A multi-node vLLM rig climbs from ~20 to 47 tok/s/user after topology tuning
TheZachMueller · x · 2026-07-21
- The author describes the final state of a “cursed” multi-node inference rig after processing 6.4M Sol tokens over 10 hours 47 minutes. - By moving from a conventional 6-GPU multi-node vLLM setup to asymmetric TP2/TP4, then a 30/48 layer split and async + NCCL, throughput climbed from about 20 tok/s/user to 47.1 tok/s/user. - The screenshot says the setup preserved exact outputs, fixed a pipeline-parallel deadlock, and reached 169 tok/s aggregate across four users.
Related event: Extreme Topology Tuning Boosts LLM Inference Throughput to 47 tok/s(2 posts)→
More from coding & agent
- Workshop to cover loop and graph engineering for AI-native software engineering — Al_Grigor · 2026-07-21
- Production agents may need a new PaaS layer built around audit and recovery — percoAi · 2026-07-21
- wikiman puts man pages, ArchWiki, DevDocs, and FreeBSD docs into one offline search index — JafarNajafov · 2026-07-21
- Open-source LangGraph coding agent only submits patches after tests pass — wusuiling-if · 2026-07-21
- An agent-friendly knowledge graph design that skips Neo4j and cuts GraphRAG costs — TheRedfather · 2026-07-21
- Factory CEO says 90% of coding tokens will run asynchronously in a future 'dark factory' — Training Data (Sequoia) · 2026-07-21