[megatron] support deepseek-v3.2

[Jintao-Huang]

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 improves the robustness of the MxQintDequantizer utility. It addresses a limitation where dequantization would fail if input matrix dimensions were not perfectly divisible by the block sizes. By introducing a fallback loop-based approach for such edge cases, the utility can now handle a wider range of tensor shapes, which is crucial for broader model compatibility and likely supports the integration of models like deepseek-v3.2 as indicated in the title.

Highlights

• Enhanced Dequantization Robustness: The MxQintDequantizer.convert method has been updated to gracefully handle cases where matrix dimensions are not perfectly divisible by the specified block sizes, preventing previous ValueError exceptions.
• Fallback Dequantization Logic: A new loop-based dequantization mechanism has been implemented as a fallback for matrices with dimensions that are not exact multiples of the block size, ensuring correct scaling in these scenarios.
• Dependency Update: The math module was imported to support ceiling calculations required for the new loop-based dequantization logic.

🧠 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

• swift/utils/dequantizer.py
  • The MxQintDequantizer.convert method was modified to no longer raise a ValueError when matrix dimensions are not perfectly divisible by block sizes.
  • A new if/else block was introduced in MxQintDequantizer.convert to provide a loop-based dequantization fallback for non-divisible dimensions.
  • The math module was imported to facilitate ceiling calculations within the new dequantization logic.

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 adds support for cases where matrix dimensions are not divisible by block sizes in Fp8Dequantizer by adding a fallback loop. My review focuses on improving this fallback mechanism for better performance and correctness, especially when handling tensors with batch dimensions.

[Jintao-Huang]

/gemini review

[gemini-code-assist]

Code Review

This pull request adds support for the deepseek-v3.2 model to the Megatron framework. The changes span documentation, model configuration, and patches to core logic for RoPE. My review identified a couple of typos in the model type string for deepseek-v3.2 which could lead to incorrect behavior. I also suggest an improvement in error handling to ensure configuration consistency by using a try...finally block. Overall, the changes are in the right direction to support the new model.

[Jintao-Huang]

/gemini review

[gemini-code-assist]

Code Review

This pull request adds support for the deepseek-v3.2 model series to Megatron. The changes are comprehensive, covering model registration, configuration updates, and documentation. A notable improvement is also made to the dequantizer to handle matrix sizes that are not divisible by the block size. Overall, the changes are well-implemented. I have a couple of minor suggestions to improve code quality and consistency.

[Jintao-Huang]

huggingface/transformers#44481