Repository navigation
[Bugfix][KV Offload] Reuse in-flight async lookup probes - #55823
Merged
Merged
Conversation
Retain submitted unresolved lookup state so replacement requests can share the existing probe. Preserve cleanup before submission and reclaim unclaimed results during flush and shutdown. Validation: git diff --check passed. Tests and commit hooks were not run at user request while the environment is being prepared. Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Alex <jihui.huang@daocloud.io>
Replace the submitted flag with explicit pending, in-flight, and resolved phases so lookup reuse and cleanup decisions are represented by the probe lifecycle. Validation: git diff --check passed. Tests and commit hooks were not run at the user's request while the environment is being prepared. Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Alex <jihui.huang@daocloud.io>
Document cached lookup verdicts and update the stale-generation test to submit the replacement probe before injecting its result. Tests: .venv/bin/python -m pytest tests/v1/kv_offload/tiering/test_async_lookup.py -v Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Alex <jihui.huang@daocloud.io>
Describe RESOLVED as a final verdict so the lifecycle comment covers both worker results and forced misses. Tests: not rerun; comment-only change. Prior targeted test run passed (18 tests). Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Alex <jihui.huang@daocloud.io>
Alex-ai-future
marked this pull request as ready for review
September 9, 2026 01:49
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Contributor
|
/ci run |
|
✅ Triggered Buildkite CI #87883 for commit |
|
✅ @Alex-ai-future, CI is now available for this PR.
|
Contributor
|
/ci run |
|
✅ Triggered Buildkite CI #88122 for commit |
Contributor
|
/ci retry |
|
✅ Queued 8 failed job(s) for retry in Buildkite CI #88122. |
orozery
approved these changes
Sep 14, 2026
ItsRoy69
pushed a commit
to ItsRoy69/vllm
that referenced
this pull request
Sep 15, 2026
…t#55823) Signed-off-by: Alex <jihui.huang@daocloud.io> Co-authored-by: OpenAI Codex <noreply@openai.com> Co-authored-by: Or Ozeri <oro@il.ibm.com>
keneoneth
pushed a commit
to keneoneth/vllm
that referenced
this pull request
Sep 16, 2026
…t#55823) Signed-off-by: Alex <jihui.huang@daocloud.io> Co-authored-by: OpenAI Codex <noreply@openai.com> Co-authored-by: Or Ozeri <oro@il.ibm.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Purpose
Fix duplicate same-key async existence probes in
AsyncLookupManager.When request A finishes while its probe is still in flight, cleanup retains the
probe state. A later request B can attach to the same probe and reuse its
verdict, so the backend is accessed once instead of once per request.
The probe lifecycle is explicitly modeled with
LookupPhase:PENDING: accumulated but not submitted.IN_FLIGHT: submitted to the worker and awaiting a result.RESOLVED: the result has been applied.This keeps probe lifecycle management separate from the cached HIT/MISS verdict
and preserves generation checks for stale results. The change is scoped to
same-key existence lookup reuse; it does not duplicate stale-result correctness,
cross-key batching, or promotion/load coalescing work.
The duplicate-probe behavior was identified in this
PR #52103 comment,
which includes a deterministic regression demonstrating the extra backend
probe.
This change does not affect model outputs or accuracy, so model evaluation is
not applicable. AI assistance was used to implement and validate this change.
Test Plan
pre-commit run --files vllm/v1/kv_offload/tiering/async_lookup.py tests/v1/kv_offload/tiering/test_async_lookup.py.venv/bin/python -m pytest tests/v1/kv_offload/tiering/test_async_lookup.py -vTest Result
18 passed, 14 warnings.torch.jit.script_methoddeprecation warnings.