Community ports DeepSelect to Hopper with CuTe DSL in TokenSpeed PR
zhyncs42 · x · 2026-10-17
The community is optimizing the TokenSpeed inference engine for DeepSeek V4.1 on Hopper. Since the official DeepSelect package ships Blackwell cubins only, contributor FC-Li opened PR #1737 adding an in-tree CuTe DSL rendition of DeepSelect (including the SM90 thread-block-cluster path), replacing FlashInfer's radix top-k for the Hopper indexer. Implementation: one CTA/cluster selects top-k of an FP32 row; the first 8192 entries are selected exactly via an 8-bit radix select over order-preserving keys. The poster notes the engine runs thousands of Hopper GPUs in production with PD disaggregation—the fastest DeepSeek V4.1 engine on Hopper yet.
Related event: TokenSpeed Engine Adds Day-0 Support for DeepSeek V4.1(2 posts)→
More from coding & agent
- Testing MiniMax Code for frontend design: daily check-in credits plus the Frontend Design skill — future_coded · 2026-10-17
- MemPalace gives coding agents local verbatim memory with 96.6% recall on LongMemEval — thisguyknowsai · 2026-10-17
- 'I built an agent and became its babysitter': a Reddit dev's honest realization — Luvena21 · 2026-10-17
- Dev builds a Gemini voice assistant that runs his Windows PC — lessons learned — 720beats · 2026-10-17
- Codex upgrades DSH and live-tests DeepSeek-V4.1-Flash 1017 — cedric_chee · 2026-10-17
- ArtCraft rebuilds Adobe suite in Rust: 13 open-source native apps, free — thisdudelikesAI · 2026-10-17