Re-quantize DeepSeek model weights to support DeepGEMM new input format #7156
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 introduces the necessary infrastructure and logic to re-quantize DeepSeek V2 model weights into a new UE8M0 FP8 format designed for compatibility with DeepGEMM. It includes utility functions, the core requantization process, and integration points within the DeepSeek V2 model loading, although the requantization step is currently disabled.
Highlights
- DeepGEMM Utility Functions: Added
alignandceil_divutility functions tosglang.math_utilswhich are copied from the DeepGEMM project. These are likely prerequisites for the new quantization format. - UE8M0 Requantization Logic: Implemented functions (
requant_weight_ue8m0_inplace,_requant_weight_ue8m0) infp8_utils.pyto dequantize existing weights and then re-quantize them into a new UE8M0 FP8 format compatible with DeepGEMM. This involves block-wise processing and scale transformation. - Block-wise FP8 Casting: Added helper functions (
per_block_cast_to_fp8,ceil_to_ue8m0) tofp8_utils.py, also copied from DeepGEMM, which handle the core logic of casting tensors totorch.float8_e4m3fnon a per-block basis, including padding and scale calculation. - DeepSeek V2 Integration: Added a new method
_weight_requant_ue8m0to theDeepseekV2Modelclass to apply the new UE8M0 requantization process to the weights of specific attention and MLP/MoE modules. - Future Activation: The call to the new
_weight_requant_ue8m0method inDeepseekV2Model.post_load_weightsis currently commented out (if False) with aTODO (pr-chain), indicating it will be enabled in a subsequent pull request.
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 functionality to re-quantize DeepSeek model weights, likely to align with a new input format required by DeepGEMM. It adds utility functions for math operations and FP8 quantization, and a new method within the DeepSeek V2 model for applying this re-quantization. The changes appear to be part of an ongoing effort, as indicated by a TODO comment and an incomplete checklist in the PR description. Key areas for attention include ensuring type safety through hints, clarifying hardcoded assumptions (like block sizes), and replacing magic numbers with named constants for better maintainability. Thorough testing of this new quantization logic will be crucial.
| @@ -1931,6 +1932,61 @@ def post_load_weights(self, is_nextn=False, weight_names=None): | |||
| self_attn.w_vc = bind_or_assign(self_attn.w_vc, w_vc.contiguous()) | |||
| self_attn.use_deep_gemm_bmm = True | |||
|
|
|||
| if False: # TODO (pr-chain) | |||
Motivation
Modifications
Checklist