[Frontend] Support configurable mm placeholder strings & flexible video sampling policies via CLI flags.

[huachenheli]

Purpose

New model additions currently suffer from two pain points:

1. Adding new multimodal models or modalities requires changing hard-coded _placeholder_str function in https://github.com/vllm-project/vllm/blob/main/vllm/entrypoints/chat_utils.py#L507. This is now configurable via --mm-placeholder-str-override so future additions can be contained within model_executor/models/ directory without modifying the vllm framework code.
2. Video input processing currently uses a hard-coded 32 frame config. We want to allow flexible & extensible media io policies in the future.

Example flag formats:
--mm-placeholder-str-override '{"video":"<|video_placeholder|>", "image": "<|image_placeholder|>"}'
--media-io-kwargs '{"video": {"num_frames": 40, "fps": 2.0, "foo": "bar"}, "image": {"foo":"bar"} }'

Test Plan

• Unit tests:
  pytest tests/engine/test_arg_utils.py
pytest tests/multimodal/test_video.py
pytest tests/multimodal/test_utils.py

• vllm serve test. Printed & verified args are passed as expected.

vllm serve Qwen/Qwen2.5-VL-7B-Instruct --port 8000 --host 0.0.0.0 --dtype bfloat16 --limit-mm-per-prompt image=5,video=5 --mm-placeholder-str-override '{"video":"<|video_placeholder|>"}' --media-io-kwargs '{"video": {"num_frames": 40, "fps": 2.0, "foo": "bar"}, "image": {"foo":"bar"} }'

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @huachenheli, 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 flexibility and extensibility of multimodal model support by introducing configurable placeholder strings and dynamic video sampling policies. These changes empower users to easily integrate new multimodal models and customize video input processing without modifying core framework code, making the system more adaptable to future model additions and diverse use cases.

Highlights

• Configurable Multimodal Placeholder Strings: Introduced a new CLI flag --mm-placeholder-str-override that allows users to specify custom placeholder strings for different modalities (e.g., video, image). This removes the need to hardcode these strings within the framework, simplifying the addition of new multimodal models.
• Flexible Video Sampling Policies: Added a new CLI flag --video-media-io-kwargs which enables passing arbitrary keyword arguments to the video processing pipeline. This allows for flexible and extensible video sampling policies, moving beyond the previously hard-coded 32-frame configuration.
• Core Configuration Updates: Updated ModelConfig and MultiModalConfig to include the new video_media_io_kwargs and mm_placeholder_str_override fields, ensuring these configurations are propagated throughout the system.
• Media Connector Enhancements: Modified the MediaConnector to accept and utilize the new video_media_io_kwargs, ensuring that video fetching and processing can leverage the newly introduced flexible policies.
• Video Loader API Extension: Extended the VideoLoader and VideoMediaIO interfaces to accept and pass through arbitrary keyword arguments, allowing custom video loaders to define and use their own specific parameters for video 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 effectively addresses the stated pain points by introducing configurable multimodal placeholder strings and flexible video sampling policies via CLI flags. The changes are well-integrated across the configuration, argument parsing, and multimodal utility layers. New test cases have been added to validate the parsing of the new CLI arguments and the propagation of video processing parameters. The overall code quality is good, with clear intent and appropriate use of dataclasses and type hints. One minor type mismatch was identified in the EngineArgs definition for mm_placeholder_str_override.

[DarkLight1337]

LGTM, thanks for working on this! @ywang96 any further comments?

[DarkLight1337]

Perhaps we can later work on a solution similar to #20179 to define the placeholder token inside the model class....