security-torch-vulnerability-valueerror-due-to-a-serious-vulnerability-issue-in-d7a183b1

THEN If you cannot upgrade torch (e.g., no wheel available for your platform), convert your PyTorch checkpoint files to safetensors format. Use an older version of transformers (e.g., 4.41.0) to load the model and resave it with safe_serialization=True. Future loading will then bypass the torch.load vulnerability check. Alternatively, upgrade PyTorch to 2.6+ if a wheel is available for your system.

# Load model (with older transformers version) and resave as safetensors
model = CLIPModel.from_pretrained("clip-ViT-B-32")
model.save_pretrained("./model_safetensors", safe_serialization=True)