ONNX Interface
For the purpose of prototyping, we're currently using pre-compiled TVM models. However in practice, our desired interface is an ONNX model specification, and compilation to TVM models should happen by Clockwork.
For the purpose of research and experimentation, this is a "nice to have" feature, but is not yet necessary.
For the purpose of the system being used in the wild, this is a necessary feature.
Questions to consider when implementing this include:
- Should we run the TVM optimization step?
- Should we reuse the TVM optimization log, since there is a lot of overlap between models in our modelzoo currently