M2T3 — Sweep KV-head sharing¶
Run ./run.ps1. The chart shows capacity, not quality. The success criterion is to predict each cache ratio before running it and explain why changing KV heads in a config without compatible trained weights is invalid.
Run ./run.ps1. The chart shows capacity, not quality. The success criterion is to predict each cache ratio before running it and explain why changing KV heads in a config without compatible trained weights is invalid.