In the newest code, why the dtype for loading models is changed to float32 rather than bfloat16?
In the newest code, why the dtype for loading models is changed to float32 rather than bfloat16?