Calling Flux.gpu downloads CUDNN artifact every first call in REPL/script

Specific Domains GPU
,
1

Pretty much as the title says. Whenever I call e.g.

Import Flux
d_gpu = Flux.gpu(Flux.Dense(10, 10))
┌ Warning: The NVIDIA driver on this system only supports up to CUDA 11.0.0.
│ For performance reasons, it is recommended to upgrade to a driver that supports CUDA 11.2 or higher.
└ @ CUDA ~/.julia/packages/CUDA/DfvRa/src/initialization.jl:70
│ Downloading artifact: CUDNN

whether in REPL or a script it starts downloading a CUDNN artifact (doesn’t show any version information).

Regarding package information. ] status returns:

Status `~/.julia/environments/v1.8/Project.toml`
  [6e4b80f9] BenchmarkTools v1.3.1
  [052768ef] CUDA v3.12.0
  [aae7a2af] DiffEqFlux v1.52.0
  [0c46a032] DifferentialEquations v7.4.0
  [31c24e10] Distributions v0.25.75
  [587475ba] Flux v0.13.6
  [c8e1da08] IterTools v1.4.0
  [033835bb] JLD2 v0.4.23
  [d3d80556] LineSearches v7.2.0
  [b2108857] Lux v0.4.24
  [961ee093] ModelingToolkit v8.23.0
  [76087f3c] NLopt v0.6.5
  [872c559c] NNlib v0.8.9
  [a00861dc] NNlibCUDA v0.2.4
  [429524aa] Optim v1.7.3
  [3bd65402] Optimisers v0.2.9
  [7f7a1694] Optimization v3.9.0
  [4e6fcdb7] OptimizationNLopt v0.1.0
  [36348300] OptimizationOptimJL v0.1.2
  [42dfb2eb] OptimizationOptimisers v0.1.0
  [842ac81e] OptimizationQuadDIRECT v0.1.0 `https://github.com/SciML/Optimization.jl:lib/OptimizationQuadDIRECT#master`
  [d330b81b] PyPlot v2.11.0
  [dae52e8d] QuadDIRECT v0.1.2 `https://github.com/timholy/QuadDIRECT.jl.git#master`
  [2913bbd2] StatsBase v0.33.21
  [d1185830] SymbolicUtils v0.19.11
  [0c5d862f] Symbolics v4.10.4
  [37e2e46d] LinearAlgebra

and nvidia-smi shows CUDA Version: 11.0 (this one’s out of my control). How can I avoid the download of the artifact every time? Any help is appreciated, if you need more information - let me know and I’ll do what I can.

2

Try running with JULIA_DEBUG=CUDA, that should print out some more details:

3

Is that in the REPL? As I’ve just tried that and get

julia> import CUDA
julia> JULIA_DEBUG=CUDA
CUDA
julia> import Flux
julia> a = Flux.Dense(10, 10)
Dense(10 => 10)     # 110 parameters
julia> b = Flux.gpu(a)
│┌ Warning: The NVIDIA driver on this system only supports up to CUDA 11.0.0.
││ For performance reasons, it is recommended to upgrade to a driver that supports CUDA 11.2 or higher. │  Downloaded artifact: CUDNN
│  Downloaded artifact: CUDNN
│┌ Warning: CUDA.jl found cuda, but did not find libcudnn. Some functionality will not be available.
│└ @ Flux ~/.julia/packages/Flux/4k0Ls/src/functor.jl:195
│Dense(10 => 10)     # 110 parameters

Hope that helps.

4

No, that’s an environment flag you need to set before starting Julia.

5

Ah, thank you, my apologies, I was not aware of that. I tried it and it doesn’t seem to give any extra information, at least based on the calls that I’ve made after setting it.

6

Can you post the output of CUDA.versioninfo()?

7

Ah, thank you, good suggestion, got some more weirdness!

julia> import CUDA
julia> CUDA.versioninfo()
┌ Warning: The NVIDIA driver on this system only supports up to CUDA 11.0.0.
│ For performance reasons, it is recommended to upgrade to a driver that supports CUDA 11.2 or higher.
└ @ CUDA ~/.julia/packages/CUDA/DfvRa/src/initialization.jl:70
CUDA toolkit 11.7, artifact installation
NVIDIA driver 450.203.3, for CUDA 11.0
CUDA driver 11.0

Libraries: 
- CUBLAS: 11.10.1
- CURAND: 10.2.10
- CUFFT: 10.7.2
- CUSOLVER: 11.3.5
- CUSPARSE: 11.7.3
- CUPTI: 17.0.0
- NVML: 11.0.0+450.203.3
  Downloaded artifact: CUDNN
  Downloaded artifact: CUDNN
- CUDNN: missing
  Downloaded artifact: CUTENSOR
- CUTENSOR: 1.4.0 (for CUDA 11.5.0)

Toolchain:
- Julia: 1.8.0
- LLVM: 13.0.1
- PTX ISA support: 3.2, 4.0, 4.1, 4.2, 4.3, 5.0, 6.0, 6.1, 6.3, 6.4, 6.5, 7.0
- Device capability support: sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61, sm_62, sm_70, sm_72, sm_75, sm_80

1 device:
  0: GeForce RTX 2060 (sm_75, 5.708 GiB / 5.795 GiB available)