Add logprob comparator for megatron runner

[fzyzcjy]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello, 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 crucial utility for validating the consistency of Megatron runner outputs by enabling the comparison of per-token log probabilities. It provides a robust mechanism to detect and report discrepancies between different runs, which is essential for ensuring model stability and correctness during development and optimization. This enhancement is a significant step in a larger series of improvements to the dumper and Megatron runner chain.

Highlights

• New Logprob Comparator Module: A new Python module, logprob_comparator.py, has been added to compare per-token log probabilities between baseline and target Megatron runner runs.
• Configurable Tolerance and Detailed Reporting: The comparator supports a configurable threshold for determining pass/fail status and provides detailed statistics on logprob differences, including max, mean, median, and percentile absolute differences, along with information about the worst-performing position.
• Data Loading and Merging: The module includes functionality to load and merge logprob entries from multiple rank JSON files, handling deduplication for identical logprobs from TP ranks and skipping invalid entries.

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.

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 a log probability comparator and associated tests. The implementation is solid and the tests are comprehensive. I've identified a minor performance improvement opportunity in the statistics calculation logic to avoid redundant computations. Overall, great work on this utility.

[gemini-code-assist]

There are a couple of minor inefficiencies in the statistics calculation:

1. statistics.mean(diffs) is called twice (once for passed and once for mean_abs_diff).
2. statistics.median(diffs) is called, which sorts the list internally, even though sorted_diffs is already available. This results in an unnecessary re-sort.

You can calculate the mean once and reuse it, and compute the median directly from the sorted list to improve performance.

    sorted_diffs = sorted(diffs)
    num = len(sorted_diffs)

    baseline_worst = baseline_entries[max_diff_key]
    target_worst = target_entries[max_diff_key]

    mean_abs_diff = statistics.mean(diffs)

    # Calculate median from the already sorted list to avoid re-sorting.
    if num % 2 == 1:
        # `num` is guaranteed to be > 0 here because `common_keys` is not empty.
        median_abs_diff = sorted_diffs[num // 2]
    else:
        median_abs_diff = (sorted_diffs[num // 2 - 1] + sorted_diffs[num // 2]) / 2

    return _CompareResult(
        passed=mean_abs_diff <= threshold,
        num_positions=num,
        max_abs_diff=max_abs_diff,
        max_diff_position=max_diff_key[1],
        max_diff_baseline_logprob=baseline_worst.logprob,
        max_diff_target_logprob=target_worst.logprob,
        max_diff_token_id=baseline_worst.token_id,
        mean_abs_diff=mean_abs_diff,
        median_abs_diff=median_abs_diff,
        p95_abs_diff=sorted_diffs[int(num * 0.95)] if num > 0 else 0.0,
        p99_abs_diff=sorted_diffs[int(num * 0.99)] if num > 0 else 0.0,
        baseline_mean_logprob=statistics.mean(baseline_logprobs),
        target_mean_logprob=statistics.mean(target_logprobs),
        threshold=threshold,
        per_position_diffs=diffs,
    )