Open cowanmeg opened 1 year ago
Can this be closed?
No, the current implementation falls back to the cpu when replacement is False.
@cowanmeg, I'll assign this to you as part of cleaning up our unassigned issues. No immediate need to work on this as far as I can tell, so feel free to put it in your backlog and prioritize as necessary.
🚀 Feature
Lower multinomial without replacement.
Motivation
Currently, we lower multinomial with replacement (https://github.com/pytorch/xla/pull/4848), but we want to accelerate both paths.
Some relevant discussion on tensorflow that might be helpful: https://github.com/tensorflow/tensorflow/issues/9260