Open adri1wald opened 4 weeks ago
Can't say I have much here. Definitely no unsafe usage in this crate. Could it be an issue of running out of memory? Have you tried running these tests without parallelization (cargo test -- --test-threads 1
)?
I'm also not too familiar with the internals of ort
, but I wonder if using the load-dynamic
feature would give more consistent results instead of the default download behavior.
Hey @nkeenan38 thanks for responding!
Definitely not running out of memory.
I imagine without parallelisation this will decrease the likelihood of seg faults but in prod this will be part of a web server so actually parallelised tests is closer to reality.
I can try load-dynamic
but honestly I'm a bit hesitant to use this crate if it's backed by ort
. I tried forked this repo and tried to replace ort
with alternatives but so far no success.
Do you think it would make sense to submit this as an issue to the ort
repo?
I've written a test suite and it seg faults every couple of runs with outputs like:
I suspect the source of the issue is in
ort
since there is no unsafe in this crate, but posting here first.System information
Reproduction below. There's some semantically nonsensical stuff in there if you look closely, but that's because I simplified it while keeping it structurally equivalent from our actual tests.