[Usage]: does extendGroupSizeTo support sub-groups where local rank ≠ global rank?

[JaqEst]

Describe your usage question

Hi, I was reading the implementation of extendGroupSizeTo and noticed that newly added ranks always get an identity local-to-global mapping:

// mooncake_backend.cpp
for (int i = oldSize; i < newSize; ++i) {
    local2global_rank_map_[i] = i;
}

But the constructor reads global_ranks_in_group from distBackendOpts to build this mapping properly, supporting cases where local rank ≠ global rank.

My question: is extendGroupSizeTo intended to be used only on full-world groups (where local rank always equals global rank)?

I haven't hit an issue in practice since all the existing tests use full-world groups, just want to understand the intended contract before relying on it in a setup where sub-groups might be involved.

Before submitting a new issue...

• Make sure you already searched for relevant issues and read the documentation

[yuechen-sys]

Hi JsqEst,

Thanks for your question. Extending an existing ProcessGroup is an interesting problem. You are correct that the current extension semantics of MooncakeBackend do not properly support sub-groups (created with new_group).

The current design of the Mooncake PG focuses mainly on fault-tolerance and recovery. Supporting split group and correctly handling sub-group extensions is on our roadmap, and we are actively developing the Mooncake PG.