I’ve been thinking about this same idea recently, so I’m glad it exists now!

The biggest benefit I see is to enable RAM constrained GPUs to perform inference of large parameter models with surprisingly high throughput. Because only a single expert is resident, the memory to compute ratio over the network is limited only by the activations, not the weights. For an Moe like kimi k3 where active parameters are 103B, we might expect to achieve performance limited only by ~5 effective tok/s per Tflop and ~1 tok/s per 100GB/s.

The more members of the network, the smaller your resident parameters are required to be. I’m not sure whether we can split layer inference into arbitrary chunks, but if so you’d be able to increase memory throughput by storing everything in GPU caches.

Of course, we expect latency to be relatively high, but that’s a tradeoff that's fine for certain circumstances.

I’m not sure whether there are any issues more with this idea, but it’s a fun one nonetheless :)