Do not initialize CUDA during precompilation

[maleadt]

This is wrong:

Flux.jl/src/losses/Losses.jl

Line 27 in 8a5b977

if CUDA.functional() include("ctc-gpu.jl") end

This should be guarded by a check if !precompiling:

Flux.jl/src/Flux.jl

Lines 59 to 61 in 8a5b977

	function __init__()
	use_cuda[] = CUDA.functional() # Can be overridden after load with `Flux.use_cuda[] = false`
	if CUDA.functional()

Additionally, this shouldn't be done during __init__, but lazily when first calling the gpu function.

[DhairyaLGandhi]

We already use the flag in gpu as well. Are you saying we skip this check here entirely and have gpu set the flag on first call and then use that consequently? Currently it's hidden behind a Ref{Bool} so we need something to say "hasn't been initialised yet", with missing or nothing

[maleadt]

I think that would be better for users with a GPU, but who don't plan to use it (so that using Flux doesn't start downloading the large CUDA artifacts). I'll explore that in the multi GPU back-end PR.

[darsnack]

In lieu of the long term fix, should we release a quick fix for the losses issue? Would using use_cuda[] instead of CUDA.functional() work?

[DhairyaLGandhi]

It's the same either way. Fixing the init would be better.

[IanButterworth]

It'd be nice to have this fixed. Downloading during precompilation is pretty awkward with the new Pkg auto precompilation, especially on a slow connection given the output during precomp is suppressed, so it's not clear that a download is happening

[IanButterworth]

Friendly bump on this