Summary: This causes issues with CUDA's teardown order when the init is separated from the total init scope.
OffloadContext