PR #37859: [Bugfix][Core] Ignore stale KV xfer callbacks after request cleanup (#37837)

• Repository: vllm-project/vllm
• Author: rohithj7
• State: open | merged: False
• Link: https://github.com/vllm-project/vllm/pull/37859

Description (problem / solution / changelog)

Purpose

#37837: in PD setups you can abort a request on the decode side while a KV transfer is still finishing. Cleanup removes the request from self.requests, then _update_from_kv_xfer_finished() gets a late finished_recving / finished_sending for that id. The code asserted the id was still in self.requests, which isn't true in that case, so the decode node dies.

Fix is to treat that as a stale callback and skip it instead of asserting. Added a couple tests for the late-notification path.

Test Plan

Linux + Python 3.12 + requirements/test.txt is what matches CI (the lockfile is Linux/CUDA oriented).

Lint:

uv pip install -r requirements/lint.txt
pre-commit install
pre-commit run --all-files

Tests (focused):

python -m pytest tests/v1/core/test_scheduler.py -v \
  -k "abort_request_waiting_for_remote_kvs or abort_request_finished_recving or ignore_late_finished"

Whole file if you feel like it:

python -m pytest tests/v1/core/test_scheduler.py -v

Commit with -s for DCO (Signed-off-by).

Test Result

Will paste pytest output here after I run the above on Linux.

Before: n/a (or describe how you repro’d it)

After:

<will add this soon>

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

Changed files