Illegal instruction (core dumped) : PyTorch 2.0 on Raspberry Pi 4.0 8gb

[P-Blackburn]

🐛 Describe the bug

Virgin install on Raspberry Pi 4.0 8gb

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
python3 -c "import torch;print(torch.__version__)"

results in:
Illegal instruction. and a core dump (on Ubuntu) and fails on the basic 64 Raspberry Pi Debian based OS...

HOWEVER:

Ignoring the installation instructions on https://pytorch.org/, and instead installing with:

pip3 install https://download.pytorch.org/whl/torch-2.0.0-cp310-cp310-manylinux2014_aarch64.whl

results in version of torch 2.0.0 that will work on the Raspberry Pi 4

Versions

executing:

wget https://raw.githubusercontent.com/pytorch/pytorch/master/torch/utils/collect_env.py
# For security purposes, please check the contents of collect_env.py before running it.
python collect_env.py

also results in Illegal instruction

(further - uninstalling torch 2.0.0 and attempting to install 1.13.1 per the instructions on https://pytorch.org/get-started/previous-versions/ i.e.

pip install torch==1.13.1+cpu torchvision==0.14.1+cpu torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cpu

results in

ERROR: Could not find a version that satisfies the requirement torch==1.13.1+cpu
ERROR: No matching distribution found for torch==1.13.1+cpu

that said an install of the form:

pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cpu

works a treat and

python3 -c "import torch;print(torch.__version__);print(torch.rand(3))"

returns what is expected:

1.13.1
tensor([0.1033, 0.6867, 0.4403])

python collect_env.py against the 1.13.1 works and returns:

Collecting environment information...
PyTorch version: 1.13.1
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Debian GNU/Linux 11 (bullseye) (aarch64)
GCC version: (Debian 10.2.1-6) 10.2.1 20210110
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.31

Python version: 3.9.2 (default, Feb 28 2021, 17:03:44) [GCC 10.2.1 20210110] (64-bit runtime)
Python platform: Linux-6.1.19-v8+-aarch64-with-glibc2.31
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: aarch64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 4
On-line CPU(s) list: 0-3
Thread(s) per core: 1
Core(s) per socket: 4
Socket(s): 1
Vendor ID: ARM
Model: 3
Model name: Cortex-A72
Stepping: r0p3
CPU max MHz: 1800.0000
CPU min MHz: 600.0000
BogoMIPS: 108.00
L1d cache: 128 KiB
L1i cache: 192 KiB
L2 cache: 1 MiB
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1: Mitigation; __user pointer sanitization
Vulnerability Spectre v2: Vulnerable
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fp asimd evtstrm crc32 cpuid

Versions of relevant libraries:
[pip3] numpy==1.19.5
[pip3] torch==1.13.1
[pip3] torchaudio==0.13.1
[pip3] torchvision==0.14.1
[conda] Could not collect

cc @ezyang @seemethere @malfet

[malfet]

@P-Blackburn can you please attach gdb to the process and print the instruction that causes an issue?
Also, can you try installing torch==1.13.0 and see if invalid instruction issue reproduces there. (I suspect it's ACL/MKLDNN integration)

[P-Blackburn]

Hi yes torch==1.13.0 works, also as I said in the main post pip3 install https://download.pytorch.org/whl/torch-2.0.0-cp310-cp310-manylinux2014_aarch64.whl 'works' with a version of PyTorch 2.0.0.

What doesn't work is pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu which essentially is the equivalent of pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/torch-2.0.0-1-cp310-cp310-manylinux2014_aarch64.whl

I am not very familiar with gdb, so please bear with me and direct me if I need to do something differently....
I installed gdb with sudo apt install gdb and installed sudo pip install pygdbmi

I then created a test.py of:

import torch
print(torch.__version__)

having then installed PyTorch with: pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

I get the following output:

pi@TRACY:~$ vi test.py
pi@TRACY:~$ python3 test.py
Illegal instruction (core dumped)
pi@TRACY:~$ gdb python3
GNU gdb (Ubuntu 12.1-3ubuntu2) 12.1
Copyright (C) 2022 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.
Type "show copying" and "show warranty" for details.
This GDB was configured as "aarch64-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<https://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
    <http://www.gnu.org/software/gdb/documentation/>.

For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from python3...
(No debugging symbols found in python3)
(gdb) run test.py
Starting program: /usr/bin/python3 test.py
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/aarch64-linux-gnu/libthread_db.so.1".

Program received signal SIGILL, Illegal instruction.
0x0000ffffed488930 in std::string::_Rep::_M_dispose(std::allocator<char> const&) [clone .part.0] () from /usr/local/lib/python3.10/dist-packages/torch/lib/../../torch.libs/libarm_compute-d27e629c.so

I removed PyTorch and installed torch==1.13.0 and also put that through gdb and the output was:

pi@TRACY:~$ python3 test.py
1.13.0
pi@TRACY:~$ gdb python3
GNU gdb (Ubuntu 12.1-3ubuntu2) 12.1
Copyright (C) 2022 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.
Type "show copying" and "show warranty" for details.
This GDB was configured as "aarch64-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<https://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
    <http://www.gnu.org/software/gdb/documentation/>.

For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from python3...
(No debugging symbols found in python3)
(gdb) run test.py
Starting program: /usr/bin/python3 test.py
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/aarch64-linux-gnu/libthread_db.so.1".
[New Thread 0xffffeb34f1a0 (LWP 12077)]
[New Thread 0xffffeab3f1a0 (LWP 12078)]
[New Thread 0xffffe632f1a0 (LWP 12079)]
1.13.0
[Thread 0xffffe632f1a0 (LWP 12079) exited]
[Thread 0xffffeab3f1a0 (LWP 12078) exited]
[Thread 0xffffeb34f1a0 (LWP 12077) exited]
[Inferior 1 (process 12075) exited normally]

[malfet]

@P-Blackburn thank you very much for detailed repro (alas, my Raspberry Pi is still running 32-bit linux, guess it's time to upgrade)
Most likely regression caused by pytorch/builder#1291 and the fact that ACL integration was enabled, see #96983
@snadampal is it really necessary to compile ACL with arch=armv8.2-a

[snadampal]

Hi @malfet , I'm surprised it's working on PT 1.13.0 but not on 2.0.0.1 because we have enabled the MKLDNN+ACL integration in 1.13.0 itself, and ACL was built witharch=armv8.2-a even for 1.13.0. Of course we have upgraded the ACL version this time. Let me check if armv8a is sufficient.

[P-Blackburn]

@malfet @snadampal I'm around for the next 12 hours so if you would like me to run any other tests, or provide you with ssh access to the PI4 - ping me. :)

[snadampal]

Hi @P-Blackburn , just to be double sure your 1.13.0 has mkldnn+acl, could you please run this on 1.13.0 installation and share the output.
$ python -c "import torch; print(torch.__version__, torch.backends.mkldnn.is_available())"

Is the illegal instruction coming from any particular gemm kernel? I'm trying to understand whether the issue is specific to PT2.0 ACL version or present even on PT1.13 ACL.

[P-Blackburn]

Hi @snadampal - I installed and tested 1.13.0 and 1.13.1 one was True the other was False

torch==1.13.0 True

peter@TRACY:~$ sudo pip install torch==1.13.0                                       Collecting torch==1.13.0
  Downloading torch-1.13.0-cp310-cp310-manylinux2014_aarch64.whl (73.2 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 73.2/73.2 MB 1.2 MB/s eta 0:00:00
Requirement already satisfied: typing-extensions in /usr/local/lib/python3.10/dist-packages (from torch==1.13.0) (4.5.0)
Installing collected packages: torch
Successfully installed torch-1.13.0
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
peter@TRACY:~$ python3 -c "import torch; print(torch.__version__, torch.backends.mkldnn.is_available())"
1.13.0 True
peter@TRACY:~$

torch==1.13.1 False

peter@TRACY:~$ sudo pip install torch==1.13.1
Collecting torch==1.13.1
  Using cached torch-1.13.1-cp310-cp310-manylinux2014_aarch64.whl (60.5 MB)
Requirement already satisfied: typing-extensions in /usr/local/lib/python3.10/dist-packages (from torch==1.13.1) (4.5.0)
Installing collected packages: torch
Successfully installed torch-1.13.1
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
peter@TRACY:~$ python -c "import torch; print(torch.__version__, torch.backends.mkldnn.is_available())"
Command 'python' not found, did you mean:
  command 'python3' from deb python3
  command 'python' from deb python-is-python3
peter@TRACY:~$ python3 -c "import torch; print(torch.__version__, torch.backends.mkldnn.is_available())"
1.13.1 False
peter@TRACY:~$

[P-Blackburn]

Hi @snadampal in answering the part of your question "Is the illegal instruction coming from any particular gemm kernel" Please do forgive me if I have not understood the question properly. However, I believe that simply trying to "import torch" when the installed version is torch-2.0.0-1-cp310-cp310-manylinux2014_aarch64.whl (which is the version installed following the instructions on the PyTorch.org installation page giving installation as: pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu immediately causes the illegal instruction and core dump

peter@TRACY:~$ sudo pip install torch
Collecting torch
  Using cached torch-2.0.0-1-cp310-cp310-manylinux2014_aarch64.whl (74.3 MB)
Requirement already satisfied: sympy in /usr/local/lib/python3.10/dist-packages (from torch) (1.11.1)
Requirement already satisfied: typing-extensions in /usr/local/lib/python3.10/dist-packages (from torch) (4.5.0)
Requirement already satisfied: filelock in /usr/local/lib/python3.10/dist-packages (from torch) (3.10.0)
Requirement already satisfied: networkx in /usr/local/lib/python3.10/dist-packages (from torch) (3.0)
Requirement already satisfied: jinja2 in /usr/lib/python3/dist-packages (from torch) (3.0.3)
Requirement already satisfied: mpmath>=0.19 in /usr/local/lib/python3.10/dist-packages (from sympy->torch) (1.3.0)
Installing collected packages: torch
Successfully installed torch-2.0.0
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
peter@TRACY:~$ python3 -c "import torch; print(torch.__version__, torch.backends.mkldnn.is_available())"
Illegal instruction (core dumped)
peter@TRACY:~$

[malfet]

@P-Blackburn it would be really nice if you can give me temp access to your Raspberry Pi (I'm using mine for home automation, so wouldn't even try to upgrade it until the weekend) Would you mind adding https://github.com/malfet.keys for next 48 hours and send me an IP address (can invite you to collaborate on a private repo, or send me an email with it)

[malfet]

Is the illegal instruction coming from any particular gemm kernel? I'm trying to understand whether the issue is specific to PT2.0 ACL version or present even on PT1.13 ACL.

@snadampal No, it is not, as visible from the backtrace posted in #97226 (comment) :

Program received signal SIGILL, Illegal instruction.
0x0000ffffed488930 in std::string::_Rep::_M_dispose(std::allocator<char> const&) [clone .part.0] () from /usr/local/lib/python3.10/dist-packages/torch/lib/../../torch.libs/libarm_compute-d27e629c.so

And instruction it crashes at are the following:

(gdb) disassemble 
Dump of assembler code for function _ZNSs4_Rep10_M_disposeERKSaIcE.part.0:
   0x0000ffffed3a8904 <+0>:	adrp	x2, 0xffffed7ec000
   0x0000ffffed3a8908 <+4>:	ldr	x2, [x2, #1336]
   0x0000ffffed3a890c <+8>:	cbnz	x2, 0xffffed3a8928 <_ZNSs4_Rep10_M_disposeERKSaIcE.part.0+36>
   0x0000ffffed3a8910 <+12>:	ldr	w2, [x0, #16]
   0x0000ffffed3a8914 <+16>:	sub	w3, w2, #0x1
   0x0000ffffed3a8918 <+20>:	str	w3, [x0, #16]
   0x0000ffffed3a891c <+24>:	cmp	w2, #0x0
   0x0000ffffed3a8920 <+28>:	b.le	0xffffed3a8938 <_ZNSs4_Rep10_M_disposeERKSaIcE.part.0+52>
   0x0000ffffed3a8924 <+32>:	ret
   0x0000ffffed3a8928 <+36>:	add	x3, x0, #0x10
   0x0000ffffed3a892c <+40>:	mov	w2, #0xffffffff            	// #-1
=> 0x0000ffffed3a8930 <+44>:	ldaddal	w2, w2, [x3]
   0x0000ffffed3a8934 <+48>:	b	0xffffed3a891c <_ZNSs4_Rep10_M_disposeERKSaIcE.part.0+24>
   0x0000ffffed3a8938 <+52>:	b	0xffffed0ce090 <_ZNSs4_Rep10_M_destroyERKSaIcE@plt>

I.e. it's likely coming from the global constructor, and that's exactly the reason why PyTorch for x86 is not compiled with AVX512 extensions by default, only part of it that is prohibited from having any global constructors that would be executed irrespective of the arch guards.

[malfet]

@P-Blackburn thank you very much, I'm in

[malfet]

Downgrading ACL to 22.05 moves crash from global constructor in the library, to ideep registration:

#0  0x0000ffffed0ca824 in arm_compute::cpuinfo::num_threads_hint() () from /home/nikita/.local/lib/python3.10/site-packages/torch/lib/../../torch.libs/libarm_compute-0268506b.so
#1  0x0000ffffed113498 in arm_compute::IScheduler::IScheduler() () from /home/nikita/.local/lib/python3.10/site-packages/torch/lib/../../torch.libs/libarm_compute-0268506b.so
#2  0x0000ffffed116c8c in arm_compute::Scheduler::get() () from /home/nikita/.local/lib/python3.10/site-packages/torch/lib/../../torch.libs/libarm_compute-0268506b.so
#3  0x0000fffff36c950c in std::call_once<dnnl::impl::cpu::aarch64::acl_thread_utils::acl_thread_bind()::{lambda()#1}>(std::once_flag&, dnnl::impl::cpu::aarch64::acl_thread_utils::acl_thread_bind()::{lambda()#1}&&)::{lambda()#2}::_FUN() () from /home/nikita/.local/lib/python3.10/site-packages/torch/lib/libtorch_cpu.so
#4  0x0000fffff7d54508 in __pthread_once_slow (once_control=0xfffff5a64050 <dnnl::impl::cpu::aarch64::acl_thread_utils::acl_thread_bind()::flag_once>, 
    init_routine=0xffffedbc0d90 <std::__once_proxy()>) at ./nptl/pthread_once.c:116
#5  0x0000fffff36c95bc in dnnl::impl::cpu::aarch64::acl_thread_utils::acl_thread_bind() () from /home/nikita/.local/lib/python3.10/site-packages/torch/lib/libtorch_cpu.so
#6  0x0000fffff2fe79ec in dnnl_engine_create () from /home/nikita/.local/lib/python3.10/site-packages/torch/lib/libtorch_cpu.so
#7  0x0000ffffef3a16d4 in ideep::engine::cpu_engine() () from /home/nikita/.local/lib/python3.10/site-packages/torch/lib/libtorch_cpu.so
#8  0x0000ffffeea0de0c in _GLOBAL__sub_I_IDeepRegistration.cpp () from /home/nikita/.local/lib/python3.10/site-packages/torch/lib/libtorch_cpu.so

which is the same ldaddal instruction:

 x/i $pc
=> 0xffffed0ca824 <_ZN11arm_compute7cpuinfo16num_threads_hintEv+2192>:	ldaddal	w2, w2, [x1]

[nSircombe]

Hi @malfet
I'm able to reproduce the illegal instruction running ACL's validation quite on Cortex-A72. I've not gone back to check earlier releases. It looks like an issue with the 'multi-isa' build, we'll have to take a closer look.

[P-Blackburn]

Hi @malfet and @nSircombe - I have also tested on the Raspberry Pi3 and found that the same illegal instruction and core dump also affects Cortex-A53.

[nSircombe]

I did bit more digging on ACL's multi_isa support - there are problems for <v8.2. Irrespective of what arch is set in the scons build, a multi_isa only has runtime selection of features since v8.2 - so there are 8.0 and 8.1 instructions (like ldaddal which I think came with v8.1) present in the multi-isa build.

[malfet]

@nSircombe, at the very least, there are no global constructor issue if one uses 22.05 vs v22.11

[malfet]

Small update: the same crash is observable on AWS A1 instances
Also, I wonder whether the crash is reproducible on M1 machines running aarch64 docker

[snadampal]

I have raised this PR (pytorch/builder#1370) to address this issue.
I have tested torchbench resnet50 and bert benchmarks on A72 and observed no crashes even with gemm execution.

[35grain]

Can confirm this is happening with the latest torch 2.0.0 version when importing the package on Raspberry Pi 4 4GB Ubuntu.

[atalman]

cc @snadampal . Thank you for cherry-picks. Please make sure to have all cherry-picks ready for April 14, 5PM PST.

This issue is in the milestones : https://github.com/pytorch/pytorch/milestone/36?closed=1, if you want to see your fix included in this minor release. Please post it as a cherry-pick into the [v2.0.1] Release Tracker.

The deadline is April 14, 5PM PST.

Only issues that have ‘cherry-picks’ will be considered for the release.

Common FAQs:

Q1: Where can I find more information on the release process and terminology?

A: pytorch/RELEASE.md at master · pytorch/pytorch · GitHub

Q2: Am I guaranteed to be included in the cherry-pick if I do above?

A: No, it is not guaranteed, the Release Team will review all submissions against the listed criteria before making the final decision on what to include on 4/17.

Q3: When is 2.1 going to be released?

A: We do not have a formal date at this time but will update the community when we do. Our immediate focus is 2.0.1. Note that 1.12 was released on 6/28/22, 1.13 on 10/28/22 and 2.0 on 3/15/23.

Q4: I missed the 4/14 5PM PST deadline, is there any option to have an extension?

A: No, in order to meet our 4/28 goal, we must hold 4/14 as our deadline and will not accept any requests after the fact. We are over communicating the timelines and process with the community to avoid such issues.

Q5: Where should I double check to see if my issue is in the cherry pick tracker?

A: [v2.0.1] Release Tracker · Issue #97272 · pytorch/pytorch · GitHub

Q6: Where can I find the Release Compatibility Matrix for PyTorch?

A: pytorch/RELEASE.md at master · pytorch/pytorch · GitHub

Please contact OSS Releng team members if you have any questions/comments. Again we appreciate everyone’s time and commitment to the community, PyTorch and 2.0 and 2.01 releases!

Please refer to this post for more details: https://dev-discuss.pytorch.org/t/pytorch-release-2-0-1-important-information/1176

[snadampal]

thanks, @atalman . I have posted this cherrypick request in the PyTorch 2.0.1 release tracker

• [v2.0.1] Release Tracker #97272 (comment)

[malfet]

Tested that 2.0.1 release candidate binaries (available at https://download.pytorch.org/whl/test/cpu ) no longer exhibit the problem on AWS A1 instances:

$ pip3 install torch --index-url https://download.pytorch.org/whl/test/cpu
$ python3 -mtorch.utils.collect_env
Collecting environment information...
PyTorch version: 2.0.1
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.2 LTS (aarch64)
GCC version: (Ubuntu 11.3.0-1ubuntu1~22.04) 11.3.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.35

Python version: 3.10.6 (main, Mar 10 2023, 10:55:28) [GCC 11.3.0] (64-bit runtime)
Python platform: Linux-5.15.0-1031-aws-aarch64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                    aarch64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
CPU(s):                          8
On-line CPU(s) list:             0-7
Vendor ID:                       ARM
Model name:                      Cortex-A72
Model:                           3
Thread(s) per core:              1
Core(s) per socket:              4
Socket(s):                       2
Stepping:                        r0p3
BogoMIPS:                        166.66
Flags:                           fp asimd evtstrm aes pmull sha1 sha2 crc32 cpuid
L1d cache:                       256 KiB (8 instances)
L1i cache:                       384 KiB (8 instances)
L2 cache:                        4 MiB (2 instances)
NUMA node(s):                    1
NUMA node0 CPU(s):               0-7
Vulnerability Itlb multihit:     Not affected
Vulnerability L1tf:              Not affected
Vulnerability Mds:               Not affected
Vulnerability Meltdown:          Not affected
Vulnerability Mmio stale data:   Not affected
Vulnerability Retbleed:          Not affected
Vulnerability Spec store bypass: Not affected
Vulnerability Spectre v1:        Mitigation; __user pointer sanitization
Vulnerability Spectre v2:        Mitigation; Branch predictor hardening, BHB
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected

Versions of relevant libraries:
[pip3] torch==2.0.1
[conda] Could not collect
$ python3 -c "import torch;x=torch.rand(3, 3);print(torch.linalg.svd(torch.mm(x,x.t())))"
<string>:1: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /root/pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)
torch.return_types.linalg_svd(
U=tensor([[-0.4932, -0.0094, -0.8699],
        [-0.8040, -0.3770,  0.4599],
        [-0.3322,  0.9262,  0.1783]]),
S=tensor([1.9125e+00, 6.0053e-01, 1.5409e-03]),
Vh=tensor([[-0.4932, -0.8040, -0.3322],
        [-0.0094, -0.3770,  0.9262],
        [-0.8699,  0.4599,  0.1783]]))

[aseok]

Facing same issue at 'import torch' with pytorch 2.0.1 rocm 5.4.2 on a pentium g4400 (no avx)