divprune

[chengtao-lv]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello @chengtao-lv, 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 integrates a new token reduction method, "DivPrune," into the llmc library. This method is designed to efficiently prune visual tokens in multimodal models, particularly LLaVA, by selecting a diverse subset of image features. The aim is to reduce computational load and improve inference efficiency without sacrificing model performance.

Highlights

• New Token Reduction Method: Introduces DivPrune, a novel token reduction technique designed to prune visual tokens based on diversity, aiming to improve efficiency in multimodal models.
• DivPrune Algorithm Implementation: Implements the core divprune function, which selects a subset of visual feature vectors by maximizing diversity using pairwise cosine similarity. This includes a divprune_post_hook for applying the reduction to model inputs.
• LLaVA Model Integration: Integrates DivPrune specifically with the LLaVA model by hooking into its prepare_inputs_labels_for_multimodal method to apply the token reduction to image features before model processing.

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 is currently in preview and 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 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 introduces DivPrune, a new token reduction method based on selecting diverse visual tokens using cosine similarity. There are several areas for improvement regarding robustness, clarity, and adherence to best practices. Key issues include potential runtime errors from division by zero and incorrect function arguments, brittleness due to hardcoded batch and argument indices, and a local import. Addressing these points will significantly improve the code's quality and maintainability.

[gemini-code-assist]

The pruning_paras argument defaults to None but is accessed on line 65 without a check, which can raise a TypeError if the function is called without this argument. Removing the default value makes it a required argument, clarifying the function's dependencies.

Suggested change

	pruning_paras=None,
	pruning_paras,

[gemini-code-assist]

This line hardcodes the batch index [0], assuming a batch size of 1. This assumption is fragile and will cause errors if the function is used with a batch size greater than 1. Add an assertion to enforce a batch size of 1.

Suggested change

	visual_tokens = inputs_embeds[0][SYS_TOKEN_LEN: SYS_TOKEN_LEN + img_feature_len]
	assert inputs_embeds.shape[0] == 1, 'Batch size must be 1'

[gemini-code-assist]

The logic here hardcodes the batch index [0] for input_ids and attention_mask, which assumes a batch size of 1. This will fail for multi-batch inputs. Add an assertion at the beginning of the wrapper to enforce a batch size of 1.

                assert input_ids.shape[0] == 1, 'Batch size must be 1'
                assert attention_mask.shape[0] == 1, 'Batch size must be 1'

[gemini-code-assist]

The calculation of norm_matrix can result in a division by zero if any vector in the input matrix has a norm of zero, leading to NaN values. Adding a small epsilon to the denominator ensures numerical stability.

Suggested change

	norm_matrix = matrix / matrix.norm(dim=1, keepdim=True)
	norm_matrix = matrix / (matrix.norm(dim=1, keepdim=True) + 1e-8)

[gemini-code-assist]

The variable name SYS_TOKEN_LEN is misleading as it stores a start index, not a length. Renaming it to image_token_start_index would improve code clarity. Update its usages in this function as well.

Suggested change

	SYS_TOKEN_LEN = pruning_paras['image_token_start_index']
	image_token_start_index = pruning_paras['image_token_start_index']

[gemini-code-assist]

Accessing arguments by their position (e.g., args[0], args[2]) is brittle and can easily break if the signature of the hooked function prepare_inputs_labels_for_multimodal changes. Accessing arguments by name would be more robust.

[gemini-code-assist]

According to PEP 8, imports should be at the top of the file. This local import makes the code's dependencies less clear. Move it to the top of the file. If llava is an optional dependency, wrap the import in a try...except ImportError block at the top level.