Specifiying the hardware topology

[tle-huu]

Hello,

I am trying to understand how IREE calculates the scheduling strategy and dispatches work on multiple devices.

If I understand correctly, by specifying multiple targets, let's say [llvm-cpu, vulkan], IREE will generate a CPU and a GPU executables for each operation.

What I have difficulty to understand is how does IREE decide which device to schedule the operation on.
More specifically, how / when does IREE know the number of different devices (n CPUs and k GPUs)?
Is it at the Stream level or is it decided dynamically at runtime?

Is it possible to specify to the compiler the different target types as well as the exact number of devices for each device type?

Thank you