DeepSeek open-sources Ascend versions of its GEMM, expert-parallel and attention kernel libraries

Why it matters

【Key points】

On September 30, 2026 DeepSeek released Huawei Ascend versions of the low-level libraries behind its own training stack. DeepGEMM-Ascend is a port of DeepGEMM that is "fully API-compatible" with it, covering BF16, FP8, FP4 GEMM, MQA logits and MegaMoE. DeepEP-Ascend provides expert-parallel all-to-all dispatch and combine for MoE training and inference, with buffer APIs aligned with the NVIDIA DeepEP. The same day FlashMLA, DeepSelect, TileKernels and DeepJIT gained Ascend kernels or backends. TileLang's Ascend 950 backend lives upstream in tile-ai/tilelang, credited there mainly to DeepSeek AI engineers.

【Verifiable】

Primary: https://github.com/deepseek-ai/DeepGEMM-Ascend and https://github.com/deepseek-ai/DeepEP-Ascend .

DeepGEMM-Ascend (Ascend 950DT, CANN 9.20): dense GEMM up to 99.8% of the hardware limit; BF16xBF16 at 4096x7168x16384 is 431 TFLOPS against a 432 limit. DeepEP-Ascend (Ascend 950DT, CANN 9.2.0): EP8 dispatch 373-375 GB/s and combine 345-347 GB/s, with sustained dispatch about 90-95% of the physical payload bandwidth limit up to EP32. FlashMLA on Ascend 950: 410 TFlops prefill (95% of hardware peak), 360 TFlops decoding (83%). Licences are MIT except DeepEP-Ascend, which has no LICENSE file; its numbers come from a PoC HDK, not a public release.

【Our read】

The layer is what matters: GEMM, expert-parallel communication, JIT and kernel libraries, not weights or an inference adapter, and the APIs mirror the NVIDIA ones, so porting resembles a package swap.

Available is not proven. No large-scale Ascend training run appears in these repos; DeepEP-Ascend lists its reduce-scatter, all-reduce and load-balancing kernels as unimplemented and PP and Engram as experimental.

The numbers resist independent reproduction: all are tied to Ascend 950/950DT and specific CANN versions, and DeepEP's bandwidth to an unreleased HDK. A repo with no licence file is strictly all rights reserved, limiting downstream commercial use until fixed.