{ai}[foss/2022b] PyTorch v2.1.2 w/ CUDA 12.0.0 #20155
Open
+390
−0
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Test report by @Flamefire |
casparvl
reviewed
Mar 19, 2024
easybuild/easyconfigs/p/PyTorch/PyTorch-2.1.2-foss-2022b-CUDA-12.0.0.eb
Outdated
Show resolved
Hide resolved
|
Test report by @Flamefire |
|
Test report by @casparvl |
|
Here and in #20156 |
|
Hm, the log contains a lot, it's a bit hard to read, but I think this is the relevant part: Error log |
|
Error in Error log: |
|
Yep that is a known issue: Reinstall your pybind11 with the latest EC |
|
Great, will do! Sorry, there are so many fixes that I often can't keep up and don't always rebuild stuff XD I'll send a new test report after the pybind11 rebuild. |
Yeah I know that is annoying, but we can't do much better than updating the existing EC(s) for such major bugs. It came up recently with someone else too so I remembered it. Side note: This is actually a good reason to run the PyTorch test suite and investigate errors: Our pybind11 version isn't (wasn't) compatible with this PyTorch version which would make it less usable as this error is likely to pop up in user code using this module. |
|
Ok, I rebuild |
|
Test report sassy-crick: |
|
Test report by @casparvl |
|
Failures are the same for
The only new one was a failure in |
|
@boegelbot please test @ generoso |
|
@casparvl: Request for testing this PR well received on login1 PR test command '
Test results coming soon (I hope)... - notification for comment with ID 2016793002 processed Message to humans: this is just bookkeeping information for me, |
|
Test report by @boegelbot |
|
@boegelbot please test @ jsc-zen3 |
|
@casparvl: Request for testing this PR well received on jsczen3l1.int.jsc-zen3.fz-juelich.de PR test command '
Test results coming soon (I hope)... - notification for comment with ID 2020037044 processed Message to humans: this is just bookkeeping information for me, |
|
Test report by @boegelbot |
Flamefire
commented
Mar 27, 2024
easybuild/easyconfigs/p/PyTorch/PyTorch-2.1.2-foss-2022b-CUDA-12.0.0.eb
Outdated
Show resolved
Hide resolved
|
Test report by @casparvl |
You are missing the patches from #19666 which are in develop |
|
Ah, let me sync your branch with develop - I'm assuming you won't mind... :) |
|
Ok, rebuild started succesfully now. Test reporting should be there somewhere tonight. I'll trigger one more rebuild on one of the test clusters for good measure. Should be good to go afterwards... |
|
@boegelbot please test @ jsc-zen3 |
|
@casparvl: Request for testing this PR well received on jsczen3l1.int.jsc-zen3.fz-juelich.de PR test command '
Test results coming soon (I hope)... - notification for comment with ID 2032212011 processed Message to humans: this is just bookkeeping information for me, |
|
Test report by @boegelbot |
|
Test report by @casparvl |
…es: PyTorch-2.1.2_add-cuda-skip-markers.patch, PyTorch-2.1.2_fix-conj-mismatch-test-failures.patch, PyTorch-2.1.2_fix-device-mesh-check.patch, PyTorch-2.1.2_fix-locale-issue-in-nvrtcCompileProgram.patch, PyTorch-2.1.2_fix-test_extension_backend-without-vectorization.patch, PyTorch-2.1.2_fix-test_memory_profiler.patch, PyTorch-2.1.2_fix-test_torchinductor-rounding.patch, PyTorch-2.1.2_fix-vsx-vector-abs.patch, PyTorch-2.1.2_fix-vsx-vector-div.patch, PyTorch-2.1.2_fix-with_temp_dir-decorator.patch, PyTorch-2.1.2_fix-wrong-device-mesh-size-in-tests.patch, PyTorch-2.1.2_relax-cuda-tolerances.patch, PyTorch-2.1.2_remove-nccl-backend-default-without-gpus.patch, PyTorch-2.1.2_skip-cpu_repro-test-without-vectorization.patch, PyTorch-2.1.2_skip-failing-test_dtensor_ops-subtests.patch, PyTorch-2.1.2_skip-test_fsdp_tp_checkpoint_integration.patch, PyTorch-2.1.2_workaround_dynamo_failure_without_nnpack.patch
Flamefire
force-pushed
the
20240319165333_new_pr_PyTorch212
branch
from
5564421
to
8bb0d57
Compare
|
Test report by @akesandgren |
Failing especially due to the seg faults: |
|
Yeah, they are a bit weird. Can't see any reason for them to have happened... |
|
Check the logs if there is any more specific message for those crashing tests. |
|
Didn't find anything useful, a bunch of tracebacks but they weren't very informative... |
|
@Flamefire This is what I got: and and |
|
ok, the segfault in test_wrap doesn't happen everytime. A cleaner stacktrace is: |
|
I also have a problem with test/distributed/test_c10d_nccl.py::NcclProcessGroupWithDispatchedCollectivesTests::test_allgather_base which also segfaults. But if I change to NCCL 2.18.3 (i.e. the one used in 2023a) that problem goes away. |
|
Using a newer NCCL also seem to eliminate the SEGV in distributed/fsdp/test_wrap I'd say we should drop this one due to the NCCL version problem. Or add this to some more tests: |
|
Also note that .github/scripts/generate_binary_build_matrix.py hints at NCCL 2.18.1 being "required", or at least what they test with. |
Those 3 suites are crashing in your report:
The test_c10d_nccl seems to be "just" a failure or at least counted as such, easy enough to skip if it fails consistently This is the only PyTorch 2.x version in 2022b, so might be worth keeping. So 3 options:
I prefer the first option. |
|
Test report by @Flamefire |
|
I ran some tests and I see random failures of whole suites with SIGSEGV (test_jit) and I opened another PR using the newer NCCL: #20520 |
Updated software
|
|
Test report by @Flamefire |
|
Test report by @Flamefire |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
(created using
eb --new-pr)