recover cp sequence before loss

[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 significantly refines the handling of log-probabilities (logps) and context parallelism (CP) within the Megatron framework, specifically targeting loss calculation. It introduces a robust mechanism to reconstruct full sequences from CP-split tensors before loss computation, centralizes logps management, and simplifies the overall loss aggregation process. These changes aim to improve the accuracy and robustness of loss calculation in distributed training setups, alongside minor utility enhancements and code cleanup.

Highlights

• Centralized Log-Probability (logps) Handling: Refactored loss calculation across multiple files to prioritize pre-computed logps and introduced logps into ModelOutput for consistent processing, reducing redundant calculations.
• Context Parallel (CP) Sequence Reconstruction: Implemented a new utility _postprocess_tensor_cp to all-gather and reconstruct full sequences from CP-split tensors, ensuring correct loss computation and gradient flow in distributed environments.
• Simplified CP Loss Aggregation: Streamlined the gather_loss_for_cp method by removing explicit all-reduce logic for context parallelism, as sequence reconstruction now handles the necessary data aggregation earlier in the pipeline.
• Enhanced Mean Collection for Dictionaries: Updated the _collect_func to correctly compute the mean for dictionary values, handling both numeric and non-numeric types gracefully.
• Code Cleanup and Robustness: Removed redundant or temporary code related to input processing in the Megatron training cookbook and improved conditions for logps concatenation to prevent errors.

🧠 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 unused input processing lines related to _inputs and _temp.
• src/twinkle/infra/init.py
  • Extended the _collect_func to support mean calculation for dictionary values, handling numeric and non-numeric types.
• src/twinkle/loss/grpo.py
  • Modified the __call__ method to prioritize logps from outputs, falling back to calculating it from logits if logps is not pre-computed.
  • Removed the explicit deletion of logits.
• src/twinkle/loss/vocab_parallel_cross_entropy.py
  • Refactored VocabParallelCrossEntropyLoss to directly use pre-computed logps for loss calculation.
  • Removed the dependency on Megatron's tensor_parallel.vocab_parallel_cross_entropy.
  • Simplified the class docstring.
• src/twinkle/model/megatron/megatron.py
  • Added a new private method _postprocess_tensor_cp to all-gather and reconstruct full sequences from context parallel split tensors.
  • Updated post_loss_function to include logps in the ModelOutput.
  • Modified forward_step_func to apply _postprocess_tensor_cp to logps and labels when context parallelism is active.
  • Adjusted the condition for concatenating logps to explicitly check if logps is not None.
• src/twinkle/model/megatron/strategy/megatron.py
  • Simplified the gather_loss_for_cp method by removing explicit all-reduce logic for context parallelism, as sequence reconstruction is now handled elsewhere.
  • Ensured that the loss value returned in the dictionary is a detached clone.

Activity

• No specific activity (comments, reviews, or progress updates) was provided in the context for this pull request.

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 refactors the loss calculation logic for context parallelism (CP) by reconstructing the full sequence from CP-split tensors before computing the loss. This simplifies the loss aggregation strategy and seems to be a good improvement. The changes are consistent across the modified files. I've provided a couple of suggestions to improve code robustness and clarity.