Closed jonathanindig closed 1 year ago
I am having some trouble recreating the issue on my machine. Based on the comments for an OpenBLAS issue I found I recommend adding -Xss4096k
to the java command in the jep script.
Thanks @bsteffensmeier - that does see to work! 🎉
I did find a similar issue (actually your comment https://github.com/ninia/jep/issues/241#issuecomment-753697396) and tried setting that on the machine I was using and it didn't work, so I gave up on that front... but @jeremyrsmith tried setting -Xss8192k
on that machine which did the trick. The machines we were using have 8 cores, so maybe they need more memory (in case anyone else stumbles upon this and 4096
doesn't work for them).
Thanks for your help!
Describe the bug Running
np.linalg.inv
with a large enough matrix (10x10 is too small, 100x100 seems to do it) in the Jep console causes a segfault.I ran into this on a variety of systems. Here's a reproduction with Docker:
This can be mitigated by setting
OPENBLAS_NUM_THREADS=1
:But this slows things down considerably (seemingly between 2x to 10x but we haven't fully benchmarked) compared to plain Python.
Any idea why this is happening with Jep and whether there's anything else we can do? Thanks!