How to manually install artifacts

I am installing Flux.jl and I have very slow access to github (~50kB/s). Now after installing the package (slow, but acceptable), it starts downloading artifact CUDA10.2. The problem is CUDA10.2 package is around 0.5GB, and could take hours to download, not to mention the risk of unstable Internet.

I have downloaded CUDA.v10.2.89.x86_64-w64-mingw32.tar.gz with my browser, the problem is what can I do so that the package installer can recognize it?

I tried reading https://julialang.github.io/Pkg.jl/dev/artifacts/#Overriding-artifact-locations-1, but it does not offer a lot of help: 1. I do not know for sure which specific artifact it is downloading (in terms of content hash or something) as Julia REPL only says “Downloading artifact: CUDA10.2”; 2. the page requires a lot of knowledge of package manager to understand.
https://discourse.julialang.org/t/install-binary-package-such-as-openblas-jll-via-very-slow-internet-connection/36335/14 has some discussion, too, but seems specific to open BLAS.

Have you seen the docs here? https://juliagpu.gitlab.io/CUDA.jl/installation/overview/#Local-installation-1

(I haven’t tried it myself but it looks relevant)

1 Like

I defined the variable to stop downloading, then possibly more trouble comes:

julia> CuArrays.version()
┌ Warning: You are using CUDNN 7.6.5 for CUDA 10.2.0 with CUDA toolkit 10.0.130; these might be incompatible.
└ @ CuArrays C:\Users\melonedo\.julia\packages\CuArrays\9n5uC\src\CuArrays.jl:142

(CUDA_PATH has already been set when I installed TensorFlow)