Fp8 compressed

[danielhanchen]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello @danielhanchen, 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 significantly improves the integration and handling of FP8 compressed tensors within the vLLM utility. It introduces robust mechanisms for identifying different FP8 quantization schemes, correctly interpreting their weight scales, and configuring the appropriate FP8 linear layers for efficient model loading and operation. The changes aim to enhance compatibility and performance when working with FP8 quantized models.

Highlights

• FP8 Compressed Tensor Support: Enhanced handling of FP8 compressed tensors by introducing logic to detect CompressedTensors quantization methods and adjust weight scale naming conventions accordingly, specifically switching from .weight_scale_inv to .weight_scale.
• Dynamic FP8 Layer Configuration: Added explicit support for compressed-tensors quantization in the convert_vllm_to_huggingface utility, allowing dynamic configuration of FP8Linear layers with appropriate block sizes derived from the quantization configuration.
• Unified FP8 Scale Handling: Refactored the loading of FP8 weight scales to check for both weight_scale and weight_scale_inv and dynamically select between FbgemmFp8Linear (for row-quantized, ndim=1) and FP8Linear (for block-quantized, ndim=2) based on the scale's dimensionality.
• vLLM Compilation Config Adjustment: The compilation_config parameter has been commented out during vLLM loading, potentially streamlining the loading process or addressing a compatibility concern.

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 FP8 compressed tensors by detecting this new quantization method, handling its specific configuration for block size, and refactoring the layer creation logic. My review focuses on improving the robustness and readability of the new code. I've suggested refactoring a type check to be less brittle, simplifying configuration parsing, and using a more specific exception handler.

[gemini-code-assist]

The check for is_compressed_linear relies on the string representation of a type, which can be brittle. This can be refactored to check the class name directly and combine the variable assignment and the conditional check into one if statement for improved robustness and conciseness.

Suggested change

	is_compressed_linear = "CompressedTensors" in str(type(getattr(proj, 'quant_method', None)))
	if is_compressed_linear:
	quant_method = getattr(proj, 'quant_method', None)
	if quant_method is not None and "CompressedTensors" in quant_method.__class__.__name__:

[gemini-code-assist]

The logic for extracting block_size from the configuration can be simplified for better readability and maintainability. Using chained .get() calls with default values makes the code more concise and less prone to errors from None values, assuming config_groups is a dictionary.

Suggested change

	config_groups = quantization_config.get('config_groups', None)
	group_0 = config_groups.get(0, None) if config_groups else None
	weights = group_0.get('weight', None) if group_0 else None
	block_size = weights.get('block_size', block_size) if weights else block_size
	config_groups = quantization_config.get('config_groups', {})
	group_0 = config_groups.get(0, {})
	weights = group_0.get('weight', {})
	block_size = weights.get('block_size', block_size)

[gemini-code-assist]

Using a bare except: is generally discouraged as it can catch unexpected exceptions (like KeyboardInterrupt or SystemExit) and hide bugs. It's better to catch the specific ImportError that is expected in this case.

Suggested change

	except:
	except ImportError: