try CUDA.synchonize

Author: simonbyrne

src/cuda.jl

@@ -1,6 +1,7 @@
 import .CUDA
 
 function Base.cconvert(::Type{MPIPtr}, buf::CUDA.CuArray{T}) where T
+    CUDA.synchronize(CUDA.CuDefaultStream())
     Base.cconvert(CUDA.CuPtr{T}, buf) # returns DeviceBuffer
 end
 

test/runtests.jl

@@ -4,6 +4,8 @@ using Test, MPI
 using DoubleFloats
 if get(ENV,"JULIA_MPI_TEST_ARRAYTYPE","") == "CuArray"
     import CUDA
+    CUDA.version()
+    CUDA.precompile_runtime()
     ArrayType = CUDA.CuArray
 else
     ArrayType = Array