KV Cache Events with Attention DP bug fix (#16030)

[kartikx]

Motivation

When using kv cache events with TP and DP-attention, only DP rank 0 emits KV cache events.

The current test test_kv_events fails with:

python test/manual/test_kv_events.py TestKvEvents.test_kv_events_attn_dp

# startup logs ...

Collected 28 events from DP rank 0
Collected 0 events from DP rank 1

# followed by assertion failure ...

Also see #16030

Modifications

I make two changes:

1. Use self.attn_tp_rank == 0 to enable KV events.
2. Some cleanup in the test files.

Accuracy Tests

The updated test passes.

python test/manual/test_kv_events.py TestKvEvents.test_kv_events_attn_dp

# startup logs ...

Collected 8 events from DP rank 0
Collected 8 events from DP rank 1
Ran 1 test in 62.988s

OK

Benchmarking and Profiling

Not applicable.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• [s] Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments (/tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci) or contact authorized users to do so.
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @kartikx, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses a critical bug in the KV cache event emission mechanism within the sglang framework, specifically impacting scenarios where Data Parallelism (DP) attention is enabled. The core issue was that only the primary DP rank was emitting KV cache events, resulting in an incomplete view of the system's state. The fix involves adjusting the condition for enabling KV cache events in the Scheduler to correctly leverage the attention tensor parallelism rank, thereby ensuring that all necessary DP ranks contribute their events. Additionally, the PR includes minor but important cleanup and robustness improvements to the associated test files.

Highlights

• KV Cache Event Emission Fix: Resolved a bug where KV cache events were only emitted by Data Parallelism (DP) rank 0 when using KV cache events with Tensor Parallelism (TP) and DP attention, leading to incomplete event collection.
• Scheduler Logic Update: Modified the Scheduler initialization to correctly enable KV cache events by checking self.attn_tp_rank == 0 instead of tp_rank == 0, ensuring proper event emission across all relevant ranks.
• Test File Refinements: Cleaned up the test_kv_events.py file by importing a default model name for testing and adding explicit resource closure (sub.close(), context.term()) in the finally block to improve test robustness.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request addresses a bug where KV cache events were only emitted by DP rank 0 when using attention with data parallelism. The fix correctly uses attn_tp_rank to identify the master process within each DP group's attention TP group, ensuring that each DP rank emits events as expected. Additionally, the pull request includes valuable cleanups in the test files, such as ensuring ZMQ resources are properly closed to prevent leaks and replacing a hardcoded model name with a constant for better maintainability. The changes are correct and improve the codebase. I have no further suggestions.

[ishandhanani]

Note that in this PR (ai-dynamo/dynamo#5097), the author imports the ZMQ event publisher from inside of sglang and starts the kv events for all attn_tp_rank==0 in each dp_rank.

In an ideal world, we treat each dp_rank as its own worker (since each has its own kv$). If I understand correctly, this PR will emit kv events from all attn_tp_rank==0 for all dp_ranks?

Have you tested multinode models (happy to help if you don't have the hardware)? Take the TP16/DP8 case for DSR1. Do we get KV events from all 8 DP ranks across both nodes?

[ishandhanani]

/tag-run-ci-label

[kartikx]

Will do multi-node testing this week

[Fridge003]

/tag-and-rerun-ci