[Utils] Offloaded cache size

[kylesayrs]

Purpose

• Make debugging offloaded cache memory easier

[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!

I've added a new utility to help debug memory usage of the offloaded cache. This new functionality provides a way to inspect the memory footprint of cached values, broken down by the device they reside on, making it easier to identify and manage memory consumption within the cache.

Highlights

• New size() method: Introduced a size() method to the Cache class in src/llmcompressor/pipelines/cache.py to calculate the memory used by cached values.
• Device-specific memory tracking: The size() method reports memory usage per torch.device, allowing for granular insights into where cached data resides.
• Recursive memory calculation: Implemented a recursive helper function _size_helper to accurately sum up memory for complex nested data structures like torch.Tensors, dataclasses, tuples, and dictionaries within the cache.

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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

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 issue 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 or fill out our survey to provide feedback.

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 a size() method to IntermediatesCache to calculate the memory usage of the cache, which is a helpful utility for debugging. The current recursive implementation has a couple of issues: it can fail with a RecursionError on deeply nested data, and it doesn't correctly calculate the size of lists. I've provided a suggestion for an iterative implementation that is more robust and correctly handles lists, which should improve both correctness and reliability.

[gemini-code-assist]

The implementation of size() has a few issues that could be improved for robustness and correctness:

1. Recursion Depth: The recursive _size_helper function can lead to a RecursionError if the cached data structures are deeply nested. An iterative approach using a queue would be more robust.
2. Incorrect list size: The size of list objects is not calculated correctly. The current implementation falls through to the else block, where sys.getsizeof() is called on the list object itself, not its contents. This will significantly underestimate memory usage for lists. Lists should be traversed like tuples.
3. Minor issues:
   • The type hint for _size_helper's return value is int, but it doesn't return anything. It should be None.
   • defaultdict(lambda: 0) can be simplified to defaultdict(int).
   • sys.getsizeof(value, 0) is a bit unusual; sys.getsizeof(value) is more idiomatic.

Here is a suggested iterative implementation that addresses these points:

    def size(self) -> Dict[torch.device, int]:
        """
        Returns the memory used by cached values, keyed by device, in bytes

        :return: dictionary mapping torch device to number of bytes in cache
        """
        sizes = defaultdict(int)
        queue = []
        for intermediates in self.batch_intermediates:
            queue.extend(intermediates.values())

        while queue:
            intermediate = queue.pop()
            value = intermediate.value

            if isinstance(value, torch.Tensor):
                sizes[value.device] += value.nbytes
            elif is_dataclass(value):
                for field in fields(value):
                    queue.append(getattr(value, field.name))
            elif isinstance(value, (list, tuple)):
                queue.extend(value)
            elif isinstance(value, dict):
                queue.extend(value.values())
            else:
                sizes[torch.device("cpu")] += sys.getsizeof(value)

        return dict(sizes)

[kylesayrs]

We've never encountered a list so far, but seems valid to add it

[brian-dellabetta]

looks good outside of gemini's recommendation