CUDA.function() == false;

[usmcamp0811]

Not exactly sure how to use this to create a Nix package that has FHS. When I try to use the example tamplate I can't get CUDA.jl to see my GPU. My GPU works on the system with Julia and CUDA but if I use the Julia that is packaged with Julia2Nix the GPU is not functional.

[samrose]

You probably need to include cudatoolkit

Here is what I had to do to create a dev env shell for pytorch that could access CUDA (for instance) https://gist.github.com/samrose/01fdcc045a262168540ba56ae95d1d26?permalink_comment_id=4816797#gistcomment-4816797

[usmcamp0811]

ah ok after some struggling I got it to work. thanks

[usmcamp0811]

After a lot more mucking around I came to realize that I still can't get CUDA to work with Julia2Nix; however this is probably because I am not building the Project correctly.

Initially I thought I had gotten things to work because I used @samrose 's shell config, but didn't realize it didn't even have Julia in it and so the shell defaulted to my system Julia which I have configured using scientific-fhs.

I am unable to generate the julia2nix.toml due to other errors. I hope that that is the root cause here.

If I try to use the julia-wrapped package the flake provides I get either one of three outcomes depending on how I configure my flake. I either get Julia failing to start because of some pathing issues (sorry I don't have the error log handy) or I get Julia to start but CUDA is not functional or finally I can get Julia to start, CUDA to report functional but I am unable to do operations on the GPU resulting in a ptxas error which I likely because I don't have cuDNN in my path or something along those lines.

ERROR: Failed to compile PTX code (ptxas exited with code 127)
Invocation arguments: --generate-line-info --verbose --gpu-name sm_86 --output-file /tmp/jl_21utE96Vkh.cubin /tmp/jl_StC1SjU3g9.ptx
Could not start dynamically linked executable: /home/mcamp/.config/julia/artifacts/913584335ab836f9781a0325178d0949c193f50b/bin/ptxas
NixOS cannot run dynamically linked executables intended for generic
linux environments out of the box. For more information, see:
https://nix.dev/permalink/stub-ld
If you think this is a bug, please file an issue and attach /tmp/jl_StC1SjU3g9.ptx
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] cufunction_compile(job::GPUCompiler.CompilerJob, ctx::LLVM.ThreadSafeContext)
    @ CUDA ~/.config/julia/packages/CUDA/BbliS/src/compiler/execution.jl:429
  [3] #228
    @ ~/.config/julia/packages/CUDA/BbliS/src/compiler/execution.jl:348 [inlined]
  [4] LLVM.ThreadSafeContext(f::CUDA.var"#228#229"{GPUCompiler.CompilerJob{GPUCompiler.PTXCompilerTarget, CUDA.CUDACompilerParams, GPUCompiler.FunctionSpec{GPUArrays.var"#broadcast_kernel#26", Tuple{CUDA.CuKernelContext, CuDeviceMatrix{Float32, 1}, Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{2}, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, typeof(*), Tuple{Int64, Base.Broadcast.Extruded{CuDeviceMatrix{Float32, 1}, Tuple{Bool, Bool}, Tuple{Int64, Int64}}}}, Int64}}}})
    @ LLVM ~/.config/julia/packages/LLVM/HykgZ/src/executionengine/ts_module.jl:14
  [5] JuliaContext(f::CUDA.var"#228#229"{GPUCompiler.CompilerJob{GPUCompiler.PTXCompilerTarget, CUDA.CUDACompilerParams, GPUCompiler.FunctionSpec{GPUArrays.var"#broadcast_kernel#26", Tuple{CUDA.CuKernelContext, CuDeviceMatrix{Float32, 1}, Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{2}, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, typeof(*), Tuple{Int64, Base.Broadcast.Extruded{CuDeviceMatrix{Float32, 1}, Tuple{Bool, Bool}, Tuple{Int64, Int64}}}}, Int64}}}})
    @ GPUCompiler ~/.config/julia/packages/GPUCompiler/S3TWf/src/driver.jl:74
  [6] cufunction_compile(job::GPUCompiler.CompilerJob)
    @ CUDA ~/.config/julia/packages/CUDA/BbliS/src/compiler/execution.jl:347
  [7] cached_compilation(cache::Dict{UInt64, Any}, job::GPUCompiler.CompilerJob, compiler::typeof(CUDA.cufunction_compile), linker::typeof(CUDA.cufunction_link))
    @ GPUCompiler ~/.config/julia/packages/GPUCompiler/S3TWf/src/cache.jl:90
  [8] cufunction(f::GPUArrays.var"#broadcast_kernel#26", tt::Type{Tuple{CUDA.CuKernelContext, CuDeviceMatrix{Float32, 1}, Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{2}, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, typeof(*), Tuple{Int64, Base.Broadcast.Extruded{CuDeviceMatrix{Float32, 1}, Tuple{Bool, Bool}, Tuple{Int64, Int64}}}}, Int64}}; name::Nothing, always_inline::Bool, kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ CUDA ~/.config/julia/packages/CUDA/BbliS/src/compiler/execution.jl:300
  [9] cufunction
    @ ~/.config/julia/packages/CUDA/BbliS/src/compiler/execution.jl:293 [inlined]
 [10] macro expansion
    @ ~/.config/julia/packages/CUDA/BbliS/src/compiler/execution.jl:102 [inlined]
 [11] #launch_heuristic#252
    @ ~/.config/julia/packages/CUDA/BbliS/src/gpuarrays.jl:17 [inlined]
 [12] launch_heuristic
    @ ~/.config/julia/packages/CUDA/BbliS/src/gpuarrays.jl:15 [inlined]
 [13] _copyto!
    @ ~/.config/julia/packages/GPUArrays/5XhED/src/host/broadcast.jl:65 [inlined]
 [14] copyto!
    @ ~/.config/julia/packages/GPUArrays/5XhED/src/host/broadcast.jl:46 [inlined]
 [15] copy
    @ ~/.config/julia/packages/GPUArrays/5XhED/src/host/broadcast.jl:37 [inlined]
 [16] materialize
    @ ./broadcast.jl:873 [inlined]
 [17] broadcast_preserving_zero_d
    @ ./broadcast.jl:862 [inlined]
 [18] *(A::Int64, B::CuArray{Float32, 2, CUDA.Mem.DeviceBuffer})
    @ Base ./arraymath.jl:21
 [19] top-level scope
    @ REPL[23]:1
 [20] top-level scope