What hardware and inference engine do you use? On my Radeon R9700, at least with llama.cpp and some others, the GPU provides as many tok/sec with a single stream as with concurrent ones, so it's just as fast and supports longer contexts to swap contexts to RAM on turn boundaries.