Previously, under mixed precision, we would autocast full precision variables to half precision, then manually cast back to full precision.
Possibly this would compile away (no idea), but is certainly very slow in eagerly. We should set autocast=False to avoid the needless type conversion loop.
Previously, under mixed precision, we would autocast full precision variables to half precision, then manually cast back to full precision.
Possibly this would compile away (no idea), but is certainly very slow in eagerly. We should set
autocast=False
to avoid the needless type conversion loop.