Resin: experimental Rust LLM inference library vibes-code onto Intel HPU in hours
charles_irl · x · 2026-10-31
ekzhang1's experimental open-source library Resin (313 stars) is a set of building blocks for LLM inference: a headless Rust engine (scheduler, KV cache, metrics) plus model-specific forward-pass kernels. In a highlighted GitHub issue, user abotsis says vLLM/vllm.cpp/llama.cpp all imposed limits on his obscure Intel/Habana HPU backend, but with Resin, Claude merged his forwards with Resin's scheduler/batcher in about 4-5 hours, achieving near-linear concurrency scaling to batch 16 — with templating and caching thrown in for free.
More from coding & agent
- repligate: Anthropic should focus on making agents more dominant — repligate · 2026-10-31
- Developer burns 967M Grok tokens in one day running 30 parallel agents — r0ck3t23 · 2026-10-31
- Chrome 157 ships HTML-in-Canvas: render live web pages into canvas with shaders — dotey · 2026-10-31
- Sketching a striking three.js scene in minutes with Claude and Tripo — nptacek · 2026-10-31
- Factory's Missions: The Emerging Primitive for Decomposing Large-Scale Agent Work — blaizedsouza · 2026-10-31
- Learner bot mines agent traces for recurring fixes and turns them into one-tap skill updates — bfrench · 2026-10-31