[ronenkat]

Purpose

Make tests/gpu/test_deepep_m2n_gpu.py fit a 4-GPU CI budget. A cold run is
dominated by the DeepEP/DeepGEMM JIT compile of each new kernel configuration,
so this removes cases whose only contribution is another compile or coverage
held elsewhere. 35 cases become 23, and none needs more than 4 GPUs.

Issue

• Related issue(s): [Feature]: Add GPU M2NConnector E2E coverage and integrate it into weekly CI #512
• Closing keyword, only if fully resolved: Closes #

Scope

• In scope: tests/gpu/test_deepep_m2n_gpu.py, tests/gpu/deepep_m2n_harness.py,
  tests/gpu/README.md; comment-only edits in deepep_m2n/connector.py and
  tests/unit/connectors/test_deepep_m2n_connector.py.
• Out of scope: connector behavior, CI pipeline wiring.

Implementation Notes

Removed cases:

Test	Removed	Why it is not needed
test_attention_tp_peers_dispatch_disjoint_slices	2F-2DPx2TP-bf16, 2F-2DPx2TP-fp8 (6 GPUs), 2F-1DPx2TP-bf16	Kept 2F-1DPx2TP FP8. The per-DP-rank split over TP peers is CPU logic, covered by test_tokens_per_attention_rank_splits_each_dp_rank_over_its_tp_peers.
test_transport_and_deepgemm_experts	1F1A-bf16, 1F1A-fp8, 1F3A-bf16	Kept 2F2A in BF16 and FP8 plus 1F3A in FP8 (pairwise, as test_two_lanes_dbo_order).
test_transport_with_fused_experts_backend	both cases	Every FFN check already runs vLLM fused_experts as the reference.
test_fp8_without_host_sync_for_small_buckets	whole test	The *_nosync standard cases skip the host sync with worst-case shapes; the threshold, ladder and precompile rules are unit-tested.
test_graph_and_eager_plans_mixed_across_ranks	whole test, plus mixed_mode_cases and ffn_graph_mode in the harness	The FFN eager fallback keeps Attention's replay flag, so ranks plan a step alike; the bucket invariant stays unit-tested.
test_dbo_pipeline_cuda_graph_capture_and_replay	both cases	test_dbo_graph_capture_reuses_sent_rows runs the same captured DBO scenario with full checks first.

Other changes: test_close_cancels_a_blocked_control_receive runs on 2F2A (reuses
compiled kernels) and checks every FFN rank; the peer-exit wall-clock assertion is
dropped (run_ranks already reports ranks alive at the budget as hung).

Test Plan

• pytest tests/unit and pytest -m "not gpu and not vllm_runtime" tests/unit
• pytest -sv -m gpu tests/gpu/test_deepep_m2n_gpu.py on 4 GPUs, cold then warm JIT caches

Test Result

vLLM 0.32.0, torch 2.13.0+cu130, afdEP-comm v1.1.0, DeepGEMM 2.8.2 (5081f85), 4x H200.

• Unit: 1619 tests, 1565 passed, 52 skipped, 2 failed; CI selection 1605 / 1552 / 51 / 2.
  Both failures (test_npu_feature_validation[gate-on-attention], [async-graph])
  also fail on main 6620336 and are unrelated.
• GPU (tests/gpu/test_deepep_m2n_gpu.py): 23/23 passed with cold JIT caches
  (17.8 min; the old 35-case suite took 32.1 min cold on the same 4x H200 setup)
  and 23/23 with warm caches (5.4 min).

Docs Impact

• Files updated: tests/gpu/README.md (GPU requirement, case count, test table).

[ronenkat]

FYI @Etelis