support AutoRound NVFP4 scheme #2169
Conversation
Signed-off-by: He, Xin3 <[email protected]>
|
👋 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. |
Summary of ChangesHello @xin3he, 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 integrates the NVFP4 quantization scheme into the Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Using Gemini Code AssistThe 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 by creating a comment using either
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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. 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
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request extends AutoRound support to include the NVFP4 scheme. The changes are well-structured, introducing support for preset schemes from auto-round and improving the integration by correctly passing tokenizer and processor information. The quantization parameter mapping has been significantly improved and refactored into a dedicated method, which now handles more parameters and includes specific logic for NVFP4. Additionally, there are good improvements to the calibration and input capturing logic. A test case for the new NVFP4 scheme is also included.
I've found one critical issue regarding the handling of None quantization parameters that could lead to non-deterministic behavior, for which I've left a specific comment with a suggested fix. Otherwise, the changes look solid and improve the AutoRoundModifier's capabilities and robustness.
| if ar_value is None: | ||
| ar_value = torch.empty(1) |
There was a problem hiding this comment.
The use of torch.empty(1) to handle None quantization parameters is unsafe. torch.empty() returns a tensor with uninitialized data, which can lead to non-deterministic behavior and incorrect results when used as a quantization parameter.
A None value should be handled based on the specific parameter's meaning. For instance, a None zero-point (zp) typically indicates symmetric quantization and should default to 0.0. For other parameters like scales, if None signifies they are not applicable for the current quantization scheme, they should be skipped.
if ar_value is None:
if ar_param_name == "zp":
# For symmetric quantization, zero point is 0
ar_value = 0.0
else:
# If other parameters are None, it likely means they are not used
# for the current quantization scheme, so we can skip them.
continueSigned-off-by: He, Xin3 <[email protected]>
Signed-off-by: He, Xin3 <[email protected]>
| model_args.config_name if model_args.config_name else model_path, | ||
| cache_dir=None, | ||
| revision=model_args.model_revision, | ||
| use_auth_token=True if model_args.use_auth_token else None, |
| ) | ||
| }, | ||
| ) | ||
| recipe_modifier_nvfp4 = AutoRoundModifier( |
There was a problem hiding this comment.
how about adding an example of NVFP4 model quantization, besides adding NVFP4 test?
|
Can vLLM load this? |
3 participants
SUMMARY:
"This PR aims to extend the support of AutoRound. supported scheme are extended from
'W4A16toW4A16, NVFP4"TEST PLAN:
"The UT of AutoRound is extended with new NVFP4 modifier recipe"