def testLimitingInfNanMonitorAlertCountWorks(self):
mock_reader = test.mock.MagicMock()
monitor = debug_events_monitors.InfNanMonitor(mock_reader, limit=3)
for i in range(10):
execution_digest = debug_events_reader.ExecutionDigest(
i * 1000, 1, "FooOp", output_tensor_device_ids=[0, 1])
execution = debug_events_reader.Execution(
execution_digest,
"worker01", ["a1", "b2", "e3"],
debug_event_pb2.TensorDebugMode.CURT_HEALTH,
graph_id=None,
input_tensor_ids=[12, 34],
output_tensor_ids=[56, 78],
debug_tensor_values=[[-1, 0],
[-1, 1]]) # [tensor_id, any_inf_nan].
monitor.on_execution(i, execution)
alerts = monitor.alerts()
self.assertLen(alerts, 3)
for i, alert in enumerate(alerts):
self.assertEqual(alert.wall_time, i * 1000)
self.assertEqual(alert.op_type, "FooOp")
self.assertEqual(alert.output_slot, 1)
# The four fields below are unavailable under CURT_HEALTH mode by design.
self.assertIsNone(alert.size)
self.assertIsNone(alert.num_neg_inf)
self.assertIsNone(alert.num_pos_inf)
self.assertIsNone(alert.num_nan)
self.assertEqual(alert.execution_index, i)
self.assertIsNone(alert.graph_execution_trace_index)
def testInfNanMonitorOnExecutionUnderHealthMode(self, tensor_debug_mode,
debug_tensor_values):
mock_reader = test.mock.MagicMock()
monitor = debug_events_monitors.InfNanMonitor(mock_reader)
execution_digest = debug_events_reader.ExecutionDigest(
1234, 1, "BarOp", output_tensor_device_ids=[0, 1])
execution = debug_events_reader.Execution(
execution_digest,
"worker01", ["a1", "b2", "e3"],
tensor_debug_mode,
graph_id=None,
input_tensor_ids=[12, 34],
output_tensor_ids=[56, 78],
debug_tensor_values=debug_tensor_values)
monitor.on_execution(60, execution)
self.assertLen(monitor.alerts(), 1)
alert = monitor.alerts()[0]
self.assertEqual(alert.wall_time, 1234)
self.assertEqual(alert.op_type, "BarOp")
self.assertEqual(alert.output_slot, 0)
self.assertEqual(alert.size, 10)
self.assertEqual(alert.num_neg_inf, 1)
self.assertEqual(alert.num_pos_inf, 2)
self.assertEqual(alert.num_nan, 3)
self.assertEqual(alert.execution_index, 60)
self.assertIsNone(alert.graph_execution_trace_index)
def testInfNanMonitorOnExecutionUnderCurtHealthMode(self):
mock_reader = test.mock.MagicMock()
monitor = debug_events_monitors.InfNanMonitor(mock_reader)
execution_digest = debug_events_reader.ExecutionDigest(
1234, 1, "FooOp", output_tensor_device_ids=[0, 1])
execution = debug_events_reader.Execution(
execution_digest,
"worker01", ["a1", "b2", "e3"],
debug_event_pb2.TensorDebugMode.CURT_HEALTH,
graph_id=None,
input_tensor_ids=[12, 34],
output_tensor_ids=[56, 78],
debug_tensor_values=[[-1, 0], [-1,
1]]) # [tensor_id, any_inf_nan].
monitor.on_execution(50, execution)
self.assertLen(monitor.alerts(), 1)
alert = monitor.alerts()[0]
self.assertEqual(alert.wall_time, 1234)
self.assertEqual(alert.op_type, "FooOp")
self.assertEqual(alert.output_slot, 1)
# The four fields below are unavailable under CURT_HEALTH mode by design.
self.assertIsNone(alert.size)
self.assertIsNone(alert.num_neg_inf)
self.assertIsNone(alert.num_pos_inf)
self.assertIsNone(alert.num_nan)
self.assertEqual(alert.execution_index, 50)
self.assertIsNone(alert.graph_execution_trace_index)
def testExecutionNoGraphNoInputButWithOutputToJson(self):
execution_digest = debug_events_reader.ExecutionDigest(
1234, 5678, "FooOp", output_tensor_device_ids=[1357])
execution = debug_events_reader.Execution(
execution_digest,
"localhost",
("a1", "b2"),
debug_event_pb2.TensorDebugMode.FULL_HEALTH,
graph_id="abcd",
input_tensor_ids=[13, 37],
output_tensor_ids=None,
debug_tensor_values=None)
json = execution.to_json()
self.jsonRoundTripCheck(json)
self.assertEqual(json["wall_time"], 1234)
self.assertEqual(json["op_type"], "FooOp")
self.assertEqual(json["output_tensor_device_ids"], (1357,))
self.assertEqual(json["host_name"], "localhost")
self.assertEqual(json["stack_frame_ids"], ("a1", "b2"))
self.assertEqual(json["tensor_debug_mode"],
debug_event_pb2.TensorDebugMode.FULL_HEALTH)
self.assertEqual(json["graph_id"], "abcd")
self.assertEqual(json["input_tensor_ids"], (13, 37))
self.assertIsNone(json["output_tensor_ids"])
self.assertIsNone(json["debug_tensor_values"])
def testExecutionDigestWithTwoOutputsToJson(self):
execution_digest = debug_events_reader.ExecutionDigest(
1234, 5678, "FooOp", output_tensor_device_ids=[1357, 2468])
json = execution_digest.to_json()
self.jsonRoundTripCheck(json)
self.assertEqual(json["wall_time"], 1234)
self.assertEqual(json["op_type"], "FooOp")
self.assertEqual(json["output_tensor_device_ids"], (1357, 2468))
def testExecutionWithNoOutputTensorsReturnsZeroForNumOutputs(
self, output_tensor_ids):
execution = debug_events_reader.Execution(
debug_events_reader.ExecutionDigest(1234, 5678, "FooOp"),
"localhost", ("a1", "b2"),
debug_event_pb2.TensorDebugMode.FULL_HEALTH,
graph_id="abcd",
input_tensor_ids=[13, 37],
output_tensor_ids=output_tensor_ids,
debug_tensor_values=None)
self.assertEqual(execution.num_outputs, 0)
def testInfNanMonitorOnExecutionUnderModeWithNoInfNanInfo(
self, tensor_debug_mode, debug_tensor_values):
mock_reader = test.mock.MagicMock()
monitor = debug_events_monitors.InfNanMonitor(mock_reader)
execution_digest = debug_events_reader.ExecutionDigest(
1234, 1, "BarOp", output_tensor_device_ids=[0, 1])
execution = debug_events_reader.Execution(
execution_digest,
"worker01", ["a1", "b2", "e3"],
tensor_debug_mode,
graph_id=None,
input_tensor_ids=[12, 34],
output_tensor_ids=[56, 78],
debug_tensor_values=debug_tensor_values)
monitor.on_execution(60, execution)
self.assertEmpty(monitor.alerts())
def testInfNanMonitorOnExecutionUnderFullTensorModeWorks(
self, tensor_value, dtype, expected_size, expected_num_neg_inf,
expected_num_pos_inf, expected_num_nan):
mock_reader = test.mock.MagicMock()
mock_reader.execution_to_tensor_values.return_value = [
np.array([[0.0, -1.0, 1.0]]),
np.array(tensor_value, dtype=dtype)
]
monitor = debug_events_monitors.InfNanMonitor(mock_reader)
execution_digest = debug_events_reader.ExecutionDigest(
1234,
1,
"__inference_bar_function_1234",
output_tensor_device_ids=[0, 1])
execution = debug_events_reader.Execution(
execution_digest,
"worker01", ["a1", "b2", "e3"],
debug_event_pb2.TensorDebugMode.FULL_TENSOR,
graph_id=None,
input_tensor_ids=[12, 34],
output_tensor_ids=[56, 78])
monitor.on_execution(70, execution)
if expected_num_neg_inf or expected_num_pos_inf or expected_num_nan:
self.assertLen(monitor.alerts(), 1)
alert = monitor.alerts()[0]
self.assertEqual(alert.wall_time, 1234)
self.assertEqual(alert.op_type, "__inference_bar_function_1234")
self.assertEqual(alert.output_slot, 1)
self.assertEqual(alert.size, expected_size)
self.assertEqual(alert.num_neg_inf, expected_num_neg_inf)
self.assertEqual(alert.num_pos_inf, expected_num_pos_inf)
self.assertEqual(alert.num_nan, expected_num_nan)
self.assertEqual(alert.execution_index, 70)
self.assertIsNone(alert.graph_execution_trace_index, 70)
else:
self.assertEmpty(monitor.alerts())
