# vipshop/cache-dit

A PyTorch-native inference engine with cache, parallelism, quantization and cpu offload for DiTs.

Repository: https://github.com/vipshop/cache-dit
Canonical: https://ross.abutalabs.com/products/cache-dit
Homepage: https://cache-dit.readthedocs.io
Language: Python
License: Apache-2.0
License Family: permissive
Topics: flux2-klein, svdquant, parallelism
Last push: 2026-09-02T06:58:35+00:00

## Health v2 (maintenance only)
Score: 82/100 (v2, computed 2026-09-03T02:39:23.370411+00:00)
- activity 100, release rhythm 88, longevity 31
- inputs: {"age_days": 447, "days_push": 0, "days_rel": 1, "gap_med": 1.0, "n_releases_24m": 97}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1267, forks 80 (observed 2026-09-03T02:15:20.342670+00:00)

## What it is
Cache-DiT is a PyTorch-native inference engine that accelerates Diffusion Transformer (DiT) models with hybrid caching, parallelism, quantization, and CPU offload. It is built on Hugging Face Diffusers and supports 40+ DiT pipeline families across NVIDIA, AMD, and Ascend hardware.

## Use cases
- speed up diffusion transformer inference
- accelerate flux image generation with caching
- run DiT models with context parallelism
- quantize diffusion models for faster inference
- offload diffusion model weights to CPU to fit in VRAM
- integrate DiT acceleration into ComfyUI or vLLM pipelines

## When to choose
- you use Diffusers DiT pipelines and need faster inference without changing models
- you need multi-GPU parallelism (context/tensor/2D/3D) for diffusion workloads
- you want cache, quantization, and offload combined in one library

## When to avoid
- you need to train or fine-tune diffusion models rather than run inference
- your models are not DiT-based or not supported by Diffusers
- you need a serving engine with its own API server rather than a library

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, gpu-computing, caching
- domain: deep-learning, image-processing, gpu-computing, machine-learning
- platform: python, windows
- tags: diffusion-transformers, dit, diffusers, inference-acceleration, quantization, cpu-offload, context-parallelism, tensor-parallelism, svdquant, video, gpu, linux

## Member repositories
- vipshop/cache-dit (main) score 82

## Provenance
- Observed fields: from GitHub, fetched 2026-09-03T02:15:20.342670+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T05:05:32.740245+00:00, confidence not recorded.
  - readme: https://github.com/vipshop/cache-dit (fetched 2026-09-03T02:15:20.342670+00:00, sha 229464764f1f)
  - registry_pypi: https://pypi.org/pypi/cache-dit/json (fetched 2026-08-29T12:17:50.952884+00:00, sha 9c0dda7dc2bd)
- Data as of 2026-08-30T08:39:29.467469+00:00.