DeepGEMM

Headlines · 2

1. 3h
   DeepGEMM adds Ascend support and new kernels, trends on GitHub

   DeepGEMM is DeepSeek's open-source, high-performance tensor-core kernel library that unifies many core LLM computation primitives — FP8/FP4/BF16 GEMMs, fused MoE with overlapped communication (Mega MoE), MQA scoring for the lightning indexer (including a sparse version) and HyperConnection (HC) — in a single CUDA codebase, with all kernels compiled at runtime via DeepJIT and no CUDA compilation at install time. According to the repository's changelog, DeepGEMM-Ascend became available on 2026.09.30 alongside new optimizations such as locality domain features, after earlier additions of the Sparse Indexer, Mega Gate, Mega mHC and MoE/Indexer optimizations. The repo also says its performance matches or exceeds expert-tuned libraries across various matrix shapes, and that it reached up to 1550 TFLOPS on H800 in April 2025.

   GitHub Trending（每日） · 🔥 26
2. 1d
   vLLM v0.31.0 expands serving, speculation, and model support

   vLLM v0.31.0 adds broad serving and model-support updates, including DeepSeek-V4.1-Flash optimizations, Model Runner V2 speculative decoding, larger-scale expert parallelism, and improved multimodal support. It also introduces the `vllm preload` weight-cache daemon for faster engine restarts and experimental CRIU-based initialized-engine snapshots, alongside several security and breaking changes.

   vLLM Releases · 🔥 4

Experience and discussion from the community

Share my DeepGEMM experienceAsk about DeepGEMM