llama.cpp PR adds missing AMD GCN MMQ config, boosting MI50/MI60 inference

pmttyji · reddit · 2026-09-12

A llama.cpp pull request (#27841) adds the missing AMD GCN MMQ configuration for the ROCm/HIP backend.

According to the contributor, the change delivers prompt processing (PP) improvements for RDNA2 cards like the MI50 and MI60, with updated pp t/s benchmarks posted in the bottom comments.

Original post →

More from Infra

Infra channel →