gqa(decode-4cases): Case 1 — Cube-SP × PE-TP (inter-CUBE lrab only; PE-TP B=1 waste) (5C.A)

Adds the fourth and final 4-cases panel: KV split S_kv-wise across
the 8 cubes (cube=row_wise, S_local = S_kv/C), replicated within
each cube (pe=replicate). PE-TP at B=1 means only PE 0 of each cube
has work; PEs 1-7 early-return (slide-11 PE-TP-at-B=1 waste). No
intra-CUBE comm; inter-CUBE 8-way reduce reuses the lrab-adapted
center-root pattern (root cube 6) — same structural cost as Case 4's
inter-CUBE phase (21 ipcq_copy). 16 tests pass (4 per case).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

tests/attention/test_milestone_gqa_decode_4cases.py

@@ -41,6 +41,7 @@ TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
 _CASE4_PANEL = "single_kv_group_decode_gqa_cube_sp_pe_sp"
 _CASE3_PANEL = "single_kv_group_decode_gqa_cube_repl_pe_sp"
 _CASE2_PANEL = "single_kv_group_decode_gqa_cube_repl_pe_tp"
+_CASE1_PANEL = "single_kv_group_decode_gqa_cube_sp_pe_tp"
 _CUBE_RE = re.compile(r"\bcube(\d+)\b")
 
 
@@ -384,6 +385,126 @@ def test_case3_single_dma_write_at_cube_0():
     )
 
 
+# ── Case 1 (Cube-SP × PE-TP) ────────────────────────────────────────
+
+
+def _run_case1_smoke(*, S_kv: int):
+    """Drive the Case 1 decode panel via the case-specific runner.
+
+    Case 1 = Cube-SP × PE-TP. K, V split S_kv-wise across the 8 cubes
+    (cube=row_wise), replicated within each cube (pe=replicate). PEs
+    nominally split on the batch dim (PE-TP); at B=1 only PE 0 of
+    each cube has work; PEs 1-7 idle. Inter-CUBE 8-way reduce via the
+    lrab-adapted center-root pattern (root cube 6); no intra-CUBE comm.
+    """
+    from kernbench.benches.milestone_gqa_decode_4cases import (
+        _run_decode_panel_cube_sp_pe_tp,
+    )
+    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
+
+    def _bench_fn(ctx):
+        _run_decode_panel_cube_sp_pe_tp(
+            ctx, panel=_CASE1_PANEL,
+            C=8, P=8,
+            T_q=1, S_kv=S_kv,
+            d_head=128, h_q=8, h_kv=1,
+        )
+
+    return run_bench(
+        topology=topo, bench_fn=_bench_fn,
+        device=resolve_device(None),
+        engine_factory=_engine_factory,
+    )
+
+
+# ── Case 1 — T1: panel registered ───────────────────────────────────
+
+
+def test_case1_panel_registered():
+    """The Case 1 panel must be in the bench's ``_PANELS`` +
+    ``_PANEL_DISPATCH`` with the expected single-KV-group dims.
+
+    Case 1: Cube-SP × PE-TP. K, V split across cubes; PEs TP on
+    batch. At B=1 only PE 0 of each cube works (PE-TP waste).
+    Inter-CUBE lrab AR; no intra-CUBE comm.
+    """
+    from kernbench.benches.milestone_gqa_decode_4cases import (
+        _PANEL_DISPATCH,
+        _PANELS,
+    )
+    assert _CASE1_PANEL in _PANELS, (
+        f"{_CASE1_PANEL!r} not in _PANELS; got {_PANELS}"
+    )
+    assert _CASE1_PANEL in _PANEL_DISPATCH
+    kind, params = _PANEL_DISPATCH[_CASE1_PANEL]
+    assert kind == "decode_cube_sp_pe_tp"
+    assert params.get("C") == 8
+    assert params.get("P") == 8
+    assert params.get("T_q") == 1
+    assert params.get("S_kv") == 131_072
+    assert params.get("d_head") == 128
+    assert params.get("h_q") == 8
+    assert params.get("h_kv") == 1
+
+
+# ── Case 1 — T2: smoke runner completes ─────────────────────────────
+
+
+def test_case1_runner_smoke():
+    """Case 1 runner drives the new kernel to completion at smoke S_kv."""
+    result = _run_case1_smoke(S_kv=8192)
+    assert result.completion.ok, (
+        f"Case 1 decode smoke at C=8 P=8 must complete; "
+        f"got {result.completion}"
+    )
+
+
+# ── Case 1 — T3: inter-CUBE lrab only (21 ipcq, no intra-CUBE) ──────
+
+
+def test_case1_inter_cube_lrab_only_ipcq():
+    """Case 1 reduce pattern: only PE 0 of each cube has work (PE-TP
+    at B=1), so NO intra-CUBE AR. Inter-CUBE 8-way reduce uses the
+    lrab-adapted center-root pattern (same structural cost as Case 4's
+    inter-CUBE phase).
+
+      Inter-CUBE lrab (sub_w=4, sub_h=2):
+        Phase 1 row reduce — 3 sends/row × 3 tensors × 2 rows = 18
+        Phase 2 col reduce — cube 2 → S × 3 tensors           =  3
+        inter-CUBE total                                      = 21
+
+      Intra-CUBE: 0 (PEs 1-7 idle, no partials to merge).
+
+      Grand total: 21.
+    """
+    result = _run_case1_smoke(S_kv=8192)
+    assert result.completion.ok
+    n_copy = _count(result.engine.op_log, "ipcq_copy")
+    assert n_copy == 21, (
+        f"Case 1 expected 21 ipcq_copy (inter-CUBE lrab only, no intra-CUBE); "
+        f"got {n_copy}"
+    )
+
+
+# ── Case 1 — T4: root at lrab center cube 6 ─────────────────────────
+
+
+def test_case1_root_at_center_cube_6():
+    """Case 1 uses the same lrab-adapted center-root reduce as Case 4's
+    inter-CUBE phase; the answer lands at the lrab center cube
+    (sub_w=4, sub_h=2 → root_col=2, root_row=1 → root_cube=6).
+    """
+    result = _run_case1_smoke(S_kv=8192)
+    assert result.completion.ok
+    cubes = _dma_write_cubes(result.engine.op_log)
+    assert cubes, "expected at least one dma_write for the final O store"
+    distinct = set(cubes)
+    assert distinct == {6}, (
+        f"Case 1 root must be the lrab center cube 6; "
+        f"got cubes={sorted(distinct)}"
+    )
+
+
 # ── Case 4 — T4 (existing, kept) ────────────────────────────────────