TRL v1.11.0 deletes its own vLLM server, cutting 1,218 lines to ~130 and moving weight sync to NCCL

GitHub·high signalXBlueskyLinkedInCopy link

Hugging Face TRL shipped v1.11.0 on 2026-08-26 replacing `trl vllm-serve`, a custom FastAPI wrapper with its own data-parallel fan-out and weight-sync worker extension, with a thin translation layer over `vllm serve`. The script drops from 1,218 to about 130 lines, fastapi/uvicorn/pydantic leave the vllm extra, and weight sync now rides vLLM's NCCL weight-transfer engine as a single packed broadcast instead of one HTTP request plus broadcast per tensor. Distillation teacher logprobs now come from `prompt_logprobs`.