Support wInt4aFp8 for moe

[Wangzheee]

SUMMARY:
Support wInt4aFp8 for moe

with compressed-tensors PR: https://github.com/vllm-project/compressed-tensors/pull/518
with sglang PR: https://github.com/sgl-project/sglang/pull/11701

[github-actions]

👋 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]

Summary of Changes

Hello @Wangzheee, 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 support for a new 4-bit integer weight quantization scheme, specifically targeting Mixture of Experts (MoE) models. The changes involve updating the core logic to recognize this new format and ensuring that the official documentation reflects its inclusion and usage.

Highlights

• New Quantization Format Support: This pull request introduces support for a new 4-bit integer weight quantization format, specifically 'W4A8 - int', which is likely intended for Mixture of Experts (MoE) models.
• Documentation Update: The docs/guides/compression_formats.md guide has been updated to include the newly supported 'W4A8 - int' format, detailing its compression method as int4_quantized.
• Compression Format Logic Enhancement: The _get_quant_compression_format function in src/llmcompressor/transformers/compression/quantization_format.py has been modified to correctly identify and return CompressionFormat.int4_quantized when processing 4-bit integer weights with a TENSOR_GROUP strategy.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

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 pull request 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. 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

1. 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. ↩

[gemini-code-assist]

Code Review

This pull request adds support for the W4A8-int quantization format. The changes look good, but I have a few suggestions to improve them. A new test case should be added to validate the new logic path for inferring this quantization format, especially since the TEST PLAN in the description is a placeholder. Additionally, a code comment on line 46 of src/llmcompressor/transformers/compression/quantization_format.py has become outdated and should be updated. Finally, there's a minor point about using is for string comparison, for which I've left an inline comment with a suggestion.

[gemini-code-assist]

For comparing string values, it's safer and more conventional to use the == operator instead of is. The is operator checks for object identity, which might work for interned strings but is not guaranteed across different Python implementations or versions. Using == ensures the comparison is always done by value.¹

Suggested change

	if weight_args.type == QuantizationType.INT.value and weight_args.num_bits == 4 and weight_args.strategy is QuantizationStrategy.TENSOR_GROUP.value:
	if weight_args.type == QuantizationType.INT.value and weight_args.num_bits == 4 and weight_args.strategy == QuantizationStrategy.TENSOR_GROUP.value:

Rules References

Footnotes

1. Use == for value equality and is for identity equality. For comparing string literals, == is preferred for robustness as string interning is an implementation detail. ↩

[kylesayrs]

Hi @Wangzheee!

Could you provide more information about the W4AFP8 scheme? Are using this just for the purpose of experimentation? Could you provide example accuracy recovery metrics? I think we should avoid providing example schemes if there are no vllm kernels to support it, maybe the example should go in the experimental folder?

[Wangzheee]

Hi @Wangzheee!

Could you provide more information about the W4AFP8 scheme? Are using this just for the purpose of experimentation? Could you provide example accuracy recovery metrics? I think we should avoid providing example schemes if there are no vllm kernels to support it, maybe the example should go in the experimental folder?

Hi～
We are currently organizing the experimental results regarding performance and accuracy, and will post the experimental results later. We are verifying its functionality on SGLang, so may be more appropriate to move it to the experimental folder. We will add the kernels to vLLM after the verification is completed.