[Quantization] Support mixed-precision compression #1713
Conversation
|
👋 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 @dsikka, 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 PR introduces support for mixed-precision compression by enabling per-module quantization format inference. It adds new logic to dynamically determine the optimal compression format for individual quantized submodules within a model, based on their specific quantization schemes and sparsity structures. This allows for more granular and efficient compression strategies.
Highlights
- Dynamic Format Inference: I've implemented a new helper function, _get_quant_method, which intelligently determines the appropriate CompressionFormat (e.g., nvfp4_pack_quantized, marlin_24, pack_quantized) for a given module based on its input/weight quantization arguments and sparsity structure.
- Per-Module Quantization Application: I've added infer_per_module_quantization, a core function that iterates through all quantized submodules in a model, applies the dynamic format inference from _get_quant_method, and updates each submodule's quantization_scheme with its specific compression format.
- Integration into Model Compression Workflow: The get_model_compressor utility now leverages infer_per_module_quantization to assign specific compression formats to individual modules, moving towards a more flexible and mixed-precision compression approach.
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 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 or fill out our survey 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 introduces support for mixed-precision quantization by allowing different compression formats for different modules. The overall approach is sound, but there are a few critical issues in the implementation that need to be addressed. Specifically, there's a logical error in how the quantization formats are inferred and used to configure the model compressor, which could lead to incorrect behavior or runtime errors. Additionally, there are some minor issues like a leftover debug print statement, a missing docstring, and a potential crash when handling sparsity structures. Addressing these points will improve the robustness and maintainability of the new functionality.
src/llmcompressor/transformers/compression/quantization_format.py
Outdated
Show resolved
Hide resolved
src/llmcompressor/transformers/compression/quantization_format.py
Outdated
Show resolved
Hide resolved
| sparsity_structure=None | ||
| if sparsity_config is None | ||
| else sparsity_config.sparsity_structure, | ||
| quantization_format: Optional[CompressionFormat] = ( |
There was a problem hiding this comment.
Type hint does not match return type of function
|
|
||
|
|
||
| def infer_quantization_format( | ||
| def _get_quant_method( |
There was a problem hiding this comment.
This function is called _get_quant_method, but returns a format
| return quant_info_weight, quant_info_inputs | ||
| if weight_scheme is None: | ||
| continue # no weight quant - nothing to compress | ||
| compression_format = _get_quant_method( |
There was a problem hiding this comment.
Why are we not inferring from scheme.format? Doesn't this override anything the user might have passed?
There was a problem hiding this comment.
Rather than waiting until the end to do inference, why not infer a format at the source?
class QuantizationScheme:
format: Optional[CompressionFormat] = None
@validate_field("format")
def validate_format(self, value):
if self.weights and self.input_activations ...
inferred_format = ...
if value is not None and value != inferred_format:
logger.warn_once("Consider using inferred scheme")
return value or inferred_formatThere was a problem hiding this comment.
LLM Compressor interprets the compression format at the time of the compression based on the quant args. This is our current lifecycle. We can still override using a global format.
We can support overriding on a per module basis in a follow-up. The goal of this PR isn't to support per-module format overriding. The goal is to support the most common pathway to start, which is where we determine the format during compresssion time.
SUMMARY:
infer_quantization_formatto beinfer_per_module_quantization_formatsuch that instead of returning a global format, a per module format is assigned to each module to be used during compression time. All unique compression formats are returnedTesting: