
-            # If we are using local attention and the current KV cache size is larger than the local attention size, we need to truncate the KV cache
-            kv_cache_size = self.max_attention_size
-            
-            # Clone cache to avoid in-place modification during gradient checkpointing
-            k_cache = kv_cache["k"].clone()
-            v_cache = kv_cache["v"].clone()
-            
-            if self.local_attn_size != -1 and (current_end > kv_cache["global_end_index"].item()) and (
-                    num_new_tokens + kv_cache["local_end_index"].item() > kv_cache_size):
-                # Calculate the number of new tokens added in this step
-                # Shift existing cache content left to discard oldest tokens
-                # Clone the source slice to avoid overlapping memory error
-                num_evicted_tokens = num_new_tokens + kv_cache["local_end_index"].item() - kv_cache_size
-                num_rolled_tokens = kv_cache_size - num_new_tokens - sink_tokens
-                k_cache[:, sink_tokens:sink_tokens + num_rolled_tokens] = \
-                    k_cache[:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone()
-                v_cache[:, sink_tokens:sink_tokens + num_rolled_tokens] = \
-                    v_cache[:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone()
-                # Insert the new keys/values at the end
-                local_end_index = kv_cache["local_end_index"].item() + current_end - \
-                    kv_cache["global_end_index"].item() - num_evicted_tokens
-                local_start_index = local_end_index - num_new_tokens
-                assert local_end_index == self.max_attention_size
-            else:
-                # Assign new keys/values directly up to current_end
-                local_end_index = kv_cache["local_end_index"].item() + current_end - kv_cache["global_end_index"].item()
-                local_start_index = local_end_index - num_new_tokens
-
-            assert local_start_index >= 0
-            q = img_q
-            k = torch.cat([k_cache[:, :local_start_index], img_k, txt_kv_cache["k_txt"]], dim=1)
-            v = torch.cat([v_cache[:, :local_start_index], img_v, txt_kv_cache["v_txt"]], dim=1)
-            img_attn = self.attn(q, k, v)
-
-            k_cache[:, local_start_index:local_end_index] = img_k
-            v_cache[:, local_start_index:local_end_index] = img_v
-            
-            updated_kv_cache = {
-                "k": k_cache,
-                "v": v_cache,
-                "global_end_index": torch.tensor([current_end], dtype=torch.long, device=k_cache.device),
-                "local_end_index": torch.tensor([local_end_index], dtype=torch.long, device=k_cache.device)
-            }
-        
-        img_attn_out, _ = self.img_attn_proj(
-            img_attn.view(batch_size, image_seq_len, -1))
-        # Use fused operation for residual connection, normalization, and modulation
-        img_mlp_input, img_residual = self.img_attn_residual_mlp_norm(
-            img, img_attn_out, img_attn_gate, img_mlp_shift, img_mlp_scale)
-
-        # Process image MLP
-        img_mlp_out = self.img_mlp(img_mlp_input)
-        img = self.img_mlp_residual(img_residual, img_mlp_out, img_mlp_gate)
-
-        return img, updated_kv_cache
-
-    def forward(
-        self,
-        txt_inference=False,
-        vision_inference=False,
-        **kwargs
-    ):
-        if txt_inference:
-            return self.forward_txt(**kwargs)
-        elif vision_inference:
-            return self.forward_vision(**kwargs)
-        else:
-            raise ValueError("txt_inference and vision_inference cannot be both False")
-
-
-class CausalHunyuanVideo15Transformer3DModel(CachableDiT):
-    r"""
-    A Transformer model for video-like data used in [HunyuanVideo1.5](https://huggingface.co/tencent/HunyuanVideo1.5).
-    """
-
-    # shard single stream, double stream blocks, and refiner_blocks
-    _fsdp_shard_conditions = HunyuanVideo15Config()._fsdp_shard_conditions
-    _compile_conditions = HunyuanVideo15Config()._compile_conditions
-    _supported_attention_backends = HunyuanVideo15Config(
-    )._supported_attention_backends
-    param_names_mapping = HunyuanVideo15Config().param_names_mapping
-    reverse_param_names_mapping = HunyuanVideo15Config(
-    ).reverse_param_names_mapping
-    lora_param_names_mapping = HunyuanVideo15Config().lora_param_names_mapping
-
-    def __init__(
-        self,
-        config: HunyuanVideo15Config,
-        hf_config: dict[str, Any],
-    ) -> None:
-        super().__init__(config=config, hf_config=hf_config)
-
-        self.hidden_size = config.hidden_size
-        self.num_attention_heads = config.num_attention_heads
-        self.num_channels_latents = config.num_channels_latents
-        self.out_channels = config.out_channels or config.in_channels
-        self.patch_size = (config.patch_size_t, config.patch_size, config.patch_size)
-
-        # 1. Latent and condition embedders
-        self.img_in = PatchEmbed(self.patch_size, 
-                                    config.in_channels, 
-                                    self.hidden_size,
-                                    prefix=f"{config.prefix}.img_in")
-        self.image_embedder = HunyuanVideo15ImageProjection(config.image_embed_dim, self.hidden_size)
-
-        self.txt_in = SingleTokenRefiner(config.text_embed_dim,
-                                                   self.hidden_size,
-                                                   config.num_attention_heads,
-                                                   depth=config.num_refiner_layers,
-                                                   dtype=None,
-                                                   prefix=f"{config.prefix}.txt_in")
-
-        self.txt_in_2 = HunyuanVideo15ByT5TextProjection(config.text_embed_2_dim, 2048, self.hidden_size)
-
-        self.time_in = HunyuanVideo15TimeEmbedding(self.hidden_size, use_meanflow=config.use_meanflow)
-
-        self.cond_type_embed = nn.Embedding(3, self.hidden_size)
-
-        # 3. Dual stream transformer blocks
-
-        self.double_blocks = nn.ModuleList(
-            [
-                MMDoubleStreamBlock(
-                    hidden_size=self.hidden_size,
-                    num_attention_heads=config.num_attention_heads,
-                    mlp_ratio=config.mlp_ratio,
-                    local_attn_size=config.local_attn_size,
-                    sink_size=config.sink_size,
-                    dtype=None,
-                    supported_attention_backends=self._supported_attention_backends,
-                    prefix=f"{config.prefix}.double_blocks.{i}"
-                )
-                for i in range(config.num_layers)
-            ]
-        )
-
-        # 5. Output projection
-        self.final_layer = FinalLayer(self.hidden_size,
-                                self.patch_size,
-                                self.out_channels,
-                                prefix=f"{config.prefix}.final_layer")
-
-        self.gradient_checkpointing = False
-
-        self.num_frame_per_block = config.num_frames_per_block
-        self.local_attn_size = config.local_attn_size
-        self.block_mask = None
-
-        self.__post_init__()
-
-    def get_text_and_mask(
-        self,
-        encoder_hidden_states: torch.Tensor,
-        encoder_hidden_states_2: torch.Tensor,
-        encoder_attention_mask: torch.Tensor,
-        encoder_attention_mask_2: torch.Tensor,
-        encoder_hidden_states_image: torch.Tensor,
-        timestep: torch.Tensor,
-    ):
-        batch_size, txt_seq_len = encoder_hidden_states.shape[0], encoder_hidden_states.shape[1]
-        # qwen text embedding
-        encoder_hidden_states = self.txt_in(encoder_hidden_states, timestep, encoder_attention_mask)
-
-        encoder_hidden_states_cond_emb = self.cond_type_embed(
-            torch.zeros_like(encoder_hidden_states[:, :, 0], dtype=torch.long)
-        )
-        encoder_hidden_states = encoder_hidden_states + encoder_hidden_states_cond_emb
-
-        # byt5 text embedding
-        encoder_hidden_states_2 = self.txt_in_2(encoder_hidden_states_2)
-
-        encoder_hidden_states_2_cond_emb = self.cond_type_embed(
-            torch.ones_like(encoder_hidden_states_2[:, :, 0], dtype=torch.long)
-        )
-        encoder_hidden_states_2 = encoder_hidden_states_2 + encoder_hidden_states_2_cond_emb
-
-        # image embed
-        encoder_hidden_states_3 = self.image_embedder(encoder_hidden_states_image)
-        is_t2v = torch.all(encoder_hidden_states_image == 0)
-        if is_t2v:
-            encoder_hidden_states_3 = encoder_hidden_states_3 * 0.0
-            encoder_attention_mask_3 = torch.zeros(
-                (batch_size, encoder_hidden_states_3.shape[1]),
-                dtype=encoder_attention_mask.dtype,
-                device=encoder_attention_mask.device,
-            )
-        else:
-            encoder_attention_mask_3 = torch.ones(
-                (batch_size, encoder_hidden_states_3.shape[1]),
-                dtype=encoder_attention_mask.dtype,
-                device=encoder_attention_mask.device,
-            )
-        encoder_hidden_states_3_cond_emb = self.cond_type_embed(
-            2
-            * torch.ones_like(
-                encoder_hidden_states_3[:, :, 0],
-                dtype=torch.long,
-            )
-        )
-        encoder_hidden_states_3 = encoder_hidden_states_3 + encoder_hidden_states_3_cond_emb
-
-        # reorder and combine text tokens: combine valid tokens first, then padding
-        encoder_attention_mask = encoder_attention_mask.bool()
-        encoder_attention_mask_2 = encoder_attention_mask_2.bool()
-        encoder_attention_mask_3 = encoder_attention_mask_3.bool()
-        new_encoder_hidden_states = []
-        new_encoder_attention_mask = []
-
-        for text, text_mask, text_2, text_mask_2, image, image_mask in zip(
-            encoder_hidden_states,
-            encoder_attention_mask,
-            encoder_hidden_states_2,
-            encoder_attention_mask_2,
-            encoder_hidden_states_3,
-            encoder_attention_mask_3,
-        ):
-            # Concatenate: [valid_image, valid_byt5, valid_mllm, invalid_image, invalid_byt5, invalid_mllm]
-            new_encoder_hidden_states.append(
-                torch.cat(
-                    [
-                        image[image_mask],  # valid image
-                        text_2[text_mask_2],  # valid byt5
-                        text[text_mask],  # valid mllm
-                        image[~image_mask],  # invalid image (zeroed)
-                        torch.zeros_like(text_2[~text_mask_2]),  # invalid byt5 (zeroed)
-                        torch.zeros_like(text[~text_mask]),  # invalid mllm (zeroed)
-                    ],
-                    dim=0,
-                )
-            )
-            # Apply same reordering to attention masks
-            new_encoder_attention_mask.append(
-                torch.cat(
-                    [
-                        image_mask[image_mask],
-                        text_mask_2[text_mask_2],
-                        text_mask[text_mask],
-                        image_mask[~image_mask],
-                        text_mask_2[~text_mask_2],
-                        text_mask[~text_mask],
-                    ],
-                    dim=0,
-                )
-            )
-
-        encoder_hidden_states = torch.stack(new_encoder_hidden_states)
-        encoder_attention_mask = torch.stack(new_encoder_attention_mask)
-        assert encoder_hidden_states.shape[0] == 1
-        return encoder_hidden_states, encoder_attention_mask
-
-
-    @staticmethod
-    def _prepare_teacher_forcing_mask(
-        device: torch.device | str, num_frames: int = 21,
-        frame_seqlen: int = 1560, num_frame_per_block=1,
-        text_seq_len: int = 0
-    ) -> BlockMask:
-        """
-        we will divide the token sequence into the following format
-        [1 latent frame] [1 latent frame] ... [1 latent frame]
-        We use flexattention to construct the attention mask
-        """
-        # debug
-        DEBUG = False
-        if DEBUG:
-            num_frames = 9
-            frame_seqlen = 256
-
-        total_length = num_frames * frame_seqlen * 2
-        total_kv_length = total_length + text_seq_len
-
-        # we do right padding to get to a multiple of 128
-        padded_length = math.ceil(total_length / 128) * 128 - total_length
-        kv_padded_length = math.ceil(total_kv_length / 128) * 128 - total_kv_length
-
-        total_length_tensor = torch.tensor(total_length, device=device)
-        total_kv_length_tensor = torch.tensor(total_kv_length, device=device)
-
-        clean_ends = num_frames * frame_seqlen
-        # for clean context frames, we can construct their flex attention mask based on a [start, end] interval
-        context_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-        # for noisy frames, we need two intervals to construct the flex attention mask [context_start, context_end] [noisy_start, noisy_end]
-        noise_context_starts = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-        noise_context_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-        noise_noise_starts = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-        noise_noise_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-
-        # Block-wise causal mask will attend to all elements that are before the end of the current chunk
-        attention_block_size = frame_seqlen * num_frame_per_block
-        frame_indices = torch.arange(
-            start=0,
-            end=num_frames * frame_seqlen,
-            step=attention_block_size,
-            device=device, dtype=torch.long
-        )
-
-        # attention for clean context frames
-        for start in frame_indices:
-            context_ends[start:start + attention_block_size] = start + attention_block_size
-
-        noisy_image_start_list = torch.arange(
-            num_frames * frame_seqlen, total_length,
-            step=attention_block_size,
-            device=device, dtype=torch.long
-        )
-        noisy_image_end_list = noisy_image_start_list + attention_block_size
-
-        # attention for noisy frames
-        for block_index, (start, end) in enumerate(zip(noisy_image_start_list, noisy_image_end_list)):
-            # attend to noisy tokens within the same block
-            noise_noise_starts[start:end] = start
-            noise_noise_ends[start:end] = end
-            # attend to context tokens in previous blocks
-            # noise_context_starts[start:end] = 0
-            noise_context_ends[start:end] = block_index * attention_block_size
-
-        def attention_mask(b, h, q_idx, kv_idx):
-            # first design the mask for clean frames
-            clean_mask = (q_idx < clean_ends) & (kv_idx < context_ends[q_idx])
-            # then design the mask for noisy frames
-            # noisy frames will attend to all clean preceeding clean frames + itself
-            C1 = (kv_idx < noise_noise_ends[q_idx]) & (kv_idx >= noise_noise_starts[q_idx])
-            C2 = (kv_idx < noise_context_ends[q_idx]) & (kv_idx >= noise_context_starts[q_idx])
-            noise_mask = (q_idx >= clean_ends) & (C1 | C2)
-
-            eye_mask = q_idx == kv_idx
-            text_mask = (kv_idx >= total_length_tensor) & (kv_idx < total_kv_length_tensor)
-            return eye_mask | clean_mask | noise_mask | text_mask
-
-        block_mask = create_block_mask(attention_mask, B=None, H=None, Q_LEN=total_length + padded_length,
-                                       KV_LEN=total_kv_length + kv_padded_length, _compile=False, device=device)
-
-        if DEBUG:
-            print(block_mask)
-            import imageio
-            import numpy as np
-            from torch.nn.attention.flex_attention import create_mask
-
-            mask = create_mask(attention_mask, B=None, H=None, Q_LEN=total_length +
-                                padded_length, KV_LEN=total_kv_length + kv_padded_length, device=device)
-            import cv2
-            mask = cv2.resize(mask[0, 0].cpu().float().numpy(), (1024, 1024))
-            imageio.imwrite("mask_%d.jpg" % (0), np.uint8(255. * mask))
-
-        return block_mask
-    
-    @staticmethod
-    def _prepare_blockwise_causal_attn_mask(
-        device: torch.device | str, num_frames: int = 21,
-        frame_seqlen: int = 1560, num_frame_per_block=1, local_attn_size=-1,
-        text_seq_len: int = 0
-    ) -> BlockMask:
-        """
-        we will divide the token sequence into the following format
-        [1 latent frame] [1 latent frame] ... [1 latent frame]
-        We use flexattention to construct the attention mask
-        """
-        total_length = num_frames * frame_seqlen
-        total_kv_length = total_length + text_seq_len
-
-        total_length_tensor = torch.tensor(total_length, device=device)
-        total_kv_length_tensor = torch.tensor(total_kv_length, device=device)
-
-        # we do right padding to get to a multiple of 128
-        padded_length = math.ceil(total_length / 128) * 128 - total_length
-        kv_padded_length = math.ceil(total_kv_length / 128) * 128 - total_kv_length
-
-        ends = torch.zeros(total_length + padded_length,
-                           device=device, dtype=torch.long)
-
-        # Block-wise causal mask will attend to all elements that are before the end of the current chunk
-        frame_indices = torch.arange(
-            # start=frame_seqlen,
-            start=0,
-            end=total_length,
-            step=frame_seqlen * num_frame_per_block,
-            device=device
-        )
-        # frame_indices = torch.cat([torch.tensor([0], device=device), frame_indices])
-
-        for i, tmp in enumerate(frame_indices):
-            # if i == 0:
-            #     ends[tmp:tmp + frame_seqlen] = tmp + frame_seqlen
-            # else:
-            ends[tmp:tmp + frame_seqlen * num_frame_per_block] = tmp + \
-                frame_seqlen * num_frame_per_block
-
-        def attention_mask(b, h, q_idx, kv_idx):
-            if local_attn_size == -1:
-                return (kv_idx < ends[q_idx]) | (q_idx == kv_idx) | ((kv_idx >= total_length_tensor) & (kv_idx < total_kv_length_tensor))
-            else:
-                return ((kv_idx < ends[q_idx]) & (kv_idx >= (ends[q_idx] - local_attn_size * frame_seqlen))) | (q_idx == kv_idx) | ((kv_idx >= total_length_tensor) & (kv_idx < total_kv_length_tensor))
-            # return ((kv_idx < total_length) & (q_idx < total_length))  | (q_idx == kv_idx) # bidirectional mask
-
-        block_mask = create_block_mask(attention_mask, B=None, H=None, Q_LEN=total_length + padded_length,
-                                       KV_LEN=total_kv_length + kv_padded_length, _compile=False, device=device)
-
-        # if not dist.is_initialized() or dist.get_rank() == 0:
-        #     print(
-        #         f" cache a block wise causal mask with block size of {num_frame_per_block} frames")
-        #     print(block_mask)
-
-        # import imageio
-        # import numpy as np
-        # from torch.nn.attention.flex_attention import create_mask
-
-        # mask = create_mask(attention_mask, B=None, H=None, Q_LEN=total_length +
-        #                    padded_length, KV_LEN=total_length + padded_length, device=device)
-        # import cv2
-        # mask = cv2.resize(mask[0, 0].cpu().float().numpy(), (1024, 1024))
-        # imageio.imwrite("mask_%d.jpg" % (0), np.uint8(255. * mask))
-
-        return block_mask
-
-    def forward_txt(
-        self,
-        encoder_hidden_states: List[torch.Tensor],
-        timestep: torch.LongTensor,
-        encoder_hidden_states_image: Optional[List[torch.Tensor]] = None,
-        encoder_attention_mask: Optional[List[torch.Tensor]] = None,
-        guidance: Optional[torch.Tensor] = None,
-        timestep_r: Optional[torch.LongTensor] = None,
-        attention_kwargs: Optional[Dict[str, Any]] = None,
-        cache_txt: bool = False,
-        **kwargs
-    ):
-        # Check that the timestep is only consisted of 0s
-        assert torch.all(timestep == 0), "Timestep for txt must be only consisted of 0s"
-
-        if cache_txt:
-            _kv_cache_new = []
-            transformer_num_layers = len(self.double_blocks)
-            for _ in range(transformer_num_layers):
-                _kv_cache_new.append(
-                    {"k_vision": None, "v_vision": None, "k_txt": None, "v_txt": None}
-                )
-
-        encoder_hidden_states_image = encoder_hidden_states_image[0]
-        encoder_hidden_states, encoder_hidden_states_2 = encoder_hidden_states
-        encoder_attention_mask, encoder_attention_mask_2 = encoder_attention_mask
-
-        # 2. Conditional embeddings
-        if timestep.dim() == 2:
-            ts_seq_len = timestep.shape[1]
-            temb = self.time_in(timestep.flatten(), timestep_r=timestep_r, timestep_seq_len=ts_seq_len)
-        else:
-            temb = self.time_in(timestep, timestep_r=timestep_r)
-        # temb is [bs, seq_len, inner_dim] if ts_seq_len is not None, otherwise [bs, inner_dim]
-
-        encoder_hidden_states, encoder_attention_mask = self.get_text_and_mask(
-            encoder_hidden_states,
-            encoder_hidden_states_2,
-            encoder_attention_mask,
-            encoder_attention_mask_2,
-            encoder_hidden_states_image,
-            timestep
-        )
-        
-        encoder_hidden_states = encoder_hidden_states[encoder_attention_mask.bool().to(encoder_hidden_states.device)].unsqueeze(0)
-
-        # 4. Transformer blocks
-        for index, block in enumerate(self.double_blocks):
-            encoder_hidden_states, t_kv = block(
-                txt_inference=True,
-                vision_inference=False,
-                txt=encoder_hidden_states,
-                vec=temb,
-                cache_txt=cache_txt,
-            )
-
-            if cache_txt:
-                _kv_cache_new[index]["k_txt"] = t_kv["k_txt"]
-                _kv_cache_new[index]["v_txt"] = t_kv["v_txt"]
-
-        if cache_txt:
-            return _kv_cache_new
-
-    def forward_vision(
-        self,
-        hidden_states: torch.Tensor,
-        timestep: torch.LongTensor,
-        guidance: Optional[torch.Tensor] = None,
-        timestep_r: Optional[torch.LongTensor] = None,
-        attention_kwargs: Optional[Dict[str, Any]] = None,
-        kv_cache: dict | None = None,
-        txt_kv_cache: list | None = None,
-        current_start: int = 0,
-        rope_start_idx: int = 0,
-        clean_hidden_states: Optional[torch.Tensor] = None,
-        aug_timestep: Optional[torch.Tensor] = None,
-    ):
-        assert txt_kv_cache is not None, "txt_kv_cache must be provided"
-
-        batch_size, num_channels, num_frames, height, width = hidden_states.shape
-        p_t, p_h, p_w = self.config.patch_size_t, self.config.patch_size, self.config.patch_size
-        post_patch_num_frames = num_frames // p_t
-        post_patch_height = height // p_h
-        post_patch_width = width // p_w
-
-        # 1. RoPE
-        # Get rotary embeddings
-        freqs_cos, freqs_sin = get_rotary_pos_embed(
-            (post_patch_num_frames, post_patch_height, post_patch_width), self.hidden_size,
-            self.num_attention_heads, self.config.rope_axes_dim, self.config.rope_theta, start_frame=rope_start_idx)
-        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
-
-        # 2. Conditional embeddings
-        if timestep.dim() == 2:
-            ts_seq_len = timestep.shape[1]
-            temb = self.time_in(timestep.flatten(), timestep_r=timestep_r, timestep_seq_len=ts_seq_len)
-        else:
-            temb = self.time_in(timestep, timestep_r=timestep_r)
-        # temb is [bs, seq_len, inner_dim] if ts_seq_len is not None, otherwise [bs, inner_dim]
-
-        hidden_states = self.img_in(hidden_states)
-
-        if clean_hidden_states is not None:
-            clean_hidden_states = self.img_in(clean_hidden_states)
-            hidden_states = torch.cat([clean_hidden_states, hidden_states], dim=1)
-
-            if aug_timestep is None:
-                aug_timestep = torch.zeros_like(timestep)
-            if aug_timestep.dim() == 2:
-                ts_seq_len = aug_timestep.shape[1]
-                aug_temb = self.time_in(aug_timestep.flatten(), timestep_r=timestep_r, timestep_seq_len=ts_seq_len)
-            else:
-                aug_temb = self.time_in(aug_timestep, timestep_r=timestep_r)
-            temb = torch.cat([aug_temb, temb], dim=1)
-
-        # Prepare block-wise causal attention mask
-        if kv_cache is None:
-            if clean_hidden_states is not None:
-                self.block_mask = self._prepare_teacher_forcing_mask(
-                    device=hidden_states.device,
-                    num_frames=num_frames,
-                    frame_seqlen=post_patch_height * post_patch_width,
-                    num_frame_per_block=self.num_frame_per_block,
-                    text_seq_len=txt_kv_cache[0]["k_txt"].shape[1]
-                )
-            else:
-                self.block_mask = self._prepare_blockwise_causal_attn_mask(
-                    device=hidden_states.device,
-                    num_frames=num_frames,
-                    frame_seqlen=post_patch_height * post_patch_width,
-                    num_frame_per_block=self.num_frame_per_block,
-                    local_attn_size=self.local_attn_size,
-                    text_seq_len=txt_kv_cache[0]["k_txt"].shape[1]
-                )
-
-        # 4. Transformer blocks
-        if torch.is_grad_enabled() and self.gradient_checkpointing:
-            for block_index, block in enumerate(self.double_blocks):
-                hidden_states, new_cache = self._gradient_checkpointing_func(
-                    block,
-                    txt_inference=False,
-                    vision_inference=True,
-                    img=hidden_states,
-                    vec=temb,
-                    freqs_cis=freqs_cis,
-                    block_mask=self.block_mask,
-                    kv_cache=kv_cache[block_index] if kv_cache is not None else None,
-                    txt_kv_cache=txt_kv_cache[block_index],
-                    current_start=current_start
-                )
-                if new_cache is not None and kv_cache is not None:
-                    for k in new_cache.keys():
-                        kv_cache[block_index][k] = new_cache[k].clone()
-
-        else:
-            for block_index, block in enumerate(self.double_blocks):
-                hidden_states, new_cache = block(
-                    txt_inference=False,
-                    vision_inference=True,
-                    img=hidden_states,
-                    vec=temb,
-                    freqs_cis=freqs_cis,
-                    block_mask=self.block_mask,
-                    kv_cache=kv_cache[block_index] if kv_cache is not None else None,
-                    txt_kv_cache=txt_kv_cache[block_index],
-                    current_start=current_start
-                )
-                if new_cache is not None and kv_cache is not None:
-                    for k in new_cache.keys():
-                        kv_cache[block_index][k] = new_cache[k].clone()
-
-        if clean_hidden_states is not None:
-            hidden_states = hidden_states[:, hidden_states.shape[1] // 2:]
-
-        # Final layer processing
-        hidden_states = self.final_layer(hidden_states, temb)
-        # Unpatchify to get original shape
-        hidden_states = unpatchify(hidden_states, post_patch_num_frames, post_patch_height, post_patch_width, self.patch_size, self.out_channels)
-
-        return hidden_states, kv_cache
-
-    def forward(
-        self,
-        txt_inference=False,
-        vision_inference=False,
-        **kwargs,
-    ):
-        if txt_inference:
-            return self.forward_txt(**kwargs)
-        elif vision_inference:
-            return self.forward_vision(**kwargs)
-        else:
-            raise ValueError("txt_inference and vision_inference cannot be both False")
@@ -88,25 +88,10 @@ class CausalWanSelfAttention(nn.Module):
            cache_start = current_start

        cos, sin = freqs_cis
+        roped_query = _apply_rotary_emb(q, cos, sin, is_neox_style=False).type_as(v)
+        roped_key = _apply_rotary_emb(k, cos, sin, is_neox_style=False).type_as(v)

        if kv_cache is None:
-            is_tf = (q.shape[1] == cos.shape[0] * 2)
-            if is_tf:
-                q_chunk = torch.chunk(q, 2, dim=1)
-                k_chunk = torch.chunk(k, 2, dim=1)
-                roped_query = []
-                roped_key = []
-                for ii in range(2):
-                    rq = _apply_rotary_emb(q_chunk[ii], cos, sin, is_neox_style=False).type_as(v)
-                    rk = _apply_rotary_emb(k_chunk[ii], cos, sin, is_neox_style=False).type_as(v)
-                    roped_query.append(rq)
-                    roped_key.append(rk)
-                roped_query = torch.cat(roped_query, dim=1)
-                roped_key = torch.cat(roped_key, dim=1)
-            else:
-                roped_query = _apply_rotary_emb(q, cos, sin, is_neox_style=False).type_as(v)
-                roped_key = _apply_rotary_emb(k, cos, sin, is_neox_style=False).type_as(v)
-
            # Padding for flex attention
            padded_length = math.ceil(q.shape[1] / 128) * 128 - q.shape[1]
            padded_roped_query = torch.cat(
@@ -135,8 +120,6 @@ class CausalWanSelfAttention(nn.Module):
                block_mask=block_mask
            )[:, :, :-padded_length].transpose(2, 1)
        else:
-            roped_query = _apply_rotary_emb(q, cos, sin, is_neox_style=False).type_as(v)
-            roped_key = _apply_rotary_emb(k, cos, sin, is_neox_style=False).type_as(v)
            frame_seqlen = q.shape[1]
            current_end = current_start + roped_query.shape[1]
            sink_tokens = self.sink_size * frame_seqlen
@@ -313,13 +296,10 @@ class CausalWanTransformerBlock(nn.Module):
                                 crossattn_cache=crossattn_cache)
        norm_hidden_states, hidden_states = self.cross_attn_residual_norm(
            hidden_states, attn_output, 1, c_shift_msa, c_scale_msa)
-        norm_hidden_states, hidden_states = norm_hidden_states.to(
-            orig_dtype), hidden_states.to(orig_dtype)

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

        return hidden_states

@@ -398,94 +378,6 @@ class CausalWanTransformer3DModel(BaseDiT):

        self.__post_init__()

-    @staticmethod
-    def _prepare_teacher_forcing_mask(
-        device: torch.device | str, num_frames: int = 21,
-        frame_seqlen: int = 1560, num_frame_per_block=1
-    ) -> BlockMask:
-        """
-        we will divide the token sequence into the following format
-        [1 latent frame] [1 latent frame] ... [1 latent frame]
-        We use flexattention to construct the attention mask
-        """
-        # debug
-        DEBUG = False
-        if DEBUG:
-            num_frames = 9
-            frame_seqlen = 256
-
-        total_length = num_frames * frame_seqlen * 2
-
-        # we do right padding to get to a multiple of 128
-        padded_length = math.ceil(total_length / 128) * 128 - total_length
-
-        clean_ends = num_frames * frame_seqlen
-        # for clean context frames, we can construct their flex attention mask based on a [start, end] interval
-        context_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-        # for noisy frames, we need two intervals to construct the flex attention mask [context_start, context_end] [noisy_start, noisy_end]
-        noise_context_starts = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-        noise_context_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-        noise_noise_starts = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-        noise_noise_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
-
-        # Block-wise causal mask will attend to all elements that are before the end of the current chunk
-        attention_block_size = frame_seqlen * num_frame_per_block
-        frame_indices = torch.arange(
-            start=0,
-            end=num_frames * frame_seqlen,
-            step=attention_block_size,
-            device=device, dtype=torch.long
-        )
-
-        # attention for clean context frames
-        for start in frame_indices:
-            context_ends[start:start + attention_block_size] = start + attention_block_size
-
-        noisy_image_start_list = torch.arange(
-            num_frames * frame_seqlen, total_length,
-            step=attention_block_size,
-            device=device, dtype=torch.long
-        )
-        noisy_image_end_list = noisy_image_start_list + attention_block_size
-
-        # attention for noisy frames
-        for block_index, (start, end) in enumerate(zip(noisy_image_start_list, noisy_image_end_list)):
-            # attend to noisy tokens within the same block
-            noise_noise_starts[start:end] = start
-            noise_noise_ends[start:end] = end
-            # attend to context tokens in previous blocks
-            # noise_context_starts[start:end] = 0
-            noise_context_ends[start:end] = block_index * attention_block_size
-
-        def attention_mask(b, h, q_idx, kv_idx):
-            # first design the mask for clean frames
-            clean_mask = (q_idx < clean_ends) & (kv_idx < context_ends[q_idx])
-            # then design the mask for noisy frames
-            # noisy frames will attend to all clean preceeding clean frames + itself
-            C1 = (kv_idx < noise_noise_ends[q_idx]) & (kv_idx >= noise_noise_starts[q_idx])
-            C2 = (kv_idx < noise_context_ends[q_idx]) & (kv_idx >= noise_context_starts[q_idx])
-            noise_mask = (q_idx >= clean_ends) & (C1 | C2)
-
-            eye_mask = q_idx == kv_idx
-            return eye_mask | clean_mask | noise_mask
-
-        block_mask = create_block_mask(attention_mask, B=None, H=None, Q_LEN=total_length + padded_length,
-                                       KV_LEN=total_length + padded_length, _compile=False, device=device)
-
-        if DEBUG:
-            print(block_mask)
-            import imageio
-            import numpy as np
-            from torch.nn.attention.flex_attention import create_mask
-
-            mask = create_mask(attention_mask, B=None, H=None, Q_LEN=total_length +
-                                padded_length, KV_LEN=total_length + padded_length, device=device)
-            import cv2
-            mask = cv2.resize(mask[0, 0].cpu().float().numpy(), (1024, 1024))
-            imageio.imwrite("mask_%d.jpg" % (0), np.uint8(255. * mask))
-
-        return block_mask
-
    @staticmethod
    def _prepare_blockwise_causal_attn_mask(
        device: torch.device | str, num_frames: int = 21,
@@ -562,7 +454,6 @@ class CausalWanTransformer3DModel(BaseDiT):
        This function will be run for num_frame times.
        Process the latent frames one by one (1560 tokens each)
        """
-        from fastvideo.platforms import current_platform

        orig_dtype = hidden_states.dtype
        if not isinstance(encoder_hidden_states, torch.Tensor):
@@ -661,7 +552,6 @@ class CausalWanTransformer3DModel(BaseDiT):
                encoder_hidden_states_image: torch.Tensor | list[torch.Tensor]
                | None = None,
                start_frame: int = 0,
-                clean_hidden_states: torch.Tensor | None = None,
                **kwargs) -> torch.Tensor:

        orig_dtype = hidden_states.dtype
@@ -673,10 +563,6 @@ class CausalWanTransformer3DModel(BaseDiT):
        else:
            encoder_hidden_states_image = None

-        if timestep.ndim == 1:
-            # [B] -> [B, F]
-            timestep = timestep.unsqueeze(1).repeat(1, hidden_states.shape[2])
-
        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
@@ -703,21 +589,13 @@ class CausalWanTransformer3DModel(BaseDiT):

        # Construct blockwise causal attn mask
        if self.block_mask is None:
-            if clean_hidden_states is not None:
-                self.block_mask = self._prepare_teacher_forcing_mask(
-                    device=hidden_states.device,
-                    num_frames=num_frames,
-                    frame_seqlen=post_patch_height * post_patch_width,
-                    num_frame_per_block=self.num_frame_per_block,
-                )
-            else:
-                self.block_mask = self._prepare_blockwise_causal_attn_mask(
-                    device=hidden_states.device,
-                    num_frames=num_frames,
-                    frame_seqlen=post_patch_height * post_patch_width,
-                    num_frame_per_block=self.num_frame_per_block,
-                    local_attn_size=self.local_attn_size
-                )
+            self.block_mask = self._prepare_blockwise_causal_attn_mask(
+                device=hidden_states.device,
+                num_frames=num_frames,
+                frame_seqlen=post_patch_height * post_patch_width,
+                num_frame_per_block=self.num_frame_per_block,
+                local_attn_size=self.local_attn_size
+            )

        hidden_states = self.patch_embedding(hidden_states)
        grid_sizes = torch.stack(
@@ -730,17 +608,6 @@ class CausalWanTransformer3DModel(BaseDiT):
                        timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
        timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=timestep.shape)

-        if clean_hidden_states is not None:
-            clean_hidden_states = self.patch_embedding(clean_hidden_states)
-            clean_hidden_states = clean_hidden_states.flatten(2).transpose(1, 2)
-            hidden_states = torch.cat([clean_hidden_states, hidden_states], dim=1)
-
-            aug_timestep = torch.zeros_like(timestep)
-            temb_clean = self.condition_embedder.time_embedder(aug_timestep.flatten(), None)
-            timestep_proj_clean = self.condition_embedder.time_modulation(temb_clean)
-            timestep_proj_clean = timestep_proj_clean.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=aug_timestep.shape)
-            timestep_proj = torch.cat([timestep_proj_clean, timestep_proj], dim=1)
-
        if encoder_hidden_states_image is not None:
            encoder_hidden_states = torch.concat(
                [encoder_hidden_states_image, encoder_hidden_states], dim=1)
@@ -764,9 +631,6 @@ class CausalWanTransformer3DModel(BaseDiT):
                                        timestep_proj, freqs_cis,
                                        block_mask=self.block_mask)

-        if clean_hidden_states is not None:
-            hidden_states = hidden_states[:, hidden_states.shape[1] // 2:]
-
        # 5. Output norm, projection & unpatchify
        temb = temb.unflatten(dim=0, sizes=timestep.shape).unsqueeze(2)
        shift, scale = (self.scale_shift_table.unsqueeze(1) + temb).chunk(2,
@@ -781,13 +645,12 @@ class CausalWanTransformer3DModel(BaseDiT):
    def forward(
        self,
        *args,
-        clean_hidden_states: torch.Tensor | None = None,
        **kwargs
    ):
        if kwargs.get('kv_cache', None) is not None:
            return self._forward_inference(*args, **kwargs)
        else:
-            return self._forward_train(*args, clean_hidden_states=clean_hidden_states, **kwargs)
+            return self._forward_train(*args, **kwargs)


    def unpatchify(self, x, grid_sizes):
@@ -814,4 +677,7 @@ 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
+        return out
+
+# Entry point for model registry
+EntryClass = CausalWanTransformer3DModel
@@ -723,4 +723,7 @@ 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
+        return hidden_states
+
+# Entry point for model registry
+EntryClass = CosmosTransformer3DModel
@@ -959,3 +959,5 @@ class Cosmos25Transformer3DModel(BaseDiT):

        return hidden_states

+# Entry point for model registry
+EntryClass = Cosmos25Transformer3DModel
@@ -940,3 +940,6 @@ 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
@@ -127,9 +127,8 @@ class HunyuanVideo15TimeEmbedding(nn.Module):
        self,
        timestep: torch.Tensor,
        timestep_r: Optional[torch.Tensor] = None,
-        timestep_seq_len: int | None = None,
    ) -> torch.Tensor:
-        timesteps_emb = self.timestep_embedder(timestep, timestep_seq_len)
+        timesteps_emb = self.timestep_embedder(timestep)

        if timestep_r is not None:
            timesteps_emb_r = self.timestep_embedder_r(timestep_r)
@@ -474,7 +473,6 @@ class HunyuanVideo15Transformer3DModel(CachableDiT):
        guidance: Optional[torch.Tensor] = None,
        timestep_r: Optional[torch.LongTensor] = None,
        attention_kwargs: Optional[Dict[str, Any]] = None,
-        **kwargs
    ):
        encoder_hidden_states_image = encoder_hidden_states_image[0]
        encoder_hidden_states, encoder_hidden_states_2 = encoder_hidden_states
@@ -496,12 +494,7 @@ class HunyuanVideo15Transformer3DModel(CachableDiT):
        freqs_cis = (freqs_cos, freqs_sin) if freqs_cos is not None else None

        # 2. Conditional embeddings
-        if timestep.dim() == 2:
-            ts_seq_len = timestep.shape[1]
-            temb = self.time_in(timestep.flatten(), timestep_r=timestep_r, timestep_seq_len=ts_seq_len)
-        else:
-            temb = self.time_in(timestep, timestep_r=timestep_r)
-        # temb is [bs, seq_len, inner_dim] if ts_seq_len is not None, otherwise [bs, inner_dim]
+        temb = self.time_in(timestep, timestep_r=timestep_r)

        hidden_states = self.img_in(hidden_states)
        hidden_states, original_seq_len = sequence_model_parallel_shard(hidden_states, dim=1)
@@ -705,7 +698,6 @@ class SingleTokenRefiner(nn.Module):
        else:
            mask_float = mask.float().unsqueeze(-1)  # [B, L, 1]
            context_aware_representations = (x * mask_float).sum(dim=1) / mask_float.sum(dim=1)
-        context_aware_representations = context_aware_representations.to(original_dtype)

        context_aware_representations = self.c_embedder(
            context_aware_representations)
@@ -858,13 +850,9 @@ class FinalLayer(nn.Module):
    def forward(self, x, c):
        # What the heck HF? Why you change the scale and shift order here???
        scale, shift = self.adaLN_modulation(c).chunk(2, dim=-1)
-        if c.dim() == 3:
-            # [bs, seq_len, inner_dim]
-            num_frames = scale.shape[1]
-            frame_seqlen = x.shape[1] // num_frames
-            x = (self.norm_final(x).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * (1.0 + scale.unsqueeze(2)) + shift.unsqueeze(2)).flatten(1, 2)
-        else:
-            # [bs, inner_dim]
-            x = self.norm_final(x) * (1.0 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+        x = self.norm_final(x) * (1.0 + scale.unsqueeze(1)) + shift.unsqueeze(1)
        x, _ = self.linear(x)
-        return x
+        return x
+
+# Entry point for model registry
+EntryClass = HunyuanVideo15Transformer3DModel
@@ -22,3 +22,6 @@ __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, # hyworld
-                    Ks_seq, # hyworld
+                    viewmats_seq,
+                    Ks_seq,
                )
        else:
            for block in self.double_blocks:
@@ -549,8 +549,8 @@ class HYWorldTransformer3DModel(HunyuanVideo15Transformer3DModel):
                    temb_txt,
                    freqs_cis,
                    seq_attention_mask,
-                    viewmats=viewmats_seq, # hyworld
-                    Ks=Ks_seq, # hyworld
+                    viewmats_seq,
+                    Ks_seq,
                )


@@ -1134,3 +1134,5 @@ class LongCatTransformer3DModel(CachableDiT):
        )
        return x

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

 from fastvideo.attention.backends.sdpa import SDPAMetadata
-from fastvideo.attention.layer import LocalAttention
+from fastvideo.attention.layer import DistributedAttention, LocalAttention
 from fastvideo.configs.models.dits import LTX2VideoConfig
-from fastvideo.forward_context import get_forward_context, set_forward_context
+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.models.dits.base import CachableDiT
 from fastvideo.platforms import AttentionBackendEnum

+logger = init_logger(__name__)
+

 def get_timestep_embedding(
    timesteps: torch.Tensor,
@@ -295,6 +306,12 @@ 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
@@ -366,6 +383,12 @@ 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,
@@ -515,6 +538,44 @@ 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:
@@ -918,6 +979,205 @@ 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__(
@@ -945,17 +1205,19 @@ 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 = LocalAttention(
+        self.attn = LTXLocalAttention(
            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 = LocalAttention(
+        self.attn_masked = LTXLocalAttention(
            num_heads=heads,
            head_size=dim_head,
+            rope_type=rope_type,
            dropout_rate=0,
            softmax_scale=None,
            causal=False,
@@ -978,15 +1240,13 @@ class LTXSelfAttention(nn.Module):
        q = self.q_norm(q)
        k = self.k_norm(k)

-        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)
-
+        # RoPE is applied inside LTXLocalAttention
        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)
@@ -1008,9 +1268,93 @@ class LTXSelfAttention(nn.Module):
                attn_metadata=attn_metadata,
                forward_batch=forward_batch,
            ):
-                out = self.attn_masked(q, k, v)
+                out = self.attn_masked(q, k, v, ltx_freqs_cis=pe, ltx_k_freqs_cis=k_pe)
        else:
-            out = self.attn(q, k, v)
+            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 = out.reshape(b, q_len, -1)
        return self.to_out(out)

@@ -1025,12 +1369,31 @@ 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:
-            self.attn1 = LTXSelfAttention(
+            # 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(
                query_dim=video.dim,
                context_dim=None,
                heads=video.heads,
@@ -1039,7 +1402,8 @@ class BasicAVTransformerBlock(torch.nn.Module):
                rope_type=rope_type,
                supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA),
            )
-            self.attn2 = LTXSelfAttention(
+            # Text cross-attention - always local (text is replicated)
+            self.attn2 = CrossAttnCls(
                query_dim=video.dim,
                context_dim=video.context_dim,
                heads=video.heads,
@@ -1052,7 +1416,17 @@ class BasicAVTransformerBlock(torch.nn.Module):
            self.scale_shift_table = torch.nn.Parameter(torch.empty(6, video.dim))

        if audio is not None:
-            self.audio_attn1 = LTXSelfAttention(
+            # 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(
                query_dim=audio.dim,
                context_dim=None,
                heads=audio.heads,
@@ -1061,7 +1435,8 @@ class BasicAVTransformerBlock(torch.nn.Module):
                rope_type=rope_type,
                supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA),
            )
-            self.audio_attn2 = LTXSelfAttention(
+            # Text cross-attention - always local (text is replicated)
+            self.audio_attn2 = CrossAttnCls(
                query_dim=audio.dim,
                context_dim=audio.context_dim,
                heads=audio.heads,
@@ -1074,6 +1449,8 @@ 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,
@@ -1083,6 +1460,8 @@ 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,
@@ -1131,7 +1510,17 @@ 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

@@ -1145,7 +1534,12 @@ 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
-            vx = vx + self.attn1(norm_vx, pe=video.positional_embeddings) * vgate_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.attn2(
                torch.nn.functional.rms_norm(vx, (vx.shape[-1],), eps=self.norm_eps),
                context=video.context,
@@ -1157,7 +1551,12 @@ 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
-            ax = ax + self.audio_attn1(norm_ax, pe=audio.positional_embeddings) * agate_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_attn2(
                torch.nn.functional.rms_norm(ax, (ax.shape[-1],), eps=self.norm_eps),
                context=audio.context,
@@ -1197,28 +1596,91 @@ 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
-                vx = vx + (
-                    self.audio_to_video_attn(
-                        vx_scaled,
-                        context=ax_scaled,
-                        pe=video.cross_positional_embeddings,
-                        k_pe=audio.cross_positional_embeddings,
+                # 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
                    )
-                    * 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
-                ax = ax + (
-                    self.video_to_audio_attn(
-                        ax_scaled,
-                        context=vx_scaled,
-                        pe=audio.cross_positional_embeddings,
-                        k_pe=video.cross_positional_embeddings,
+                # 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
                    )
-                    * gate_out_v2a
-                )

        if run_vx:
            vshift_mlp, vscale_mlp, vgate_mlp = self.get_ada_values(
@@ -1289,6 +1751,8 @@ 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
@@ -1340,6 +1804,8 @@ 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(
@@ -1469,6 +1935,8 @@ 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(
@@ -1490,6 +1958,7 @@ 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(
@@ -1498,6 +1967,8 @@ 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)
            ]
@@ -1507,9 +1978,16 @@ 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, audio = block(
+                video=video,
+                audio=audio,
+                video_attention_mask=video_attention_mask,
+                audio_attention_mask=audio_attention_mask,
+            )
        return video, audio

    def _process_output(
@@ -1533,7 +2011,17 @@ 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"
@@ -1551,7 +2039,12 @@ 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_out, audio_out = self._process_transformer_blocks(
+            video_args,
+            audio_args,
+            video_attention_mask=video_attention_mask,
+            audio_attention_mask=audio_attention_mask,
+        )

        vx = (
            self._process_output(
@@ -1588,6 +2081,24 @@ 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,
@@ -1612,6 +2123,8 @@ 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])
@@ -1627,6 +2140,7 @@ 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", "")
@@ -1702,10 +2216,12 @@ class LTX2Transformer3DModel(CachableDiT):
        if isinstance(encoder_hidden_states, list):
            encoder_hidden_states = encoder_hidden_states[0]

-        video_shape = VideoLatentShape.from_torch_shape(hidden_states.shape)
-        positions = self.patchifier.get_patch_grid_bounds(
-            video_shape, device=hidden_states.device
-        )
+        # 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
        fps = None
        try:
            forward_ctx = get_forward_context()
@@ -1717,17 +2233,60 @@ 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)
-        latents = self.patchifier.patchify(hidden_states)
+
+        # 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)
+
        video_modality = Modality(
            enabled=True,
            latent=latents,
-            timesteps=timestep,
+            timesteps=video_timestep,
            positions=positions,
            context=encoder_hidden_states,
            context_mask=encoder_attention_mask,
@@ -1745,19 +2304,45 @@ 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_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_positions = self.audio_patchifier.get_patch_grid_bounds(
-                audio_shape, device=audio_hidden_states.device)
-            audio_latents = self.audio_patchifier.patchify(
-                audio_hidden_states)
+                audio_shape, device=audio_hidden_states.device
+            )
+
            audio_modality = Modality(
                enabled=True,
                latent=audio_latents,
-                timesteps=audio_timestep,
+                timesteps=sharded_audio_timestep,
                positions=audio_positions,
                context=audio_encoder_hidden_states,
                context_mask=audio_encoder_attention_mask,
@@ -1775,10 +2360,16 @@ 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,
@@ -1791,6 +2382,20 @@ 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:
@@ -1798,3 +2403,6 @@ 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
@@ -9,3 +9,6 @@ __all__ = [
    "CausalMatrixGameTransformerBlock",
    "ActionModule",
 ]
+
+# Entry point for model registry
+EntryClass = [MatrixGameWanModel, CausalMatrixGameWanModel]
@@ -9,20 +9,32 @@ 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
@@ -41,11 +53,12 @@ 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)
@@ -53,7 +66,7 @@ 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,
    ):
@@ -61,17 +74,25 @@ 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):
@@ -87,7 +108,7 @@ 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
@@ -101,7 +122,7 @@ 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)

@@ -112,19 +133,20 @@ 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 {}

@@ -140,7 +162,8 @@ 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
@@ -161,28 +184,28 @@ 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
@@ -204,7 +227,7 @@ class MatrixGameTransformerBlock(nn.Module):
        temb: torch.Tensor,
        freqs_cis: tuple[torch.Tensor, torch.Tensor],
        # Action Module specific args
-        grid_sizes: torch.Tensor | None = None,
+        grid_sizes: torch.Tensor,
        mouse_cond: torch.Tensor | None = None,
        keyboard_cond: torch.Tensor | None = None,
    ) -> torch.Tensor:
@@ -214,9 +237,16 @@ 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)
@@ -227,13 +257,20 @@ 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)
@@ -249,9 +286,10 @@ 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)
@@ -260,18 +298,24 @@ 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:
@@ -280,9 +324,12 @@ class MatrixGameTransformerBlock(nn.Module):
                # ActionModule implementation takes hidden_states directly
                hidden_states = self.action_model(
                    hidden_states,
-                    grid_sizes[0], grid_sizes[1], grid_sizes[2],
-                    mouse_cond, keyboard_cond,
-                    num_frame_per_block=grid_sizes[0],
+                    int(grid_sizes[0]),
+                    int(grid_sizes[1]),
+                    int(grid_sizes[2]),
+                    mouse_cond,
+                    keyboard_cond,
+                    num_frame_per_block=int(grid_sizes[0]),
                )
        # =================================================

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

        return hidden_states

+
 _DEFAULT_MATRIXGAME_CONFIG = MatrixGameWanVideoConfig()


@@ -302,15 +350,23 @@ 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
@@ -320,10 +376,12 @@ 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(
@@ -333,57 +391,73 @@ 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
@@ -393,18 +467,25 @@ 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)
@@ -412,8 +493,14 @@ 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:
@@ -429,25 +516,30 @@ 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,
@@ -456,19 +548,27 @@ 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
@@ -685,3 +685,6 @@ 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
@@ -8,8 +8,6 @@ import numpy as np
 import torch
 import torch.nn as nn

-from einops import repeat
-
 import fastvideo.envs as envs
 from fastvideo.attention import (DistributedAttention, DistributedAttention_VSA,
                                 LocalAttention)
@@ -154,32 +152,6 @@ class WanSelfAttention(nn.Module):
        """
        pass

-class WanGanCrossAttention(WanSelfAttention):
-
-    def forward(self, x, context, context_lens=None, crossattn_cache=None):
-        r"""
-        Args:
-            x(Tensor): Shape [B, L1, C]
-            context(Tensor): Shape [B, L2, C]
-            context_lens(Tensor): Shape [B]
-            crossattn_cache (List[dict], *optional*): Contains the cached key and value tensors for context embedding.
-        """
-        b, n, d = x.size(0), self.num_heads, self.head_dim
-
-        # compute query, key, value
-        qq = self.norm_q(self.to_q(context)[0]).view(b, 1, -1, d)
-
-        kk = self.norm_k(self.to_k(x)[0]).view(b, -1, n, d)
-        vv = self.to_v(x)[0].view(b, -1, n, d)
-
-        # compute attention
-        x = self.attn(qq, kk, vv)
-
-        # output
-        x = x.flatten(2)
-        x, _ = self.to_out(x)
-        return x
-

 class WanT2VCrossAttention(WanSelfAttention):

@@ -273,98 +245,6 @@ class WanI2VCrossAttention(WanSelfAttention):
        x, _ = self.to_out(x)
        return x

-class GanAttentionBlock(nn.Module):
-
-    def __init__(self,
-                 dim=1536,
-                 ffn_dim=8192,
-                 num_heads=12,
-                 window_size=(-1, -1),
-                 qk_norm=True,
-                 cross_attn_norm=True,
-                 eps=1e-6):
-        super().__init__()
-        self.dim = dim
-        self.ffn_dim = ffn_dim
-        self.num_heads = num_heads
-        self.window_size = window_size
-        self.qk_norm = qk_norm
-        self.cross_attn_norm = cross_attn_norm
-        self.eps = eps
-
-        # layers
-        # self.norm1 = WanLayerNorm(dim, eps)
-        # self.self_attn = WanSelfAttention(dim, num_heads, window_size, qk_norm,
-        #   eps)
-        self.norm3 = FP32LayerNorm(
-            dim, eps,
-            elementwise_affine=True) if cross_attn_norm else nn.Identity()
-
-        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=False)
-        self.ffn = nn.Sequential(
-            nn.Linear(dim, ffn_dim), nn.GELU(approximate='tanh'),
-            nn.Linear(ffn_dim, dim))
-
-        self.cross_attn = WanGanCrossAttention(dim, num_heads,
-                                               (-1, -1),
-                                               qk_norm,
-                                               eps)
-
-        # modulation
-        # self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
-
-    def forward(
-        self,
-        hidden_states,
-        encoder_hidden_states,
-        # seq_lens,
-        # grid_sizes,
-        # freqs,
-        # context,
-        # context_lens,
-    ):
-        r"""
-        Args:
-            x(Tensor): Shape [B, L, C]
-            e(Tensor): Shape [B, 6, C]
-            seq_lens(Tensor): Shape [B], length of each sequence in batch
-            grid_sizes(Tensor): Shape [B, 3], the second dimension contains (F, H, W)
-            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
-        """
-        # assert e.dtype == torch.float32
-        # with amp.autocast(dtype=torch.float32):
-        # e = (self.modulation + e).chunk(6, dim=1)
-        # assert e[0].dtype == torch.float32
-
-        # # self-attention
-        # y = self.self_attn(
-        #     self.norm1(x) * (1 + e[1]) + e[0], seq_lens, grid_sizes,
-        #     freqs)
-        # # with amp.autocast(dtype=torch.float32):
-        # x = x + y * e[2]
-
-        # cross-attention & ffn function
-        def cross_attn_ffn(x, context):
-            token = context + self.cross_attn(self.norm3(x), context)
-            y = self.ffn(self.norm2(token)) + token  # * (1 + e[4]) + e[3])
-            # with amp.autocast(dtype=torch.float32):
-            # x = x + y * e[5]
-            return y
-
-        x = cross_attn_ffn(hidden_states, encoder_hidden_states)
-        return x
-
-class RegisterTokens(nn.Module):
-    def __init__(self, num_registers: int, dim: int):
-        super().__init__()
-        self.register_tokens = nn.Parameter(torch.randn(num_registers, dim) * 0.02)
-        self.rms_norm = RMSNorm(dim, eps=1e-6)
-
-    def forward(self):
-        return self.rms_norm(self.register_tokens)
-
-    def reset_parameters(self):
-        nn.init.normal_(self.register_tokens, std=0.02)

 class WanTransformerBlock(nn.Module):

@@ -750,26 +630,6 @@ class WanTransformer3DModel(CachableDiT):
        self.cnt = 0
        self.__post_init__()

-    def adding_cls_branch(self, atten_dim=1536, num_class=4, time_embed_dim=0) -> None:
-        self._cls_pred_branch = nn.Sequential(
-            # Input: [B, 384, 21, 60, 104]
-            nn.LayerNorm(atten_dim * 3 + time_embed_dim),
-            nn.Linear(atten_dim * 3 + time_embed_dim, 1536),
-            nn.SiLU(),
-            nn.Linear(atten_dim, num_class)
-        )
-        self._cls_pred_branch.requires_grad_(True)
-        num_registers = 3
-        self._register_tokens = RegisterTokens(num_registers=num_registers, dim=atten_dim)
-        self._register_tokens.requires_grad_(True)
-        
-        gan_ca_blocks = []
-        for _ in range(num_registers):
-            block = GanAttentionBlock()
-            gan_ca_blocks.append(block)
-        self._gan_ca_blocks = nn.ModuleList(gan_ca_blocks)
-        self._gan_ca_blocks.requires_grad_(True)
-
    def forward(self,
                hidden_states: torch.Tensor,
                encoder_hidden_states: torch.Tensor | list[torch.Tensor],
@@ -777,8 +637,6 @@ class WanTransformer3DModel(CachableDiT):
                encoder_hidden_states_image: torch.Tensor | list[torch.Tensor]
                | None = None,
                guidance=None,
-                classify_mode=False,
-                concat_time_embeddings=False,
                **kwargs) -> torch.Tensor:
        forward_batch = get_forward_context().forward_batch
        enable_teacache = forward_batch is not None and forward_batch.enable_teacache
@@ -869,16 +727,6 @@ class WanTransformer3DModel(CachableDiT):

        assert encoder_hidden_states.dtype == orig_dtype

-        final_x = None
-        if classify_mode:
-            assert self._register_tokens is not None
-            assert self._gan_ca_blocks is not None
-            assert self._cls_pred_branch is not None
-
-            final_x = []
-            registers = repeat(self._register_tokens(), "n d -> b n d", b=hidden_states.shape[0])
-
-        gan_idx = 0
        # 4. Transformer blocks
        # if caching is enabled, we might be able to skip the forward pass
        should_skip_forward = self.should_skip_forward_for_cached_states(
@@ -892,33 +740,19 @@ class WanTransformer3DModel(CachableDiT):
            if enable_teacache:
                original_hidden_states = hidden_states.clone()

-            for ii, block in enumerate(self.blocks):
-                if torch.is_grad_enabled() and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                for block in self.blocks:
                    hidden_states = self._gradient_checkpointing_func(
                        block, hidden_states, encoder_hidden_states,
                        timestep_proj, freqs_cis, attention_mask)
-                else:
+            else:
+                for block in self.blocks:
                    hidden_states = block(hidden_states, encoder_hidden_states,
-                                            timestep_proj, freqs_cis, attention_mask)
-
-                if classify_mode and ii in [13, 21, 29]:
-                    gan_token = registers[:, gan_idx: gan_idx + 1]
-                    final_x.append(self._gan_ca_blocks[gan_idx](hidden_states, gan_token))
-                    gan_idx += 1
+                                              timestep_proj, freqs_cis, attention_mask)
            # if teacache is enabled, we need to cache the original hidden states

            if enable_teacache:
                self.maybe_cache_states(hidden_states, original_hidden_states)
-
-        if classify_mode:
-            final_x = torch.cat(final_x, dim=1)
-            if concat_time_embeddings:
-                cls_input = torch.cat([final_x.float(), (10 * temb[:, None, :]).float()], dim=1).view(final_x.shape[0], -1)
-                final_x = self._cls_pred_branch(cls_input)
-            else:
-                cls_input = final_x.float().view(final_x.shape[0], -1)
-                final_x = self._cls_pred_branch(cls_input)
-
        # 5. Output norm, projection & unpatchify
        if temb.dim() == 3:
            # batch_size, seq_len, inner_dim (wan 2.2 ti2v)
@@ -944,8 +778,6 @@ class WanTransformer3DModel(CachableDiT):
        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)

-        if classify_mode:
-            return output, torch.nan_to_num(final_x)
        return output

    def maybe_cache_states(self, hidden_states: torch.Tensor,
@@ -1039,4 +871,6 @@ 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
@@ -631,3 +631,6 @@ class CLIPVisionModel(ImageEncoder):
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params
+
+# Entry point for model registry
+EntryClass = [CLIPTextModel, CLIPVisionModel]
@@ -561,3 +561,6 @@ 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
@@ -429,3 +429,6 @@ class LlamaModel(TextEncoder):
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params
+
+# Entry point for model registry
+EntryClass = LlamaModel
@@ -385,3 +385,6 @@ 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
@@ -351,3 +351,5 @@ class Reason1TextEncoder(TextEncoder):
            tensor.std(dim=-1, keepdim=True) + 1e-8
        )

+# Entry point for model registry
+EntryClass = Reason1TextEncoder
@@ -417,3 +417,6 @@ class SiglipVisionModel(ImageEncoder):
            loaded_params.add(name)
            
        return loaded_params
+
+# Entry point for model registry
+EntryClass = SiglipVisionModel
@@ -583,3 +583,6 @@ 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
@@ -719,3 +719,6 @@ class UMT5EncoderModel(TextEncoder):
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params
+
+# Entry point for model registry
+EntryClass = [UMT5EncoderModel, T5EncoderModel]
@@ -79,6 +79,7 @@ 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"),
@@ -90,6 +91,8 @@ 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:
@@ -391,8 +394,6 @@ class TextEncoderLoader(ComponentLoader):

            from fastvideo.platforms import current_platform

-            logger.info("Loading text encoder with cpu_offload: %s", use_cpu_offload)
-
            if use_cpu_offload:
                pin_cpu_memory = fastvideo_args.pin_cpu_memory and is_pin_memory_available()
                # Disable FSDP for MPS as it's not compatible
@@ -560,9 +561,11 @@ 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)
-        config.pop("_name_or_path", None)
        class_name = config.pop("_class_name")
-        assert class_name is not None, "Model config does not contain a _class_name attribute. Only diffusers format is supported."
+        config.pop("_name_or_path", None)
+        assert class_name is not None, (
+            "Model config does not contain a _class_name attribute. Only diffusers format is supported."
+        )
        fastvideo_args.model_paths["vae"] = model_path

        from fastvideo.platforms import current_platform
@@ -730,9 +733,9 @@ class TransformerLoader(ComponentLoader):
    def load(self, model_path: str, fastvideo_args: FastVideoArgs):
        """Load the transformer based on the model path, and inference args."""
        config = get_diffusers_config(model=model_path)
-        config.pop("_name_or_path", None)
        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. "
@@ -747,7 +750,7 @@ class TransformerLoader(ComponentLoader):
        fastvideo_args.model_paths["transformer"] = model_path

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

        model_cls, _ = ModelRegistry.resolve_model_cls(cls_name)
@@ -809,11 +812,6 @@ class TransformerLoader(ComponentLoader):
            or cls_name == "Cosmos25Transformer3DModel"
            or getattr(fastvideo_args.pipeline_config, "prefix", "") == "Cosmos25"
        )
-
-        add_cls_branch = False
-        if hasattr(fastvideo_args, "_adding_cls_branch") and fastvideo_args._adding_cls_branch:
-            strict_load = False
-            add_cls_branch = True
        model = maybe_load_fsdp_model(
            model_cls=model_cls,
            init_params={"config": dit_config, "hf_config": hf_config},
@@ -833,11 +831,10 @@ class TransformerLoader(ComponentLoader):
            training_mode=fastvideo_args.training_mode,
            enable_torch_compile=fastvideo_args.enable_torch_compile,
            torch_compile_kwargs=fastvideo_args.torch_compile_kwargs,
-            add_cls_branch=add_cls_branch,
        )

        total_params = sum(p.numel() for p in model.parameters())
-        logger.info("Loaded model with %.2fB parameters, with cpu_offload: %s", total_params / 1e9, fastvideo_args.dit_cpu_offload)
+        logger.info("Loaded model with %.2fB parameters", total_params / 1e9)

        assert next(model.parameters()).dtype == default_dtype, (
            "Model dtype does not match default dtype"
@@ -875,11 +872,9 @@ class SchedulerLoader(ComponentLoader):

        scheduler_cls, _ = ModelRegistry.resolve_model_cls(class_name)

-        if fastvideo_args.pipeline_config.flow_shift is not None and "shift" in config:
-            config["shift"] = fastvideo_args.pipeline_config.flow_shift
        scheduler = scheduler_cls(**config)
-        # if fastvideo_args.pipeline_config.flow_shift is not None:
-        #     scheduler.set_shift(fastvideo_args.pipeline_config.flow_shift)
+        if fastvideo_args.pipeline_config.flow_shift is not None:
+            scheduler.set_shift(fastvideo_args.pipeline_config.flow_shift)
        if fastvideo_args.pipeline_config.timesteps_scale is not None:
            scheduler.set_timesteps_scale(
                fastvideo_args.pipeline_config.timesteps_scale
@@ -887,6 +882,50 @@ 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."""

@@ -75,7 +75,6 @@ def maybe_load_fsdp_model(
    pin_cpu_memory: bool = True,
    enable_torch_compile: bool = False,
    torch_compile_kwargs: dict[str, Any] | None = None,
-    add_cls_branch: bool = False,
 ) -> torch.nn.Module:
    """
    Load the model with FSDP if is training, else load the model without FSDP.
@@ -97,8 +96,6 @@ def maybe_load_fsdp_model(
    logger.info("Loading model with default_dtype: %s", default_dtype)
    with set_default_dtype(default_dtype), torch.device("meta"):
        model = model_cls(**init_params)
-    if add_cls_branch:
-        model.adding_cls_branch(atten_dim=1536, num_class=1, time_embed_dim=0)

    # Check if we should use FSDP
    use_fsdp = training_mode or fsdp_inference
@@ -343,7 +340,7 @@ def load_model_from_full_model_state_dict(
                       unused_keys)

    # List of allowed parameter name patterns
-    ALLOWED_NEW_PARAM_PATTERNS = ["gate_compress", "proj_l", "_cls_pred_branch", "_register_tokens", "_gan_ca_blocks"]  # Can be extended as needed
+    ALLOWED_NEW_PARAM_PATTERNS = ["gate_compress", "proj_l"]  # 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):
@@ -356,16 +353,14 @@ def load_model_from_full_model_state_dict(
        meta_sharded_param = meta_sd.get(new_param_name)
        if not hasattr(meta_sharded_param, "device_mesh"):
            # Initialize with zeros
-            # sharded_tensor = torch.zeros_like(meta_sharded_param,
-            #                                   device=device,
-            #                                   dtype=param_dtype)
-            sharded_tensor = meta_sharded_param.to(device=device, dtype=param_dtype)
+            sharded_tensor = torch.zeros_like(meta_sharded_param,
+                                              device=device,
+                                              dtype=param_dtype)
        else:
            # Initialize with zeros and distribute
-            # full_tensor = torch.zeros_like(meta_sharded_param,
-            #                                device=device,
-            #                                dtype=param_dtype)
-            full_tensor = meta_sharded_param.to(device=device, dtype=param_dtype)
+            full_tensor = torch.zeros_like(meta_sharded_param,
+                                           device=device,
+                                           dtype=param_dtype)
            sharded_tensor = distribute_tensor(
                full_tensor,
                meta_sharded_param.device_mesh,
@@ -78,10 +78,9 @@ def hf_to_custom_state_dict(
    for source_param_name, full_tensor in hf_param_sd:  # type: ignore
        target_param_name, merge_index, num_params_to_merge = param_names_mapping(
            source_param_name)
-        if merge_index is None:
-            reverse_param_names_mapping[target_param_name] = (source_param_name,
-                                                            merge_index,
-                                                            num_params_to_merge)
+        reverse_param_names_mapping[target_param_name] = (source_param_name,
+                                                          merge_index,
+                                                          num_params_to_merge)
        if merge_index is not None:
            to_merge_params[target_param_name][merge_index] = full_tensor
            if len(to_merge_params[target_param_name]) == num_params_to_merge:
@@ -1,6 +1,7 @@
 # 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
@@ -28,8 +29,6 @@ _TEXT_TO_VIDEO_DIT_MODELS = {
    ("dits", "hunyuanvideo15", "HunyuanVideo15Transformer3DModel"),
    "HYWorldTransformer3DModel":
    ("dits", "hyworld", "HYWorldTransformer3DModel"),
-    "CausalHunyuanVideo15Transformer3DModel":
-    ("dits", "causal_hunyuanvideo15", "CausalHunyuanVideo15Transformer3DModel"),
    "WanTransformer3DModel": ("dits", "wanvideo", "WanTransformer3DModel"),
    "CausalWanTransformer3DModel": ("dits", "causal_wanvideo", "CausalWanTransformer3DModel"),
    "StepVideoModel": ("dits", "stepvideo", "StepVideoModel"),
@@ -100,7 +99,12 @@ _SCHEDULERS = {
    ("schedulers", "scheduling_rcm", "RCMScheduler"),
 }

-_FAST_VIDEO_MODELS = {
+_UPSAMPLERS = {
+    "SRTo720pUpsampler": ("upsamplers", "hunyuan15", "SRTo720pUpsampler"),
+    "SRTo1080pUpsampler": ("upsamplers", "hunyuan15", "SRTo1080pUpsampler"),
+}
+
+_LEGACY_FAST_VIDEO_MODELS = {
    **_TEXT_TO_VIDEO_DIT_MODELS,
    **_IMAGE_TO_VIDEO_DIT_MODELS,
    **_TEXT_ENCODER_MODELS,
@@ -108,8 +112,102 @@ _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")
@@ -341,7 +439,8 @@ class _ModelRegistry:
 ModelRegistry = _ModelRegistry({
    model_arch:
    _LazyRegisteredModel(
-        module_name=f"fastvideo.models.{component_name}.{mod_relname}",
+        module_name=(f"fastvideo.models.{component_name}.{mod_relname}"
+                     if mod_relname else f"fastvideo.models.{component_name}"),
        component_name=component_name,
        class_name=cls_name,
    )
@@ -155,8 +155,8 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin,

        self.sigmas = sigmas.to(
            "cpu")  # to avoid too much CPU/GPU communication
-        self.sigma_min = sigma_min if sigma_min is not None else self.sigmas[-1].item()
-        self.sigma_max = sigma_max if sigma_max is not None else self.sigmas[0].item()
+        self.sigma_min = self.sigmas[-1].item()
+        self.sigma_max = self.sigmas[0].item()

        BaseScheduler.__init__(self)

@@ -289,7 +289,6 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin,
        sigmas: list[float] | None = None,
        mu: float | None = None,
        timesteps: list[float] | None = None,
-        extra_one_step: bool = False,
    ) -> None:
        """
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
@@ -351,10 +350,7 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin,
            if timesteps_array is None:
                t_max = self._sigma_to_t(self.sigma_max)
                t_min = self._sigma_to_t(self.sigma_min)
-                if extra_one_step:
-                    timesteps_array = np.linspace(t_max, t_min, num_inference_steps + 1)[:-1]
-                else:
-                    timesteps_array = np.linspace(t_max, t_min, num_inference_steps)
+                timesteps_array = np.linspace(t_max, t_min, num_inference_steps)
            sigmas_array = timesteps_array / self.config.num_train_timesteps
        else:
            sigmas_array = np.array(sigmas).astype(np.float32)
@@ -648,7 +644,7 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin,
        self,
        clean_latent: torch.Tensor,
        noise: torch.Tensor,
-        timestep: torch.Tensor,
+        timestep: torch.IntTensor,
    ) -> torch.Tensor:

        """
@@ -689,3 +685,6 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin,

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

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

    def __len__(self) -> int:
        return self.config.num_train_timesteps
+
+# Entry point for model registry
+EntryClass = RCMScheduler
@@ -69,9 +69,6 @@ class SelfForcingFlowMatchScheduler(BaseScheduler, ConfigMixin, SchedulerMixin):
            # use this scheduler for ODE trajectory
            timestep = timestep.unsqueeze(0)

-        model_output = model_output.permute(0, 2, 1, 3, 4)
-        sample = sample.permute(0, 2, 1, 3, 4)
-
        self.sigmas = self.sigmas.to(model_output.device)
        self.timesteps = self.timesteps.to(model_output.device)
        timestep = timestep.to(model_output.device)
@@ -84,8 +81,7 @@ class SelfForcingFlowMatchScheduler(BaseScheduler, ConfigMixin, SchedulerMixin):
                self.inverse_timesteps or self.reverse_sigmas) else 0
        else:
            sigma_ = self.sigmas[timestep_id + 1].reshape(-1, 1, 1, 1)
-        prev_sample = sample.flatten(0, 1) + model_output.flatten(0, 1) * (sigma_ - sigma)
-        prev_sample = prev_sample.unflatten(0, model_output.shape[:2]).permute(0, 2, 1, 3, 4)
+        prev_sample = sample + model_output * (sigma_ - sigma)
        if isinstance(prev_sample, torch.Tensor | float) and not return_dict:
            return (prev_sample, )
        return SelfForcingFlowMatchSchedulerOutput(prev_sample=prev_sample)
@@ -164,4 +160,6 @@ class SelfForcingFlowMatchScheduler(BaseScheduler, ConfigMixin, SchedulerMixin):

    def set_shift(self, shift: float) -> None:
        self.shift = shift
-    
+
+# Entry point for model registry
+EntryClass = SelfForcingFlowMatchScheduler
@@ -1094,4 +1094,7 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):
        return noisy_samples

    def __len__(self):
-        return self.config.num_train_timesteps
+        return self.config.num_train_timesteps
+
+# Entry point for model registry
+EntryClass = UniPCMultistepScheduler
@@ -0,0 +1,170 @@
+# Licensed under the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5/blob/main/LICENSE
+#
+# Unless and only to the extent required by applicable law, the Tencent Hunyuan works and any
+# output and results therefrom are provided "AS IS" without any express or implied warranties of
+# any kind including any warranties of title, merchantability, noninfringement, course of dealing,
+# usage of trade, or fitness for a particular purpose. You are solely responsible for determining the
+# appropriateness of using, reproducing, modifying, performing, displaying or distributing any of
+# the Tencent Hunyuan works or outputs and assume any and all risks associated with your or a
+# third party's use or distribution of any of the Tencent Hunyuan works or outputs and your exercise
+# of rights and permissions under this agreement.
+# See the License for the specific language governing permissions and limitations under the License.
+
+from collections.abc import Sequence
+from dataclasses import dataclass
+from enum import Enum
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from torch import Tensor
+
+from fastvideo.models.vaes.hunyuan15vae import (
+    HunyuanVideo15CausalConv3d,
+    HunyuanVideo15RMS_norm,
+)
+
+from fastvideo.layers.activation import get_act_fn
+from fastvideo.configs.models.upsamplers import SRTo720pUpsamplerConfig, SRTo1080pUpsamplerConfig
+
+class HunyuanVideo15ResnetBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int | None = None,
+        non_linearity: str = "swish",
+    ) -> None:
+        super().__init__()
+        out_channels = out_channels or in_channels
+
+        self.nonlinearity = get_act_fn(non_linearity)
+
+        self.norm1 = HunyuanVideo15RMS_norm(in_channels, images=False)
+        self.conv1 = HunyuanVideo15CausalConv3d(in_channels, out_channels, kernel_size=3)
+
+        self.norm2 = HunyuanVideo15RMS_norm(out_channels, images=False)
+        self.conv2 = HunyuanVideo15CausalConv3d(out_channels, out_channels, kernel_size=3)
+
+        self.nin_shortcut = None
+        if in_channels != out_channels:
+            self.nin_shortcut = nn.Conv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        residual = hidden_states
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv1(hidden_states)
+
+        hidden_states = self.norm2(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.nin_shortcut is not None:
+            residual = self.nin_shortcut(residual)
+
+        return hidden_states + residual
+
+class SRResidualCausalBlock3D(nn.Module):
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block = nn.Sequential(
+            HunyuanVideo15CausalConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            HunyuanVideo15CausalConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            HunyuanVideo15CausalConv3d(channels, channels, kernel_size=3),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.block(x)
+
+
+class SRTo720pUpsampler(nn.Module):
+
+    def __init__(
+        self,
+        config: SRTo720pUpsamplerConfig,
+    ):
+        super().__init__()
+        self.in_conv = HunyuanVideo15CausalConv3d(config.in_channels, config.hidden_channels, kernel_size=3)
+        self.blocks = nn.ModuleList([SRResidualCausalBlock3D(config.hidden_channels) for _ in range(config.num_blocks)])
+        self.out_conv = HunyuanVideo15CausalConv3d(config.hidden_channels, config.out_channels, kernel_size=3)
+        self.global_residual = bool(config.global_residual)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        residual = x
+        y = self.in_conv(x)
+        for blk in self.blocks:
+            y = blk(y)
+        y = self.out_conv(y)
+        if self.global_residual and (y.shape == residual.shape):
+            y = y + residual
+        return y
+
+
+class SRTo1080pUpsampler(nn.Module):
+
+    def __init__(
+        self,
+        config: SRTo1080pUpsamplerConfig,
+    ):
+        super().__init__()
+        self.num_res_blocks = config.num_res_blocks
+        self.block_out_channels = config.block_out_channels
+        self.z_channels = config.z_channels
+
+        block_in = config.block_out_channels[0]
+        self.conv_in = HunyuanVideo15CausalConv3d(config.z_channels, block_in, kernel_size=3)
+
+        self.up = nn.ModuleList()
+        for i_level, ch in enumerate(config.block_out_channels):
+            block = nn.ModuleList()
+            block_out = ch
+            for _ in range(self.num_res_blocks + 1):
+                block.append(HunyuanVideo15ResnetBlock(in_channels=block_in, out_channels=block_out))
+                block_in = block_out
+            up = nn.Module()
+            up.block = block
+
+            self.up.append(up)
+
+        self.norm_out = HunyuanVideo15RMS_norm(block_in, images=False)
+        self.conv_out = HunyuanVideo15CausalConv3d(block_in, config.out_channels, kernel_size=3)
+
+        self.gradient_checkpointing = False
+        self.is_residual = config.is_residual
+
+    def forward(self, z: Tensor, target_shape: Sequence[int] = None) -> Tensor:
+        """
+        Args:
+            z: (B, C, T, H, W)
+            target_shape: (H, W)
+        """
+        if target_shape is not None and z.shape[-2:] != target_shape:
+            bsz = z.shape[0]
+            z = rearrange(z, "b c f h w -> (b f) c h w")
+            z = F.interpolate(z, size=target_shape, mode="bilinear", align_corners=False)
+            z = rearrange(z, "(b f) c h w -> b c f h w", b=bsz)
+
+        # z to block_in
+        repeats = self.block_out_channels[0] // (self.z_channels)
+        h = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1)
+
+        # upsampling
+        for i_level in range(len(self.block_out_channels)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+            if hasattr(self.up[i_level], "upsample"):
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = get_act_fn("swish")(h)
+        h = self.conv_out(h)
+        return h
@@ -663,7 +663,7 @@ class AutoencoderKLHunyuanVideo15(nn.Module, ParallelTiledVAE):
        # When decoding spatially large video latents, the memory requirement is very high. By breaking the video latent
        # frames spatially into smaller tiles and performing multiple forward passes for decoding, and then blending the
        # intermediate tiles together, the memory requirement can be lowered.
-        self.use_tiling = True
+        self.use_tiling = False

        # The minimal tile height and width for spatial tiling to be used
        self.tile_sample_min_height = 256
@@ -700,4 +700,7 @@ class AutoencoderKLHunyuanVideo15(nn.Module, ParallelTiledVAE):
        else:
            z = posterior.mode()
        dec = self.decode(z)
-        return dec
+        return dec
+
+# Entry point for model registry
+EntryClass = AutoencoderKLHunyuanVideo15
@@ -850,3 +850,6 @@ class AutoencoderKLHunyuanVideo(nn.Module, ParallelTiledVAE):
            z = posterior.mode()
        dec = self.decode(z)
        return dec
+
+# Entry point for model registry
+EntryClass = AutoencoderKLHunyuanVideo
@@ -15,15 +15,977 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-from fastvideo.models.vaes.hunyuan15vae import AutoencoderKLHunyuanVideo15
-from fastvideo.configs.models.vaes import Hunyuan15VAEConfig
+from typing import List, Optional, Tuple, Union

-class AutoencoderKLHYWorld(AutoencoderKLHunyuanVideo15):
-    # TODO(mingjia): add temporal caching support for HYWorld VAE
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from einops import rearrange
+
+from fastvideo.layers.activation import get_act_fn
+from fastvideo.configs.models.vaes import Hunyuan15VAEConfig
+from fastvideo.models.vaes.common import ParallelTiledVAE
+from fastvideo.models.vaes.hunyuan15vae import (
+    HunyuanVideo15RMS_norm as HYWorldRMS_norm,
+    HunyuanVideo15AttnBlock as HYWorldAttnBlock,
+)
+
+# Cache size for temporal feature caching (number of frames to cache)
+CACHE_T = 2
+
+class HYWorldCausalConv3d(nn.Module):
+    """Causal Conv3d with optional cache support for temporal feature caching."""
+    
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, Tuple[int, int, int]] = 3,
+        stride: Union[int, Tuple[int, int, int]] = 1,
+        padding: Union[int, Tuple[int, int, int]] = 0,
+        dilation: Union[int, Tuple[int, int, int]] = 1,
+        bias: bool = True,
+        pad_mode: str = "replicate",
+    ) -> None:
+        super().__init__()
+
+        kernel_size = (kernel_size, kernel_size, kernel_size) if isinstance(kernel_size, int) else kernel_size
+
+        self.pad_mode = pad_mode
+        # Padding format: (W_left, W_right, H_left, H_right, T_left, T_right)
+        self.time_causal_padding = (
+            kernel_size[0] // 2,  # W_left (spatial)
+            kernel_size[0] // 2,  # W_right (spatial)
+            kernel_size[1] // 2,  # H_left (spatial)
+            kernel_size[1] // 2,  # H_right (spatial)
+            kernel_size[2] - 1,   # T_left (temporal causal padding)
+            0,                    # T_right (no future padding for causal)
+        )
+
+        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias=bias)
+
+    def forward(self, hidden_states: torch.Tensor, cache_x: Optional[torch.Tensor] = None) -> torch.Tensor:
+        """
+        Forward pass with optional temporal caching.
+        
+        Args:
+            hidden_states: Input tensor of shape (B, C, T, H, W)
+            cache_x: Optional cached frames from previous chunk, shape (B, C, CACHE_T, H, W)
+                    When provided, uses cached frames instead of padding for temporal dimension.
+        """
+        padding = list(self.time_causal_padding)
+        
+        if cache_x is not None and self.time_causal_padding[4] > 0:  # Has temporal padding and cache
+            cache_x = cache_x.to(hidden_states.device)
+            # Concatenate cached frames with current input on temporal dimension
+            hidden_states = torch.cat([cache_x, hidden_states], dim=2)
+            # Reduce temporal padding since we have cached frames
+            padding[4] -= cache_x.shape[2]
+        
+        hidden_states = F.pad(hidden_states, padding, mode=self.pad_mode)
+        return self.conv(hidden_states)
+
+
+class HYWorldUpsample(nn.Module):
+    """Hierarchical upsampling with temporal/spatial support and optional caching."""
+    
+    def __init__(self, in_channels: int, out_channels: int, add_temporal_upsample: bool = True):
+        super().__init__()
+        factor = 2 * 2 * 2 if add_temporal_upsample else 1 * 2 * 2
+        self.conv = HYWorldCausalConv3d(in_channels, out_channels * factor, kernel_size=3)
+        self.add_temporal_upsample = add_temporal_upsample
+        self.repeats = factor * out_channels // in_channels
+
+    def forward(
+        self, 
+        x: torch.Tensor,
+        feat_cache: Optional[List[Optional[torch.Tensor]]] = None,
+        feat_idx: Optional[List[int]] = None,
+        first_chunk: bool = False,
+    ):
+        """
+        Forward pass with optional temporal caching.
+        
+        Args:
+            x: Input tensor of shape (B, C, T, H, W)
+            feat_cache: List of cached features for each conv layer
+            feat_idx: List containing current cache index [idx]
+            first_chunk: Whether this is the first chunk (affects temporal upsample behavior)
+        """
+        r1 = 2 if self.add_temporal_upsample else 1
+        if feat_cache is not None and feat_idx is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x],
+                    dim=2,
+                )
+            h = self.conv(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            h = self.conv(x)
+
+        if self.add_temporal_upsample:
+            if first_chunk:
+                # First chunk: only spatial upsample
+                h = rearrange(h, "b (r2 r3 c) f h w -> b c f (h r2) (w r3)", r2=2, r3=2)
+                h = h[:, : h.shape[1] // 2]
+                # Compute the shortcut part
+                shortcut = rearrange(x, "b (r2 r3 c) f h w -> b c f (h r2) (w r3)", r2=2, r3=2)
+                shortcut = shortcut.repeat_interleave(repeats=self.repeats // 2, dim=1)
+            elif feat_cache is None and x.shape[2] > 1:
+                # No cache and multiple frames: separate first frame and rest
+                h_first = h[:, :, :1, :, :]
+                h_rest = h[:, :, 1:, :, :]
+                x_first = x[:, :, :1, :, :]
+                x_rest = x[:, :, 1:, :, :]
+
+                # First frame: only spatial upsample
+                h_first = rearrange(h_first, "b (r2 r3 c) f h w -> b c f (h r2) (w r3)", r2=2, r3=2)
+                h_first = h_first[:, : h_first.shape[1] // 2]
+                shortcut_first = rearrange(x_first, "b (r2 r3 c) f h w -> b c f (h r2) (w r3)", r2=2, r3=2)
+                shortcut_first = shortcut_first.repeat_interleave(repeats=self.repeats // 2, dim=1)
+                out_first = h_first + shortcut_first
+
+                # Remaining frames: spatio-temporal upsample
+                h_rest = rearrange(h_rest, "b (r1 r2 r3 c) f h w -> b c (f r1) (h r2) (w r3)", r1=r1, r2=2, r3=2)
+                shortcut_rest = rearrange(x_rest, "b (r1 r2 r3 c) f h w -> b c (f r1) (h r2) (w r3)", r1=r1, r2=2, r3=2)
+                shortcut_rest = shortcut_rest.repeat_interleave(repeats=self.repeats, dim=1)
+                out_rest = h_rest + shortcut_rest
+
+                return torch.cat([out_first, out_rest], dim=2)
+            else:
+                # Subsequent chunks with cache: spatio-temporal upsample
+                h = rearrange(h, "b (r1 r2 r3 c) f h w -> b c (f r1) (h r2) (w r3)", r1=r1, r2=2, r3=2)
+                shortcut = rearrange(x, "b (r1 r2 r3 c) f h w -> b c (f r1) (h r2) (w r3)", r1=r1, r2=2, r3=2)
+                shortcut = shortcut.repeat_interleave(repeats=self.repeats, dim=1)
+        else:
+            h = rearrange(h, "b (r1 r2 r3 c) f h w -> b c (f r1) (h r2) (w r3)", r1=r1, r2=2, r3=2)
+            shortcut = x.repeat_interleave(repeats=self.repeats, dim=1)
+            shortcut = rearrange(shortcut, "b (r1 r2 r3 c) f h w -> b c (f r1) (h r2) (w r3)", r1=r1, r2=2, r3=2)
+        
+        return h + shortcut
+
+
+class HYWorldDownsample(nn.Module):
+    """Hierarchical downsampling with temporal/spatial support and optional caching."""
+    
+    def __init__(self, in_channels: int, out_channels: int, add_temporal_downsample: bool = True):
+        super().__init__()
+        factor = 2 * 2 * 2 if add_temporal_downsample else 1 * 2 * 2
+        self.conv = HYWorldCausalConv3d(in_channels, out_channels // factor, kernel_size=3)
+
+        self.add_temporal_downsample = add_temporal_downsample
+        self.group_size = factor * in_channels // out_channels
+
+    def forward(
+        self, 
+        x: torch.Tensor,
+        feat_cache: Optional[List[Optional[torch.Tensor]]] = None,
+        feat_idx: Optional[List[int]] = None,
+    ):
+        """
+        Forward pass with optional temporal caching.
+        
+        Args:
+            x: Input tensor of shape (B, C, T, H, W)
+            feat_cache: List of cached features for each conv layer
+            feat_idx: List containing current cache index [idx]
+        """
+        r1 = 2 if self.add_temporal_downsample else 1
+        
+        # Apply conv with caching
+        if feat_cache is not None and feat_idx is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x],
+                    dim=2,
+                )
+            h = self.conv(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            h = self.conv(x) # Change the channel, ready for spatial or temporal downsample
+        
+        if self.add_temporal_downsample:
+            if x.shape[2] == 1:
+                # Single frame: only spatial downsample
+                h = rearrange(h, "b c f (h r2) (w r3) -> b (r2 r3 c) f h w", r2=2, r3=2)
+                h = torch.cat([h, h], dim=1)
+                # Compute the shortcut part
+                shortcut = rearrange(x, "b c f (h r2) (w r3) -> b (r2 r3 c) f h w", r2=2, r3=2)
+                B, C, T, H, W = shortcut.shape
+                shortcut = shortcut.view(B, h.shape[1], self.group_size // 2, T, H, W).mean(dim=2)
+            else:
+                # Multiple frames: full spatio-temporal downsample
+                h = rearrange(h, "b c (f r1) (h r2) (w r3) -> b (r1 r2 r3 c) f h w", r1=r1, r2=2, r3=2)
+                # Shortcut computation
+                shortcut = rearrange(x, "b c (f r1) (h r2) (w r3) -> b (r1 r2 r3 c) f h w", r1=r1, r2=2, r3=2)
+                B, C, T, H, W = shortcut.shape
+                shortcut = shortcut.view(B, h.shape[1], self.group_size, T, H, W).mean(dim=2)
+        else:
+            h = rearrange(h, "b c (f r1) (h r2) (w r3) -> b (r1 r2 r3 c) f h w", r1=r1, r2=2, r3=2)
+            shortcut = rearrange(x, "b c (f r1) (h r2) (w r3) -> b (r1 r2 r3 c) f h w", r1=r1, r2=2, r3=2)
+            B, C, T, H, W = shortcut.shape
+            shortcut = shortcut.view(B, h.shape[1], self.group_size, T, H, W).mean(dim=2)
+
+        return h + shortcut
+
+
+class HYWorldResnetBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        non_linearity: str = "swish",
+    ) -> None:
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+
+        self.nonlinearity = get_act_fn(non_linearity)
+
+        self.norm1 = HYWorldRMS_norm(in_channels, images=False)
+        self.conv1 = HYWorldCausalConv3d(in_channels, out_channels, kernel_size=3)
+
+        self.norm2 = HYWorldRMS_norm(out_channels, images=False)
+        self.conv2 = HYWorldCausalConv3d(out_channels, out_channels, kernel_size=3)
+        if in_channels != out_channels:
+            self.nin_shortcut = nn.Conv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+    
+    def forward(
+        self, 
+        hidden_states: torch.Tensor,
+        feat_cache: Optional[List[Optional[torch.Tensor]]] = None,
+        feat_idx: Optional[List[int]] = None,
+    ) -> torch.Tensor:
+        """
+        Forward pass with optional temporal feature caching.
+        
+        Args:
+            hidden_states: Input tensor of shape (B, C, T, H, W)
+            feat_cache: List of cached features for each conv layer
+            feat_idx: List containing current cache index [idx]
+        """
+        residual = hidden_states
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        
+        # apply the feature cacheing mechanism
+        if feat_cache is not None and feat_idx is not None:
+            # Retrieve the current layer index.
+            idx = feat_idx[0]
+
+            # Clone the last CACHE_T frames from the current input to store for the next step.
+            cache_x = hidden_states[:, :, -CACHE_T:, :, :].clone()
+            
+            # Handle boundary conditions: if the current temporal chunk is too short (< 2 frames)
+            # and we have a previous cache, prepend the last frame of the previous cache.
+            # This ensures sufficient temporal context for the convolution kernel.
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # actually this means chunk 1 after chunk 0
+                cache_x = torch.cat(
+                    [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x],
+                    dim=2,
+                )
+            
+            # Apply the convolution layer using `hidden_states`and the *previous* cached state.
+            hidden_states = self.conv1(hidden_states, feat_cache[idx])
+            
+            # Update the cache for this layer with the newly prepared context ('cache_x').
+            feat_cache[idx] = cache_x
+
+            # Increment the global layer index.
+            feat_idx[0] += 1
+        else:
+            hidden_states = self.conv1(hidden_states)
+
+        hidden_states = self.norm2(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        # Second conv with caching
+        if feat_cache is not None and feat_idx is not None:
+            idx = feat_idx[0]
+            cache_x = hidden_states[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x],
+                    dim=2,
+                )
+            hidden_states = self.conv2(hidden_states, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            hidden_states = self.conv2(hidden_states)
+
+        if self.in_channels != self.out_channels:
+            residual = self.nin_shortcut(residual)
+
+        return hidden_states + residual
+
+
+class HYWorldMidBlock(nn.Module):
+    """Mid block with attention and resnet blocks, with optional caching support."""
+    
+    def __init__(
+        self,
+        in_channels: int,
+        num_layers: int = 1,
+        add_attention: bool = True,
+    ) -> None:
+        super().__init__()
+        self.add_attention = add_attention
+
+        # There is always at least one resnet
+        resnets = [
+            HYWorldResnetBlock(
+                in_channels=in_channels,
+                out_channels=in_channels,
+            )
+        ]
+        attentions = []
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                attentions.append(HYWorldAttnBlock(in_channels))
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                HYWorldResnetBlock(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self, 
+        hidden_states: torch.Tensor,
+        feat_cache: Optional[List[Optional[torch.Tensor]]] = None,
+        feat_idx: Optional[List[int]] = None,
+    ) -> torch.Tensor:
+        """
+        Forward pass with optional temporal caching.
+        
+        Args:
+            hidden_states: Input tensor of shape (B, C, T, H, W)
+            feat_cache: List of cached features for each conv layer
+            feat_idx: List containing current cache index [idx]
+        """
+        hidden_states = self.resnets[0](hidden_states, feat_cache, feat_idx)
+
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                hidden_states = attn(hidden_states)
+            hidden_states = resnet(hidden_states, feat_cache, feat_idx)
+
+        return hidden_states
+
+
+class HYWorldDownBlock3D(nn.Module):
+    """Down block with resnet blocks and optional downsampling, with caching support."""
+    
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_layers: int = 1,
+        downsample_out_channels: Optional[int] = None,
+        add_temporal_downsample: int = True,
+    ) -> None:
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                HYWorldResnetBlock(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if downsample_out_channels is not None:
+            self.downsamplers = nn.ModuleList(
+                [
+                    HYWorldDownsample(
+                        out_channels,
+                        out_channels=downsample_out_channels,
+                        add_temporal_downsample=add_temporal_downsample,
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self, 
+        hidden_states: torch.Tensor,
+        feat_cache: Optional[List[Optional[torch.Tensor]]] = None,
+        feat_idx: Optional[List[int]] = None,
+    ) -> torch.Tensor:
+        """
+        Forward pass with optional temporal caching.
+        
+        Args:
+            hidden_states: Input tensor of shape (B, C, T, H, W)
+            feat_cache: List of cached features for each conv layer
+            feat_idx: List containing current cache index [idx]
+        """
+        for resnet in self.resnets:
+            hidden_states = resnet(hidden_states, feat_cache, feat_idx)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, feat_cache, feat_idx)
+
+        return hidden_states
+
+
+class HYWorldUpBlock3D(nn.Module):
+    """Up block with resnet blocks and optional upsampling, with caching support."""
+    
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_layers: int = 1,
+        upsample_out_channels: Optional[int] = None,
+        add_temporal_upsample: bool = True,
+    ) -> None:
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            input_channels = in_channels if i == 0 else out_channels
+
+            resnets.append(
+                HYWorldResnetBlock(
+                    in_channels=input_channels,
+                    out_channels=out_channels,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if upsample_out_channels is not None:
+            self.upsamplers = nn.ModuleList(
+                [
+                    HYWorldUpsample(
+                        out_channels,
+                        out_channels=upsample_out_channels,
+                        add_temporal_upsample=add_temporal_upsample,
+                    )
+                ]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self, 
+        hidden_states: torch.Tensor,
+        feat_cache: Optional[List[Optional[torch.Tensor]]] = None,
+        feat_idx: Optional[List[int]] = None,
+        first_chunk: bool = False,
+    ) -> torch.Tensor:
+        """
+        Forward pass with optional temporal caching.
+        
+        Args:
+            hidden_states: Input tensor of shape (B, C, T, H, W)
+            feat_cache: List of cached features for each conv layer
+            feat_idx: List containing current cache index [idx]
+            first_chunk: Whether this is the first chunk (for upsampling behavior)
+        """
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for resnet in self.resnets:
+                hidden_states = self._gradient_checkpointing_func(resnet, hidden_states)
+        else:
+            for resnet in self.resnets:
+                hidden_states = resnet(hidden_states, feat_cache, feat_idx)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, feat_cache, feat_idx, first_chunk)
+
+        return hidden_states
+
+
+class HYWorldEncoder3D(nn.Module):
+    r"""
+    3D vae encoder for HunyuanImageRefiner with optional temporal caching.
+    """

    def __init__(
        self,
-        config: Hunyuan15VAEConfig,
+        in_channels: int = 3,
+        out_channels: int = 64,
+        block_out_channels: Tuple[int, ...] = (128, 256, 512, 1024, 1024),
+        layers_per_block: int = 2,
+        temporal_compression_ratio: int = 4,
+        spatial_compression_ratio: int = 16,
+        downsample_match_channel: bool = True,
    ) -> None:
-        AutoencoderKLHunyuanVideo15.__init__(self, config)
+        super().__init__()

+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.group_size = block_out_channels[-1] // self.out_channels
+        self.block_out_channels = block_out_channels
+
+        self.conv_in = HYWorldCausalConv3d(in_channels, block_out_channels[0], kernel_size=3)
+        self.mid_block = None
+        self.down_blocks = nn.ModuleList([])
+
+        input_channel = block_out_channels[0]
+        for i in range(len(block_out_channels)):
+            add_spatial_downsample = i < np.log2(spatial_compression_ratio)
+            output_channel = block_out_channels[i]
+            if not add_spatial_downsample:
+                down_block = HYWorldDownBlock3D(
+                    num_layers=layers_per_block,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    downsample_out_channels=None,
+                    add_temporal_downsample=False,
+                )
+                input_channel = output_channel
+            else:
+                add_temporal_downsample = i >= np.log2(spatial_compression_ratio // temporal_compression_ratio)
+                downsample_out_channels = block_out_channels[i + 1] if downsample_match_channel else output_channel
+                down_block = HYWorldDownBlock3D(
+                    num_layers=layers_per_block,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    downsample_out_channels=downsample_out_channels,
+                    add_temporal_downsample=add_temporal_downsample,
+                )
+                input_channel = downsample_out_channels
+
+            self.down_blocks.append(down_block)
+
+        self.mid_block = HYWorldMidBlock(in_channels=block_out_channels[-1])
+
+        self.norm_out = HYWorldRMS_norm(block_out_channels[-1], images=False)
+        self.conv_act = nn.SiLU()
+        self.conv_out = HYWorldCausalConv3d(block_out_channels[-1], out_channels, kernel_size=3)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self, 
+        hidden_states: torch.Tensor,
+        feat_cache: Optional[List[Optional[torch.Tensor]]] = None,
+        feat_idx: Optional[List[int]] = None,
+    ) -> torch.Tensor:
+        """
+        Forward pass with optional temporal caching.
+        
+        Args:
+            hidden_states: Input tensor of shape (B, C, T, H, W)
+            feat_cache: List of cached features for each conv layer
+            feat_idx: List containing current cache index [idx]
+        """
+        # conv_in with caching
+        if feat_cache is not None and feat_idx is not None:
+            idx = feat_idx[0]
+            cache_x = hidden_states[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x],
+                    dim=2,
+                )
+            hidden_states = self.conv_in(hidden_states, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            hidden_states = self.conv_in(hidden_states)
+
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for down_block in self.down_blocks:
+                hidden_states = self._gradient_checkpointing_func(down_block, hidden_states)
+            hidden_states = self._gradient_checkpointing_func(self.mid_block, hidden_states)
+        else:
+            for down_block in self.down_blocks:
+                hidden_states = down_block(hidden_states, feat_cache, feat_idx)
+            hidden_states = self.mid_block(hidden_states, feat_cache, feat_idx)
+
+        batch_size, _, frame, height, width = hidden_states.shape
+        short_cut = hidden_states.view(batch_size, -1, self.group_size, frame, height, width).mean(dim=2)
+
+        hidden_states = self.norm_out(hidden_states)
+        hidden_states = self.conv_act(hidden_states)
+
+        # conv_out with caching
+        if feat_cache is not None and feat_idx is not None:
+            idx = feat_idx[0]
+            cache_x = hidden_states[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x],
+                    dim=2,
+                )
+            hidden_states = self.conv_out(hidden_states, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            hidden_states = self.conv_out(hidden_states)
+
+        hidden_states += short_cut
+
+        return hidden_states
+
+
+class HYWorldDecoder3D(nn.Module):
+    r"""
+    Causal decoder for 3D video-like data used for HunyuanImage-1.5 Refiner with optional temporal caching.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 32,
+        out_channels: int = 3,
+        block_out_channels: Tuple[int, ...] = (1024, 1024, 512, 256, 128),
+        layers_per_block: int = 2,
+        spatial_compression_ratio: int = 16,
+        temporal_compression_ratio: int = 4,
+        upsample_match_channel: bool = True,
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.block_out_channels = block_out_channels
+        self.repeat = block_out_channels[0] // self.in_channels
+
+        self.conv_in = HYWorldCausalConv3d(self.in_channels, block_out_channels[0], kernel_size=3)
+        self.up_blocks = nn.ModuleList([])
+
+        # mid
+        self.mid_block = HYWorldMidBlock(in_channels=block_out_channels[0])
+
+        # up
+        input_channel = block_out_channels[0]
+        for i in range(len(block_out_channels)):
+            output_channel = block_out_channels[i]
+
+            add_spatial_upsample = i < np.log2(spatial_compression_ratio)
+            add_temporal_upsample = i < np.log2(temporal_compression_ratio)
+            if add_spatial_upsample or add_temporal_upsample:
+                upsample_out_channels = block_out_channels[i + 1] if upsample_match_channel else output_channel
+                up_block = HYWorldUpBlock3D(
+                    num_layers=self.layers_per_block + 1,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    upsample_out_channels=upsample_out_channels,
+                    add_temporal_upsample=add_temporal_upsample,
+                )
+                input_channel = upsample_out_channels
+            else:
+                up_block = HYWorldUpBlock3D(
+                    num_layers=self.layers_per_block + 1,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    upsample_out_channels=None,
+                    add_temporal_upsample=False,
+                )
+                input_channel = output_channel
+
+            self.up_blocks.append(up_block)
+
+        # out
+        self.norm_out = HYWorldRMS_norm(block_out_channels[-1], images=False)
+        self.conv_act = nn.SiLU()
+        self.conv_out = HYWorldCausalConv3d(block_out_channels[-1], out_channels, kernel_size=3)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self, 
+        hidden_states: torch.Tensor,
+        feat_cache: Optional[List[Optional[torch.Tensor]]] = None,
+        feat_idx: Optional[List[int]] = None,
+        first_chunk: bool = False,
+    ) -> torch.Tensor:
+        """
+        Forward pass with optional temporal caching.
+        
+        Args:
+            hidden_states: Input tensor of shape (B, C, T, H, W)
+            feat_cache: List of cached features for each conv layer
+            feat_idx: List containing current cache index [idx]
+            first_chunk: Whether this is the first chunk (for upsampling behavior)
+        """
+        # conv_in with caching
+        if feat_cache is not None and feat_idx is not None:
+            idx = feat_idx[0]
+            cache_x = hidden_states[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x],
+                    dim=2,
+                )
+            hidden_states = self.conv_in(hidden_states, feat_cache[idx]) + hidden_states.repeat_interleave(
+                repeats=self.repeat, dim=1
+            )
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            hidden_states = self.conv_in(hidden_states) + hidden_states.repeat_interleave(repeats=self.repeat, dim=1)
+
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            hidden_states = self._gradient_checkpointing_func(self.mid_block, hidden_states)
+            for up_block in self.up_blocks:
+                hidden_states = self._gradient_checkpointing_func(up_block, hidden_states)
+        else:
+            hidden_states = self.mid_block(hidden_states, feat_cache, feat_idx)
+            for up_block in self.up_blocks:
+                hidden_states = up_block(hidden_states, feat_cache, feat_idx, first_chunk)
+
+        # post-process
+        hidden_states = self.norm_out(hidden_states)
+        hidden_states = self.conv_act(hidden_states)
+
+        # conv_out with caching
+        if feat_cache is not None and feat_idx is not None:
+            idx = feat_idx[0]
+            cache_x = hidden_states[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x],
+                    dim=2,
+                )
+            hidden_states = self.conv_out(hidden_states, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            hidden_states = self.conv_out(hidden_states)
+        
+        return hidden_states
+
+
+class AutoencoderKLHYWorld(nn.Module, ParallelTiledVAE):
+    r"""
+    A VAE model with KL loss for encoding videos into latents and decoding latent representations into videos.
+    Revised from HunyuanVideo-1.5 with temporal caching support for HY-WorldPlay integration.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        config: Hunyuan15VAEConfig, # HYWorld use the same VAE architecture as HunyuanVideo-1.5
+    ) -> None:
+        nn.Module.__init__(self)
+        ParallelTiledVAE.__init__(self, config)
+
+        self.encoder: Optional[HYWorldEncoder3D] = None
+        self.decoder: Optional[HYWorldDecoder3D] = None
+
+        if config.load_encoder:
+            self.encoder = HYWorldEncoder3D(
+                in_channels=config.in_channels,
+                out_channels=config.latent_channels * 2,
+                block_out_channels=config.block_out_channels,
+                layers_per_block=config.layers_per_block,
+                temporal_compression_ratio=config.temporal_compression_ratio,
+                spatial_compression_ratio=config.spatial_compression_ratio,
+                downsample_match_channel=config.downsample_match_channel,
+            )
+
+        if config.load_decoder:
+            self.decoder = HYWorldDecoder3D(
+                in_channels=config.latent_channels,
+                out_channels=config.out_channels,
+                block_out_channels=list(reversed(config.block_out_channels)),
+                layers_per_block=config.layers_per_block,
+                temporal_compression_ratio=config.temporal_compression_ratio,
+                spatial_compression_ratio=config.spatial_compression_ratio,
+                upsample_match_channel=config.upsample_match_channel,
+            )
+
+        # TODO: Add spatial tiling.
+        self.use_tiling = False
+
+        # The minimal tile height and width for spatial tiling to be used
+        self.tile_sample_min_height = 256
+        self.tile_sample_min_width = 256
+        self.tile_sample_min_num_frames = 2000 # Fill in a random large number, as hy1.5 vae does not use temporal tiling
+
+        # Cache-related attributes (initialized in clear_cache)
+        self._conv_num: int = 0
+        self._conv_idx: List[int] = [0]
+        self._feat_map: List[Optional[torch.Tensor]] = []
+        self._enc_conv_num: int = 0
+        self._enc_conv_idx: List[int] = [0]
+        self._enc_feat_map: List[Optional[torch.Tensor]] = []
+
+        # Precompute and cache conv counts for encoder and decoder for clear_cache speedup
+        self._cached_conv_counts = {
+            "decoder": (
+                sum(1 for m in self.decoder.modules() if isinstance(m, HYWorldCausalConv3d))
+                if self.decoder is not None
+                else 0
+            ),
+            "encoder": (
+                sum(1 for m in self.encoder.modules() if isinstance(m, HYWorldCausalConv3d))
+                if self.encoder is not None
+                else 0
+            ),
+        }
+
+    def clear_cache(self) -> None:
+        """
+        Initialize/clear the feature cache for chunk-based encoding/decoding.
+        
+        This should be called before starting a new encode/decode sequence to ensure
+        the cache is properly initialized.
+        """
+        # Cache for decoder
+        self._conv_num = self._cached_conv_counts["decoder"]
+        self._conv_idx = [0]
+        self._feat_map: List[Optional[torch.Tensor]] = [None] * self._conv_num
+        
+        # Cache for encoder
+        self._enc_conv_num = self._cached_conv_counts["encoder"]
+        self._enc_conv_idx = [0]
+        self._enc_feat_map: List[Optional[torch.Tensor]] = [None] * self._enc_conv_num
+
+    def _encode(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Encode video with temporal caching for chunk-based processing.
+        
+        This processes video in chunks (first frame separately, then 4 frames at a time)
+        while maintaining temporal context through caching. This matches the HY-WorldPlay
+        behavior for memory-efficient long video encoding.
+        
+        Args:
+            x: Input video tensor of shape (B, C, T, H, W)
+            
+        Returns:
+            Encoded latent tensor
+        """
+        assert self.encoder is not None, "Encoder not loaded"
+        _, _, num_frame, _, _ = x.shape
+
+        self.clear_cache()
+        
+        # Process in chunks: first frame alone, then groups of 4 frames
+        iter_ = 1 + (num_frame - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                # First frame
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+            else:
+                # Subsequent frames in groups of 4
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1) : 1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+                out = torch.cat([out, out_], dim=2)
+
+        self.clear_cache()
+        return out
+
+    def _decode(self, z: torch.Tensor) -> torch.Tensor:
+        """
+        Decode latents with temporal caching for chunk-based processing.
+        
+        This processes latents one frame at a time while maintaining temporal context
+        through caching. This matches the HY-WorldPlay behavior for memory-efficient
+        long video decoding.
+        
+        Args:
+            z: Latent tensor of shape (B, C, T, H, W)
+            
+        Returns:
+            Decoded video tensor
+        """
+        assert self.decoder is not None, "Decoder not loaded"
+        _, _, num_frame, _, _ = z.shape
+
+        self.clear_cache()
+        
+        # Process one frame at a time with caching
+        for i in range(num_frame):
+            self._conv_idx = [0]
+            if i == 0:
+                # First frame with first_chunk=True
+                out = self.decoder(
+                    z[:, :, i : i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
+                    first_chunk=True,
+                )
+            else:
+                # Subsequent frames
+                out_ = self.decoder(
+                    z[:, :, i : i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
+                    first_chunk=False,
+                )
+                out = torch.cat([out, out_], dim=2)
+
+        self.clear_cache()
+        return out
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> torch.Tensor:
+        r"""
+        Args:
+            sample (`torch.Tensor`): Input sample.
+            sample_posterior (`bool`, *optional*, defaults to `False`):
+                Whether to sample from the posterior.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+            generator (`torch.Generator`, *optional*):
+                Generator for sampling.
+        """
+        x = sample
+
+        # encode() uses temporal caching by default (via _encode)
+        posterior = self.encode(x)
+
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+
+        # decode() uses temporal caching by default (via _decode)
+        dec = self.decode(z)
+
+        return dec
+
+# Entry point for model registry
+EntryClass = AutoencoderKLHYWorld
@@ -1802,3 +1802,6 @@ class LTX2CausalVideoAutoencoder(nn.Module):
        if previous_chunk is not None:
            previous_weights = previous_weights.clamp(min=1e-8)
            yield previous_chunk / previous_weights
+
+# Entry point for model registry
+EntryClass = LTX2CausalVideoAutoencoder
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
JerryZhou54	8cb2ae9d27	ckpt	2026-02-09 05:50:48 +00:00
XOR-op	f76efe798e	[bugfix]: `_compile_conditions` regression (#1077 )	2026-02-06 18:57:43 -08:00
Zhang Peiyuan	09f455233e	[misc] readme small fix (#1076 )	2026-02-06 16:57:48 -08:00
XOR-op	aea300f690	[perf]: use CUDA IPC in multiproc executor to avoid serialization overhead (#1061 )	2026-02-06 19:43:18 -05:00
Hao Zhang	b92219f6a6	more fix and relocate STA arguments to pipeline config (#1073 )	2026-02-06 13:38:35 -08:00
Jinzhe Pan	a321b95a8a	[Fix] remove video ratio limitation (#1069 )	2026-02-05 20:33:08 -08:00
Wei Zhou	98308db7e0	[Feature] [Hy1.5] Support HY1.5 super-resolution pipeline for 1080p videos (#1046 )	2026-02-05 16:39:20 -08:00
Hao Zhang	c1e18f6722	Some minor fixes (#1068 )	2026-02-05 16:28:00 -08:00
William Lin	75e193a2c9	[core] Refactor and centralize our registry for models, pipelines, and sampling params (#1066 )	2026-02-05 14:40:30 -08:00
XOR-op	d6e0a7d0dd	[refactor] Action module (#1065 )	2026-02-05 13:56:35 -08:00
William Linandgemini-code-assist[bot]	7fc5f241da	[misc] Fix naming instruction in runpod.md (#1067 ) Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>	2026-02-05 13:45:49 -08:00
Mingjia Huo	aae48a7e90	[feat] HYworld VAE with cache (#1057 )	2026-02-05 04:18:17 -08:00
William Lin	88f38eb0f4	[misc] upgrade torch to 2.10 (#1048 )	2026-02-05 04:15:22 -08:00
Shao Duan	d750b463dc	Added Sequence Parallelism for LTX-2 Distilled (#1036 )	2026-02-04 15:25:30 -08:00
KyleShaoandWill Lin	e10b26a3d8	[feat] Add Cosmos 2.5 I2W/V2W support (staged pipeline + examples) (#1021 ) Co-authored-by: Will Lin <wlsaidhi@gmail.com>	2026-02-04 14:19:30 -08:00
XOR-op	74636ba246	[chore]: use higher precision timestamp in logging (#1062 )	2026-02-04 13:22:59 -08:00
Wei Zhou	7e2f3f14e7	[Bugfix] [Wan I2V] Fix CLIP Image encoder config (#1063 )	2026-02-04 13:20:19 -08:00
Kaiqin Kong	caa1c402ba	[bugfix] Double Normalization in Preprocessing Dataset (#1055 )	2026-01-31 11:11:59 -08:00
XOR-op	38a6bd93d3	[chore]: update sageattn3 installation instructions (#1050 )	2026-01-29 15:45:52 -08:00
				`@@ -1 +0,0 @@`
				`The raw mixkit dataset is downloaded and preprocessed following instructions from https://github.com/tianweiy/CausVid/tree/master`