VLLM TRIPLES MINIMAX M3 INFERENCE THROUGHPUT ON AMD'S MI355X THROUGH SYSTEMATIC BOTTLENECK CHASING

Summary from the TLDR InfoSec newsletter

AMD and Embedded LLM teams took MiniMax M3's day-0 vLLM implementation from 109.1 to 342.4 output tokens/s/GPU on AMD's Instinct MI355X, a 3.14x gain, by repeatedly asking five questions as the bottleneck kept shifting: what local shape reached each rank after tensor-parallel sharding, what work repeats every layer, which bytes actually move, whether a claimed fast path actually executed, and which queue grows once kernels stop being the limiter. Big wins came from fusing the model's shared expert into the same grouped GEMMs as routed experts (up to 30% faster at low concurrency), and reusing sparse-attention block selections across adjacent layers instead of recomputing per layer.