it's not a matter of performance but expressibility. const y = await remote_model_a(x) // different machine const z = await remote_model_b(y) // different

it's not a matter of performance but expressibility. const y = await remote_model_a(x) // different machine const z = await remote_model_b(y) // different machine await z.backward() is trivially pipelined when run in parallel. With multithreading suddenly the backing C++ library has to be aware of this and figure things out for you