libtorch_c_api not defined in Torch.Wrapper`

[kadir-gunel]

Hello,

I am still unable to install this lib. I have fresh install of Julia v1.11.3.

The main error I am getting is :

 PkgPrecompileError: The following 1 direct dependency failed to precompile:

Torch

Failed to precompile Torch [6a2ea274-3061-11ea-0d63-ff850051a295] to "/home/kguenel/.julia/compiled/v1.11/Torch/jl_c11DHW".
ERROR: LoadError: UndefVarError: `libtorch_c_api` not defined in `Torch.Wrapper`
Stacktrace:
 [1] at_grad_set_enabled(arg1::Base.RefValue{Int32}, arg2::Int64)
   @ Torch.Wrapper ~/.julia/packages/Torch/QTrgw/src/wrapper.jl:15
 [2] top-level scope
   @ ~/.julia/packages/Torch/QTrgw/src/tensor.jl:8
 [3] include(mod::Module, _path::String)
   @ Base ./Base.jl:557
 [4] include(x::String)
   @ Torch ~/.julia/packages/Torch/QTrgw/src/Torch.jl:1
 [5] top-level scope
   @ ~/.julia/packages/Torch/QTrgw/src/Torch.jl:24
 [6] include
   @ ./Base.jl:557 [inlined]
 [7] include_package_for_output(pkg::Base.PkgId, input::String, depot_path::Vector{String}, dl_load_path::Vector{String}, load_path::Vector{String}, concrete_deps::Vector{Pair{Base.PkgId, UInt128}}, source::Nothing)
   @ Base ./loading.jl:2881
 [8] top-level scope
   @ stdin:6
in expression starting at /home/kguenel/.julia/packages/Torch/QTrgw/src/tensor.jl:7
in expression starting at /home/kguenel/.julia/packages/Torch/QTrgw/src/Torch.jl:1
in expression starting at stdin:```

What should I do? 


B.R.

[stemann]

Are you on x86_64 Linux without a CUDA GPU?

If so, then you might be affected by #20 (comment)

[kadir-gunel]

No, I have CUDA gpus. And the driver is also installed. I can use flux with gpu, or CUDA.jl itself.

[stemann]

Alright - you likely need to include CUDA_Runtime_jll in your project/environment, and set the CUDA version preference to a supported CUDA version, cf. https://cuda.juliagpu.org/stable/installation/overview/#Specifying-the-CUDA-version

It should definitely work on CUDA 11.3 (but I hope it also works on CUDA 11.8).

[kadir-gunel]

Based on your suggestions, I found out that CUDA.jl left support for CUDA 11.3 in version 5. And If I want to continue with 11.8 it gives error. But since Torch.jl works with CUDA 11.3, I gave a shot to [email protected] and [email protected] . I also downgraded julia version to 1.8. With this configuration ([email protected], [email protected] with v11.3 and of course Torch.jl), I still get eror :

[Info: Precompiling Torch [ 6a2ea274-3061-11ea-0d63-ff850051a295]
ERROR: LoadError: InitError: could not load library "/home/kg/.julia/artifacts/df751aad1e240ec7b9143977/lib/libtorch.so/"
libnvToolsExt.so.1 : cannot open shared object file : No such file or directory

[kadir-gunel]

Ok, now the issue is not related with Torch.jl . It was due to the missing system package inside Rocky Linux. By installing openblas-dev and lapack packages everything seems to be working properly. I can form n-dimensional tensors in gpu with cuda 11.3 and cuda 11.8.

Any julia package above 1.8.x series to the latest version 1.11.x can work with Torch.jl, no worries.

I really missed this package. It is great to see it working again!

B.R.

[stemann]

Great to hear you got it working!

(though you should not need to adjust any system packages)

[stemann]

It's interesting that you got the issue with libnvToolsExt, as there is (some level of) explicit handling for this in https://github.com/JuliaPackaging/Yggdrasil/blob/7deefacab0cacfe5ccf7d9e437bdeea554dcf1de/T/Torch/build_tarballs.jl#L249

Can you reproduce this error on Julia 1.10 or 1.11? Or is it just on 1.8?

[kadir-gunel]

I think it was general, because after I installed those packages, it stopped giving errors.

Additionally, I am facing issues when try to send resnet example to torch. But I think it is better to create a new issue for that.