JuliaPackaging
WIP: [Torch] Renamed Torch to TorchCUDA, and added a CPU-only Torch
Currently a WIP. Aims at resolving https://github.com/FluxML/Torch.jl/issues/20.
Similar approach as for ONNXRuntime - separate binaries for CPU-only and CUDA etc. (CPU: #4369, CUDA: #4386).
Relates to: #1529
Does this need to be a different package? Or can we turn this into a variant?
Does this need to be a different package? Or can we turn this into a variant?
If eventually both CPU, CUDA and ROCm builds are available, wouldn't the different sets of dependencies necessitate different packages?
You still have the same issue that if Torch.jl includes Torch_jll, TorchCUDA_jll,... all dependencies must be fulfilled. Better to use a platform tag and platform specific dependencies.
You still have the same issue that if Torch.jl includes
Torch_jll, TorchCUDA_jll,...all dependencies must be fulfilled. Better to use a platform tag and platform specific dependencies.
Right, I see. Is there some discussion/registry of defined/reserved platform tags? I.e. that "cuda" is reserved for CUDA and so on for e.g. "amd_rocm", "intel_oneapi", "apple_coreml", ...