DeepGEMM-Ascend and DeepEP-Ascend

First-party Huawei Ascend ports of DeepSeek's two core kernel libraries, published a day apart at the end of September 2026 and validated on Ascend 950 hardware with CANN 9.2 and torch_npu. Both keep the CUDA versions' public APIs, so the same code runs on either stack, and both compile their device kernels at runtime through DeepJIT, the dual-target JIT runtime DeepSeek released three weeks earlier. Together with DeepSeek-V4, trained on Ascend, they complete the lab's second, non-NVIDIA stack in public.

DeepGEMM-Ascend (MIT, ~363 stars at filing) is "a port of DeepGEMM to the HUAWEI Ascend platform… fully API-compatible" and covers BF16, FP8 and FP4 GEMM, MQA logits for the Lightning Indexer, MegaMoE and the mHC prenorm kernel. It wraps Ascend's matrix multiply-add primitives so kernels avoid fractal layouts, alignment and address arithmetic, and leans on Ascend-specific tricks — sparse data loading, coroutine-based pipelining. On Ascend 950DT the README reports dense GEMM at 98.3–99.8% of the hardware limit across types (BF16 431 of 432 TFLOPS; FP8 861 of 865; FP4 1,701 of 1,730 at M=4096, N=7168, K=16384), grouped MoE GEMM at 818–862 TFLOPS, and an MQA-logits kernel that is FIX-pipe rather than compute bound at 99% utilization. One scaling-factor difference from NVIDIA is documented: UE8M0 pairs are packed into an int16 along K and stored MN-major.

DeepEP-Ascend (~177 stars; no LICENSE file in the repository at filing) provides the expert-parallel all-to-all for MoE dispatch and combine, including FP8 dispatch and deferred epilogues, with work-in-progress pipeline-parallel send/receive, Bucket collectives for context and data parallelism, and remote memory access through Engram. Its Ascend C kernels ride Huawei's HCCL/HCOMM, UBMEM and URMA stack. Measured on Ascend 950DT at 16,384 tokens per rank, hidden 7168 and top-6 routing over 256 experts, dispatch reaches 373–375 GB/s at EP8 and 313–320 GB/s at EP128, combine 345–347 down to 272–278 GB/s; sustained dispatch is "roughly 90–95% of the physical payload bandwidth limit for EP sizes up to 32," with larger EP sizes and combine still under optimization. The README is explicit that its numbers came from a proof-of-concept Huawei HDK supplied to DeepSeek with manual configuration, not a public release. DeepGEMM-Ascend's acknowledgements thank Huawei for engineering support; DeepEP-Ascend's thank the asc-comm team.