Open cortner opened 2 years ago
This is still a problem in J1.7.1, and occurs only when running as part of the test suite, not when run "manually" within VS Code, or by simply executing the test file as a script.
Some updates - it seems this occurs only with O3 and forcing bounds checking. How weird is that? Also, one moment it is reproducible, next it is not... I don't like it.
... and now I can't reproduce it anymore...
This still appears to happen, now opened a discussion at Discourse but nobody seems interested.
no longer clear that Zygote is the culprit. segfaults happen randomly all over the place. It seems that using Rosetta-Julia 1.7 resolves the issue for now? But maybe this is random too?
test_admodel.jl
has a segfault when Zygoting the loss w.r.t. the model. (indicated in comments onmain
branch)