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)→

Original post →

More from coding & agent

coding & agent channel →