[bugfix] fix text_position_ids

[Jintao-Huang]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello @Jintao-Huang, 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 addresses a bug where the text_position_ids tensor was not correctly shaped, leading to dimension mismatch errors when combined with other position IDs in the GLM and Qwen model templates. The changes ensure proper tensor shape alignment, enhancing the robustness of position ID generation.

Highlights

• Fixing text_position_ids shape: Corrected the shape of text_position_ids in the _get_position_ids function for GLM and Qwen templates. Previously, text_position_ids was a 1D tensor, which caused dimension mismatch issues when concatenated with other position_ids that might have more dimensions. The fix uses .expand() to ensure text_position_ids has compatible dimensions.
• Tensor dimension consistency: Ensured that the text_position_ids tensor is properly expanded to match the dimensions of other position_ids before concatenation, resolving a potential runtime error related to incompatible tensor shapes.

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. 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 fixes a bug in the calculation of text_position_ids across multiple templates (GLM4_5VTemplate, Qwen2VLTemplate, Qwen2_5OmniTemplate). The previous implementation did not correctly handle batched inputs (batch size > 1), which is now resolved by using expand() to match the batch dimension of the multimodal position IDs.

My review focuses on improving code maintainability by addressing duplicated logic and enhancing robustness by explicitly setting the device for new tensor creation. I've suggested refactoring the repeated code into a shared utility and adding the device argument to torch.arange in all occurrences.

[gemini-code-assist]

While this change correctly fixes the issue for batch processing, this logic is duplicated in other templates (e.g., Qwen2VLTemplate, Qwen2_5OmniTemplate). Consider refactoring this into a shared utility function to improve maintainability.

Additionally, it's safer to specify the device when creating a new tensor to prevent potential mismatches.

Suggested change

	text_position_ids = torch.arange(inputs['input_ids'].shape[-1]).expand(1, *position_ids.shape[1:])
	return torch.concat([text_position_ids, position_ids], dim=0)
	text_position_ids = torch.arange(inputs['input_ids'].shape[-1], device=inputs['input_ids'].device).expand(1, *position_ids.shape[1:])
	return torch.concat([text_position_ids, position_ids], dim=0)

[gemini-code-assist]

This logic is duplicated across multiple templates. To improve maintainability, it would be beneficial to extract it into a shared helper function. See also GLM4_5VTemplate and Qwen2_5OmniTemplate.

Also, explicitly setting the tensor device is a good practice to avoid potential device mismatches.

Suggested change

	text_position_ids = torch.arange(inputs['input_ids'].shape[-1]).expand(1, *position_ids.shape[1:])
	return torch.concat([text_position_ids, position_ids], dim=0)
	text_position_ids = torch.arange(inputs['input_ids'].shape[-1], device=inputs['input_ids'].device).expand(1, *position_ids.shape[1:])
	return torch.concat([text_position_ids, position_ids], dim=0)

[gemini-code-assist]

This is another instance of duplicated logic for handling text_position_ids. Refactoring into a common utility would be ideal for maintainability.

It's also recommended to specify the device for torch.arange and use the local input_ids variable for consistency.

Suggested change

	text_position_ids = torch.arange(inputs['input_ids'].shape[-1]).expand(1, *position_ids.shape[1:])
	return torch.concat([text_position_ids, position_ids], dim=0)
	text_position_ids = torch.arange(input_ids.shape[-1], device=input_ids.device).expand(1, *position_ids.shape[1:])
	return torch.concat([text_position_ids, position_ids], dim=0)