# luminal-ai/luminal

Inference at the speed of light.

Repository: https://github.com/luminal-ai/luminal
Canonical: https://ross.abutalabs.com/products/luminal
Homepage: https://luminal.com
Language: Rust
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T21:29:57+00:00

## Health v2 (maintenance only)
Score: 78/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 50, longevity 81
- inputs: {"age_days": 1142, "days_push": 7, "days_rel": 121, "gap_med": null, "n_releases_24m": 1}
- flags: prerelease_only
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2956, forks 225 (observed 2026-08-28T04:07:32.150955+00:00)

## What it is
Luminal is a high-performance general-purpose ML inference compiler written in Rust that lowers models to a minimal 15-op dataflow IR and compiles them to optimized GPU/ASIC kernels. It integrates natively with PyTorch as a torch.compile backend and also offers a Rust tensor API, running models like Llama 3 8B near theoretical peak performance on H100s.

## Use cases
- run llama 3 8b locally on cuda with maximum speed
- compile pytorch models to optimized gpu kernels
- serve llm inference at near-theoretical-peak throughput on h100
- avoid heavyweight ml frameworks with a minimal rust tensor library
- automatically discover kernel fusions like flash attention without hand-written ops
- compile transformers and convnets ahead of time for gpu or asic

## When to choose
- you need maximum inference throughput on NVIDIA GPUs and are comfortable with Rust or PyTorch integration
- you want a small, hackable ML compiler core instead of a large framework
- you want automatic kernel optimization (fusion, tiling, scheduling) via search rather than heuristics

## When to avoid
- you need broad production support for many accelerators, quantization formats, or serving features today
- you rely on a mature ecosystem of ops, model zoos, and tooling like PyTorch or vLLM
- you need CPU-only or non-CUDA hardware support as a primary target

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, compiler, machine-learning, deep-learning, gpu-computing
- domain: machine-learning, deep-learning, large-language-models, compilers, gpu-computing, performance
- platform: rust, python, windows
- tags: inference-compiler, cuda, pytorch-backend, kernel-fusion, tensor-api, aot-compilation, gpu, linux, macos

## Member repositories
- luminal-ai/luminal (main) score 78

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:32.150955+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 2026-08-30T07:32:26.736704+00:00, confidence not recorded.
  - readme: https://github.com/luminal-ai/luminal (fetched 2026-08-28T04:07:32.150955+00:00, sha 773f1b821a9c)
  - homepage: https://luminal.com (fetched 2026-08-29T09:47:35.947111+00:00, sha 122a3f903e07)
- Data as of 2026-08-30T08:39:29.467469+00:00.