This is awesome! Thank you very much. I'll wait for the device then. 😎

It seems, that this is indeed impossible due to MLIR dialect for Structured Control Flow. https://llvm.discourse.group/t/scf-whileop-type-inference-with-dynamic-shapes/2529

The problem is with broadcasted template. Tensor with "static" shape works OK.

extension MPSDataType {     func description() -> String {         switch(self) {         case .float32:             return "f32"         case .int32:             return "i32"         // other...         default:             return "?"         }     } }

Thank you, AncientCoder. I had a different use case in mind. You proposed using pre-made architectures that can be trained on the device (classifiers, regressors, etc.), but I meant something much lower-level than this. For example, I would like to be able to create *.mlmodel that takes 2 inputs and just adds them together. Or any architecture using available ops. I can write MIL definition, but there is no way of turning it to the mlmodel/mlpackage in Swift, only using coremltools in Python. It seems that MPSGraph is probably the best fit for my needs right now.