Support DeepSeekV3-style block FP8 quantization #1607
Conversation
Signed-off-by: mgoin <michael@neuralmagic.com>
|
👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review. Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed. |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @mgoin, 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 comprehensive support for DeepSeekV3-style block FP8 quantization, a technique designed to further compress large language models for more efficient inference. The changes encompass the fundamental implementation of block-wise quantization, robust handling of quantization parameters, updated documentation, and a practical example to guide users in applying this new scheme.
Highlights
- New Quantization Scheme: Introduced support for
W8A8-FP8_BLOCKquantization. This scheme applies block-wise FP8 quantization to weights (typically in 128x128 tiles) and dynamic per-token-group (128) FP8 quantization for activations. A key benefit is that it does not require a calibration dataset. - Block Quantization Implementation: The core logic for the
BLOCKquantization strategy has been implemented withinsrc/llmcompressor/observers/base.py. This involves calculating and storing individual scales and zero points for each defined block within a tensor, replacing a previousNotImplementedError. - Dynamic Parameter Handling: The
call_observerfunction insrc/llmcompressor/modifiers/quantization/calibration.pyhas been updated to correctly register and update scale and zero-point parameters. This change specifically addresses the varying shapes of these parameters when block-wise quantization is applied. - Example and Documentation: A new example script (
examples/quantization_w8a8_fp8/fp8_block_example.py) has been added to demonstrate how to apply theFP8_BLOCKscheme to a model. Corresponding documentation has been updated indocs/schemes.mdto describe this new quantization method. - Test Coverage: A new test case has been added in
tests/llmcompressor/modifiers/quantization/test_base.pyto ensure that the block quantization configuration, including theblock_structureparameter, is correctly parsed and resolved by theGPTQModifier.
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
This pull request adds support for DeepSeekV3-style block FP8 quantization, including the necessary observer logic, calibration handling for dynamic parameter shapes, a new example, and documentation. The changes are well-implemented. My feedback includes suggestions to improve clarity in the documentation, fix inaccuracies in the example script, and refactor a small piece of duplicated code for better maintainability.
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
There was a problem hiding this comment.
Thanks for adding this! We had some users requesting block support in
Couple comments for @kylesayrs , but otherwise LGTM
| module.register_parameter( | ||
| scale_name, torch.nn.Parameter(updated_scale.clone()) | ||
| ) |
There was a problem hiding this comment.
@kylesayrs do we need these (this and L141-143) to use register_offload_parameter instead?
| zp_name = f"{base_name}_zero_point" | ||
| if not hasattr(module, scale_name) or getattr(module, scale_name).shape != updated_scale.shape: | ||
| if hasattr(module, scale_name): | ||
| delattr(module, scale_name) |
There was a problem hiding this comment.
@kylesayrs do we need this and L140 to be delete_offload_parameter?
Signed-off-by: mgoin <michael@neuralmagic.com>
SUMMARY:
Blocked on CT support in neuralmagic/compressed-tensors#372
TEST PLAN: