Open jpmorgan98 opened 4 months ago
So this is odd. After an initial compilation some of the tests that had previously failed are passing using cached kernels. Again I still think that this has to do with compiler issues but we will see....
Ok so I think I was running into similar issues with the roc port and the soultion was a specific version of libgcc-ng
which is installed when conda install gxx
@braxtoncuneo can you comment on if this is the same issue you are seeing on Lassen?
@braxtoncuneo can you comment on if this is the same issue you are seeing on Lassen?
Reproduced my segfault. This is what I got:
__ __ ____ __ ____ ____
| \/ |/ ___|/ /_ _ \ / ___|
| |\/| | | /_ / | | | |
| | | | |___ / /| |_| | |___
|_| |_|\____|// |____/ \____|
Mode | Numba
Algorithm | History-based
MPI Processes | 1
OpenMP Threads | 1
Now running TNT...
Segmentation fault
So in the OSU CI machine cretin numba problems would copmile but fail to run. This happened on a number of the regression tests as well that where passing in the gh action runner. The full error is here:
Whenever I see errors like
lib64/libc.so.6
my mind immediately goes to incompatible compiler issues. First thing I tried asand that fixed it for some problems but still resulted in a seg fault for others specifically in the regression tests. I am running this in a manual terminal right now but eventually this will be the env that we do gh actions on for GPU regression testing. I am going to try other modules that have g++ and maybe look at llvm versions.
One thing to emphasize is this does seem like a runtime issue, not a compilation failure