Open NohTow opened 1 week ago
This PR changes the loading logic to rely on the pytorch_model.bin file that is always there both locally and remotely. cc @jessiejuachon
pytorch_model.bin
Tests failing are not related to the PR but to this. We should either wait for ST to create the version with the merged fixes or set the version of transformers to pre-4.46.
68 fixed the local loading of Stanford models, but the issue is that the safe tensor is only created when uploaded to HF through the dedicated bot.
This PR changes the loading logic to rely on the
pytorch_model.bin
file that is always there both locally and remotely. cc @jessiejuachon