vipshopvipshopOfficial·7 Agent Skills Included

cache-dit

Speed up diffusion model inference with caching, parallelism, and quantization

Accelerates diffusion transformer inference for image, video, and audio models using hybrid caching, context and tensor parallelism, quantization, and CPU offload. Eliminates slow generation times and manual GPU tuning across 120+ Diffusers model variants. Includes agent skills for writing CUDA, Triton, and CUTLASS kernels, migrating operators, and integrating new DiT models with validated testing workflows.
npx skills add vipshop/cache-dit --all -g -y

All Skills in This Repository (7)

Pure Emerald Level Indicators

Frequently Asked Questions

FAQPage Schema
How to install cache-dit?

Run `npx skills add vipshop/cache-dit --all -g -y` in your terminal to install all skills in this suite globally.

How to speed up diffusion model inference?

Cache-DiT applies hybrid caching, context parallelism, and quantization on top of Diffusers pipelines, delivering up to 9x speedup without changing your model code.

Which models does cache-dit support?

It supports over 40 DiT pipeline families and 120+ variants from HuggingFace Diffusers, including FLUX, Qwen-Image, CogVideoX, Wan, and HunyuanVideo.

Can AI agents write GPU kernels with these skills?

Yes. The bundled skills guide agents through writing and optimizing CUDA, Triton, CuTe DSL, and CUTLASS kernels with architecture-specific profiling and validation steps.

Does cache-dit work with vLLM and ComfyUI?

Yes. It integrates natively with SGLang Diffusion, vLLM-Omni, TensorRT-LLM, and ComfyUI, and runs on NVIDIA, AMD, and Ascend hardware.

Related Repositories in Software Engineering

View All in Software Engineering