SIGBUS error when doing inference with a torchscript model with CPU

C++
1

Hi I have a speech recognition model torchscript and a C++ engine that I’ve written to perform inference.

I’m noticing that some of the time when I run my model, I get SIGBUS exceptions with stack traces that don’t have much information.

Here is one of those stack traces that I got from boost stacktrace

FAILED WITH SIGNAL: 7
 0# handler(int) at /home/psridhar/speech-engine/bin/eval_session.cpp:79
 1# 0x00007FB303B2DFD0 in /lib/x86_64-linux-gnu/libc.so.6
 2# mkl_blas_avx512_sgemm_kernel_0 in /home/psridhar/speech-engine/third_party/libtorch/lib/libtorch_cpu.so

I got my libtorch from https://download.pytorch.org/libtorch/nightly/cpu/libtorch-cxx11-abi-shared-with-deps-latest.zip on March 10th.

Anyone have any idea why this is happening / how I can debug further?

2

Based on the stack trace it seems that MKL is raising the SIGBUS error in an AVX512 instruction.
If I’m not mistaken, this error could be raised, if the expected memory alignment is not met, so it could point towards an internal MKL error.
Would you mind posting this issue on GitHub with (if possible) an executable code snippet?