mojolearn.numeric_profile

mojolearn.numeric_profile(use='inference')

The numeric profile used when a call names none.

A numeric profile is the number format the matrix products of the neural models compute in. It is independent of the numeric mode: under numeric_mode="identical" every profile gives the same bits across vendors. The default is "fp32_v1" for inference and training unless the MOJOLEARN_NUMERIC_PROFILE environment variable or mojolearn.set_numeric_profile() sets another inference default. The inference default applies only to the families it covers (the transformer models of mojolearn.models); other families use "fp32_v1".

Parameters:

use ({“inference”, “training”}, default “inference”) – Which default to return.

Returns:

str – The profile name.

Examples

>>> import mojolearn
>>> mojolearn.numeric_profile()
'fp32_v1'