Minor speed up block_quant_dequant #6814
Conversation
There was a problem hiding this comment.
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!
Summary of Changes
Hello! gemini-code-assist here, providing a summary of this pull request to help everyone get up to speed quickly.
This PR aims to significantly speed up the block_quant_dequant function within the FP8 quantization utilities. The primary motivation, based on the title, is performance optimization for this specific operation. The changes involve replacing a manual, loop-based implementation of block dequantization with a more efficient tensor-based approach leveraging the einops library.
Highlights
- Performance Optimization: The core change is a rewrite of the
block_quant_dequantfunction to replace nested Python loops and manual tiling with a single, potentially faster, tensor operation usingeinops.repeat. - Code Simplification: The implementation of
block_quant_dequantis simplified by removing the explicit loop structure and tile handling, resulting in more concise code. - New Dependency: The
einopslibrary is added as a dependency to facilitate the tensor repetition needed for the new dequantization logic. - Input Validation: Assertions were added to
block_quant_dequantto ensure that the input tensor dimensions are perfectly divisible by the specified block size, which is required for the neweinops-based approach.
Changelog
- python/sglang/srt/layers/quantization/fp8_utils.py
- Removed unused imports
osandcurses.flash. - Added import for the
einopslibrary. - Rewrote the
block_quant_dequantfunction to useeinops.repeatand tensor multiplication for dequantization, replacing the previous loop-based implementation. - Added assertions to
block_quant_dequantto check for divisibility of input dimensions by block size.
- Removed unused imports
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.
Loops run slow and long,
Tensor math makes code strong,
Speed up, dequantize!
Footnotes
-
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. ↩
There was a problem hiding this comment.
Code Review
This pull request aims to speed up the block_quant_dequant function by replacing a loop-based implementation with a vectorized approach using einops. This is a good optimization strategy and should lead to performance improvements.
The main point to clarify is the change in input validation: the new implementation asserts that input tensor dimensions must be perfectly divisible by the block sizes, whereas the previous implementation appeared to handle non-perfectly divisible inputs. Please confirm if this change in behavior is intentional. If so, the function's docstring should be updated to reflect this new precondition.
Summary of Findings
- Change in Input Validation for
block_quant_dequant: The functionblock_quant_dequantnow explicitly asserts that the input tensor dimensionsnandkmust be perfectly divisible byblock_nandblock_krespectively. The previous implementation appeared to handle non-perfectly divisible inputs. This change needs clarification to ensure it's intentional and to update documentation accordingly if it is. - Unused Imports Removed: The imports
osandcurses.flashwere removed. This is a good cleanup as they were unused in the context of the changed function. (Severity: low, not commented due to review settings)
Merge Readiness
This pull request introduces a good performance optimization by vectorizing the block_quant_dequant function. However, there's a significant change in input validation that needs clarification. If the new assertions for input divisibility are intended and all callers adhere to this, the change is acceptable, but the docstring should be updated. If the previous behavior of handling non-divisible inputs needs to be maintained, the current changes would be breaking.
I recommend addressing the point about input validation before merging. I am unable to approve pull requests, so please ensure other reviewers approve this change after the clarification.
| assert n % block_n == 0 | ||
| assert k % block_k == 0 |
There was a problem hiding this comment.
The new implementation introduces assertions n % block_n == 0 and k % block_k == 0. This is a stricter requirement compared to the previous version, which used min() in slicing (e.g., min((j + 1) * block_n, n)) and could handle cases where x_q_block dimensions were not perfectly divisible by block_size (as long as x_s had the correct number of scales for the tiles).
Could you clarify if this change in behavior is intentional?
- If
x_q_blockis always guaranteed to have dimensions divisible byblock_sizein all call sites, then these assertions make an implicit precondition explicit, which is good. In this case, it would be beneficial to update the function's docstring (lines 374-378) to mention these new preconditions forx_q_blockandx_sshapes. - If the function was previously expected to handle non-divisible shapes, this change might break existing callers or alter the function's contract. If the old behavior needs to be preserved, the
einopsapproach might require padding or a different vectorization strategy.
This change is important as it affects the function's input contract.
|
NO einops for multi-platform compatibility XD |
|
@Alcanderian Ah, that one does not support AMD? I thought it is a wrapper around pytorch... |
|
einops removed |
Motivation
Modifications
Checklist