[bugfix] fix logps with PP

[hjh0119]

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 resolves a bug in the Megatron model's pipeline parallelism implementation where log probabilities were being computed or accessed incorrectly across different pipeline stages. The core change ensures that logps are exclusively handled by the final stage of the pipeline, aligning with the correct behavior for loss calculation in such distributed environments. Additionally, the PR includes minor code cleanup by removing unnecessary imports.

Highlights

• Log Probability Calculation Fix: Corrected the calculation of log probabilities (logps) in the Megatron model's forward step. The fix ensures that logps are only processed on the last pipeline stage, which is crucial for accurate loss computation in pipeline-parallel setups.
• Code Cleanup: Removed two unused import statements from cookbook/megatron/tp.py, improving code cleanliness and reducing potential clutter.

🧠 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.

Changelog

• cookbook/megatron/tp.py
  • Removed import for input_feature_to_datum.
  • Removed import for TwinkleCompatModelBase.
• src/twinkle/model/megatron/megatron.py
  • Modified the condition for logps calculation to include a check for mpu.is_pipeline_last_stage().

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 a bugfix for calculating log probabilities (logps) in a pipeline parallel (PP) setup. The change in src/twinkle/model/megatron/megatron.py correctly ensures that logps are computed only on the last pipeline stage by adding mpu.is_pipeline_last_stage() to the condition. This is the correct approach, as only the final stage has access to the full logits required for this calculation. Additionally, the pull request cleans up unused imports in cookbook/megatron/tp.py. The changes are well-implemented and I have no further suggestions for improvement.