"Illegal instruction (core dumped)" when importing PyTorch

1

When I try to import pytorch, I keep getting “Illegal instruction (core dumped)”. I’ve tried CUDA 12.1, Ubuntu 22.04 and different installation methods.

Current setup:
Ubuntu 20.04
Nvidia RTX 3060
CUDA 11.8
Python 3.8.10

On this forum I found that the output of gdb might be useful, which is:

(gdb) run test.py
Starting program: /home/sphynx/Documents/test/.venv/bin/python3 test.py
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[Detaching after fork from child process 13834]

Program received signal SIGILL, Illegal instruction.
0x00007fffef02c732 in void std::vector<dnnl_post_ops::entry_t, std::allocator<dnnl_post_ops::entry_t> >::_M_realloc_insert<>(__gnu_cxx::__normal_iterator<dnnl_post_ops::entry_t*, std::vector<dnnl_post_ops::entry_t, std::allocator<dnnl_post_ops::entry_t> > >) () from /home/sphynx/Documents/test/.venv/lib/python3.8/site-packages/torch/lib/libtorch_cpu.so

Any help is appreciated.

2

For anyone encountering this in the future. For me, installing PyTorch with

pip3 install https://download.pytorch.org/whl/cu118/torch-2.0.0%2Bcu118-cp38-cp38-linux_x86_64.whl

instead of the command mentioned on the PyTorch website (Start Locally) solved the problem.

3

dnnl points to Intel’s oneDNN lib and does not seem to be supported on your CPU. Downgrading PyTorch might be a workaround but not a solution. Which CPU are you using?

4

Thanks for your reply. I am using an AMD Phenom II X4 965.

5

I don’t believe this CPU supports AVX, while it might be a requirement.
CC @malfet do you know if and which AVX requirement the current binaries have?