fix: allow pass image kwargs to image processor
#90
by
bigmoyan - opened
No description provided.
Why we need this?
(1) From a design standpoint, I see no justification for why image_processor does not accept images_kwargs.
(2) This leads to a critical bug: passing an image with height=3 and a specific width causes the vLLM service to crash immediately. The crash occurs because height=3 images resolve to different dimensions during embedding calculation versus token calculation, causing a length mismatch between the computed token count and the actual embeddings produced. Without support for **image_kwargs, we have no workaround from the client side.
I will submit a PR to vLLM to fix this issue on the vLLM side, but the prerequisite is that image_processor must accept externally passed **kwargs.
bigmoyan changed pull request status to
open