Broken in transformers 4.53 >

#7
by Sorenmc - opened

When attempting to download the model with .from_pretrained I'm met with the following error:

ValueError: Supplied state dict for model.visual.blocks.31.mlp.down_proj.weight does not contain bitsandbytes__* and possibly other quantized_stats components.

This is related to the architecture overhaul huggingface made for vision language models, and can possibly be fixed by doing something like what unsloth does in their patching:

https://github.com/mmathew23/unsloth-zoo/blob/0235a7ff83d5a2ef4a0a43af025e82e4cda0cae4/unsloth_zoo/patching_utils.py#L447

Sign up or log in to comment