vLLM turns on node-shared Engram tables by default and falls back instead of failing startup

GitHub·medium signalXBlueskyLinkedInCopy link

PR #57651 (merged 2026-09-21) makes `EngramConfig.dp_shared_memory` resolve to enabled automatically when CPU offload is on, `data_parallel_size > 1` and elastic EP is off — previously it had to be requested by hand and hard-failed startup when its requirements were not met. The concrete number that makes the fallback necessary: DeepSeek-V4.1-Flash's full Engram tables are 188.8 GiB against Docker's default 64 MiB `/dev/shm`, so the check compares the tmpfs total (identical on every rank, so no collective needed) and degrades to DP head sharding with a `warning_once` pointing at `--shm-size` / `--ipc=host`. GSM8K 5-shot on one GB200 node with no `--engram-config` flag scored 0.931 both flexible and strict.