add node_rank label

[Freddy9527]

Motivation

Modifications

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

[gemini-code-assist]

Summary of Changes

Hello @Freddy9527, 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 enhances the metrics collection system by incorporating the node rank. This change is crucial for accurately monitoring and debugging performance in distributed environments where the system spans multiple nodes. The scheduler now calculates the node rank and passes it to the metrics collector, which uses it to label and organize collected statistics.

Highlights

• Metrics Labeling: Added the node_rank label to the metrics collected by the scheduler. This allows for better monitoring and analysis of performance and resource usage on a per-node basis in distributed setups.
• Metrics Collection Structure: Modified the internal storage of labeled gauges in the SchedulerMetricsCollector to be keyed by a tuple (node_rank, dp_rank) instead of just dp_rank. This ensures metrics are correctly associated with the specific node and data parallel rank.
• Stats Aggregation: Included the node_rank in the tensor used to aggregate scheduler statistics across processes, increasing the tensor size from 15 to 16 elements. This allows the receiving process (rank 0) to identify the node rank associated with each set of statistics.

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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

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 issue 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 is currently in preview and 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 to provide feedback.

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 configureGemini 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 introduces a node_rank label to the metrics system, primarily affecting how SchedulerMetricsCollector is initialized and how it labels metrics in multi-node environments. Key changes include calculating node_rank in the scheduler and passing it to the collector, and updating the collector to use this new rank information.

My review has identified a few points:

• A hardcoded value for gpus_per_node in scheduler.py could limit flexibility.
• There's a potentially critical issue in metrics/collector.py where the gen_throughput Gauge was removed but is still referenced, which will likely cause a runtime error.
• An outdated error message in metrics/collector.py needs updating.
• A minor redundancy in label handling was also noted.

Addressing these points will help ensure the robustness and clarity of the new metrics labeling.

[gemini-code-assist]

The number of GPUs per node (gpus_per_node) is hardcoded to 8. This value might not be suitable for all deployment environments, as the number of GPUs per node can vary significantly across different hardware setups.

To enhance flexibility and adaptability, it's recommended to make gpus_per_node configurable. This could be achieved by:

• Adding it as a server argument.
• Reading it from an environment variable.

This change would allow the metrics collection to accurately reflect node_rank in diverse infrastructures.

[gemini-code-assist]

The conditional block:

if "node_rank" not in labels:
    labels["node_rank"] = str(node_rank)

appears to be effectively dead code. The labels dictionary, when passed to the SchedulerMetricsCollector constructor from scheduler.py (line 568), already includes the node_rank key with its string value: "node_rank": str(node_rank).

Consequently, the condition "node_rank" not in labels on line 278 will always evaluate to false, meaning line 279 will never be executed. This block can likely be removed for improved code clarity and conciseness.

[gemini-code-assist]

The error message string in the ValueError still refers to an expected size of 15 for local_info:
f"local_info.size(0) != 15: {local_info.size(0)}"

However, the actual check performed on line 343 is if local_info.size(0) != 16:. To avoid confusion and ensure the error message accurately reflects the condition being checked, it should be updated to refer to the new expected size of 16.

Suggested change

	raise ValueError(f"local_info.size(0) != 15: {local_info.size(0)}")
	raise ValueError(f"local_info.size(0) != 16: {local_info.size(0)}")