Mochi-Diffusers-xDiT

This project is based on the Mochi algorithm of the diffuser and optimized and accelerated for multi GPUs inference using the Xufser framework.

Installation

To set up the environment and install the required dependencies, follow these steps:

1. Create and activate the Conda environment:

conda create -n mochi-xfuser python=3.10
conda activate mochi-xfuser

2. Install PyTorch and related libraries:

pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118

3. Install the xDiT package:

cd xDiT
pip install -e .

[Optional] You can install flash-att==2.6.3 by this link:

flash-attn==2.6.3

pip install flash_attn-2.6.3+cu118torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

4. Download the required files:

You can download the necessary files from the following link:

Download diffusers-mochi.zip

Then extract and navigate into the diffusers-mochi folder:

cd diffusers-mochi
pip install -e .

5. Download the model parameters:

You can download the model parameters from this link:

feizhengcong/mochi-1-preview-diffusers

Alternatively, you can use the following command to download them:

pip install -U huggingface_hub
huggingface-cli download --resume-download feizhengcong/mochi-1-preview-diffusers --local-dir feizhengcong/mochi-1-preview-diffusers --local-dir-use-symlinks False

Usage

Once the setup is complete, you can run the example with the following command:

torchrun --nproc_per_node=4 examples/mochi_example.py --model 'feizhengcong/mochi-1-preview-diffusers' --ulysses_degree 2 --ring_degree 1 --use_cfg_parallel --height 480 --width 848 --num_frames 163 --prompt "your prompt" --num_inference_steps 50 --seed 42

Related Links

• xDiT
• Mochi
• Mochi's Diffusers Integration Contributors:
  • Zhengcong Fei