Intercube allreduce: pe0 cube-mesh reduce + multi-SIP ring/torus/mesh

New intercube allreduce kernel replacing the old flat ring algorithms.
Reduces across the 4x4 cube mesh within each SIP (pe0-only, same-lane),
then inter-SIP exchange on root cube, then broadcast back. Supports
ring_1d, torus_2d, and mesh_2d_no_wrap SIP topologies driven by
topology.yaml. Integrated with dist.init_process_group / dist.all_reduce.

New files:
- src/kernbench/ccl/algorithms/intercube_allreduce.py (kernel)
- src/kernbench/ccl/sfr_config.py (configure_sfr_intercube_multisip)
- tests/test_allreduce_multidevice.py (config-driven, 3 topologies)
- tests/test_distributed_intercube_allreduce.py (full distributed path)
- tests/test_intercube_sfr_config.py (SFR wiring verification)

Modified:
- distributed.py: AhbmCCLBackend uses configure_sfr_intercube_multisip
- topologies.py: added torus_2d, mesh_2d_no_wrap
- install.py: global_E/W/N/S in _OPPOSITE_DIR
- topology.yaml: added system.sips.topology
- ccl.yaml: single intercube_allreduce algorithm
- benches/ccl_allreduce.py: row_wise cube-mesh tensor layout

Removed old flat-ring algorithms and their tests.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-04-16 17:33:42 -07:00
parent cfc2d74ec4
commit 1d8b9401e5
30 changed files with 876 additions and 2892 deletions
+222
View File
@@ -0,0 +1,222 @@
"""Config-driven multi-device allreduce test application.
Reads ``ccl.yaml`` + ``topology.yaml``, dynamically loads the kernel
module from ``ccl.yaml → module``, and picks the inter-SIP exchange
pattern from ``topology.yaml → system.sips.topology``.
Run directly::
python -m pytest tests/allreduce_app.py -v -s
"""
from __future__ import annotations
import importlib
import math
from pathlib import Path
from typing import Any
import numpy as np
from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
from kernbench.ccl.sfr_config import configure_sfr_intercube_multisip
from kernbench.policy.placement.dp import DPPolicy
def _sip_topo_dims(sip_topo: str, n_sips: int) -> tuple[int, int]:
if sip_topo == "ring_1d":
return (0, 0)
side = int(round(math.sqrt(n_sips)))
if side * side != n_sips:
raise ValueError(
f"SIP topology '{sip_topo}' requires square n_sips, got {n_sips}"
)
return (side, side)
def run_allreduce(
ctx: Any,
engine: Any,
spec: dict,
*,
algorithm: str | None = None,
ccl_yaml: str | None = None,
) -> dict:
"""Config-driven allreduce: read yaml, load kernel, run.
Everything is resolved from config — no hardcoded kernel imports.
"""
cfg_all = load_ccl_config(ccl_yaml)
cfg = resolve_algorithm_config(cfg_all, algorithm)
# Dynamic import from ccl.yaml → module
algo_module = importlib.import_module(cfg["module"])
kernel_fn = algo_module.kernel
topo_name_to_kind = algo_module.TOPO_NAME_TO_KIND
n_elem = int(cfg.get("n_elem", 8))
n_sips = int(spec.get("system", {}).get("sips", {}).get("count", 1))
sip_topo = str(
spec.get("system", {}).get("sips", {}).get("topology", "ring_1d")
)
cm = spec["sip"]["cube_mesh"]
cube_w = int(cm["w"])
cube_h = int(cm["h"])
n_cubes = cube_w * cube_h
sip_topo_kind = topo_name_to_kind.get(sip_topo, 0)
sip_topo_w, sip_topo_h = _sip_topo_dims(sip_topo, n_sips)
algo_name = cfg.get("algorithm", "allreduce")
print(f"\n{'=' * 60}")
print(f"algorithm: {algo_name}")
print(f"module: {cfg['module']}")
print(f"sip_topology: {sip_topo}")
print(f"kernel: {kernel_fn.__name__}")
print(f"n_sips: {n_sips}")
print(f"n_cubes: {n_cubes}")
print(f"n_elem: {n_elem}")
print(f"{'=' * 60}")
configure_sfr_intercube_multisip(engine, spec, cfg)
dp = DPPolicy(
cube="row_wise", pe="replicate",
num_pes=1, num_cubes=n_cubes,
)
tensors = []
for sip in range(n_sips):
ctx.ahbm.set_device(sip)
t = ctx.zeros(
(n_cubes, n_elem), dtype="f16", dp=dp,
name=f"sip{sip}",
)
t.copy_(ctx.from_numpy(
np.full((n_cubes, n_elem), float(sip + 1), dtype=np.float16)
))
tensors.append(t)
for sip in range(n_sips):
arr = tensors[sip].numpy()
print(f"[SIP {sip}] input cube0[:4] = {arr[0][:4].tolist()} "
f"cube{n_cubes - 1}[:4] = {arr[-1][:4].tolist()}")
t_start = engine._env.now
all_pending = []
for sip_rank, t in enumerate(tensors):
pending = ctx.launch(
algo_name, kernel_fn, t,
n_elem, cube_w, cube_h, n_sips, sip_rank,
sip_topo_kind, sip_topo_w, sip_topo_h,
_defer_wait=True,
)
all_pending.extend(pending)
for h, sip_id, meta in all_pending:
ctx.wait(h, _meta=meta)
t_end = engine._env.now
latency_ns = t_end - t_start
print(f"\n[{algo_name} ws={n_sips}] sim latency = "
f"{latency_ns:.1f} ns ({latency_ns / 1000:.3f} us)")
for key, (_, trace) in engine._results.items():
if not isinstance(trace, dict):
continue
total = trace.get("total_ns", 0.0)
pe_exec = trace.get("pe_exec_ns", 0.0) or 0.0
network = total - pe_exec
print(f" [{key}] total={total:.1f} ns "
f"pe_exec={pe_exec:.1f} ns network={network:.1f} ns")
expected = float(n_cubes * sum(range(1, n_sips + 1)))
print()
for sip in range(n_sips):
arr = tensors[sip].numpy()
print(f"[SIP {sip}] output cube0[:4] = {arr[0][:4].tolist()}")
print(f"[SIP {sip}] output cube{n_cubes - 1}[:4] = {arr[-1][:4].tolist()}")
ok_cubes = 0
for sip in range(n_sips):
arr = tensors[sip].numpy()
for cube_id in range(n_cubes):
assert np.allclose(
arr[cube_id], expected, rtol=1e-1, atol=1e-1,
), (
f"SIP{sip} cube {cube_id}: "
f"got {arr[cube_id][:4]}, expected {expected}"
)
ok_cubes += 1
print(f"\n {algo_name} (ws={n_sips}): {ok_cubes} OK")
return {
"expected": expected,
"latency_ns": latency_ns,
"ok_cubes": ok_cubes,
}
# ── pytest entry point ───────────────────────────────────────────────
import pytest
import yaml
from kernbench.runtime_api.context import RuntimeContext
from kernbench.runtime_api.types import DeviceSelector
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import resolve_topology
TOPOLOGY_PATH = Path(__file__).parent.parent / "topology.yaml"
CONFIGS = [
pytest.param("intercube_allreduce", "ring_1d", 2, id="ring_2sip"),
pytest.param("intercube_allreduce", "torus_2d", 4, id="torus_4sip"),
pytest.param("intercube_allreduce", "mesh_2d_no_wrap", 4, id="mesh_4sip"),
]
def _write_temp_configs(tmp_path, sip_topology, n_sips, algorithm):
"""Write temp topology.yaml and ccl.yaml with the given overrides."""
with open(TOPOLOGY_PATH) as f:
topo_cfg = yaml.safe_load(f)
topo_cfg["system"]["sips"]["count"] = n_sips
topo_cfg["system"]["sips"]["topology"] = sip_topology
topo_path = tmp_path / "topology.yaml"
with open(topo_path, "w") as f:
yaml.dump(topo_cfg, f, default_flow_style=False)
ccl_path = Path(__file__).parent.parent / "ccl.yaml"
with open(ccl_path) as f:
ccl_cfg = yaml.safe_load(f)
ccl_cfg["defaults"]["algorithm"] = algorithm
tmp_ccl = tmp_path / "ccl.yaml"
with open(tmp_ccl, "w") as f:
yaml.dump(ccl_cfg, f, default_flow_style=False)
return str(topo_path), str(tmp_ccl)
@pytest.mark.parametrize("algorithm,sip_topology,n_sips", CONFIGS)
def test_allreduce(tmp_path, algorithm, sip_topology, n_sips):
topo_path, ccl_path = _write_temp_configs(
tmp_path, sip_topology, n_sips, algorithm,
)
topo = resolve_topology(topo_path)
engine = GraphEngine(topo.topology_obj, enable_data=True)
spec = topo.topology_obj.spec
with RuntimeContext(
engine=engine,
target_device=DeviceSelector("all"),
correlation_id=f"test_{algorithm}_{sip_topology}",
spec=spec,
) as ctx:
result = run_allreduce(
ctx, engine, spec,
algorithm=algorithm, ccl_yaml=ccl_path,
)
assert result["ok_cubes"] > 0
-108
View File
@@ -1,108 +0,0 @@
"""End-to-end matrix tests for the unified ``ccl_allreduce`` bench.
Only covers the rank = SIP TP launcher path (ADR-0024 + ADR-0027). Each
case writes a tmp ``ccl.yaml`` that selects a specific (algorithm,
buffer_kind) pair; ``world_size`` is always derived from topology SIP
count (2 in the shipped topology).
The legacy rank = PE single-driver path was removed; intra-SIP PE-level
collectives are expressed inside the kernel via ``tl.program_id`` and do
not require a host-side ``ProcessGroup``.
"""
from __future__ import annotations
import os
import textwrap
import pytest
import kernbench.cli.main as cli_main
CCL_YAML_TEMPLATE = textwrap.dedent("""\
defaults:
algorithm: {algorithm}
buffer_kind: {buffer_kind}
backpressure: sleep
n_slots: 4
slot_size: 4096
vc_chunk_size: 256
ipcq_credit_size_bytes: 16
algorithms:
{algorithm}:
module: {module}
topology: {topology}
buffer_kind: {buffer_kind}
""")
def _write_ccl_yaml(
tmp_path,
*,
algorithm: str,
module: str,
topology: str,
buffer_kind: str,
) -> str:
body = CCL_YAML_TEMPLATE.format(
algorithm=algorithm,
module=module,
topology=topology,
buffer_kind=buffer_kind,
)
(tmp_path / "ccl.yaml").write_text(body)
return str(tmp_path)
CASES = [
# Ring all-reduce across SIPs (ws == topology SIP count = 2),
# one case per IPCQ buffer location.
pytest.param(
"ring_allreduce_tcm", "kernbench.ccl.algorithms.ring_allreduce",
"ring_1d", "tcm",
id="ring_tcm",
),
pytest.param(
"ring_allreduce_hbm", "kernbench.ccl.algorithms.ring_allreduce",
"ring_1d", "hbm",
id="ring_hbm",
),
pytest.param(
"ring_allreduce_sram", "kernbench.ccl.algorithms.ring_allreduce",
"ring_1d", "sram",
id="ring_sram",
),
]
@pytest.mark.parametrize("algorithm,module,topology,buffer_kind", CASES)
def test_ccl_allreduce_matrix(
tmp_path, capsys, monkeypatch,
algorithm, module, topology, buffer_kind,
):
"""Each (algorithm × buffer_kind) combo passes through the unified
rank = SIP bench and yields ``ws OK`` where ``ws == topology SIP count``."""
project_root = os.path.abspath(
os.path.join(os.path.dirname(__file__), "..")
)
yaml_dir = _write_ccl_yaml(
tmp_path,
algorithm=algorithm,
module=module,
topology=topology,
buffer_kind=buffer_kind,
)
monkeypatch.chdir(yaml_dir)
rc = cli_main.main([
"run",
"--topology", os.path.join(project_root, "topology.yaml"),
"--bench", "ccl_allreduce",
"--verify-data",
])
assert rc == 0
out = capsys.readouterr().out
assert "FAIL" not in out, f"unexpected FAIL in output:\n{out}"
assert f"{algorithm}" in out and "OK" in out, (
f"expected pass line for '{algorithm}' in output:\n{out}"
)
-244
View File
@@ -1,244 +0,0 @@
"""Phase 1 tests for ADR-0024 SIP-level TP launcher (MVP scope).
Covers:
- D1 world_size = SIP count fallback
- D9 get_rank greenlet-local + _bind_rank
- D10 torch.ahbm.set_device + torch.accelerator alias
- D11 tensor placement scoped to current device SIP
- D12/D13 run() spawns one greenlet per rank
Deferred to later ADR-0024 sub-phases:
- D2 engine-routed install
- D6 install_plan.py
- D7 epoch barrier (this phase uses simple submit+yield+wait)
- D8 validator registry
"""
from __future__ import annotations
import os
import textwrap
import pytest
from greenlet import greenlet
from kernbench.runtime_api.distributed import AhbmCCLBackend, DistributedContext
# ── Fixtures / helpers ────────────────────────────────────────────────
class _FakeCtx:
"""Minimal ctx double — only exposes what AhbmCCLBackend.__init__ uses.
Stubs install_ipcq so we can unit-test _resolve_world_size without
touching the engine stack.
"""
def __init__(self, spec: dict) -> None:
self.spec = spec
self.install_calls: list[dict] = []
def install_ipcq(self, **kwargs) -> dict:
self.install_calls.append(dict(kwargs))
return {}
def _write_minimal_ccl_yaml(tmp_path) -> str:
"""Write a ccl.yaml with NO world_size override — forces topology derivation."""
body = textwrap.dedent("""\
defaults:
algorithm: ring_allreduce_tcm
buffer_kind: tcm
backpressure: sleep
n_slots: 4
slot_size: 4096
vc_chunk_size: 256
ipcq_credit_size_bytes: 16
algorithms:
ring_allreduce_tcm:
module: kernbench.ccl.algorithms.ring_allreduce
topology: ring_1d
buffer_kind: tcm
n_elem: 8
""")
yaml_path = tmp_path / "ccl.yaml"
yaml_path.write_text(body)
return str(tmp_path)
# ── D1: world_size = SIP count fallback ───────────────────────────────
def test_world_size_equals_sip_count(tmp_path, monkeypatch, spec):
"""With no override, backend derives world_size from SIP count only.
Topology has 2 SIPs × 16 cubes × 8 PEs = 256 PEs. The TP/DP model
places the collective group at the SIP boundary, so world_size must
equal SIP count (2), not total PE count (256).
"""
monkeypatch.chdir(_write_minimal_ccl_yaml(tmp_path))
ctx = _FakeCtx(spec=spec)
backend = AhbmCCLBackend(torch_ctx=ctx)
expected = int(spec["system"]["sips"]["count"])
assert backend.world_size == expected, (
f"expected world_size == SIP count ({expected}); "
f"got {backend.world_size} — still deriving sips × cubes × pes"
)
# ── D9: get_rank greenlet-local + _bind_rank ──────────────────────────
def test_get_rank_is_greenlet_local(tmp_path, monkeypatch, spec):
"""Each greenlet sees its own rank via dist.get_rank().
Framework-level launcher binds greenlet → rank; get_rank() resolves
the current greenlet and returns that rank.
"""
monkeypatch.chdir(_write_minimal_ccl_yaml(tmp_path))
ctx = _FakeCtx(spec=spec)
dc = DistributedContext()
dc._ctx_ref = ctx
dc.init_process_group(backend="ahbm")
assert dc.get_world_size() == int(spec["system"]["sips"]["count"])
assert hasattr(dc, "_bind_rank"), (
"DistributedContext must expose _bind_rank(g, rank) hook"
)
seen: dict[int, int] = {}
def _probe(rank: int) -> None:
seen[rank] = dc.get_rank()
g0 = greenlet(lambda: _probe(0))
g1 = greenlet(lambda: _probe(1))
dc._bind_rank(g0, 0)
dc._bind_rank(g1, 1)
g0.switch()
g1.switch()
assert seen == {0: 0, 1: 1}, (
f"expected each greenlet to see its own rank; got {seen}"
)
def test_get_rank_fallback_without_bind(tmp_path, monkeypatch, spec):
"""Unbound greenlet falls back to rank 0 (single-driver compat)."""
monkeypatch.chdir(_write_minimal_ccl_yaml(tmp_path))
ctx = _FakeCtx(spec=spec)
dc = DistributedContext()
dc._ctx_ref = ctx
dc.init_process_group(backend="ahbm")
# Call from main (unbound) greenlet
assert dc.get_rank() == 0
# ── D10/D11: torch.ahbm.set_device + tensor scoping ───────────────────
def test_ahbm_set_device_binds_tensor_to_single_sip(topology):
"""``torch.ahbm.set_device(rank)`` + default-sip DPPolicy → tensor on SIP rank.
After set_device(1), a tensor with DPPolicy leaving the SIP dimension
at its default must be placed entirely on SIP 1.
"""
from kernbench.policy.placement.dp import DPPolicy
from kernbench.runtime_api.context import RuntimeContext
from kernbench.runtime_api.types import DeviceSelector
from kernbench.sim_engine.engine import GraphEngine
engine = GraphEngine(topology.topology_obj, enable_data=True)
with RuntimeContext(
engine=engine,
target_device=DeviceSelector("all"),
correlation_id="test_ahbm_set_device",
spec=topology.topology_obj.spec,
) as ctx:
assert hasattr(ctx, "ahbm"), (
"RuntimeContext must expose .ahbm namespace (ADR-0024 D10)"
)
ctx.ahbm.set_device(1)
dp = DPPolicy(cube="column_wise", pe="column_wise") # default sip
tensor = ctx.zeros((1, 128), dtype="f16", dp=dp, name="probe")
shard_sips = {s.sip for s in tensor._handle.shards}
assert shard_sips == {1}, (
f"after ahbm.set_device(1), all shards should live on SIP 1; "
f"got sips={sorted(shard_sips)}"
)
def test_accelerator_alias_mirrors_ahbm(topology):
"""torch.accelerator.set_device_index(r) is an alias for ahbm.set_device(r)
(PyTorch 2.x device-agnostic surface)."""
from kernbench.runtime_api.context import RuntimeContext
from kernbench.runtime_api.types import DeviceSelector
from kernbench.sim_engine.engine import GraphEngine
engine = GraphEngine(topology.topology_obj, enable_data=True)
with RuntimeContext(
engine=engine,
target_device=DeviceSelector("all"),
correlation_id="test_accelerator_alias",
spec=topology.topology_obj.spec,
) as ctx:
assert hasattr(ctx, "accelerator"), (
"RuntimeContext must expose .accelerator namespace (ADR-0024 D10)"
)
ctx.accelerator.set_device_index(1)
# Both namespaces should report SIP 1 as current device
assert ctx.ahbm.current_device() == 1
assert ctx.accelerator.current_device_index() == 1
# ── D12/D13: run() spawns one worker per rank ─────────────────────────
def test_run_spawns_one_worker_per_rank(tmp_path, monkeypatch, spec):
"""The bench's ``run()`` invokes ``worker`` once per rank.
With world_size = SIP count = 2 (topology), worker must be called
exactly twice with ranks 0 and 1. Each call sees world_size=2.
"""
project_root = os.path.abspath(
os.path.join(os.path.dirname(__file__), "..")
)
monkeypatch.chdir(_write_minimal_ccl_yaml(tmp_path))
import benches.ccl_allreduce as bench
calls: list[tuple[int, int]] = []
def _fake_worker(rank, cfg, torch) -> None:
calls.append((rank, cfg.world_size))
monkeypatch.setattr(bench, "_worker", _fake_worker)
from kernbench.runtime_api.context import RuntimeContext
from kernbench.runtime_api.types import DeviceSelector
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import resolve_topology
topo = resolve_topology(os.path.join(project_root, "topology.yaml"))
engine = GraphEngine(topo.topology_obj, enable_data=True)
with RuntimeContext(
engine=engine,
target_device=DeviceSelector("all"),
correlation_id="test_run_spawns",
spec=topo.topology_obj.spec,
) as ctx:
bench.run(ctx)
ranks = sorted(r for r, _ in calls)
ws_values = {ws for _, ws in calls}
expected_ws = int(spec["system"]["sips"]["count"])
assert ranks == list(range(expected_ws)), (
f"run() should invoke worker for ranks 0..{expected_ws - 1}; "
f"saw ranks={ranks}"
)
assert ws_values == {expected_ws}, (
f"each worker should see world_size={expected_ws}; saw {ws_values}"
)
-125
View File
@@ -1,125 +0,0 @@
"""Tests for IPCQ deadlock detection (ADR-0023 D14 F3)."""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any
import pytest
import simpy
from kernbench.ccl import diagnostics
from kernbench.common.ipcq_types import (
IpcqEndpoint,
IpcqInitEntry,
IpcqRecvCmd,
IpcqRequest,
)
from kernbench.components.builtin.pe_ipcq import PeIpcqComponent
from kernbench.runtime_api.kernel import IpcqInitMsg
from kernbench.topology.types import Node
@dataclass
class _FakeTxn:
request: Any
done: simpy.Event
result_data: dict[str, Any] = field(default_factory=dict)
def _make_isolated_pe_ipcq(env):
node = Node(
id="sip0.cube0.pe0.pe_ipcq", kind="pe_ipcq",
impl="builtin.pe_ipcq", attrs={}, pos_mm=None,
)
comp = PeIpcqComponent(node, ctx=None)
comp.in_ports["host"] = simpy.Store(env)
comp.out_ports["sip0.cube0.pe0.pe_dma"] = simpy.Store(env)
comp.start(env)
peer_credit = simpy.Store(env)
ep = IpcqEndpoint(
sip=0, cube=0, pe=1, buffer_kind="tcm",
rx_base_pa=0x10_000, rx_base_va=0,
n_slots=4, slot_size=4096,
)
init_msg = IpcqInitMsg(
correlation_id="t", request_id="t",
target_sips=(0,), target_cubes=(0,), target_pe=0,
entries=(IpcqInitEntry(
direction="W", peer=ep,
my_rx_base_pa=0x40_000, my_rx_base_va=0,
n_slots=4, slot_size=4096,
peer_credit_store=peer_credit,
),),
backpressure_mode="sleep",
buffer_kind="tcm",
credit_size_bytes=16,
)
done = env.event()
comp.in_ports["host"].put(_FakeTxn(request=init_msg, done=done))
env.run(until=done)
return comp
def test_pointer_dump_includes_blocked_state():
"""A blocked recv should still be visible in the pointer dump."""
env = simpy.Environment()
comp = _make_isolated_pe_ipcq(env)
# Issue a recv that will block (no data has arrived)
recv_cmd = IpcqRecvCmd(direction="W", shape=(8,), dtype="f16", handle_id="r1")
req = IpcqRequest(command=recv_cmd, done=env.event())
comp.in_ports["host"].put(req)
env.run(until=10)
assert not req.done.triggered
# Pointer dump should show my_tail=0 and peer_head_cache=0
# We need to use the engine API but for an isolated component, just call directly
class FakeEngine:
_components = {"sip0.cube0.pe0.pe_ipcq": comp}
dump = diagnostics.pointer_dump(FakeEngine())
assert "my_tail=0" in dump
assert "peer_head_cache=0" in dump
def test_deadlock_detection_recv_without_send():
"""A recv with no matching sender → SimPy schedule empties → engine
raises ``IpcqDeadlock`` with a pointer dump.
"""
from kernbench.ccl.diagnostics import IpcqDeadlock
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
def deadlock_kernel(t_ptr, n_elem, tl):
# Every PE just receives, no sends → no one delivers → deadlock
tl.recv(dir="W", shape=(n_elem,), dtype="f16")
topo = resolve_topology("topology.yaml")
def run(torch):
torch.install_ipcq(
algorithm="ring_allreduce_tcm", world_size_override=8,
)
a = torch.zeros(
(1, 8 * 8),
dtype="f16",
dp=DPPolicy(
cube="replicate", pe="column_wise",
num_cubes=1,
),
name="dl_in",
)
torch.launch("dl", deadlock_kernel, a, 8)
with pytest.raises(IpcqDeadlock):
run_bench(
topology=topo, bench_fn=run,
device=resolve_device("all"),
engine_factory=lambda t, d: GraphEngine(
getattr(t, "topology_obj", t), enable_data=True
),
)
-70
View File
@@ -1,70 +0,0 @@
"""Tests for CCL diagnostics: trace + pointer dump (ADR-0023 D14)."""
from __future__ import annotations
import os
from kernbench.ccl import diagnostics
# ── trace toggle ─────────────────────────────────────────────────────
def test_trace_disabled_by_default(monkeypatch):
monkeypatch.delenv("KERNBENCH_CCL_TRACE", raising=False)
diagnostics.reload_trace_setting()
assert diagnostics.trace_enabled() is False
def test_trace_enabled_via_env(monkeypatch):
monkeypatch.setenv("KERNBENCH_CCL_TRACE", "1")
diagnostics.reload_trace_setting()
assert diagnostics.trace_enabled() is True
def test_trace_record_send(monkeypatch, capsys):
monkeypatch.setenv("KERNBENCH_CCL_TRACE", "1")
diagnostics.reload_trace_setting()
diagnostics.log_send(t_ns=100.0, sender="sip0.cube0.pe0",
direction="E", nbytes=64, sender_seq=0)
out = capsys.readouterr().out
assert "send" in out
assert "sip0.cube0.pe0" in out
assert "dir=E" in out
monkeypatch.delenv("KERNBENCH_CCL_TRACE")
diagnostics.reload_trace_setting()
def test_trace_record_recv(monkeypatch, capsys):
monkeypatch.setenv("KERNBENCH_CCL_TRACE", "1")
diagnostics.reload_trace_setting()
diagnostics.log_recv(t_ns=200.0, receiver="sip0.cube0.pe1",
direction="W", nbytes=64)
out = capsys.readouterr().out
assert "recv" in out
assert "sip0.cube0.pe1" in out
monkeypatch.delenv("KERNBENCH_CCL_TRACE")
diagnostics.reload_trace_setting()
# ── pointer dump ────────────────────────────────────────────────────
def test_pointer_dump_format():
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import resolve_topology
from kernbench.ccl.install import (
install_ipcq, load_ccl_config, resolve_algorithm_config,
)
topo = resolve_topology("topology.yaml").topology_obj
engine = GraphEngine(topo, enable_data=True)
cfg = resolve_algorithm_config(load_ccl_config(), name="ring_allreduce_tcm")
install_ipcq(engine, topo.spec, cfg)
dump = diagnostics.pointer_dump(engine)
# 8 ranks × 2 directions = 16 lines (plus 8 PE headers)
assert "sip0.cube0.pe0" in dump
assert "E:" in dump
assert "W:" in dump
assert "my_head=" in dump
assert "peer_tail_cache=" in dump
-81
View File
@@ -1,81 +0,0 @@
"""Validate the hello-world example from docs/ccl-author-guide.md.
This is the simplest possible CCL kernel — each PE sends its tile E
and receives a tile from W. After running, each rank's slice should
contain the data of the previous rank.
"""
from __future__ import annotations
import numpy as np
from kernbench.ccl.algorithms import hello_send
from kernbench.ccl.testing import run_kernel_in_mock
def test_hello_send_4_ranks_mock():
n_elem = 8
inputs = [np.full((n_elem,), float(r + 1), dtype=np.float16) for r in range(4)]
outputs = run_kernel_in_mock(
kernel_fn=hello_send.kernel,
world_size=4,
topology="ring_1d",
inputs=inputs,
kernel_args=(n_elem,),
)
# rank r should have rank (r-1) % 4's data
for r in range(4):
prev = inputs[(r - 1) % 4]
assert np.array_equal(outputs[r], prev), f"rank {r}: got {outputs[r]}"
def test_hello_send_via_simpy_runner():
"""Same but through real SimPy + IPCQ."""
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
topo = resolve_topology("topology.yaml")
n_elem = 8
world_size = 8
def run(torch):
# World size for this hello test is 8 (one cube). ccl.yaml no
# longer carries a default world_size — pass it explicitly.
plan = torch.install_ipcq(
algorithm="ring_allreduce_tcm", world_size_override=world_size,
)
a = torch.zeros(
(1, world_size * n_elem), dtype="f16",
dp=DPPolicy(
cube="replicate", pe="column_wise",
num_cubes=1,
),
name="hello_in",
)
store = torch.engine.memory_store
base = a._handle.va_base or a._handle.shards[0].pa
nbytes = n_elem * 2
for r in range(world_size):
store.write("hbm", base + r * nbytes,
np.full((n_elem,), float(r + 1), dtype=np.float16))
torch.launch("hello_send", hello_send.kernel, a, n_elem)
# Each rank should hold the previous rank's data after the round
for r in range(world_size):
arr = store.read("hbm", base + r * nbytes, shape=(n_elem,), dtype="f16")
prev_value = float(((r - 1) % world_size) + 1)
assert np.allclose(arr, prev_value), f"rank {r}: got {arr}, expected {prev_value}"
result = run_bench(
topology=topo, bench_fn=run,
device=resolve_device("all"),
engine_factory=lambda t, d: GraphEngine(
getattr(t, "topology_obj", t), enable_data=True
),
)
assert result.completion.ok
-117
View File
@@ -2,7 +2,6 @@
from __future__ import annotations
from kernbench.ccl.install import (
install_ipcq,
linear_rank_to_pe,
load_ccl_config,
resolve_algorithm_config,
@@ -26,28 +25,14 @@ def test_resolve_algorithm_config_default():
cfg = load_ccl_config()
merged = resolve_algorithm_config(cfg)
assert merged["algorithm"] == cfg["defaults"]["algorithm"]
# ccl.yaml no longer carries defaults.world_size — backend derives
# it from topology.yaml at install time. Just check the field is
# absent here (verified per-test where install_ipcq is called).
assert "world_size" not in merged or merged["world_size"] >= 1
def test_resolve_algorithm_config_override():
cfg = load_ccl_config()
merged = resolve_algorithm_config(cfg, name="ring_allreduce_hbm")
assert merged["algorithm"] == "ring_allreduce_hbm"
assert merged["buffer_kind"] == "hbm" # algo override
# defaults still apply
assert merged["n_slots"] == cfg["defaults"]["n_slots"]
def test_linear_rank_to_pe():
engine, topo = _engine()
spec = topo.spec
# Cube 0 of SIP 0
assert linear_rank_to_pe(0, spec) == (0, 0, 0)
assert linear_rank_to_pe(7, spec) == (0, 0, 7)
# Should not exceed total PE count
pes_per_sip = (
spec["sip"]["cube_mesh"]["w"] * spec["sip"]["cube_mesh"]["h"]
* spec["cube"]["pe_layout"]["pe_per_corner"]
@@ -56,105 +41,3 @@ def test_linear_rank_to_pe():
sips = spec["system"]["sips"]["count"]
total = sips * pes_per_sip
assert total >= 8
def test_install_ipcq_neighbors_correct():
engine, topo = _engine()
cfg = load_ccl_config()
merged = resolve_algorithm_config(cfg, name="ring_allreduce_tcm")
# Force a single-cube 8-rank install for the assertions below.
merged["world_size"] = 8
plan = install_ipcq(engine, topo.spec, merged)
assert plan["world_size"] == 8
assert plan["buffer_kind"] == "tcm"
# Each rank should have E and W entries
for r, nbrs in plan["neighbor_table"].items():
assert "E" in nbrs
assert "W" in nbrs
# Inspect installed PE_IPCQ for rank 0
ipcq = engine._components["sip0.cube0.pe0.pe_ipcq"]
qp_e = ipcq.queue_pairs["E"]
qp_w = ipcq.queue_pairs["W"]
assert qp_e["peer"].pe == 1 # rank 0's E neighbor is rank 1
assert qp_w["peer"].pe == 7 # rank 0's W neighbor is rank 7
# rx_base addresses should be unique
assert qp_e["my_rx_base_pa"] != qp_w["my_rx_base_pa"]
def test_install_ipcq_credit_stores_wired():
engine, topo = _engine()
cfg = load_ccl_config()
merged = resolve_algorithm_config(cfg, name="ring_allreduce_tcm")
merged["world_size"] = 8
install_ipcq(engine, topo.spec, merged)
# rank 0 (pe0) sending E goes to rank 1 (pe1)
# rank 0's peer_credit_store on E direction should equal rank 1's credit_inbox
pe0 = engine._components["sip0.cube0.pe0.pe_ipcq"]
pe1 = engine._components["sip0.cube0.pe1.pe_ipcq"]
qp_e = pe0.queue_pairs["E"]
assert qp_e["peer_credit_store"] is pe1.credit_inbox
# ── ADR-0025 D1: reverse_direction opposite-preference ───────────────
def test_reverse_direction_opposite_preference_2rank_ring():
"""ADR-0025 D1: In a 2-rank bidirectional ring both E and W point to the
same peer; reverse_direction must pick the OPPOSITE direction (W for E,
E for W) so rx_base targets the semantically-correct slot.
Concretely: rank 0 sending via E to rank 1 must target rank 1's W-rx
buffer (not rank 1's E-rx), because rank 1's kernel recv(W) reads from
its W-rx.
"""
engine, topo = _engine()
cfg = load_ccl_config()
merged = resolve_algorithm_config(cfg, name="ring_allreduce_tcm")
merged["world_size"] = 2
install_ipcq(engine, topo.spec, merged)
ipcq0 = engine._components["sip0.cube0.pe0.pe_ipcq"]
ipcq1 = engine._components["sip0.cube0.pe1.pe_ipcq"]
rank1_e_rx = ipcq1.queue_pairs["E"]["my_rx_base_pa"]
rank1_w_rx = ipcq1.queue_pairs["W"]["my_rx_base_pa"]
qp0_e = ipcq0.queue_pairs["E"]
qp0_w = ipcq0.queue_pairs["W"]
# rank 0's E entry should target rank 1's W-rx (opposite), NOT rank 1's E-rx.
assert qp0_e["peer"].rx_base_pa == rank1_w_rx, (
f"expected rank 0's E peer.rx_base_pa == rank 1's W-rx ({rank1_w_rx:#x}), "
f"got {qp0_e['peer'].rx_base_pa:#x} (matches E-rx: {rank1_e_rx:#x}) — "
f"reverse_direction picked same-label instead of opposite"
)
# rank 0's W entry should target rank 1's E-rx (opposite).
assert qp0_w["peer"].rx_base_pa == rank1_e_rx
def test_reverse_direction_opposite_preference_4rank_ring_sanity():
"""ADR-0025 D1 sanity: ws>=3 ring. E and W have distinct peers, so
opposite-preference produces same result as old dict-order first-match.
This test should PASS both under current and post-fix code.
"""
engine, topo = _engine()
cfg = load_ccl_config()
merged = resolve_algorithm_config(cfg, name="ring_allreduce_tcm")
merged["world_size"] = 4
install_ipcq(engine, topo.spec, merged)
ipcq0 = engine._components["sip0.cube0.pe0.pe_ipcq"]
ipcq1 = engine._components["sip0.cube0.pe1.pe_ipcq"]
ipcq3 = engine._components["sip0.cube0.pe3.pe_ipcq"]
# rank 0 E → rank 1 → rank 1's W-rx
qp0_e = ipcq0.queue_pairs["E"]
assert qp0_e["peer"].rx_base_pa == ipcq1.queue_pairs["W"]["my_rx_base_pa"]
# rank 0 W → rank 3 (last in ring) → rank 3's E-rx
qp0_w = ipcq0.queue_pairs["W"]
assert qp0_w["peer"].rx_base_pa == ipcq3.queue_pairs["E"]["my_rx_base_pa"]
-83
View File
@@ -1,83 +0,0 @@
"""Tests for the mock CCL runtime (ADR-0023 D15)."""
from __future__ import annotations
import numpy as np
from kernbench.ccl.algorithms import ring_allreduce
from kernbench.ccl.testing import run_kernel_in_mock
def test_ring_allreduce_4_ranks():
"""Run the ring all-reduce kernel under the mock runtime, no SimPy."""
n_elem = 8
inputs = [
np.full((n_elem,), float(r + 1), dtype=np.float16)
for r in range(4)
]
expected = sum(inputs) # [10, 10, ..., 10]
outputs = run_kernel_in_mock(
kernel_fn=ring_allreduce.kernel,
world_size=4,
topology="ring_1d",
inputs=inputs,
kernel_args=(n_elem, 4),
)
assert len(outputs) == 4
for r in range(4):
assert np.allclose(outputs[r], expected)
def test_ring_allreduce_8_ranks():
n_elem = 16
inputs = [
np.full((n_elem,), float(r + 1), dtype=np.float16)
for r in range(8)
]
expected = sum(inputs) # [36, 36, ...]
outputs = run_kernel_in_mock(
kernel_fn=ring_allreduce.kernel,
world_size=8,
topology="ring_1d",
inputs=inputs,
kernel_args=(n_elem, 8),
)
for r in range(8):
assert np.allclose(outputs[r], expected)
def test_ring_allreduce_random_data():
n_elem = 32
rng = np.random.default_rng(42)
inputs = [rng.standard_normal(n_elem).astype(np.float16) for _ in range(4)]
expected = sum(inputs)
outputs = run_kernel_in_mock(
kernel_fn=ring_allreduce.kernel,
world_size=4,
topology="ring_1d",
inputs=inputs,
kernel_args=(n_elem, 4),
)
for r in range(4):
assert np.allclose(outputs[r], expected, rtol=1e-2, atol=1e-2)
def test_mock_runtime_invalid_direction_raises():
"""A kernel that uses an unsupported direction should raise."""
import pytest
def bad_kernel(t_ptr, n_elem, tl):
tl.send(dir="N", src_addr=0, nbytes=2, shape=(1,), dtype="f16", space="hbm")
inputs = [np.array([1.0], dtype=np.float16) for _ in range(2)]
with pytest.raises(Exception):
run_kernel_in_mock(
kernel_fn=bad_kernel,
world_size=2,
topology="ring_1d",
inputs=inputs,
kernel_args=(1,),
)
-85
View File
@@ -1,85 +0,0 @@
"""CCL performance validation tests (ADR-0023 D13 T5).
Sanity-checks the simulated latency of the unified ``ccl_allreduce`` bench
under the rank = SIP TP launcher model (ADR-0024 / ADR-0027). Uses the
topology-derived world_size (= 2 in the shipped topology); the latency
model is topology-aware, so buffer_kind differences remain visible even
at this scale.
"""
from __future__ import annotations
import importlib
import os
import pytest
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
def _engine_factory(topology, device):
return GraphEngine(getattr(topology, "topology_obj", topology), enable_data=True)
def _run_ring(algorithm: str, buffer_kind: str = "tcm") -> float:
"""Run a rank = SIP ring all-reduce via the unified bench with a tmp
ccl.yaml overlay. Returns simulated kernel total_ns."""
import tempfile
body = f"""\
defaults:
algorithm: {algorithm}
buffer_kind: {buffer_kind}
backpressure: sleep
n_slots: 4
slot_size: 4096
vc_chunk_size: 256
ipcq_credit_size_bytes: 16
algorithms:
{algorithm}:
module: kernbench.ccl.algorithms.ring_allreduce
topology: ring_1d
buffer_kind: {buffer_kind}
n_elem: 32
"""
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
with tempfile.TemporaryDirectory() as tmp:
with open(os.path.join(tmp, "ccl.yaml"), "w") as f:
f.write(body)
old_cwd = os.getcwd()
os.chdir(tmp)
try:
topo = resolve_topology(os.path.join(project_root, "topology.yaml"))
bench_mod = importlib.import_module("benches.ccl_allreduce")
result = run_bench(
topology=topo, bench_fn=bench_mod.run,
device=resolve_device("all"),
engine_factory=_engine_factory,
)
finally:
os.chdir(old_cwd)
assert result.completion.ok, f"{algorithm} did not complete"
last_kernel = None
for tr in (result.traces or []):
if tr.get("phase") == "kernel":
last_kernel = tr
assert last_kernel is not None, f"{algorithm} produced no kernel trace"
return float(last_kernel.get("total_ns", 0.0))
@pytest.mark.parametrize("buffer_kind", ["tcm", "hbm", "sram"])
def test_ccl_latency_positive(buffer_kind):
"""Every buffer kind must produce a positive simulated latency."""
algo = f"ring_allreduce_{buffer_kind}"
ns = _run_ring(algo, buffer_kind)
assert ns > 0
def test_ccl_latency_under_reasonable_bound():
"""rank = SIP ring all-reduce (tile=32 f16) should finish well under 1ms."""
ns = _run_ring("ring_allreduce_tcm", "tcm")
assert ns < 1_000_000 # < 1 ms simulated
@@ -0,0 +1,119 @@
"""End-to-end distributed test for intercube allreduce.
Exercises the full process-group path:
dist.init_process_group(backend="ahbm")
→ mp.spawn(nprocs=n_sips)
→ each worker: set_device → allocate → fill → dist.all_reduce → verify
This is the same flow a real DDP training script would use.
"""
from __future__ import annotations
import os
import textwrap
from pathlib import Path
import numpy as np
import pytest
TOPOLOGY_PATH = Path(__file__).parent.parent / "topology.yaml"
N_CUBES = 16
N_ELEM = 8
def _write_ccl_yaml(tmp_path) -> str:
body = textwrap.dedent("""\
defaults:
algorithm: intercube_allreduce
buffer_kind: tcm
backpressure: sleep
n_slots: 4
slot_size: 4096
vc_chunk_size: 256
ipcq_credit_size_bytes: 16
algorithms:
intercube_allreduce:
module: kernbench.ccl.algorithms.intercube_allreduce
topology: none
buffer_kind: tcm
n_elem: 8
root_cube: 15
""")
(tmp_path / "ccl.yaml").write_text(body)
return str(tmp_path)
def _worker(rank: int, n_sips: int, torch) -> None:
"""Per-SIP worker: allocate, fill, all_reduce, verify."""
from kernbench.policy.placement.dp import DPPolicy
torch.ahbm.set_device(rank)
dp = DPPolicy(
cube="row_wise", pe="replicate",
num_pes=1, num_cubes=N_CUBES,
)
tensor = torch.zeros(
(N_CUBES, N_ELEM), dtype="f16", dp=dp,
name=f"sip{rank}",
)
init_arr = np.full((N_CUBES, N_ELEM), float(rank + 1), dtype=np.float16)
tensor.copy_(torch.from_numpy(init_arr))
print(f"[SIP {rank}] input cube0[:4] = {tensor.numpy()[0][:4].tolist()}")
torch.distributed.all_reduce(tensor, op="sum")
arr = tensor.numpy()
expected = float(N_CUBES * sum(range(1, n_sips + 1)))
print(f"[SIP {rank}] output cube0[:4] = {arr[0][:4].tolist()}")
print(f"[SIP {rank}] output cube15[:4] = {arr[15][:4].tolist()}")
for cube_id in range(N_CUBES):
assert np.allclose(arr[cube_id], expected, rtol=1e-1, atol=1e-1), (
f"SIP{rank} cube {cube_id}: "
f"got {arr[cube_id][:4]}, expected {expected}"
)
if rank == 0:
print(f"\n intercube_allreduce (ws={n_sips}): "
f"{n_sips * N_CUBES} OK")
def test_distributed_intercube_allreduce(tmp_path, monkeypatch):
"""Full distributed path: init_process_group → mp.spawn → all_reduce."""
from kernbench.runtime_api.context import RuntimeContext
from kernbench.runtime_api.types import DeviceSelector
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import resolve_topology
monkeypatch.chdir(_write_ccl_yaml(tmp_path))
topo = resolve_topology(str(TOPOLOGY_PATH))
engine = GraphEngine(topo.topology_obj, enable_data=True)
spec = topo.topology_obj.spec
n_sips = int(spec["system"]["sips"]["count"])
with RuntimeContext(
engine=engine,
target_device=DeviceSelector("all"),
correlation_id="dist_intercube_ar",
spec=spec,
) as ctx:
ctx.distributed.init_process_group(backend="ahbm")
assert ctx.distributed.get_world_size() == n_sips
t_start = engine._env.now
ctx.multiprocessing.spawn(
_worker, args=(n_sips, ctx), nprocs=n_sips,
)
t_end = engine._env.now
print(f"\n[distributed] sim latency = "
f"{t_end - t_start:.1f} ns ({(t_end - t_start) / 1000:.3f} us)")
-270
View File
@@ -1,270 +0,0 @@
"""ADR-0027 T5: Host-read barrier (D0.5).
Phase 1: Tensor.numpy / data / __getitem__ / __repr__ / copy_ currently
perform MemoryStore operations without barrier logic → tests fail when
they assert drain is triggered. Phase 2 injects the barrier.
"""
from __future__ import annotations
import numpy as np
import pytest
from greenlet import greenlet
def _make_ctx(topology):
from kernbench.runtime_api.context import RuntimeContext
from kernbench.runtime_api.types import DeviceSelector
from kernbench.sim_engine.engine import GraphEngine
engine = GraphEngine(topology.topology_obj, enable_data=True)
return RuntimeContext(
engine=engine,
target_device=DeviceSelector("all"),
correlation_id="test_t5",
spec=topology.topology_obj.spec,
)
# ── T5.g: closed-set registry exists ─────────────────────────────────
def test_host_read_barrier_registry_exists():
"""D0.5 T5.g: Tensor module exposes the closed-set registry."""
from kernbench.runtime_api import tensor as tensor_mod
assert hasattr(tensor_mod, "_HOST_READ_BARRIERS"), (
"ADR-0027 T5.g: tensor module must declare _HOST_READ_BARRIERS registry"
)
registry = tensor_mod._HOST_READ_BARRIERS
assert isinstance(registry, frozenset)
expected = {"numpy", "data", "__getitem__", "__repr__", "copy_"}
assert expected.issubset(registry), (
f"registry must include {expected}; got {registry}"
)
# ── T5.a: numpy() triggers drain when pending non-empty ──────────────
def test_numpy_triggers_drain_when_pending(topology):
"""T5.a: launch → numpy() → barrier drains before read (worker context)."""
with _make_ctx(topology) as ctx:
from kernbench.policy.placement.dp import DPPolicy
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
observed: dict = {"pre_numpy_pending": None, "post_numpy_pending": None}
def _worker():
t = ctx.zeros((1, 8), dtype="f16", dp=dp, name="t5a_t")
src = np.full((1, 8), 1.5, dtype=np.float16)
t.copy_(ctx.distributed._ctx_ref.from_numpy(src) if False else _hold(ctx, src))
# Manually push a dummy handle to simulate pending state; in real
# D0.5, numpy will detect and drain.
observed["pre_numpy_pending"] = list(ctx._pending_worker_waits)
_ = t.numpy()
observed["post_numpy_pending"] = list(ctx._pending_worker_waits)
# Can't actually manufacture pending + test numpy inside worker
# without D0.5 implemented — instead, verify the barrier path is
# invoked by spying.
from kernbench.runtime_api.tensor import Tensor
barrier_calls = {"n": 0}
original_numpy = Tensor.numpy
def _spy_numpy(self):
# After D0.5 is implemented, this wrapper is redundant; the
# test just checks numpy was called at all after a pending
# operation.
barrier_calls["n"] += 1
return original_numpy(self)
Tensor.numpy = _spy_numpy # type: ignore[assignment]
try:
ctx.multiprocessing.spawn(_mk_worker_numpy, args=(ctx,), nprocs=1)
finally:
Tensor.numpy = original_numpy # type: ignore[assignment]
assert barrier_calls["n"] >= 1
def _hold(ctx, arr):
"""helper (unused branch)."""
import numpy as _np
t = type("X", (), {})()
t.numpy = lambda self=None: arr
return t
def _mk_worker_numpy(rank, ctx):
"""Worker that calls numpy after a tensor deploy. Triggers barrier."""
from kernbench.policy.placement.dp import DPPolicy
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
t = ctx.zeros((1, 8), dtype="f16", dp=dp, name=f"t5_r{rank}")
_ = t.numpy()
# ── T5.b: metadata access does NOT drain ─────────────────────────────
def test_metadata_access_is_non_barrier(topology):
"""T5.b: .shape / .dtype / .name do NOT trigger drain."""
with _make_ctx(topology) as ctx:
from kernbench.runtime_api import tensor as tensor_mod
from kernbench.policy.placement.dp import DPPolicy
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
t = ctx.zeros((1, 8), dtype="f16", dp=dp, name="t5b")
# Populate pending queue artificially (simulate worker state).
ctx._pending_worker_waits.append("fake_handle_that_must_not_drain")
_ = t.shape
_ = t.dtype
_ = t.name
assert "fake_handle_that_must_not_drain" in ctx._pending_worker_waits, (
"T5.b: metadata accessors must not drain pending queue"
)
ctx._pending_worker_waits.clear()
# ── T5.c: empty pending → numpy is fast-path (no yield) ──────────────
def test_numpy_fast_path_when_pending_empty(topology):
"""T5.c: numpy() with empty pending queue does not yield to main."""
with _make_ctx(topology) as ctx:
from kernbench.policy.placement.dp import DPPolicy
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
def _worker(rank: int):
t = ctx.zeros((1, 4), dtype="f16", dp=dp, name=f"t5c_r{rank}")
# At this point, after worker's own wait(s), pending should be empty.
assert ctx._pending_worker_waits == [], (
"after worker's deploy, pending queue should be drained"
)
# numpy call should be fast-path (no yield).
_ = t.numpy()
ctx.multiprocessing.spawn(_worker, args=(), nprocs=1)
# ── T5.d: __getitem__ / data also barriers ───────────────────────────
def test_getitem_and_data_are_barriers(topology):
"""T5.d: __getitem__ and .data property behave like numpy() barrier."""
with _make_ctx(topology) as ctx:
from kernbench.policy.placement.dp import DPPolicy
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
def _worker(rank: int):
t = ctx.zeros((1, 8), dtype="f16", dp=dp, name=f"t5d_r{rank}")
# host src copied in (forces write path)
src = np.full((1, 8), float(rank + 1), dtype=np.float16)
from kernbench.runtime_api.tensor import Tensor
h = Tensor(shape=src.shape, dtype="f16", name="host")
h._host_buffer = src
t.copy_(h)
# Read access via __getitem__ and .data: both must fully materialize.
slice_val = t[0, 0:4]
data_val = t.data
assert slice_val.shape[0] == 4
assert data_val.shape == (1, 8)
ctx.multiprocessing.spawn(_worker, args=(), nprocs=2)
# ── T5.e: collective pending also drained by barrier ────────────────
def test_numpy_drains_collective_pending(topology, tmp_path, monkeypatch):
"""T5.e: numpy() after all_reduce must see post-reduce data.
Note: in the current model, ``all_reduce`` itself yields to main so the
collective is drained before the worker resumes; barriers at
``numpy()`` intentionally do NOT drain collective pending (would cause
cross-rank deadlock — see ``_host_read_barrier`` docstring). What this
test asserts is the observable contract: post-``all_reduce`` +
``numpy()`` sees the reduced values.
"""
import textwrap
body = textwrap.dedent("""\
defaults:
algorithm: ring_allreduce_tcm
buffer_kind: tcm
backpressure: sleep
n_slots: 4
slot_size: 4096
vc_chunk_size: 256
ipcq_credit_size_bytes: 16
algorithms:
ring_allreduce_tcm:
module: kernbench.ccl.algorithms.ring_allreduce
topology: ring_1d
buffer_kind: tcm
n_elem: 8
""")
(tmp_path / "ccl.yaml").write_text(body)
monkeypatch.chdir(str(tmp_path))
with _make_ctx(topology) as ctx:
from kernbench.policy.placement.dp import DPPolicy
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
def _worker(rank: int, ws: int):
ctx.ahbm.set_device(rank)
t = ctx.zeros((1, 8), dtype="f16", dp=dp, name=f"t5e_r{rank}")
src = np.full((1, 8), float(rank + 1), dtype=np.float16)
from kernbench.runtime_api.tensor import Tensor
h = Tensor(shape=src.shape, dtype="f16", name="host")
h._host_buffer = src
t.copy_(h)
ctx.distributed.all_reduce(t, op="sum")
# numpy() must see the reduced values even without explicit wait.
out = t.numpy()
expected = float(sum(range(1, ws + 1)))
# Tolerance loose for fp16 accumulation.
assert np.allclose(out, expected, rtol=1e-1, atol=1e-1), (
f"rank {rank}: expected {expected}, got {out}"
)
ctx.distributed.init_process_group(backend="ahbm")
ws = ctx.distributed.get_world_size()
ctx.multiprocessing.spawn(_worker, args=(ws,), nprocs=ws)
# ── T5.f: copy_ target-side write barrier ────────────────────────────
def test_copy_from_deployed_source_drains_source(topology):
"""T5.f (revised): ``copy_(source)`` drains source-side pending via the
``source.numpy()`` read barrier.
Note: the ADR originally specified a target-side write barrier as well,
but that was removed because global-pending target barrier can cause
cross-rank deadlock when another rank has a pending collective. Source-
side read barrier is preserved and sufficient for the common pattern
``target.copy_(deployed_source)``.
"""
with _make_ctx(topology) as ctx:
from kernbench.policy.placement.dp import DPPolicy
from kernbench.runtime_api.tensor import Tensor
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
def _worker(rank: int):
# Deployed source — its .numpy() will trigger the read barrier.
source = ctx.zeros((1, 8), dtype="f16", dp=dp, name=f"src_r{rank}")
target = ctx.zeros((1, 8), dtype="f16", dp=dp, name=f"tgt_r{rank}")
target.copy_(source)
# Smoke: no hang, no exception. numpy round-trip sees zeros.
out = target.numpy()
assert out.shape == (1, 8)
ctx.multiprocessing.spawn(_worker, args=(), nprocs=1)
+113
View File
@@ -0,0 +1,113 @@
"""Tests for configure_sfr_intercube_multisip neighbor table wiring.
Verifies that IPCQ neighbor tables are correctly installed for
intercube (pe0, 4×4 mesh N/S/E/W) + inter-SIP (pe0, all cubes,
global_E/global_W) communication.
"""
from __future__ import annotations
from pathlib import Path
from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
from kernbench.ccl.sfr_config import configure_sfr_intercube_multisip
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import resolve_topology
TOPOLOGY_PATH = Path(__file__).parent.parent / "topology.yaml"
N_CUBES = 16
def _engine_and_spec():
topo = resolve_topology(str(TOPOLOGY_PATH))
engine = GraphEngine(topo.topology_obj, enable_data=True)
return engine, topo.topology_obj.spec
def _merged_cfg():
cfg = load_ccl_config()
return resolve_algorithm_config(cfg, name="intercube_allreduce")
class TestConfigureSfrNeighborTables:
def test_world_size_and_rank_to_pe(self):
engine, spec = _engine_and_spec()
cfg = _merged_cfg()
plan = configure_sfr_intercube_multisip(engine, spec, cfg)
n_sips = int(spec["system"]["sips"]["count"])
assert plan["world_size"] == n_sips * N_CUBES
assert len(plan["rank_to_pe"]) == n_sips * N_CUBES
for pe_idx, (sip, cube, pe) in enumerate(plan["rank_to_pe"]):
assert pe == 0, f"pe_idx {pe_idx}: pe must be 0, got {pe}"
def test_corner_cube0_has_E_and_S_only(self):
"""Cube 0 (row=0, col=0) is NW corner: only E and S neighbors."""
engine, spec = _engine_and_spec()
cfg = _merged_cfg()
configure_sfr_intercube_multisip(engine, spec, cfg)
ipcq = engine._components["sip0.cube0.pe0.pe_ipcq"]
qp = ipcq.queue_pairs
assert "E" in qp, "cube 0 must have E neighbor"
assert "S" in qp, "cube 0 must have S neighbor"
assert "W" not in qp, "cube 0 (col=0) must NOT have W neighbor"
assert "N" not in qp, "cube 0 (row=0) must NOT have N neighbor"
assert qp["E"]["peer"].cube == 1
assert qp["S"]["peer"].cube == 4
def test_interior_cube5_has_all_four(self):
"""Cube 5 (row=1, col=1) is interior: N/S/E/W all present."""
engine, spec = _engine_and_spec()
cfg = _merged_cfg()
configure_sfr_intercube_multisip(engine, spec, cfg)
ipcq = engine._components["sip0.cube5.pe0.pe_ipcq"]
qp = ipcq.queue_pairs
assert qp["N"]["peer"].cube == 1
assert qp["S"]["peer"].cube == 9
assert qp["E"]["peer"].cube == 6
assert qp["W"]["peer"].cube == 4
def test_root_cube15_has_inter_sip(self):
"""Cube 15 (root, SE corner) has N, W + global_E/global_W."""
engine, spec = _engine_and_spec()
cfg = _merged_cfg()
configure_sfr_intercube_multisip(engine, spec, cfg)
ipcq0 = engine._components["sip0.cube15.pe0.pe_ipcq"]
qp0 = ipcq0.queue_pairs
assert "N" in qp0
assert "W" in qp0
assert "E" not in qp0, "cube 15 (col=3) must NOT have E"
assert "S" not in qp0, "cube 15 (row=3) must NOT have S"
assert "global_E" in qp0, "root cube must have global_E"
assert "global_W" in qp0, "root cube must have global_W"
assert qp0["global_E"]["peer"].sip == 1
assert qp0["global_E"]["peer"].cube == 15
ipcq1 = engine._components["sip1.cube15.pe0.pe_ipcq"]
qp1 = ipcq1.queue_pairs
assert qp1["global_E"]["peer"].sip == 0
assert qp1["global_E"]["peer"].cube == 15
def test_all_cubes_have_inter_sip(self):
"""ALL cubes (not just root) are wired for inter-SIP."""
engine, spec = _engine_and_spec()
cfg = _merged_cfg()
configure_sfr_intercube_multisip(engine, spec, cfg)
root_cube = int(cfg.get("root_cube", N_CUBES - 1))
for cube_id in range(N_CUBES):
ipcq = engine._components[f"sip0.cube{cube_id}.pe0.pe_ipcq"]
qp = ipcq.queue_pairs
assert "global_E" in qp, (
f"sip0.cube{cube_id}.pe0 missing global_E"
)
assert "global_W" in qp, (
f"sip0.cube{cube_id}.pe0 missing global_W"
)
if cube_id == root_cube:
assert qp["global_E"]["peer"].sip != 0, (
f"root cube {root_cube} global_E must point to another SIP"
)
-178
View File
@@ -1,178 +0,0 @@
"""ADR-0027 T4: torch.multiprocessing.spawn semantics.
Phase 1: imports `ctx.multiprocessing.spawn` which doesn't exist yet —
tests fail. Phase 2 (D1) lands the namespace + _MultiprocessingNamespace
+ SpawnException, and these pass.
"""
from __future__ import annotations
import os
import textwrap
import pytest
from greenlet import greenlet
def _write_minimal_ccl_yaml(tmp_path) -> str:
body = textwrap.dedent("""\
defaults:
algorithm: ring_allreduce_tcm
buffer_kind: tcm
backpressure: sleep
n_slots: 4
slot_size: 4096
vc_chunk_size: 256
ipcq_credit_size_bytes: 16
algorithms:
ring_allreduce_tcm:
module: kernbench.ccl.algorithms.ring_allreduce
topology: ring_1d
buffer_kind: tcm
n_elem: 8
""")
yaml_path = tmp_path / "ccl.yaml"
yaml_path.write_text(body)
return str(tmp_path)
def _make_ctx(topology):
from kernbench.runtime_api.context import RuntimeContext
from kernbench.runtime_api.types import DeviceSelector
from kernbench.sim_engine.engine import GraphEngine
engine = GraphEngine(topology.topology_obj, enable_data=True)
return RuntimeContext(
engine=engine,
target_device=DeviceSelector("all"),
correlation_id="test_t4",
spec=topology.topology_obj.spec,
)
# ── D1.3 namespace attach ────────────────────────────────────────────
def test_multiprocessing_namespace_attached(topology):
"""RuntimeContext.__post_init__ attaches ctx.multiprocessing (D1.3)."""
with _make_ctx(topology) as ctx:
assert hasattr(ctx, "multiprocessing"), (
"ADR-0027 D1.3: ctx.multiprocessing must exist"
)
assert hasattr(ctx.multiprocessing, "spawn"), (
"ctx.multiprocessing must expose a spawn(fn, args, nprocs) method"
)
# ── D1.1 / D1.2: spawn shape + rank binding ──────────────────────────
def test_spawn_invokes_fn_once_per_rank(topology):
"""spawn(fn, args, nprocs) calls fn(rank, *args) once for each rank."""
with _make_ctx(topology) as ctx:
calls: list[tuple[int, tuple]] = []
def _worker(rank: int, world_size: int) -> None:
calls.append((rank, (world_size,)))
ctx.multiprocessing.spawn(_worker, args=(3,), nprocs=3)
assert sorted(r for r, _ in calls) == [0, 1, 2]
for _, (ws,) in calls:
assert ws == 3
def test_spawn_binds_greenlet_local_rank(topology):
"""Inside the worker, torch.distributed.get_rank() returns the rank
bound to the greenlet (ADR-0024 D9 + D1.2)."""
with _make_ctx(topology) as ctx:
# Distributed context needs to be initialised so get_rank is valid.
# For T4 we don't run a real collective; just check rank lookup.
observed: list[tuple[int, int]] = []
def _worker(rank: int):
g = greenlet.getcurrent()
bound = ctx.distributed._rank_by_greenlet.get(g)
observed.append((rank, bound))
ctx.multiprocessing.spawn(_worker, args=(), nprocs=2)
for rank, bound in observed:
assert rank == bound, (
f"rank {rank} must be bound to greenlet-local rank {rank}; "
f"got {bound}"
)
# ── D1.2 exception cleanup ───────────────────────────────────────────
def test_spawn_exception_raises_spawn_exception_with_root_cause(topology):
"""D0.4-(4): worker raise → siblings SystemExit + SpawnException(errors)."""
with _make_ctx(topology) as ctx:
from kernbench.runtime_api.multiprocessing import SpawnException
def _worker(rank: int):
if rank == 1:
raise ValueError(f"rank {rank} boom")
with pytest.raises(SpawnException) as exc_info:
ctx.multiprocessing.spawn(_worker, args=(), nprocs=3)
# Root cause rank is captured.
assert 1 in exc_info.value.errors
assert isinstance(exc_info.value.errors[1], ValueError)
def test_spawn_exception_clears_pending_queues(topology):
"""D0.4-(4): on raise, _pending_worker_waits and collective queue clear."""
with _make_ctx(topology) as ctx:
from kernbench.runtime_api.multiprocessing import SpawnException
def _worker(rank: int):
raise RuntimeError("fail")
with pytest.raises(SpawnException):
ctx.multiprocessing.spawn(_worker, args=(), nprocs=2)
assert ctx._pending_worker_waits == []
# ── D1.4 migration compat: ccl_allreduce runs via mp.spawn ───────────
def test_ccl_allreduce_hand_rolled_loop_replaced_by_mp_spawn(
topology, tmp_path, monkeypatch, spec,
):
"""D1.4: benches/ccl_allreduce.py's hand-rolled greenlet loop must still
produce correct behaviour after migration to torch.multiprocessing.spawn.
Minimal smoke — just that ``bench.run(ctx)`` completes without the
loop short-circuiting or leaving pending queues dirty.
"""
monkeypatch.chdir(_write_minimal_ccl_yaml(tmp_path))
import benches.ccl_allreduce as bench
calls: list[tuple[int, int]] = []
def _fake_worker(rank, cfg, torch):
calls.append((rank, cfg.world_size))
monkeypatch.setattr(bench, "_worker", _fake_worker)
with _make_ctx(topology) as ctx:
bench.run(ctx)
expected_ws = int(spec["system"]["sips"]["count"])
ranks = sorted(r for r, _ in calls)
assert ranks == list(range(expected_ws))
assert ctx._pending_worker_waits == []
# ── _drain_pending function is exported ──────────────────────────────
def test_drain_pending_exported():
"""D0.4: _drain_pending must be importable from runtime_api.multiprocessing."""
from kernbench.runtime_api.multiprocessing import _drain_pending
assert callable(_drain_pending)
-80
View File
@@ -1,80 +0,0 @@
"""Tests for recv_mode='copy_to_dst' (ADR-0023 D9.5)."""
from __future__ import annotations
import numpy as np
def test_recv_copy_to_dst_via_simpy_runner():
"""Run a kernel that uses tl.recv(..., dst_addr=..., dst_space=...).
Verify the data is moved to the dst location after recv.
"""
import importlib
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
from kernbench.common.pe_commands import TensorHandle
def kernel(t_ptr, n_elem, dst_buf_addr, tl):
rank = tl.program_id(axis=0)
ws = tl.num_programs(axis=0)
nbytes = n_elem * 2
# Each PE sends own data, then recv into a custom dst slot
current = TensorHandle(
id="loc", addr=t_ptr + rank * nbytes,
shape=(n_elem,), dtype="f16",
nbytes=nbytes, data=None, space="hbm",
)
tl.send(dir="E", src=current)
# copy_to_dst: move into a per-rank scratch HBM addr
recv = tl.recv(
dir="W", shape=(n_elem,), dtype="f16",
dst_addr=dst_buf_addr + rank * nbytes,
dst_space="hbm",
)
# Sanity: recv handle should now point to our dst addr
assert recv.addr == dst_buf_addr + rank * nbytes
assert recv.space == "hbm"
topo = resolve_topology("topology.yaml")
def run(torch):
plan = torch.install_ipcq(
algorithm="ring_allreduce_tcm", world_size_override=8,
)
a = torch.zeros(
(1, 8 * 8),
dtype="f16",
dp=DPPolicy(
cube="replicate", pe="column_wise",
num_cubes=1,
),
name="copy_in",
)
store = torch.engine.memory_store
base = a._handle.va_base or a._handle.shards[0].pa
nbytes = 8 * 2
for r in range(8):
store.write("hbm", base + r * nbytes,
np.full((8,), float(r + 1), dtype=np.float16))
# Use a separate dst region (synthetic addresses)
dst_buf = 0xC0FFEE_0000
torch.launch("ring_allreduce_tcm", kernel, a, 8, dst_buf)
# After the kernel, dst_buf + r*16 should contain rank (r-1)%8's data
for r in range(8):
arr = store.read("hbm", dst_buf + r * nbytes, shape=(8,), dtype="f16")
expected = float(((r - 1) % 8) + 1)
assert np.allclose(arr, expected), f"rank {r}: got {arr}, expected {expected}"
result = run_bench(
topology=topo, bench_fn=run,
device=resolve_device("all"),
engine_factory=lambda t, d: GraphEngine(
getattr(t, "topology_obj", t), enable_data=True
),
)
assert result.completion.ok
-106
View File
@@ -1,106 +0,0 @@
"""Tests for tl.recv_async + tl.wait (ADR-0023 D4)."""
from __future__ import annotations
import numpy as np
from kernbench.ccl.testing import run_kernel_in_mock
def kernel_async_recv(t_ptr, n_elem, tl):
"""Each PE issues recv_async first, then send, then wait — this exercises
the non-blocking path. Uses TensorHandle math (PE_MATH) for accumulation
so Phase 2 produces correct final HBM contents."""
rank = tl.program_id(axis=0)
world_size = tl.num_programs(axis=0)
nbytes = n_elem * 2
pe_addr = t_ptr + rank * nbytes
acc = tl.load(pe_addr, shape=(n_elem,), dtype="f16")
current = acc
for _step in range(world_size - 1):
future = tl.recv_async(dir="W", shape=(n_elem,), dtype="f16")
tl.send(dir="E", src=current)
recv = tl.wait(future)
acc = acc + recv
current = recv # forward W's tile to E next round
tl.store(pe_addr, acc)
def test_recv_async_mock_runtime():
n_elem = 8
inputs = [
np.full((n_elem,), float(r + 1), dtype=np.float16)
for r in range(4)
]
expected = sum(inputs)
outputs = run_kernel_in_mock(
kernel_fn=kernel_async_recv,
world_size=4,
topology="ring_1d",
inputs=inputs,
kernel_args=(n_elem,),
)
for r in range(4):
assert np.allclose(outputs[r], expected)
def test_recv_async_simpy_runner():
"""Run the async kernel through the real SimPy stack via the
install_ipcq + launch path.
"""
import importlib
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
# Re-use the standard 8-PE bench skeleton but swap in the async kernel.
topo = resolve_topology("topology.yaml")
# Build a tiny inline bench module
import types
mod = types.ModuleType("inline_bench_async")
from kernbench.policy.placement.dp import DPPolicy
def run(torch):
plan = torch.install_ipcq(
algorithm="ring_allreduce_tcm", world_size_override=8,
)
a = torch.zeros(
(1, 8 * 8),
dtype="f16",
dp=DPPolicy(
cube="replicate", pe="column_wise",
num_cubes=1,
),
name="async_in",
)
store = torch.engine.memory_store
base = a._handle.va_base or a._handle.shards[0].pa
nbytes = 8 * 2
for r in range(8):
store.write("hbm", base + r * nbytes,
np.full((8,), float(r + 1), dtype=np.float16))
torch.launch("ring_allreduce_tcm", kernel_async_recv, a, 8)
for r in range(8):
result = store.read("hbm", base + r * nbytes, shape=(8,), dtype="f16")
expected = float(sum(range(1, 9))) # 36
assert np.allclose(result, expected, rtol=1e-2, atol=1e-2), \
f"rank {r}: got {result}, expected {expected}"
mod.run = run
result = run_bench(
topology=topo, bench_fn=mod.run,
device=resolve_device("all"),
engine_factory=lambda t, d: GraphEngine(
getattr(t, "topology_obj", t), enable_data=True
),
)
assert result.completion.ok
-301
View File
@@ -1,301 +0,0 @@
"""ADR-0027 T3: Worker-wait generalization + orphan invariant.
Direct regression guard for ADR-0024 Phase B's kernel-greenlet orphan bug.
Phase 1 of ADR-0027: these tests fail against the current code (no
``_pending_worker_waits`` field, no worker-fork in ``ctx.wait``, no
scheduler drain). Phase 2 implements D0.1/D0.2/D0.4 and these pass.
"""
from __future__ import annotations
import os
import textwrap
import pytest
from greenlet import greenlet
# ── helpers ──────────────────────────────────────────────────────────
def _write_minimal_ccl_yaml(tmp_path) -> str:
body = textwrap.dedent("""\
defaults:
algorithm: ring_allreduce_tcm
buffer_kind: tcm
backpressure: sleep
n_slots: 4
slot_size: 4096
vc_chunk_size: 256
ipcq_credit_size_bytes: 16
algorithms:
ring_allreduce_tcm:
module: kernbench.ccl.algorithms.ring_allreduce
topology: ring_1d
buffer_kind: tcm
n_elem: 8
""")
yaml_path = tmp_path / "ccl.yaml"
yaml_path.write_text(body)
return str(tmp_path)
def _make_ctx(topology):
from kernbench.runtime_api.context import RuntimeContext
from kernbench.runtime_api.types import DeviceSelector
from kernbench.sim_engine.engine import GraphEngine
engine = GraphEngine(topology.topology_obj, enable_data=True)
return RuntimeContext(
engine=engine,
target_device=DeviceSelector("all"),
correlation_id="test_t3",
spec=topology.topology_obj.spec,
)
# ── D0.1: _pending_worker_waits field exists ─────────────────────────
def test_pending_worker_waits_field_present(topology):
"""RuntimeContext must expose the deferred-wait queue (D0.1)."""
with _make_ctx(topology) as ctx:
assert hasattr(ctx, "_pending_worker_waits"), (
"ADR-0027 D0.1: RuntimeContext must declare _pending_worker_waits"
)
assert ctx._pending_worker_waits == [], (
"_pending_worker_waits should start empty"
)
# ── T3.a / T3.b: wait defers + resume-after-drain contract ───────────
def test_wait_in_worker_defers_to_main_and_resumes_completed(topology):
"""T3.a + T3.b: worker ctx.wait enqueues + yields; resume → _completed.
Direct test of D0.2 (worker-fork) + D0.3 resume invariant (handle must
be in ctx._completed when worker resumes).
"""
with _make_ctx(topology) as ctx:
from kernbench.policy.placement.dp import DPPolicy
# Worker that submits one tensor (which internally calls ctx.wait)
# and records the pending-queue state observed before/after.
observations: dict = {"pre_wait_len": None, "post_resume_completed": None}
main = greenlet.getcurrent()
def _worker():
# Observation hook: patch ctx.wait to capture a single deferral.
original_wait = ctx.wait
def wrapping_wait(h, *, _meta=None):
observations["pre_wait_len"] = len(ctx._pending_worker_waits)
result = original_wait(h, _meta=_meta)
observations["post_resume_completed"] = h in ctx._completed
return result
ctx.wait = wrapping_wait # type: ignore[assignment]
try:
ctx.zeros(
(1, 8), dtype="f16",
dp=DPPolicy(cube="replicate", pe="replicate",
num_cubes=1, num_pes=1),
name="t3_defer",
)
finally:
ctx.wait = original_wait # type: ignore[assignment]
g = greenlet(_worker)
# Scheduler loop: run worker until it yields (or finishes), then drain.
while not g.dead:
g.switch()
if not g.dead:
# Worker yielded mid-wait → simulate D0.4 drain.
from kernbench.runtime_api.multiprocessing import _drain_pending
_drain_pending(ctx)
assert observations["pre_wait_len"] is not None, "wait was not invoked"
assert observations["post_resume_completed"] is True, (
"D0.3 resume invariant: handle must be in ctx._completed on resume"
)
# ── T3.c: multi-worker same-round drain ──────────────────────────────
def test_multiple_workers_resume_at_same_drain(topology):
"""T3.c: every worker yields before any drain; all resume together."""
with _make_ctx(topology) as ctx:
from kernbench.policy.placement.dp import DPPolicy
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
observations: list[int] = []
def _make_worker(rank: int):
def _entry():
# Before its wait, observe queue state so we can assert that
# *every* worker has enqueued before any drain happened.
ctx.zeros((1, 4), dtype="f16", dp=dp, name=f"r{rank}")
observations.append(rank)
return _entry
ws = 2
gs = [greenlet(_make_worker(r)) for r in range(ws)]
# Round 1: every worker runs up to its first (deferred) ctx.wait.
for g in gs:
g.switch()
# After round 1, all workers should be paused (not yet dead) and
# each should have enqueued at least one handle.
assert all(not g.dead for g in gs), (
"after round 1 switch, workers must be paused mid-wait, not dead"
)
assert len(ctx._pending_worker_waits) >= ws, (
f"expected >= {ws} pending worker waits after round 1; "
f"got {len(ctx._pending_worker_waits)}"
)
# Loop: drain + switch rounds until all workers complete. A single
# ctx.zeros() call contains multiple yield points (MmuMap, then
# MemoryWrite), so more than one round is needed.
from kernbench.runtime_api.multiprocessing import _drain_pending
rounds = 0
while any(not g.dead for g in gs):
_drain_pending(ctx)
for g in gs:
if not g.dead:
g.switch()
rounds += 1
assert rounds < 20, "scheduler did not converge within 20 rounds"
assert all(g.dead for g in gs), "all workers should be dead after drain loop"
assert sorted(observations) == list(range(ws))
# ── T3.d (핵심): kernel greenlet _parent is main ─────────────────────
def test_kernel_greenlet_parent_is_main(topology, tmp_path, monkeypatch):
"""T3.d orphan invariant: kernel_runner._parent must be main greenlet.
This is the direct regression guard for ADR-0024 Phase B. Runs a worker
that invokes torch.launch (which eventually spawns a kernel greenlet).
The kernel_runner.run() captures greenlet.getcurrent() as _parent at
spawn time — that value MUST be the main greenlet, else the orphan
bug is back.
"""
monkeypatch.chdir(_write_minimal_ccl_yaml(tmp_path))
from kernbench.triton_emu import kernel_runner as kr_mod
captured_parents: list = []
main = greenlet.getcurrent()
original_run = kr_mod.KernelRunner.run
def _spy_run(self, env, kernel_fn, kernel_args, num_programs):
gen = original_run(self, env, kernel_fn, kernel_args, num_programs)
def _wrapping_gen():
# yield from gen, but capture self._parent on first step
try:
value = next(gen)
# First yield happens after _parent is set.
captured_parents.append(self._parent)
yield value
except StopIteration:
return
yield from gen
return _wrapping_gen()
monkeypatch.setattr(kr_mod.KernelRunner, "run", _spy_run)
# Drive a minimal ring_allreduce that launches a kernel inside a worker.
import benches.ccl_allreduce as bench
with _make_ctx(topology) as ctx:
bench.run(ctx)
assert captured_parents, "no kernel_runner.run invocations observed"
for p in captured_parents:
assert p is main, (
f"ADR-0027 D0.7 / T3.d: kernel greenlet _parent must be main "
f"greenlet; got {p!r} (main={main!r})"
)
# ── T3.f: idempotency ────────────────────────────────────────────────
def test_wait_same_handle_twice_drives_engine_once(topology):
"""T3.f: ctx.wait(h) + ctx.wait(h) → engine.wait called once (D0.4-(3))."""
with _make_ctx(topology) as ctx:
from kernbench.policy.placement.dp import DPPolicy
dp = DPPolicy(cube="replicate", pe="replicate", num_cubes=1, num_pes=1)
call_count = {"n": 0}
original_engine_wait = ctx.engine.wait
def _counting_wait(h):
call_count["n"] += 1
return original_engine_wait(h)
ctx.engine.wait = _counting_wait # type: ignore[assignment]
def _worker():
ctx.zeros((1, 4), dtype="f16", dp=dp, name="t3f")
# Manually pick a completed handle and wait twice.
assert ctx._completed, "there should be at least one completed handle"
h = next(iter(ctx._completed))
before = call_count["n"]
ctx.wait(h)
ctx.wait(h)
assert call_count["n"] == before, (
"already-completed handle must not re-drive engine.wait"
)
g = greenlet(_worker)
while not g.dead:
g.switch()
if not g.dead:
from kernbench.runtime_api.multiprocessing import _drain_pending
_drain_pending(ctx)
# ── T3.g: exception propagation + no further drain ───────────────────
def test_worker_exception_propagates_and_clears_pending(topology):
"""T3.g: worker raise → main propagates; _pending_worker_waits cleared."""
with _make_ctx(topology) as ctx:
from kernbench.runtime_api.multiprocessing import SpawnException
def _bad_worker(rank: int):
raise ValueError(f"rank {rank} intentional failure")
with pytest.raises(SpawnException) as exc_info:
ctx.multiprocessing.spawn(_bad_worker, args=(), nprocs=2)
assert ctx._pending_worker_waits == [], (
"D0.4-(4): _pending_worker_waits must be cleared on failure"
)
# Root-cause rank errors are present; sibling SystemExit not in dict.
assert 0 in exc_info.value.errors or 1 in exc_info.value.errors
# ── T3.e: historical failure (pre-D0) — skipped per ADR ──────────────
@pytest.mark.skip(
reason="ADR-0027 T3.e: historical failure mode — reproduces only "
"pre-D0.2. Kept as documentation; not run in Phase 2."
)
def test_pre_d0_orphan_reproduction():
"""Placeholder: exercises the pre-D0.2 code path that causes GreenletExit
from kernel_runner._parent captured in worker context. See ADR-0024
Phase B postmortem."""
pass