Using Torch.jl without a GPU?

[DilumAluthge]

Would it be possible to add support for using Torch.jl on a machine without a GPU?

[DilumAluthge]

For example, this is what I currently get when I try to run import Torch in a machine without a GPU:

julia> import Torch
[ Info: Precompiling Torch [6a2ea274-3061-11ea-0d63-ff850051a295]
ERROR: LoadError: LoadError: could not load library "libdoeye_caml"
dlopen(libdoeye_caml.dylib, 1): image not found
Stacktrace:
 [1] macro expansion at /Users/dilum/.julia/packages/Torch/Q8Y45/src/error.jl:12 [inlined]
 [2] at_grad_set_enabled(::Int64) at /Users/dilum/.julia/packages/Torch/Q8Y45/src/wrap/libdoeye_caml_generated.jl:70
 [3] top-level scope at /Users/dilum/.julia/packages/Torch/Q8Y45/src/tensor.jl:6
 [4] include(::Function, ::Module, ::String) at ./Base.jl:380
 [5] include at ./Base.jl:368 [inlined]
 [6] include(::String) at /Users/dilum/.julia/packages/Torch/Q8Y45/src/Torch.jl:1
 [7] top-level scope at /Users/dilum/.julia/packages/Torch/Q8Y45/src/Torch.jl:25
 [8] include(::Function, ::Module, ::String) at ./Base.jl:380
 [9] include(::Module, ::String) at ./Base.jl:368
 [10] top-level scope at none:2
 [11] eval at ./boot.jl:331 [inlined]
 [12] eval(::Expr) at ./client.jl:467
 [13] top-level scope at ./none:3
in expression starting at /Users/dilum/.julia/packages/Torch/Q8Y45/src/tensor.jl:6
in expression starting at /Users/dilum/.julia/packages/Torch/Q8Y45/src/Torch.jl:25
ERROR: Failed to precompile Torch [6a2ea274-3061-11ea-0d63-ff850051a295] to /Users/dilum/.julia/compiled/v1.6/Torch/2cR1S_xGAhl.ji.
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] compilecache(::Base.PkgId, ::String) at ./loading.jl:1290
 [3] _require(::Base.PkgId) at ./loading.jl:1030
 [4] require(::Base.PkgId) at ./loading.jl:928
 [5] require(::Module, ::Symbol) at ./loading.jl:923

[DhairyaLGandhi]

Yes, this is due to using the GPU binaries from torch and having the wrapper rely on cuda being available. We just need to build a Torch_cpu artifact to do this correctly.

[Sundaravelpandian]

@DhairyaLGandhi Is the following error also linked to the absence of GPU?

`julia> using Torch
[ Info: Precompiling Torch [6a2ea274-3061-11ea-0d63-ff850051a295]
ERROR: InitError: could not load library "/home/ssing/.julia/artifacts/d6ce2ca09ab00964151aaeae71179deb8f9800d1/lib/libtorch.so"

/home/ssing/.julia/artifacts/d6ce2ca09ab00964151aaeae71179deb8f9800d1/lib/libtorch.so: undefined symbol: _ZN3c1016C10FlagsRegistryB5cxx11Ev
Stacktrace:
[1] dlopen(::String, ::UInt32; throw_error::Bool) at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.4/Libdl/src/Libdl.jl:109
[2] dlopen at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.4/Libdl/src/Libdl.jl:109 [inlined] (repeats 2 times)
[3] init() at /home/ssing/.julia/packages/Torch_jll/sFQc0/src/wrappers/x86_64-linux-gnu-cxx11.jl:64
[4] _include_from_serialized(::String, ::Array{Any,1}) at ./loading.jl:697
[5] _require_search_from_serialized(::Base.PkgId, ::String) at ./loading.jl:781
[6] _tryrequire_from_serialized(::Base.PkgId, ::UInt64, ::Nothing) at ./loading.jl:712
[7] _require_from_serialized(::String) at ./loading.jl:743
[8] _require(::Base.PkgId) at ./loading.jl:1039
[9] require(::Base.PkgId) at ./loading.jl:927
[10] require(::Module, ::Symbol) at ./loading.jl:922
during initialization of module Torch_jll`

[freddycct]

I mainly develop on my laptop, it doesn't have a cuda gpu, so.... i can't import this unless the CPU version is compiled for.

[ordicker]

I tried to compile cpu-only version of pytorch and Torch.jl, using the wizard.
I have download this pytorch version: https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-1.5.1%2Bcpu.zip

I followed this "script": https://github.com/JuliaPackaging/Yggdrasil/blob/master/T/Torch/build_tarballs.jl

I ran this line:
cmake -DCMAKE_PREFIX_PATH=$prefix -DTorch_DIR=$prefix/share/cmake/Torch ..

but got this error
Re-run cmake no build system arguments CMake Error at /usr/share/cmake/Modules/FindPackageHandleStandardArgs.cmake:137 (message): Could NOT find CUDNN (missing: CUDNN_INCLUDE_DIR CUDNN_LIBRARIES) Call Stack (most recent call first): /usr/share/cmake/Modules/FindPackageHandleStandardArgs.cmake:378 (_FPHSA_FAILURE_MESSAGE) FindCUDNN.cmake:6 (find_package_handle_standard_args) CMakeLists.txt:6 (find_package)

[DhairyaLGandhi]

You might need to make it so that the CUDA deps don't leak into the build process. The CMakeLists, and the CUDA related packages would need to be put behind a boolean GPU flag.

[ordicker]

I get that, but I don't understand which line caused the CUDA deps.

[stemann]

This will need the Torch_jll package to be split into Torch_jll (CPU), and Torch_CUDA_jll: JuliaPackaging/Yggdrasil#9785 (comment)

Likewise for TorchCAPI_jll.