Without circular at first (This PR)
Adding circular
See https://github.com/google/maxtext/pull/677 for details on tests run in addition to the ones added in the PR. Note that without circular pipelining the bubble is usually large (or microbatch_size is small) so hard to get good performance
Full PR with circular support is https://github.com/google/maxtext/pull/677 - we have decided to split up the PR into two
Without circular at first (This PR) Adding circular See https://github.com/google/maxtext/pull/677 for details on tests run in addition to the ones added in the PR. Note that without circular pipelining the bubble is usually large (or microbatch_size is small) so hard to get good performance