def __init__(self, container_strategy, num_gpus_per_worker):
super(ParameterServerExtended, self).__init__(container_strategy)
self._num_gpus_per_worker = num_gpus_per_worker
self._initialize_local(num_gpus_per_worker)
# We typically don't need to do all-reduce in this strategy.
self._cross_device_ops = (
cross_device_ops_lib.ReductionToOneDeviceCrossDeviceOps(
reduce_to_device=_LOCAL_CPU))
def __init__(self,
container_strategy,
cluster_resolver=TFConfigClusterResolver()):
super(ParameterServerStrategyExtended, self).__init__(container_strategy)
self._initialize_strategy(cluster_resolver)
# We typically don't need to do all-reduce in this strategy.
self._cross_device_ops = (
cross_device_ops_lib.ReductionToOneDeviceCrossDeviceOps(
reduce_to_device=_LOCAL_CPU))
def _get_cross_device_ops(self):
if self._cross_device_ops is None:
self._cross_device_ops = (
cross_device_ops_lib.ReductionToOneDeviceCrossDeviceOps())
return self._cross_device_ops
class SingleWorkerCrossDeviceOpsTest(CrossDeviceOpsTestBase):
# TODO(yuefengz): decouple the num_gpus check from distribution in
# combinations module so that we can pass in devices instead of a distribution
# strategy.
reduction_to_one_combinations = combinations.combine(
cross_device_ops=[
combinations.NamedObject(
"DefaultReductionToOneDeviceCrossDeviceOps",
cross_device_ops_lib.ReductionToOneDeviceCrossDeviceOps()),
combinations.NamedObject(
"ReductionToCPUDeviceCrossDeviceOps",
cross_device_ops_lib.ReductionToOneDeviceCrossDeviceOps(
reduce_to_device=_cpu_device)),
combinations.NamedObject(
"AccumulateNCrossDeviceOp",
cross_device_ops_lib.ReductionToOneDeviceCrossDeviceOps(
accumulation_fn=math_ops.accumulate_n)),
],
distribution=[
combinations.one_device_strategy,
combinations.mirrored_strategy_with_gpu_and_cpu,
combinations.mirrored_strategy_with_two_gpus,
combinations.core_mirrored_strategy_with_gpu_and_cpu,
combinations.core_mirrored_strategy_with_two_gpus
],
mode=["graph", "eager"])
allreduce_combinations = combinations.combine(
cross_device_ops=[
combinations.NamedObject(
"AllReduce",
cross_device_ops_lib.AllReduceCrossDeviceOps("nccl", 1, 0, 0)),
combinations.NamedObject(
"HierarchicalCopy",
cross_device_ops_lib.AllReduceCrossDeviceOps(
"hierarchical_copy", 8, 0, 0)),
combinations.NamedObject(
"AllReduceNoGradientRepacking",
cross_device_ops_lib.AllReduceCrossDeviceOps("nccl", 0, 0, 0)),
combinations.NamedObject(
"HierarchicalCopyAggregateSmallTensors",
cross_device_ops_lib.AllReduceCrossDeviceOps(
"hierarchical_copy", 0, 100, 10))
],
distribution=[combinations.mirrored_strategy_with_two_gpus,
combinations.core_mirrored_strategy_with_two_gpus],
mode=["graph", "eager"])
@combinations.generate(reduction_to_one_combinations + allreduce_combinations)
def testReductionAndBroadcast(self, cross_device_ops, distribution):
with distribution.scope():
self._testReductionAndBroadcast(cross_device_ops, distribution)
def testChooseAlgorithm(self):
device_links = [[1, 2, 3, 4], [0, 2, 3, 5], [0, 1, 3, 6], [0, 1, 2, 7],
[0, 5, 6, 7], [1, 4, 6, 7], [2, 4, 5, 7], [3, 4, 5, 6]]
result = cross_device_ops_lib._choose_all_reduce_algorithm(device_links)
self.assertIsInstance(result, cross_device_ops_lib.AllReduceCrossDeviceOps)
self.assertEqual(result._all_reduce_alg, "hierarchical_copy")
self.assertEqual(result._num_packs, 8)
# if there are only 4 devices
device_links = [[1, 2, 3, 4], [0, 2, 3, 5], [0, 1, 3, 6], [0, 1, 2, 7]]
result = cross_device_ops_lib._choose_all_reduce_algorithm(device_links)
self.assertIsInstance(result, cross_device_ops_lib.AllReduceCrossDeviceOps)
self.assertEqual(result._all_reduce_alg, "nccl")
self.assertEqual(result._num_packs, 1)
# if devices links contain each device itself
device_links = [[0, 1, 2, 3, 4], [0, 1, 2, 3, 5], [0, 1, 2, 3, 6],
[0, 1, 2, 3, 7], [0, 4, 5, 6, 7], [1, 4, 5, 6, 7],
[2, 4, 5, 6, 7], [3, 4, 5, 6, 7]]
result = cross_device_ops_lib._choose_all_reduce_algorithm(device_links)
self.assertIsInstance(result, cross_device_ops_lib.AllReduceCrossDeviceOps)
self.assertEqual(result._all_reduce_alg, "hierarchical_copy")
self.assertEqual(result._num_packs, 8)
# if not dgx1-like links
device_links = [[0, 2, 3, 5], [0, 1, 3, 6], [0, 1, 2, 7], [0, 5, 6, 7],
[1, 4, 6, 7], [2, 4, 5, 7], [3, 4, 5, 6], [1, 2, 3, 4]]
result = cross_device_ops_lib._choose_all_reduce_algorithm(device_links)
self.assertIsInstance(result, cross_device_ops_lib.AllReduceCrossDeviceOps)
self.assertEqual(result._all_reduce_alg, "nccl")
self.assertEqual(result._num_packs, 1)
@combinations.generate(combinations.combine(
mode=["graph", "eager"],
required_gpus=1))
def testSimpleReduceWithIndexedSlices(self):
devices = ["/cpu:0", "/gpu:0"]
t0 = _make_indexed_slices([[1., 2.]], [1], [5, 2], devices[0])
t1 = _make_indexed_slices([[3., 4.], [5., 6.]], [1, 3], [5, 2], devices[1])
per_replica = value_lib.PerReplica({devices[0]: t0, devices[1]: t1})
result = cross_device_ops_lib._simple_reduce(
per_replica, devices[0], math_ops.add_n, reduce_util.ReduceOp.SUM)
# Test that the result is semantically equal to both the concatenated
# IndexedSlices with and without duplicate indices.
total_with_dups = _make_indexed_slices(
[[1., 2.], [3., 4.], [5., 6.]], [1, 1, 3], [5, 2], devices[0])
total_without_dups = _make_indexed_slices(
[[4., 6.], [5., 6.]], [1, 3], [5, 2], devices[0])
self._assert_indexed_slices_equal(total_with_dups, result)
self._assert_indexed_slices_equal(total_without_dups, result)
@combinations.generate(
combinations.combine(
cross_device_ops_instance=[
combinations.NamedObject(
"ReductionToOneDeviceCrossDeviceOps",
cross_device_ops_lib.ReductionToOneDeviceCrossDeviceOps()),
combinations.NamedObject(
"AllReduceCrossDeviceOps",
cross_device_ops_lib.AllReduceCrossDeviceOps())
],
reduce_op=[reduce_util.ReduceOp.SUM, reduce_util.ReduceOp.MEAN],
batch_reduce=[True, False],
mode=["graph", "eager"],
required_gpus=1))
def testIndexedSlicesAllReduce(self, cross_device_ops_instance, reduce_op,
batch_reduce):
devices = ["/cpu:0", "/gpu:0"]
dense_shape = [5, 2]
t0 = _make_indexed_slices([[1., 2.]], [1], dense_shape, devices[0])
t1 = _make_indexed_slices(
[[3., 4.], [5., 6.]], [1, 3], dense_shape, devices[1])
per_replica = value_lib.PerReplica({devices[0]: t0, devices[1]: t1})
if batch_reduce:
result = cross_device_ops_instance.batch_reduce(
reduce_op, [(per_replica, devices)])
else:
result = cross_device_ops_instance.reduce(
reduce_op, per_replica, devices)
total_indices_with_dups = [1, 1, 3]
total_indices_without_dups = [1, 3]
if reduce_op == reduce_util.ReduceOp.SUM:
total_values_with_dups = [[1., 2.], [3., 4.], [5., 6.]]
total_values_without_dups = [[4., 6.], [5., 6.]]
else:
assert reduce_op == reduce_util.ReduceOp.MEAN
total_values_with_dups = [[0.5, 1.], [1.5, 2.], [2.5, 3.]]
total_values_without_dups = [[2., 3.], [2.5, 3.]]
total_mirrored_with_dups = _make_mirrored_indexed_slices(
devices, total_values_with_dups, total_indices_with_dups, dense_shape)
total_mirrored_without_dups = _make_mirrored_indexed_slices(
devices, total_values_without_dups, total_indices_without_dups,
dense_shape)
# Test that the result is semantically equal to both the concatenated
# IndexedSlices, as well as when the duplicate indices are summed up.
if batch_reduce:
total_mirrored_with_dups = [total_mirrored_with_dups]
total_mirrored_without_dups = [total_mirrored_without_dups]
self._assert_values_equal(total_mirrored_with_dups, result)
self._assert_values_equal(total_mirrored_without_dups, result)
