This PR fixes #65 by disabling numba caching due to a race condition in writing the cache when using MPI. The performance hit of re-generating the numba compilation every time is minor compared to the overall runtime of extractions and pixsims.
Curiously, I was only able to reproduce the failure with installed versions of specter. Simply having a bare clone in my $PYTHONPATH worked (though I couldn't find if/where it was writing the cache in that case). I have confirmed that turning off that cache works with installed copies of specter (albeit with low-number-statistics on the runs).
This PR fixes #65 by disabling numba caching due to a race condition in writing the cache when using MPI. The performance hit of re-generating the numba compilation every time is minor compared to the overall runtime of extractions and pixsims.
Curiously, I was only able to reproduce the failure with installed versions of specter. Simply having a bare clone in my $PYTHONPATH worked (though I couldn't find if/where it was writing the cache in that case). I have confirmed that turning off that cache works with installed copies of specter (albeit with low-number-statistics on the runs).