ctranslate2: don't set default nvcc arch flags

[danielfullmer]

This fixes a build failure when using cudaPackages_13:

Unsupported gpu architecture 'compute_53'

Build failure originally reproduced with:

nix build -f =(echo 'with (import (builtins.getFlake "nixpkgs/nixos-unstable")  { config = { cudaSupport = true; allowUnfree=true; }; }).extend (final: prev: { cudaPackages = final.cudaPackages_13; }); ctranslate2' )

Things done

• Built on platform:
  • x86_64-linux
  • aarch64-linux
  • aarch64-darwin
• Tested, as applicable:
  • NixOS tests in nixos/tests.
  • Package tests at passthru.tests.
  • Tests in lib/tests or pkgs/test for functions and "core" functionality.
• Ran nixpkgs-review on this PR. See nixpkgs-review usage.
• Tested basic functionality of all binary files, usually in ./result/bin/.
• Nixpkgs Release Notes
  • Package update: when the change is major or breaking.
• NixOS Release Notes
  • Module addition: when adding a new NixOS module.
  • Module update: when the change is significant.
• Fits CONTRIBUTING.md, pkgs/README.md, maintainers/README.md and other READMEs.
• Follows the automation/AI policy.

[ConnorBaker]

How are the desired set of CUDA capabilities passed to CMake?

I would have expected setting CUDA_ARCH_LIST in cmakeFlagsArray (contains whitespace, cmakeFlags if structured attributes are enabled) using some of the values from cudaPackages.flags.

[danielfullmer]

They use the deprecated FindCuda cmake module, specifically the cuda_select_nvcc_arch_flags() described here: https://cmake.org/cmake/help/latest/module/FindCUDA.html

https://github.com/OpenNMT/CTranslate2/blob/e23a925508612a754e46242cb07a6c84bef56521/CMakeLists.txt#L541

As far as I can see, the format cuda_select_nvcc_arch_flags() accepts doesn't match any currently available thing under flags.*. cmakeCudaArchitecturesString is for CMAKE_CUDA_ARCHITECTURES, which is slightly different than cuda_select_nvcc_arch_flags. I could try to come up with some logic to set cuda_select_nvcc_arch_flags correctly, but since it's deprecated, I didn't want to put too much unnecessary effort into it if it's going away.

By the way, is there a standard for when/if we should set the CUDA architectures in build systems? Is it just if it's convenient and someone remembers to? Or are you trying to be more complete?

[ConnorBaker]

This should work:

cmakeFlags = ... ++ [ (lib.cmakeFeature "CUDA_ARCH_LIST" (builtins.toString cudaPackages.flags.cudaCapabilities) ]

(use cudaPackages.flags.cudaCapabilities instead of config.cudaCapabilities because the default set of capabilities depends on the version of the CUDA package set being used and so can't be populated as a global default value; also the order of resolution of config in the fixed point is odd).

CUDA architectures should always be provided to the build system when CUDA support is enabled.

[danielfullmer]

I get the following error when setting CUDA_ARCH_LIST:

CMake Error at /nix/store/k1b82rs72af124wmq6nb00nhw5i7ikp7-cmake-4.1.2/share/cmake-4.1/Modules/FindCUDA/select_compute_arch.cmake:245 (message):
  Unknown CUDA Architecture Name 10.0 in CUDA_SELECT_NVCC_ARCH_FLAGS
Call Stack (most recent call first):
  CMakeLists.txt:541 (cuda_select_nvcc_arch_flags)


CMake Error at /nix/store/k1b82rs72af124wmq6nb00nhw5i7ikp7-cmake-4.1.2/share/cmake-4.1/Modules/FindCUDA/select_compute_arch.cmake:249 (message):
  arch_bin wasn't set for some reason
Call Stack (most recent call first):
  CMakeLists.txt:541 (cuda_select_nvcc_arch_flags)


CMake Error at /nix/store/k1b82rs72af124wmq6nb00nhw5i7ikp7-cmake-4.1.2/share/cmake-4.1/Modules/FindCUDA/select_compute_arch.cmake:245 (message):
  Unknown CUDA Architecture Name 10.3 in CUDA_SELECT_NVCC_ARCH_FLAGS
Call Stack (most recent call first):
  CMakeLists.txt:541 (cuda_select_nvcc_arch_flags)


CMake Error at /nix/store/k1b82rs72af124wmq6nb00nhw5i7ikp7-cmake-4.1.2/share/cmake-4.1/Modules/FindCUDA/select_compute_arch.cmake:249 (message):
  arch_bin wasn't set for some reason
Call Stack (most recent call first):
  CMakeLists.txt:541 (cuda_select_nvcc_arch_flags)


CMake Error at /nix/store/k1b82rs72af124wmq6nb00nhw5i7ikp7-cmake-4.1.2/share/cmake-4.1/Modules/FindCUDA/select_compute_arch.cmake:245 (message):
  Unknown CUDA Architecture Name 12.0 in CUDA_SELECT_NVCC_ARCH_FLAGS
Call Stack (most recent call first):
  CMakeLists.txt:541 (cuda_select_nvcc_arch_flags)


CMake Error at /nix/store/k1b82rs72af124wmq6nb00nhw5i7ikp7-cmake-4.1.2/share/cmake-4.1/Modules/FindCUDA/select_compute_arch.cmake:249 (message):
  arch_bin wasn't set for some reason
Call Stack (most recent call first):
  CMakeLists.txt:541 (cuda_select_nvcc_arch_flags)


CMake Error at /nix/store/k1b82rs72af124wmq6nb00nhw5i7ikp7-cmake-4.1.2/share/cmake-4.1/Modules/FindCUDA/select_compute_arch.cmake:245 (message):
  Unknown CUDA Architecture Name 12.1 in CUDA_SELECT_NVCC_ARCH_FLAGS
Call Stack (most recent call first):
  CMakeLists.txt:541 (cuda_select_nvcc_arch_flags)


CMake Error at /nix/store/k1b82rs72af124wmq6nb00nhw5i7ikp7-cmake-4.1.2/share/cmake-4.1/Modules/FindCUDA/select_compute_arch.cmake:249 (message):
  arch_bin wasn't set for some reason
Call Stack (most recent call first):
  CMakeLists.txt:541 (cuda_select_nvcc_arch_flags)

It looks like FindCUDA doesn't even accept double digit major version numbers based on the regex here. I assume they just arent putting in any additional effort to FindCUDA given that it's deprecated.

[danielfullmer]

Also, it looks like there's ongoing work to migrate to the newer cmake support here: OpenNMT/CTranslate2#2067

[mweinelt]

Based on the build failure issues linked I'm inclined to revert.

[mweinelt]

Feel free to resubmit when the CUDA issues are settled.