llama.cpp Local Inference Debugging: Sudden 50% TPS Drop

campaigner_ · reddit · 2026-08-02

A developer encountered a bizarre issue where their local inference speed (TPS) using llama.cpp suddenly halved on an RTX 3070 + i7-11700 setup.

The Issue:

Previously, running Qwen and Gemma models maintained a stable 26-30 t/s. Recently, the speed dropped to 12-13 t/s at startup, and reducing the context size from 131k to 32k didn't help.

The author shares their full startup parameters (including --cpu-moe, --no-mmap, --chat-template-kwargs, etc.) to ask the community for help in troubleshooting whether the bottleneck is due to parameter conflicts or underlying updates.

Related event: Windows Reset Fixes Halved Local LLM Inference Speeds on RTX 3070(2 posts)→

Original post →

More from Infra

Infra channel →