Skip to content

FIX: NixlConnector: do not skip short do_remote_prefill requests #18590

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
wants to merge 2 commits into from
Closed

FIX: NixlConnector: do not skip short do_remote_prefill requests #18590

wants to merge 2 commits into from

Conversation

juncgu
Copy link

@juncgu juncgu commented May 23, 2025 •
edited by github-actions bot
Loading

The original issue is #18490, and #18429 wanted to fix it.

But skipping short do_remote_prefill request in connector's metadata will also skip sending the notification to the prefill worker for releasing remote kv pages.

The modifications here are

  1. revoke the changes from [Bugfix][Failing Test] Fix nixl connector test when promt size < block size #18429
  2. update test_nixl_connector.py::test_prompt_less_than_block_size accordingly

FIX #18591

@github-actions GitHub Actions
Copy link

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

@mergify mergify bot added the v1 label May 23, 2025
wwl2755
wwl2755 approved these changes May 23, 2025
View reviewed changes
Copy link
Contributor

@wwl2755 wwl2755 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM! Thanks for pointing out! Sorry for the oversight on existing logic of handling 0 block situation.

@juncgu
tweaks
7d5333a 
Signed-off-by: Juncheng Gu <[email protected]>
@njhill
Copy link
Member

njhill commented Jun 2, 2025

This overlaps with #18632 but it would be good to include your test changes.

@njhill
Copy link
Member

njhill commented Jun 2, 2025

@juncgu I've pulled your test updates into #18632 (still with you as the commit author).

@juncgu
Copy link
Author

juncgu commented Jun 2, 2025

@juncgu I've pulled your test updates into #18632 (still with you as the commit author).

Thanks, @njhill. Then, I will close this PR.

@juncgu juncgu closed this Jun 2, 2025
@njhill
Copy link
Member

njhill commented Jun 3, 2025

@juncgu actually I reverted the test change because it was failing and I don't think it was correct.

When the number of tokens is < block_size, so that 0 blocks are to be sent, they are immediately freed and so the P worker doesn't require the D worker to make a request to free them. For this reason, the D worker side of the connector (correctly) does nothing the metadata contains no blocks.

@juncgu
Copy link
Author

juncgu commented Jun 3, 2025

@juncgu actually I reverted the test change because it was failing and I don't think it was correct.

When the number of tokens is < block_size, so that 0 blocks are to be sent, they are immediately freed and so the P worker doesn't require the D worker to make a request to free them. For this reason, the D worker side of the connector (correctly) does nothing the metadata contains no blocks.

Yes, you are right, the current scheduler / connector_scheduler do not need the changes in the unit test.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@wwl2755 wwl2755 wwl2755 approved these changes

Assignees
No one assigned
Labels
v1
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

[Bug]: NixlConnector should not skip short do_remote_prefill requests in connector metadata
3 participants
@juncgu @njhill @wwl2755