You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
bfloat16 has virtually the same range as float32, just with less precision but this is by design. If you don't have memory issues it would (almost) always be preferable to load data as float32.
Checking all data values before loading the data could also slow down the loading process, but we could consider adding an option to check all values and warn if they are out-of-range of the target datatype.
Perhaps it would be good to check if the data contains big values and propose scaling it down to avoid floating point accuracy issues.
The text was updated successfully, but these errors were encountered: