Repository navigation
[CI] Stop requiring both API servers to record weight sync metrics - #59521
Merged
aoshen02 merged 2 commits intoOct 4, 2026
Merged
Conversation
aarushjain29
requested review from
AndreasKaratzas,
DarkLight1337,
NickLucche,
aarnphm,
aoshen02 and
robertgshaw2-redhat
as code owners
October 1, 2026 00:23
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
aarushjain29
force-pushed
the
rocm/fix-weight-sync-metrics-flake
branch
from
October 1, 2026 00:23
5ae77da to
901be4a
Compare
Contributor
|
This pull request has merge conflicts that must be resolved before it can be |
test_weight_sync_metrics_aggregate_across_api_servers asserted that both API servers recorded a weight operation. The API servers accept from one shared listening socket, so the kernel can hand every connection to one process: AMD CI nightly build 13954 sent all 64 syncs allowed by vllm-project#59507 to the same server and still failed. Which server handles a request is not something the test controls, and summing across processes is already covered deterministically by test_in_flight_is_summed_across_api_server_processes. Send a fixed 8 syncs and keep the count and in-flight checks, which hold whichever server records. Signed-off-by: aarushjain29 <aarushi.jain2@amd.com>
aarushjain29
force-pushed
the
rocm/fix-weight-sync-metrics-flake
branch
from
October 1, 2026 16:29
901be4a to
bc68c25
Compare
aarushjain29
changed the title
Member
|
/ci run |
|
✅ Triggered Buildkite CI #92457 for commit |
Member
|
We tested this one 30 times in a loop cause we got again this one today: |
|
✅ @aarushjain29, CI is now available for this PR.
|
aoshen02
approved these changes
Oct 4, 2026
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
3 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.
The test asserted that both API servers recorded a weight operation.
The API servers share one listening socket, so the kernel can send every connection to one server. It was sending all 64 syncs allowed by #59507 to one server and still failed.
Adding up metrics across processes is already covered by test_in_flight_is_summed_across_api_server_processes, so this drops the check and keeps the count and in-flight checks with a fixed 8 syncs.
Verified locally on MI300X: 6/6 passes, including a run where only one API server handled requests