I checked the redistribution code, left-to-right order is not completely correct.
The Partial-> Replicate can be right-to-left order:

As long as we have Shard in src placement, we can trigger right-to-left:

from torch.utils._debug_mode import DebugMode
class DistributeWithPartialTest(DTensorTestBase):
    @property
    def world_size(self) -> int:
        return 8

    def _extract_redistribute_trace_from_debug_mode(self, s: str) -> str:
        import re

        match = re.search(r"trace:\s*(.*)\)", s)
        if match:
            trace_str = match.group(1)
            return trace_str
        else:
            return ""

    @with_comms
    def test_tmp(self):
        mesh = init_device_mesh(self.device_type, (2, 2, 2))
        input_data = torch.randn((8, 8, 8), device=self.device_type)
        dt = DTensor.from_local(input_data, mesh, (Partial('sum'), Shard(0), Partial('max')))
        with DebugMode(record_torchfunction=False) as debug_mode:
            dt2 = dt.redistribute(mesh, (Replicate(), Replicate(), Replicate()))
        trace_str = self._extract_redistribute_trace_from_debug_mode(debug_mode.debug_string())
        print(trace_str)

The redistribution path is: P(sum)S(0)P(max)->P(sum)S(0)R->P(sum)RR->RRR. This made me think that there is a bug with the greedy redistribution algorithm when handling Partial with variant. We can't arbitrarily switch between left-to-right or right-to-left order. I think right-to-left should be the correct order to handle Partial->Replicate.

That sounds good to me. However don't move too fast on it, we are discussing just banning mixed partial types existing in the same placement and then we don't need to define an
order

oh, this PR was not even supposed to include the changes to partial order. it was a rebasing mistake. I will remove that code. I have another PR up now anyway that partially bans mixed partials. If people are OK with that direction, i'll extend that and ensure we don't allow mixed partials at all parts of the DTensor stack. then this ordering stuff becomes unnecessary.

Yes, "banning mixed Partial" sounds good to me!

Aha, I see. I left the comment in README file too early. This PR is enforcing the left-to-right order when there exists multiple Partial variants.

If the condition dst_partial_dims and dst_partial_dims != src_partial_dims doesn't hold, then we will use the default greedy code to handle partial, which can either be left-to-right or reverse. We also need to enforce that order.