CUDA's installer hangs for VS 2022 17.13 and later #5282
Labels
Comments
StephanTLavavej
added
external
This was referenced Mar 3, 2025
N-Storm
added a commit
to N-Storm/mfaktc
that referenced
this issue
Mar 3, 2025
This commit provides a workaround for these issues: microsoft/STL#5282 Jimver/cuda-toolkit#382 And fixes builds for older CUDA versions 8.x - 10.x as well.
N-Storm
added a commit
to N-Storm/mfaktc
that referenced
this issue
Mar 3, 2025
This commit marks version 0.23.2 (see primesearch#14). * Fix incorrect `cuda_version` reference and order of CUDA version / CC columns in the table generated in the release notes. * Fix `mfaktc.ini` copy on Windows CI/CD builds: GitHub Actions runs an unusual environment for Windows builds, where GNU Make has `SHELL` set to PowerShell (otherwise, GNU Make from MSYS2 would try to launch the Bash shell, which fails to build without heavy modifications to the Makefile). For some reason, Make fails to spawn the `copy` command, probably because it's just a shortcut for the `Copy-Item` command. Anyway, simply copying this file to the upper directory before launching Make fixes this issue, as Make will skip this step once it has already been completed. This way, we don't have to patch `Makefile.win` or come up with other workarounds. * Changes suggested by @tdulcet on primesearch#14 added. Thanks! These include: * Set action `fail-fast` to `false`. This allows other jobs to continue if one fails. * Commented out CUDA versions to leave only the highest `.patch` version per `major.minor` version. * "Code quality" improvements to the workflow & helper script. * Fix broken CUDA installation with newer MSVC versions. This provides a workaround for these issues: microsoft/STL#5282 Jimver/cuda-toolkit#382 A recent update to GitHub Actions' Windows runner images updated the MSVC version, causing the CUDA Toolkit installer to hang during the VS integration component installation. A workaround was added to install only the `nvcc` and `cudart` components. Additionally, the CUDA binaries directory must be added to `PATH` from the workflow since it was originally set by the broken component installation. * Added CUDA 8.x / 9.x / 10.x Windows builds. There is an option to run the MSVC 14.0 (2015) build environment, which is installed as a component alongside MSVC 2019 on the GitHub Actions runner.
N-Storm
added a commit
to N-Storm/mfaktc
that referenced
this issue
Mar 3, 2025
This commit marks version 0.23.2 (see primesearch#14). * Fix incorrect `cuda_version` reference and order of CUDA version / CC columns in the table generated in the release notes. * Fix `mfaktc.ini` copy on Windows CI/CD builds: GitHub Actions runs an unusual environment for Windows builds, where GNU Make has `SHELL` set to PowerShell (otherwise, GNU Make from MSYS2 would try to launch the Bash shell, which fails to build without heavy modifications to the Makefile). For some reason, Make fails to spawn the `copy` command, probably because it's just a shortcut for the `Copy-Item` command. Anyway, simply copying this file to the upper directory before launching Make fixes this issue, as Make will skip this step once it has already been completed. This way, we don't have to patch `Makefile.win` or come up with other workarounds. * Changes suggested by @tdulcet on primesearch#14 added. Thanks! These include: * Set action `fail-fast` to `false`. This allows other jobs to continue if one fails. * Commented out CUDA versions to leave only the highest `.patch` version per `major.minor` version. * "Code quality" improvements to the workflow & helper script. * Fix broken CUDA installation with newer MSVC versions. This provides a workaround for these issues: microsoft/STL#5282 Jimver/cuda-toolkit#382 A recent update to GitHub Actions' Windows runner images updated the MSVC version, causing the CUDA Toolkit installer to hang during the VS integration component installation. A workaround was added to install only the `nvcc` and `cudart` components. Additionally, the CUDA binaries directory must be added to `PATH` from the workflow since it was originally set by the broken component installation. * Added CUDA 8.x / 9.x / 10.x Windows builds. There is an option to run the MSVC 14.0 (2015) build environment, which is installed as a component alongside MSVC 2019 on the GitHub Actions runner.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Known issue: https://forums.developer.nvidia.com/t/stuck-installing-nsight-on-cuda-12-8-0-571-96-windows/323732/2
(Update: This is a VS bug, DevCom-10841757, expected to be fixed in VS 2022 17.14 Preview 2.)
I encountered this when updating the toolset (from Preview 2) to VS 2022 17.13 Preview 3 in #5247 but didn't understand the cause then.
Unfortunately, my "solution" of installing just 'nvcc_12.4' didn't work, but I didn't realize it.
nvccwas no longer being found on the path, causing the test harness to skipGH_000639_nvcc_include_all, and I didn't think to look for this.I can't figure out how
nvccis normally added to the developer command prompt path (i.e. if there's a particular subpackage that will do it, or only the full installation). As a workaround, I'm going to try to manually add it to the path. I also need to install the CUDA runtime libraries.This is a tracking issue to remember what's going on here. I'm not sure if this will be fixed on the CUDA side or the VS side. Either way, my workaround should be removed at that time, and if CUDA is the one that's fixed, we should strongly consider increasing our minimum required CUDA version.
The text was updated successfully, but these errors were encountered: