security-dependency-upgrade-user-gets-valueerror-requiring-torch-2-6-when-load-a5752b8a

THEN Convert the checkpoint to safetensors format, which avoids the security vulnerability and does not require torch 2.6. This can be done by loading the model with an older transformers version on a system that supports torch 2.6 (e.g., Colab), then saving it in safetensors format. Alternatively, downgrade transformers to version 4.41.0 and pin numpy to 1.26.4 to bypass the check temporarily.

# Example: load and resave a model in safetensors on a compatible machine
from transformers import AutoModel
model = AutoModel.from_pretrained("model-name", use_safetensors=False)
model.save_pretrained("./converted-model", safe_serialization=True)