vLLM V1 修復後實現後端對等
huggingface 工程師在強化學習工作負載中恢復了 vllm v0.8.5 和 v1.0.18.1 間的後端一致性。四個具體修復解決了 grpo 訓練期間的滾動日誌概率、運行時默認值和權重更新處理差異。
Engineers achieved parity between vLLM V0.8.5 and V0.18.1 by resolving four critical backend discrepancies in an online RL setup. The fixes addressed processed rollout logprob semantics, explicit runtime defaults for prefix caching and async scheduling, and the inflight weight-update path. This established a stable baseline for V1, enabling objective-level corrections without semantic interference.