Files
kernbench2/benches/ccl_allreduce.py
T
ywkang 357cab525b ADR-0026: DPPolicy intra-device only + ShardSpec structural coords
DPPolicy no longer carries a cross-SIP axis. SIP-level placement is
solely controlled by torch.ahbm.set_device(rank) (ADR-0024); DPPolicy
itself describes only the cube × PE layout within one SIP. ShardSpec
switches to structural (sip, cube, pe) coordinates; the flat pe_index
field/property is fully removed — silent drift between global-flat and
SIP-local interpretations was a foot-gun flagged by ADR-0024 D11.

Breaking API (explicit TypeError / AttributeError):
- DPPolicy(sip=...) / DPPolicy(num_sips=...) -> TypeError
- ShardSpec.pe_index -> AttributeError
- ShardSpec(pe_index=...) -> TypeError
- resolve_dp_policy now takes target_sip= (required), no num_sips.

Downstream migration:
- PE allocator dict keyed by (sip, cube, pe) tuples, in both
  _ensure_allocators and _free_tensor. deploy_tensor uses tuple lookup.
- _create_tensor passes target_sip=current_sip; post-hoc pe_index
  shifting removed entirely.
- launch._compute_local_shape drops the dp.sip branch.
- Internal resolvers (column_wise / row_wise / replicate / tiled_*)
  return _LocalPeShard (cube-local identifier) instead of ShardSpec —
  resolve_dp_policy lifts them to full structural coords.

Tests:
- New tests/test_adr0026_dppolicy_intra_device.py (12 tests) pins the
  contract end-to-end.
- test_sip_parallel.py rewritten: SIP composition now modeled as two
  resolve_dp_policy(target_sip=...) calls (ADR-0024 launcher style).
- Call-site migration: test_tensor, test_va_integration, test_va_offset,
  test_runtime_api_tensor, test_tl_recv_async, test_ccl_* and benches
  gemm_single_pe, gpt3_qkv, va_offset_verify, ccl_allreduce (legacy
  branch) all use intra-device DPPolicy and structural ShardSpec.

Result: 523 passed, 1 strict xfail (ring_default_ws — unchanged
ADR-0024 Phase B blocker; architectural fix deferred to ADR-0027).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-14 13:02:19 -07:00

188 lines
7.7 KiB
Python

"""CCL all-reduce bench (ADR-0024 Phase A).
Driven entirely by ``ccl.yaml`` + ``topology.yaml``:
- ``defaults.algorithm`` in ``ccl.yaml`` picks which kernel to run.
- ``world_size`` resolution: explicit override in ccl.yaml > defaults >
topology's SIP count. ADR-0024 D1: topology fallback is the SIP count
(each rank = one SIP, TP boundary).
- ``run()`` is hybrid:
- If ``world_size == topology SIP count`` (the intended new path):
spawn one greenlet per rank, bind it via ``dist._bind_rank``, and
each worker calls ``torch.ahbm.set_device(rank)`` + runs its portion
of the collective. Cross-rank IPCQ exchange handles the reduce.
- Legacy path (``world_size > SIP count``, via explicit ccl.yaml
override): single worker at rank 0 with the full tensor distributed
across all participating PEs via ``_derive_dp``. Retained for
backward compatibility with existing kernel / topology tests.
"""
from __future__ import annotations
import numpy as np
from greenlet import greenlet
from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
from kernbench.policy.placement.dp import DPPolicy
# Default per-rank tile size if ccl.yaml doesn't override it.
DEFAULT_N_ELEM = 32
def _derive_dp(spec: dict, world_size: int) -> DPPolicy:
"""Legacy DPPolicy for world_size > SIP count (rank = flat PE index).
Used only in the ccl.yaml-override path so the existing matrix tests
with explicit world_size (8, 16, 7 etc.) keep working. ADR-0026:
DPPolicy is intra-device only, so this legacy path now always stays
within a single SIP and distributes the override world_size across
that SIP's cubes and PEs.
"""
pl = spec["cube"]["pe_layout"]
pes_per_cube = int(pl["pe_per_corner"]) * len(pl["corners"])
cm = spec["sip"]["cube_mesh"]
cubes_per_sip = int(cm["w"]) * int(cm["h"])
if world_size <= pes_per_cube:
return DPPolicy(
cube="replicate", pe="column_wise",
num_cubes=1, num_pes=world_size,
)
if world_size <= cubes_per_sip * pes_per_cube:
return DPPolicy(
cube="column_wise", pe="column_wise",
num_cubes=world_size // pes_per_cube,
)
return DPPolicy(cube="column_wise", pe="column_wise")
def worker(rank: int, world_size: int, torch) -> None:
"""Per-rank worker (new TP path) OR single-worker legacy driver.
Behaviour depends on whether this call originates from the
multi-greenlet launcher (new path) or from the legacy single-call
fallback; distinguished by which ``dp`` layout applies.
"""
cfg = resolve_algorithm_config(load_ccl_config())
algo_name = cfg["algorithm"]
n_elem = int(cfg.get("n_elem", DEFAULT_N_ELEM))
spec = torch.spec or {}
n_sips = int(spec.get("system", {}).get("sips", {}).get("count", 1))
if world_size == n_sips:
# ADR-0024 new path: rank = SIP, worker sees its SIP's
# representative PE via torch.ahbm.set_device.
torch.ahbm.set_device(rank)
dp = DPPolicy(cube="replicate", pe="replicate",
num_cubes=1, num_pes=1)
tensor = torch.zeros(
(1, n_elem), dtype="f16", dp=dp, name=f"ccl_in_r{rank}",
)
# Each rank initialises its tile with (rank + 1); after all_reduce
# every rank sees sum(1..world_size).
init = np.full((1, n_elem), float(rank + 1), dtype=np.float16)
tensor.copy_(torch.from_numpy(init))
torch.distributed.all_reduce(tensor, op="sum")
result = tensor.numpy()
expected = float(sum(range(1, world_size + 1)))
all_ok = bool(np.allclose(result, expected, rtol=1e-1, atol=1e-1))
if rank == 0:
if all_ok:
print(f" {algo_name} (ws={world_size}): {world_size} OK")
else:
print(
f" [FAIL] rank {rank} "
f"(ws={world_size}, algo={algo_name}): "
f"got mean={float(result.reshape(-1).mean()):.3f}, "
f"expected={expected:.3f}"
)
print(
f" {algo_name} (ws={world_size}): "
f"0 OK / {world_size} FAIL"
)
return
# Legacy path: world_size overridden via ccl.yaml to exceed SIP count.
# Single-worker at rank 0; whole tensor distributed across all
# participating PEs using the derived DPPolicy. Matches pre-ADR-0024
# behaviour.
dp = _derive_dp(spec, world_size)
tensor = torch.zeros(
(1, world_size * n_elem), dtype="f16", dp=dp, name="ccl_in",
)
init = np.zeros((1, world_size * n_elem), dtype=np.float16)
for r in range(world_size):
init[0, r * n_elem : (r + 1) * n_elem] = float(r + 1)
tensor.copy_(torch.from_numpy(init))
torch.distributed.all_reduce(tensor, op="sum")
result = tensor.numpy()
expected = float(sum(range(1, world_size + 1)))
all_ok = bool(np.allclose(result, expected, rtol=1e-1, atol=1e-1))
if rank == 0:
if all_ok:
print(f" {algo_name} (ws={world_size}): {world_size} OK")
else:
flat = result.reshape(-1)
n_fail = 0
for r in range(world_size):
slice_r = flat[r * n_elem : (r + 1) * n_elem]
if not np.allclose(slice_r, expected, rtol=1e-1, atol=1e-1):
n_fail += 1
if n_fail <= 5:
print(
f" [FAIL] rank {r} "
f"(ws={world_size}, algo={algo_name}): "
f"got mean={float(slice_r.mean()):.3f}, "
f"expected={expected:.3f}"
)
print(
f" {algo_name} (ws={world_size}): "
f"{world_size - n_fail} OK / {n_fail} FAIL"
)
def run(torch) -> None:
"""CLI entry — dispatch to multi-greenlet path when ws == SIP count,
else fall back to single-worker legacy path for ccl.yaml override compat.
"""
dist = torch.distributed
dist.init_process_group(backend="ahbm")
world_size = dist.get_world_size()
spec = torch.spec or {}
n_sips = int(spec.get("system", {}).get("sips", {}).get("count", 1))
if world_size == n_sips:
# ADR-0024 D12/D13: one greenlet per rank. After each scheduler
# round, the main greenlet drains any pending collective handles
# (ADR-0024 D7) — this must happen in the main context, not inside
# a worker, so env.run is invoked with main as the current greenlet
# and kernel_runner's spawned kernel greenlets correctly get main
# as their parent.
backend = dist._backend
gs: list[greenlet] = []
for rank in range(world_size):
def _entry(r: int = rank) -> None:
worker(r, world_size, torch)
g = greenlet(_entry)
dist._bind_rank(g, rank)
gs.append(g)
while True:
alive = [g for g in gs if not g.dead]
if not alive:
break
for g in alive:
if not g.dead:
g.switch()
# Drain pending collective handles. All sibling workers have
# either submitted (and yielded) or completed; their kernels
# are live in the SimPy queue, ready to exchange via IPCQ.
pending = backend._pending_collective_handles
if pending:
for h, _sip_id, meta in pending:
torch.wait(h, _meta=meta)
backend._pending_collective_handles = []
else:
# Legacy single-worker path (ccl.yaml world_size override).
worker(rank=dist.get_rank(), world_size=world_size, torch=torch)