Any chance to bump the CUDA compatibility to 5? I'm seeing no issues in my particular use case at least.