gqa: reorganize benches into gqa_helpers/ subpackage; drop legacy headline

Splits the GQA helpers into a dedicated subpackage to make room for the
prefill 4-cases study (next commit) and a single umbrella bench
(milestone-1h-gqa, after that).

Layout:
  benches/gqa_helpers/
    long_ctx/    — decode 4-cases kernels + sweep runner
    short_ctx/   — prefill/decode short-context kernels
    shared/      — _gqa_panel_helpers + decode_opt2 (context-agnostic)

The registry audit now skips subpackages so gqa_helpers/ (without a
leading underscore) doesn't get audited for @bench decorators.

Also drops the legacy milestone-gqa-headline bench, its
_gqa_attention_prefill_long kernel, 6 dependent prefill tests, the
paper_gqa_latency.py report harness, and the 3 stale headline-derived
PNGs the §6 wire-up referenced (paper will re-pull from the new
1H_milestone_output/gqa/long_ctx/ once §6 is updated).

The _ccl_cfg and _summarize_op_log helpers used to live in the
headline bench; extracted them to gqa_helpers/shared/_gqa_panel_helpers.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
2026-06-16 13:05:41 -07:00
parent 359a0eaa44
commit e45626c036
32 changed files with 178 additions and 1674 deletions
@@ -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_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 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__":