For the record: I oppose adding a second partial state here. We should have 100% coverage for boxing wrappers in c10 if we add it. If you can't handle an op, disable it from c10 instead.

The reason for the two-step process is that it allows us to delete globalATenDispatch earlier. Once all ops are in stage 1, we can already delete it and making them boxed is a problem inside the c10 dispatcher. If we force all ops to immediately be box-able, it will significantly delay the deletion of globalATenDispatch and there will be a longer time we have to deal with multiple dispatchers.

My original plan was what you're describing - keep the c10 dispatcher pure and only move what can be moved, add features and operators step-by-step. But you and @dzhulgakov convinced me that having multiple dispatchers is a major issue for other development flows (like multi dispatch or lazy) and that it's important to get down to only one dispatcher earlier than that and that it's ok if the c10 dispatcher has both unboxed_only and regular ops for a while.

OK

This pull request has been merged in 5016796.