Update CUDA to version 13.4.1 - #10844
Conversation
The NVIDIA drivers are now packaged separately. Patch the cuda-gdb shell script rather than adding a second wrapper.
|
A new Pull Request was created by @fwyzard for branch IB/CMSSW_20_1_X/master. @akritkbehera, @cmsbuild, @iarspider, @raoatifshad, @smuzaffar can you please review it and eventually sign? Thanks. |
|
cms-bot internal usage |
|
@smuzaffar @akritkbehera, if this works for GCC 14 in C++20 mode, how do I test it with GCC 15 and GCC 16 in C++23 mode ? |
|
enable gpu |
|
please test |
|
please test for el9_amd64_gcc16 with cms-sw/cms-bot#2838 |
|
please test for el9_amd64_gcc15 with cms-sw/cms-bot#2838 |
|
-1 Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-e693b7/55790/summary.html Failed External BuildI found compilation error when building: cwd: /data/cmsbld/jenkins/workspace/ib-run-pr-tests/testBuildDir/BUILD/el9_amd64_gcc16/external/py3-torch-cuda/2.13.0-c725d6eced9d91b59d33ddc3ba0b3142/external_py3-torch-cuda_2.13.0-c725d6eced9d91b59d33ddc3ba0b3142-1-build/cmsdist-pip-src/torch-2.13.0
Building wheel for torch (pyproject.toml): finished with status 'error'
ERROR: Failed building wheel for torch
Failed to build torch
ERROR: Failed to build one or more wheels
error: Bad exit status from /data/cmsbld/jenkins/workspace/ib-run-pr-tests/testBuildDir/tmp/rpm-tmp.xJbBlY (%build)
RPM build warnings:
Macro expanded in comment on line 674: %{pkginstroot}/bin/*
Macro expanded in comment on line 679: %{pkginstroot}/${PYTHON3_LIB_SITE_PACKAGES}
|
|
The failure with GCC 16 is a general CUDA issue, already reported here: https://forums.developer.nvidia.com/t/nvcc-13-4-rejects-conforming-libstdc-string-with-g-16/382852 . |
|
-1 Failed Tests: UnitTests nvidia_l40sUnitTests Failed Unit TestsI found 10 errors in the following unit tests: ---> test unitTestsGroup_1_u2 had ERRORS ---> test unitTestsGroup_4_u20 had ERRORS ---> test TestFWCoreIntegrationStandalone had ERRORS and more ... Comparison SummarySummary:
AMD_MI300X Comparison SummarySummary:
AMD_W7900 Comparison SummarySummary:
NVIDIA_H100 Comparison SummarySummary:
NVIDIA_L40S Comparison SummarySummary:
NVIDIA_T4 Comparison SummarySummary:
|
|
-1 Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-e693b7/55801/summary.html Failed External BuildI found compilation error when building: cwd: /data/cmsbld/jenkins/workspace/ib-run-pr-tests/testBuildDir/BUILD/el9_amd64_gcc15/external/py3-torch-cuda/2.13.0-69c963f7f69016829584583110c7f92b/external_py3-torch-cuda_2.13.0-69c963f7f69016829584583110c7f92b-1-build/cmsdist-pip-src/torch-2.13.0
Building wheel for torch (pyproject.toml): finished with status 'error'
ERROR: Failed building wheel for torch
Failed to build torch
ERROR: Failed to build one or more wheels
error: Bad exit status from /data/cmsbld/jenkins/workspace/ib-run-pr-tests/testBuildDir/tmp/rpm-tmp.JwEzSB (%build)
RPM build warnings:
Macro expanded in comment on line 674: %{pkginstroot}/bin/*
Macro expanded in comment on line 679: %{pkginstroot}/${PYTHON3_LIB_SITE_PACKAGES}
|
|
torch still fails with |
Is that because of gcc15 or because of c++23 ? |
|
I think this is due to c++23, last time we tested cuda update with CPP23/GCC14 and we have seen the same ICE for that. |
|
Would it be useful to have CUDA with gcc15 or gcc16 in c++20 mode ? |
|
What I mean is that, given that CUDA still does not really support c++23, we could at least update the compiler to gcc 15 or gcc 16 and remain in c++20 mode 🤷🏻♂️ |
|
I have no objections moving default gcc15/gcc16 to c++20. May be we can add extra CPP23 IBs for gcc15 and gcc16 just to make sure we build the sw stack with c++23 and newer compilers |
OK, let's bring this up at the next Core Software meeting.
Yes, absolutely. |
FWIW I'm fine with the suggestion (in case you don't want to wait for the next core meeting that might take a while).
This sounds good to me as well. |
|
ok, I will change GCC15/16 OBs to use default c++20 and also open CPP23 IBs for GCC15/16 |
Update CUDA to version 13.4.1 (general availability).
Major changes since CUDA 13.3:
Patch the
cuda-gdbshell script rather than adding a second wrapper.