A GLM5.2 production deploy hit a permanent vLLM scheduler deadlock combining MTP with a KV connector

GitHub·high signalXBlueskyLinkedInCopy link

vLLM PR #57104, merged 2026-09-16T22:22Z, fixes a scheduler deadlock under KV memory pressure when multi-token prediction runs alongside a KV connector, found in a production GLM5.2 deployment. The PR body walks a 7-block worked example: two requests are admitted to `skipped_waiting` as `WAITING_FOR_REMOTE_KVS`, request A's transfer lands and it is promoted to `WAITING`, but its MTP block (allocated only after the async load) needs one block when zero remain, so `allocate_slots` returns None and the loop breaks with the queues in exactly their prior state, forever.