Summary

DeepGEMM Ascend is DeepSeek's efficient matrix-operation library for Huawei Ascend hardware. It exposes DeepGEMM-compatible APIs over Ascend matrix primitives while handling platform layouts, alignment, and optimized data movement.

Key Features

DeepGEMM API compatibility.
BF16 matrix multiplication.
FP8 kernels.
FP4 kernels.
MQA logits.
MegaMoE support.
Ascend MAD abstraction.
Coroutine-based pipelining.

About DeepGEMM Ascend


AI infrastructure developers can port compatible model kernels and reuse familiar workflows on Ascend accelerators. Core capabilities include DeepGEMM API compatibility, BF16 matrix multiplication, FP8 kernels, and FP4 kernels.


Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include MQA logits, MegaMoE support, Ascend MAD abstraction, and Coroutine-based pipelining.

Get more likes & reach the top of search results by adding this button on your site!

Embed button preview - Light theme
Embed button preview - Dark theme
TurboType Banner

Subscribe to the AI Search Newsletter

Get top updates in AI to your inbox every weekend. It's free!