| Documentation | Community | Contribution | Discord |
onediff is an out-of-the-box acceleration library for diffusion models, it provides:
- Out-of-the-box acceleration for popular UIs/libs(such as HF diffusers and ComfyUI)
- PyTorch code compilation tools and strong optimized GPU Kernels for diffusion models
- [2024/07/23] 🚀 Up to 1.7x Speedup for Kolors: Kolors Acceleration Report
- [2024/06/18] 🚀 Acceleration for DiT models: SD3 Acceleration Report, PixArt Acceleration Report, and Latte Acceleration Report
- [2024/04/13] 🚀 OneDiff 1.0 is released (Acceleration of SD & SVD with one line of code)
- [2024/01/12] 🚀 Accelerating Stable Video Diffusion 3x faster with OneDiff DeepCache + Int8
- [2023/12/19] 🚀 Accelerating SDXL 3x faster with DeepCache and OneDiff
We're hiring! If you are interested in working on onediff at SiliconFlow, we have roles open for Interns and Engineers in Beijing (near Tsinghua University).
If you have contributed significantly to open-source software and are interested in remote work, you can contact us at [email protected] with onediff in the email title.
onediff is the abbreviation of "one line of code to accelerate diffusion models".
- Model stabilityai/stable-diffusion-xl-base-1.0;
- Image size 1024*1024, batch size 1, steps 30;
- NVIDIA A100 80G SXM4;
- Model stabilityai/stable-video-diffusion-img2vid-xt;
- Image size 576*1024, batch size 1, steps 25, decoder chunk size 5;
- NVIDIA A100 80G SXM4;
Note that we haven't got a way to run SVD with TensorRT on Feb 29 2024.
We also maintain a repository for benchmarking the quality of generation after acceleration: odeval
- Create an issue
- Chat in Discord:
- Community and Feedback
- Linux
- If you want to use onediff on Windows, please use it under WSL.
- The guide to install onediff in WSL2.
- NVIDIA GPUs
Note: You can choose the latest versions you want for diffusers or transformers.
python3 -m pip install "torch" "transformers==4.27.1" "diffusers[torch]==0.19.3"
When considering the choice between OneFlow and Nexfort, either one is optional, and only one is needed.
-
For DiT structural models or H100 devices, it is recommended to use Nexfort.
-
For all other cases, it is recommended to use OneFlow. Note that optimizations within OneFlow will gradually transition to Nexfort in the future.
Install Nexfort is Optional. The detailed introduction of Nexfort is here.
python3 -m pip install -U torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 torchao==0.1
python3 -m pip install -U nexfortInstall OneFlow is Optional.
NOTE: We have updated OneFlow frequently for onediff, so please install OneFlow by the links below.
-
CUDA 11.8
For NA/EU users
python3 -m pip install -U --pre oneflow -f https://github.com/siliconflow/oneflow_releases/releases/expanded_assets/community_cu118
For CN users
python3 -m pip install -U --pre oneflow -f https://oneflow-pro.oss-cn-beijing.aliyuncs.com/branch/community/cu118
Click to get OneFlow packages for other CUDA versions.
-
CUDA 12.1
For NA/EU users
python3 -m pip install -U --pre oneflow -f https://github.com/siliconflow/oneflow_releases/releases/expanded_assets/community_cu121
For CN users
python3 -m pip install -U --pre oneflow -f https://oneflow-pro.oss-cn-beijing.aliyuncs.com/branch/community/cu121
-
CUDA 12.2
For NA/EU users
python3 -m pip install -U --pre oneflow -f https://github.com/siliconflow/oneflow_releases/releases/expanded_assets/community_cu122
For CN users
python3 -m pip install -U --pre oneflow -f https://oneflow-pro.oss-cn-beijing.aliyuncs.com/branch/community/cu122
- From PyPI
python3 -m pip install --pre onediff
- From source
git clone https://github.com/siliconflow/onediff.git
cd onediff && python3 -m pip install -e .
Or install for development:
# install for dev
cd onediff && python3 -m pip install -e '.[dev]'
# code formatting and linting
pip3 install pre-commit
pre-commit install
pre-commit run --all-files
NOTE: If you intend to utilize plugins for ComfyUI/StableDiffusion-WebUI, we highly recommend installing OneDiff from the source rather than PyPI. This is necessary as you'll need to manually copy (or create a soft link) for the relevant code into the extension folder of these UIs/Libs.
| Functionality | Details |
|---|---|
| Compiling Time | About 1 minute (SDXL) |
| Deployment Methods | Plug and Play |
| Dynamic Image Size Support | Support with no overhead |
| Model Support | SD1.5~2.1, SDXL, SDXL Turbo, etc. |
| Algorithm Support | SD standard workflow, LoRA, ControlNet, SVD, InstantID, SDXL Lightning, etc. |
| SD Framework Support | ComfyUI, Diffusers, SD-webui |
| Save & Load Accelerated Models | Yes |
| Time of LoRA Switching | Hundreds of milliseconds |
| LoRA Occupancy | Tens of MB to hundreds of MB. |
| Device Support | NVIDIA GPU 3090 RTX/4090 RTX/A100/A800/A10 etc. (Compatibility with Ascend in progress) |
onediff supports the acceleration for SOTA models.
- stable: release for public usage, and has long-term support;
- beta: release for professional usage, and has long-term support;
- alpha: early release for expert usage, and should be careful to use;
| AIGC Type | Models | HF diffusers | ComfyUI | SD web UI | |||
|---|---|---|---|---|---|---|---|
| Community | Enterprise | Community | Enterprise | Community | Enterprise | ||
| Image | SD 1.5 | stable | stable | stable | stable | stable | stable |
| SD 2.1 | stable | stable | stable | stable | stable | stable | |
| SDXL | stable | stable | stable | stable | stable | stable | |
| LoRA | stable | stable | stable | ||||
| ControlNet | stable | stable | |||||
| SDXL Turbo | stable | stable | |||||
| LCM | stable | stable | |||||
| SDXL DeepCache | alpha | alpha | alpha | alpha | |||
| InstantID | beta | beta | |||||
| Video | SVD(stable Video Diffusion) | stable | stable | stable | stable | ||
| SVD DeepCache | alpha | alpha | alpha | alpha |
Compile and save the compiled result offline, then load it online for serving
- Save and Load the compiled graph
- Compile at one device(such as device 0), then use the compiled result to other device(such as device 1~7). Change device of the compiled graph to do multi-process serving
If you want to do distributed inference, you can use onediff's compiler to do single-device acceleration in a distributed inference engine such as xDiT
If you need Enterprise-level Support for your system or business, you can email us at [email protected], or contact us through the website: https://siliconflow.cn/pricing
| Onediff Enterprise Solution | |
|---|---|
| More extreme compiler optimization for diffusion process | Usually another 20%~30% or more performance gain |
| End-to-end workflow speedup solutions | Sometimes 200%~300% performance gain |
| End-to-end workflow deployment solutions | Workflow to online model API |
| Technical support for deployment | High priority support |
@misc{2022onediff,
author={OneDiff Contributors},
title = {OneDiff: An out-of-the-box acceleration library for diffusion models},
year = {2022},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/siliconflow/onediff}}
}