[Model] Add moondream vision language model

[vikhyat]

This pull request adds support for the moondream vision language model, added similar to how the LLaVA implementation works.

Usage:

    llm = LLM(
        model="vikhyatk/moondream2",
        trust_remote_code=True,
        image_input_type="pixel_values",
        image_token_id=50256,
        image_input_shape="1,3,378,378",
        image_feature_size=729,
    )

   output = llm.generate(
        [("<|endoftext|>" * 729) + "\n\nQuestion: Describe this image.\n\nAnswer:"],
        multi_modal_data=MultiModalData(
            type=MultiModalData.Type.IMAGE, data=image_pixels
        ),
        sampling_params=sampling_params,
    )

---

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Model] for adding a new model or improving an existing model. Model name should appear in the title.
• [Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
• [Kernel] for changes affecting CUDA kernels or other compute kernels.
• [Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
• [Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• We adhere to Google Python style guide and Google C++ style guide.
• Pass all linter checks. Please use format.sh to format your code.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
• Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

• After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
• After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
• After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
• Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

[DarkLight1337]

Thanks for expanding the list of supported multimodal models!

Could you add a test to ensure that the model is consistent with its HuggingFace implementation? (Similar to test_llava.py)

To facilitate #4194, it would be great if you could also wrap the image preprocessing logic inside of a HuggingFace image processor so it can be loaded by AutoImageProcessor.

[DarkLight1337]

#4197 is very close to getting merged. @vikhyat can you address the issues mentioned above? Meanwhile, I can work on registering the model in vLLM to support multimodal input (as described in the docs added by #4197).

[ywang96]

Hello! @vikhyat Thank you for the contribution and I'm excited to have this model supported in vLLM. We have done a significant amount of refactoring on vLLM for supporting multimodal models in the last 2 months - please see updates in #4194.

If you don't mind, I'm happy to update this PR accordingly so we can get this merged. Let me know!

[vikhyat]

Hi @ywang96! I had this in my backlog but I'm currently swamped trying to get a new version of the model trained and released, so it has been hard to find the time. I would definitely hugely appreciate it if you can help out!

[github-actions]

This pull request has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this pull request should remain open. Thank you!

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. @vikhyat please rebase it. https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[ywang96]

Ooops - Feel free to ignore the comment from the bot. (I do still need to spend some time working on this model...)

[nph4rd]

👀

[ValeryNo]

Very relevant now that a major update has been released today. Great model

[ywang96]

Very relevant now that a major update has been released today. Great model

@ValeryNo I've spoken to vik about this and spent some time myself on this model, but unfortunately we don't support PrefixLM attention as of now (but will likely do via FlexAttention in the future).