Unable to use local CUDA runtime toolkit

Specific Domains GPU
1

Hi All,

I am trying use local CUDA runtime (12.4) toolkit, but I get the following error.


└ @ CUDA_Runtime_Discovery C:\Users\mohitt\.julia\packages\CUDA_Runtime_Discovery\ZGnCZ\src\CUDA_Runtime_Discovery.jl:139
┌ Debug: Could not discover CUDA toolkit
│   exception =
│    Could not find library 'nvToolsExt' in your local CUDA installation.
│    Stacktrace:
│      [1] error(s::String)
│        @ Base .\error.jl:35
│      [2] get_library(dirs::Vector{String}, name::String; optional::Bool)

I have all the required library installed in CUDA toolkit, and CUDA runtime is able to discover them, except for this one. Here NVTX - NVIDIA Tools Extension Library, it is mentioned that NVTX is no longer a compiled library (it is now header-only). This makes sense, as I could not find Nsight NVTX installation option in the NVIDIA CUDA 12.4 installer.

julia> versioninfo()
Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 16 × 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores
Environment:
  JULIA_DEBUG = CUDA_Runtime_Discovery
  JULIA_DIR = C:\Users\AppData\Local\Programs\julia-1.10.0\lib\julia
  JULIA_IMAGE_THREADS = 4
  JULIA_NUM_THREADS = 1
2

You’re right, this can be removed from CUDA_Runtime_Discovery.jl. Can you file an issue there?

Also, what’s the reason you’re using a local toolkit? This is not recommended.

3

Thank you for the reply, I have filled the issue.

CUDA.jl (always) works well with the runtime artifact. I am exploring the possibility of not downloading the artifact and rather using the local installation, to reduce the size of the compiled app.