when a plain Python stack is enough for Qwen

[fastbench]

That matches what I was seeing. The pattern around MoE serving has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.