You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I noticed that with certain versions of hugging face hub, above version 0.19 I think that models loaded to HF get saved as safetensors instead of .bin files by default and we don't currently support loading a safetensor model in our from_pretrained method, I don't know a lot about safetensors apart from skimming through the documentation page but seems it might be quicker to load models with them, be a bit more secure and also gives you a bit more model transparency on HF (shows some metadata without opening the file in the HF UI)
Wanted to open a discussion here on whether it would be worth moving to saving models in this new format and supporting loading them? Or at the very least adding support for loading them so we can have an option of which format we save them in? Keen to hear your thoughts @peterdudfield@dfulu@AUdaltsova, thanks
The text was updated successfully, but these errors were encountered:
I noticed that with certain versions of hugging face hub, above version 0.19 I think that models loaded to HF get saved as safetensors instead of .bin files by default and we don't currently support loading a safetensor model in our from_pretrained method, I don't know a lot about safetensors apart from skimming through the documentation page but seems it might be quicker to load models with them, be a bit more secure and also gives you a bit more model transparency on HF (shows some metadata without opening the file in the HF UI)
Wanted to open a discussion here on whether it would be worth moving to saving models in this new format and supporting loading them? Or at the very least adding support for loading them so we can have an option of which format we save them in? Keen to hear your thoughts @peterdudfield @dfulu @AUdaltsova, thanks
The text was updated successfully, but these errors were encountered: