vLLM 0.32.1rc0 adds diskless HTTP weight transfer

vLLM v0.32.1rc0 introduces an HTTP Weight-Transfer Protocol that streams updated weights from trainer nodes directly into inference workers' VRAM, bypassing checkpoint files on NFS, Ceph or Lustre. The change removes storage I/O saturation and GPU idle time in online RL pipelines, where cluster utilization often fell below 40%.

• Weights stream over HTTP to the /update_weights endpoint with no disk writes
• Previously 14B–70B checkpoints were dumped every few iterations, saturating storage
• Idle rollout GPUs drove aggregate cluster utilization below 40%
• Chunked layer streaming and mTLS are recommended for production