Saving Transformers model as binary instead of safetensors format?

[StatsGary]

I am trying to Torchserve (https://pytorch.org/serve/) a Transformers model to be deployed as a prediction API on GCP as a custom prediction service.

However, torchserve requires that the model be in either .bin format or .pt format. Is there a way to deactivate safetensors and return the .bin PyTorch file?

I have tried to use torch.save(model.state_dict(), PATH) but have been unsuccessful thus far? Anyone encountered similar issues?

[StatsGary]

➡️💡 SOLUTION ➡️💡

I dig some digging of the parameters of the save_pretrained and trainer methods and you can actually 🔧 turn off the storing of the safetensors format models, by using the following example:

model.save_pretrained(<PATH_TO_SAVE_MODEL>, safe_serialization=False) #Replace path

[Narsil]

Could luck trying to secure your servers from malicious actors ! :)