[Performance]: does vllm support tensor data for requests？

[zswodegit]

Proposal to improve performance

For my current project, I need to accelerate image-text inference. I want to cache images locally first, and then directly send the tensor values when calling the API. However, it seems that image data can only be sent in Base64 format. Is there any way to send tensor-type data instead?

Report of performance regression

No response

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

The output of `python collect_env.py`

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[DarkLight1337]

By "tensor values", do you mean the original image pixels in NCHW format? Has the image been processed yet (e.g. resize, normalization)?