Support new DeepGEMM input format in silu_and_mul_masked_post_quant_fwd #7153
Conversation
There was a problem hiding this comment.
Summary of Changes
Hello @fzyzcjy, 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 integrates a specific scaling method required for a "DeepGEMM format" into the existing post-quantization kernel for the silu_and_mul operation. This is achieved by adding a new flag to the kernel and its wrapper function, enabling conditional application of a power-of-2 based scaling logic.
Highlights
- Kernel Modification: Added a new boolean constant parameter
SCALE_UE8M0to the_silu_and_mul_post_quant_kernelfunction signature. - Conditional Scaling Logic: Implemented a conditional block within the kernel to apply a different scaling method (
tl.exp2(tl.ceil(tl.log2(tl.abs(output_s))))) when theSCALE_UE8M0flag is true. This likely aligns the scaling with requirements for a specific DeepGEMM format. - Wrapper Function Update: Updated the
silu_and_mul_masked_post_quant_fwdfunction to accept a new boolean parameterscale_ue8m0(defaulting to False) and pass it down to the kernel call.
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 is currently in preview and 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 to provide feedback.
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
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configureGemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces a new quantization scaling option, SCALE_UE8M0, to the _silu_and_mul_post_quant_kernel Triton kernel and its corresponding Python wrapper. When enabled, this option rounds the quantization scale (output_s) up to the nearest power of two. This change appears to be aimed at supporting a specific "DeepGEMM format" which likely benefits from power-of-two scales.
The changes are well-localized and the core logic for the new scaling method is clear. My review includes a suggestion for a minor optimization by potentially removing a redundant tl.abs call in the Triton kernel and a recommendation to update the function's docstring to reflect the new parameter.
It would also be beneficial for the author to fill out the "Motivation" and "Modifications" sections in the pull request description to provide more context for these changes.
| @@ -319,6 +320,8 @@ def _silu_and_mul_post_quant_kernel( | |||
| gate_up = up * gate | |||
| _absmax = tl.maximum(tl.max(tl.abs(gate_up)), 1e-10) | |||
| output_s = _absmax / fp8_max | |||
| if SCALE_UE8M0: | |||
| output_s = tl.exp2(tl.ceil(tl.log2(tl.abs(output_s)))) | |||
There was a problem hiding this comment.
The output_s variable is calculated as _absmax / fp8_max. Given that _absmax = tl.maximum(tl.max(tl.abs(gate_up)), 1e-10) is always positive (>= 1e-10) and fp8_max is also positive, output_s itself should always be positive.
Therefore, the tl.abs(output_s) call here might be redundant. Removing it could offer a micro-optimization and slightly improve code clarity if output_s is indeed guaranteed to be positive.
| output_s = tl.exp2(tl.ceil(tl.log2(tl.abs(output_s)))) | |
| output_s = tl.exp2(tl.ceil(tl.log2(output_s))) |
Motivation
Modifications
Checklist