safetensor percision question :)
#9
by
SynysterSocks
- opened
Hi,
absolutly fantastic work :).
appologies for the question, is the transformer safetensor uploaded to huggingface stored in full fp32 or bf16?
thank you :).
Thank you for your usage and support!
The uploaded transformer is in BF16 data type, which is the same data type we used during training. Therefore, I think there should be no precision loss due to data type conversion.