Warning: Package cuDNN not found in current path

Specific Domains GPU
,
1

I get this warning while using Flux on a CUDA GPU:

┌ Warning: Package cuDNN not found in current path.
│ - Run `import Pkg; Pkg.add("cuDNN")` to install the cuDNN package, then restart julia.
│ - If cuDNN is not installed, some Flux functionalities will not be available when running on the GPU.
└ @ FluxCUDAExt ~/.julia/packages/Flux/u7QSl/ext/FluxCUDAExt/FluxCUDAExt.jl:56

I tried pkg> add cuDNN, but nothing happened. Plus, I thought cuDNN was included in CUDA. What should I do?

The script generates the above warning (and other errors) in the following environment:

julia> VERSION
v"1.9.3"

(@v1.9) pkg> st CUDA
Status `~/.julia/environments/v1.9/Project.toml`
  [052768ef] CUDA v5.0.0

(@v1.9) pkg> st Flux
Status `~/.julia/environments/v1.9/Project.toml`
  [587475ba] Flux v0.14.6

The same script runs fine in an older environment:

julia> VERSION
v"1.8.5"

(@v1.8) pkg> st CUDA
Status `~/.julia/environments/v1.8/Project.toml`
⌅ [052768ef] CUDA v4.0.1

(@v1.8) pkg> st Flux
Status `~/.julia/environments/v1.8/Project.toml`
⌅ [587475ba] Flux v0.13.14
1 Like
2

I found a relevant part in the Flux documentation:

NVIDIA GPU support requires the packages CUDA.jl and cuDNN.jl to be installed in the environment. In the julia REPL, type ] add CUDA, cuDNN to install them.

Not sure what was wrong, but for some reason nothing happened when I tried to install cuDNN only by pkg> add cuDNN at the time of the original posting. After finding out the above part in the documentation, I executed pkg> add CUDA, cuDNN as instructed, and it started installing cuDNN.

1 Like
3

I didn’t find any documentation for cuDNN.jl. Do you see that? I even check on JuliaGPU but no hope. Here is the best I can, but CUDNN seems different to cuDNN.

4

Then you must have cuDNN.jl installed already.

cuDNN.jl is part of the CUDA.jl repository, https://github.com/JuliaGPU/CUDA.jl/tree/master/lib/cudnn.

5

I (and colleagues) regularly get this message as well when using a package that in turns uses cuDNN (and has it in its dependencies).
pkg> st -m also displays cuDNN.

6

Yeah I see this if a dependency brings in CUDA and cuDNN is not imported at the current directory top level (even if the dependency also brings in cuDNN and it’s in the Manifest.toml). This ends up being super annoying in Distributed contexts as it can print repeatedly in Distributed tasks.

I haven’t found an easy way to suppress the warning though..