DeepSelect

TopK kernels for the DeepSeek Sparse Attention lightning indexer (bf16, top-k up to 4,096) and for samplers, running 2 to 20 times faster than torch.topk; the selection step that V3.2, V4, and V4.1-Flash run on every token. Released September 9, 2026 under MIT with a v1.0.0 tag and a design deep-dive the next day; 274 stars at filing.