[V1] Refactor model executable interface for multimodal models

[ywang96]

This PR refactors the interface of all multimodal language models for V1 VLM re-arch and torch.compile support. In particular, all multimodal model implementations on vLLM will need to meet the following requirements:

• get_multimodal_embeddings(**kwargs) implemented in XYZModel or XYZForConditionalGeneration
• get_input_embeddings(input_ids, multimodal_embeddings) implemented in XYZModel or XYZForConditionalGeneration to output input embeddings to be passed to the language backbone.
• Backward compatibility added for V0 in XYZModel or XYZForConditionalGeneration until V0 is fully deprecated.

List of LMMs to be worked on

• BLIP2
• Chameleon
• GLM
• Fuyu
• Intern-VL
• Llava-next
• Llava-next-video
• Llava-one-vision
• Molmo
• Paligemma
• Pixtral
• Qwen2VL
• Qwen2Audio
• Ultravox

This PR is a prerequisite of applying #9871 to all multimodal models on vLLM.

[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.

🚀

[ywang96]

This PR is mostly done but with a few caveats:

1. For MiniCPMV, the multimodal embedding generation relies on input ids. It's not so clear how it can be separated from input text embedding generation. We need to tag model author to see how to separate them.
2. For Molmo, the original implementation requires the final multimodal embedding length to match input embedding length becasue of the (+) operation. I'll leave optimizing this to a later PR.
3. This PR does not consider mllama.

All the other models included in this PR have been tested with example scripts to ensure v0 compatibility. Full v1 compatibility will be worked in the next PR.

[DarkLight1337]

I'm getting this error for internvl tests (tests/models/decoder_only/vision_language/test_models.py::test_single_image_models[intern_vl-test_case52]):

self = InternVLChatModel(
  (vision_model): InternVisionPatchModel(
    (embeddings): InternVisionEmbeddings(
      (patch_em...048, bias=True)
    (2): GELU(approximate='none')
    (3): Linear(in_features=2048, out_features=2048, bias=True)
  )
)
input_ids = tensor([92546, 92546, 92546,  ..., 92546, 92546, 92546], device='cuda:0')

    def _get_visual_token_mask(self, input_ids: torch.Tensor) -> torch.Tensor:
        if self.is_mono:
>           visual_token_mask = (
                input_ids == self.img_context_token_id).reshape(-1, 1)
E           AttributeError: 'bool' object has no attribute 'reshape'

vllm/model_executor/models/internvl.py:639: AttributeError

I think it's because self.img_context_token_id hasn't been set yet, so it resolves to None in the comparison expression.

[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]

I'm getting this error for internvl tests (tests/models/decoder_only/vision_language/test_models.py::test_single_image_models[intern_vl-test_case52]):

self = InternVLChatModel(
  (vision_model): InternVisionPatchModel(
    (embeddings): InternVisionEmbeddings(
      (patch_em...048, bias=True)
    (2): GELU(approximate='none')
    (3): Linear(in_features=2048, out_features=2048, bias=True)
  )
)
input_ids = tensor([92546, 92546, 92546,  ..., 92546, 92546, 92546], device='cuda:0')

    def _get_visual_token_mask(self, input_ids: torch.Tensor) -> torch.Tensor:
        if self.is_mono:
>           visual_token_mask = (
                input_ids == self.img_context_token_id).reshape(-1, 1)
E           AttributeError: 'bool' object has no attribute 'reshape'

vllm/model_executor/models/internvl.py:639: AttributeError

I think it's because self.img_context_token_id hasn't been set yet, so it resolves to None in the comparison expression.

I pushed a change which I think should be a clean fix for this. Let me know what you think!