Currently we require a config path, checkpoint path, and specify whether to return decoder
We should use a huggingface style interface, Model.from_pretrained, and specify a single URI with a config path and checkpoint path.
When training we should save the model and its config file to the output path every checkpoint step.
- Optional: Prefer safetensors.
Currently we require a config path, checkpoint path, and specify whether to return decoder
We should use a huggingface style interface,
Model.from_pretrained, and specify a single URI with a config path and checkpoint path.When training we should save the model and its config file to the output path every checkpoint step.