[Bugfix] Fix Mllama interleaved images input support

[Isotr0py]

Slack Discussion: https://vllm-dev.slack.com/archives/C07QCGVDNUF/p1743003881430129?thread_ts=1743001000.770969&cid=C07QCGVDNUF

• Fix broken Mllama interleaved images input support
• Rewrite multi-images example to use interleaved images input

cc @heheda12345

[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.

🚀

[Isotr0py]

Interleaved image test should pass now.

[heheda12345]

Will it be better to put this hack in get_replacement_mllama?

[heheda12345]

Interleaved image test should pass now.

I ran the tests in my local env and get the following errors:

FAILED test_mllama.py::test_models_interleaved_images[_Backend.XFORMERS-5-128-bfloat16-meta-llama/Llama-3.2-11B-Vision-Instruct] - AttributeError: 'list' object has no attribute 'shape'
FAILED test_mllama.py::test_models_interleaved_images[_Backend.FLASH_ATTN-5-128-bfloat16-meta-llama/Llama-3.2-11B-Vision-Instruct] - AttributeError: 'list' object has no attribute 'shape'

I'm expecting this PR to fix it #14883.

How can you pass the interleaved image tests?

[Isotr0py]

I ran the tests in my local env and get the following errors:

Hmmm, that's weird... I re-ran the test locally again, and it can still pass. The only change I made is changing the test to use tensor_parallel_size=2 because I can't load the model with only one GPU, which should not affect the test...

[heheda12345]

Discussed with @Isotr0py offline. This PR can pass the tests in test_mllama.py if we set max_num_seqs=1 to vllm_runner. We can merge this PR first and wait for #14883 to support max_num_seqs>1.