Can llama-server do round-robin load balancing across 3 AMD MI50s?

HlddenDreck · reddit · 2026-09-01

A user running llama-server locally on three AMD MI50s found tensor parallelism too slow due to PCIe 3.0 bandwidth and cards on different NUMA nodes. They want each GPU running the same model with server-side round-robin request distribution (request 1 to card 1, request 2 to card 2), without client-side multi-port load balancing. They ask whether llama-server alone can do this or if middleware is needed.

Original post →

More from Infra

Infra channel →