Currently the Python send_recv benchmark does not support AM, this needs to be added. Furthermore, in UCX-Py there is an inconsistency with receiving AM when compared to ucx_perftest, where UCX-Py always reallocates memory regardless of the --reuse-alloc being specified, see rapidsai/ucx-py#1125 (comment) for more details, a way to handle that is to register an AM allocator that always returns the same memory buffer.
Currently the Python
send_recvbenchmark does not support AM, this needs to be added. Furthermore, in UCX-Py there is an inconsistency with receiving AM when compared toucx_perftest, where UCX-Py always reallocates memory regardless of the--reuse-allocbeing specified, see rapidsai/ucx-py#1125 (comment) for more details, a way to handle that is to register an AM allocator that always returns the same memory buffer.