2026-10-05 · America/Los_Angeles · 社区动态 · #20

Llama-server kept re-processing the full prompt between OpenCode turns with Qwen3.6 35B-A3B on a 16GB card

Every turn in OpenCode started with a pause, and on most turns it was long and got longer as the session went on. Generation was fine once it started. The server is llama-server on my desktop, running Qwen3.6 35B-A3B at Q4_K_M on a 4060 Ti 16GB with most of the experts on the CPU through --n-cpu-moe, and OpenCode runs on my laptop. I assumed it was the slot, since the card runs at PCIe 4.0 x8 and I had just read a thread about 16GB cards on half the lanes. With most of the experts on the CPU, llama.cpp copies them to the GPU for every batch of prompt it processes, and raising -ub for bigger…

热度 44.6 / 100；排名与评分保留该期记录。