multimodal chuncked prefill select needed multimodal objs.

[hiworldwzj]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello @hiworldwzj, 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 significantly enhances the efficiency of multimodal model prefill operations, especially in chunked scenarios, by introducing a specialized Triton kernel that intelligently selects only the multimodal objects whose tokens are relevant to the current processing chunk. Furthermore, it provides a new configuration option to perform visual model inference directly within the main process, eliminating the overhead of inter-process communication for multimodal data. To support this in-process mode, a robust image embedding cache manager has been implemented, ensuring efficient memory utilization and management of image embeddings.

Highlights

• Optimized Multimodal Prefill: Introduced a new Triton kernel (mark_multimodal_obj) to efficiently identify and mark only the necessary multimodal objects (images/audios) for processing during chunked prefill. This avoids unnecessary computations by skipping irrelevant multimodal tokens.
• In-Process Multimodal Inference Option: Added a new command-line argument (--disable_extra_process_for_multimodal) that allows visual model inference to occur directly within the main process, removing the need for a separate visual process and its associated inter-process communication overhead.
• Image Embedding Cache Management: Implemented a new ImageCacheManager to handle the caching and lifecycle of image embeddings. This manager uses both GPU and CPU caches (with LRU eviction for the CPU cache) to optimize memory usage, particularly when multimodal inference is performed in-process.
• Integrated Multimodal Object Selection: Integrated the new multimodal object marking logic and the in-process inference option across various core components, including InferStateInfo, pre-layer inference modules (Gemma3, Qwen-VL, InternVL), and the server's request handling, ensuring consistent behavior and efficiency gains.

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

The pull request introduces several changes to support multimodal chunked prefill, including marking multimodal objects, caching image embeddings, and managing resources. The changes look good overall, but there are a few issues that need to be addressed, including a potential out-of-bounds write in the triton kernel, an inefficiency in the qwen_vl prefill logic, and incorrect filtering of image UUIDs in the infer batch.