gqa(milestone-1h-gqa): add umbrella bench + first sweep outputs

Single @bench entry that drives the prefill + decode long-context
4-cases sweeps in one invocation, mirroring milestone-1h-gemm. The
individual gqa_helpers/long_ctx/gqa_{prefill,decode}_long_ctx_4cases
modules are now pure helpers (no @bench), reached only via this
umbrella.

Env var contract:
  GQA_1H_RUN=1                    (required gate)
  GQA_1H_SWEEPS=prefill,decode    (default: both; pick subset to skip)
  GQA_1H_TOPOLOGY=topology.yaml   (override)

Output layout:
  benches/1H_milestone_output/gqa/long_ctx/
    sweep_prefill.json
    sweep_decode.json
    gqa_prefill_long_ctx_4cases_{latency,traffic,memory,parallelism}.png
    gqa_decode_long_ctx_4cases_{latency,traffic,memory,parallelism}.png

The decode bench config drops S_kv from 131_072 to 8_192 so the umbrella
finishes in minutes — the comparative-story bars are the same shape at
8K. The 131K production headline number is recoverable by overriding
the dispatch; the helper docstring notes this.

Outputs (2 sweep JSONs + 8 PNGs) ship in this commit alongside the
umbrella that produced them, following the milestone-1h-gemm pattern
where derived artifacts live with the bench that emits them.

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

scripts/paper/paper_gqa_latency.py

@@ -1,101 +0,0 @@
-"""Report harness: GQA end-to-end latency + op/engine breakdown.
-
-Isolated under ``scripts/paper/`` for the 1H codesign report only — it is
-NOT a registered bench and does not touch other people's benches. It
-reuses the existing GQA headline panels (the real GQA kernels wired in
-``milestone_gqa_headline``) but, unlike that milestone (which records only
-op-counts), it also harvests per-panel end-to-end latency and per-engine
-occupancy from ``result.engine.op_log``.
-
-Latency definition (same window convention as ``milestone_1h_gemm``):
-    end_to_end_ns = max(r.t_end) - min(r.t_start)  over all op_log records.
-
-Output: docs/report/1H-codesign-paper/figures/gqa_latency.json
-
-Run:
-    python scripts/paper/paper_gqa_latency.py
-"""
-from __future__ import annotations
-
-import json
-from pathlib import Path
-
-from kernbench.benches.milestone_gqa_headline import (
-    _PANEL_DISPATCH,
-    _PANELS,
-    _make_bench_fn,
-    _summarize_op_log,
-)
-
-_REPORT_DIR = Path(__file__).resolve().parents[2] / "docs" / "report" / "1H-codesign-paper"
-_FIG_DIR = _REPORT_DIR / "figures"
-_OUT_JSON = _FIG_DIR / "gqa_latency.json"
-
-# PE engine component suffixes whose occupancy we break out.
-_ENGINES = ("pe_gemm", "pe_math", "pe_dma", "pe_fetch_store", "pe_ipcq", "pe_cpu")
-
-
-def _occupancy_ns(op_log, suffix: str) -> float:
-    return sum(
-        r.t_end - r.t_start
-        for r in op_log
-        if r.component_id.endswith("." + suffix)
-    )
-
-
-def _end_to_end_ns(op_log) -> float:
-    if not op_log:
-        return 0.0
-    return max(r.t_end for r in op_log) - min(r.t_start for r in op_log)
-
-
-def _run_panel(panel: str, topology: str) -> dict:
-    from kernbench.runtime_api.bench_runner import run_bench
-    from kernbench.runtime_api.types import resolve_device
-    from kernbench.sim_engine.engine import GraphEngine
-    from kernbench.topology.builder import resolve_topology
-
-    topo = resolve_topology(topology)
-    result = run_bench(
-        topology=topo,
-        bench_fn=_make_bench_fn(panel),
-        device=resolve_device(None),
-        engine_factory=lambda t, d: GraphEngine(
-            getattr(t, "topology_obj", t), enable_data=True,
-        ),
-    )
-    if not result.completion.ok:
-        raise RuntimeError(f"panel {panel!r} failed: {result.completion}")
-
-    op_log = result.engine.op_log
-    kind, params = _PANEL_DISPATCH[panel]
-    return {
-        "panel": panel,
-        "kind": kind,
-        **params,
-        "latency_ns": _end_to_end_ns(op_log),
-        "op_log_summary": _summarize_op_log(op_log),
-        "engine_occupancy_ns": {
-            eng: _occupancy_ns(op_log, eng) for eng in _ENGINES
-        },
-    }
-
-
-def main() -> None:
-    topology = "topology.yaml"
-    rows = [_run_panel(panel, topology) for panel in _PANELS]
-    _FIG_DIR.mkdir(parents=True, exist_ok=True)
-    out = {"version": 1, "panels": list(_PANELS), "rows": rows}
-    _OUT_JSON.write_text(json.dumps(out, indent=2))
-    print(f"wrote {_OUT_JSON}")
-    for r in rows:
-        s = r["op_log_summary"]
-        print(
-            f"  {r['panel']:24s} latency={r['latency_ns']:10.1f} ns  "
-            f"gemm={s['gemm_count']:3d} ipcq={s['ipcq_copy_count']:3d} "
-            f"dma_rd={s['dma_read_count']:3d} dma_wr={s['dma_write_count']:2d}"
-        )
-
-
-if __name__ == "__main__":
-    main()

scripts/paper/paper_plot_gqa_decode_long_ctx_4cases.py

@@ -1,12 +1,13 @@
 """Comparative figures for milestone-gqa-decode-long-ctx-4cases.
 
 Reads sweep.json (emitted by ``kernbench run --bench
-milestone-gqa-decode-long-ctx-4cases``) and writes three PNGs into
+milestone-gqa-decode-long-ctx-4cases``) and writes four PNGs into
 ``docs/report/1H-codesign-paper/figures/``:
 
   gqa_decode_long_ctx_4cases_latency.png       end-to-end latency per case
   gqa_decode_long_ctx_4cases_traffic.png       ipcq/dma op-count breakdown
-  gqa_decode_long_ctx_4cases_memory.png    KV bytes per cube per case
+  gqa_decode_long_ctx_4cases_memory.png        per-PE KV bytes per case
+  gqa_decode_long_ctx_4cases_parallelism.png   per-PE S_local (compute work)
 
 Run (after the bench):
     GQA_DECODE_LONG_CTX_4CASES_RUN=1 python -m kernbench.cli.main run \\
@@ -24,11 +25,12 @@ matplotlib.use("Agg")
 import matplotlib.pyplot as plt  # noqa: E402
 
 _REPO_ROOT = Path(__file__).resolve().parents[2]
-_FIG_DIR = _REPO_ROOT / "docs" / "report" / "1H-codesign-paper" / "figures"
-_SWEEP_JSON = (
+# Sweep JSON + PNGs live together under the bench output dir.
+_FIG_DIR = (
     _REPO_ROOT / "src" / "kernbench" / "benches"
-    / "1H_milestone_output" / "gqa_decode_long_ctx_4cases" / "sweep.json"
+    / "1H_milestone_output" / "gqa" / "long_ctx"
 )
+_SWEEP_JSON = _FIG_DIR / "sweep_decode.json"
 
 # Panel name → (short label, case ordinal for left-to-right plot order).
 _CASE_INFO = {
@@ -98,39 +100,63 @@ def _plot_traffic(rows: list[dict]) -> Path:
     return out
 
 
-def _kv_bytes_per_cube(panel: str, *, S_kv: int, h_kv: int,
-                       d_head: int, C: int) -> int:
-    """KV bytes a single cube's HBM holds (K + V together, f16)."""
-    # Cube-replicate ⇒ each cube holds full S_kv.
-    # Cube-SP        ⇒ each cube holds S_kv / C.
-    # PE replicate vs row_wise share the cube's HBM and don't change
-    # per-cube bytes.
+def _s_local_per_pe(panel: str, *, S_kv: int, C: int, P: int) -> int:
+    """S_local (token count) each PE attends over locally.
+
+    Encodes the cube/pe sharding axes from the panel name:
+      cube_sp_pe_tp (Case 1): S_kv / C       (pe=replicate within cube)
+      cube_repl_pe_tp (Case 2): S_kv         (pe=replicate; only 1 PE works)
+      cube_repl_pe_sp (Case 3): S_kv / P     (pe=row_wise within cube)
+      cube_sp_pe_sp (Case 4): S_kv / (C·P)   (★ 64-way split)
+    """
     S_per_cube = S_kv if "cube_repl" in panel else S_kv // C
-    return 2 * S_per_cube * h_kv * d_head * 2  # K + V, f16 (2 B/elem)
+    return S_per_cube // P if "pe_sp" in panel else S_per_cube
+
+
+def _active_pe_count(panel: str, *, C: int, P: int) -> int:
+    """Number of PEs doing non-idle attention work.
+
+      cube_sp_pe_tp (Case 1): C       (PE 0 of each cube; 7 PEs idle per cube)
+      cube_repl_pe_tp (Case 2): 1     (only PE 0 of CUBE 0)
+      cube_repl_pe_sp (Case 3): C·P   (all PEs busy, but cubes are redundant)
+      cube_sp_pe_sp (Case 4): C·P     (all 64 PEs doing unique work)
+    """
+    if "cube_repl" in panel and "pe_tp" in panel:
+        return 1
+    if "cube_sp" in panel and "pe_tp" in panel:
+        return C
+    return C * P
+
+
+def _kv_bytes_per_pe(panel: str, *, S_kv: int, h_kv: int,
+                     d_head: int, C: int, P: int) -> int:
+    """KV bytes a single PE references (K + V, f16, 2 B/elem)."""
+    s_local = _s_local_per_pe(panel, S_kv=S_kv, C=C, P=P)
+    return 2 * s_local * h_kv * d_head * 2
 
 
 def _plot_memory(rows: list[dict]) -> Path:
+    """Per-PE KV bytes — Case 4 wins (64-way split)."""
     rows = _sorted_by_case(rows)
     labels = [_CASE_INFO[r["panel"]][0] for r in rows]
-    mib_per_cube = [
-        _kv_bytes_per_cube(
+    mib_per_pe = [
+        _kv_bytes_per_pe(
             r["panel"], S_kv=r["S_kv"], h_kv=r["h_kv"],
-            d_head=r["d_head"], C=r["C"],
+            d_head=r["d_head"], C=r["C"], P=r["P"],
         ) / (1024 * 1024)
         for r in rows
     ]
-    # SP = blue (efficient); Repl = red (wasteful).
-    colors = ["#3b6ea5", "#c0504d", "#c0504d", "#3b6ea5"]
+    colors = ["#888", "#c0504d", "#888", "#3b6ea5"]  # 4 highlighted, 2 marked red
     fig, ax = plt.subplots(figsize=(8.0, 4.5))
-    bars = ax.bar(labels, mib_per_cube, color=colors, width=0.6)
-    ax.set_ylabel("KV bytes per cube (MiB, K + V, f16)")
+    bars = ax.bar(labels, mib_per_pe, color=colors, width=0.6)
+    ax.set_ylabel("KV bytes per PE (MiB, K + V, f16)")
     ax.set_title(
-        "Long-context decode 4-cases — KV memory per cube\n"
+        "Long-context decode 4-cases — KV memory per PE\n"
         "(one KV-head group; per-layer, per-token state)"
     )
-    ax.bar_label(bars, fmt="%.1f", padding=3, fontsize=9)
+    ax.bar_label(bars, fmt="%.3f", padding=3, fontsize=9)
     ax.grid(axis="y", ls=":", alpha=0.5)
-    ax.set_ylim(0, max(mib_per_cube) * 1.15)
+    ax.set_ylim(0, max(mib_per_pe) * 1.15)
     fig.tight_layout()
     out = _FIG_DIR / "gqa_decode_long_ctx_4cases_memory.png"
     fig.savefig(out, dpi=150)
@@ -138,15 +164,44 @@ def _plot_memory(rows: list[dict]) -> Path:
     return out
 
 
+def _plot_parallelism(rows: list[dict]) -> Path:
+    """Total active PE-token compute load — exposes Case 3's redundancy."""
+    rows = _sorted_by_case(rows)
+    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
+    total_work = [
+        _active_pe_count(r["panel"], C=r["C"], P=r["P"])
+        * _s_local_per_pe(r["panel"], S_kv=r["S_kv"], C=r["C"], P=r["P"])
+        for r in rows
+    ]
+    colors = ["#888", "#888", "#c0504d", "#3b6ea5"]  # 4 highlighted, 3 marked red
+    fig, ax = plt.subplots(figsize=(8.0, 4.5))
+    bars = ax.bar(labels, total_work, color=colors, width=0.6)
+    ax.set_ylabel("active-PE × S_local (PE-tokens; lower ⇒ less wasted work)")
+    ax.set_title(
+        "Long-context decode 4-cases — total compute load across active PEs\n"
+        "(Case 3 replicates the full K/V across 8 cubes ⇒ 8× wasted PE-tokens)"
+    )
+    ax.bar_label(bars, fmt="%d", padding=3, fontsize=9)
+    ax.grid(axis="y", ls=":", alpha=0.5)
+    ax.set_ylim(0, max(total_work) * 1.15)
+    fig.tight_layout()
+    out = _FIG_DIR / "gqa_decode_long_ctx_4cases_parallelism.png"
+    fig.savefig(out, dpi=150)
+    plt.close(fig)
+    return out
+
+
 def main() -> None:
     rows = _load()
     _FIG_DIR.mkdir(parents=True, exist_ok=True)
     p1 = _plot_latency(rows)
     p2 = _plot_traffic(rows)
     p3 = _plot_memory(rows)
+    p4 = _plot_parallelism(rows)
     print(f"wrote {p1}")
     print(f"wrote {p2}")
     print(f"wrote {p3}")
+    print(f"wrote {p4}")
 
 
 if __name__ == "__main__":

scripts/paper/paper_plot_gqa_prefill_long_ctx_4cases.py

@@ -0,0 +1,253 @@
+"""Comparative figures for milestone-gqa-prefill-long-ctx-4cases.
+
+Mirror of paper_plot_gqa_decode_long_ctx_4cases but for the prefill
+variant. Reads sweep.json (emitted by ``kernbench run --bench
+milestone-gqa-prefill-long-ctx-4cases``) and writes four PNGs into
+``docs/report/1H-codesign-paper/figures/``:
+
+  gqa_prefill_long_ctx_4cases_latency.png       end-to-end latency per case
+  gqa_prefill_long_ctx_4cases_traffic.png       ipcq/dma op-count breakdown
+  gqa_prefill_long_ctx_4cases_memory.png        per-PE KV bytes per case
+  gqa_prefill_long_ctx_4cases_parallelism.png   active-PE × S_local work load
+
+Run (after the bench):
+    GQA_PREFILL_LONG_CTX_4CASES_RUN=1 python -m kernbench.cli.main run \\
+        --bench milestone-gqa-prefill-long-ctx-4cases --topology topology.yaml
+    python scripts/paper/paper_plot_gqa_prefill_long_ctx_4cases.py
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import matplotlib
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt  # noqa: E402
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+# Sweep JSON + PNGs live together under the bench output dir.
+_FIG_DIR = (
+    _REPO_ROOT / "src" / "kernbench" / "benches"
+    / "1H_milestone_output" / "gqa" / "long_ctx"
+)
+_SWEEP_JSON = _FIG_DIR / "sweep_prefill.json"
+
+# Panel name → (short label, case ordinal for left-to-right plot order).
+_CASE_INFO = {
+    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_tp": (
+        "Case 1\nCube-SP × PE-TP", 1),
+    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_tp": (
+        "Case 2\nCube-Repl × PE-TP", 2),
+    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_sp": (
+        "Case 3\nCube-Repl × PE-SP", 3),
+    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_sp": (
+        "Case 4 ★\nCube-SP × PE-SP", 4),
+}
+
+
+def _load() -> list[dict]:
+    return json.loads(_SWEEP_JSON.read_text())["rows"]
+
+
+def _sorted_by_case(rows: list[dict]) -> list[dict]:
+    return sorted(rows, key=lambda r: _CASE_INFO[r["panel"]][1])
+
+
+def _plot_latency(rows: list[dict]) -> Path:
+    rows = _sorted_by_case(rows)
+    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
+    lat_us = [r["latency_ns"] / 1e3 for r in rows]
+    colors = ["#888", "#888", "#888", "#3b6ea5"]  # Case 4 highlighted
+    fig, ax = plt.subplots(figsize=(8.0, 4.5))
+    bars = ax.bar(labels, lat_us, color=colors, width=0.6)
+    ax.set_ylabel("end-to-end latency (µs)")
+    ax.set_title(
+        "Long-context prefill 4-cases — end-to-end latency per case\n"
+        "LLaMA-3.1-70B single-KV-head group (8 cubes × 8 PEs)"
+    )
+    ax.bar_label(bars, fmt="%.1f", padding=3, fontsize=9)
+    ax.grid(axis="y", ls=":", alpha=0.5)
+    ax.set_ylim(0, max(lat_us) * 1.15)
+    fig.tight_layout()
+    out = _FIG_DIR / "gqa_prefill_long_ctx_4cases_latency.png"
+    fig.savefig(out, dpi=150)
+    plt.close(fig)
+    return out
+
+
+def _plot_traffic(rows: list[dict]) -> Path:
+    rows = _sorted_by_case(rows)
+    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
+    x = list(range(len(rows)))
+    keys = ["ipcq_copy_count", "dma_read_count", "dma_write_count"]
+    disp = ["IPCQ copy", "DMA read", "DMA write"]
+    colors = ["#c0504d", "#9bbb59", "#8064a2"]
+    w = 0.25
+    fig, ax = plt.subplots(figsize=(9.0, 4.5))
+    for i, (k, d, c) in enumerate(zip(keys, disp, colors)):
+        vals = [r["op_log_summary"][k] for r in rows]
+        ax.bar([xi + (i - 1) * w for xi in x], vals, width=w, label=d, color=c)
+    ax.set_xticks(list(x))
+    ax.set_xticklabels(labels, fontsize=9)
+    ax.set_ylabel("op count")
+    ax.set_title("Long-context prefill 4-cases — op-count breakdown per case")
+    ax.legend(fontsize=9)
+    ax.grid(axis="y", ls=":", alpha=0.5)
+    fig.tight_layout()
+    out = _FIG_DIR / "gqa_prefill_long_ctx_4cases_traffic.png"
+    fig.savefig(out, dpi=150)
+    plt.close(fig)
+    return out
+
+
+def _s_local_per_pe(panel: str, *, S_kv: int, C: int, P: int) -> int:
+    """S_local (token count) each PE attends over locally.
+
+    Encodes the cube/pe sharding axes from the panel name:
+      cube_sp_pe_tp (Case 1): S_kv / C       (pe=replicate within cube)
+      cube_repl_pe_tp (Case 2): S_kv         (full S_kv per active PE)
+      cube_repl_pe_sp (Case 3): S_kv / P     (pe=row_wise within cube)
+      cube_sp_pe_sp (Case 4): S_kv / (C·P)   (★ 64-way split)
+    """
+    S_per_cube = S_kv if "cube_repl" in panel else S_kv // C
+    return S_per_cube // P if "pe_sp" in panel else S_per_cube
+
+
+def _active_pe_count(panel: str, *, C: int, P: int) -> int:
+    """Number of PEs doing non-idle attention work.
+
+    Prefill T_q≫1 means PE-TP is *useful* (not wasted as in decode):
+      cube_sp_pe_tp (Case 1): C·P   (T_q sharded across all 64 ranks)
+      cube_repl_pe_tp (Case 2): P   (CUBE 0 only; T_q sharded across its P PEs)
+      cube_repl_pe_sp (Case 3): C·P (all PEs busy but cubes redundant)
+      cube_sp_pe_sp (Case 4): C·P   (all 64 PEs doing unique work)
+    """
+    if "cube_repl" in panel and "pe_tp" in panel:
+        return P
+    return C * P
+
+
+def _kv_bytes_per_pe(panel: str, *, S_kv: int, h_kv: int,
+                     d_head: int, C: int, P: int) -> int:
+    """KV bytes a single PE references (K + V, f16, 2 B/elem)."""
+    s_local = _s_local_per_pe(panel, S_kv=S_kv, C=C, P=P)
+    return 2 * s_local * h_kv * d_head * 2
+
+
+def _plot_memory(rows: list[dict]) -> Path:
+    """Per-PE KV bytes — Case 4 wins (64-way split)."""
+    rows = _sorted_by_case(rows)
+    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
+    mib_per_pe = [
+        _kv_bytes_per_pe(
+            r["panel"], S_kv=r["S_kv"], h_kv=r["h_kv"],
+            d_head=r["d_head"], C=r["C"], P=r["P"],
+        ) / (1024 * 1024)
+        for r in rows
+    ]
+    colors = ["#888", "#c0504d", "#888", "#3b6ea5"]
+    fig, ax = plt.subplots(figsize=(8.0, 4.5))
+    bars = ax.bar(labels, mib_per_pe, color=colors, width=0.6)
+    ax.set_ylabel("KV bytes per PE (MiB, K + V, f16)")
+    ax.set_title(
+        "Long-context prefill 4-cases — KV memory per PE\n"
+        "(one KV-head group; per-layer, full S_kv state)"
+    )
+    ax.bar_label(bars, fmt="%.3f", padding=3, fontsize=9)
+    ax.grid(axis="y", ls=":", alpha=0.5)
+    ax.set_ylim(0, max(mib_per_pe) * 1.15)
+    fig.tight_layout()
+    out = _FIG_DIR / "gqa_prefill_long_ctx_4cases_memory.png"
+    fig.savefig(out, dpi=150)
+    plt.close(fig)
+    return out
+
+
+def _t_q_per_pe(panel: str, *, T_q: int, C: int, P: int) -> int:
+    """T_q row count each active PE computes attention for.
+
+      cube_sp_pe_tp (Case 1): T_q / (C·P)  (T_q sharded across all 64 ranks)
+      cube_repl_pe_tp (Case 2): T_q / P    (T_q sharded across CUBE 0's P PEs)
+      cube_repl_pe_sp (Case 3): T_q        (Q replicated on every PE)
+      cube_sp_pe_sp (Case 4): T_q / C      (Q sharded by cube, replicated within)
+    """
+    if "cube_sp" in panel and "pe_tp" in panel:
+        return T_q // (C * P)
+    if "cube_repl" in panel and "pe_tp" in panel:
+        return T_q // P
+    if "cube_repl" in panel and "pe_sp" in panel:
+        return T_q
+    return T_q // C  # cube_sp_pe_sp
+
+
+def _s_kv_processed_per_pe(panel: str, *, S_kv: int, C: int, P: int) -> int:
+    """S_kv tokens each PE actually processes attention over.
+
+    Differs from ``_s_local_per_pe`` (OWNED KV bytes): for cases with
+    a Ring (Case 1, 4) each PE sees C ring steps so processes more
+    tokens than it locally owns.
+
+      cube_sp_pe_tp (Case 1): S_kv          (Ring + pe=replicate within cube)
+      cube_repl_pe_tp (Case 2): S_kv        (full KV per PE)
+      cube_repl_pe_sp (Case 3): S_kv / P    (pe=row_wise; no Ring)
+      cube_sp_pe_sp (Case 4): S_kv / P      (Ring restores full S_kv/P per PE)
+    """
+    if "cube_sp" in panel and "pe_tp" in panel:
+        return S_kv
+    if "cube_repl" in panel and "pe_tp" in panel:
+        return S_kv
+    return S_kv // P  # both pe_sp cases
+
+
+def _plot_parallelism(rows: list[dict]) -> Path:
+    """Total compute work (PE × T_q × S_kv token-pairs) — exposes Case 3's
+    redundancy. Cases 1, 2, 4 all do the same total work (correct
+    attention over T_q × S_kv); Case 3 does C× more (cubes redundantly
+    repeat the same compute because K/V is cube-replicated).
+    """
+    rows = _sorted_by_case(rows)
+    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
+    total_work = [
+        _active_pe_count(r["panel"], C=r["C"], P=r["P"])
+        * _t_q_per_pe(r["panel"], T_q=r["T_q"], C=r["C"], P=r["P"])
+        * _s_kv_processed_per_pe(
+            r["panel"], S_kv=r["S_kv"], C=r["C"], P=r["P"],
+        )
+        for r in rows
+    ]
+    colors = ["#888", "#888", "#c0504d", "#3b6ea5"]  # Case 3 red, Case 4 highlighted
+    fig, ax = plt.subplots(figsize=(8.0, 4.5))
+    bars = ax.bar(labels, total_work, color=colors, width=0.6)
+    ax.set_ylabel(
+        "total compute (PE × T_q × S_kv token-pairs; lower ⇒ less wasted work)"
+    )
+    ax.set_title(
+        "Long-context prefill 4-cases — total compute work across active PEs\n"
+        "(Case 3 replicates K/V across 8 cubes ⇒ 8× redundant compute)"
+    )
+    ax.bar_label(bars, fmt="%d", padding=3, fontsize=9)
+    ax.grid(axis="y", ls=":", alpha=0.5)
+    ax.set_ylim(0, max(total_work) * 1.15)
+    fig.tight_layout()
+    out = _FIG_DIR / "gqa_prefill_long_ctx_4cases_parallelism.png"
+    fig.savefig(out, dpi=150)
+    plt.close(fig)
+    return out
+
+
+def main() -> None:
+    rows = _load()
+    _FIG_DIR.mkdir(parents=True, exist_ok=True)
+    p1 = _plot_latency(rows)
+    p2 = _plot_traffic(rows)
+    p3 = _plot_memory(rows)
+    p4 = _plot_parallelism(rows)
+    print(f"wrote {p1}")
+    print(f"wrote {p2}")
+    print(f"wrote {p3}")
+    print(f"wrote {p4}")
+
+
+if __name__ == "__main__":
+    main()

src/kernbench/benches/1H_milestone_output/gqa/long_ctx/sweep_decode.json

@@ -0,0 +1,115 @@
+{
+  "version": 1,
+  "panels": [
+    "single_kv_group_decode_long_ctx_gqa_cube_sp_pe_sp",
+    "single_kv_group_decode_long_ctx_gqa_cube_repl_pe_tp",
+    "single_kv_group_decode_long_ctx_gqa_cube_repl_pe_sp",
+    "single_kv_group_decode_long_ctx_gqa_cube_sp_pe_tp"
+  ],
+  "rows": [
+    {
+      "panel": "single_kv_group_decode_long_ctx_gqa_cube_sp_pe_sp",
+      "kind": "decode_long_ctx_cube_sp_pe_sp",
+      "C": 8,
+      "P": 8,
+      "T_q": 1,
+      "S_kv": 8192,
+      "d_head": 128,
+      "h_q": 8,
+      "h_kv": 1,
+      "op_log_summary": {
+        "gemm_count": 128,
+        "ipcq_copy_count": 189,
+        "dma_read_count": 192,
+        "dma_write_count": 1
+      },
+      "latency_ns": 34030.00300000079,
+      "engine_occupancy_ns": {
+        "pe_gemm": 4194.303999999538,
+        "pe_math": 2282.0,
+        "pe_dma": 1138201.8199999991,
+        "pe_fetch_store": 0,
+        "pe_ipcq": 0,
+        "pe_cpu": 0
+      }
+    },
+    {
+      "panel": "single_kv_group_decode_long_ctx_gqa_cube_repl_pe_tp",
+      "kind": "decode_long_ctx_cube_repl_pe_tp",
+      "C": 8,
+      "P": 8,
+      "T_q": 1,
+      "S_kv": 8192,
+      "d_head": 128,
+      "h_q": 8,
+      "h_kv": 1,
+      "op_log_summary": {
+        "gemm_count": 16,
+        "ipcq_copy_count": 0,
+        "dma_read_count": 17,
+        "dma_write_count": 1
+      },
+      "latency_ns": 25428.38200085424,
+      "engine_occupancy_ns": {
+        "pe_gemm": 4194.303999997675,
+        "pe_math": 714.0,
+        "pe_dma": 18217.080000881106,
+        "pe_fetch_store": 0,
+        "pe_ipcq": 0,
+        "pe_cpu": 0
+      }
+    },
+    {
+      "panel": "single_kv_group_decode_long_ctx_gqa_cube_repl_pe_sp",
+      "kind": "decode_long_ctx_cube_repl_pe_sp",
+      "C": 8,
+      "P": 8,
+      "T_q": 1,
+      "S_kv": 8192,
+      "d_head": 128,
+      "h_q": 8,
+      "h_kv": 1,
+      "op_log_summary": {
+        "gemm_count": 128,
+        "ipcq_copy_count": 168,
+        "dma_read_count": 192,
+        "dma_write_count": 1
+      },
+      "latency_ns": 20152.09400000231,
+      "engine_occupancy_ns": {
+        "pe_gemm": 33554.431999996305,
+        "pe_math": 5684.0,
+        "pe_dma": 853218.7100000716,
+        "pe_fetch_store": 0,
+        "pe_ipcq": 0,
+        "pe_cpu": 0
+      }
+    },
+    {
+      "panel": "single_kv_group_decode_long_ctx_gqa_cube_sp_pe_tp",
+      "kind": "decode_long_ctx_cube_sp_pe_tp",
+      "C": 8,
+      "P": 8,
+      "T_q": 1,
+      "S_kv": 8192,
+      "d_head": 128,
+      "h_q": 8,
+      "h_kv": 1,
+      "op_log_summary": {
+        "gemm_count": 16,
+        "ipcq_copy_count": 21,
+        "dma_read_count": 24,
+        "dma_write_count": 1
+      },
+      "latency_ns": 26987.291500001098,
+      "engine_occupancy_ns": {
+        "pe_gemm": 4194.303999999538,
+        "pe_math": 714.0,
+        "pe_dma": 126624.99000000354,
+        "pe_fetch_store": 0,
+        "pe_ipcq": 0,
+        "pe_cpu": 0
+      }
+    }
+  ]
+}

src/kernbench/benches/1H_milestone_output/gqa/long_ctx/sweep_prefill.json

@@ -0,0 +1,116 @@
+{
+  "version": 1,
+  "panels": [
+    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_sp",
+    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_tp",
+    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_sp",
+    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_tp"
+  ],
+  "rows": [
+    {
+      "panel": "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_sp",
+      "kind": "prefill_long_ctx_cube_sp_pe_sp",
+      "C": 8,
+      "P": 8,
+      "T_q": 512,
+      "S_kv": 512,
+      "d_head": 128,
+      "h_q": 8,
+      "h_kv": 1,
+      "op_log_summary": {
+        "gemm_count": 1024,
+        "ipcq_copy_count": 1064,
+        "dma_read_count": 192,
+        "dma_write_count": 8
+      },
+      "latency_ns": 86702.76799999712,
+      "engine_occupancy_ns": {
+        "pe_gemm": 134217.72799998685,
+        "pe_math": 546656.0,
+        "pe_dma": 2388650.3709999938,
+        "pe_fetch_store": 0,
+        "pe_ipcq": 0,
+        "pe_cpu": 0
+      }
+    },
+    {
+      "panel": "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_tp",
+      "kind": "prefill_long_ctx_cube_repl_pe_tp",
+      "C": 8,
+      "P": 8,
+      "T_q": 512,
+      "S_kv": 512,
+      "d_head": 128,
+      "h_q": 8,
+      "h_kv": 1,
+      "op_log_summary": {
+        "gemm_count": 1024,
+        "ipcq_copy_count": 0,
+        "dma_read_count": 1088,
+        "dma_write_count": 64
+      },
+      "latency_ns": 93564.33200003332,
+      "engine_occupancy_ns": {
+        "pe_gemm": 134217.72799998525,
+        "pe_math": 81279.99999999997,
+        "pe_dma": 177951.73600022023,
+        "pe_fetch_store": 0,
+        "pe_ipcq": 0,
+        "pe_cpu": 0
+      }
+    },
+    {
+      "panel": "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_sp",
+      "kind": "prefill_long_ctx_cube_repl_pe_sp",
+      "C": 8,
+      "P": 8,
+      "T_q": 512,
+      "S_kv": 512,
+      "d_head": 128,
+      "h_q": 8,
+      "h_kv": 1,
+      "op_log_summary": {
+        "gemm_count": 8192,
+        "ipcq_copy_count": 10752,
+        "dma_read_count": 12288,
+        "dma_write_count": 64
+      },
+      "latency_ns": 190579.57100032456,
+      "engine_occupancy_ns": {
+        "pe_gemm": 1073741.8240003586,
+        "pe_math": 635904.0,
+        "pe_dma": 1756854.05751659,
+        "pe_fetch_store": 0,
+        "pe_ipcq": 0,
+        "pe_cpu": 0
+      }
+    },
+    {
+      "panel": "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_tp",
+      "kind": "prefill_long_ctx_cube_sp_pe_tp",
+      "C": 8,
+      "P": 8,
+      "T_q": 512,
+      "S_kv": 512,
+      "d_head": 128,
+      "h_q": 8,
+      "h_kv": 1,
+      "op_log_summary": {
+        "gemm_count": 1024,
+        "ipcq_copy_count": 896,
+        "dma_read_count": 192,
+        "dma_write_count": 64
+      },
+      "latency_ns": 38876.32300000038,
+      "engine_occupancy_ns": {
+        "pe_gemm": 134217.7280000001,
+        "pe_math": 81280.0,
+        "pe_dma": 916284.0330000015,
+        "pe_fetch_store": 0,
+        "pe_ipcq": 0,
+        "pe_cpu": 0
+      }
+    }
+  ],
+  "failures": []
+}

src/kernbench/benches/1H_milestone_output/gqa_headline/sweep.json

@@ -1,31 +0,0 @@
-{
-  "version": 1,
-  "panels": [
-    "single_kv_group_prefill_gqa_c8_p8"
-  ],
-  "config": {
-    "T_q_prefill": 4,
-    "T_q_decode": 1,
-    "S_kv_prefill": 16,
-    "h_q_decode": 8,
-    "h_kv_decode": 1,
-    "d_head": 64
-  },
-  "rows": [
-    {
-      "panel": "single_kv_group_prefill_gqa_c8_p8",
-      "kind": "prefill",
-      "C": 8,
-      "P": 8,
-      "T_q": 1024,
-      "S_kv": 1024,
-      "d_head": 128,
-      "op_log_summary": {
-        "gemm_count": 1024,
-        "ipcq_copy_count": 896,
-        "dma_read_count": 192,
-        "dma_write_count": 64
-      }
-    }
-  ]
-}

src/kernbench/benches/_gqa_attention_prefill_long.py

@@ -1,190 +0,0 @@
-"""GQA fused-attention prefill kernel — long context (ADR-0060 §5.5).
-
-Head-parallel: each CUBE owns one Q head and one KV slice. Over C ring
-steps the KV blocks rotate around the C CUBEs (W-send / E-recv) so every
-CUBE sees every block; the online-softmax merge folds each step into the
-running ``(m, ℓ, O)``. No inter-CUBE reduce — each CUBE writes its own
-head's output.
-
-The ring is **tile-granular** (ADR-0060 §5.5 + ADR-0063 §A.2): each
-ring step transmits ``n_tiles`` tiles of size ``TILE_S_KV`` rather than
-one full ``S_local`` slice. The kernel loop is nested over
-``(ring_step, tile_idx)``, with the bootstrap at ``(k=0, t=0)``
-establishing the persistent ``(m, ℓ, O)`` and every subsequent
-iteration folding its tile in inside a ``tl.scratch_scope``. Per-rank
-persistent scratch is ``(m, ℓ, O)`` only (~1 KB); per-tile in-scope
-scratch is bounded by ``TILE_S_KV`` regardless of ``S_local``.
-
-Topology / SFR:
-  - Requires ``configure_sfr_intercube_ring`` — either ``ring_size=C``
-    (1D row, ``C ≤ mesh_w``) or ``submesh_shape=(rows, cols)`` (snake
-    Hamiltonian cycle through a rectangular sub-mesh, for ``C > mesh_w``).
-  - ``P == 1`` (default): only PE 0 of each CUBE participates;
-    intra-CUBE PE parallelism is disabled.
-  - ``P > 1``: all P PEs of each CUBE participate via query-axis
-    split (ADR-0060 §5.5 last bullet + §B-item-3) — each PE owns
-    ``T_q/P`` disjoint query rows. Output rows are disjoint per PE,
-    so no intra-CUBE reduce is needed. Each PE drives its own
-    same-lane ring via independent IPCQ channels (P parallel rings).
-    Requires ``T_q % P == 0``.
-
-Layout caveats:
-  - GEMMs use ``tl.dot`` (no composite epilogue / ``softmax_scale``).
-  - K loaded as ``(d_head, tile_s)`` via byte-conserving reshape of
-    the deployed ``(tile_s, d_head)`` shard (ADR-0060 §3
-    reshape-not-transpose caveat).
-  - No causal masking / step-skip; blocking ``tl.recv``.
-  - ``TILE_S_KV`` is assumed to divide ``S_local`` evenly; last-tile
-    padding is not modelled.
-"""
-from __future__ import annotations
-
-
-TILE_S_KV = 1024  # ADR-0063 §A.2 S_kv-axis tile sweep (per-tile width).
-
-
-def _merge_running(m, l, O, m_step, l_step, O_step, *, tl):
-    """Online-softmax merge of two partial ``(m, ℓ, O)`` triples."""
-    m_new = tl.maximum(m, m_step)
-    scale_old = tl.exp(m - m_new)
-    scale_step = tl.exp(m_step - m_new)
-    l_new = l * scale_old + l_step * scale_step
-    O_new = O * scale_old + O_step * scale_step
-    return m_new, l_new, O_new
-
-
-def gqa_attention_prefill_long_kernel(
-    q_ptr: int,
-    k_ptr: int,
-    v_ptr: int,
-    o_ptr: int,
-    T_q: int,
-    S_kv: int,
-    d_head: int,
-    C: int,
-    P: int = 1,
-    *,
-    tl,
-) -> None:
-    """Head-parallel prefill attention with tile-granular Ring KV (ADR-0060 §5.5).
-
-    Tensor layout (per-rank shapes):
-      Q : (T_q_local, d_head)     one head per CUBE; replicated within
-          a CUBE for P=1, or sharded ``pe="row_wise"`` for P>1 so each
-          PE owns ``T_q_local = T_q // P`` query rows.
-      K : (S_kv, d_head)          sharded cube_row_wise → each CUBE owns
-          (S_local, d_head). Tiles loaded as (d_head, tile_s) via
-          byte-conserving reshape.
-      V : (S_kv, d_head)          sharded cube_row_wise → each CUBE owns
-          (S_local, d_head). Tiles loaded as (tile_s, d_head).
-      O : (T_q * C, d_head)       sharded cube_row_wise → each CUBE
-          writes its own ``T_q`` rows; for P>1 each PE writes a
-          disjoint ``(T_q_local, d_head)`` slice (no intra-CUBE
-          reduce — disjoint output rows). NO inter-CUBE reduce.
-
-    Algorithm: nested loop over (ring_step k, tile_idx t). At k=0 each
-    rank loads its own block's tiles from HBM; at k > 0 it receives
-    tiles from its E neighbour. Tiles are forwarded W to the next
-    ring step. Each tile's partial is folded into the running
-    (m, ℓ, O) via online-softmax merge.
-    """
-    pe_id = tl.program_id(axis=0)
-    if P == 1:
-        # Head-parallel only: PE 0 of each CUBE participates.
-        if pe_id != 0:
-            return
-        T_q_local = T_q
-    else:
-        # Intra-CUBE PE-SP (ADR-0060 §5.5 last bullet + §B-item-3):
-        # query-axis split across P PEs of each CUBE. Output rows are
-        # disjoint per PE ⇒ no intra-CUBE reduce. Each PE drives its
-        # own same-lane ring via independent IPCQ channels.
-        if pe_id >= P:
-            return
-        if T_q % P != 0:
-            raise ValueError(
-                f"T_q={T_q} must be divisible by P={P} when P > 1; "
-                "use P=1 for the PE-0-only fallback path."
-            )
-        T_q_local = T_q // P
-
-    S_local = S_kv // C
-    n_tiles = (S_local + TILE_S_KV - 1) // TILE_S_KV
-    KV_ROW_BYTES = d_head * 2  # f16
-    Q = tl.load(q_ptr, shape=(T_q_local, d_head), dtype="f16")
-
-    # ── Bootstrap: (t=0, k=0) — load my own tile 0, establish (m, ℓ, O) ──
-    #
-    # Cannot be folded into the (t, k) loop (kernbench-only limitation):
-    #   - persistent (m, ℓ, O) must live OUTSIDE ``tl.scratch_scope``,
-    #     otherwise scope teardown discards them before the next tile's
-    #     merge can read them;
-    #   - kernbench has no scratch-backed initializer — ``tl.zeros`` /
-    #     ``tl.full`` return addr=0 handles with no backing storage, so
-    #     they cannot be overwritten via ``tl.copy_to`` to seed (-inf, 0, 0);
-    #   - the ring-step ordering also requires k=0 to send its loaded
-    #     tile W before any k>0 iteration can recv from E, so the very
-    #     first (t=0, k=0) step has to run outside the scoped loop body
-    #     where the send is conditional on ``C > 1``.
-    # Triton port: limitation does not apply (SSA tensors stay live
-    # across iterations); the bootstrap collapses into the loop.
-    tile_s = min(TILE_S_KV, S_local)
-    K_t = tl.load(k_ptr, shape=(d_head, tile_s), dtype="f16")
-    V_t = tl.load(v_ptr, shape=(tile_s, d_head), dtype="f16")
-    if C > 1:
-        tl.send(dir="W", src=K_t)
-        tl.send(dir="W", src=V_t)
-    scores = tl.dot(Q, K_t)
-    m = tl.max(scores, axis=-1)
-    exp_scores = tl.exp(scores - m)
-    l = tl.sum(exp_scores, axis=-1)
-    O = tl.dot(exp_scores, V_t)
-
-    # ── Nested loop: outer tile, inner ring step ──
-    # Each tile propagates all the way through the ring before the next
-    # tile starts; IPCQ in-flight depth stays at 1 per direction.
-    #
-    # Per outer ``t``:
-    #   k=0      : load my own tile ``t`` from HBM
-    #   k=1..C-1 : receive tile ``t`` of a peer's block from E
-    #              (sent by my E neighbour during their previous k step)
-    #   k<C-1    : forward the tile to W
-    #
-    # The ``(t=0, k=0)`` iteration is the bootstrap above; the loop skips
-    # it via ``k_start``. Every other iteration uses ``scratch_scope`` +
-    # ``copy_to`` to recycle per-tile intermediates.
-    #
-    # Triton port: drop ``with tl.scratch_scope():`` and replace each
-    # ``copy_to`` with a Python rebind (e.g. ``m = m_new``).
-    for t in range(n_tiles):
-        tile_start = t * TILE_S_KV
-        tile_s = min(TILE_S_KV, S_local - tile_start)
-        k_start = 1 if t == 0 else 0
-        for k in range(k_start, C):
-            with tl.scratch_scope():
-                if k == 0:
-                    K_t = tl.load(k_ptr + tile_start * KV_ROW_BYTES,
-                                  shape=(d_head, tile_s), dtype="f16")
-                    V_t = tl.load(v_ptr + tile_start * KV_ROW_BYTES,
-                                  shape=(tile_s, d_head), dtype="f16")
-                else:
-                    K_t = tl.recv(dir="E", shape=(d_head, tile_s), dtype="f16")
-                    V_t = tl.recv(dir="E", shape=(tile_s, d_head), dtype="f16")
-                if k < C - 1:
-                    tl.send(dir="W", src=K_t)
-                    tl.send(dir="W", src=V_t)
-                scores = tl.dot(Q, K_t)
-                m_step = tl.max(scores, axis=-1)
-                exp_scores = tl.exp(scores - m_step)
-                l_step = tl.sum(exp_scores, axis=-1)
-                O_step = tl.dot(exp_scores, V_t)
-                m_new, l_new, O_new = _merge_running(
-                    m, l, O, m_step, l_step, O_step, tl=tl,
-                )
-                tl.copy_to(m, m_new)
-                tl.copy_to(l, l_new)
-                tl.copy_to(O, O_new)
-
-    # ── Final normalise + store (each CUBE writes its own head) ──
-    O_final = O / l
-    tl.store(o_ptr, O_final)

src/kernbench/benches/gqa_helpers/__init__.py

@@ -0,0 +1,9 @@
+"""GQA attention kernels, panel runners, and helpers (internal subpackage).
+
+The bench-package auto-discovery in ``kernbench.benches.registry``
+skips subpackages, so this folder is not audited for ``@bench``
+decorators. The only public-facing GQA bench is
+``kernbench.benches.milestone_1h_gqa`` (umbrella that drives all
+panels by importing this subpackage's ``gqa_decode_long_ctx_4cases``
+and ``gqa_prefill_long_ctx_4cases`` modules).
+"""

src/kernbench/benches/gqa_helpers/long_ctx/__init__.py

@@ -0,0 +1 @@
+"""GQA long-context kernels + 4-cases comparative-study runners."""

src/kernbench/benches/gqa_helpers/long_ctx/_gqa_attention_prefill_long_ctx_cube_repl_pe_sp.py

@@ -0,0 +1,177 @@
+"""GQA prefill kernel — Case 3 (Cube-Repl × PE-SP) at single-KV-head group.
+
+Per GQA_full_deck.pptx slide 11 (prefill counterpart, T_q ≫ 1):
+  - K, V replicated across all 8 cubes (the 8× memory waste).
+  - PEs SP on S_kv inside each cube: each PE attends to its
+    ``S_local = S_kv / P`` slice.
+  - Q replicated on every rank (every cube does full T_q × S_local).
+  - Intra-CUBE 8-way reduce on the partial ``(m, ℓ, O)`` triple
+    per Q-tile (row chain along intra_W + col bridge along intra_N
+    over the 2×4 PE grid).
+  - NO inter-CUBE communication — every cube ends with the full
+    answer (8× redundant compute is the inherent cost of Case 3).
+  - Designated writer: only PE 0 of CUBE 0 stores ``O`` to avoid
+    8 redundant DMA writes.
+
+Q-axis tiling (vs decode Case 3): an outer Q-tile loop wraps the
+entire per-tile pipeline (Q load → bootstrap → S_kv-tile fold →
+intra-CUBE AR → store) in ``tl.scratch_scope``. Peak scratch is
+bounded by ``TILE_Q × TILE_S_KV`` instead of ``T_q × S_kv``, so the
+kernel runs at large T_q without exceeding the ~1 MB per-PE budget.
+Each Q-tile's (m, ℓ, O) is independent of other Q-tiles, so no
+state leaks across iterations.
+
+Tensor layout:
+  Q : (T_q, h_q · d_head) replicated on every rank; loaded per
+      Q-tile as ``(TILE_Q, d_head)`` slices (G is folded into the
+      T_q row axis, total ``G·T_q`` rows).
+  K : (S_kv, h_kv · d_head) with cube=replicate, pe=row_wise — each
+      PE owns its ``(S_local, h_kv·d_head)`` shard within its cube.
+  V : same as K.
+  O : (T_q, h_q · d_head) — only PE 0 of CUBE 0 stores, one
+      ``(TILE_Q, d_head)`` slice per outer Q-tile iteration.
+
+Topology / SFR:
+  - Requires intra_* lanes; ``configure_sfr_intercube_ring`` (with the
+    snake submesh) or ``configure_sfr_intercube_multisip`` both install
+    them.
+"""
+from __future__ import annotations
+
+
+TILE_S_KV = 64  # per-tile S_kv width (small to keep ``scores`` bounded).
+TILE_Q = 64     # per-tile Q row count (outer-loop tile size).
+
+
+def _merge_running(m_local, l_local, O_local, m_other, l_other, O_other, *, tl):
+    """Online-softmax merge of two partial ``(m, ℓ, O)`` triples."""
+    m_new = tl.maximum(m_local, m_other)
+    scale_old = tl.exp(m_local - m_new)
+    scale_new = tl.exp(m_other - m_new)
+    l_new = l_local * scale_old + l_other * scale_new
+    O_new = O_local * scale_old + O_other * scale_new
+    return m_new, l_new, O_new
+
+
+def gqa_attention_prefill_long_ctx_cube_repl_pe_sp_kernel(
+    q_ptr: int,
+    k_ptr: int,
+    v_ptr: int,
+    o_ptr: int,
+    T_q: int,
+    S_kv: int,
+    h_q: int,
+    h_kv: int,
+    d_head: int,
+    C: int,
+    P: int,
+    *,
+    tl,
+) -> None:
+    """Case-3 prefill: PE-SP S_kv split + Q-tiled intra-CUBE AR."""
+    G = h_q // h_kv
+    S_local = S_kv // P  # each PE owns S_kv/P (cube=replicate, pe=row_wise)
+    pe_id = tl.program_id(axis=0)
+    cube_id = tl.program_id(axis=1)
+
+    PE_GRID_COLS = 4
+    pe_col = pe_id % PE_GRID_COLS
+    pe_row = pe_id // PE_GRID_COLS
+    pe_cols_used = min(PE_GRID_COLS, P)
+    pe_rows_used = (P + PE_GRID_COLS - 1) // PE_GRID_COLS
+
+    KV_ROW_BYTES = d_head * 2  # f16
+    Q_ROW_BYTES = d_head * 2   # G is folded into the row axis, not d_head
+    total_q_rows = G * T_q
+    n_q_tiles = (total_q_rows + TILE_Q - 1) // TILE_Q
+    n_kv_tiles = (S_local + TILE_S_KV - 1) // TILE_S_KV
+
+    # ── Outer Q-tile loop ──
+    # Per Q-tile: load Q-slice, run full S_kv attention, intra-CUBE AR,
+    # store. All scratch is recycled at outer-scope exit.
+    for q_idx in range(n_q_tiles):
+        q_start = q_idx * TILE_Q
+        q_rows = min(TILE_Q, total_q_rows - q_start)
+        with tl.scratch_scope():
+            Q = tl.load(q_ptr + q_start * Q_ROW_BYTES,
+                        shape=(q_rows, d_head), dtype="f16")
+
+            # ── Bootstrap (S_kv tile 0) ──
+            # K_T, V, scores, exp_scores live in the OUTER (Q-tile)
+            # scope. They survive the inner S_kv-tile loop (each
+            # iteration's allocations are recycled) and are freed
+            # together when the outer scope exits.
+            tile_s0 = min(TILE_S_KV, S_local)
+            K_T = tl.load(k_ptr, shape=(d_head, tile_s0), dtype="f16")
+            V = tl.load(v_ptr, shape=(tile_s0, d_head), dtype="f16")
+            scores = tl.dot(Q, K_T)
+            m_local = tl.max(scores, axis=-1)
+            exp_scores = tl.exp(scores - m_local)
+            l_local = tl.sum(exp_scores, axis=-1)
+            O_local = tl.dot(exp_scores, V)
+
+            # ── S_kv tiles 1..N: fold via online-softmax merge ──
+            for tile_idx in range(1, n_kv_tiles):
+                tile_start = tile_idx * TILE_S_KV
+                tile_s = min(TILE_S_KV, S_local - tile_start)
+                with tl.scratch_scope():
+                    K_T_t = tl.load(k_ptr + tile_start * KV_ROW_BYTES,
+                                    shape=(d_head, tile_s), dtype="f16")
+                    V_t = tl.load(v_ptr + tile_start * KV_ROW_BYTES,
+                                  shape=(tile_s, d_head), dtype="f16")
+                    scores_t = tl.dot(Q, K_T_t)
+                    m_tile = tl.max(scores_t, axis=-1)
+                    exp_scores_t = tl.exp(scores_t - m_tile)
+                    l_tile = tl.sum(exp_scores_t, axis=-1)
+                    O_tile = tl.dot(exp_scores_t, V_t)
+                    m_new, l_new, O_new = _merge_running(
+                        m_local, l_local, O_local,
+                        m_tile, l_tile, O_tile, tl=tl,
+                    )
+                    tl.copy_to(m_local, m_new)
+                    tl.copy_to(l_local, l_new)
+                    tl.copy_to(O_local, O_new)
+
+            # ── Intra-CUBE 8-way reduce-to-PE0 (row chain + col bridge) ──
+            # (m_local, l_local, O_local) are still alive in the outer
+            # scope; the send/recv operate on them directly.
+            if pe_cols_used > 1:
+                if pe_col < pe_cols_used - 1:
+                    with tl.scratch_scope():
+                        m_other = tl.recv(dir="intra_E", shape=m_local.shape, dtype="f16")
+                        l_other = tl.recv(dir="intra_E", shape=l_local.shape, dtype="f16")
+                        O_other = tl.recv(dir="intra_E", shape=O_local.shape, dtype="f16")
+                        m_new, l_new, O_new = _merge_running(
+                            m_local, l_local, O_local,
+                            m_other, l_other, O_other, tl=tl,
+                        )
+                        tl.copy_to(m_local, m_new)
+                        tl.copy_to(l_local, l_new)
+                        tl.copy_to(O_local, O_new)
+                if pe_col > 0:
+                    tl.send(dir="intra_W", src=m_local)
+                    tl.send(dir="intra_W", src=l_local)
+                    tl.send(dir="intra_W", src=O_local)
+
+            if pe_col == 0 and pe_rows_used > 1:
+                if pe_row < pe_rows_used - 1:
+                    with tl.scratch_scope():
+                        m_other = tl.recv(dir="intra_S", shape=m_local.shape, dtype="f16")
+                        l_other = tl.recv(dir="intra_S", shape=l_local.shape, dtype="f16")
+                        O_other = tl.recv(dir="intra_S", shape=O_local.shape, dtype="f16")
+                        m_new, l_new, O_new = _merge_running(
+                            m_local, l_local, O_local,
+                            m_other, l_other, O_other, tl=tl,
+                        )
+                        tl.copy_to(m_local, m_new)
+                        tl.copy_to(l_local, l_new)
+                        tl.copy_to(O_local, O_new)
+                if pe_row > 0:
+                    tl.send(dir="intra_N", src=m_local)
+                    tl.send(dir="intra_N", src=l_local)
+                    tl.send(dir="intra_N", src=O_local)
+
+            # ── Store this Q-tile (designated writer: cube 0, PE 0) ──
+            if pe_id == 0 and cube_id == 0:
+                O_final = O_local / l_local
+                tl.store(o_ptr + q_start * Q_ROW_BYTES, O_final)

src/kernbench/benches/gqa_helpers/long_ctx/_gqa_attention_prefill_long_ctx_cube_repl_pe_tp.py

@@ -0,0 +1,128 @@
+"""GQA prefill kernel — Case 2 (Cube-Repl × PE-TP) at single-KV-head group.
+
+Per GQA_full_deck.pptx slide 11 (prefill counterpart, T_q ≫ 1):
+  - K, V replicated across all 8 cubes (the 8× memory waste — this is
+    the inherent cost Case 2 demonstrates). PE-replicate within each
+    cube too: every PE on every cube holds an HBM copy.
+  - PE-TP on T_q: only CUBE 0 has work; within CUBE 0 the P PEs split
+    the T_q axis disjointly. CUBEs 1..7 idle (slide-11 "PE-TP only
+    uses one cube; replicate is pure waste here").
+  - NO inter-rank communication (each active rank has full KV and
+    disjoint T_q rows — slide 11 lists comm cost as "none").
+
+Distinction from decode Case 2: decode T_q=1 makes PE-TP wasteful
+(only PE 0 of CUBE 0 works → 1 of 64 ranks active). Prefill T_q≫1
+makes PE-TP useful — all P PEs of CUBE 0 work on disjoint T_q rows
+(P of 64 ranks active; CUBEs 1..7 are pure memory waste).
+
+Q-axis tiling: an outer Q-tile loop wraps the entire per-tile
+pipeline (Q load → bootstrap → S_kv-tile fold → store) in
+``tl.scratch_scope``. Peak scratch is bounded by
+``TILE_Q × TILE_S_KV`` instead of ``T_q_local × S_kv``, so the
+kernel runs at large T_q without exceeding the ~1 MB per-PE budget.
+Each Q-tile is independent (disjoint output rows; no inter-tile state).
+
+Tensor layout:
+  Q : (T_q, h_q · d_head) cube=replicate, pe=row_wise on T_q. Only
+      CUBE 0's PEs load — each owns ``G · T_q_local`` total rows with
+      ``T_q_local = T_q / P``; loaded per Q-tile.
+  K : (S_kv, h_kv · d_head) replicated on every rank.
+  V : (S_kv, h_kv · d_head) replicated on every rank.
+  O : (T_q, h_q · d_head) cube=replicate, pe=row_wise on T_q —
+      only CUBE 0's PEs write their disjoint ``T_q_local`` rows,
+      one ``(TILE_Q, d_head)`` slice per outer iteration.
+
+Topology / SFR:
+  - SFR setup is benign (no inter-rank sends/recvs); any of the
+    ``configure_sfr_intercube_*`` helpers works.
+
+Requires ``T_q % P == 0``.
+"""
+from __future__ import annotations
+
+
+TILE_S_KV = 64  # per-tile S_kv width (small to keep ``scores`` bounded).
+TILE_Q = 64     # per-tile Q row count (outer-loop tile size).
+
+
+def gqa_attention_prefill_long_ctx_cube_repl_pe_tp_kernel(
+    q_ptr: int,
+    k_ptr: int,
+    v_ptr: int,
+    o_ptr: int,
+    T_q: int,
+    S_kv: int,
+    h_q: int,
+    h_kv: int,
+    d_head: int,
+    C: int,
+    P: int,
+    *,
+    tl,
+) -> None:
+    """Case-2 prefill: CUBE 0 only; Q-tiled PE-TP; full KV per rank; no comm."""
+    if T_q % P != 0:
+        raise ValueError(
+            f"Case 2 prefill requires T_q={T_q} divisible by P={P}"
+        )
+
+    cube_id = tl.program_id(axis=1)
+    # CUBE 0 only — CUBEs 1..7 are pure memory-replication waste, no work.
+    if cube_id != 0:
+        return
+
+    G = h_q // h_kv
+    T_q_local = T_q // P
+    n_kv_tiles = (S_kv + TILE_S_KV - 1) // TILE_S_KV
+    KV_ROW_BYTES = d_head * 2  # f16
+    Q_ROW_BYTES = d_head * 2
+    total_q_rows = G * T_q_local
+    n_q_tiles = (total_q_rows + TILE_Q - 1) // TILE_Q
+
+    # ── Outer Q-tile loop ──
+    # Per Q-tile: load Q-slice, run full S_kv attention, store. All
+    # scratch is recycled at outer-scope exit, so peak usage is bounded
+    # by TILE_Q × TILE_S_KV.
+    for q_idx in range(n_q_tiles):
+        q_start = q_idx * TILE_Q
+        q_rows = min(TILE_Q, total_q_rows - q_start)
+        with tl.scratch_scope():
+            Q = tl.load(q_ptr + q_start * Q_ROW_BYTES,
+                        shape=(q_rows, d_head), dtype="f16")
+
+            # ── Bootstrap (S_kv tile 0) ──
+            tile_s0 = min(TILE_S_KV, S_kv)
+            K_T = tl.load(k_ptr, shape=(d_head, tile_s0), dtype="f16")
+            V = tl.load(v_ptr, shape=(tile_s0, d_head), dtype="f16")
+            scores = tl.dot(Q, K_T)
+            m_local = tl.max(scores, axis=-1)
+            exp_scores = tl.exp(scores - m_local)
+            l_local = tl.sum(exp_scores, axis=-1)
+            O_local = tl.dot(exp_scores, V)
+
+            # ── S_kv tiles 1..N: fold via online-softmax merge ──
+            for tile_idx in range(1, n_kv_tiles):
+                tile_start = tile_idx * TILE_S_KV
+                tile_s = min(TILE_S_KV, S_kv - tile_start)
+                with tl.scratch_scope():
+                    K_T_t = tl.load(k_ptr + tile_start * KV_ROW_BYTES,
+                                    shape=(d_head, tile_s), dtype="f16")
+                    V_t = tl.load(v_ptr + tile_start * KV_ROW_BYTES,
+                                  shape=(tile_s, d_head), dtype="f16")
+                    scores_t = tl.dot(Q, K_T_t)
+                    m_tile = tl.max(scores_t, axis=-1)
+                    exp_scores_t = tl.exp(scores_t - m_tile)
+                    l_tile = tl.sum(exp_scores_t, axis=-1)
+                    O_tile = tl.dot(exp_scores_t, V_t)
+                    m_new = tl.maximum(m_local, m_tile)
+                    scale_old = tl.exp(m_local - m_new)
+                    scale_new = tl.exp(m_tile - m_new)
+                    l_new = l_local * scale_old + l_tile * scale_new
+                    O_new = O_local * scale_old + O_tile * scale_new
+                    tl.copy_to(m_local, m_new)
+                    tl.copy_to(l_local, l_new)
+                    tl.copy_to(O_local, O_new)
+
+            # ── Store this Q-tile (each active PE writes its slice) ──
+            O_final = O_local / l_local
+            tl.store(o_ptr + q_start * Q_ROW_BYTES, O_final)

src/kernbench/benches/gqa_helpers/long_ctx/_gqa_attention_prefill_long_ctx_cube_sp_pe_sp.py

@@ -0,0 +1,186 @@
+"""GQA prefill kernel — Case 4 (Cube-SP × PE-SP) at single-KV-head group. ★
+
+Per GQA_full_deck.pptx slide 11 (prefill counterpart, T_q ≫ 1):
+  - K, V split 64-way (cube=row_wise, pe=row_wise on S_kv); each rank
+    owns ``S_local_pe = S_kv / (C·P)``.
+  - Q, O split T_q-wise across cubes (cube=row_wise on T_q),
+    pe=replicate within each cube — each cube owns
+    ``T_q_cube = T_q / C`` rows, all P PEs of the cube hold a copy.
+  - Inter-CUBE Ring KV (snake over 4×2 sub-mesh, ADR-0060 §5.5):
+    P parallel rings — each PE drives its own same-lane ring via
+    independent IPCQ channels. Over C ring steps each PE has
+    processed its strided lane of S_kv (= ``S_kv / P`` tokens) for
+    the cube's T_q_cube query rows.
+  - Intra-CUBE 8-way reduce on the partial ``(m, ℓ, O)`` triple
+    (row chain along intra_W + col bridge along intra_N over the
+    2×4 PE grid) — combines the P PE-lanes per cube so cube's PE 0
+    holds the full S_kv attention for its T_q_cube rows.
+  - NO inter-CUBE reduce — each cube writes its own disjoint
+    T_q_cube rows of O (PE 0 of each cube is the writer).
+
+Distinction from decode Case 4: decode T_q=1 forces an inter-CUBE
+lrab reduce-to-root over (m,ℓ,O). Prefill T_q≫1 allows Q to be
+T_q-sharded across cubes, so disjoint output rows avoid the
+inter-cube reduce — the prefill-native pattern (per user's
+"Ring KV" choice).
+
+Tensor layout:
+  Q : (T_q, h_q · d_head) sharded cube=row_wise on T_q, pe=replicate
+      within cube; each rank loads ``(G · T_q_cube, d_head)``.
+  K : (S_kv, h_kv · d_head) cube=row_wise + pe=row_wise on S_kv;
+      each rank owns ``(S_local_pe, h_kv·d_head)``.
+  V : same as K.
+  O : (T_q, h_q · d_head) sharded same as Q — PE 0 of each cube
+      writes its ``(G · T_q_cube, d_head)`` slice; PEs 1..P-1 idle
+      the write.
+
+Topology / SFR:
+  - Requires ``configure_sfr_intercube_ring(submesh_shape=(2, 4))`` —
+    that helper installs both the snake E/W ring across 8 cubes and
+    the intra_* lanes needed for the intra-CUBE reduce.
+
+Requires ``T_q % C == 0`` and ``S_kv % (C · P) == 0``.
+"""
+from __future__ import annotations
+
+
+TILE_S_KV = 1024  # ADR-0063 §A.2 S_kv-axis tile sweep (per-tile width).
+
+
+def _merge_running(m_local, l_local, O_local, m_other, l_other, O_other, *, tl):
+    """Online-softmax merge of two partial ``(m, ℓ, O)`` triples."""
+    m_new = tl.maximum(m_local, m_other)
+    scale_old = tl.exp(m_local - m_new)
+    scale_new = tl.exp(m_other - m_new)
+    l_new = l_local * scale_old + l_other * scale_new
+    O_new = O_local * scale_old + O_other * scale_new
+    return m_new, l_new, O_new
+
+
+def gqa_attention_prefill_long_ctx_cube_sp_pe_sp_kernel(
+    q_ptr: int,
+    k_ptr: int,
+    v_ptr: int,
+    o_ptr: int,
+    T_q: int,
+    S_kv: int,
+    h_q: int,
+    h_kv: int,
+    d_head: int,
+    C: int,
+    P: int,
+    *,
+    tl,
+) -> None:
+    """Case-4 prefill: Ring KV across cubes + intra-CUBE AR; disjoint T_q/C rows per cube."""
+    if T_q % C != 0:
+        raise ValueError(
+            f"Case 4 prefill requires T_q={T_q} divisible by C={C}"
+        )
+    if S_kv % (C * P) != 0:
+        raise ValueError(
+            f"Case 4 prefill requires S_kv={S_kv} divisible by C·P={C * P}"
+        )
+
+    G = h_q // h_kv
+    T_q_cube = T_q // C
+    S_local_pe = S_kv // (C * P)
+    KV_ROW_BYTES = d_head * 2  # f16
+    n_tiles = (S_local_pe + TILE_S_KV - 1) // TILE_S_KV
+
+    pe_id = tl.program_id(axis=0)
+
+    # Q is this cube's T_q slice (pe=replicate within cube ⇒ every PE
+    # loads the same G·T_q_cube rows).
+    Q = tl.load(q_ptr, shape=(G * T_q_cube, d_head), dtype="f16")
+
+    # ── Bootstrap (t=0, k=0): load own rank's tile 0; send W for ring ──
+    # Persistent (m, ℓ, O) must live OUTSIDE tl.scratch_scope (kernbench
+    # scope teardown discards in-scope tensors).
+    tile_s = min(TILE_S_KV, S_local_pe)
+    K_T = tl.load(k_ptr, shape=(d_head, tile_s), dtype="f16")
+    V = tl.load(v_ptr, shape=(tile_s, d_head), dtype="f16")
+    if C > 1:
+        tl.send(dir="W", src=K_T)
+        tl.send(dir="W", src=V)
+    scores = tl.dot(Q, K_T)
+    m_local = tl.max(scores, axis=-1)
+    exp_scores = tl.exp(scores - m_local)
+    l_local = tl.sum(exp_scores, axis=-1)
+    O_local = tl.dot(exp_scores, V)
+
+    # ── Nested loop (outer tile, inner ring step) — P parallel rings ──
+    for t in range(n_tiles):
+        tile_start = t * TILE_S_KV
+        tile_s = min(TILE_S_KV, S_local_pe - tile_start)
+        k_start = 1 if t == 0 else 0
+        for k in range(k_start, C):
+            with tl.scratch_scope():
+                if k == 0:
+                    K_T_t = tl.load(k_ptr + tile_start * KV_ROW_BYTES,
+                                    shape=(d_head, tile_s), dtype="f16")
+                    V_t = tl.load(v_ptr + tile_start * KV_ROW_BYTES,
+                                  shape=(tile_s, d_head), dtype="f16")
+                else:
+                    K_T_t = tl.recv(dir="E", shape=(d_head, tile_s), dtype="f16")
+                    V_t = tl.recv(dir="E", shape=(tile_s, d_head), dtype="f16")
+                if k < C - 1:
+                    tl.send(dir="W", src=K_T_t)
+                    tl.send(dir="W", src=V_t)
+                scores_t = tl.dot(Q, K_T_t)
+                m_step = tl.max(scores_t, axis=-1)
+                exp_scores_t = tl.exp(scores_t - m_step)
+                l_step = tl.sum(exp_scores_t, axis=-1)
+                O_step = tl.dot(exp_scores_t, V_t)
+                m_new, l_new, O_new = _merge_running(
+                    m_local, l_local, O_local, m_step, l_step, O_step, tl=tl,
+                )
+                tl.copy_to(m_local, m_new)
+                tl.copy_to(l_local, l_new)
+                tl.copy_to(O_local, O_new)
+
+    # ── Intra-CUBE 8-way reduce-to-PE0 (row chain + col bridge) ──
+    PE_GRID_COLS = 4
+    pe_col = pe_id % PE_GRID_COLS
+    pe_row = pe_id // PE_GRID_COLS
+    pe_cols_used = min(PE_GRID_COLS, P)
+    pe_rows_used = (P + PE_GRID_COLS - 1) // PE_GRID_COLS
+
+    if pe_cols_used > 1:
+        if pe_col < pe_cols_used - 1:
+            with tl.scratch_scope():
+                m_other = tl.recv(dir="intra_E", shape=m_local.shape, dtype="f16")
+                l_other = tl.recv(dir="intra_E", shape=l_local.shape, dtype="f16")
+                O_other = tl.recv(dir="intra_E", shape=O_local.shape, dtype="f16")
+                m_new, l_new, O_new = _merge_running(
+                    m_local, l_local, O_local, m_other, l_other, O_other, tl=tl,
+                )
+                tl.copy_to(m_local, m_new)
+                tl.copy_to(l_local, l_new)
+                tl.copy_to(O_local, O_new)
+        if pe_col > 0:
+            tl.send(dir="intra_W", src=m_local)
+            tl.send(dir="intra_W", src=l_local)
+            tl.send(dir="intra_W", src=O_local)
+
+    if pe_col == 0 and pe_rows_used > 1:
+        if pe_row < pe_rows_used - 1:
+            with tl.scratch_scope():
+                m_other = tl.recv(dir="intra_S", shape=m_local.shape, dtype="f16")
+                l_other = tl.recv(dir="intra_S", shape=l_local.shape, dtype="f16")
+                O_other = tl.recv(dir="intra_S", shape=O_local.shape, dtype="f16")
+                m_new, l_new, O_new = _merge_running(
+                    m_local, l_local, O_local, m_other, l_other, O_other, tl=tl,
+                )
+                tl.copy_to(m_local, m_new)
+                tl.copy_to(l_local, l_new)
+                tl.copy_to(O_local, O_new)
+        if pe_row > 0:
+            tl.send(dir="intra_N", src=m_local)
+            tl.send(dir="intra_N", src=l_local)
+            tl.send(dir="intra_N", src=O_local)
+
+    # ── Final normalise + store (PE 0 of each cube writes its T_q_cube rows) ──
+    if pe_id == 0:
+        O_final = O_local / l_local
+        tl.store(o_ptr, O_final)

src/kernbench/benches/gqa_helpers/long_ctx/_gqa_attention_prefill_long_ctx_cube_sp_pe_tp.py

@@ -0,0 +1,136 @@
+"""GQA prefill kernel — Case 1 (Cube-SP × PE-TP) at single-KV-head group.
+
+Per GQA_full_deck.pptx slide 11 (prefill counterpart, T_q ≫ 1):
+  - K, V split S_kv-wise across the 8 cubes (cube=row_wise);
+    replicated within each cube (pe=replicate). Each cube owns
+    ``S_local = S_kv / C``.
+  - Q, O split T_q-wise across cubes AND across PEs intra-cube
+    (cube=row_wise, pe=row_wise on T_q). Each rank owns disjoint
+    ``T_q_local = T_q / (C · P)`` query rows.
+  - Inter-CUBE Ring KV (snake over 4×2 sub-mesh, ADR-0060 §5.5):
+    P parallel rings — each PE drives its own same-lane ring via
+    independent IPCQ channels. Over C ring steps the cube-owned
+    K/V blocks rotate through all C cubes, so every rank sees
+    every block.
+  - NO inter-CUBE reduce — each rank writes its own disjoint
+    T_q_local rows of O.
+  - NO intra-CUBE reduce — PEs are disjoint on T_q.
+
+Distinction from decode Case 1: decode T_q=1 makes PE-TP wasteful
+(only PE 0 of each cube works). Prefill T_q≫1 makes PE-TP useful —
+all 64 ranks work on disjoint Q rows.
+
+Tensor layout:
+  Q : (T_q, h_q · d_head) sharded (cube_row_wise, pe_row_wise) on T_q;
+      each rank loads ``(G · T_q_local, d_head)``.
+  K : (S_kv, h_kv · d_head) with cube=row_wise, pe=replicate — each
+      cube's PE 0..P-1 each hold an HBM copy of the cube's
+      ``(S_local, h_kv·d_head)`` shard.
+  V : same as K.
+  O : (T_q, h_q · d_head) sharded same as Q — each rank writes its
+      own ``(G · T_q_local, d_head)`` slice.
+
+Topology / SFR:
+  - Requires ``configure_sfr_intercube_ring(submesh_shape=(2, 4))``
+    for the snake E/W ring across 8 cubes (wrap from cube 3 ↔ cube 7).
+
+Requires ``T_q % (C · P) == 0``.
+"""
+from __future__ import annotations
+
+
+TILE_S_KV = 1024  # ADR-0063 §A.2 S_kv-axis tile sweep (per-tile width).
+
+
+def _merge_running(m_local, l_local, O_local, m_other, l_other, O_other, *, tl):
+    """Online-softmax merge of two partial ``(m, ℓ, O)`` triples."""
+    m_new = tl.maximum(m_local, m_other)
+    scale_old = tl.exp(m_local - m_new)
+    scale_new = tl.exp(m_other - m_new)
+    l_new = l_local * scale_old + l_other * scale_new
+    O_new = O_local * scale_old + O_other * scale_new
+    return m_new, l_new, O_new
+
+
+def gqa_attention_prefill_long_ctx_cube_sp_pe_tp_kernel(
+    q_ptr: int,
+    k_ptr: int,
+    v_ptr: int,
+    o_ptr: int,
+    T_q: int,
+    S_kv: int,
+    h_q: int,
+    h_kv: int,
+    d_head: int,
+    C: int,
+    P: int,
+    *,
+    tl,
+) -> None:
+    """Case-1 prefill: Ring KV across cubes; disjoint T_q rows per rank."""
+    if T_q % (C * P) != 0:
+        raise ValueError(
+            f"Case 1 prefill requires T_q={T_q} divisible by C·P={C * P}"
+        )
+
+    G = h_q // h_kv
+    T_q_local = T_q // (C * P)
+    S_local = S_kv // C
+    KV_ROW_BYTES = d_head * 2  # f16
+    n_tiles = (S_local + TILE_S_KV - 1) // TILE_S_KV
+
+    # Q is this rank's own disjoint T_q slice (G·T_q_local rows for the
+    # G-grouped attention).
+    Q = tl.load(q_ptr, shape=(G * T_q_local, d_head), dtype="f16")
+
+    # ── Bootstrap (t=0, k=0): load own cube's tile 0; send W for ring ──
+    # Persistent (m, ℓ, O) must live OUTSIDE tl.scratch_scope (kernbench
+    # scope teardown discards in-scope tensors). Mirrors decode Cases
+    # 1, 3, 4 and the existing prefill_long bootstrap pattern.
+    tile_s = min(TILE_S_KV, S_local)
+    K_T = tl.load(k_ptr, shape=(d_head, tile_s), dtype="f16")
+    V = tl.load(v_ptr, shape=(tile_s, d_head), dtype="f16")
+    if C > 1:
+        tl.send(dir="W", src=K_T)
+        tl.send(dir="W", src=V)
+    scores = tl.dot(Q, K_T)
+    m_local = tl.max(scores, axis=-1)
+    exp_scores = tl.exp(scores - m_local)
+    l_local = tl.sum(exp_scores, axis=-1)
+    O_local = tl.dot(exp_scores, V)
+
+    # ── Nested loop (outer tile, inner ring step) ──
+    # Each tile propagates around the ring before the next tile starts;
+    # IPCQ in-flight depth stays at 1 per direction per PE-lane.
+    for t in range(n_tiles):
+        tile_start = t * TILE_S_KV
+        tile_s = min(TILE_S_KV, S_local - tile_start)
+        k_start = 1 if t == 0 else 0
+        for k in range(k_start, C):
+            with tl.scratch_scope():
+                if k == 0:
+                    K_T_t = tl.load(k_ptr + tile_start * KV_ROW_BYTES,
+                                    shape=(d_head, tile_s), dtype="f16")
+                    V_t = tl.load(v_ptr + tile_start * KV_ROW_BYTES,
+                                  shape=(tile_s, d_head), dtype="f16")
+                else:
+                    K_T_t = tl.recv(dir="E", shape=(d_head, tile_s), dtype="f16")
+                    V_t = tl.recv(dir="E", shape=(tile_s, d_head), dtype="f16")
+                if k < C - 1:
+                    tl.send(dir="W", src=K_T_t)
+                    tl.send(dir="W", src=V_t)
+                scores_t = tl.dot(Q, K_T_t)
+                m_step = tl.max(scores_t, axis=-1)
+                exp_scores_t = tl.exp(scores_t - m_step)
+                l_step = tl.sum(exp_scores_t, axis=-1)
+                O_step = tl.dot(exp_scores_t, V_t)
+                m_new, l_new, O_new = _merge_running(
+                    m_local, l_local, O_local, m_step, l_step, O_step, tl=tl,
+                )
+                tl.copy_to(m_local, m_new)
+                tl.copy_to(l_local, l_new)
+                tl.copy_to(O_local, O_new)
+
+    # ── Final normalise + store (every rank writes its own T_q_local rows) ──
+    O_final = O_local / l_local
+    tl.store(o_ptr, O_final)

src/kernbench/benches/gqa_helpers/long_ctx/gqa_decode_long_ctx_4cases.py

@@ -26,32 +26,32 @@ import json
 import os
 from pathlib import Path
 
-from kernbench.benches._gqa_attention_decode_long_ctx_cube_repl_pe_sp import (
+from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_decode_long_ctx_cube_repl_pe_sp import (
     gqa_attention_decode_long_ctx_cube_repl_pe_sp_kernel,
 )
-from kernbench.benches._gqa_attention_decode_long_ctx_cube_repl_pe_tp import (
+from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_decode_long_ctx_cube_repl_pe_tp import (
     gqa_attention_decode_long_ctx_cube_repl_pe_tp_kernel,
 )
-from kernbench.benches._gqa_attention_decode_long_ctx_cube_sp_pe_sp import (
+from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_decode_long_ctx_cube_sp_pe_sp import (
     gqa_attention_decode_long_ctx_cube_sp_pe_sp_kernel,
 )
-from kernbench.benches._gqa_attention_decode_long_ctx_cube_sp_pe_tp import (
+from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_decode_long_ctx_cube_sp_pe_tp import (
     gqa_attention_decode_long_ctx_cube_sp_pe_tp_kernel,
 )
-from kernbench.benches.milestone_gqa_headline import (
+from kernbench.benches.gqa_helpers.shared._gqa_panel_helpers import (
     _ccl_cfg,
     _summarize_op_log,
 )
-from kernbench.benches.registry import bench
 from kernbench.ccl.sfr_config import configure_sfr_intercube_multisip
 from kernbench.policy.placement.dp import DPPolicy
 
+# File is at benches/gqa_helpers/long_ctx/ — go up 2 parents to reach
+# benches/, then into 1H_milestone_output/gqa/long_ctx/.
 _OUTPUT_DIR = (
-    Path(__file__).resolve().parent
-    / "1H_milestone_output"
-    / "gqa_decode_long_ctx_4cases"
+    Path(__file__).resolve().parents[2]
+    / "1H_milestone_output" / "gqa" / "long_ctx"
 )
-_SWEEP_JSON = _OUTPUT_DIR / "sweep.json"
+_SWEEP_JSON = _OUTPUT_DIR / "sweep_decode.json"
 
 
 # ── Panel registry ───────────────────────────────────────────────────
@@ -75,7 +75,7 @@ _PANEL_DISPATCH: dict[str, tuple[str, dict]] = {
         # The sub_w=4/sub_h=2 lrab center-root geometry (root cube 6) is
         # baked into the Case-4 wrapper kernel.
         "C": 8, "P": 8,
-        "T_q": 1, "S_kv": 131_072,
+        "T_q": 1, "S_kv": 8_192,
         "d_head": 128, "h_q": 8, "h_kv": 1,
     }),
     "single_kv_group_decode_long_ctx_gqa_cube_repl_pe_tp": ("decode_long_ctx_cube_repl_pe_tp", {
@@ -83,7 +83,7 @@ _PANEL_DISPATCH: dict[str, tuple[str, dict]] = {
         # on batch. For B=1 only one rank works (slide-11 PE-TP waste).
         # No inter-rank communication.
         "C": 8, "P": 8,
-        "T_q": 1, "S_kv": 131_072,
+        "T_q": 1, "S_kv": 8_192,
         "d_head": 128, "h_q": 8, "h_kv": 1,
     }),
     "single_kv_group_decode_long_ctx_gqa_cube_repl_pe_sp": ("decode_long_ctx_cube_repl_pe_sp", {
@@ -91,7 +91,7 @@ _PANEL_DISPATCH: dict[str, tuple[str, dict]] = {
         # within each cube. Intra-CUBE 8-way reduce; no inter-CUBE comm
         # (every cube ends with full answer; designated writer = cube 0).
         "C": 8, "P": 8,
-        "T_q": 1, "S_kv": 131_072,
+        "T_q": 1, "S_kv": 8_192,
         "d_head": 128, "h_q": 8, "h_kv": 1,
     }),
     "single_kv_group_decode_long_ctx_gqa_cube_sp_pe_tp": ("decode_long_ctx_cube_sp_pe_tp", {
@@ -99,7 +99,7 @@ _PANEL_DISPATCH: dict[str, tuple[str, dict]] = {
         # PEs TP on batch — at B=1 only PE 0 of each cube works (PE-TP
         # waste). Inter-CUBE lrab AR (root = cube 6); no intra-CUBE comm.
         "C": 8, "P": 8,
-        "T_q": 1, "S_kv": 131_072,
+        "T_q": 1, "S_kv": 8_192,
         "d_head": 128, "h_q": 8, "h_kv": 1,
     }),
 }
@@ -316,31 +316,12 @@ def _run_panel(panel: str, topology: str) -> dict:
     }
 
 
-# ── Bench entry ──────────────────────────────────────────────────────
+# ── Sweep entry (called by the umbrella ``milestone_1h_gqa``) ────────
 
 
-@bench(
-    name="milestone-gqa-decode-long-ctx-4cases",
-    description=(
-        "Long-context decode 4-cases comparative study on the "
-        "LLaMA-3.1-70B single-KV-head group (8 cubes × 8 PEs)."
-    ),
-)
-def run(torch) -> None:
-    """Drive the registered decode case panels; write sweep.json.
-
-    Gated by GQA_DECODE_LONG_CTX_4CASES_RUN=1.
-    """
+def run_sweep(topology: str = "topology.yaml") -> int:
+    """Drive all decode case panels; write sweep.json. Returns row count."""
     _OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
-    if not os.environ.get("GQA_DECODE_LONG_CTX_4CASES_RUN"):
-        raise RuntimeError(
-            "milestone-gqa-decode-long-ctx-4cases needs "
-            "GQA_DECODE_LONG_CTX_4CASES_RUN=1."
-        )
-
-    topology = os.environ.get(
-        "GQA_DECODE_LONG_CTX_4CASES_TOPOLOGY", "topology.yaml",
-    )
     rows = [_run_panel(panel, topology) for panel in _PANELS]
     sweep = {
         "version": 1,
@@ -348,6 +329,5 @@ def run(torch) -> None:
         "rows": rows,
     }
     _SWEEP_JSON.write_text(json.dumps(sweep, indent=2))
-    print(
-        f"  milestone-gqa-decode-long-ctx-4cases: {len(rows)} rows -> {_SWEEP_JSON}"
-    )
+    print(f"  gqa-decode-long-ctx-4cases: {len(rows)} rows -> {_SWEEP_JSON}")
+    return len(rows)

src/kernbench/benches/gqa_helpers/long_ctx/gqa_prefill_long_ctx_4cases.py

@@ -0,0 +1,384 @@
+"""milestone-gqa-prefill-long-ctx-4cases: long-context prefill 4-cases study.
+
+Prefill counterpart to milestone-gqa-decode-long-ctx-4cases. Per
+GQA_full_deck.pptx slides 11-17: 4 KV-cache sharding strategies on
+the LLaMA-3.1-70B single-KV-head group (1 KV head, 8 Q heads,
+d_head=128) on 8 cubes × 8 PEs.
+
+Bench size T_q=S_kv=512 (equal — prompt length matches K/V length
+for prefill). Cases 2 and 3 use Q-axis tiling (``TILE_Q``) so the
+per-PE scratch is bounded by ``TILE_Q × TILE_S_KV`` regardless of
+T_q — much larger sizes are possible at the cost of bench
+wall-clock (Case 2's single-cube serial path dominates).
+
+    Case 1  Cube-SP / PE-TP  → KV split S_kv across cubes (Ring),
+                               Q/O split T_q across all 64 ranks
+    Case 2  Cube-Repl / PE-TP → full KV per cube; only CUBE 0's P PEs
+                               split T_q; CUBEs 1..7 are pure memory waste
+    Case 3  Cube-Repl / PE-SP → full KV per cube; PEs SP on S_kv
+                               (intra-cube AR); 8× redundant compute
+    Case 4  Cube-SP / PE-SP  → KV split 64-way + Q split T_q across cubes;
+                               Ring KV + intra-cube AR per cube  ★ optimal
+
+Each case is a separate panel. The bench drives all panels in one
+invocation and writes per-panel op_log_summary + latency to sweep.json
+so the comparative analysis (latency, comm volume, parallelism,
+memory) can be generated from a single sweep.
+
+Distinction from decode: decode T_q=1 makes PE-TP cases wasteful (1
+active PE). Prefill T_q≫1 makes PE-TP useful (full parallelism). The
+narrative differs — the figures emphasise Case 2/3 memory waste and
+Case 3 compute redundancy, not PE-TP-at-B=1 wasted ranks.
+
+Per-panel try/except in ``run()`` writes whichever cases succeed —
+the bench tolerates per-panel scratch / SFR failures so sweep.json
+always lands with at least the successful rows + a ``failures`` list.
+
+Gated by ``GQA_PREFILL_LONG_CTX_4CASES_RUN=1``.
+"""
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+
+from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_prefill_long_ctx_cube_repl_pe_sp import (
+    gqa_attention_prefill_long_ctx_cube_repl_pe_sp_kernel,
+)
+from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_prefill_long_ctx_cube_repl_pe_tp import (
+    gqa_attention_prefill_long_ctx_cube_repl_pe_tp_kernel,
+)
+from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_prefill_long_ctx_cube_sp_pe_sp import (
+    gqa_attention_prefill_long_ctx_cube_sp_pe_sp_kernel,
+)
+from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_prefill_long_ctx_cube_sp_pe_tp import (
+    gqa_attention_prefill_long_ctx_cube_sp_pe_tp_kernel,
+)
+from kernbench.benches.gqa_helpers.shared._gqa_panel_helpers import (
+    _ccl_cfg,
+    _summarize_op_log,
+)
+from kernbench.ccl.sfr_config import configure_sfr_intercube_ring
+from kernbench.policy.placement.dp import DPPolicy
+
+# File is at benches/gqa_helpers/long_ctx/ — go up 2 parents to reach
+# benches/, then into 1H_milestone_output/gqa/long_ctx/.
+_OUTPUT_DIR = (
+    Path(__file__).resolve().parents[2]
+    / "1H_milestone_output" / "gqa" / "long_ctx"
+)
+_SWEEP_JSON = _OUTPUT_DIR / "sweep_prefill.json"
+
+
+# ── Panel registry ───────────────────────────────────────────────────
+
+
+_PANELS = (
+    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_sp",      # Case 4 ★ optimal
+    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_tp",    # Case 2
+    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_sp",    # Case 3
+    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_tp",      # Case 1
+)
+
+# Each entry: (kind, panel-specific params).
+# LLaMA-3.1-70B single-KV-head group target:
+#   1 KV head, h_q = 8 (G = 8 group), d_head = 128
+#   8 cubes (head-parallel group), 8 PEs/cube
+#   T_q = S_kv = 4096 (long-context prefill)
+_PANEL_DISPATCH: dict[str, tuple[str, dict]] = {
+    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_sp": ("prefill_long_ctx_cube_sp_pe_sp", {
+        # Case 4: KV split 64-way + Q split T_q across cubes; Ring KV
+        # + intra-CUBE AR per cube. PE 0 of each cube writes its
+        # T_q/C disjoint output rows.
+        "C": 8, "P": 8,
+        "T_q": 512, "S_kv": 512,
+        "d_head": 128, "h_q": 8, "h_kv": 1,
+    }),
+    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_tp": ("prefill_long_ctx_cube_repl_pe_tp", {
+        # Case 2: K, V replicated everywhere (8× memory waste); within
+        # CUBE 0 the P PEs split T_q disjointly. CUBEs 1..7 idle.
+        # No inter-rank communication.
+        "C": 8, "P": 8,
+        "T_q": 512, "S_kv": 512,
+        "d_head": 128, "h_q": 8, "h_kv": 1,
+    }),
+    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_sp": ("prefill_long_ctx_cube_repl_pe_sp", {
+        # Case 3: K, V replicated per cube (8× memory); PEs SP on S_kv
+        # within each cube. Intra-CUBE 8-way reduce; no inter-CUBE comm
+        # (every cube does redundant T_q × S_kv compute; designated
+        # writer = cube 0).
+        "C": 8, "P": 8,
+        "T_q": 512, "S_kv": 512,
+        "d_head": 128, "h_q": 8, "h_kv": 1,
+    }),
+    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_tp": ("prefill_long_ctx_cube_sp_pe_tp", {
+        # Case 1: K, V split S_kv across cubes (S_local = S_kv/C per
+        # cube); Q/O split T_q across all 64 ranks. Inter-CUBE snake
+        # Ring KV (P parallel rings); no intra-CUBE comm.
+        "C": 8, "P": 8,
+        "T_q": 512, "S_kv": 512,
+        "d_head": 128, "h_q": 8, "h_kv": 1,
+    }),
+}
+
+
+# ── Per-panel runner ─────────────────────────────────────────────────
+
+
+def _ring_sfr(ctx):
+    """Install snake Ring SFR (also provides intra_* lanes for AR cases).
+
+    Snake submesh of shape (2, 4) over the default 4×4 cube mesh
+    yields a Hamiltonian cycle through all 8 cubes
+    (0→1→2→3→7→6→5→4→0), with intra_* lanes available for the
+    Case 3/4 intra-CUBE reduce.
+    """
+    configure_sfr_intercube_ring(
+        ctx.engine, ctx.spec, _ccl_cfg(),
+        submesh_shape=(2, 4),
+    )
+
+
+def _run_prefill_panel_long_ctx_cube_repl_pe_tp(
+    ctx, *, panel: str, C: int, P: int,
+    T_q: int, S_kv: int,
+    d_head: int, h_q: int, h_kv: int,
+) -> None:
+    """Case 2 runner: K, V replicated everywhere; CUBE 0 PE-TP on T_q.
+
+    DPPolicy models the cluster-wide 8× memory waste — every rank
+    holds full K, V in its HBM region. Within CUBE 0, P PEs split
+    T_q (so P of 64 ranks actually compute). CUBEs 1..7 idle.
+    """
+    _ring_sfr(ctx)
+    dp_repl = DPPolicy(cube="replicate", pe="replicate",
+                       num_cubes=C, num_pes=P)
+    dp_qo = DPPolicy(cube="replicate", pe="row_wise",
+                     num_cubes=C, num_pes=P)
+    q = ctx.zeros((T_q, h_q * d_head),
+                  dtype="f16", dp=dp_qo, name=f"{panel}_q")
+    k = ctx.zeros((S_kv, h_kv * d_head),
+                  dtype="f16", dp=dp_repl, name=f"{panel}_k")
+    v = ctx.zeros((S_kv, h_kv * d_head),
+                  dtype="f16", dp=dp_repl, name=f"{panel}_v")
+    o = ctx.empty((T_q, h_q * d_head),
+                  dtype="f16", dp=dp_qo, name=f"{panel}_o")
+    ctx.launch(
+        panel, gqa_attention_prefill_long_ctx_cube_repl_pe_tp_kernel,
+        q, k, v, o,
+        T_q, S_kv, h_q, h_kv, d_head, C, P,
+        _auto_dim_remap=False,
+    )
+
+
+def _run_prefill_panel_long_ctx_cube_repl_pe_sp(
+    ctx, *, panel: str, C: int, P: int,
+    T_q: int, S_kv: int,
+    d_head: int, h_q: int, h_kv: int,
+) -> None:
+    """Case 3 runner: K, V replicated per cube; PEs SP on S_kv intra-cube.
+
+    DPPolicy models the cluster-wide 8× memory waste — every cube
+    holds full K, V; intra-cube splits the S_kv axis row_wise across
+    8 PEs. Every rank holds full Q (cube=replicate, pe=replicate).
+    The kernel does an intra-CUBE 8-way reduce on (m, ℓ, O); only
+    cube 0's PE 0 writes the output.
+    """
+    _ring_sfr(ctx)
+    dp_full = DPPolicy(cube="replicate", pe="replicate",
+                       num_cubes=C, num_pes=P)
+    dp_kv = DPPolicy(cube="replicate", pe="row_wise",
+                     num_cubes=C, num_pes=P)
+    q = ctx.zeros((T_q, h_q * d_head),
+                  dtype="f16", dp=dp_full, name=f"{panel}_q")
+    k = ctx.zeros((S_kv, h_kv * d_head),
+                  dtype="f16", dp=dp_kv, name=f"{panel}_k")
+    v = ctx.zeros((S_kv, h_kv * d_head),
+                  dtype="f16", dp=dp_kv, name=f"{panel}_v")
+    o = ctx.empty((T_q, h_q * d_head),
+                  dtype="f16", dp=dp_full, name=f"{panel}_o")
+    ctx.launch(
+        panel, gqa_attention_prefill_long_ctx_cube_repl_pe_sp_kernel,
+        q, k, v, o,
+        T_q, S_kv, h_q, h_kv, d_head, C, P,
+        _auto_dim_remap=False,
+    )
+
+
+def _run_prefill_panel_long_ctx_cube_sp_pe_tp(
+    ctx, *, panel: str, C: int, P: int,
+    T_q: int, S_kv: int,
+    d_head: int, h_q: int, h_kv: int,
+) -> None:
+    """Case 1 runner: K, V split S_kv across cubes (Ring); Q/O split T_q 64-way.
+
+    DPPolicy: K, V are cube=row_wise on S_kv (S_local = S_kv/C per
+    cube), pe=replicate within cube. Q, O are cube=row_wise +
+    pe=row_wise on T_q — every rank owns T_q/(C·P) disjoint Q rows.
+    The kernel runs P parallel Rings through the snake submesh
+    (each PE has its own IPCQ lane); no intra-CUBE reduce.
+    """
+    _ring_sfr(ctx)
+    dp_qo = DPPolicy(cube="row_wise", pe="row_wise",
+                     num_cubes=C, num_pes=P)
+    dp_kv = DPPolicy(cube="row_wise", pe="replicate",
+                     num_cubes=C, num_pes=P)
+    q = ctx.zeros((T_q, h_q * d_head),
+                  dtype="f16", dp=dp_qo, name=f"{panel}_q")
+    k = ctx.zeros((S_kv, h_kv * d_head),
+                  dtype="f16", dp=dp_kv, name=f"{panel}_k")
+    v = ctx.zeros((S_kv, h_kv * d_head),
+                  dtype="f16", dp=dp_kv, name=f"{panel}_v")
+    o = ctx.empty((T_q, h_q * d_head),
+                  dtype="f16", dp=dp_qo, name=f"{panel}_o")
+    ctx.launch(
+        panel, gqa_attention_prefill_long_ctx_cube_sp_pe_tp_kernel,
+        q, k, v, o,
+        T_q, S_kv, h_q, h_kv, d_head, C, P,
+        _auto_dim_remap=False,
+    )
+
+
+def _run_prefill_panel_long_ctx_cube_sp_pe_sp(
+    ctx, *, panel: str, C: int, P: int,
+    T_q: int, S_kv: int,
+    d_head: int, h_q: int, h_kv: int,
+) -> None:
+    """Case 4 runner: K, V split 64-way; Q split T_q across cubes.
+
+    DPPolicy: K, V are cube=row_wise + pe=row_wise on S_kv — each
+    rank owns S_local_pe = S_kv/(C·P). Q, O are cube=row_wise on
+    T_q (T_q_cube = T_q/C per cube), pe=replicate within cube. Ring
+    KV through the snake submesh (P parallel lanes per cube); after
+    the ring, an intra-CUBE 8-way AR combines PE lanes so cube's
+    PE 0 has the full attention for its T_q_cube rows. PE 0 of each
+    cube writes its disjoint output rows (no inter-cube reduce).
+    """
+    _ring_sfr(ctx)
+    dp_qo = DPPolicy(cube="row_wise", pe="replicate",
+                     num_cubes=C, num_pes=P)
+    dp_kv = DPPolicy(cube="row_wise", pe="row_wise",
+                     num_cubes=C, num_pes=P)
+    q = ctx.zeros((T_q, h_q * d_head),
+                  dtype="f16", dp=dp_qo, name=f"{panel}_q")
+    k = ctx.zeros((S_kv, h_kv * d_head),
+                  dtype="f16", dp=dp_kv, name=f"{panel}_k")
+    v = ctx.zeros((S_kv, h_kv * d_head),
+                  dtype="f16", dp=dp_kv, name=f"{panel}_v")
+    o = ctx.empty((T_q, h_q * d_head),
+                  dtype="f16", dp=dp_qo, name=f"{panel}_o")
+    ctx.launch(
+        panel, gqa_attention_prefill_long_ctx_cube_sp_pe_sp_kernel,
+        q, k, v, o,
+        T_q, S_kv, h_q, h_kv, d_head, C, P,
+        _auto_dim_remap=False,
+    )
+
+
+def _make_bench_fn(panel: str):
+    kind, params = _PANEL_DISPATCH[panel]
+
+    def _bench_fn(ctx):
+        if kind == "prefill_long_ctx_cube_sp_pe_sp":
+            _run_prefill_panel_long_ctx_cube_sp_pe_sp(ctx, panel=panel, **params)
+        elif kind == "prefill_long_ctx_cube_repl_pe_tp":
+            _run_prefill_panel_long_ctx_cube_repl_pe_tp(ctx, panel=panel, **params)
+        elif kind == "prefill_long_ctx_cube_repl_pe_sp":
+            _run_prefill_panel_long_ctx_cube_repl_pe_sp(ctx, panel=panel, **params)
+        elif kind == "prefill_long_ctx_cube_sp_pe_tp":
+            _run_prefill_panel_long_ctx_cube_sp_pe_tp(ctx, panel=panel, **params)
+        else:
+            raise RuntimeError(
+                f"milestone-gqa-prefill-long-ctx-4cases panel {panel!r} has "
+                f"unsupported kind={kind!r}."
+            )
+    return _bench_fn
+
+
+# ── Panel metrics (sweep.json carries these for the comparative plot) ─
+
+
+_ENGINE_SUFFIXES = (
+    "pe_gemm", "pe_math", "pe_dma", "pe_fetch_store", "pe_ipcq", "pe_cpu",
+)
+
+
+def _end_to_end_ns(op_log) -> float:
+    """End-to-end window: ``max(t_end) - min(t_start)`` over all records."""
+    if not op_log:
+        return 0.0
+    return max(r.t_end for r in op_log) - min(r.t_start for r in op_log)
+
+
+def _engine_occupancy_ns(op_log) -> dict[str, float]:
+    """Per-engine summed occupancy (component_id suffix match)."""
+    return {
+        eng: sum(
+            r.t_end - r.t_start
+            for r in op_log
+            if r.component_id.endswith("." + eng)
+        )
+        for eng in _ENGINE_SUFFIXES
+    }
+
+
+def _run_panel(panel: str, topology: str) -> dict:
+    from kernbench.runtime_api.bench_runner import run_bench
+    from kernbench.runtime_api.types import resolve_device
+    from kernbench.sim_engine.engine import GraphEngine
+    from kernbench.topology.builder import resolve_topology
+
+    topo = resolve_topology(topology)
+    result = run_bench(
+        topology=topo, bench_fn=_make_bench_fn(panel),
+        device=resolve_device(None),
+        engine_factory=lambda t, d: GraphEngine(
+            getattr(t, "topology_obj", t), enable_data=True,
+        ),
+    )
+    if not result.completion.ok:
+        raise RuntimeError(
+            f"milestone-gqa-prefill-long-ctx-4cases panel {panel!r} failed: "
+            f"{result.completion}"
+        )
+    kind, params = _PANEL_DISPATCH[panel]
+    op_log = result.engine.op_log
+    return {
+        "panel": panel,
+        "kind": kind,
+        **params,
+        "op_log_summary": _summarize_op_log(op_log),
+        "latency_ns": _end_to_end_ns(op_log),
+        "engine_occupancy_ns": _engine_occupancy_ns(op_log),
+    }
+
+
+# ── Sweep entry (called by the umbrella ``milestone_1h_gqa``) ────────
+
+
+def run_sweep(topology: str = "topology.yaml") -> int:
+    """Drive all prefill case panels; write sweep.json. Returns row count.
+
+    Per-panel try/except: even with Q-axis tiling, very large T_q or
+    unusual config combinations can hit the per-PE scratch budget. We
+    keep going so sweep.json carries whichever panels succeed; failures
+    land in a ``failures`` list for follow-up.
+    """
+    _OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
+    rows: list[dict] = []
+    failures: list[dict] = []
+    for panel in _PANELS:
+        try:
+            rows.append(_run_panel(panel, topology))
+        except Exception as e:
+            print(f"  panel {panel!r} FAILED: {e}")
+            failures.append({"panel": panel, "error": str(e)})
+    sweep = {
+        "version": 1,
+        "panels": list(_PANELS),
+        "rows": rows,
+        "failures": failures,
+    }
+    _SWEEP_JSON.write_text(json.dumps(sweep, indent=2))
+    print(f"  gqa-prefill-long-ctx-4cases: {len(rows)} rows -> {_SWEEP_JSON}")
+    return len(rows)

src/kernbench/benches/gqa_helpers/shared/__init__.py

@@ -0,0 +1,5 @@
+"""GQA helpers shared across long-ctx and short-ctx variants.
+
+Includes the panel-metrics helpers (``_gqa_panel_helpers``) and the
+context-agnostic decode-opt2 kernel.
+"""

src/kernbench/benches/gqa_helpers/shared/_gqa_panel_helpers.py

@@ -0,0 +1,40 @@
+"""Shared helpers for the GQA 4-cases milestone benches.
+
+Used by both ``milestone_gqa_decode_long_ctx_4cases`` and
+``milestone_gqa_prefill_long_ctx_4cases``. Underscore-prefixed so the
+benches package auto-discovery (which skips ``_*.py``) doesn't try to
+audit it for a ``@bench`` decorator.
+"""
+from __future__ import annotations
+
+from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
+
+
+def _ccl_cfg() -> dict:
+    """Resolve the lrab hierarchical AllReduce CCL config."""
+    return resolve_algorithm_config(
+        load_ccl_config(), name="lrab_hierarchical_allreduce",
+    )
+
+
+def _summarize_op_log(op_log) -> dict[str, int]:
+    """Per-panel op_log counts (gemm, ipcq_copy, dma_read, dma_write)."""
+    gemm_count = 0
+    ipcq_copy_count = 0
+    dma_read_count = 0
+    dma_write_count = 0
+    for r in op_log:
+        if r.op_kind == "gemm":
+            gemm_count += 1
+        elif r.op_name == "dma_read":
+            dma_read_count += 1
+        elif r.op_name == "dma_write":
+            dma_write_count += 1
+        elif r.op_name == "ipcq_copy":
+            ipcq_copy_count += 1
+    return {
+        "gemm_count": gemm_count,
+        "ipcq_copy_count": ipcq_copy_count,
+        "dma_read_count": dma_read_count,
+        "dma_write_count": dma_write_count,
+    }

src/kernbench/benches/gqa_helpers/short_ctx/__init__.py

@@ -0,0 +1 @@
+"""GQA short-context kernels."""

src/kernbench/benches/milestone_1h_gqa.py

@@ -0,0 +1,78 @@
+"""milestone-1h-gqa: umbrella GQA bench for the 1H code-sign milestone.
+
+Single ``@bench`` entry that drives all GQA panels by delegating to
+internal helpers under ``kernbench.benches.gqa_helpers``. Mirrors the
+``milestone_1h_gemm`` umbrella pattern.
+
+Currently exercises (long-context only — short-context panels are
+future work):
+  - 4-cases prefill comparative study (gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases)
+  - 4-cases decode comparative study  (gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases)
+
+Each sub-sweep writes its own ``sweep_{prefill,decode}.json`` into the
+shared output dir ``benches/1H_milestone_output/gqa/gqa_long_ctx/``.
+Selection via the env var ``GQA_1H_SWEEPS=prefill,decode`` (default
+runs both). Toggle individual sweeps with ``GQA_1H_SWEEPS=prefill``
+or ``GQA_1H_SWEEPS=decode``.
+
+Gated by ``GQA_1H_RUN=1`` to keep CI fast.
+"""
+from __future__ import annotations
+
+import os
+
+from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
+    run_sweep as _run_decode_sweep,
+)
+from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+    run_sweep as _run_prefill_sweep,
+)
+from kernbench.benches.registry import bench
+from kernbench.policy.placement.dp import DPPolicy
+
+
+@bench(
+    name="milestone-1h-gqa",
+    description=(
+        "Umbrella GQA milestone — drives long-context prefill + decode "
+        "4-cases sweeps on the LLaMA-3.1-70B single-KV-head group "
+        "(8 cubes × 8 PEs). Gated by GQA_1H_RUN=1."
+    ),
+)
+def run(torch) -> None:
+    """Drive selected GQA sub-sweeps; each writes its own sweep_*.json.
+
+    Env vars:
+      GQA_1H_RUN=1                    (required gate)
+      GQA_1H_TOPOLOGY=topology.yaml   (override topology path)
+      GQA_1H_SWEEPS=prefill,decode    (default: both; comma-separated)
+    """
+    if not os.environ.get("GQA_1H_RUN"):
+        raise RuntimeError("milestone-1h-gqa needs GQA_1H_RUN=1.")
+
+    topology = os.environ.get("GQA_1H_TOPOLOGY", "topology.yaml")
+    requested = os.environ.get("GQA_1H_SWEEPS", "prefill,decode")
+    sweeps = [s.strip() for s in requested.split(",") if s.strip()]
+
+    runners = {
+        "prefill": _run_prefill_sweep,
+        "decode": _run_decode_sweep,
+    }
+    unknown = [s for s in sweeps if s not in runners]
+    if unknown:
+        raise RuntimeError(
+            f"GQA_1H_SWEEPS contains unknown sweep(s) {unknown}; "
+            f"valid: {sorted(runners)}"
+        )
+
+    for s in sweeps:
+        runners[s](topology)
+
+    # Sentinel tensor (ADR-0045 D4 / ADR-0054 D2 carve-out) — the
+    # sub-sweeps each spin up their own GraphEngine via ``run_bench``,
+    # so this outer @bench needs to submit at least one request.
+    torch.zeros(
+        (1, 1), dtype="f16",
+        dp=DPPolicy(cube="row_wise", pe="replicate", num_cubes=1, num_pes=1),
+        name="milestone_1h_gqa_sentinel",
+    )

src/kernbench/benches/milestone_gqa_headline.py

@@ -1,231 +0,0 @@
-"""milestone-gqa-headline bench: real GQA prefill — single-KV-group target.
-
-Drives the LLaMA-3.1-70B single-KV-head-group prefill panel through
-``_gqa_attention_prefill_long`` (C=8 snake Ring KV + intra-CUBE PE-SP,
-all 64 ranks), writing ``op_log_summary`` into ``sweep.json``.
-Independent from the existing ``milestone-gqa-llama70b`` validation-scale
-bench (which stays on the legacy baseline kernels).
-
-Decode panels live in ``milestone_gqa_decode_long_ctx_4cases.py`` (the
-4-cases long-context decode comparative study).
-
-Restrictions:
-  - Single SIP (default ``topology.yaml`` 4×4 cube mesh; 4-SIP
-    headline deferred per ADR-0060 §B-item-1)
-  - T_q = S_kv = 1024 (scratch-limited; 32K headline awaits Q-axis
-    kernel tiling)
-  - No figure renderers (defer to a separate cycle)
-
-Panels:
-  single_kv_group_prefill_gqa_c8_p8: prefill C=8 snake Ring KV +
-                                     intra-CUBE PE-SP, T_q=S_kv=1K,
-                                     d_head=128 — the LLaMA-3.1-70B
-                                     single-KV-group prefill target.
-
-Gated by ``GQA_HEADLINE_RUN=1``.
-"""
-from __future__ import annotations
-
-import json
-import os
-from pathlib import Path
-
-from kernbench.benches._gqa_attention_prefill_long import gqa_attention_prefill_long_kernel
-from kernbench.benches.registry import bench
-from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
-from kernbench.ccl.sfr_config import configure_sfr_intercube_ring
-from kernbench.policy.placement.dp import DPPolicy
-
-_OUTPUT_DIR = Path(__file__).resolve().parent / "1H_milestone_output" / "gqa_headline"
-_SWEEP_JSON = _OUTPUT_DIR / "sweep.json"
-
-# ── Panel configs ────────────────────────────────────────────────────
-
-_DTYPE = "f16"
-_D_HEAD = 64
-_T_Q_PREFILL = 4
-_S_KV_PREFILL = 16
-
-_PANELS = (
-    "single_kv_group_prefill_gqa_c8_p8",
-)
-
-# Each entry: (kind, panel-specific params)
-_PANEL_DISPATCH: dict[str, tuple[str, dict]] = {
-    "single_kv_group_prefill_gqa_c8_p8": ("prefill", {
-        "C": 8, "P": 8,
-        # T_q = S_kv = 1024 (one-shot long-context prefill, scratch-limited).
-        # The bootstrap section of the prefill kernel leaves K_t/V_t/scores/
-        # exp_scores persistent (outside tl.scratch_scope), inflating the
-        # baseline. At T_q=2048 the peak hits ~1.05 MB vs 1.0 MB budget; at
-        # 1024 it fits comfortably. The true LLaMA 32K headline awaits a
-        # future increment to (a) add Q-axis tiling and (b) move the
-        # bootstrap into scratch discipline.
-        "T_q": 1_024, "S_kv": 1_024,
-        "d_head": 128,
-    }),
-}
-
-
-def _ccl_cfg():
-    return resolve_algorithm_config(
-        load_ccl_config(), name="lrab_hierarchical_allreduce",
-    )
-
-
-# ── Per-kind launch helpers ──────────────────────────────────────────
-
-
-def _run_prefill_panel(
-    ctx, *, panel: str, C: int, S_kv: int,
-    P: int = 1,
-    T_q: int = _T_Q_PREFILL,
-    d_head: int = _D_HEAD,
-) -> None:
-    if C > 1:
-        mesh_w = int(ctx.spec["sip"]["cube_mesh"]["w"])
-        if C <= mesh_w:
-            configure_sfr_intercube_ring(
-                ctx.engine, ctx.spec, _ccl_cfg(), ring_size=C,
-            )
-        else:
-            if C % mesh_w != 0:
-                raise ValueError(
-                    f"C={C} > mesh_w={mesh_w} requires C divisible by mesh_w"
-                )
-            configure_sfr_intercube_ring(
-                ctx.engine, ctx.spec, _ccl_cfg(),
-                submesh_shape=(C // mesh_w, mesh_w),
-            )
-    # Q and O switch to pe="row_wise" when intra-CUBE PE-SP is active
-    # (ADR-0060 §5.5 + §B-item-3: disjoint query-row split across PEs).
-    q_pe = "row_wise" if P > 1 else "replicate"
-    o_pe = "row_wise" if P > 1 else "replicate"
-    dp_q = DPPolicy(cube="replicate", pe=q_pe,
-                    num_cubes=C, num_pes=P)
-    dp_kv = DPPolicy(cube="row_wise" if C > 1 else "replicate",
-                     pe="replicate", num_cubes=C, num_pes=P)
-    dp_o = DPPolicy(cube="row_wise" if C > 1 else "replicate",
-                    pe=o_pe, num_cubes=C, num_pes=P)
-    q = ctx.zeros((T_q, d_head),
-                  dtype=_DTYPE, dp=dp_q, name=f"{panel}_q")
-    k = ctx.zeros((S_kv, d_head),
-                  dtype=_DTYPE, dp=dp_kv, name=f"{panel}_k")
-    v = ctx.zeros((S_kv, d_head),
-                  dtype=_DTYPE, dp=dp_kv, name=f"{panel}_v")
-    o = ctx.empty((T_q * C, d_head),
-                  dtype=_DTYPE, dp=dp_o, name=f"{panel}_o")
-    ctx.launch(
-        panel, gqa_attention_prefill_long_kernel,
-        q, k, v, o,
-        T_q, S_kv, d_head, C, P,
-        _auto_dim_remap=False,
-    )
-
-
-def _make_bench_fn(panel: str):
-    kind, params = _PANEL_DISPATCH[panel]
-    assert kind == "prefill", (
-        f"milestone-gqa-headline only registers prefill panels; got {kind!r}"
-    )
-
-    def _bench_fn(ctx):
-        _run_prefill_panel(ctx, panel=panel, **params)
-
-    return _bench_fn
-
-
-# ── Op-log summary ──────────────────────────────────────────────────
-
-
-def _summarize_op_log(op_log) -> dict[str, int]:
-    """Per-panel op_log counts (gemm, ipcq_copy, dma_read, dma_write)."""
-    gemm_count = 0
-    ipcq_copy_count = 0
-    dma_read_count = 0
-    dma_write_count = 0
-    for r in op_log:
-        if r.op_kind == "gemm":
-            gemm_count += 1
-        elif r.op_name == "dma_read":
-            dma_read_count += 1
-        elif r.op_name == "dma_write":
-            dma_write_count += 1
-        elif r.op_name == "ipcq_copy":
-            ipcq_copy_count += 1
-    return {
-        "gemm_count": gemm_count,
-        "ipcq_copy_count": ipcq_copy_count,
-        "dma_read_count": dma_read_count,
-        "dma_write_count": dma_write_count,
-    }
-
-
-def _run_panel(panel: str, topology: str) -> dict:
-    """Run one panel in a fresh engine; return its row dict."""
-    from kernbench.runtime_api.bench_runner import run_bench
-    from kernbench.runtime_api.types import resolve_device
-    from kernbench.sim_engine.engine import GraphEngine
-    from kernbench.topology.builder import resolve_topology
-
-    topo = resolve_topology(topology)
-    result = run_bench(
-        topology=topo, bench_fn=_make_bench_fn(panel),
-        device=resolve_device(None),
-        engine_factory=lambda t, d: GraphEngine(
-            getattr(t, "topology_obj", t), enable_data=True,
-        ),
-    )
-    if not result.completion.ok:
-        raise RuntimeError(
-            f"milestone-gqa-headline panel {panel!r} failed: "
-            f"{result.completion}"
-        )
-    kind, params = _PANEL_DISPATCH[panel]
-    return {
-        "panel": panel,
-        "kind": kind,
-        **params,
-        "op_log_summary": _summarize_op_log(result.engine.op_log),
-    }
-
-
-# ── Bench entry ──────────────────────────────────────────────────────
-
-
-@bench(
-    name="milestone-gqa-headline",
-    description="Headline GQA prefill milestone — LLaMA-3.1-70B single-KV-group target (C=8, P=8).",
-)
-def run(torch) -> None:
-    """Drive the headline prefill panel; write sweep.json.
-
-    Gated by GQA_HEADLINE_RUN=1.
-    """
-    _OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
-    if not os.environ.get("GQA_HEADLINE_RUN"):
-        raise RuntimeError(
-            "milestone-gqa-headline needs GQA_HEADLINE_RUN=1."
-        )
-
-    topology = os.environ.get("GQA_HEADLINE_TOPOLOGY", "topology.yaml")
-    rows = [_run_panel(panel, topology) for panel in _PANELS]
-    sweep = {
-        "version": 1,
-        "panels": list(_PANELS),
-        "config": {
-            "T_q_prefill": _T_Q_PREFILL,
-            "S_kv_prefill": _S_KV_PREFILL,
-            "d_head": _D_HEAD,
-        },
-        "rows": rows,
-    }
-    _SWEEP_JSON.write_text(json.dumps(sweep, indent=2))
-    print(f"  milestone-gqa-headline: {len(rows)} rows -> {_SWEEP_JSON}")
-
-    # Sentinel tensor (ADR-0045 D4 / ADR-0054 D2 carve-out).
-    torch.zeros(
-        (1, 1), dtype="f16",
-        dp=DPPolicy(cube="row_wise", pe="replicate", num_cubes=1, num_pes=1),
-        name="milestone_gqa_headline_sentinel",
-    )

src/kernbench/benches/registry.py

@@ -99,7 +99,11 @@ def _audit_modules(imported: list[str], registered: set[str]) -> None:
 def _eager_import_and_audit(pkg_path: list[str], pkg_name: str) -> None:
     imported: list[str] = []
     for m in iter_modules(pkg_path):
-        if m.name == "registry" or m.name.startswith("_"):
+        # Skip helper-name modules, the registry itself, and any
+        # subpackages (audit policy applies to ``.py`` bench files
+        # only — subpackages are explicitly imported by the benches
+        # that need them).
+        if m.ispkg or m.name == "registry" or m.name.startswith("_"):
             continue
         mod = import_module(f"{pkg_name}.{m.name}")
         imported.append(mod.__name__)

tests/attention/test_gqa_decode_opt2.py

@@ -151,7 +151,7 @@ def test_k_before_v_in_opt2_plan():
 def test_opt2_bench_completes_oplog_mode():
     from pathlib import Path
 
-    from kernbench.benches._gqa_attention_decode_opt2 import (  # noqa: F401
+    from kernbench.benches.gqa_helpers.shared._gqa_attention_decode_opt2 import (  # noqa: F401
         gqa_attention_decode_opt2_kernel,
     )
     from kernbench.policy.placement.dp import DPPolicy

tests/attention/test_gqa_prefill.py

@@ -1,118 +0,0 @@
-"""Phase 1 spec test for P6a GQA prefill kernel (head-parallel, C=1 baseline).
-
-P6a introduces ``_gqa_attention_prefill_long.py`` with the head-parallel structure (one
-Q head per CUBE, per-CUBE distributed output, no reduce). C=1 is the
-degenerate case — no Ring KV, no IPCQ traffic. Validates kernel
-structure and T_q > 1 attention.
-
-P6b (deferred) adds the Ring KV rotation for C > 1, which needs either
-a new SFR install function (intra_* + wrapped E/W at CUBE level) or a
-topology-specific config — separate design call.
-
-The prefill kernel differs from decode (P1a/P2a/P2b) in three ways
-(ADR-0060 §5.5 / TL;DR):
-  1. Q has T_q > 1 rows (not just decode's single timestep).
-  2. Head-parallel placement: each CUBE owns ONE Q head — no M-fold.
-  3. Each CUBE writes its own head's output — NO reduce.
-
-Phase 1 (this commit): tests only — production code lands in Phase 2.
-"""
-from __future__ import annotations
-
-from pathlib import Path
-
-from kernbench.benches._gqa_attention_prefill_long import gqa_attention_prefill_long_kernel  # noqa: F401
-from kernbench.policy.placement.dp import DPPolicy
-from kernbench.runtime_api.bench_runner import run_bench
-from kernbench.runtime_api.types import resolve_device
-from kernbench.sim_engine.engine import GraphEngine
-from kernbench.topology.builder import resolve_topology
-
-TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
-
-D_HEAD = 64
-DTYPE = "f16"
-
-
-def _engine_factory(t, d):
-    return GraphEngine(getattr(t, "topology_obj", t), enable_data=True)
-
-
-def _run_prefill(*, T_q: int, S_kv: int, C: int = 1):
-    """C=1 head-parallel prefill: single CUBE owns the one head + full KV."""
-    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
-
-    def _bench_fn(ctx):
-        dp = DPPolicy(cube="replicate", pe="replicate",
-                      num_cubes=C, num_pes=1)
-        # Q: (T_q, d_head) — one head per CUBE (head-parallel; for C=1
-        # only one head total). 2D layout matches what the kernel loads.
-        # P6b will use a 3D (h_q, T_q, d_head) Q with cube_row_wise
-        # sharding so each CUBE owns its head.
-        q = ctx.zeros((T_q, D_HEAD), dtype=DTYPE, dp=dp,
-                      name=f"q_t{T_q}_c{C}")
-        # K, V: full local for C=1 (no ring). Kernel loads K as
-        # (d_head, S_kv) via byte-conserving reshape.
-        k = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp,
-                      name=f"k_t{T_q}_c{C}")
-        v = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp,
-                      name=f"v_t{T_q}_c{C}")
-        # O: (T_q, d_head) — per-CUBE distributed output.
-        o = ctx.empty((T_q, D_HEAD), dtype=DTYPE, dp=dp,
-                      name=f"o_t{T_q}_c{C}")
-        ctx.launch(
-            f"gqa_prefill_p6a_t{T_q}_s{S_kv}_c{C}",
-            gqa_attention_prefill_long_kernel,
-            q, k, v, o,
-            T_q, S_kv, D_HEAD, C,
-            _auto_dim_remap=False,
-        )
-
-    return run_bench(
-        topology=topo, bench_fn=_bench_fn,
-        device=resolve_device(None),
-        engine_factory=_engine_factory,
-    )
-
-
-def _count(op_log, name: str) -> int:
-    return sum(1 for r in op_log if r.op_name == name)
-
-
-def test_prefill_c_one_t_q_one_completes():
-    """C=1, T_q=1: smallest workload (decode-like)."""
-    result = _run_prefill(T_q=1, S_kv=16, C=1)
-    assert result.completion.ok, (
-        f"prefill C=1 T_q=1 failed: {result.completion}"
-    )
-
-
-def test_prefill_c_one_t_q_four_completes():
-    """C=1, T_q=4: real prefill (Q has multiple rows) — distinguishes
-    prefill from decode (T_q=1)."""
-    result = _run_prefill(T_q=4, S_kv=16, C=1)
-    assert result.completion.ok, (
-        f"prefill C=1 T_q=4 failed: {result.completion}"
-    )
-
-
-def test_prefill_c_one_no_ipcq_traffic():
-    """C=1: no ring step, no IPCQ traffic."""
-    result = _run_prefill(T_q=4, S_kv=16, C=1)
-    assert result.completion.ok
-    n_copy = _count(result.engine.op_log, "ipcq_copy")
-    assert n_copy == 0, (
-        f"C=1 must have no IPCQ traffic (no ring); got {n_copy}"
-    )
-
-
-def test_prefill_c_one_one_dma_write():
-    """C=1, one head: exactly one dma_write (per-CUBE distributed output;
-    no reduce). For C > 1 in P6b this becomes dma_write_count == C."""
-    result = _run_prefill(T_q=4, S_kv=16, C=1)
-    assert result.completion.ok
-    n_writes = _count(result.engine.op_log, "dma_write")
-    assert n_writes == 1, (
-        f"C=1 prefill: expected 1 dma_write (one head per cube); "
-        f"got {n_writes}"
-    )

tests/attention/test_gqa_prefill_long_c8_snake.py

@@ -1,154 +0,0 @@
-"""Tests for prefill_long Ring KV at C=8 over a 2×4 snake sub-mesh.
-
-ADR-0060 §5.5 design target for the single-KV-group LLaMA-3.1-70B
-configuration: ``C = G = 8`` (one Q head per CUBE), Ring KV rotates
-the 8 KV slices around all 8 CUBEs.
-
-Increment 1 wired ``configure_sfr_intercube_ring(submesh_shape=(2, 4))``
-to map the kernel's logical E/W to a snake/serpentine 1-hop path
-through the top 2×4 sub-mesh of the 4×4 SIP CUBE mesh. The kernel
-itself uses only logical ``dir="W"`` / ``dir="E"`` — the snake is
-transparent at kernel level.
-
-This file verifies the assembly works end-to-end at C=8:
-  T1  kernel completes (no scratch overflow, no deadlock)
-  T2  per-CUBE head-parallel output (8 dma_writes, one per CUBE)
-  T3  tile-granular ring traffic at C=8 follows the same
-      ``(C-1)·n_tiles·2·C`` formula as the existing tile-ring tests
-"""
-from __future__ import annotations
-
-from pathlib import Path
-
-from kernbench.benches._gqa_attention_prefill_long import gqa_attention_prefill_long_kernel  # noqa: F401
-from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
-from kernbench.ccl.sfr_config import configure_sfr_intercube_ring
-from kernbench.policy.placement.dp import DPPolicy
-from kernbench.runtime_api.bench_runner import run_bench
-from kernbench.runtime_api.types import resolve_device
-from kernbench.sim_engine.engine import GraphEngine
-from kernbench.topology.builder import resolve_topology
-
-TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
-
-D_HEAD = 64
-DTYPE = "f16"
-
-
-def _ccl_cfg():
-    return resolve_algorithm_config(
-        load_ccl_config(), name="lrab_hierarchical_allreduce",
-    )
-
-
-def _engine_factory(t, d):
-    return GraphEngine(getattr(t, "topology_obj", t), enable_data=True)
-
-
-def _count(op_log, name: str) -> int:
-    return sum(1 for r in op_log if r.op_name == name)
-
-
-def _run_prefill_c8_snake(*, T_q: int, S_kv: int):
-    """Head-parallel prefill at C=8 with snake-mapped Ring KV over the
-    top 2×4 sub-mesh of the 4×4 SIP CUBE mesh."""
-    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
-    C = 8
-
-    def _bench_fn(ctx):
-        configure_sfr_intercube_ring(
-            ctx.engine, ctx.spec, _ccl_cfg(),
-            submesh_shape=(2, 4),
-        )
-        dp_q = DPPolicy(cube="replicate", pe="replicate",
-                        num_cubes=C, num_pes=1)
-        dp_kv = DPPolicy(cube="row_wise", pe="replicate",
-                         num_cubes=C, num_pes=1)
-        dp_o = DPPolicy(cube="row_wise", pe="replicate",
-                        num_cubes=C, num_pes=1)
-        q = ctx.zeros((T_q, D_HEAD), dtype=DTYPE, dp=dp_q,
-                      name=f"q_c8_snake_t{T_q}_s{S_kv}")
-        k = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp_kv,
-                      name=f"k_c8_snake_t{T_q}_s{S_kv}")
-        v = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp_kv,
-                      name=f"v_c8_snake_t{T_q}_s{S_kv}")
-        o = ctx.empty((T_q * C, D_HEAD), dtype=DTYPE, dp=dp_o,
-                      name=f"o_c8_snake_t{T_q}_s{S_kv}")
-        ctx.launch(
-            f"gqa_prefill_long_c8_snake_t{T_q}_s{S_kv}",
-            gqa_attention_prefill_long_kernel,
-            q, k, v, o,
-            T_q, S_kv, D_HEAD, C,
-            _auto_dim_remap=False,
-        )
-
-    return run_bench(
-        topology=topo, bench_fn=_bench_fn,
-        device=resolve_device(None),
-        engine_factory=_engine_factory,
-    )
-
-
-# ── T1: kernel completes end-to-end at C=8 ───────────────────────────
-
-
-def test_prefill_long_c8_snake_completes():
-    """ADR-0060 §5.5 design target: prefill Ring KV at C=G=8 over the
-    2×4 snake sub-mesh. With zero inputs, output is trivially zero;
-    this test guards that the kernel reaches completion without
-    deadlock, scratch overflow, or routing failure.
-
-    Configuration: T_q=4, S_kv=8192 → S_local=1024, n_tiles=1
-    (TILE_S_KV=1024). Bounded scratch.
-    """
-    result = _run_prefill_c8_snake(T_q=4, S_kv=8192)
-    assert result.completion.ok, (
-        f"prefill at C=8 with snake ring must complete; "
-        f"got {result.completion}"
-    )
-
-
-# ── T2: 8 dma_writes (head-parallel, one head per CUBE) ──────────────
-
-
-def test_prefill_long_c8_snake_distributed_output_count():
-    """Head-parallel: each CUBE owns one Q head and writes its own
-    head's output (T_q, d_head) rows. No inter-CUBE reduce on the
-    output side — so ``dma_write_count == C == 8``.
-
-    This is the C=8 analogue of
-    ``test_prefill_long_tile_ring_dma_write_count`` at C=4.
-    """
-    result = _run_prefill_c8_snake(T_q=4, S_kv=8192)
-    assert result.completion.ok
-    n_writes = _count(result.engine.op_log, "dma_write")
-    assert n_writes == 8, (
-        f"head-parallel C=8: expected 8 dma_writes (one per CUBE); "
-        f"got {n_writes}"
-    )
-
-
-# ── T3: tile-granular ring ipcq_copy count scales as (C-1)·n_tiles·2·C
-
-
-def test_prefill_long_c8_snake_tile_ipcq_count():
-    """ADR-0060 §5.5.1 tile-granular ring: per-CUBE send count is
-    ``2·n_tiles·(C-1)`` (K + V per tile per ring step). Aggregated
-    across C CUBEs, total ipcq_copy = ``(C-1)·n_tiles·2·C``.
-
-    Configuration: T_q=4, S_kv=8192, C=8 → S_local=1024, n_tiles=1
-    (TILE_S_KV=1024). Expected total = ``7·1·2·8 = 112``.
-
-    Verifies that the snake-mapped 1-hop physical links carry the
-    same logical ring traffic as the existing C=4 1D-row ring tests.
-    """
-    C = 8
-    n_tiles = 1  # S_local=1024 / TILE_S_KV=1024
-    result = _run_prefill_c8_snake(T_q=4, S_kv=8192)
-    assert result.completion.ok
-    n_copy = _count(result.engine.op_log, "ipcq_copy")
-    expected = (C - 1) * n_tiles * 2 * C
-    assert n_copy == expected, (
-        f"tile-granular ring at C=8: expected {expected} ipcq_copy "
-        f"((C-1)·n_tiles·2·C = {C - 1}·{n_tiles}·2·{C}); got {n_copy}"
-    )

tests/attention/test_gqa_prefill_long_pe_sp.py

@@ -1,233 +0,0 @@
-"""Tests for intra-CUBE PE-SP in prefill_long (query-axis split).
-
-ADR-0060 §5.5 last bullet + §B-item-3: split the head's query rows
-``[T_q, d_head]`` across the ``P`` PEs of each CUBE so all P PEs work
-in parallel. Output rows are disjoint across PEs ⇒ no intra-CUBE
-reduce needed.
-
-Same-lane SFR wiring: PE ``i`` in CUBE A has its own E/W ring link to
-PE ``i`` in CUBE B (the snake's prev/next). All P PEs of a CUBE see
-the same K, V (HBM-resident, ``pe="replicate"``) ⇒ P parallel rings
-run in lockstep, each PE rotating its own K/V copies via its own IPCQ
-channels.
-
-Activation contract:
-  - ``P == 1`` (default; omitted from launch args) → existing
-    PE-0-only behavior. Byte-for-byte unchanged.
-  - ``P > 1`` → all P PEs active; each handles ``T_q // P`` query
-    rows. Requires ``T_q % P == 0`` (degenerate T_q < P is rejected
-    — caller must use ``P=1`` for that workload, ADR-0060 §B-item-3).
-
-Phase 1: tests only — production code lands in Phase 2.
-T1, T2, T3, T5 fail today (TypeError: kernel signature has no P).
-T4 passes today as the backward-compat anchor.
-"""
-from __future__ import annotations
-
-from pathlib import Path
-
-import pytest
-
-from kernbench.benches._gqa_attention_prefill_long import gqa_attention_prefill_long_kernel  # noqa: F401
-from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
-from kernbench.ccl.sfr_config import configure_sfr_intercube_ring
-from kernbench.policy.placement.dp import DPPolicy
-from kernbench.runtime_api.bench_runner import run_bench
-from kernbench.runtime_api.types import resolve_device
-from kernbench.sim_engine.engine import GraphEngine
-from kernbench.topology.builder import resolve_topology
-
-TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
-
-D_HEAD = 64
-DTYPE = "f16"
-
-
-def _ccl_cfg():
-    return resolve_algorithm_config(
-        load_ccl_config(), name="lrab_hierarchical_allreduce",
-    )
-
-
-def _engine_factory(t, d):
-    return GraphEngine(getattr(t, "topology_obj", t), enable_data=True)
-
-
-def _count(op_log, name: str) -> int:
-    return sum(1 for r in op_log if r.op_name == name)
-
-
-def _run_prefill_pe_sp(
-    *, T_q: int, S_kv: int, C: int, P: int | None,
-    snake: bool,
-):
-    """Drive a prefill_long launch with optional intra-CUBE PE-SP.
-
-    ``P=None`` → omit P from the launch args (exercises the kernel's
-    default behaviour).
-    ``snake=True`` → install the 2×4 snake ring SFR (Increment 1);
-    else install the 1D-row ring at ``ring_size=C``.
-    """
-    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
-
-    def _bench_fn(ctx):
-        if snake:
-            configure_sfr_intercube_ring(
-                ctx.engine, ctx.spec, _ccl_cfg(),
-                submesh_shape=(2, 4),
-            )
-        else:
-            configure_sfr_intercube_ring(
-                ctx.engine, ctx.spec, _ccl_cfg(),
-                ring_size=C,
-            )
-        num_pes = P if (P is not None and P > 1) else 1
-        # When PE-SP is active, Q and O are split row-wise across PEs;
-        # K, V remain replicated within a CUBE.
-        q_pe = "row_wise" if num_pes > 1 else "replicate"
-        o_pe = "row_wise" if num_pes > 1 else "replicate"
-        dp_q = DPPolicy(cube="replicate", pe=q_pe,
-                        num_cubes=C, num_pes=num_pes)
-        dp_kv = DPPolicy(cube="row_wise" if C > 1 else "replicate",
-                         pe="replicate", num_cubes=C, num_pes=num_pes)
-        dp_o = DPPolicy(cube="row_wise" if C > 1 else "replicate",
-                        pe=o_pe, num_cubes=C, num_pes=num_pes)
-        suffix = f"c{C}_p{P}_t{T_q}_s{S_kv}"
-        q = ctx.zeros((T_q, D_HEAD), dtype=DTYPE, dp=dp_q,
-                      name=f"q_pe_sp_{suffix}")
-        k = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp_kv,
-                      name=f"k_pe_sp_{suffix}")
-        v = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp_kv,
-                      name=f"v_pe_sp_{suffix}")
-        o = ctx.empty((T_q * C, D_HEAD), dtype=DTYPE, dp=dp_o,
-                      name=f"o_pe_sp_{suffix}")
-        launch_args = [q, k, v, o, T_q, S_kv, D_HEAD, C]
-        if P is not None:
-            launch_args.append(P)
-        ctx.launch(
-            f"gqa_prefill_long_pe_sp_{suffix}",
-            gqa_attention_prefill_long_kernel,
-            *launch_args,
-            _auto_dim_remap=False,
-        )
-
-    return run_bench(
-        topology=topo, bench_fn=_bench_fn,
-        device=resolve_device(None),
-        engine_factory=_engine_factory,
-    )
-
-
-# ── T1: C=8 P=8 completes end-to-end ─────────────────────────────────
-
-
-def test_prefill_c8_p8_pe_sp_completes():
-    """ADR-0060 §5.5 + §B-item-3 design target: 64 ranks active
-    (8 CUBEs × 8 PEs), query-axis split across PEs, snake-mapped
-    Ring KV. Guards no scratch overflow, no deadlock across the
-    P parallel same-lane rings.
-    """
-    result = _run_prefill_pe_sp(
-        T_q=8, S_kv=8192, C=8, P=8, snake=True,
-    )
-    assert result.completion.ok, (
-        f"prefill at C=8, P=8 (PE-SP) must complete; "
-        f"got {result.completion}"
-    )
-
-
-# ── T2: 64 dma_writes (one per PE, disjoint query rows) ──────────────
-
-
-def test_prefill_c8_p8_pe_sp_64_dma_writes():
-    """With query-axis split, each PE owns ``T_q/P = 1`` query row
-    and stores its own ``(1, d_head)`` slice. Across all 64 ranks
-    (8 CUBEs × 8 PEs), ``dma_write_count == 64``.
-
-    This is the structural signal that PE-SP is actually wired:
-    PE-0-only would give 8 dma_writes (one per CUBE).
-    """
-    result = _run_prefill_pe_sp(
-        T_q=8, S_kv=8192, C=8, P=8, snake=True,
-    )
-    assert result.completion.ok
-    n_writes = _count(result.engine.op_log, "dma_write")
-    assert n_writes == 64, (
-        f"PE-SP at C=8, P=8: expected 64 dma_writes "
-        f"(8 cubes × 8 PEs, each writes its T_q/P=1 query row); "
-        f"got {n_writes}"
-    )
-
-
-# ── T3: P parallel rings — ipcq_copy scales by P ─────────────────────
-
-
-def test_prefill_c8_p8_pe_sp_ring_ipcq_count():
-    """Per-PE same-lane rings: each PE runs its own ring traffic via
-    its own IPCQ channels. Total inter-CUBE ipcq_copy:
-    ``(C-1) · n_tiles · 2 · C · P`` (= existing C=8 formula × P).
-
-    Configuration: T_q=8, S_kv=8192, C=8 → S_local=1024, n_tiles=1
-    (TILE_S_KV=1024). Expected = ``7·1·2·8·8`` = **896**.
-    """
-    C = 8
-    P = 8
-    n_tiles = 1
-    result = _run_prefill_pe_sp(
-        T_q=8, S_kv=8192, C=C, P=P, snake=True,
-    )
-    assert result.completion.ok
-    n_copy = _count(result.engine.op_log, "ipcq_copy")
-    expected = (C - 1) * n_tiles * 2 * C * P
-    assert n_copy == expected, (
-        f"PE-SP parallel rings at C=8, P=8: expected {expected} "
-        f"ipcq_copy ((C-1)·n_tiles·2·C·P = "
-        f"{C - 1}·{n_tiles}·2·{C}·{P}); got {n_copy}"
-    )
-
-
-# ── T4: backward-compat — P omitted → existing PE-0-only behaviour ───
-
-
-def test_prefill_p_default_1_backward_compat():
-    """When ``P`` is omitted from the launch args, the kernel must
-    behave identically to today: only PE 0 of each CUBE participates;
-    one head per CUBE; one dma_write per CUBE.
-
-    At C=4 this gives 4 dma_writes (matches the existing
-    ``test_prefill_long_tile_ring_dma_write_count``). Phase 2 must
-    not regress this path.
-
-    Passes today AND after Phase 2.
-    """
-    C = 4
-    result = _run_prefill_pe_sp(
-        T_q=4, S_kv=8192, C=C, P=None, snake=False,
-    )
-    assert result.completion.ok, (
-        f"prefill at C=4 with default P (PE-0-only) must complete; "
-        f"got {result.completion}"
-    )
-    n_writes = _count(result.engine.op_log, "dma_write")
-    assert n_writes == C, (
-        f"default P=1 (PE-0-only): expected {C} dma_writes "
-        f"(one per CUBE); got {n_writes}"
-    )
-
-
-# ── T5: validation — T_q must be divisible by P when P > 1 ───────────
-
-
-def test_prefill_pe_sp_rejects_non_divisible_t_q():
-    """ADR-0060 §B-item-3 fallback (KV-block split + intra-CUBE
-    reduce for T_q < P) is deferred. Callers must request ``P=1`` for
-    workloads where T_q < P or T_q % P != 0.
-
-    C=4, P=8, T_q=4 violates ``T_q % P == 0`` (and also T_q < P).
-    The kernel must raise ValueError with a clear error message
-    rather than silently producing wrong results.
-    """
-    with pytest.raises((ValueError, AssertionError), match=r"T_q"):
-        _run_prefill_pe_sp(
-            T_q=4, S_kv=8192, C=4, P=8, snake=False,
-        )

tests/attention/test_gqa_prefill_long_tile_ring.py

@@ -1,162 +0,0 @@
-"""Phase 1 spec test for P3c: tile-granular Ring KV in prefill_long
-(ADR-0060 §5.5 amendment + ADR-0063 §A.2).
-
-Today's prefill_long ring sends and receives full ``(d_head, S_local)``
-KV slices per step. Step 0's local-attention intermediates also live
-outside any ``scratch_scope`` (they're loaded as full-slice ``Kc``,
-``Vc`` and feed ``scores``, ``exp_scores`` as persistent allocations).
-At larger ``S_local`` both the step-0 leak and the ring step's
-in-scope intermediates grow linearly with ``S_local``, and at
-``S_local = 32K`` (S_kv=128K, C=4, T_q=4) the peak overflows the
-1 MiB pool.
-
-P3c converts the ring to **tile-granular**: a nested loop
-``for k in range(C): for t in range(n_tiles): ...`` where each
-iteration sends/recvs one ``(d_head, TILE_S_KV)`` tile (and its V
-counterpart). The persistent state shrinks to ``(m, ℓ, O)`` only
-(~1 KB); per-tile in-scope scratch is bounded by
-``TILE_S_KV`` regardless of ``S_local``. Ceiling lifted.
-
-Trade-off: the per-CUBE send count grows from ``2·(C-1)`` to
-``2·n_tiles·(C-1)``. At ``n_tiles=1`` (small ``S_local``) the count is
-unchanged, so the existing ``test_prefill_ring_c_*`` tests at
-``S_kv ∈ {16, 32}`` still pass.
-
-Phase 1 (this commit): tests only — production code lands in Phase 2.
-T1 fails today with a ``TLContext scratch overflow``; T2 fails today
-with the slice-granular ipcq_copy count; T5 passes today and is a
-regression guard for the per-CUBE output write count.
-"""
-from __future__ import annotations
-
-from pathlib import Path
-
-from kernbench.benches._gqa_attention_prefill_long import gqa_attention_prefill_long_kernel  # noqa: F401
-from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
-from kernbench.ccl.sfr_config import configure_sfr_intercube_ring
-from kernbench.policy.placement.dp import DPPolicy
-from kernbench.runtime_api.bench_runner import run_bench
-from kernbench.runtime_api.types import resolve_device
-from kernbench.sim_engine.engine import GraphEngine
-from kernbench.topology.builder import resolve_topology
-
-TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
-
-D_HEAD = 64
-DTYPE = "f16"
-
-
-def _ccl_cfg():
-    return resolve_algorithm_config(
-        load_ccl_config(), name="lrab_hierarchical_allreduce",
-    )
-
-
-def _engine_factory(t, d):
-    return GraphEngine(getattr(t, "topology_obj", t), enable_data=True)
-
-
-def _count(op_log, name: str) -> int:
-    return sum(1 for r in op_log if r.op_name == name)
-
-
-def _run_prefill_ring(*, T_q: int, S_kv: int, C: int):
-    """Head-parallel prefill with Ring KV across C CUBEs."""
-    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
-
-    def _bench_fn(ctx):
-        configure_sfr_intercube_ring(
-            ctx.engine, ctx.spec, _ccl_cfg(), ring_size=C,
-        )
-        dp_q = DPPolicy(cube="replicate", pe="replicate",
-                        num_cubes=C, num_pes=1)
-        dp_kv = DPPolicy(cube="row_wise" if C > 1 else "replicate",
-                         pe="replicate", num_cubes=C, num_pes=1)
-        dp_o = DPPolicy(cube="row_wise" if C > 1 else "replicate",
-                        pe="replicate", num_cubes=C, num_pes=1)
-        q = ctx.zeros((T_q, D_HEAD), dtype=DTYPE, dp=dp_q,
-                      name=f"q_tlr_t{T_q}_c{C}_s{S_kv}")
-        k = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp_kv,
-                      name=f"k_tlr_t{T_q}_c{C}_s{S_kv}")
-        v = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp_kv,
-                      name=f"v_tlr_t{T_q}_c{C}_s{S_kv}")
-        o = ctx.empty((T_q * C, D_HEAD), dtype=DTYPE, dp=dp_o,
-                      name=f"o_tlr_t{T_q}_c{C}_s{S_kv}")
-        ctx.launch(
-            f"gqa_prefill_long_tile_ring_t{T_q}_c{C}_s{S_kv}",
-            gqa_attention_prefill_long_kernel,
-            q, k, v, o,
-            T_q, S_kv, D_HEAD, C,
-            _auto_dim_remap=False,
-        )
-
-    return run_bench(
-        topology=topo, bench_fn=_bench_fn,
-        device=resolve_device(None),
-        engine_factory=_engine_factory,
-    )
-
-
-# ── T1: 128K ceiling lift ────────────────────────────────────────────
-
-
-def test_prefill_long_context_128k_completes():
-    """ADR-0063 §A.2 headline ceiling lift. At S_kv=128K with C=4,
-    S_local=32K. Today: step-0 score-stack (~768 KB persistent) + ring
-    scope (~768 KB) → 1.5 MB peak → TLContext scratch overflow. After
-    P3c the persistent state is just ``(m, ℓ, O)`` (≈ 1 KB) and the
-    per-tile in-scope scratch is bounded by TILE_S_KV.
-    """
-    result = _run_prefill_ring(T_q=4, S_kv=131_072, C=4)
-    assert result.completion.ok, (
-        f"prefill_long at S_kv=128K must complete after tile-granular "
-        f"ring lands; got {result.completion}"
-    )
-
-
-# ── T2: tile-granular ipcq_copy count ────────────────────────────────
-
-
-def test_prefill_long_tile_granular_ipcq_count():
-    """ADR-0060 §5.5 amendment: with tile-granular sends, the per-CUBE
-    send count grows from ``2·(C-1)`` to ``2·n_tiles·(C-1)``. Aggregated
-    across all C CUBEs the total ipcq_copy becomes
-    ``2·n_tiles·(C-1)·C``.
-
-    Config: T_q=4, S_kv=4096, C=2 → S_local=2048, n_tiles=2 (with
-    TILE_S_KV=1024). Today: slice-granular total =
-    ``(C-1)·2·C = 4``. After P3c: tile-granular total =
-    ``(C-1)·n_tiles·2·C = 8``.
-    """
-    C = 2
-    n_tiles = 2  # S_local=2048 / TILE_S_KV=1024
-    result = _run_prefill_ring(T_q=4, S_kv=4096, C=C)
-    assert result.completion.ok, (
-        f"prefill_long multi-tile ring must complete; got {result.completion}"
-    )
-    n_copy = _count(result.engine.op_log, "ipcq_copy")
-    expected = (C - 1) * n_tiles * 2 * C
-    assert n_copy == expected, (
-        f"tile-granular ring: expected {expected} ipcq_copy "
-        f"((C-1)·n_tiles·2·C = {C - 1}·{n_tiles}·2·{C}); got {n_copy}"
-    )
-
-
-# ── T5: per-CUBE distributed output unchanged (regression guard) ─────
-
-
-def test_prefill_long_tile_ring_dma_write_count():
-    """ADR-0060 §5.5: per-CUBE distributed output must hold under the
-    tile-granular ring rewrite. Each CUBE still writes its own head's
-    output (no inter-CUBE reduce); dma_write_count == C.
-
-    Today passes; must continue to pass after P3c.
-    """
-    C = 4
-    result = _run_prefill_ring(T_q=4, S_kv=4096, C=C)
-    assert result.completion.ok
-    n_writes = _count(result.engine.op_log, "dma_write")
-    assert n_writes == C, (
-        f"per-CUBE distributed output: expected {C} dma_writes (one per "
-        f"CUBE); got {n_writes}"
-    )

tests/attention/test_gqa_prefill_ring.py

@@ -1,161 +0,0 @@
-"""Phase 1 spec test for P6b GQA prefill Ring KV (head-parallel, C>1).
-
-P6b adds the Ring KV rotation (ADR-0060 §5.5) to the prefill kernel.
-Each CUBE owns one Q head + its KV slice; over C ring steps the KV
-blocks rotate around the C CUBEs so every CUBE sees every block. The
-running ``(m, ℓ, O)`` is folded inside the loop. No reduce — each CUBE
-writes its own head's output.
-
-Requires a new SFR install ``configure_sfr_intercube_ring(ring_size=C)``
-that wires:
-  - ``intra_*``  : 2×4 PE grid within a CUBE (same as multisip)
-  - ``E/W``      : 1D ring of CUBEs 0..ring_size-1 WITH WRAP
-                    (symmetric to ``configure_sfr_intracube_pe_ring``,
-                     applied at CUBE level)
-  - ``global_*`` : SIP-level (same as multisip)
-
-The kernel ring body:
-  for step in range(1, C):
-      tl.send(dir="W", src=Kc)
-      tl.send(dir="W", src=Vc)
-      Kc = tl.recv(dir="E", ...)
-      Vc = tl.recv(dir="E", ...)
-      ... local partial + online-softmax merge into (m, ℓ, O) ...
-
-Per CUBE per step: 2 sends (K, V) → 2 ``ipcq_copy``. Across all CUBEs:
-``(C-1) * 2 * C`` total ipcq_copy.
-
-Restriction in P6b first cut: ``C ∈ {1, 2, 4}`` (single row of the 4×4
-cube mesh). C=8 ring would span rows — follow-on.
-
-Phase 1 (this commit): tests only — production code lands in Phase 2.
-"""
-from __future__ import annotations
-
-from pathlib import Path
-
-from kernbench.benches._gqa_attention_prefill_long import gqa_attention_prefill_long_kernel  # noqa: F401
-from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
-from kernbench.ccl.sfr_config import configure_sfr_intercube_ring  # noqa: F401 (Phase 2)
-from kernbench.policy.placement.dp import DPPolicy
-from kernbench.runtime_api.bench_runner import run_bench
-from kernbench.runtime_api.types import resolve_device
-from kernbench.sim_engine.engine import GraphEngine
-from kernbench.topology.builder import resolve_topology
-
-TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
-
-D_HEAD = 64
-DTYPE = "f16"
-
-
-def _ccl_cfg():
-    return resolve_algorithm_config(
-        load_ccl_config(), name="lrab_hierarchical_allreduce",
-    )
-
-
-def _engine_factory(t, d):
-    return GraphEngine(getattr(t, "topology_obj", t), enable_data=True)
-
-
-def _run_prefill_ring(*, T_q: int, S_kv: int, C: int):
-    """Head-parallel prefill with Ring KV across C CUBEs."""
-    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
-
-    def _bench_fn(ctx):
-        # P6b: new SFR install with cube-level ring wrap.
-        configure_sfr_intercube_ring(
-            ctx.engine, ctx.spec, _ccl_cfg(), ring_size=C,
-        )
-        # Q replicated on every CUBE (zeros for testing; per-CUBE head
-        # indexing is implicit). KV sequence-sharded by CUBE. O
-        # distributed — each CUBE writes its slice of (T_q*C, d_head).
-        dp_q = DPPolicy(cube="replicate", pe="replicate",
-                        num_cubes=C, num_pes=1)
-        dp_kv = DPPolicy(cube="row_wise", pe="replicate",
-                         num_cubes=C, num_pes=1)
-        dp_o = DPPolicy(cube="row_wise", pe="replicate",
-                        num_cubes=C, num_pes=1)
-        q = ctx.zeros((T_q, D_HEAD), dtype=DTYPE, dp=dp_q,
-                      name=f"q_t{T_q}_c{C}_ring")
-        k = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp_kv,
-                      name=f"k_t{T_q}_c{C}_ring")
-        v = ctx.zeros((S_kv, D_HEAD), dtype=DTYPE, dp=dp_kv,
-                      name=f"v_t{T_q}_c{C}_ring")
-        # O: (T_q * C, d_head), each CUBE writes (T_q, d_head) at its
-        # slice. dma_write_count = C (per-CUBE distributed output).
-        o = ctx.empty((T_q * C, D_HEAD), dtype=DTYPE, dp=dp_o,
-                      name=f"o_t{T_q}_c{C}_ring")
-        ctx.launch(
-            f"gqa_prefill_ring_t{T_q}_s{S_kv}_c{C}",
-            gqa_attention_prefill_long_kernel,
-            q, k, v, o,
-            T_q, S_kv, D_HEAD, C,
-            _auto_dim_remap=False,
-        )
-
-    return run_bench(
-        topology=topo, bench_fn=_bench_fn,
-        device=resolve_device(None),
-        engine_factory=_engine_factory,
-    )
-
-
-def _count(op_log, name: str) -> int:
-    return sum(1 for r in op_log if r.op_name == name)
-
-
-# ── C=2 Ring KV ────────────────────────────────────────────────────────
-
-
-def test_prefill_ring_c_two_completes():
-    """C=2: 1 ring step rotates KV between the 2 CUBEs."""
-    result = _run_prefill_ring(T_q=4, S_kv=16, C=2)
-    assert result.completion.ok, (
-        f"prefill ring C=2 failed: {result.completion}"
-    )
-
-
-def test_prefill_ring_c_two_distributed_output():
-    """C=2: per-CUBE distributed output, no reduce → dma_write_count == 2."""
-    result = _run_prefill_ring(T_q=4, S_kv=16, C=2)
-    assert result.completion.ok
-    n_writes = _count(result.engine.op_log, "dma_write")
-    assert n_writes == 2, (
-        f"C=2 prefill: expected 2 dma_write (one per CUBE); got {n_writes}"
-    )
-
-
-def test_prefill_ring_c_two_ipcq_count():
-    """C=2: 1 ring step × 2 handles (K, V) × 2 CUBEs = 4 ipcq_copy."""
-    result = _run_prefill_ring(T_q=4, S_kv=16, C=2)
-    assert result.completion.ok
-    n_copy = _count(result.engine.op_log, "ipcq_copy")
-    expected = (2 - 1) * 2 * 2
-    assert n_copy == expected, (
-        f"C=2 ring: expected {expected} ipcq_copy "
-        f"((C-1)·2·C = 1·2·2); got {n_copy}"
-    )
-
-
-# ── C=4 Ring KV (combined assertions) ─────────────────────────────────
-
-
-def test_prefill_ring_c_four_combined():
-    """C=4: 3 ring steps rotate KV around 4 CUBEs.
-    Expected: completes; 4 dma_writes; (4-1)·2·4 = 24 ipcq_copy."""
-    result = _run_prefill_ring(T_q=4, S_kv=32, C=4)
-    assert result.completion.ok, (
-        f"prefill ring C=4 failed: {result.completion}"
-    )
-    n_writes = _count(result.engine.op_log, "dma_write")
-    assert n_writes == 4, (
-        f"C=4 prefill: expected 4 dma_write (one per CUBE); got {n_writes}"
-    )
-    n_copy = _count(result.engine.op_log, "ipcq_copy")
-    expected = (4 - 1) * 2 * 4
-    assert n_copy == expected, (
-        f"C=4 ring: expected {expected} ipcq_copy "
-        f"((C-1)·2·C = 3·2·4); got {n_copy}"
-    )

tests/attention/test_gqa_short_context.py

@@ -76,7 +76,7 @@ def _run_decode_short(*, h_q: int, h_kv: int, kv_per_cube: int,
       O: replicated; each group root writes the full byte-conserving
           (h_q·T_q, D_HEAD) result.
     """
-    from kernbench.benches._gqa_attention_decode_short import gqa_attention_decode_short_kernel  # Phase 2
+    from kernbench.benches.gqa_helpers.short_ctx._gqa_attention_decode_short import gqa_attention_decode_short_kernel  # Phase 2
 
     topo = resolve_topology(str(TOPOLOGY_DEFAULT))
 
@@ -196,7 +196,7 @@ def _run_prefill_short(*, h_kv: int, kv_per_cube: int,
     Layout: same head-stacked K/V scheme as decode short, with Q/O
     replicated and byte-conserving reshape inside the kernel.
     """
-    from kernbench.benches._gqa_attention_prefill_short import gqa_attention_prefill_short_kernel  # Phase 2
+    from kernbench.benches.gqa_helpers.short_ctx._gqa_attention_prefill_short import gqa_attention_prefill_short_kernel  # Phase 2
 
     topo = resolve_topology(str(TOPOLOGY_DEFAULT))
 

tests/attention/test_milestone_gqa_decode_long_ctx_4cases.py

@@ -68,7 +68,7 @@ def _run_case4_smoke(*, S_kv: int):
     ``S_kv=128K`` runs come from ``kernbench run --bench
     milestone-gqa-decode-long-ctx-4cases``, not pytest.
     """
-    from kernbench.benches.milestone_gqa_decode_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _run_decode_panel_long_ctx_cube_sp_pe_sp,
     )
     topo = resolve_topology(str(TOPOLOGY_DEFAULT))
@@ -99,13 +99,13 @@ def test_case4_panel_registered():
       C = 8                 (head-parallel CUBE Group)
       P = 8                 (intra-CUBE PE-SP)
       T_q = 1               (decode: one new token per pass)
-      S_kv = 131_072        (LLaMA long-context decode target)
+      S_kv = 8_192        (LLaMA long-context decode target)
       d_head = 128, h_q = 8, h_kv = 1
 
     The lrab sub_w=4 / sub_h=2 geometry is baked into the Case 4
     wrapper kernel; it is not a panel parameter.
     """
-    from kernbench.benches.milestone_gqa_decode_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _PANEL_DISPATCH,
         _PANELS,
     )
@@ -120,7 +120,7 @@ def test_case4_panel_registered():
     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("S_kv") == 8_192
     assert params.get("d_head") == 128
     assert params.get("h_q") == 8
     assert params.get("h_kv") == 1
@@ -197,7 +197,7 @@ def _run_case2_smoke(*, S_kv: int):
     slide-11 memory waste); for B=1 only one rank does the work; no
     inter-rank comm.
     """
-    from kernbench.benches.milestone_gqa_decode_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _run_decode_panel_long_ctx_cube_repl_pe_tp,
     )
     topo = resolve_topology(str(TOPOLOGY_DEFAULT))
@@ -229,7 +229,7 @@ def test_case2_panel_registered():
     For B=1 only one rank works (PEs 1-7 idle — slide-11 calls
     out this PE-TP waste).
     """
-    from kernbench.benches.milestone_gqa_decode_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _PANEL_DISPATCH,
         _PANELS,
     )
@@ -242,7 +242,7 @@ def test_case2_panel_registered():
     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("S_kv") == 8_192
     assert params.get("d_head") == 128
     assert params.get("h_q") == 8
     assert params.get("h_kv") == 1
@@ -303,7 +303,7 @@ def _run_case3_smoke(*, S_kv: int):
     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_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _run_decode_panel_long_ctx_cube_repl_pe_sp,
     )
     topo = resolve_topology(str(TOPOLOGY_DEFAULT))
@@ -333,7 +333,7 @@ def test_case3_panel_registered():
     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_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _PANEL_DISPATCH,
         _PANELS,
     )
@@ -346,7 +346,7 @@ def test_case3_panel_registered():
     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("S_kv") == 8_192
     assert params.get("d_head") == 128
     assert params.get("h_q") == 8
     assert params.get("h_kv") == 1
@@ -423,7 +423,7 @@ def _run_case1_smoke(*, S_kv: int):
     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_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _run_decode_panel_long_ctx_cube_sp_pe_tp,
     )
     topo = resolve_topology(str(TOPOLOGY_DEFAULT))
@@ -454,7 +454,7 @@ def test_case1_panel_registered():
     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_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _PANEL_DISPATCH,
         _PANELS,
     )
@@ -467,7 +467,7 @@ def test_case1_panel_registered():
     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("S_kv") == 8_192
     assert params.get("d_head") == 128
     assert params.get("h_q") == 8
     assert params.get("h_kv") == 1
@@ -548,7 +548,7 @@ def test_panel_metrics_helpers_present_and_correct():
     Helpers exercised via the Case 2 smoke runner's op_log to avoid
     paying the 128K-config ``_run_panel`` runtime.
     """
-    from kernbench.benches.milestone_gqa_decode_long_ctx_4cases import (
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases import (
         _end_to_end_ns,
         _engine_occupancy_ns,
     )
@@ -579,7 +579,7 @@ def test_run_panel_returns_latency_and_engine_occupancy(monkeypatch):
     Uses ``monkeypatch.setitem`` on ``_PANEL_DISPATCH`` to lower S_kv to
     ``_SMOKE_S_KV`` just for this test — avoids the 131K runtime.
     """
-    import kernbench.benches.milestone_gqa_decode_long_ctx_4cases as mod
+    import kernbench.benches.gqa_helpers.long_ctx.gqa_decode_long_ctx_4cases as mod
 
     orig_kind, orig_params = mod._PANEL_DISPATCH[_CASE2_PANEL]
     fast_params = {**orig_params, "S_kv": _SMOKE_S_KV}

tests/attention/test_milestone_gqa_headline.py

@@ -1,89 +0,0 @@
-"""Tests for the milestone-gqa-headline bench.
-
-The bench wires the new ``_gqa_attention_prefill_long`` kernel into a
-single-KV-group milestone panel (LLaMA-3.1-70B target, C=8 P=8). The
-4 legacy C=1 / C=4 panels (single_user_*, multi_user_*) were removed
-once pytest regression covered those configurations comprehensively
-and the comparative decode work moved into the
-``milestone-gqa-decode-4cases`` bench.
-
-Single SIP scope (default ``topology.yaml`` 4×4 cube mesh; 4-SIP
-headline deferred per ADR-0060 §B-item-1).
-"""
-from __future__ import annotations
-
-import json
-
-import kernbench.benches.milestone_gqa_headline as bench_mod  # noqa: F401 (Phase 2)
-from kernbench.benches.registry import resolve
-from kernbench.runtime_api.bench_runner import run_bench
-from kernbench.runtime_api.types import resolve_device
-from kernbench.sim_engine.engine import GraphEngine
-from kernbench.topology.builder import resolve_topology
-
-BENCH_NAME = "milestone-gqa-headline"
-
-PANELS = (
-    "single_kv_group_prefill_gqa_c8_p8",
-)
-
-
-def _run_validation():
-    topo = resolve_topology("topology.yaml")
-    return run_bench(
-        topology=topo,
-        bench_fn=resolve(BENCH_NAME).run,
-        device=resolve_device(None),
-        engine_factory=lambda t, d: GraphEngine(
-            getattr(t, "topology_obj", t), enable_data=True,
-        ),
-    )
-
-
-def _sweep_json(monkeypatch) -> dict:
-    monkeypatch.setenv("GQA_HEADLINE_RUN", "1")
-    out = bench_mod._OUTPUT_DIR / "sweep.json"
-    if not out.exists():
-        result = _run_validation()
-        assert result.completion.ok, result.completion
-    assert out.exists(), f"missing {out}"
-    return json.loads(out.read_text())
-
-
-# ── Registration ───────────────────────────────────────────────────────
-
-
-def test_bench_registered():
-    spec = resolve(BENCH_NAME)
-    assert spec.name == BENCH_NAME
-    assert callable(spec.run)
-    assert spec.description.strip(), "description must be non-empty"
-
-
-# ── Validation run completes ──────────────────────────────────────────
-
-
-def test_validation_run_completes_ok(monkeypatch):
-    monkeypatch.setenv("GQA_HEADLINE_RUN", "1")
-    result = _run_validation()
-    assert result.completion.ok, (
-        f"headline validation run failed: {result.completion}"
-    )
-
-
-# ── sweep.json shape ──────────────────────────────────────────────────
-
-
-def test_sweep_json_has_expected_panels(monkeypatch):
-    data = _sweep_json(monkeypatch)
-    assert set(data["panels"]) == set(PANELS), (
-        f"panels mismatch: expected {set(PANELS)}, got {set(data['panels'])}"
-    )
-    assert len(data["rows"]) == len(PANELS)
-    assert {r["panel"] for r in data["rows"]} == set(PANELS)
-
-
-# Per-panel architectural assertions for the surviving single-KV-group
-# panel live in tests/attention/test_milestone_gqa_single_kv_group_prefill_panel.py.
-# Decode architectural assertions live in
-# tests/attention/test_milestone_gqa_decode_4cases.py.

tests/attention/test_milestone_gqa_prefill_long_ctx_4cases.py

@@ -0,0 +1,471 @@
+"""Tests for the long-context prefill 4-cases comparative-study bench.
+
+Prefill counterpart to test_milestone_gqa_decode_long_ctx_4cases. The
+4 cases differ in how KV cache is sharded across the 8 cubes and 8
+PEs of a single KV-head group on LLaMA-3.1-70B GQA:
+
+    Case 1  Cube-SP / PE-TP  → KV split by S_kv across cubes (Ring);
+                               Q/O split T_q across all 64 ranks
+    Case 2  Cube-Repl / PE-TP → full KV per cube; only CUBE 0's P PEs
+                               split T_q (CUBEs 1..7 = pure memory waste)
+    Case 3  Cube-Repl / PE-SP → full KV per cube; PEs SP on S_kv
+                               (intra-cube AR); 8× redundant compute
+    Case 4  Cube-SP / PE-SP  → KV split 64-way + Q split T_q across cubes;
+                               Ring KV + intra-cube AR per cube  ★ optimal
+
+Smoke sizes (T_q=256, S_kv=2048) keep the per-PE scratch comfortably
+inside budget; the headline T_q=S_kv=4096 runs come from
+``kernbench run --bench milestone-gqa-prefill-long-ctx-4cases``, not
+pytest.
+"""
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+from kernbench.runtime_api.bench_runner import run_bench
+from kernbench.runtime_api.types import resolve_device
+from kernbench.sim_engine.engine import GraphEngine
+from kernbench.topology.builder import resolve_topology
+
+TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
+
+_CASE4_PANEL = "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_sp"
+_CASE3_PANEL = "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_sp"
+_CASE2_PANEL = "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_tp"
+_CASE1_PANEL = "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_tp"
+_CUBE_RE = re.compile(r"\bcube(\d+)\b")
+
+# Smoke sizes: T_q=256 (divisible by C·P=64 for Case 1, by P=8 for
+# Case 2, by C=8 for Case 4) and S_kv=2048 (≥ 2·TILE_S_KV so the
+# per-tile fold loop runs; Case 1/4 have S_local sufficient to exercise
+# the Ring). Assertions are size-independent. The headline T_q=S_kv=4096
+# bench runs come from the env-gated bench, not pytest.
+# Smoke sizes — kept small so pytest stays fast:
+#   T_q=64  : minimum that satisfies Case 1's ``T_q % (C·P) == 0``.
+#   S_kv=512: enough to exercise multi-tile fold for Cases 2/3
+#             (TILE_S_KV=64 → 8 tiles for Case 2; 1 tile/PE for Case 3
+#             after intra-cube S_kv split) without burning minutes on
+#             Case 2's single-rank serial path.
+# Cases 2/3 also use TILE_S_KV=64 in their kernels to keep the
+# persistent ``scores`` tensor inside the 1 MB scratch budget at
+# T_q=64. Headline T_q=S_kv=4096 runs come from the env-gated bench,
+# not pytest.
+_SMOKE_T_Q = 64
+_SMOKE_S_KV = 512
+
+_HEADLINE_T_Q = 512
+_HEADLINE_S_KV = 512
+
+
+def _engine_factory(t, d):
+    return GraphEngine(getattr(t, "topology_obj", t), enable_data=True)
+
+
+def _count(op_log, name: str) -> int:
+    return sum(1 for r in op_log if r.op_name == name)
+
+
+def _dma_write_cubes(op_log) -> list[int]:
+    cubes: list[int] = []
+    for r in op_log:
+        if r.op_name != "dma_write":
+            continue
+        m = _CUBE_RE.search(r.component_id)
+        if m is not None:
+            cubes.append(int(m.group(1)))
+    return cubes
+
+
+# ── Case 4 (Cube-SP × PE-SP) — ★ optimal ────────────────────────────
+
+
+def _run_case4_smoke(*, T_q: int, S_kv: int):
+    """Drive the Case 4 prefill panel via the case-specific runner."""
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _run_prefill_panel_long_ctx_cube_sp_pe_sp,
+    )
+    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
+
+    def _bench_fn(ctx):
+        _run_prefill_panel_long_ctx_cube_sp_pe_sp(
+            ctx, panel=_CASE4_PANEL,
+            C=8, P=8,
+            T_q=T_q, 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,
+    )
+
+
+def test_case4_panel_registered():
+    """Case 4 panel must be in ``_PANELS`` + ``_PANEL_DISPATCH`` with
+    the LLaMA-3.1-70B single-KV-group prefill dims (T_q=S_kv=4096).
+    """
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _PANEL_DISPATCH,
+        _PANELS,
+    )
+    assert _CASE4_PANEL in _PANELS, f"{_CASE4_PANEL!r} not in {_PANELS}"
+    assert _CASE4_PANEL in _PANEL_DISPATCH
+    kind, params = _PANEL_DISPATCH[_CASE4_PANEL]
+    assert kind == "prefill_long_ctx_cube_sp_pe_sp"
+    assert params == {
+        "C": 8, "P": 8,
+        "T_q": _HEADLINE_T_Q, "S_kv": _HEADLINE_S_KV,
+        "d_head": 128, "h_q": 8, "h_kv": 1,
+    }
+
+
+def test_case4_runner_smoke():
+    """Case 4 runner drives the new kernel to completion at smoke sizes."""
+    result = _run_case4_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok, (
+        f"Case 4 prefill smoke at C=8 P=8 must complete; "
+        f"got {result.completion}"
+    )
+
+
+def test_case4_writers_one_per_cube():
+    """Case 4 prefill: each cube writes its own disjoint T_q/C rows
+    (PE 0 of each cube). Output cubes must cover all 8.
+    """
+    result = _run_case4_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    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 == set(range(8)), (
+        f"Case 4 expected PE 0 of every cube to write; got cubes={sorted(distinct)}"
+    )
+
+
+def test_case4_ipcq_copy_positive():
+    """Case 4 prefill has both inter-CUBE Ring and intra-CUBE AR
+    traffic; total ipcq_copy must be strictly positive.
+    """
+    result = _run_case4_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok
+    n_copy = _count(result.engine.op_log, "ipcq_copy")
+    assert n_copy > 0, f"Case 4 must have Ring + intra-CUBE comm; got 0 ipcq_copy"
+
+
+# ── Case 2 (Cube-Repl × PE-TP) ──────────────────────────────────────
+
+
+def _run_case2_smoke(*, T_q: int, S_kv: int):
+    """Drive the Case 2 prefill panel via the case-specific runner."""
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _run_prefill_panel_long_ctx_cube_repl_pe_tp,
+    )
+    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
+
+    def _bench_fn(ctx):
+        _run_prefill_panel_long_ctx_cube_repl_pe_tp(
+            ctx, panel=_CASE2_PANEL,
+            C=8, P=8,
+            T_q=T_q, 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,
+    )
+
+
+def test_case2_panel_registered():
+    """Case 2 panel registered with single-KV-group prefill dims."""
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _PANEL_DISPATCH,
+        _PANELS,
+    )
+    assert _CASE2_PANEL in _PANELS, f"{_CASE2_PANEL!r} not in {_PANELS}"
+    assert _CASE2_PANEL in _PANEL_DISPATCH
+    kind, params = _PANEL_DISPATCH[_CASE2_PANEL]
+    assert kind == "prefill_long_ctx_cube_repl_pe_tp"
+    assert params == {
+        "C": 8, "P": 8,
+        "T_q": _HEADLINE_T_Q, "S_kv": _HEADLINE_S_KV,
+        "d_head": 128, "h_q": 8, "h_kv": 1,
+    }
+
+
+def test_case2_runner_smoke():
+    """Case 2 runner drives the new kernel to completion at smoke sizes."""
+    result = _run_case2_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok, (
+        f"Case 2 prefill smoke at C=8 P=8 must complete; "
+        f"got {result.completion}"
+    )
+
+
+def test_case2_zero_ipcq_copy_no_comm():
+    """Case 2's defining property: full KV per rank ⇒ NO inter-rank
+    communication. Slide 11 lists comm cost as 'none'.
+    """
+    result = _run_case2_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok
+    n_copy = _count(result.engine.op_log, "ipcq_copy")
+    assert n_copy == 0, (
+        f"Case 2 must have zero inter-rank comm; got ipcq_copy={n_copy}"
+    )
+
+
+def test_case2_writers_only_cube_0():
+    """Case 2: only CUBE 0 active (PEs 0..P-1 each writing their T_q/P
+    rows). All dma_writes come from cube 0.
+    """
+    result = _run_case2_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    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 2 writers must be cube 0 only; got cubes={sorted(distinct)}"
+    )
+
+
+# ── Case 3 (Cube-Repl × PE-SP) ──────────────────────────────────────
+
+
+def _run_case3_smoke(*, T_q: int, S_kv: int):
+    """Drive the Case 3 prefill panel via the case-specific runner."""
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _run_prefill_panel_long_ctx_cube_repl_pe_sp,
+    )
+    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
+
+    def _bench_fn(ctx):
+        _run_prefill_panel_long_ctx_cube_repl_pe_sp(
+            ctx, panel=_CASE3_PANEL,
+            C=8, P=8,
+            T_q=T_q, 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,
+    )
+
+
+def test_case3_panel_registered():
+    """Case 3 panel registered with single-KV-group prefill dims."""
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _PANEL_DISPATCH,
+        _PANELS,
+    )
+    assert _CASE3_PANEL in _PANELS, f"{_CASE3_PANEL!r} not in {_PANELS}"
+    assert _CASE3_PANEL in _PANEL_DISPATCH
+    kind, params = _PANEL_DISPATCH[_CASE3_PANEL]
+    assert kind == "prefill_long_ctx_cube_repl_pe_sp"
+    assert params == {
+        "C": 8, "P": 8,
+        "T_q": _HEADLINE_T_Q, "S_kv": _HEADLINE_S_KV,
+        "d_head": 128, "h_q": 8, "h_kv": 1,
+    }
+
+
+def test_case3_runner_smoke():
+    """Case 3 runner drives the new kernel to completion at smoke sizes."""
+    result = _run_case3_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok, (
+        f"Case 3 prefill smoke at C=8 P=8 must complete; "
+        f"got {result.completion}"
+    )
+
+
+def test_case3_intra_cube_ar_only_ipcq():
+    """Case 3 reduce pattern: per-CUBE 8-way PE-SP AR (same structural
+    cost as decode Case 3's intra-CUBE phase = 21 ipcq_copy per cube),
+    and NO inter-CUBE traffic (each cube has a full copy of KV).
+
+    Q-axis tiling tradeoff: prefill Case 3 wraps the intra-CUBE AR
+    INSIDE the outer Q-tile loop (the kernel can't hold full
+    (m, ℓ, O) for the whole T_q in scratch). So the AR runs once
+    per Q-tile → total ipcq = n_q_tiles × per-tile cost.
+
+      per-CUBE intra (2×4 PE grid), PER Q-TILE:
+        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
+        × n_q_tiles (= G·T_q / TILE_Q = 8·64/64 = 8)   = 1344
+
+      inter-CUBE: 0 (replicated KV ⇒ no AllReduce needed).
+    """
+    from kernbench.benches.gqa_helpers.long_ctx._gqa_attention_prefill_long_ctx_cube_repl_pe_sp import (
+        TILE_Q,
+    )
+
+    result = _run_case3_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok
+    G = 8  # h_q / h_kv
+    n_q_tiles = (G * _SMOKE_T_Q + TILE_Q - 1) // TILE_Q
+    expected = n_q_tiles * 168
+    n_copy = _count(result.engine.op_log, "ipcq_copy")
+    assert n_copy == expected, (
+        f"Case 3 expected {expected} ipcq_copy "
+        f"({n_q_tiles} Q-tiles × 168 per Q-tile intra-CUBE AR); "
+        f"got {n_copy}"
+    )
+
+
+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.
+    """
+    result = _run_case3_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    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; got cubes={sorted(distinct)}"
+    )
+
+
+# ── Case 1 (Cube-SP × PE-TP) ────────────────────────────────────────
+
+
+def _run_case1_smoke(*, T_q: int, S_kv: int):
+    """Drive the Case 1 prefill panel via the case-specific runner."""
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _run_prefill_panel_long_ctx_cube_sp_pe_tp,
+    )
+    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
+
+    def _bench_fn(ctx):
+        _run_prefill_panel_long_ctx_cube_sp_pe_tp(
+            ctx, panel=_CASE1_PANEL,
+            C=8, P=8,
+            T_q=T_q, 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,
+    )
+
+
+def test_case1_panel_registered():
+    """Case 1 panel registered with single-KV-group prefill dims."""
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _PANEL_DISPATCH,
+        _PANELS,
+    )
+    assert _CASE1_PANEL in _PANELS, f"{_CASE1_PANEL!r} not in {_PANELS}"
+    assert _CASE1_PANEL in _PANEL_DISPATCH
+    kind, params = _PANEL_DISPATCH[_CASE1_PANEL]
+    assert kind == "prefill_long_ctx_cube_sp_pe_tp"
+    assert params == {
+        "C": 8, "P": 8,
+        "T_q": _HEADLINE_T_Q, "S_kv": _HEADLINE_S_KV,
+        "d_head": 128, "h_q": 8, "h_kv": 1,
+    }
+
+
+def test_case1_runner_smoke():
+    """Case 1 runner drives the new kernel to completion at smoke sizes."""
+    result = _run_case1_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok, (
+        f"Case 1 prefill smoke at C=8 P=8 must complete; "
+        f"got {result.completion}"
+    )
+
+
+def test_case1_ipcq_copy_positive_ring_only():
+    """Case 1 has inter-CUBE Ring (P parallel lanes) only; total ipcq
+    must be positive. No intra-CUBE comm (PEs are disjoint on T_q).
+    """
+    result = _run_case1_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok
+    n_copy = _count(result.engine.op_log, "ipcq_copy")
+    assert n_copy > 0, f"Case 1 must have Ring KV comm; got 0 ipcq_copy"
+
+
+def test_case1_writers_all_64_ranks():
+    """Case 1 prefill: every rank owns disjoint T_q/(C·P) rows and
+    writes them. All 8 cubes appear in dma_write component ids.
+    """
+    result = _run_case1_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    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 == set(range(8)), (
+        f"Case 1 expected every cube to write; got cubes={sorted(distinct)}"
+    )
+
+
+# ── Panel-metrics helpers + _run_panel wiring ───────────────────────
+
+
+_EXPECTED_ENGINES = {
+    "pe_gemm", "pe_math", "pe_dma", "pe_fetch_store", "pe_ipcq", "pe_cpu",
+}
+
+
+def test_panel_metrics_helpers_present_and_correct():
+    """Bench file must expose ``_end_to_end_ns`` and
+    ``_engine_occupancy_ns`` for the comparative plot script.
+    Exercised via the Case 2 smoke runner's op_log (cheapest case).
+    """
+    from kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases import (
+        _end_to_end_ns,
+        _engine_occupancy_ns,
+    )
+    result = _run_case2_smoke(T_q=_SMOKE_T_Q, S_kv=_SMOKE_S_KV)
+    assert result.completion.ok
+    op_log = result.engine.op_log
+
+    lat = _end_to_end_ns(op_log)
+    assert lat > 0, f"expected positive end-to-end latency; got {lat}"
+
+    occ = _engine_occupancy_ns(op_log)
+    assert isinstance(occ, dict)
+    assert set(occ.keys()) >= _EXPECTED_ENGINES, (
+        f"engine_occupancy_ns missing required engines; "
+        f"got {set(occ.keys())}"
+    )
+    assert occ["pe_gemm"] > 0, (
+        f"expected pe_gemm occupancy > 0; got {occ['pe_gemm']}"
+    )
+
+
+def test_run_panel_returns_latency_and_engine_occupancy(monkeypatch):
+    """``_run_panel`` must include ``latency_ns`` + ``engine_occupancy_ns``
+    alongside ``op_log_summary`` so sweep.json carries them for the
+    comparative plot script. Uses monkeypatch to swap the Case 2 panel's
+    T_q/S_kv to smoke sizes for speed.
+    """
+    import kernbench.benches.gqa_helpers.long_ctx.gqa_prefill_long_ctx_4cases as mod
+
+    orig_kind, orig_params = mod._PANEL_DISPATCH[_CASE2_PANEL]
+    fast_params = {**orig_params, "T_q": _SMOKE_T_Q, "S_kv": _SMOKE_S_KV}
+    monkeypatch.setitem(
+        mod._PANEL_DISPATCH, _CASE2_PANEL, (orig_kind, fast_params),
+    )
+
+    row = mod._run_panel(_CASE2_PANEL, str(TOPOLOGY_DEFAULT))
+    assert "op_log_summary" in row
+    assert "latency_ns" in row, (
+        f"_run_panel row missing latency_ns; keys={sorted(row.keys())}"
+    )
+    assert row["latency_ns"] > 0
+    assert "engine_occupancy_ns" in row, (
+        f"_run_panel row missing engine_occupancy_ns; "
+        f"keys={sorted(row.keys())}"
+    )
+    assert isinstance(row["engine_occupancy_ns"], dict)
+    assert set(row["engine_occupancy_ns"].keys()) >= _EXPECTED_ENGINES

tests/attention/test_milestone_gqa_single_kv_group_prefill_panel.py

@@ -1,121 +0,0 @@
-"""Tests for the single-KV-group prefill panel in milestone_gqa_headline.
-
-The single-KV-group (LLaMA-3.1-70B target) prefill panel wires together
-all Increment 1–4 work in a single bench panel:
-  - C=8 head-parallel + Ring KV via snake-mapped 2×4 sub-mesh (Inc 1, 3)
-  - P=8 intra-CUBE PE-SP (Inc 4)
-  - d_head=128 (LLaMA-3.1-70B), one-shot prefill T_q = S_kv = 1K
-    (scratch-limited; LLaMA 32K headline awaits Q-axis kernel tiling)
-
-This file verifies the bench-config wiring:
-  T1  the new ``single_kv_group_prefill_gqa_c8_p8`` panel is registered
-      with the expected dims in ``_PANELS`` + ``_PANEL_DISPATCH``.
-  T2  ``_run_prefill_panel`` accepts the new ``P``, ``T_q``, ``d_head``
-      kwargs and drives the prefill kernel to completion at the
-      milestone-target ``(C, P) = (8, 8)``. Uses smaller T_q/S_kv to
-      keep test time bounded — full 32K runs come from
-      ``kernbench run milestone-gqa-headline``.
-  T3  Existing prefill panels still work when ``_run_prefill_panel``
-      is called without the new kwargs (backward compat anchor).
-
-Phase 1: tests only — production code lands in Phase 2.
-T1 fails today (panel not registered).
-T2 fails today (helper doesn't accept the new kwargs).
-T3 passes today as the backward-compat anchor.
-"""
-from __future__ import annotations
-
-from pathlib import Path
-
-from kernbench.benches.milestone_gqa_headline import (  # noqa: F401
-    _PANEL_DISPATCH,
-    _PANELS,
-    _run_prefill_panel,
-)
-from kernbench.runtime_api.bench_runner import run_bench
-from kernbench.runtime_api.types import resolve_device
-from kernbench.sim_engine.engine import GraphEngine
-from kernbench.topology.builder import resolve_topology
-
-TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
-
-
-def _engine_factory(t, d):
-    return GraphEngine(getattr(t, "topology_obj", t), enable_data=True)
-
-
-# ── T1: new panel registered ─────────────────────────────────────────
-
-
-def test_single_kv_group_prefill_panel_registered():
-    """The ``single_kv_group_prefill_gqa_c8_p8`` panel must be in ``_PANELS`` and
-    its dispatch entry must have the expected LLaMA-3.1-70B params.
-
-    Headline config (scratch-limited; LLaMA 32K headline awaits
-    Q-axis kernel tiling):
-      C = 8        (head-parallel, snake sub-mesh)
-      P = 8        (intra-CUBE PE-SP)
-      T_q = 1_024  (one-shot long-context prefill)
-      S_kv = 1_024
-      (scratch-limited; LLaMA 32K headline awaits Q-axis kernel tiling)
-      d_head = 128 (LLaMA-3.1-70B)
-    """
-    panel_name = "single_kv_group_prefill_gqa_c8_p8"
-    assert panel_name in _PANELS, (
-        f"{panel_name!r} not in _PANELS; got {_PANELS}"
-    )
-    assert panel_name in _PANEL_DISPATCH, (
-        f"{panel_name!r} not in _PANEL_DISPATCH"
-    )
-    kind, params = _PANEL_DISPATCH[panel_name]
-    assert kind == "prefill", f"kind={kind!r}, expected 'prefill'"
-    assert params.get("C") == 8, f"C={params.get('C')}, expected 8"
-    assert params.get("P") == 8, f"P={params.get('P')}, expected 8"
-    assert params.get("T_q") == 1_024, (
-        f"T_q={params.get('T_q')}, expected 1_024"
-    )
-    assert params.get("S_kv") == 1_024, (
-        f"S_kv={params.get('S_kv')}, expected 1_024"
-    )
-    assert params.get("d_head") == 128, (
-        f"d_head={params.get('d_head')}, expected 128"
-    )
-
-
-# ── T2: helper drives the C=8, P=8 prefill kernel to completion ──────
-
-
-def test_single_kv_group_prefill_panel_runner_smoke():
-    """``_run_prefill_panel`` must accept the new ``P``, ``T_q``,
-    ``d_head`` kwargs and successfully launch the prefill kernel at
-    ``(C, P) = (8, 8)`` with the snake-mapped 2×4 SFR.
-
-    Uses **smaller** T_q/S_kv than the headline panel so the test
-    completes quickly. The headline 32K dims are exercised via
-    ``kernbench run milestone-gqa-headline``, not pytest.
-    """
-    topo = resolve_topology(str(TOPOLOGY_DEFAULT))
-
-    def _bench_fn(ctx):
-        _run_prefill_panel(
-            ctx,
-            panel="single_kv_group_prefill_gqa_c8_p8",
-            C=8, P=8,
-            T_q=8, S_kv=8192,   # smoke dims, not the 32K headline
-            d_head=128,
-        )
-
-    result = run_bench(
-        topology=topo, bench_fn=_bench_fn,
-        device=resolve_device(None),
-        engine_factory=_engine_factory,
-    )
-    assert result.completion.ok, (
-        f"single_kv_group prefill panel smoke at C=8 P=8 must complete; "
-        f"got {result.completion}"
-    )
-
-
-# Backward-compat test removed when the legacy multi_user_prefill_gqa
-# panel was dropped; signature-extension contract is now exercised
-# implicitly by the live single_kv_group_prefill_gqa_c8_p8 panel.