Currently the migraphx torch.compile backend relies on aot_export_joint_simple to generate an aten graph that can be interpreted to migraphx IR. There are some limitations to this API, particularly for handling graphs that mutate input tensors.
The AotAutograd plugin (used by Inductor) is better built to handle these cases. We should create a migraphx backend that uses aot autograd for generating the aten IR. Initially, this can be an optional/experimental backend to only be used with particular models that require this functionallity.