[V1] Only print cudagraph tqdm on rank 0 with is_global_first_rank

[mgoin]

Purpose

#19501 (comment)

Test Plan

Test Result

Before:

vllm serve meta-llama/Llama-3.1-8B-Instruct -tp 2
...
Capturing CUDA graphs:   0%|                                   | 0/67 [00:00<?, ?it/s]Capturing CUDA graphs: 100%|████████████████████████████████████| 67/67 [00:21<00:00,  3.07it/s]
(VllmWorker rank=0 pid=2326676) INFO 06-11 22:23:54 [custom_all_reduce.py:196] Registering 4355 cuda graph addresses
Capturing CUDA graphs: 100%|████████████████████████████████████| 67/67 [00:22<00:00,  3.03it/s]
(VllmWorker rank=1 pid=2326677) INFO 06-11 22:23:55 [custom_all_reduce.py:196] Registering 4355 cuda graph addresses
(VllmWorker rank=1 pid=2326677) INFO 06-11 22:23:55 [gpu_model_runner.py:2051] Graph capturing finished in 22 secs, took 0.60 GiB
(VllmWorker rank=0 pid=2326676) INFO 06-11 22:23:55 [gpu_model_runner.py:2051] Graph capturing finished in 22 secs, took 0.60 GiB

After:

vllm serve meta-llama/Llama-3.1-8B-Instruct -tp 2
...
Capturing CUDA graph shapes: 100%|██████████████████████████████████████████| 67/67 [00:21<00:00,  3.07it/s]
(VllmWorker rank=1 pid=2321732) INFO 06-11 22:22:10 [custom_all_reduce.py:196] Registering 4355 cuda graph addresses
(VllmWorker rank=0 pid=2321731) INFO 06-11 22:22:10 [custom_all_reduce.py:196] Registering 4355 cuda graph addresses
(VllmWorker rank=1 pid=2321732) INFO 06-11 22:22:10 [gpu_model_runner.py:2056] Graph capturing finished in 22 secs, took 0.60 GiB
(VllmWorker rank=0 pid=2321731) INFO 06-11 22:22:10 [gpu_model_runner.py:2056] Graph capturing finished in 22 secs, took 0.60 GiB

[github-actions]

👋 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.

🚀

[gemini-code-assist]

Summary of Changes

Hello @mgoin, 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 noisy console output in distributed VLLM runs by ensuring that the progress bar shown during CUDA graph capture is only displayed by the global rank 0 process. This is achieved by introducing a new helper function to check for the global first rank and using it to conditionally wrap the graph capture loop with tqdm.

Highlights

• Helper Function: Introduced a new helper function is_global_first_rank in vllm/distributed/parallel_state.py to determine if the current process is the global rank 0 across all parallelism dimensions (TP, PP, DP, etc.). This provides a robust way to identify the primary process for logging or UI updates.
• Conditional Progress Bar: Modified the CUDA graph capture logic in vllm/v1/worker/gpu_model_runner.py to only display the tqdm progress bar on the global first rank. This prevents duplicate progress bars from appearing when running with multiple workers/ranks, cleaning up the console output.

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 useful change to prevent multiple progress bars during CUDA graph capturing in distributed settings, enhancing log readability. The core logic involves a new is_global_first_rank utility.

The primary suggestion focuses on improving the error handling in this new utility function by making the exception catching more specific and considering logging for better diagnosability. This will make the function more robust for potential future uses beyond controlling tqdm visibility.

[gemini-code-assist]

The use of a bare except Exception: can mask various underlying issues from torch.distributed or the _WORLD state. This might lead to unexpected behavior if, for example, multiple processes erroneously assume they are the global first rank due to an unrelated exception during rank determination.

To improve robustness and diagnosability, consider the following:

1. Capture the exception instance: Change except Exception: to except Exception as e:. This allows access to the specific error.
2. Log the exception: Logging the error e would be beneficial for debugging. This helps in understanding why rank determination might have failed. (Assuming a logger is available or can be set up in this module, similar to other vllm.distributed files).
3. Docstring update: The current docstring explains that the function returns True if distributed is not initialized. It would be helpful to also document the behavior of returning True in case of any other exception during rank checking, if this is the intended safe default for all scenarios.

If returning True on any error is a deliberate and safe design choice for all potential uses of this function, explicitly stating this rationale in a comment within the except block or the docstring would improve clarity.

Suggested change

	except Exception:
	# If anything goes wrong, assume this is the first rank
	return True
	except Exception as e: # Capture the specific exception instance.
	# Consider logging 'e' for debugging to understand potential failures.
	# For example, if a logger is configured:
	# logger.warning("is_global_first_rank() encountered an error, defaulting to True: %s", e, exc_info=True)
	# If anything goes wrong, assume this is the first rank
	return True

[russellb]

logging a debug message at least would be nice here

[njhill]

The trouble with using global first rank is that we now have cases where different ranks for the same deployment may run in different places, via separate CLI launches. For example multi-node DP.

I'll try to think of how we can restrict this to running once per top-level vllm serve invocation.

This is still probably better than having it run in every rank though.

[russellb]

seems like an improvement, even if there's more that could be done (based on nick's comment)

[russellb]

logging a debug message at least would be nice here

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @mgoin.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @mgoin.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork