Not fully tested yet, just raising to start reviewing.
Should be fully backwards compatible on API, and make safetensors a properly optional dependancy (loading should work with the pure python version, saving requires the module but the default training script behaviour is to skip and only write pytorch models if it's not available)
Not fully tested yet, just raising to start reviewing.
Should be fully backwards compatible on API, and make safetensors a properly optional dependancy (loading should work with the pure python version, saving requires the module but the default training script behaviour is to skip and only write pytorch models if it's not available)