#8989th of 105 in ML frameworks and libraries
deepgemm
CUDA tensor core kernels for FP8, FP4 and BF16 GEMMs and fused MoE
- Stars
- 9.1k
- License
- MIT
- Last commit
- Oct 2026
- Last release
- Oct 2025
- Language
- Cuda
6,244 on 1 Apr 20269,066 on 21 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
DeepGEMM is a CUDA kernel library for NVIDIA SM90 and SM100 GPUs, called from Python through PyTorch. It provides dense and grouped FP8/FP4/BF16 GEMMs, MQA logits kernels for the DeepSeek V3.2 indexer, and a fused Mega MoE kernel that overlaps expert-parallel communication with compute. Kernels are compiled at runtime by its own JIT, so installation needs no CUDA compilation.
Who it is for: Engineers writing LLM inference or training code on Hopper and Blackwell GPUs
Strengths
- Covers dense, contiguous and masked grouped GEMMs, MQA logits and fused Mega MoE in one library
- Runtime JIT compilation, so install does not compile CUDA kernels
- Small set of core kernels without heavy CUTLASS templates, readable for learning
- MIT license; supports both SM90 and SM100 GPUs
Weaknesses
- Only NVIDIA SM90 or SM100 GPUs are supported
- Needs CUDA Toolkit 12.9+, PyTorch 2.3+ and a C++20 compiler; Mega MoE needs PyTorch 2.9+
- Caller must handle input transposition, FP8 casting and scaling-factor layout
- SM90 supports only the NT memory layout
What it needs
- GPU required
- Needs CUDA Toolkit 12.9+, PyTorch 2.3+, CUTLASS 4.0+, Python 3.8+
- Models: FP8, FP4, BF16, DeepSeek V3.2 MQA indexer
Also in ML frameworks and libraries
See all 105| Rank | Project | Score |
|---|---|---|
| 1 | transformersPython library defining open models for text, vision, audio and multimodal tasks | 94 out of 100 |
| 2 | pytorchPython tensor library with GPU support and tape-based autograd for deep learning | 94 out of 100 |
| 3 | ultralyticsPython library and CLI for training and running YOLO vision models | 93 out of 100 |
| #4 | pytorch-image-modelsPyTorch library of image models, pretrained weights and reference training scripts | 92 out of 100 |
| #5 | supervisionPython toolkit for loading, annotating and counting computer vision detections | 91 out of 100 |