ywkang/kernbench2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

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

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

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Mukesh Garg
2026-06-15 15:30:00 -07:00
parent ae942f6959
commit 0ef4fde5d8
3 changed files with 407 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/kernbench/benches/_gqa_attention_decode_cube_sp_pe_tp.py
+238
View File
@@ -0,0 +1,238 @@
"""GQA decode kernel — Case 1 (Cube-SP × PE-TP).
Per GQA_full_deck.pptx slide 11:
- K, V split S_kv-wise across the 8 cubes (cube=row_wise);
replicated within each cube (pe=replicate). Each active rank
owns S_local = S_kv / C.
- PEs nominally split on the batch dim (PE-TP). At B=1 (the
slide-11 single-user default) only PE 0 of each cube has work;
PEs 1-7 of every cube idle (PE-TP-at-B=1 waste).
- NO intra-CUBE communication (only PE 0 holds a valid partial).
- Inter-CUBE 8-way reduce on (m, ℓ, O) via the lrab-adapted
center-root pattern (ADR-0060 §4.2): Phase 1 row reduce
converges at root_col; Phase 2 col reduce on root_col converges
at root_row. For sub_w=4, sub_h=2: root_col=2, root_row=1,
root_cube=6 (lrab geometric center).
Tensor layout:
Q : (T_q, h_q · d_head) replicated on every rank.
K : (S_kv, h_kv · d_head) with cube=row_wise, pe=replicate —
each cube's PE 0 owns the full (S_local, h_kv·d_head) shard.
V : same as K.
O : (T_q, h_q · d_head) — only PE 0 of CUBE 6 stores.
Topology / SFR:
- Requires ``configure_sfr_intercube_multisip`` for the E/W/N/S
inter-CUBE lanes used by the lrab reduce.
"""
from __future__ import annotations
from kernbench.benches._gqa_attention_decode_long import _merge_running
TILE_S_KV = 1024 # ADR-0063 §A.2 S_kv-axis tile sweep (per-tile width).
def gqa_attention_decode_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 decode: PE 0 per cube does attention; inter-CUBE lrab AR only."""
# B=1 single-user decode + PE-TP: only PE 0 per cube has work.
pe_id = tl.program_id(axis=0)
cube_id = tl.program_id(axis=1)
if pe_id != 0:
return
# Cube-SP geometry: 4×2 sub-mesh, lrab center-root cube = 6.
sub_w = 4
sub_h = C // sub_w
if sub_w < 2 or sub_h < 2 or sub_w * sub_h != C:
raise ValueError(
f"Case 1 requires C decomposable as sub_w=4, sub_h>=2; got C={C}"
)
root_col = sub_w // 2
root_row = sub_h // 2
root_cube = root_row * sub_w + root_col
G = h_q // h_kv
S_local = S_kv // C # cube=row_wise, pe=replicate ⇒ S_local per cube
KV_ROW_BYTES = d_head * 2 # f16
# ── Local attention (S_kv-axis tile sweep, ADR-0063 §A.2) ──
Q = tl.load(q_ptr, shape=(G * T_q, d_head), dtype="f16")
n_tiles = (S_local + TILE_S_KV - 1) // TILE_S_KV
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)
centered = scores - m_local
exp_scores = tl.exp(centered)
l_local = tl.sum(exp_scores, axis=-1)
O_local = tl.dot(exp_scores, V)
for tile_idx in range(1, n_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)
centered_t = scores_t - m_tile
exp_scores_t = tl.exp(centered_t)
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)
# ── Inter-CUBE lrab-adapted center-root reduce (ADR-0060 §4.2) ──
row = cube_id // sub_w
col = cube_id % sub_w
# Phase 1: row reduce — converge at col == root_col.
if col == 0 and root_col > 0:
tl.send(dir="E", src=m_local)
tl.send(dir="E", src=l_local)
tl.send(dir="E", src=O_local)
elif 0 < col < root_col:
with tl.scratch_scope():
m_other = tl.recv(dir="W", shape=m_local.shape, dtype="f16")
l_other = tl.recv(dir="W", shape=l_local.shape, dtype="f16")
O_other = tl.recv(dir="W", 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)
tl.send(dir="E", src=m_local)
tl.send(dir="E", src=l_local)
tl.send(dir="E", src=O_local)
elif col == root_col:
if root_col > 0:
with tl.scratch_scope():
m_other = tl.recv(dir="W", shape=m_local.shape, dtype="f16")
l_other = tl.recv(dir="W", shape=l_local.shape, dtype="f16")
O_other = tl.recv(dir="W", 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 sub_w - 1 > root_col:
with tl.scratch_scope():
m_other = tl.recv(dir="E", shape=m_local.shape, dtype="f16")
l_other = tl.recv(dir="E", shape=l_local.shape, dtype="f16")
O_other = tl.recv(dir="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)
elif root_col < col < sub_w - 1:
with tl.scratch_scope():
m_other = tl.recv(dir="E", shape=m_local.shape, dtype="f16")
l_other = tl.recv(dir="E", shape=l_local.shape, dtype="f16")
O_other = tl.recv(dir="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)
tl.send(dir="W", src=m_local)
tl.send(dir="W", src=l_local)
tl.send(dir="W", src=O_local)
elif col == sub_w - 1 and sub_w - 1 > root_col:
tl.send(dir="W", src=m_local)
tl.send(dir="W", src=l_local)
tl.send(dir="W", src=O_local)
# Phase 2: col reduce on col == root_col — converge at row == root_row.
if col == root_col:
if row == 0 and root_row > 0:
tl.send(dir="S", src=m_local)
tl.send(dir="S", src=l_local)
tl.send(dir="S", src=O_local)
elif 0 < row < root_row:
with tl.scratch_scope():
m_other = tl.recv(dir="N", shape=m_local.shape, dtype="f16")
l_other = tl.recv(dir="N", shape=l_local.shape, dtype="f16")
O_other = tl.recv(dir="N", 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)
tl.send(dir="S", src=m_local)
tl.send(dir="S", src=l_local)
tl.send(dir="S", src=O_local)
elif row == root_row:
if root_row > 0:
with tl.scratch_scope():
m_other = tl.recv(dir="N", shape=m_local.shape, dtype="f16")
l_other = tl.recv(dir="N", shape=l_local.shape, dtype="f16")
O_other = tl.recv(dir="N", 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 sub_h - 1 > root_row:
with tl.scratch_scope():
m_other = tl.recv(dir="S", shape=m_local.shape, dtype="f16")
l_other = tl.recv(dir="S", shape=l_local.shape, dtype="f16")
O_other = tl.recv(dir="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)
elif root_row < row < sub_h - 1:
with tl.scratch_scope():
m_other = tl.recv(dir="S", shape=m_local.shape, dtype="f16")
l_other = tl.recv(dir="S", shape=l_local.shape, dtype="f16")
O_other = tl.recv(dir="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)
tl.send(dir="N", src=m_local)
tl.send(dir="N", src=l_local)
tl.send(dir="N", src=O_local)
elif row == sub_h - 1 and sub_h - 1 > root_row:
tl.send(dir="N", src=m_local)
tl.send(dir="N", src=l_local)
tl.send(dir="N", src=O_local)
# ── Final normalise + store (root only: PE 0 of cube 6) ──
if cube_id == root_cube:
O_final = O_local / l_local
tl.store(o_ptr, O_final)
src/kernbench/benches/milestone_gqa_decode_4cases.py
+48 -1
View File
@@ -38,6 +38,9 @@ from kernbench.benches._gqa_attention_decode_cube_repl_pe_sp import (
from kernbench.benches._gqa_attention_decode_cube_repl_pe_tp import (
gqa_attention_decode_cube_repl_pe_tp_kernel,
)
from kernbench.benches._gqa_attention_decode_cube_sp_pe_tp import (
gqa_attention_decode_cube_sp_pe_tp_kernel,
)
from kernbench.benches.milestone_gqa_headline import (
_ccl_cfg,
_run_decode_panel,
@@ -62,7 +65,7 @@ _PANELS = (
"single_kv_group_decode_gqa_cube_sp_pe_sp", # Case 4 ★ optimal
"single_kv_group_decode_gqa_cube_repl_pe_tp", # Case 2 (no comm; 8× memory)
"single_kv_group_decode_gqa_cube_repl_pe_sp", # Case 3 (intra-CUBE AR only; 8× memory)
# Case 1 to be added by 5C.A
"single_kv_group_decode_gqa_cube_sp_pe_tp", # Case 1 (inter-CUBE lrab only; PE-TP B=1 waste)
)
# Each entry: (kind, panel-specific params).
@@ -94,6 +97,14 @@ _PANEL_DISPATCH: dict[str, tuple[str, dict]] = {
"T_q": 1, "S_kv": 131_072,
"d_head": 128, "h_q": 8, "h_kv": 1,
}),
"single_kv_group_decode_gqa_cube_sp_pe_tp": ("decode_cube_sp_pe_tp", {
# Case 1: K, V split across cubes (S_local = S_kv/C per cube);
# 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,
"d_head": 128, "h_q": 8, "h_kv": 1,
}),
}
@@ -164,6 +175,40 @@ def _run_decode_panel_cube_repl_pe_sp(
)
def _run_decode_panel_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 across cubes; PE-TP single-rank work at B=1.
DPPolicy: K, V are cube=row_wise (S_local = S_kv/C per cube),
pe=replicate (full S_local within each cube). At B=1, the kernel
runs only on PE 0 of every cube; PEs 1-7 early-return. PE 0 per
cube does local attention on S_local, then participates in the
inter-CUBE lrab AR; PE 0 of cube 6 (lrab center) writes O.
"""
configure_sfr_intercube_multisip(ctx.engine, ctx.spec, _ccl_cfg())
dp_full = DPPolicy(cube="replicate", pe="replicate",
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_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_decode_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 _make_bench_fn(panel: str):
kind, params = _PANEL_DISPATCH[panel]
@@ -174,6 +219,8 @@ def _make_bench_fn(panel: str):
_run_decode_panel_cube_repl_pe_tp(ctx, panel=panel, **params)
elif kind == "decode_cube_repl_pe_sp":
_run_decode_panel_cube_repl_pe_sp(ctx, panel=panel, **params)
elif kind == "decode_cube_sp_pe_tp":
_run_decode_panel_cube_sp_pe_tp(ctx, panel=panel, **params)
else:
raise RuntimeError(
f"milestone-gqa-decode-4cases panel {panel!r} has "