[Connor-XY]

Why. FullyShardedDataParallelV1 fills an unset tp or expt_tp in pg_collection with
dist.new_group(ranks=[rank]), a different rank list on every rank. Group creation must be identical on
all ranks: this works only while every rank takes the same branch, and otherwise later groups mismatch or
hang. Dense models also create an expert TP group they never use, and a collection that lacks a needed
group fails with an AttributeError on None deep in mesh construction.

What.

• Create single-rank TP groups once, with new_subgroups_by_enumeration over every rank, and only for
  the axes the model uses.
• Check the groups the configuration needs (dp_cp; intra_dp_cp, inter_dist_opt for HSDP; ep,
  expt_dp, intra_expt_dp for MoE) and raise ValueError naming the missing ones.
• Read collection fields directly. Comment the global branch as the migration fallback; its groups are
  unchanged.

The collection path keeps the replicate data-parallel fields. With GTP weight remat off they hold the
same ranks as the fallback's GTP-inclusive groups.

Tests. New test_mcore_fsdp_dist_index.py (4 GPUs): a HyperCommGrid collection laid out unlike the global grid,
with every parallel_state accessor patched to raise; identical group creation on every rank;
ValueError for missing groups; explicit-vs-fallback rank parity on eight standard-grid layouts. Before the
2026-11-01 sync, 10 of the 20 cases failed on main, and the existing MFSDP v1 tests matched main on 4 and 2 GPUs.
In the integration run, this PR's test file shows no regressions against main: all 20 cases of
test_mcore_fsdp_dist_index.py pass on 4 GPUs, and main does not have the file.

Compatibility. No change for the standard grid or for collections that set the needed groups. A collection missing one
now raises ValueError instead of AttributeError.

The branch was synced with main 77391e7 on 2026-11-01 by a merge commit (no conflicts), so existing commits

Later on 2026-11-01 the branch also merged main bfab98f. Its #8471 moves FullyShardedDataParallelV1 and FullyShardedDataParallelV2 from mcore_fsdp_adapter.py into mcore_adapter_v1.py and mcore_adapter_v2.py, and keeps mcore_fsdp_adapter.py as the factory that re-exports them. This PR's edits to those classes are applied unchanged where the code moved, and the process-group allowlist follows the move. On 4 GPUs against its new base, the Megatron-FSDP adapter tests and this PR's own tests (7 files) show no regressions: 78 passed here, 58 on the base. The two CUDA-graph end-to-end cases of test_mcore_fully_sharded_data_parallel.py fail on both sides, as they do on main.
and review links stay valid.

Part of #6307.

🤖 Generated with Claude Code

[Connor-XY]

/ok to test 528b9f6

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[Connor-XY]

Superseded by #8894, which carries this change as commit c2dbd1d under the same title, with "Originally #8051." in its message. Closing so review happens in one place. The branch stays on the fork.