Source code for horovod.tensorflow

# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
# Modifications copyright (C) 2019 Uber Technologies, Inc.
# Modifications copyright Microsoft
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
# pylint: disable=g-short-docstring-punctuation

import os
import warnings

from horovod.common.util import check_extension, gpu_available

check_extension('horovod.tensorflow', 'HOROVOD_WITH_TENSORFLOW', __file__, 'mpi_lib')

from horovod.tensorflow import elastic
from horovod.tensorflow.compression import Compression
from horovod.tensorflow.functions import allgather_object, broadcast_object, broadcast_object_fn, broadcast_variables
from horovod.tensorflow.mpi_ops import allgather, broadcast, _allreduce, alltoall
from horovod.tensorflow.mpi_ops import init, shutdown
from horovod.tensorflow.mpi_ops import is_initialized, start_timeline, stop_timeline
from horovod.tensorflow.mpi_ops import size, local_size, rank, local_rank, is_homogeneous
from horovod.tensorflow.mpi_ops import rank_op, local_rank_op, size_op, local_size_op
from horovod.tensorflow.mpi_ops import mpi_threads_supported, mpi_enabled, mpi_built
from horovod.tensorflow.mpi_ops import gloo_enabled, gloo_built
from horovod.tensorflow.mpi_ops import nccl_built, ddl_built, ccl_built, cuda_built, rocm_built
from horovod.tensorflow.mpi_ops import Average, Sum, Adasum
from horovod.tensorflow.mpi_ops import handle_average_backwards_compatibility, check_num_rank_power_of_2
from horovod.tensorflow.util import _executing_eagerly, _make_subgraph, _cache
from horovod.tensorflow.mpi_ops import join
from horovod.tensorflow.sync_batch_norm import SyncBatchNormalization

import tensorflow as tf

# @DEKHTIARJonathan: Do not remove, this fixes issues:
# - https://github.com/tensorflow/tensorflow/issues/38516
# - https://github.com/tensorflow/tensorflow/issues/39894
if tf.__version__.startswith('2.2.'):
  from tensorflow.python.keras.mixed_precision.experimental import device_compatibility_check
  device_compatibility_check.log_device_compatibility_check = lambda policy_name, skip_local: None


[docs]def allreduce(tensor, average=None, device_dense='', device_sparse='', compression=Compression.none, op=None, prescale_factor=1.0, postscale_factor=1.0, name=None): """Perform an allreduce on a tf.Tensor or tf.IndexedSlices. This function performs a bandwidth-optimal ring allreduce on the input tensor. If the input is an tf.IndexedSlices, the function instead does an allgather on the values and the indices, effectively doing an allreduce on the represented tensor. Arguments: tensor: tf.Tensor, tf.Variable, or tf.IndexedSlices to reduce. The shape of the input must be identical across all ranks. average: .. warning:: .. deprecated:: 0.19.0 Use `op` instead. Will be removed in v0.21.0. device_dense: Device to be used for dense tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. device_sparse: Device to be used for sparse tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. compression: Compression algorithm used to reduce the amount of data sent and received by each worker node. Defaults to not using compression. op: The reduction operation to combine tensors across different ranks. Defaults to Average if None is given. prescale_factor: Multiplicative factor to scale tensor before allreduce. postscale_factor: Multiplicative factor to scale tensor after allreduce. name: A name of the allreduce operation Returns: A tensor of the same shape and type as `tensor`, summed across all processes. """ op = handle_average_backwards_compatibility(op, average) if isinstance(tensor, tf.IndexedSlices): # TODO: Need to fix this to actuall call Adasum if op == Adasum: raise NotImplementedError('The Adasum reduction does not currently support sparse tensors. As a ' 'workaround please pass sparse_as_dense=True to DistributedOptimizer') with tf.device(device_sparse): # For IndexedSlices, do two allgathers instead of an allreduce. horovod_size = tf.cast(size_op() if int(os.environ.get("HOROVOD_ELASTIC", 0)) else size(), dtype=tensor.values.dtype) values = allgather(tensor.values) indices = allgather(tensor.indices) # To make this operation into an average, divide allgathered values by # the Horovod size. new_values = (values / horovod_size) if op == Average else values return tf.IndexedSlices(new_values, indices, dense_shape=tensor.dense_shape) else: average_in_framework = False if rocm_built(): # For ROCm, perform averaging at framework level average_in_framework = op == Average or op == Adasum op = Sum if op == Average else op with tf.device(device_dense): horovod_size = tf.cast(size_op() if int(os.environ.get("HOROVOD_ELASTIC", 0)) else size(), dtype=tensor.dtype) tensor_compressed, ctx = compression.compress(tensor) summed_tensor_compressed = _allreduce(tensor_compressed, op=op, prescale_factor=prescale_factor, postscale_factor=postscale_factor, name=name) summed_tensor = compression.decompress(summed_tensor_compressed, ctx) if op == Adasum: if 'CPU' not in tensor.device and gpu_available('tensorflow'): if nccl_built(): if not is_homogeneous: raise NotImplementedError( 'Running GPU Adasum on heterogeneous cluster is not supported yet.') elif not check_num_rank_power_of_2(int(size() / local_size())): raise NotImplementedError( 'Running GPU Adasum with non-power of 2 nodes is not supported yet.') if rocm_built(): horovod_local_size = tf.cast(local_size_op() if int(os.environ.get("HOROVOD_ELASTIC", 0)) else local_size(), dtype=tensor.dtype) new_tensor = summed_tensor / horovod_local_size else: new_tensor = summed_tensor else: warnings.warn('Adasum reduction does not currently support GPU reduction using MPI. Tensors ' 'are copied to CPU memory instead. To use Adasum for GPU reduction, please ' 'compile Horovod with HOROVOD_GPU_OPERATIONS=NCCL.') new_tensor = summed_tensor else: if not check_num_rank_power_of_2(size()): raise NotImplementedError('Running Adasum with non-power of 2 ranks is not supported yet.') new_tensor = summed_tensor else: if rocm_built(): new_tensor = (summed_tensor / horovod_size) if average_in_framework else summed_tensor else: new_tensor = summed_tensor return new_tensor
def _allreduce_cond(tensor, *args, **kwargs): def allreduce_fn(): return allreduce(tensor, *args, **kwargs) def id_fn(): return tensor return tf.cond((size_op() > 1) if int(os.environ.get("HOROVOD_ELASTIC", 0)) else tf.convert_to_tensor(size() > 1), allreduce_fn, id_fn) try: _global_variables = tf.compat.v1.global_variables except AttributeError: try: _global_variables = tf.global_variables except AttributeError: _global_variables = None if _global_variables is not None:
[docs] def broadcast_global_variables(root_rank): """Broadcasts all global variables from root rank to all other processes. **NOTE:** deprecated in TensorFlow 2.0. Arguments: root_rank: rank of the process from which global variables will be broadcasted to all other processes. """ if _executing_eagerly(): raise RuntimeError( "hvd.broadcast_global_variables() does not support eager execution. " "Please use `hvd.broadcast_variables(<model/optimizer variables>)` instead." ) return broadcast_variables(_global_variables(), root_rank)
try: _get_default_graph = tf.compat.v1.get_default_graph except AttributeError: try: _get_default_graph = tf.get_default_graph except AttributeError: _get_default_graph = None try: _SessionRunHook = tf.estimator.SessionRunHook except AttributeError: try: _SessionRunHook = tf.train.SessionRunHook except AttributeError: _SessionRunHook = None if _SessionRunHook is not None and _get_default_graph is not None:
[docs] class BroadcastGlobalVariablesHook(_SessionRunHook): """ SessionRunHook that will broadcast all global variables from root rank to all other processes during initialization. This is necessary to ensure consistent initialization of all workers when training is started with random weights or restored from a checkpoint. **NOTE:** deprecated in TensorFlow 2.0. """ def __init__(self, root_rank, device=''): """Construct a new BroadcastGlobalVariablesHook that will broadcast all global variables from root rank to all other processes during initialization. Args: root_rank: Rank that will send data, other ranks will receive data. device: Device to be used for broadcasting. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. """ super(BroadcastGlobalVariablesHook, self).__init__() self.root_rank = root_rank self.bcast_op = None self.device = device def begin(self): if not self.bcast_op or self.bcast_op.graph != _get_default_graph(): with tf.device(self.device): self.bcast_op = broadcast_global_variables(self.root_rank) def after_create_session(self, session, coord): session.run(self.bcast_op)
@_cache def _make_allreduce_grads_fn(name, device_dense, device_sparse, compression, sparse_as_dense, op, gradient_predivide_factor): if op == Average: # Split average operation across pre/postscale factors # C++ backend will apply additional 1 / size() factor to postscale_factor for op == Average. prescale_factor = 1.0 / gradient_predivide_factor postscale_factor = gradient_predivide_factor else: prescale_factor = 1.0 postscale_factor = 1.0 def allreduce_grads(grads): with tf.name_scope(name + "_Allreduce"): if sparse_as_dense: grads = [tf.convert_to_tensor(grad) if grad is not None and isinstance(grad, tf.IndexedSlices) else grad for grad in grads] return [_allreduce_cond(grad, device_dense=device_dense, device_sparse=device_sparse, compression=compression, op=op, prescale_factor=prescale_factor, postscale_factor=postscale_factor) if grad is not None else grad for grad in grads] if _executing_eagerly(): return _make_subgraph(allreduce_grads) else: return allreduce_grads try: # TensorFlow 2.x _LegacyOptimizer = tf.compat.v1.train.Optimizer except AttributeError: try: # TensorFlow 1.x _LegacyOptimizer = tf.train.Optimizer except AttributeError: # Future TensorFlow versions _LegacyOptimizer = None if _LegacyOptimizer is not None: class _DistributedOptimizer(_LegacyOptimizer): """An optimizer that wraps another tf.Optimizer, using an allreduce to combine gradient values before applying gradients to model weights.""" def __init__(self, optimizer, name=None, use_locking=False, device_dense='', device_sparse='', compression=Compression.none, sparse_as_dense=False, op=Average, gradient_predivide_factor=1.0): if name is None: name = "Distributed{}".format(type(optimizer).__name__) super(_DistributedOptimizer, self).__init__(name=name, use_locking=use_locking) self._optimizer = optimizer self._allreduce_grads = _make_allreduce_grads_fn( name, device_dense, device_sparse, compression, sparse_as_dense, op, gradient_predivide_factor) def compute_gradients(self, *args, **kwargs): """Compute gradients of all trainable variables. See Optimizer.compute_gradients() for more info. In DistributedOptimizer, compute_gradients() is overriden to also allreduce the gradients before returning them. """ gradients = self._optimizer.compute_gradients(*args, **kwargs) grads, vars = zip(*gradients) avg_grads = self._allreduce_grads(grads) return list(zip(avg_grads, vars)) def apply_gradients(self, *args, **kwargs): """Calls this same method on the underlying optimizer.""" return self._optimizer.apply_gradients(*args, **kwargs) def get_slot(self, *args, **kwargs): """Calls this same method on the underlying optimizer.""" return self._optimizer.get_slot(*args, **kwargs) def get_slot_names(self, *args, **kwargs): """Calls this same method on the underlying optimizer.""" return self._optimizer.get_slot_names(*args, **kwargs) def variables(self, *args, **kwargs): """Calls this same method on the underlying optimizer.""" return self._optimizer.variables(*args, **kwargs) class _DistributedAdasumOptimizer(_LegacyOptimizer): """An optimizer that wraps another tf.Optimizer, using an allreduce to combine model deltas after applying gradients to model weights.""" def __init__(self, optimizer, name=None, use_locking=False, device_dense='', device_sparse='', compression=Compression.none, backward_passes_per_step=1): if name is None: name = "DistributedDelta{}".format(type(optimizer).__name__) super(_DistributedAdasumOptimizer, self).__init__(name=name, use_locking=use_locking) self._optimizer = optimizer self._name = name self._device_dense = device_dense self._device_sparse = device_sparse self._compression = compression self._backward_passes_per_step = backward_passes_per_step def _prepare(self): self._step_count = tf.get_variable( name="step_count", shape=[], dtype=tf.int64, trainable=False, initializer=tf.zeros_initializer) self._is_first_step = tf.cast(tf.math.equal(self._step_count, 0), dtype=tf.bool) self._is_comm_step = tf.cast(tf.math.equal(self._step_count % self._backward_passes_per_step, self._backward_passes_per_step - 1), dtype=tf.bool) def _apply_shared(self, var, get_update_op): start_slot = self._get_or_make_slot(var, "delta_start") # initialize start on the first step assign_op = tf.cond(self._is_first_step, lambda: start_slot.assign(var, use_locking=self.use_locking).op, tf.no_op) with tf.control_dependencies([assign_op]): update_op = get_update_op() with tf.control_dependencies([update_op]): def update(): # delta = var - start local_delta = var.assign_sub(start_slot, use_locking=self.use_locking) # reuse var's memory # delta = allreduce (delta) global_delta = allreduce(local_delta, device_dense=self._device_dense, device_sparse=self._device_sparse, compression=self._compression, op=Adasum) # start = start + delta new_start = start_slot.assign_add(global_delta, use_locking=self.use_locking) # var = start return var.assign(new_start, use_locking=self.use_locking).op # if its a communication step, then apply logic above # if its not a communication step then just have the underlying # optimizer update the model parameters according to its logic return tf.cond(self._is_comm_step, update, tf.no_op) def _apply_dense(self, grad, var): return self._apply_shared(var, lambda: self._optimizer._apply_dense(grad, var)) def _resource_apply_dense(self, grad, handle): return self._apply_shared(handle, lambda: self._optimizer._resource_apply_dense(grad, handle)) def _apply_sparse(self, grad, var): return self._apply_shared(var, lambda: self._optimizer._apply_sparse(grad, var)) def _resource_apply_sparse(self, grad, handle, indices): return self._apply_shared(handle, lambda: self._optimizer._resource_apply_sparse(grad, handle, indices)) def _finish(self, update_ops, name_scope): with tf.control_dependencies(update_ops): return tf.assign_add(self._step_count, 1) def compute_gradients(self, *args, **kwargs): """Compute gradients of all trainable variables. See Optimizer.compute_gradients() for more info. """ return self._optimizer.compute_gradients(*args, **kwargs) def apply_gradients(self, *args, **kwargs): """Calls this same method on the underlying optimizer.""" return self._optimizer.apply_gradients(*args, **kwargs) def get_slot(self, var, name): """Calls this same method on the underlying optimizer.""" tmp = super(_DistributedAdasumOptimizer, self).get_slot(var, name) if tmp is not None: return tmp return self._optimizer.get_slot(var, name) def get_slot_names(self): """Appends local slot names to those of the underlying optimizer.""" return super(_DistributedAdasumOptimizer, self).get_slot_names() +\ self._optimizer.get_slot_names() def variables(self, *args, **kwargs): """Calls this same method on the underlying optimizer.""" return self._optimizer.variables(*args, **kwargs)
[docs]def DistributedOptimizer(optimizer, name=None, use_locking=False, device_dense='', device_sparse='', compression=Compression.none, sparse_as_dense=False, backward_passes_per_step=1, op=Average, gradient_predivide_factor=1.0): """Construct a new DistributedOptimizer, which uses another optimizer under the hood for computing single-process gradient values and applying gradient updates after the gradient values have been combined across all the Horovod ranks. Args: optimizer: Optimizer to use for computing gradients and applying updates. name: Optional name prefix for the operations created when applying gradients. Defaults to "Distributed" followed by the provided optimizer type. use_locking: Whether to use locking when updating variables. See Optimizer.__init__ for more info. device_dense: Device to be used for dense tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. device_sparse: Device to be used for sparse tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. compression: Compression algorithm used during allreduce to reduce the amount of data sent during each parameter update step. Defaults to not using compression. sparse_as_dense: Treat all sparse gradients as dense tensors. This can help improve performance and memory utilization if the original sparse gradient has high density. Defaults to false. backward_passes_per_step: Number of backward passes to perform before calling hvd.allreduce. This allows accumulating updates over multiple mini-batches before reducing and applying them. op: The reduction operation to use when combining gradients across different ranks. gradient_predivide_factor: If op == Average, gradient_predivide_factor splits the averaging before and after the sum. Gradients are scaled by 1.0 / gradient_predivide_factor before the sum and gradient_predivide_factor / size after the sum. """ if gradient_predivide_factor != 1.0: if rocm_built(): raise ValueError('gradient_predivide_factor not supported yet with ROCm') if op != Average: raise ValueError('gradient_predivide_factor not supported with op != Average') if isinstance(optimizer, _LegacyOptimizer): if op == Adasum: return _DistributedAdasumOptimizer(optimizer, name, use_locking, device_dense, device_sparse, compression, backward_passes_per_step) else: if backward_passes_per_step > 1: raise ValueError('backward_passes_per_step>1 is not supported yet with ' 'op != Adasum') return _DistributedOptimizer(optimizer, name, use_locking, device_dense, device_sparse, compression, sparse_as_dense, op, gradient_predivide_factor) elif isinstance(optimizer, tf.keras.optimizers.Optimizer): if op == Adasum: raise ValueError('op == Adasum is not supported yet with Keras') if backward_passes_per_step > 1: raise ValueError('backward_passes_per_step > 1 is not supported yet with Keras') import horovod.tensorflow.keras as hvd_k return hvd_k.DistributedOptimizer(optimizer, name, device_dense, device_sparse, compression, sparse_as_dense, gradient_predivide_factor) else: raise ValueError('Provided optimizer doesn\'t inherit from either legacy ' 'TensorFlow or Keras optimizer: %s' % optimizer)
if hasattr(tf, 'GradientTape'): class _DistributedGradientTape(tf.GradientTape): def __init__(self, tape, device_dense, device_sparse, compression, sparse_as_dense, op, gradient_predivide_factor, persistent=False, watch_accessed_variables=True): if hasattr(tape, '_watch_accessed_variables'): super(self.__class__, self).__init__(persistent, watch_accessed_variables) else: super(self.__class__, self).__init__(persistent) self._tape = tape self._allreduce_grads = _make_allreduce_grads_fn( 'DistributedGradientTape', device_dense, device_sparse, compression, sparse_as_dense, op, gradient_predivide_factor) def gradient(self, target, sources, output_gradients=None): gradients = super(self.__class__, self).gradient(target, sources, output_gradients) return self._allreduce_grads(gradients)
[docs] def DistributedGradientTape(gradtape, device_dense='', device_sparse='', compression=Compression.none, sparse_as_dense=False, op=Average, gradient_predivide_factor=1.0): """A tape that wraps another tf.GradientTape, using an allreduce to combine gradient values before applying gradients to model weights. Args: gradtape: GradientTape to use for computing gradients and applying updates. device_dense: Device to be used for dense tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. device_sparse: Device to be used for sparse tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. compression: Compression algorithm used during allreduce to reduce the amount of data sent during each parameter update step. Defaults to not using compression. sparse_as_dense: Treat all sparse gradients as dense tensors. This can help improve performance and memory utilization if the original sparse gradient has high density. Defaults to false. op: The reduction operation to use when combining gradients across different ranks. gradient_predivide_factor: If op == Average, gradient_predivide_factor splits the averaging before and after the sum. Gradients are scaled by 1.0 / gradient_predivide_factor before the sum and gradient_predivide_factor / size after the sum. """ if gradient_predivide_factor != 1.0: if rocm_built(): raise ValueError('gradient_predivide_factor not supported yet with ROCm') if op != Average: raise ValueError('gradient_predivide_factor not supported with op != Average') cls = type(gradtape.__class__.__name__, (gradtape.__class__,), dict(_DistributedGradientTape.__dict__)) if hasattr(gradtape, '_watch_accessed_variables'): return cls(gradtape._tape, device_dense, device_sparse, compression, sparse_as_dense, op, gradient_predivide_factor, gradtape._persistent, gradtape._watch_accessed_variables) else: return cls(gradtape._tape, device_dense, device_sparse, compression, sparse_as_dense, op, gradient_predivide_factor, gradtape._persistent)