replace assertions by highlight in JIT tests to simple regex matches in PyTorch v1.10.0 tests

[lexming]

(created using eb --new-pr)

Workaround for issue pytorch/pytorch#72516

Some tests fail despite working as intended because the assertion by highlight fails to properly catch the highlighted part in the error message. Replacing those assertions with a simple regex match solves the issue.

[lexming]

@boegelbot please test @ generoso
CORE_CNT=16

[boegelbot]

@lexming: Request for testing this PR well received on login1

PR test command 'EB_PR=15073 EB_ARGS= /opt/software/slurm/bin/sbatch --job-name test_PR_15073 --ntasks="16" ~/boegelbot/eb_from_pr_upload_generoso.sh' executed!

• exit code: 0
• output:

Submitted batch job 8213

Test results coming soon (I hope)...

- notification for comment with ID 1057494985 processed

Message to humans: this is just bookkeeping information for me,
it is of no use to you (unless you think I have a bug, which I don't).

[boegelbot]

Test report by @boegelbot
SUCCESS
Build succeeded for 2 out of 2 (2 easyconfigs in total)
cnx1 - Linux Rocky Linux 8.5, x86_64, Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz (haswell), Python 3.6.8
See https://gist.github.com/4a739d215a3ac59e6524e8f7447a26be for a full test report.

[casparvl]

Looks reasonable to me. Any idea why we haven't hit this with the original PR: does it just happen on certain systems and/or under certain circumstances?

Out of personal interest: do you happen to know if this is still in PyTorch-1.11 as well? I've created an EasyConfig for 1.11-rc4 that I'd like to contribute once the final release is out. Just wondering if I should include this patch right away :)

Edit: to answer my own question: I looked at 1.11.0-rc5 and at the LOC that your patch changed. It's still relevant, also for PT 1.11.0. But, I did not hit this issue myself when installing based on my 1.11-rc4 EasyConfig. So I guess that juts leaves my first question: any idea when one hits this particular problem (and when not)? :)

[casparvl]

Test report by @casparvl
SUCCESS
Build succeeded for 2 out of 2 (2 easyconfigs in total)
software2.lisa.surfsara.nl - Linux debian 10.11, x86_64, Intel(R) Xeon(R) Bronze 3104 CPU @ 1.70GHz, 4 x NVIDIA NVIDIA TITAN V, 470.103.01, Python 2.7.16
See https://gist.github.com/a41714cc77d8b49fa58038ad2fc3f29f for a full test report.

[lexming]

@casparvl this is a good a question, but I'm afraid I do not have a clear answer. I also wonder how nobody else hit this issue before me. On our side, these failures are systematic, the same tests fail in every build. I have the suspicion that it might be related to using a temp dir in a shared filesystem. But I have no real proof. I'll try to run a couple of tests to shed some more light.

[casparvl]

@lexming Any updates on this? To me, the patch seems quite low impact (it's "only" about a test, and my understanding is that it still checks the results). At the same time, I saw the original issue on the PyTorch issue tracker, and saw the devs weren't too keen on making this change themselves - though I couldn't 100% understand their argumentation.

[boegel]

Test report by @boegel
SUCCESS
Build succeeded for 25 out of 25 (2 easyconfigs in total)
node3303.joltik.os - Linux RHEL 8.4, x86_64, Intel(R) Xeon(R) Gold 6242 CPU @ 2.80GHz (cascadelake), 1 x NVIDIA Tesla V100-SXM2-32GB, 510.47.03, Python 3.6.8
See https://gist.github.com/10c53d53aa64cd5cd97c9866c78caf04 for a full test report.

[boegel]

Test report by @boegel
FAILED
Build succeeded for 0 out of 2 (2 easyconfigs in total)
node3900.accelgor.os - Linux RHEL 8.4, x86_64, AMD EPYC 7413 24-Core Processor (zen3), 1 x NVIDIA NVIDIA A100-SXM4-80GB, 510.47.03, Python 3.6.8
See https://gist.github.com/ac5f3313d590bf093b1a3d42a6250af6 for a full test report.

[lexming]

I ran a few more tests and I can confirm that these tests only fail whenever the build is carried out on a shared filesystem (GPFS). So, another workaround is to execute eb with --tmpdir in a local filesystem.

UPDATE: my previous test worked because it got patched with this patch actually. The installation from current develop does fail on a local temp directory, you can see in the snipped below an example failure where the installation directory is in a shared filesystem, but --tmpdir is in /tmp

======================================================================
ERROR: test_del (jit.test_builtins.TestBuiltins)
----------------------------------------------------------------------
RuntimeError: 
undefined value a:
  File "/theia/scratch/brussel/vo/000/bvo00005/vsc10122/easybuild/install/skylake/build/PyTorch/1.10.0/foss-2021a/pytorch/test/jit/test_builtins.py", line 94
                a = x ** 2
                del a
                return a
                       ~ <--- HERE


During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/theia/scratch/brussel/vo/000/bvo00005/vsc10122/easybuild/install/skylake/build/PyTorch/1.10.0/foss-2021a/pytorch/test/jit/test_builtins.py", line 91, in test_del
    def fn(x):
  File "/tmp/eb-HcdqX4/tmp8BKhmT/lib/python3.9/site-packages/torch/testing/_internal/jit_utils.py", line 92, in __exit__
    FileCheck().check_source_highlighted(self.highlight).run(str(value))
RuntimeError: Expected to find "a"highlighted but it is not.

undefined value a:
  File "/theia/scratch/brussel/vo/000/bvo00005/vsc10122/easybuild/install/skylake/build/PyTorch/1.10.0/foss-2021a/pytorch/test/jit/test_builtins.py", line 94
                a = x ** 2

Full build log: https://gist.github.com/lexming/72be3627f389cabdd4ebc95361d1f79c

So my suspicions about the storage were not correct. I'll provide more info in the issue upstream, but this patch is still mandatory for us.

[Flamefire]

Test report by @Flamefire
FAILED
Build succeeded for 10 out of 12 (2 easyconfigs in total)
taurusi8011 - Linux CentOS Linux 7.9.2009, x86_64, AMD EPYC 7352 24-Core Processor (zen2), 8 x NVIDIA NVIDIA A100-SXM4-40GB, 470.57.02, Python 2.7.5
See https://gist.github.com/ad64041f1808e9e201812dceb96bbb78 for a full test report.

[Flamefire]

Test report by @Flamefire
FAILED
Build succeeded for 11 out of 12 (2 easyconfigs in total)
taurusa14 - Linux CentOS Linux 7.7.1908, x86_64, Intel(R) Xeon(R) CPU E5-2603 v4 @ 1.70GHz (broadwell), 3 x NVIDIA GeForce GTX 1080 Ti, 460.32.03, Python 2.7.5
See https://gist.github.com/c8ea0b28cd860527b37ce0cea33f8d86 for a full test report.

[lexming]

Abandoned. Patch will stay available at https://github.com/lexming/easybuild-easyconfigs/tree/vub-hpc/easybuild/easyconfigs/p/PyTorch

[Flamefire]

Maybe rebase this after #15904 was merged which got the other tests working