Fix for llama.cpp 100% CPU single core usage despite full GPU offload

MelodicRecognition7 · reddit · 2026-08-19

A Reddit user shared a bug report and patch for llama.cpp regarding 100% CPU single-core usage even when the model is fully offloaded to GPU. The proposed fix, which sets the cudaDeviceScheduleBlockingSync flag, successfully reduces CPU usage (to 30-50% during decode) but results in approximately a 6-8% drop in generation throughput (TPS). The author invites better solutions from the community.

Original post →

More from Infra

Infra channel →