vLLM Large Scale Serving: DeepSeek @ 2.2k tok/s/H200 with Wide-EP

Introduction

Read in full here: