File size: 17,516 Bytes
b025706 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 | # SPDX-FileCopyrightText: © 2025 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
import contextlib
import fcntl
import os
import time
import pytest
import ttnn
# ==============================================================================
# Device Lock - Coordinates exclusive access to TT devices across processes
# ==============================================================================
_TT_DEVICE_LOCK_PATH = os.environ.get("TT_DEVICE_LOCK_PATH", "/tmp/tt_device.lock")
_TT_DEVICE_LOCK_TIMEOUT = float(os.environ.get("TT_DEVICE_LOCK_TIMEOUT", "60")) # 1 min default
class DeviceLockTimeout(Exception):
"""Raised when acquiring the device lock times out."""
# todo)) the UMD already provides a lock mechanism -- use it instead of this?
@contextlib.contextmanager
def tt_device_lock(lock_path: str = _TT_DEVICE_LOCK_PATH, timeout: float = _TT_DEVICE_LOCK_TIMEOUT):
"""
Context manager for exclusive access to TT devices.
Uses flock for cross-process coordination. Blocks until lock is acquired
or timeout is reached.
Usage:
with tt_device_lock():
mesh = ttnn.open_mesh_device(...)
# ... do work ...
ttnn.close_mesh_device(mesh)
Debug stuck locks with: lsof /tmp/tt_device.lock
Environment variables:
TT_DEVICE_LOCK_PATH: Override lock file path (default: /tmp/tt_device.lock)
TT_DEVICE_LOCK_TIMEOUT: Override timeout in seconds (default: 300)
"""
lock_dir = os.path.dirname(lock_path)
if lock_dir and not os.path.exists(lock_dir):
os.makedirs(lock_dir, exist_ok=True)
lock_file = open(lock_path, "a+") # open the file in append mode to avoid truncation race condition among processes
start_time = time.monotonic()
lock_acquired = False
try:
# Poll for lock with timeout
logged_waiting = False
while True:
try:
fcntl.flock(lock_file, fcntl.LOCK_EX | fcntl.LOCK_NB)
lock_acquired = True
break
except BlockingIOError:
pass # Lock held by another process
if not logged_waiting:
print(f"[tt_device_lock] Waiting for device lock (held by another process)...")
print(f"[tt_device_lock] Debug with: lsof {lock_path}")
logged_waiting = True
if time.monotonic() - start_time >= timeout:
lock_file.close()
raise DeviceLockTimeout(
f"Timed out after {timeout}s waiting for device lock. " f"Check: lsof {lock_path}"
)
time.sleep(1) # sleep for 1 second to avoid busy-waiting
if logged_waiting:
print(f"[tt_device_lock] Lock acquired after {time.monotonic() - start_time:.1f}s")
# Write PID for debugging
lock_file.truncate(0) # clear the file
lock_file.write(f"{os.getpid()}\n")
lock_file.flush()
yield
finally:
if lock_acquired:
fcntl.flock(lock_file, fcntl.LOCK_UN)
lock_file.close()
def pytest_collection_modifyitems(config, items):
"""Deselect tests where ttnn_mesh_device fixture doesn't match mesh_shape param.
This enables tests to use cross-product parametrization (all meshes × all cases)
while only running the valid combinations, without noisy skip messages.
"""
selected = []
deselected = []
for item in items:
if not hasattr(item, "callspec"):
selected.append(item)
continue
params = item.callspec.params
fixture_mesh = params.get("ttnn_mesh_device")
required_mesh = params.get("mesh_shape")
# Keep test if no mesh_shape param or if meshes match
if required_mesh is None or fixture_mesh == required_mesh:
selected.append(item)
else:
deselected.append(item)
items[:] = selected
if deselected:
config.hook.pytest_deselected(items=deselected)
# UMD PCIe hang / board-off-bus. Matched on the invariant tail of:
# Read 0xffffffff over PCIe ID 13: the board should be reset.
# Same class as sweeps_runner._DEVICE_FATAL_SIGNATURES — do not skip these as
# "device unavailable" or coverage fail-under will fire after host-only tests.
_DEVICE_FATAL_SIGNATURES = (
"the board should be reset",
"0xffffffff",
"pciehang",
)
def _is_device_fatal_error(exc: BaseException) -> bool:
"""Return True if ``exc`` indicates a hung board that needs a reset."""
msg = str(exc).lower()
name = type(exc).__name__.lower()
return any(sig in msg or sig in name for sig in _DEVICE_FATAL_SIGNATURES)
@pytest.fixture(scope="module")
def ttnn_mesh_device(request):
"""Create and yield a mesh device for a given mesh shape, cleanup on teardown."""
if not hasattr(request, "param"):
pytest.skip(f"{__file__}: mesh_device fixture called without parametrization")
mesh_device_name = os.environ.get("MESH_DEVICE", "").strip().upper()
blackhole_selected = mesh_device_name in {"P150", "P300", "P150X4"}
if ttnn.device.is_blackhole() and not blackhole_selected:
pytest.skip(f"{__file__}: select Blackhole explicitly with MESH_DEVICE=P150, P300, or P150x4")
if blackhole_selected and not ttnn.device.is_blackhole():
pytest.skip(f"{__file__}: MESH_DEVICE={mesh_device_name} requires a Blackhole device")
# request.param is either a Sequence of ints or a dict with fabric_config and etc.
params = getattr(request, "param", tuple())
if isinstance(params, tuple):
mesh_shape = params
updated_params = dict()
else:
try:
updated_params = params.copy()
mesh_shape = updated_params.pop("mesh_shape")
except Exception as e:
pytest.skip(f"{__file__}: mesh_shape is required: {e}")
# Pre-check: if no devices at all, skip without invoking C++ open.
# Transient driver/UMD issues are "device unavailable"; a hung PCIe bus is not.
try:
num_pcie = ttnn.get_num_pcie_devices()
except Exception as e:
if _is_device_fatal_error(e):
pytest.fail(f"{__file__}: Device hung / needs reset, not a test skip: {e}")
pytest.skip(f"{__file__}: Unable to query TT devices on this system: {e}")
if isinstance(num_pcie, int) and num_pcie == 0:
pytest.skip(f"{__file__}: No TT devices detected on this system")
# Pre-check: skip shapes that cannot fit into the SystemMesh to avoid native exceptions
sys_desc = ttnn._ttnn.multi_device.SystemMeshDescriptor() # type: ignore[attr-defined]
sys_shape = tuple(sys_desc.shape())
req_shape = tuple(mesh_shape)
if blackhole_selected and req_shape == (1, 4) and not _is_physical_p150x4_cluster(ttnn.cluster.get_cluster_type()):
pytest.skip(
"Exact P150x4 hardware coverage requires a physical P150_X4 or P300_X2 cluster; "
"other submeshes are not SKU-equivalent"
)
if blackhole_selected and mesh_device_name == "P300" and req_shape == (1, 2):
cluster_type = ttnn.cluster.get_cluster_type()
if cluster_type not in (ttnn.cluster.ClusterType.P150_X2, ttnn.cluster.ClusterType.P300_X2):
pytest.skip(
"P300 or P300-equivalent development coverage requires a directly connected physical "
"two-chip Blackhole cluster; "
f"got {cluster_type}"
)
allowed = _allowed_req_shapes_for_system(sys_shape, blackhole_selected=blackhole_selected)
if req_shape not in allowed:
pytest.skip(
f"{__file__}: Requested mesh {req_shape} unsupported on system {sys_shape}. "
f"Allowed for this system: {allowed}"
)
parent_shape = _pick_parent_shape_for_submesh(sys_shape, req_shape)
# config fabric config
fabric_config = updated_params.pop("fabric_config", None)
if parent_shape == (1, 1):
# Single device does not need fabric config.
pass
else:
# Select the default fabric topology for the logical mesh requested by the test.
# A submesh still requires opening the full system parent, but its workload topology
# determines whether that parent must provide Ring or Linear routes.
# Select the default fabric topology for the logical mesh requested by the test.
# A submesh still requires opening the full system parent, but its workload topology
# determines whether that parent must provide Ring or Linear routes.
if fabric_config is None:
fabric_config = _default_fabric_config(req_shape)
fabric_config = _default_fabric_config(req_shape)
# set all other input arguments to default values by top-level conftest.py
ttnn.set_fabric_config(
fabric_config, ttnn.FabricReliabilityMode.STRICT_INIT, None, ttnn.FabricTensixConfig.DISABLED
)
# config dispatch core to default values by conftest.py
updated_params["dispatch_core_config"] = ttnn.DispatchCoreConfig(type=None, axis=None, fabric_tensix_config=None)
# If a test requests a submesh of a larger system mesh (e.g. request 2x4 on a 8x4 system),
# fabric cannot be initialized on only the subset of devices. In that case, open the full
# system mesh first, then return the "first" submesh. We intentionally rely on the default
# offset behavior here (i.e. no explicit offset selection).
parent_device = None
submesh_device = None
# Acquire exclusive lock to prevent concurrent device access across processes
with tt_device_lock():
try:
if req_shape != parent_shape:
parent_device = ttnn.open_mesh_device(mesh_shape=ttnn.MeshShape(parent_shape), **updated_params)
submesh_device = parent_device.create_submesh(ttnn.MeshShape(req_shape))
yield submesh_device
else:
parent_device = ttnn.open_mesh_device(mesh_shape=ttnn.MeshShape(parent_shape), **updated_params)
yield parent_device
except Exception as e:
# Focused BH qualification nodes are required gates. Exceptions raised by the test body
# cross the fixture's ``yield`` and must remain failures rather than becoming skips.
# A hung PCIe bus is likewise not a skip: later coverage fail-under would hide it.
# Retain the legacy skip behavior for other non-opted-in WH errors.
if blackhole_selected or _is_device_fatal_error(e):
raise
pytest.skip(f"{__file__}: Mesh device unavailable or unsupported for this configuration: {e}")
finally:
if submesh_device is not None:
ttnn.close_mesh_device(submesh_device)
if parent_device is not None:
ttnn.close_mesh_device(parent_device)
if fabric_config:
ttnn.set_fabric_config(ttnn.FabricConfig.DISABLED)
del parent_device
@pytest.fixture(scope="session")
def require_blackhole_mesh_device():
"""Skip a Blackhole-only test unless the requested SKU is explicit."""
mesh_device_name = os.environ.get("MESH_DEVICE", "").strip().upper()
if mesh_device_name not in {"P150", "P150X4"}:
pytest.skip("Blackhole-only test requires MESH_DEVICE=P150 or P150x4")
return mesh_device_name
def _default_fabric_config(mesh_shape: tuple[int, int]) -> ttnn.FabricConfig | None:
"""Select the generic fabric topology for the requested logical mesh."""
if mesh_shape == (1, 1):
return None
num_devices = mesh_shape[0] * mesh_shape[1]
if mesh_shape[0] == 1 and num_devices >= 8:
return ttnn.FabricConfig.FABRIC_1D_RING
return ttnn.FabricConfig.FABRIC_1D
def _is_physical_p150x4_cluster(cluster_type) -> bool:
"""Accept the two physical four-die BH systems that expose logical P150x4."""
return cluster_type in (ttnn.cluster.ClusterType.P150_X4, ttnn.cluster.ClusterType.P300_X2)
def _allowed_req_shapes_for_system(
sys_shape: tuple[int, int], *, blackhole_selected: bool = False
) -> set[tuple[int, int]]:
# todo)) Different cluster has potentially different physical interconnects (in terms of number of links, topology, etc.).
# Thus, a tuple of ints may not be enough to fingerprint the parent/system mesh device. We need to use a more sophisticated fingerprinting mechanism so we can base the allowed list of (sub)mesh shapes on the parent/system mesh device.
# [INFO] The most robust way to identify the underlying system is to use ttnn.cluster.get_cluster_type(), which returns a ClusterType enum that precisely identifies your hardware configuration. cluster.cpp:16-37
_CANDIDATE_REQ_SHAPES = {
(1, 1): ((1, 1),),
(1, 2): ((1, 2), (1, 1)),
# A 2-chip N300 does not always enumerate as (1, 2): auto-discovery on some hosts (e.g. the
# wh_n300 CI runners) reports the same two chips transposed, as (2, 1). Both are the same
# hardware, so a (2, 1) system must serve the (1, 2) request every N300 demo makes — a request
# that uses all devices is opened in the requested *view* by _pick_parent_shape_for_submesh.
# Without this entry the lookup below misses, `allowed` comes back empty, and EVERY test on
# such a host skips.
(2, 1): ((1, 2), (2, 1), (1, 1)),
# Blackhole P150x4 may enumerate as a line or square. Focused tests use
# the same logical 1x4 view as the 1D modules.
(1, 4): ((1, 4), (1, 2), (1, 1)),
(4, 1): ((1, 4), (4, 1), (1, 2), (1, 1)),
(2, 2): ((2, 2), (1, 4), (1, 2), (1, 1)),
(2, 4): ((2, 4), (1, 8), (1, 4), (1, 2), (1, 1)),
(8, 4): ((8, 4), (4, 8), (1, 8), (1, 4), (1, 2), (1, 1)),
# [INFO] add more system shapes here
}
allowed: set[tuple[int, int]] = set()
if sys_shape in _CANDIDATE_REQ_SHAPES:
for mesh_shape in _CANDIDATE_REQ_SHAPES[sys_shape]:
allowed.add(mesh_shape)
# A physical 2x2 Blackhole quietbox is exposed to the in-scope 1D models
# as the canonical P150x4 view. Preserve generic/non-BH square requests;
# only the opted-in BH path rejects model-visible (2,2).
if blackhole_selected and sys_shape == (2, 2):
allowed.discard((2, 2))
return allowed
def _pick_parent_shape_for_submesh(system_shape: tuple[int, int], requested_shape: tuple[int, int]) -> tuple[int, int]:
# For multi-device workloads we always open the full system mesh (fabric cannot be launched on a subset),
# but we may choose the *orientation* of the full mesh such that the requested submesh fits with the
# default offset (i.e. "first submesh").
if requested_shape == (1, 1):
return (1, 1)
# If the request uses all devices, treat it as a "full-mesh view" shape and open the parent mesh in that view.
# This enables shapes like (1,32) on a system whose SystemMeshDescriptor reports (8,4).
system_num_devices = system_shape[0] * system_shape[1]
requested_num_devices = requested_shape[0] * requested_shape[1]
if requested_num_devices == system_num_devices:
return requested_shape
if requested_shape[0] <= system_shape[0] and requested_shape[1] <= system_shape[1]:
return system_shape
rotated = (system_shape[1], system_shape[0])
if requested_shape[0] <= rotated[0] and requested_shape[1] <= rotated[1]:
return rotated
# No orientation can fit this request without an explicit offset / mapping.
pytest.skip(
f"{__file__}: Requested submesh {requested_shape} does not fit within system mesh {system_shape} "
f"(or its rotated view {rotated}) with default offset."
)
def _host_is_galaxy_cluster() -> bool | None:
"""Whether this host is a Galaxy, or None when the cluster type cannot be determined.
ttnn.cluster.get_cluster_type() returns a ClusterType that fingerprints the hardware
directly, which _allowed_req_shapes_for_system documents as the robust alternative to
inferring the system from a mesh-shape tuple. Note that a hand-wired 32-chip rig
reports CUSTOM rather than GALAXY, so it is not matched here.
"""
try:
galaxy_cluster_types = {
ttnn.cluster.ClusterType.GALAXY,
ttnn.cluster.ClusterType.TG,
ttnn.cluster.ClusterType.BLACKHOLE_GALAXY,
}
return ttnn.cluster.get_cluster_type() in galaxy_cluster_types
except Exception:
return None
@pytest.fixture(scope="module")
def skip_on_galaxy_system():
"""Skip a module whose meshes are 1D-only when the host system is a Galaxy.
The 1D module suites request shapes like (1, 8), which _allowed_req_shapes_for_system
permits on a Galaxy (8, 4) as well as on a T3K (2, 4). They are targeted at the T3K,
so on a Galaxy they would consume scarce hardware to re-run LLMBox coverage. Gate is
opt-in per module rather than applied in _allowed_req_shapes_for_system, because
test_auto_compose.py deliberately exercises 1D shapes on a Galaxy.
Queried at fixture setup rather than collection time: probing the device while
collecting has deadlocked nested-pytest runs before. An indeterminate cluster type
does not skip; the mesh fixture makes the call instead.
"""
if _host_is_galaxy_cluster():
pytest.skip("1D module suites are T3K-targeted; host cluster type is a Galaxy")
|