Validates artifacts

[finbarrtimbers]

We ran into issues on Olmo 3 where we were supposedly checkpointing, but the checkpoints weren't happening. This validates that the artifacts (checkpoints, rollouts) actually exist.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[gemini-code-assist]

Summary of Changes

Hello @finbarrtimbers, 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 introduces a critical set of artifact verification mechanisms to prevent silent failures during model training and saving. By adding explicit checks for the existence and integrity of model checkpoints, DeepSpeed state checkpoints, and rollout traces, it ensures that saved artifacts are valid and complete. This directly addresses past issues where checkpointing was reported as successful but the artifacts were found to be missing or corrupted, thereby improving the reliability of the training process.

Highlights

• Artifact Verification Functions: Introduced new functions (verify_model_checkpoint, verify_deepspeed_state_checkpoint, verify_traces) to validate the integrity and existence of saved model checkpoints, DeepSpeed state checkpoints, and rollout traces.
• Integrated Verification: Integrated these verification checks into the model saving (maybe_save_checkpoint, save_final_model) and DeepSpeed checkpointing workflows to ensure artifacts are valid upon creation.
• Verification Logging: Added a log_verification function to output verification results to the console and optionally to Weights & Biases (wandb) for better monitoring.
• Wandb Tag Handling: Adjusted the setup_experiment_tracking function to correctly handle args.exp_name as an optional tag for wandb, preventing potential errors when exp_name is not provided.
• Type Hint Refinement: Refined the type hint for loss_denominator in ExperimentConfig from str | float to str and added a type ignore comment to address a type checking issue.

🧠 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 introduces validation for artifacts like model checkpoints, DeepSpeed states, and training traces to ensure they are created correctly. The changes look good and address the problem described. I've provided a few suggestions to improve the conciseness and readability of the new verification functions. Additionally, I've identified a type hint inconsistency in grpo_utils.py that should be corrected.

[gemini-code-assist]

This function can be simplified for better readability and efficiency.

1. The check for rollouts_path.exists() is redundant because metadata_file.exists() implicitly checks for the parent directory's existence.
2. Using any() with a generator expression is more memory-efficient than creating a list with glob() and checking its length, as any() will stop at the first match.

    rollouts_path = pathlib.Path(rollouts_save_path)
    metadata_file = rollouts_path / f'{run_name}_metadata.jsonl'
    if not metadata_file.exists():
        return False

    return any(rollouts_path.glob(f'{run_name}_rollouts_*.jsonl'))

[gemini-code-assist]

This type: ignore is only necessary because the type hint for loss_denominator was incorrectly changed to str. If you restore the type hint to str | float as suggested in the other comment, this ignore comment can be removed.

Suggested change

	self.loss_denominator = utils.get_denominator(self.loss_denominator) # type: ignore[assignment]
	self.loss_denominator = utils.get_denominator(self.loss_denominator)