Building the vLLM RDNA2 fork for 4× Radeon Pro V620: build, packaging, and benchmarking notes
A full day bringing vllm-rdna HEAD (gfx1030) into production on 4x V620s: the 7-architecture build trap, docker commit's entrypoint bug, +/-3.5% bench noise, the exact final config, and the branch's headline number I never reproduced.…