Cyrus Leung

Note that I have split out some of the changes in `conftest.py` (that do not depend on multimodal data) to #5118. I have also split out the commit for using...

Note that I have removed `multi_modal_data` from the legacy LLM API since this PR will break existing code that uses it anyway (due to the new method of constructing `MultiModalData`).

I have enabled tensors to be passed to `ImagePixelData` to skip preprocessing. However, it seems that the pixel tensor has slightly different preprocessing than HuggingFace's `LlavaProcessor`, causing the test to...

We have added [documentation](https://docs.vllm.ai/en/stable/getting_started/debugging.html) for this situation in #5430. Please take a look.

OP has been fixed by #5450. > `NameError: name 'vllm_ops' is not defined` Fixed by #5009

#5140 looks promising as a first step.

Ref. #5816

Note that by setting `gpu_memory_utilization` to a low value for tests that only use small models, we can now run multiple files in `entrypoints_tests` at the same time. However, I...

> @DarkLight1337 Left comments & questions - PTAL! Thanks for your time! I have responded to your comments.

@ywang96 I have finished addressing this round of comments.