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

gqa: single-KV-group LLaMA-3.1-70B prefill milestone (Increments 1-5)

Browse Source 
End-to-end wires C=8 P=8 d_head=128 prefill with snake-ring inter-CUBE
SFR, intra-CUBE PE-SP (all 64 ranks active), and the milestone bench
panel. Decode kernel gains lrab-adapted center-root reduce for the
2×4 sub-mesh per ADR-0060 §4.2.

Increment 1 — SFR multi-row snake
  src/kernbench/ccl/sfr_config.py: configure_sfr_intercube_ring gains
  submesh_shape / submesh_origin kwargs; installs a Hamiltonian snake
  ring through a rectangular sub-mesh (every hop is 1-hop physical
  neighbour). Backward-compat: 1D-row behaviour preserved when
  submesh_shape is None.
  tests/test_intercube_snake_ring.py (12 tests)

Increment 2 — Decode lrab-adapted center-root reduce
  src/kernbench/benches/_gqa_attention_decode_long.py: new sub_w param
  (default 0 = existing 1D-chain). sub_w >= 2 selects the ADR-0060
  §4.2 prescribed lrab-adapted Phase 1+2 reduce (bidirectional row +
  bidirectional col converge to the center cube), with log-sum-exp
  _merge_running replacing the plain + of lrab.
  tests/attention/test_gqa_decode_long_2d_reduce.py (4 tests)

Increment 3 — Prefill kernel at C=8 (no production change)
  Verified by inspection that the existing prefill_long kernel +
  Increment 1's snake SFR already work at C=8 without any kernel
  edit. The kernel speaks logical W/E; the snake routes it.
  tests/attention/test_gqa_prefill_long_c8_snake.py (3 tests)

Increment 4 — Intra-CUBE PE-SP in prefill (all 64 ranks)
  src/kernbench/benches/_gqa_attention_prefill_long.py: new P param
  (default 1 = existing PE-0-only). P > 1 splits T_q query-axis-wise
  across the P PEs of each CUBE; output rows are disjoint per PE so
  no intra-CUBE reduce is needed; each PE drives its own same-lane
  ring (P parallel rings).
  tests/attention/test_gqa_prefill_long_pe_sp.py (5 tests)

Increment 5 — LLaMA-scale milestone bench panel
  src/kernbench/benches/milestone_gqa_headline.py: new panel
  single_kv_group_prefill_gqa_c8_p8 (C=8, P=8, T_q=S_kv=32K,
  d_head=128). _run_prefill_panel extended with P/T_q/d_head
  defaults; routes snake SFR when C > mesh_w.
  tests/attention/test_milestone_gqa_single_kv_group_prefill_panel.py (3 tests)

Total: 4 production files modified, 5 new test files, 27 new tests.
Followed the Phase 1/2 protocol per CLAUDE.md throughout.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Mukesh Garg
2026-06-15 13:42:31 -07:00
parent 9270d3435a
commit 9e1242039b
9 changed files with 1332 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/kernbench/benches/_gqa_attention_decode_long.py
+176 -8
View File
@@ -4,14 +4,29 @@ Each rank holds an ``S_local = S_kv / (C·P)`` slice of K, V and the full
Q (replicated). The local attention is computed via an S_kv-axis tile
sweep (ADR-0063 §A.2) so per-rank scratch is bounded by ``TILE_S_KV``
regardless of ``S_local``. The partial ``(m, ℓ, O)`` is then reduced
to PE 0 of CUBE 0 via a 2-level chain (intra-CUBE row+col, then
inter-CUBE), and the root writes the final output.
via a 2-level pattern (intra-CUBE row+col, then inter-CUBE), and the
root writes the final output.
Inter-CUBE reduce (selected by the ``sub_w`` launch arg):
- ``sub_w == 0`` (default): 1D chain along ``W``, root at CUBE 0.
Used for ``C ∈ {1, 4}`` panels.
- ``sub_w >= 2``: ADR-0060 §4.2 prescribed **lrab-adapted center-root
mesh reduce** over a ``sub_w × sub_h`` sub-mesh (``sub_h = C //
sub_w``). Phase 1 row reduce converges at ``root_col = sub_w//2``;
Phase 2 col reduce on ``root_col`` converges at
``root_row = sub_h//2``. Root cube id:
``root_row * sub_w + root_col``. Used for the C=8 single-KV-group
LLaMA-3.1-70B target (sub_w=4, sub_h=2, root=cube 6).
Requires ``sub_w >= 2 and sub_h >= 2`` — degenerate 1×N / N×1
layouts must use the 1D-chain path with ``sub_w=0``.
Topology / SFR:
- Requires ``configure_sfr_intercube_multisip`` when ``P > 1`` or
``C > 1`` (provides disjoint ``intra_*`` and ``E/W/N/S`` namespaces).
- Intra-CUBE PEs are arranged as a 2×4 grid (no wrap).
- Inter-CUBE CUBEs are arranged as a 1D row (no wrap).
- Inter-CUBE CUBEs: 1D row (``sub_w == 0``) or rectangular sub-mesh
of the SIP's 4×4 CUBE mesh starting at origin (0, 0) with
``sub_w == mesh_w`` (``sub_w >= 2``).
Layout caveats:
- GEMMs use ``tl.dot`` (no composite epilogue / ``softmax_scale``).
@@ -47,10 +62,11 @@ def gqa_attention_decode_long_kernel(
d_head: int,
C: int,
P: int,
sub_w: int = 0,
*,
tl,
) -> None:
"""GQA decode with M-fold + S_kv tile sweep + 2-level chain reduce-to-root.
"""GQA decode with M-fold + S_kv tile sweep + 2-level reduce-to-root.
Tensor layout:
Q : (T_q, h_q · d_head) replicated on every rank; loaded as
@@ -59,8 +75,26 @@ def gqa_attention_decode_long_kernel(
loads its (d_head, S_local) slice via byte-conserving reshape.
V : (S_kv, h_kv · d_head) sharded row_wise by (cube, pe); each rank
loads its (S_local, d_head) slice.
O : (T_q, h_q · d_head) — only PE 0 of CUBE 0 stores.
O : (T_q, h_q · d_head) — only PE 0 of the root CUBE stores
(CUBE 0 for ``sub_w=0``; geometric center cube for ``sub_w>=2``).
"""
if sub_w > 0:
sub_h = C // sub_w
if sub_w < 2 or sub_h < 2 or sub_w * sub_h != C:
raise ValueError(
f"sub_w={sub_w} requires sub_w>=2 and sub_h>=2 and "
f"sub_w*sub_h==C; got sub_h={sub_h}, C={C}. "
"Use sub_w=0 for the 1D-chain path."
)
root_col = sub_w // 2
root_row = sub_h // 2
root_cube = root_row * sub_w + root_col
else:
sub_h = 0
root_col = 0
root_row = 0
root_cube = 0
G = h_q // h_kv
n_ranks = C * P
S_local = S_kv // n_ranks
@@ -161,8 +195,9 @@ def gqa_attention_decode_long_kernel(
tl.send(dir="intra_N", src=l_local)
tl.send(dir="intra_N", src=O_local)
# Level-1 inter-CUBE chain (along W, leftward; only PE 0 of each CUBE).
if pe_id == 0 and C > 1:
# Level-1 inter-CUBE reduce (only PE 0 of each CUBE participates).
if pe_id == 0 and C > 1 and sub_w == 0:
# 1D chain along W, leftward; root at CUBE 0.
if cube_id < C - 1:
with tl.scratch_scope():
m_other = tl.recv(dir="E", shape=m_local.shape, dtype="f16")
@@ -178,8 +213,141 @@ def gqa_attention_decode_long_kernel(
tl.send(dir="W", src=m_local)
tl.send(dir="W", src=l_local)
tl.send(dir="W", src=O_local)
elif pe_id == 0 and sub_w > 0:
# ── ADR-0060 §4.2 lrab-adapted center-root mesh reduce ──
# Adapts Phases 1-2 of lrab_hierarchical_allreduce.py:
# bidirectional row reduce converges at root_col, bidirectional
# col reduce on root_col converges at root_row. Reduce-only
# (Phases 3-5 dropped: no inter-SIP, no broadcast — attention
# needs the answer at one rank). Plain ``+`` replaced with the
# log-sum-exp ``_merge_running`` for (m, ℓ, O).
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) ──
if pe_id == 0 and cube_id == 0:
if pe_id == 0 and cube_id == root_cube:
O_final = O_local / l_local
tl.store(o_ptr, O_final)