Practical overview
When it fits
Use Chipmunk when integrating inference acceleration into a diffusion-transformer workflow and the target model and GPU environment fit the repository's requirements.
Who it is for
- Engineers optimizing diffusion-transformer inference pipelines.
- Researchers exploring dynamic sparsity and sparse-kernel optimization for image or video generation.
Documented uses
- Reduce attention computation in diffusion-transformer image generation.
- Reduce MLP computation during diffusion-transformer inference.
- Apply sparse-kernel techniques to supported video-generation workflows.
- Adapt the method to another diffusion-transformer codebase using the repository tutorial.
Requirements & limits
The repository is public and MIT-licensed.
- The repository's kernel guidance specifies Hopper GPUs and CUDA Toolkit 12.4 or later.
- The quickstart uses Python 3.11 and recommends CUDA 12.8.
- The first few generations can be slower because of PyTorch compiler cold starts; the README says speedups begin from generation three.
- Reported speedups depend on the specific model, GPU and generation configuration.
- The HunyuanVideo example may spend up to about five minutes initializing pinned CPU memory buffers for just-in-time offloading.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the Chipmunk environment and build its kernels
Install Chipmunk and build its kernels using the repository's documented quickstart.
Before you start
- A Linux environment, as identified in the project platform metadata.
- Python 3.11 and Conda.
- CUDA Toolkit 12.4 or later; the quickstart installs CUDA 12.8.
- A Hopper GPU for the documented kernels.
Steps
Clone the repository with its submodules: the command below.
git clone https://github.com/sandyresearch/chipmunk --recurse-submodules --shallow-submodules --depth 1Create and activate a Conda environment with Python 3.11:
conda create -n chipmunk python=3.11 -yandconda activate chipmunk.Install CUDA 12.8 with
conda install cuda==12.8.0 -c nvidia -y.Install dependencies and build the kernels with the command below.
pip install -e . --no-build-isolation
Expected outcome
The documented environment and project installation are set up for the subsequent model-specific examples.
Guide limitations
- The kernel guidance specifies Hopper GPUs and CUDA Toolkit 12.4 or later.
- This setup recipe does not cover running a generation; use a model-specific example for that next stage.
Run the documented FLUX.1-dev image-generation example
Use the repository's FLUX.1-dev example command with a prompt and Chipmunk configuration.
Before you start
- The Chipmunk repository and its documented environment setup.
- A FLUX.1-dev model setup and the example's required dependencies.
- A CUDA Toolkit version and GPU meeting the repository's kernel requirements.
Steps
Follow the repository quickstart to clone Chipmunk and install its dependencies and kernels.
Run the documented example command from the repository:
cd examples/flux && pip install -e . && python -m flux.cli --name flux-dev --loop --prompt "A very cute cartoon chipmunk dressed up as a ninja holding katanas" --chipmunk-config ./chipmunk-config.yml.
Expected outcome
The documented FLUX.1-dev example is invoked with Chipmunk configuration and a text prompt.
Guide limitations
- The README warns that the first few image or video generations may be slower because of PyTorch compiler cold-start overhead; it says speedups should begin at generation three.
- The example command's prompt is illustrative; performance depends on the model and hardware configuration.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- sandyresearch/chipmunk · Official homepage title · retained in its source language