# deepseek-ai/DeepGEMM

DeepGEMM: clean and efficient BLAS kernel library on GPU

Repository: https://github.com/deepseek-ai/DeepGEMM
Canonical: https://ross.abutalabs.com/products/deepgemm
Language: Cuda
License: MIT
License Family: permissive
Last push: 2026-08-11T08:21:15+00:00

## Health v2 (maintenance only)
Score: 84/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 97, release rhythm 93, longevity 40
- inputs: {"age_days": 566, "days_push": 22, "days_rel": 44, "gap_med": 19, "n_releases_24m": 10}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 7738, forks 1216 (observed 2026-08-28T04:10:02.694559+00:00)

## What it is
DeepGEMM is a high-performance CUDA BLAS kernel library for NVIDIA tensor cores, providing FP8, FP4, and BF16 GEMMs plus fused MoE and other LLM computation primitives with runtime JIT compilation. It is designed to be clean and simple, matching or exceeding expert-tuned libraries on Hopper (SM90) and Blackwell (SM100) GPUs.

## Use cases
- speed up fp8 gemm kernels for llm inference on h100
- efficient moe fused kernels with overlapped communication
- learn nvidia gpu kernel optimization techniques
- bf16 matrix multiplication on sm90 gpus
- integrate fast gemm into a training framework
- mqa scoring kernels for lightning indexer

## When to choose
- you need near-peak tensor core GEMM performance on SM90/SM100 NVIDIA GPUs
- you want a small, readable CUDA codebase for studying kernel optimization
- you need FP8/FP4 GEMMs or fused MoE kernels for LLM workloads without heavy template frameworks

## When to avoid
- you need BLAS on non-NVIDIA hardware or pre-Hopper GPUs
- you need a general-purpose BLAS with broad routine coverage beyond GEMM-style operations
- you require a stable API with long-term compatibility guarantees

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, gpu-computing, llm-inference, llm-training, benchmarking
- domain: deep-learning, large-language-models, gpu-computing, machine-learning, developer-tools
- platform: python, cpp
- tags: cuda, gemm, blas, tensor-cores, fp8, moe, jit-compilation, kernel-optimization, linux, gpu

## Member repositories
- deepseek-ai/DeepGEMM (main) score 84

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:02.694559+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 2026-08-29T17:35:11.299743+00:00, confidence not recorded.
  - readme: https://github.com/deepseek-ai/DeepGEMM (fetched 2026-08-28T04:10:02.694559+00:00, sha 815cb9b45aae)
- Data as of 2026-08-30T08:39:29.467469+00:00.