vLLM 官方博客（网页）·· 2026-09-30AI score50
追踪瓶颈：在 AMD Instinct MI355X 上优化 MiniMax M3
Following the Bottleneck: Optimizing MiniMax M3 on AMD Instinct MI355XSep 10, 2026·16 min readA performance model for LLM serving: inspect local shapes, remove repeated work, verify data movement and dispatch, then follow the queue.
Not yet translated; shown in Chinese.
AI guide
vLLM 针对 AMD Instinct MI355X 优化 MiniMax M3 推理性能，在并发 32 下将 MXFP8 吞吐提升 3.14 倍至 342.4 output tokens/s/GPU。
Source:vLLM 官方博客（网页） · vllm.ai