Implement tensor.size(int) for BatchedTensor (#40028)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40028
We have this call native::size directly. Some alternatives I considered
were:
- Call VariableType::size directly. That seems isomorphic to what we're
doing now.
- when creating a BatchedTensor from a regular tensor, put all of the
keys on that tensor into the BatchedTensor's dispatch key set and use
the dispatcher fallthrough mechanism. That seems weird because
BatchedTensor is a tensor wrapper and also error prone because if
BatchedTensor gets the VariableType key, there's a chance that if
something goes wrong, an autogradmeta gets created on it...
Test Plan: - `./build/bin/vmap_test`
Differential Revision: D22070655
Pulled By: zou3519
fbshipit-source-id: 18530579ad41f3c4f96589da41eb24a46caf7af9