Cyrus Leung

> In my case, it was items in the messages array containing extra fields... date, and _id... This is the core of the breaking change that occurred after 0.4.1. The...

Can you add a test to check the model's consistency with its HuggingFace counterpart? That would be the most straightforward way to verify the correctness of your implementation. Afterwards, please...

By the way, there have been a few changes to multimodal models in the past few weeks, so you should merge `main` into your branch first.

Almost forgot - please also update `vllm.multimodal.get_full_image_text_prompt` with your model, otherwise, it won't be available via OpenAI API.

> By the way, May I ask one more question? In our MiniCPMV, there's a mechanism to slice the input image. In our previous implementation, we put the slicing into...

Thanks for the update! > MiniCPM-V will slice image when a image sent to it. So it needs a mechanism to change the length of input tokens. As this is...

#5276 has been merged, so you can start to make changes accordingly to support dynamic number of image tokens. You may refer to [this guide](https://docs.vllm.ai/en/latest/dev/multimodal/adding_multimodal_model.html) for more details.

Closing as it is superceded by #4540, which introduces `ExecuteModelRequest` to contain the keyword arguments, fixing #4568 in a more comprehensive manner.

This seems to be the same issue as #1200

Is this issue fixed in the current version of vLLM?