[Bug]: multiprocessing KeyError from `cache[rtype].remove(name)` about vllm HOT 4 CLOSED

@njhill please check #5512 . That benign error comes from Python's incorrect resource tracking

from vllm.

I don't think your code crashes after this benign error. see #5468

from vllm.

You are in fact correct: of course, immediately after posting this bug report, I found out that vLLM is taking a while to connect and it's being killed by its supervisor.

Sorry for the annoyance. Maybe it'll save someone else from the red herring in the future...?

from vllm.

Though benign, it's still a bit strange / not ideal that these errors appear in the log. So this might still be a legit issue to track.

FWIW it does not seem to happen when fork is used instead of spawn (VLLM_WORKER_MULTIPROC_METHOD=fork). Looks like we'll probably make this the default soon anyhow: #5669.

from vllm.