# mirage-project/mirage

Mirage Persistent Kernel: Compiling LLMs into a MegaKernel

Repository: https://github.com/mirage-project/mirage
Canonical: https://ross.abutalabs.com/products/mirage-project-mirage
Homepage: https://mirage-project.readthedocs.io/
Language: Cuda
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T15:41:44+00:00

## Health v2 (maintenance only)
Score: 83/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 75, longevity 60
- inputs: {"age_days": 847, "days_push": 7, "days_rel": 89, "gap_med": 71, "n_releases_24m": 6}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2452, forks 242 (observed 2026-08-28T04:06:52.980972+00:00)

## What it is
Mirage Persistent Kernel (MPK) is a compiler and runtime that transforms multi-GPU LLM inference into a single fused megakernel, reducing inference latency by 1.2x to 6.7x. It compiles Hugging Face models like Qwen3 into a persistent GPU kernel with minimal developer effort.

## Use cases
- compile LLMs into a single megakernel for low-latency inference
- reduce multi-GPU LLM inference latency without manual kernel tuning
- fuse computation and communication across GPUs into one kernel launch
- profile and visualize GPU task execution timelines for LLM inference
- serve Hugging Face models like Qwen3 with a persistent kernel runtime

## When to choose
- you need minimal inference latency for LLM serving on NVIDIA GPUs
- you want to avoid hand-writing fused CUDA kernels for multi-GPU inference
- you're comfortable building from source and working with a research-grade compiler

## When to avoid
- you need a production-ready inference server with broad model support out of the box
- you don't have NVIDIA GPUs or multi-GPU setups
- you need pre-built wheels or stable APIs, which are still in progress

## Facets
- artifact type: library
- maturity: active
- function: compiler, llm-inference, gpu-computing, machine-learning, interpreter
- domain: large-language-models, gpu-computing, compilers, machine-learning, performance
- platform: python
- tags: megakernel, kernel-fusion, multi-gpu, inference-optimization, cuda, persistent-kernel, huggingface, gpu, linux

## Member repositories
- mirage-project/mirage (main) score 83

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:52.980972+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 2026-08-30T02:30:10.629968+00:00, confidence not recorded.
  - readme: https://github.com/mirage-project/mirage (fetched 2026-08-28T04:06:52.980972+00:00, sha e14230550c59)
- Data as of 2026-08-30T08:39:29.467469+00:00.