per-token scaling for fp8 linear

[akaitsuki-ii]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello @akaitsuki-ii, 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 enhances the fp8_linear function by introducing a per-token scaling mechanism for the input. This change improves the numerical stability and precision of FP8 linear operations by dynamically adjusting the scaling factor based on the input data's characteristics. The input is now pre-scaled before conversion to FP8, and the overall implementation of the scaled matrix multiplication has been simplified for consistency.

Highlights

• Dynamic Per-Token Scaling: The fp8_linear function now dynamically computes scale_a for each token (or row after reshaping) based on the maximum absolute value of the input. This replaces the previous static scale_a values, enabling more granular and precise scaling tailored to the input data's range.
• Input Pre-Scaling for Numerical Stability: The input tensor is now explicitly scaled down by the newly calculated scale_a (plus a small epsilon for stability) before being converted to the FP8 data type. This pre-scaling step is crucial for preventing overflow and maintaining numerical stability during the FP8 computation.
• Simplified _scaled_mm Call: The logic within the fp8_linear function has been streamlined by removing the conditional block that handled different input shapes. The input is now consistently reshaped at the beginning, allowing for a single, unified call to torch._scaled_mm, simplifying the code and ensuring consistent behavior.

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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

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 issue 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 per-token scaling for FP8 linear operations, which is a good improvement for quantization accuracy. The code is also cleaner due to the removal of conditional logic for input tensor shapes. I have provided a couple of suggestions for minor improvements: one to simplify the creation of the scale_b tensor, and another to remove a redundant operation for numerical stability, which will make the code cleaner and slightly more efficient.

[xuyxu]

LGTM