Support cyclic dependency graph

[junrushao-openai]

A common usecase in ML compiler we have is to depict producer-consumer relationship as a dependency graph. This graph will generally contain cycles because each node will have to keep track of both its producers and consumers. Is there a way to do this in MLC?

[potatomashed]

Hmm there's no fundamental challenge supporting this particular data structure in Python if we handle them carefully. Could you elaborate and share examples which problems you encounter?

A common foot gun is that one should almost certainly use mlc_object._mlc_address instead of id(mlc_object). I was thinking to introduce a shorter method mlc_object.id() as a shortcut, but wasn't quite sure about naming.

It's common utility though in a compiler infra to keep track of producer-consumer relationship, and the latest torch.fx.Graph seem to design the API moderately well. Seems interesting thing to do to replicate its features in mlc core :P

[junrushao-openai]

Yes I think fx.Graph API looks sufficiently reasonable

[potatomashed]

Fixed by #65

[junrushao]

I tested the performance of this implementation on a graph that consists of 1M nodes:

• My own [link] implementation: 13.02 sec
• MLC’s implementation: 1.23 sec
• Benchmark script [link]