[Test] Skip (again) CUDA integration tests when they are known to fail

[giordano]

CUDA integration tests are still failing on non-CUDA systems in Julia v1.11 due to #614, but they were brought back in #667. With this change we skip running those tests when we expect them to fail.

Note: as mentioned in #614 (comment) these tests should finally be working again in Julia v1.11.4.

[wsmoses]

I think we should go for a different fix here and instead change the precompile reactantcudaext.jl to only run on 11.4+, which I believe should fix this while still running tests

[wsmoses]

The Apple cuda tests are currently only disabled pending JuliaGPU/GPUCompiler.jl#660

[giordano]

I think we should go for a different fix here and instead change the precompile reactantcudaext.jl to only run on 11.4+, which I believe should fix this while still running tests

OK, I can confirm

diff --git a/ext/ReactantCUDAExt.jl b/ext/ReactantCUDAExt.jl
index 391be0b..d8ccbec 100644
--- a/ext/ReactantCUDAExt.jl
+++ b/ext/ReactantCUDAExt.jl
@@ -1,5 +1,7 @@
 module ReactantCUDAExt
 
+__precompile__(!(v"1.11.0" <= VERSION < v"1.11.4"))
+
 using CUDA
 using Reactant:
     Reactant, TracedRArray, AnyTracedRArray, AnyConcreteRArray, MLIR, TracedRNumber

would also make this work on Julia v1.11.3, the annoying thing is that it makes loading Reactant a lot slower for everybody because the extension needs to be precompiled all times on load, and this seems to be quite slow:

julia> using CUDA

julia> @time @eval using Reactant
AssertionError("Could not find registered platform with name: \"cuda\". Available platform names are: ")
Precompiling ReactantCUDAExt...
  ? Reactant → ReactantCUDAExt
[ Info: Precompiling ReactantCUDAExt [91356394-6239-50cb-9e6b-6f1f3f798cde] (cache misses: incompatible header (4), mismatched flags (2))
┌ Info: Skipping precompilation due to precompilable error. Importing ReactantCUDAExt [91356394-6239-50cb-9e6b-6f1f3f798cde].
└   exception = Error when precompiling module, potentially caused by a __precompile__(false) declaration in the module.
 32.698447 seconds (39.38 M allocations: 2.111 GiB, 4.15% gc time, 92.60% compilation time: 83% of which was recompilation)

I'm happy to make this change if you think it's worth it, but I outlined the drawback!

[wsmoses]

@giordano ah I didn’t mean that I meant this:

Reactant.jl/ext/ReactantCUDAExt.jl

Line 955 in 1e6037f

@static if !Sys.isapple() && Sys.ARCH != :aarch64

[wsmoses]

Also incidentally we can now remove the aarch64 and is functional checks now that CPU codegen works (though obviously it will still require an unreleased Julia)