Implement propagate_error argument

[kylesayrs]

Changes

• Add propagate_error argument to sequential pipeline.
  • When True, the sequential pipeline propagates quantized outputs (outputs after module optimization/GPTQ apply)
  • When False, the sequential pipeline propagates unquantized outputs (full precision)
  • GPTAQ authors report slightly better accuracy recovery when propagate_error=False. This aligns with the preliminary testing I've done myself, but more work will be required to make this a default argument.

[gemini-code-assist]

Summary of Changes

Hello @kylesayrs, 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 enhances the sequential pipeline by introducing a propagate_error argument, allowing fine-grained control over how intermediate layer outputs are handled during the quantization calibration process. This new option enables users to choose between propagating quantized or unquantized outputs, offering a mechanism to potentially improve model accuracy recovery, aligning with recent findings in quantization research.

Highlights

• New propagate_error Argument: Introduced a new propagate_error boolean argument to the sequential pipeline configuration, defaulting to True.
• Output Propagation Control: This argument controls whether quantized (when True) or unquantized (when False) layer outputs are propagated as inputs to subsequent layers during calibration.
• Accuracy Recovery: Preliminary testing and research (GPTAQ) suggest that setting propagate_error to False can lead to better accuracy recovery by propagating unquantized outputs.

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 a propagate_error argument to the sequential pipeline, allowing control over whether to propagate quantized or unquantized outputs between subgraphs. The implementation correctly reflects this new option. My review focuses on improving the efficiency of the pipeline by avoiding redundant computations. I've identified two places where an efficiency check for the last subgraph was removed and should be restored. I also pointed out an opportunity to skip an entire processing loop when it's not needed, which would significantly improve performance in that scenario.

[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.

[brian-dellabetta]

nice! nit -- batch_idx is a bit more readable than b_idx

[kylesayrs]

@brian-dellabetta I agree, but I think b_idx is more readable than putting the loop on two separate lines 🙁