Tools
vLLM's Rust frontend stops returning 404 to RL weight-sync clients
HTTPVLLMWeightSyncClient was getting 404 on every lifecycle call through the Rust frontend, blocking existing HTTP trainers from syncing an online inference model. PR #56567 (merged 2026-09-14) exposes the dev-only init, start, draft start, update, finish and weight-version endpoints over the engine-core client RPCs already added by the merged gRPC support in #51316, sets an optional version only after the finish RPC succeeds, and supports an omitted finish body. Validation on RTX 3090 with OPT-125M covered unpacked and packed IPC at TP=1 and NCCL at TP=2 with a separate trainer GPU, with 128 generated tokens matching Hugging Face exactly after syncing from dummy to real weights.
Source
↳ Follow the thread