[2/n]decouple quantization implementation from vLLM dependency#8112
[2/n]decouple quantization implementation from vLLM dependency#8112zhyncs merged 21 commits intosgl-project:mainfrom
Conversation
Co-authored-by: walker-ai <yiyun.wyt@antgroup.com> Co-authored-by: yinfan98 <1106310035@qq.com>
There was a problem hiding this comment.
Summary of Changes
Hello @AniZpZ, 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 enhances the project's quantization capabilities by decoupling existing Marlin implementations from a specific Mixture of Experts dependency and introducing new, generalized GPTQ and Marlin GEMM kernels. This refactoring broadens the applicability of these optimized quantization techniques across the codebase, improving efficiency for various low-bit inference scenarios, including new support for FP4 quantization.
Highlights
- Quantization Decoupling: Marlin-based quantization implementations (GPTQ, AWQ repack) are moved from a MoE-specific path (
csrc/moe/marlin_moe_wna16) to a more general GEMM path (csrc/gemm/marlin), indicating a broader applicability beyond Mixture of Experts. - New GPTQ Kernels: Introduces comprehensive GPTQ GEMM kernels (2-bit, 3-bit, 4-bit, 8-bit) adapted from
exllamav2andGPTQ-for-LLaMa, including weight reconstruction and shuffling functionalities. - Marlin GEMM Integration: Integrates the core Marlin GEMM implementation, enabling highly optimized quantized matrix multiplication with support for various quantization types (INT4, INT8, FP4, FP8) and configurations (group size, act-order).
- FP4 Quantization Support: Adds support for FP4 (E2M1) floating-point quantization, expanding the range of supported low-bit quantization schemes.
- CUDA Utilities: Introduces new CUDA utility files for atomic operations on half types, matrix views for quantized data, and specialized quantization/dequantization (QDQ) functions for different bit-widths.
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 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
The pull request decouples the quantization implementation from the vLLM dependency. The changes involve adding new CUDA kernels for GPTQ and AWQ, refactoring existing code, and updating CMakeLists.txt. The code introduces new matrix view classes and implements dequantization and matrix multiplication kernels for different bit widths (2, 3, 4, and 8 bits). Additionally, it includes functions for shuffling and reconstructing quantized weights. The changes aim to provide a more modular and efficient implementation of quantization techniques.
|
@AniZpZ @Hongbosherlock please fix the conflicts thanks |
…roject#8112) Co-authored-by: walker-ai <yiyun.wyt@antgroup.com> Co-authored-by: leoneo <1320612015@qq.com>
…roject#8112) Co-authored-by: walker-ai <yiyun.wyt@antgroup.com> Co-authored-by: leoneo <1320612015@qq.com>
Motivation
The primary goal of this change is to enhance the consistency and stability of SGLang's quantization features. By decoupling the quantization implementation from its vLLM dependency, we aim to make the module easier to maintain and more portable.
Full realization of this goal will involve several subsequent PRs; this particular PR addresses the marlin kernel issues.
Modifications
Checklist