Broadcast native FLA CP chain state from the group's global owner rank

[bradhilton]

Fixes #919.

torch.distributed.broadcast takes a global rank as src. Two collectives in src/art/megatron/gdn/fla_cp.py passed context-parallel-local indices instead: _broadcast_chain_final_state used cp_size - 1 and _suffix_summary_exclusive_and_full used 0. Those coincide with global ranks only when the CP group is range(cp_size), i.e. TP 1 and DP 1. With TP 2 on four ranks the CP groups are {0, 2} and {1, 3}, so the first GDN forward raised ValueError: Global rank 1 is not part of group; DP × CP fails the same way for the second group.

Change: translate through dist.get_global_rank(group, local_index) at both sites. No behavior change for TP 1 / DP 1 (the translation is the identity there). The neighbouring _exchange_summary is untouched: all_to_all_single split lists are group-local by design.

Test: tests/unit/test_fla_cp_group_ranks.py spawns four gloo CPU ranks with strided ({0,2}/{1,3}) and offset ({0,1}/{2,3}) CP groups and checks that every member receives the last member's chain state and that the full-chain suffix summary agrees across the group. It passes on this branch and fails on main with the reported error.

GPU validation: in progress on Qwen/Qwen3.8-27B (48 GDN + 16 attention layers), 4×H200, rank-1 LoRA probe forward/backward, with ART_MEGATRON_TENSOR_MODEL_PARALLEL_SIZE=2 ART_MEGATRON_CONTEXT_PARALLEL_SIZE=2 and with DP 2 × CP 2; the same workload failed on the first forward before this change. I will post step timing and peak memory here when the lanes finish.

🤖 Generated with Claude Code

[bradhilton]

Review text removed on 2026-09-17; the consolidated review record is in the summary comment on this PR.

[bradhilton]

Review text removed on 2026-09-17; the consolidated review record is in the summary comment on this PR.

[bradhilton]

Consolidated review record (automated agent review traffic removed 2026-09-17)

Change. Fixes #919: two collectives in src/art/megatron/gdn/fla_cp.py passed CP-local indices to torch.distributed.broadcast, which takes global ranks, so GDN context parallelism combined with TP or DP failed on the first forward. Both sites now translate through dist.get_global_rank; TP 1 / DP 1 behavior is unchanged. Adds a four-rank Gloo test with strided and offset CP groups.

Review. Kang routed independent reviews to McCarthy and Minsky; both returned CLEAR for source correctness at ba959aa, noting the PR edits art.megatron so merge eligibility was Brad's decision. Verified: quality-checks, trainer-rank-gpu-validation and the 2xH200 run were green; Minsky ran the new test (2 passed at head, 2 failed on base with the reported error); McCarthy's 36-case routing fixture passed at head and failed 12 on base; author-reported 4xH200 validation on Qwen3.8-27B (CP2xTP2 45.8 s/step, DP2xCP2 33.6 s/step) at the byte-identical parent commit.

Deferred / follow-ups.

• Nonblocking: the Gloo test asserts cross-rank agreement, not an independently computed suffix summary; eight-rank and CP>2-with-TP topologies remain unexercised.