Following the Bottleneck: Optimizing MiniMax M3 on AMD Instinct MI355X

A performance model for LLM serving: inspect local shapes, remove repeated work, verify data movement and dispatch, then follow the queue.