Why is a 7B model with a file size of over 30GB?
#10
by
Saaiet
- opened
Just curious. Since the torch.dtype in the config.json file is bfloat16. it stands to reason that the model file should only be about 16GB.
We uploaded the model in fp32 precision, which results in a larger file size. The config file remains the default one provided by the official release.