gqa(decode-4cases): Case 3 — Cube-Repl × PE-SP (intra-CUBE AR only; 8× memory) (5C.C)

Adds the third 4-cases panel: KV replicated per cube (8× memory
waste), PEs SP on S_kv within each cube, intra-CUBE 8-way reduce on
(m, ℓ, O), and no inter-CUBE comm (every cube ends with full answer;
designated writer = cube 0). Reuses the row-chain + col-bridge
intra-CUBE pattern that anchors Case 4 (21 ipcq_copy per cube × 8
cubes = 168 total). 12 tests pass (4 Case 4 + 4 Case 2 + 4 Case 3).

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

tests/attention/test_milestone_gqa_decode_4cases.py

@@ -39,6 +39,7 @@ from kernbench.topology.builder import resolve_topology
 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"
 _CUBE_RE = re.compile(r"\bcube(\d+)\b")
 
@@ -265,6 +266,124 @@ def test_case2_single_dma_write_at_cube_0():
     )
 
 
+# ── Case 3 (Cube-Repl × PE-SP) ──────────────────────────────────────
+
+
+def _run_case3_smoke(*, S_kv: int):
+    """Drive the Case 3 decode panel via the case-specific runner.
+
+    Case 3 = Cube-Repl × PE-SP. K, V replicated per cube; S_kv split
+    8-way across PEs within each cube. Intra-CUBE 8-way reduce on
+    (m, ℓ, O); no inter-CUBE comm (every cube ends with full answer).
+    """
+    from kernbench.benches.milestone_gqa_decode_4cases import (
+        _run_decode_panel_cube_repl_pe_sp,
+    )
+    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
+
+    def _bench_fn(ctx):
+        _run_decode_panel_cube_repl_pe_sp(
+            ctx, panel=_CASE3_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 3 — T1: panel registered ───────────────────────────────────
+
+
+def test_case3_panel_registered():
+    """The Case 3 panel must be in the bench's ``_PANELS`` +
+    ``_PANEL_DISPATCH`` with the expected single-KV-group dims.
+
+    Case 3: Cube-Repl × PE-SP. K, V replicated per cube; PEs SP on
+    S_kv. Intra-CUBE 8-way AR; no inter-CUBE comm.
+    """
+    from kernbench.benches.milestone_gqa_decode_4cases import (
+        _PANEL_DISPATCH,
+        _PANELS,
+    )
+    assert _CASE3_PANEL in _PANELS, (
+        f"{_CASE3_PANEL!r} not in _PANELS; got {_PANELS}"
+    )
+    assert _CASE3_PANEL in _PANEL_DISPATCH
+    kind, params = _PANEL_DISPATCH[_CASE3_PANEL]
+    assert kind == "decode_cube_repl_pe_sp"
+    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 3 — T2: smoke runner completes ─────────────────────────────
+
+
+def test_case3_runner_smoke():
+    """Case 3 runner drives the new kernel to completion at smoke S_kv."""
+    result = _run_case3_smoke(S_kv=8192)
+    assert result.completion.ok, (
+        f"Case 3 decode smoke at C=8 P=8 must complete; "
+        f"got {result.completion}"
+    )
+
+
+# ── Case 3 — T3: intra-CUBE-only AR (168 ipcq, no inter-CUBE) ───────
+
+
+def test_case3_intra_cube_ar_only_ipcq():
+    """Case 3 reduce pattern: per-CUBE 8-way PE-SP AR (same structural
+    cost as Case 4's intra-CUBE phase = 21 ipcq_copy per cube), and
+    NO inter-CUBE traffic (each cube has a full copy of KV).
+
+      per-CUBE intra (2×4 PE grid):
+        row chain along intra_W: cols 1,2,3 each row × 2 rows ×
+          3 tensors (m, ℓ, O)                          = 18
+        col bridge along intra_N: pe4 only × 3 tensors =  3
+        per-CUBE intra total                           = 21
+        × 8 CUBEs                                      = 168
+
+      inter-CUBE: 0 (replicated KV ⇒ no AllReduce needed).
+
+      Grand total: 168.
+    """
+    result = _run_case3_smoke(S_kv=8192)
+    assert result.completion.ok
+    n_copy = _count(result.engine.op_log, "ipcq_copy")
+    assert n_copy == 168, (
+        f"Case 3 expected 168 ipcq_copy (intra-CUBE only, no inter-CUBE); "
+        f"got {n_copy}"
+    )
+
+
+# ── Case 3 — T4: single dma_write from cube 0 (designated writer) ───
+
+
+def test_case3_single_dma_write_at_cube_0():
+    """Every cube ends with the full answer after intra-CUBE AR; only
+    the designated writer (cube 0, PE 0) stores O to avoid 8 redundant
+    DMAs.
+    """
+    result = _run_case3_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 == {0}, (
+        f"Case 3 designated writer must be cube 0; "
+        f"got cubes={sorted(distinct)}"
+    )
+
+
 # ── Case 4 — T4 (existing, kept) ────────────────────────────────────