disaggregated inference thoughts on M3 Max

[kernelmike]

that fixed it for me too.

Interesting, I hadn't tried that. 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.