Files
kernbench2/tests/test_ccl_deadlock_detection.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

126 lines
3.9 KiB
Python

"""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
),
)