jax-am icon indicating copy to clipboard operation
jax-am copied to clipboard

Published 20 hours ago • verified publisher icon tianjuxue

PETSc with original ad_wrapper is computing wrong gradients

Open SNMS95 opened this issue 1 year ago • 1 comments

This has to be fixed!

We may need to keep short gradient checks for each function to detect where the leak is happening.

SNMS95 avatar Jul 22 '23 07:07 SNMS95

I think this is a problem with PETSc not converging. If you look at the logging information, the solver did not converge at all. So gradient computation will definitely be wrong.

I have a feeling that this is due to PETSc using 32-bit rather than 64-bit. If you have some time, please explore. Thanks.

tianjuxue avatar Jul 22 '23 14:07 tianjuxue