I see support for CUDA streams in CUDA ggml implementation (for example, here .cpp/blob/master/ggml/src/ggml-cuda/softmax.cu#L172 and here .cpp/blob/master/ggml/src/ggml-cuda/common.cuh#L674 ), but it is in vain since ggml_backend_cuda_context.stream() always return stream #0: .cpp/blob/master/ggml/src/ggml-cuda/common.cuh#L683

Am I right? There is no way to use CUDA streams in llamacpp?