Optimizers with high atols should likely be using np.complex128 #5536
Labels
area/transformers
kind/bug-report
Something doesn't seem to work.
triage/accepted
there is consensus amongst maintainers that this is a real bug or a reasonable feature to add
We have flakes on tests that have quite high atols due to the move from
np.complex128
tonp.complex64
default. See #5535 for example. These are sometimes optimizers that are very numerically unstable. We should likely usenp.complex128
for these optimizers (say ones with atols of 2e-5 or worse)The text was updated successfully, but these errors were encountered: