[CI/Build] Add Model Tests for PixtralHF

[mgoin]

Adds a simple vLLM vs HF test for PixtralHF models

Tested locally with

pytest tests/models/decoder_only/vision_language/test_models.py -k "pixtral" -s -v
...
============================= 8 passed, 109 deselected, 77 warnings in 657.77s (0:10:57) ==============================

[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]

Hey @mgoin can you hold this PR until we merge #9372?

[mgoin]

Ah okay @ywang96, sure it can wait

[DarkLight1337]

Suggested change

	prompt_formatter=lambda img_prompt: f"<s>[INST]{img_prompt}[/INST]",
	prompt_formatter=lambda img_prompt: f"<s>[INST]\n{img_prompt}[/INST]",

[DarkLight1337]

Actually based on the example script, there should be a newline here

[mgoin]

@DarkLight1337 From the HF example, the newline is actually supposed to be between the prompt and the images i.e.

PROMPT = "<s>[INST]Describe the images.\n[IMG][IMG][IMG][IMG][/INST]"

vllm/examples/offline_inference_vision_language.py

Line 295 in 77f7ef2

prompt = f"<s>[INST]{question}\n[IMG][/INST]"

I wasn't sure how to do this with this structure, but it wasn't important to get matching vLLM vs HF

[DarkLight1337]

Oh, I see now. You can use the img_idx_to_prompt argument for using [IMG] token instead of <image>. We need to update the test framework to be able to support custom joining between text and image parts though.

[mgoin]

Yeah I did use img_idx_to_prompt=lambda idx: "[IMG]", and that worked thankfully. Luckily the newline isn't crucial for correctness, so that feature can wait