InternLM 开源 AdaptiveGEMM：支持不同 Group M 长度的 FP8 GEMM
GitHub - InternLM/AdaptiveGEMM: AdaptiveGEMM: FP8 GEMM with Adaptation to Various Lengths of Group M · GitHub
InternLM 发布 AdaptiveGEMM，在 DeepGEMM 基础上支持不同长度的 group M，并新增 Group K GEMM 支持，每组 k 需补齐到 128x。该项目要求 Hopper 架构 GPU（sm_90a）、Python 3.8+、CUDA 12.3+（推荐 12.8+）、PyTorch 2.1+ 和 CUTLASS 3.6+。
正文
当前显示原文，中文全文译文尚未就绪。
InternLM; AdaptiveGEMM
Repository navigation
Code; Pull requests; Actions; Projects; Security and quality; Insights
NotificationsYou must be signed in to change notification settings; Fork4 (4); Star3 (3)You must be signed in to star a repository
Repository files navigation
README; License
Adaptation to Various Lengths of Group M: We have enhanced the repository to support different lengths of group M, providing more flexibility for diverse use cases; Support for Group K GEMM: We have also added support for Group K GEMM operations, expanding the functionality of the original DeepGEMM. NOTE: k in each group should be padded to 128x
Quick start
Requirements
Hopper architecture GPUs, sm_90a must be supported; Python 3.8 or above; CUDA 12.3 or above
But we highly recommend 12.8 or above for the best performance; PyTorch 2.1 or above; CUTLASS 3.6 or above (could be cloned by Git submodule)
If you find our work useful, please cite:
Original deepgemm
To get detailed information of original deepgeem, please refer to https://github.com/deepseek-ai/DeepGEMM