Granite4 FP8 Block Quantization

[krishnateja95]

SUMMARY:
"Recipe for FP8 Block quantization of Granite4-small model (https://huggingface.co/ibm-granite/granite-4.0-h-small)"

[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 @krishnateja95, 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 enables FP8 block quantization for the Granite4-small model by introducing a new example and essential architectural adaptations. It provides a mechanism to convert the model's Mixture-of-Experts (MoE) layers into a format amenable to quantization and then repackages the resulting weights for optimized storage and compatibility with high-performance inference frameworks like vLLM. This work streamlines the process of deploying quantized Granite models.

Highlights

• FP8 Block Quantization Example: A new example script has been added to demonstrate FP8 block quantization for the ibm-granite/granite-4.0-h-small model, showcasing the full quantization workflow from model loading to saving and sample generation.
• MoE Expert Conversion: Introduced a utility function replace_granite_moe_with_linear that converts GraniteMoeHybridParallelExperts modules into individual nn.Linear layers, which is a necessary step to enable proper quantization of these complex MoE architectures.
• 3D Expert Packing for vLLM Compatibility: A new pack_3d_experts function has been implemented to transform the model's expert weights from a per-expert storage format to a stacked 3D tensor representation. This restructuring is crucial for compatibility with inference engines like vLLM and for more efficient storage.
• Dynamic Config Update: The pack_3d_experts function also includes logic to automatically update the config.json file, specifically renaming mamba.in_proj and mamba.out_proj entries to mixer.in_proj and mixer.out_proj within the quantization_config.ignore list, ensuring consistency for quantized models.

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 introduces FP8 block quantization for the Granite4-small model, including a new example script and utility functions. The core logic involves replacing MoE expert layers with standard linear layers for quantization and then packing them back into a 3D tensor format for compatibility. My review focuses on improving code quality, robustness, and fixing a bug that prevents the example from running. I've suggested removing unused imports, renaming a function for consistency and correctness, using isinstance for more robust type checking, and refactoring a large function for better maintainability. Overall, the changes are a good addition, and with these adjustments, the code will be more robust and easier to maintain.

[dsikka]

Couple of key issues:

1. We should be using the moe_context: https://github.com/vllm-project/llm-compressor/blob/main/src/llmcompressor/modeling/moe_context.py. We have many examples you can use that apply the context
2. All saving logic should be applied by compressed-tensors. Seems like you're doing a lot of custom logic

[dsikka]

Please use this context: https://github.com/vllm-project/llm-compressor/blob/main/src/llmcompressor/modeling/moe_context.py

[krishnateja95]

@dsikka This saving logic is specific to granite4 small model to ensure the fp8 block quantized model is compatible with vLLM.