Issue evidence · Serve · vllm-project/vllm

KV offload back-pressure ignores failed secondary-tier stores, so a full filesystem tier keeps receiving writes

GitHub issue: https://github.com/vllm-project/vllm/issues/61775

Confirmed from source (not run)

TieringOffloadingManager only reports successful primary-to-secondary store completions to the back-pressure detector, so a tier that rejects writes never triggers back-pressure.

What Badgr ran

Read vllm/v1/kv_offload/tiering/manager.py on vLLM main (2026-10-15). Not run: no filesystem tier was filled.

What came back

  • At line 545 the back-pressure update sits under "if completed_job.success:", as the issue quotes.
  • Failed jobs take the other branch and are not fed to the detector.

Code reading only (not run): the issue's reproduction is a unit test in the repo that was not executed here.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence