Fix megatron loss

[tastelikefeet]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Write the detail information belongs to this PR.

Experiment results

Paste your experiment result here(if needed).

[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 addresses a bug in the Megatron model's loss calculation mechanism. It introduces explicit per-token loss calculation, refines the aggregation logic within the forward pass to correctly account for the number of tokens, and modifies the loss reduction strategy to ensure accurate normalization of the loss value. These changes collectively ensure that the model's loss is computed and averaged correctly, which is crucial for stable training.

Highlights

• Loss Calculation Configuration: Enabled explicit per-token loss calculation within the Megatron model's argument configuration, ensuring that loss is computed on a per-token basis.
• Loss Aggregation Logic: Refactored the forward_step_func to correctly accumulate loss by retrieving num_tokens from the loss dictionary and updated the final loss division to use this token count for accurate normalization.
• Loss Reduction Strategy: Modified the reduce_loss method in the Megatron strategy to return the num_tokens alongside the loss, facilitating proper aggregation and normalization across distributed training setups.

🧠 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

• src/twinkle/model/megatron/args.py
  • Enabled calculate_per_token_loss in the base model configuration.
• src/twinkle/model/megatron/megatron.py
  • Modified forward_step_func to accumulate num_tokens from the loss dictionary.
  • Updated the loss division logic to use count or 1 for robustness.
  • Added a ValueError for unexpected torch.Tensor loss types.
• src/twinkle/model/megatron/strategy/megatron.py
  • Adjusted the reduce_loss method to return local_loss, count, and a dictionary including num_tokens.
  • Converted local_count to torch.int64 before clamping.

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 aims to fix the loss calculation in Megatron models by averaging over tokens instead of microbatches. The changes in strategy/megatron.py and args.py correctly support this. However, the implementation in megatron.py introduces a critical bug. It incorrectly processes the return value from the forward pass, which will lead to an incorrect loss of 0 during training and will break the forward_only path. I've provided a detailed comment with a suggested fix.