FP8 range for E4M3 dtype

Inference FAQ
1

E4M3 data type is supposed to have a maximum value of 448. When I test that data type in 1.16.2, I see the maximum value is 240. Why is there such a discrepancy?

To reproduce the issue, you can use:

‘’’
import torch
import habana_frameworks.torch.core as htcore
import habana_frameworks.torch.hpex.experimental.transformer_engine as te

x = torch.arange(0, 448., device=‘hpu’).to(torch.float8_e4m3fn)
print(x)
‘’’

1 Like
2

Thanks for posting, we are looking into this

3

Hi, any update about this range?

4

This is being worked upon by internal teams. Once they provide guidelines on which fp8 version is supported exactly, will post here.

1 Like