DeepSeek open-sources DeepSelect, a TopK kernel for its sparse attention that it says runs 2 to 20x faster than torch.topk
GitHub·medium signal
deepseek-ai/DeepSelect, created 2026-09-09 and at v1.0.0 on 09-10, implements the TopK kernel behind DeepSeek Sparse Attention's lightning indexer (used in V3.2, V4 and V4.1) plus a sampler TopK, and claims 2 to 20x over torch.topk. It targets two cases: bf16 indexer scores at any batch and vocab size, and fp32 sampling over roughly 128K vocabularies, with topk capped at 4096. It is the third DeepSeek infra repo in four days and had 279 stars at the time of this run.