Chase Riley Roberts

Great! We should allow 0 as a dangling leg then. Erroring out when seeing multiple 0s should fix any (unlikely) porting issues.

Numpy array should always work given you use `backend.convert_to_tensor`. Nested lists behave differently for each backend so that might be harder.

Agreed. I think we'd be ok to to fall back on however the underlying array does it's `__repr__` + some extra info?

Yeah that seems like it would be difficult to add. Is there any reason you can't just manually reuse nodes? It's not automated, but that should work?

Eh, sounds more like a pytorch issue than an us issue. I assume you hit this from writing a test?

Indexing outside of the shape of the array will throw NaNs silently in JAX, that's likely the issue.

Let me try and solve this one. I haven't had to do a bug fix in a while (Because everything we write is so great!).

Great idea! If you write a benchmark I can run it on a V100 like last time.

I took a look, I'm not sure I see the value in this. In what situations does this array class help?