[V1] Initial support of multimodal models for V1 re-arch

[ywang96]

This PR adds V1 support for selected image language models

• Intern-VL
• Molmo
• Pixtral
• Qwen2-VL (missing MRope implementation, will address in a separate PR) separated to [V1] Add V1 support of Qwen2-VL  #12128

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[DarkLight1337]

Model changes look good to me overall. We can consider passing the placeholder indices to get_input_embeddings so that the relationship between input_ids and multimodal_embeddings becomes more explicit.

We should add tests to ensure that these models actually work in V1.

[ywang96]

We can consider passing the placeholder indices to get_input_embeddings so that the relationship between input_ids and multimodal_embeddings becomes more explicit.

Hmm I don't think that alone will help actually because V1 assumes chunked prefill by default. Unless we're passing additional information such as computed tokens per request, the model runner doesn't know the original indices of input_ids in their corresponding requests.

[WoosukKwon]

LGTM! It'd be nice if @DarkLight1337 can also take another look.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ywang96.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[ywang96]

It seems that we have tests related to using image embeddings as input, indicating usage of this feature from the community (which will be a breaking change in V1 since we cannot take batched image embedding directly as input), thus I'm separating Qwen2VL out to unblock this PR.

[DarkLight1337]

LGTM! It'd be nice if @DarkLight1337 can also take another look.

Sorry for the late reply, it looks good to me.