[Tracing] Dispatch after tracing

[kylesayrs]

Purpose

• Avoid complications from dispatch+tracing by dispatching after tracing
• Prepare for TorchOffloader to replace accelerate

Background

Originally, we dispatched the model for sequential onloading before tracing the model. This was so that any onloading behavior would also be reflected in the trace. However, accelerate's onloading implementation is not traceable anyways (we skip it via wrapping sequential targets), so clearly dispatching has no interaction with tracing.

There is a patch required to get TorchOffloader to work with tracing, but it's easier to avoid the problem entirely and trace the model on cpu (no actual cpu calculations are performed during tracing).

All of the tracing tests were not dispatching the model anyways.

Changes

• Dispatch after tracing

Testing

• Tested the following models e2e:
  • Meta-Llama-3-8B-Instruct
• Traced Llama-3.3-70B-Instruct in 0.07 seconds (used to take 0.33 seconds)
• Tracing tests pass

[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 @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 refactors the model preparation pipeline by changing the sequence of dispatching and tracing operations. The primary goal is to resolve complexities that arose from dispatching the model before tracing, especially since the onloading behavior is not traceable. By dispatching the model after tracing, the change simplifies the process, improves tracing speed, and ensures that tracing occurs on the CPU without actual CPU computations, leading to a more efficient and robust workflow.

Highlights

• Dispatching and Tracing Order: The order of model dispatching and tracing has been reversed. Previously, dispatching occurred before tracing, but it has been moved to after tracing to avoid complications and align with the fact that accelerate's onloading is not traceable.
• Performance Improvement: This reordering significantly improves tracing performance. For example, tracing Llama-3.3-70B-Instruct now completes in 0.07 seconds, a substantial reduction from the previous 0.33 seconds.

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 refactors the sequential calibration pipeline to perform model dispatching after tracing, rather than before. The reasoning provided is sound: since accelerate's on-loading is not traceable, tracing the model on CPU and then dispatching it simplifies the process and avoids potential complications. This change also yields a significant performance improvement in tracing time, as noted in the description. The code modification is a straightforward reordering of operations and appears correct. The fact that existing tests already covered this execution path increases confidence in the change's correctness and stability. Overall, this is a good improvement for both correctness and performance.