Serving Qwen3.5-0.8B with a 0.016% weight edit: measure the loops, pin the sampling on the server

An edit to 64 columns in two layers ended most thinking loops: 62% right vs 8% on one RTX 4090. What to measure and which settings matter.

Read Original

Related