megengine.amp.low_prec_dtype¶

low_prec_dtype = <property object>¶

Get or set amp autocast mode’s lower precision dtype. It will change the target dtype in tensor casting for better speed and memory. Default: float16.

实际案例

import megengine as mge
mge.amp.low_prec_dtype = "float16"

megengine.amp.high_prec_dtype

megengine.dtr