I’m using the Docker image vllm/vllm-openai:latest
I’m currently using the mistralai/Mistral-Small-4-119B-2603 model and I’m having an issue with vLLM version 0.23.0.
It shows an error when launching:
ValueError: Failed to apply MistralCommonPixtralProcessor on data={'text': '', 'images': [<PIL.Image.Image image mode=RGB size=1540x1540 at 0x7B05C8517110>]} with kwargs={'return_tensors': None}
As if the PIXTRAL part isn’t being taken into account.
If I disable the image processing part with the parameter --limit-mm-per-prompt '{"image":0}, it works.
Does anyone else have the same issue?
Should I necessarily downgrade, or is there a patch available?
This is a known issue in vLLM 0.22.1 and 0.23.0 when running Mistral models with Pixtral vision (like Mistral-Small-4-119B-2603): the processor passes return_tensors=None to the HuggingFace processor, which expects ‘pt’, causing a ValueError during dummy profiling. Disabling vision with --limit-mm-per-prompt ‘{“image”:0}’ is a temporary workaround, but image inputs won’t work until the bug is fixed. A patch is being developed, and similar issues have been addressed for related models, but as of now, there is no official release with a fix—downgrading to a version before 0.22.1 or waiting for the patch is recommended if you need vision support. See the discussion and workaround in the linked issue for updates and PR references.
Would you like more detail on the technical cause or tracking the patch status?
