In reply to @keiichi:matrix.org
when using localai 2.15 from unstable and even after a reboot I get ggml_cuda_init: failed to initialize CUDA: CUDA driver is a stub library. It's a bit random but if anyone has a tip, I take it. nvidia-smi output looks fine

In reply to @glepage:matrix.org

I am packaging this: https://github.com/EricLBuehler/mistral.rs?tab=readme-ov-file#installation-and-build
You can see that it support several variations for building (CUDA, metal, mkl...)

-> What should be the approach ? Adding cudaSupport ? metalSupport ? mklSupport ?

In reply to @ss:someonex.net
Does it allow enabling multiple features at once?

In reply to @gjvnq:matrix.org
Hey, can I ask for help with compiling alice-vision on NixOS?

In reply to @ss:someonex.net
Looking more closely, I'd guess the issue is somewhere around __has_include(<Imath/half.h>) in ${openimageio.dev}/include/OpenImageIO/half.h

Yeah, I had already figured out it but the bug issue is that I don't know what is the "right" way to include the definition of the half type.

To make matters worse, I've tried to compile AliceVision on a docker container and using the official compilation scripts and yet the thing keeps failing. This means I can't even look at how the thing is supposed to compile.

I'm at a bit of a loss as for how to proceed, but I suspect that I'll have to either ask the original authors for help or carefully read the cmake compilation scripts in order to look for potential sources of the error.

Theoretically AliceVision has a nice CI pipeline but I can't see their build history so I don't even know how useful their CI scripts are.

In reply to @keiichi:matrix.org
I dont seem to have any cuda library in /run/opengl-driver/. Should I add anything into hardware.opengl.extraPackages ?

Huh, not sure how that's working for you since I can't get it to build:

$ cat ~/.config/nixpkgs/config.nix 
{
  allowAliases = false;
  allowBroken = false;
  allowUnfree = true;
  checkMeta = true;
  cudaCapabilities = [ "7.5" ];
  cudaSupport = true;
}
$ nix run --impure -L --show-trace --builders '' --max-jobs 1 github:nixos/nixpkgs/c7b821ba2e1e635ba5a76d299af62821cbcb09f3#local-ai
...
go-stable-diffusion> CMake Error at /nix/store/q1nssraba326p2kp6627hldd2bhg254c-cmake-3.29.2/share/cmake-3.29/Modules/FindCUDA.cmake:883 (message):
go-stable-diffusion>   Specify CUDA_TOOLKIT_ROOT_DIR
go-stable-diffusion> Call Stack (most recent call first):
go-stable-diffusion>   /nix/store/r008mflixfchlfscby4h0mjgqvz059pw-opencv-4.9.0/lib/cmake/opencv4/OpenCVConfig.cmake:86 (find_package)
go-stable-diffusion>   /nix/store/r008mflixfchlfscby4h0mjgqvz059pw-opencv-4.9.0/lib/cmake/opencv4/OpenCVConfig.cmake:108 (find_host_package)
go-stable-diffusion>   examples/CMakeLists.txt:17 (find_package)

In reply to @trexd:matrix.org
Do you guys have any recommendations on setting up a cache for https://github.com/hasktorch/hasktorch-skeleton/pull/9 ? The haskell part of the build already takes quite a long time but throw CUDA in the mix and I think build times would quickly get out of hand. Either way I've also never setup a cache via cachix or hercules-ci before so I don't know what the limits on build times are either.