# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
# Modifications copyright (C) 2019 Uber Technologies, Inc.
# Modifications copyright Microsoft
# Modifications copyright (C) 2020, NVIDIA CORPORATION. All rights reserved.
#
# 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.
# =============================================================================
"""Inter-process communication using MPI."""
import re
import tensorflow as tf
from tensorflow.python.framework import load_library
from tensorflow.python.framework import ops
from tensorflow.python.platform import resource_loader
from horovod.common.util import check_installed_version, get_ext_suffix, \
get_average_backwards_compatibility_fun, gpu_available, num_rank_is_power_2
from horovod.common.basics import HorovodBasics as _HorovodBasics
from horovod.common.process_sets import _setup as _setup_process_sets
from horovod.common.process_sets import ProcessSet, global_process_set, add_process_set, remove_process_set, \
_temp_process_set_object
from horovod.tensorflow.util import _executing_eagerly
def _load_library(name):
"""Loads a .so file containing the specified operators.
Args:
name: The name of the .so file to load.
Raises:
NotFoundError if were not able to load .so file.
"""
filename = resource_loader.get_path_to_datafile(name)
library = load_library.load_op_library(filename)
return library
# Check possible symbol not found error from tensorflow version mismatch
try:
MPI_LIB = _load_library('mpi_lib' + get_ext_suffix())
except Exception as e:
check_installed_version('tensorflow', tf.__version__, e)
raise e
else:
check_installed_version('tensorflow', tf.__version__)
_basics = _HorovodBasics(__file__, 'mpi_lib')
# import basic methods
shutdown = _basics.shutdown
is_initialized = _basics.is_initialized
start_timeline = _basics.start_timeline
stop_timeline = _basics.stop_timeline
size = _basics.size
local_size = _basics.local_size
cross_size = _basics.cross_size
rank = _basics.rank
local_rank = _basics.local_rank
cross_rank = _basics.cross_rank
mpi_threads_supported = _basics.mpi_threads_supported
mpi_enabled = _basics.mpi_enabled
mpi_built = _basics.mpi_built
gloo_enabled = _basics.gloo_enabled
gloo_built = _basics.gloo_built
nccl_built = _basics.nccl_built
ddl_built = _basics.ddl_built
ccl_built = _basics.ccl_built
cuda_built = _basics.cuda_built
rocm_built = _basics.rocm_built
# import reduction op values
Average = _basics.Average
Sum = _basics.Sum
Adasum = _basics.Adasum
Min = _basics.Min
Max = _basics.Max
Product = _basics.Product
def init(*args, **kwargs):
_basics.init(*args, **kwargs)
# Call set up again to make sure the basics is in sync
_setup_process_sets(_basics)
is_homogeneous = _basics.is_homogeneous
handle_average_backwards_compatibility = get_average_backwards_compatibility_fun(_basics)
check_num_rank_power_of_2 = num_rank_is_power_2
_setup_process_sets(_basics)
# This function will create a default device map which includes all visible devices.
# Please run this function in a subprocess
def _check_has_gpu():
import tensorflow as tf
return tf.test.is_gpu_available()
def _normalize_name(name):
"""Normalizes operation name to TensorFlow rules."""
return re.sub('[^a-zA-Z0-9_]', '_', name)
def _allreduce(tensor, name=None, op=Sum, prescale_factor=1.0, postscale_factor=1.0,
ignore_name_scope=False, process_set=global_process_set):
"""An op which reduces an input tensor over all the Horovod processes. The
default reduction is a sum.
The reduction operation is keyed by the name of the op. The tensor type and
shape must be the same on all Horovod processes for a given name. The reduction
will not start until all processes are ready to send and receive the tensor.
Returns:
A tensor of the same shape and type as `tensor`, summed across all
processes.
"""
if name is None and not _executing_eagerly():
name = 'HorovodAllreduce_%s' % _normalize_name(tensor.name)
return MPI_LIB.horovod_allreduce(tensor, name=name, reduce_op=op,
prescale_factor=prescale_factor,
postscale_factor=postscale_factor,
ignore_name_scope=ignore_name_scope,
process_set_id=process_set.process_set_id)
@ops.RegisterGradient('HorovodAllreduce')
def _allreduce_grad(op, grad):
"""Gradient for allreduce op.
Args:
op: An operation.
grad: `Tensor` gradient with respect to the output of the op.
Returns:
The gradient with respect to the input of the op.
"""
reduce_op = op.get_attr('reduce_op')
prescale_factor = op.get_attr('prescale_factor')
postscale_factor = op.get_attr('postscale_factor')
ignore_name_scope = op.get_attr('ignore_name_scope')
process_set_id = op.get_attr('process_set_id')
return _allreduce(grad, op=reduce_op, prescale_factor=prescale_factor,
postscale_factor=postscale_factor,
ignore_name_scope=ignore_name_scope,
process_set=_temp_process_set_object(process_set_id))
def _grouped_allreduce(tensors, name=None, op=Sum, prescale_factor=1.0, postscale_factor=1.0,
ignore_name_scope=False, process_set=global_process_set):
"""An op which reduces input tensors over all the Horovod processes. The
default reduction is a sum.
The reduction operations are keyed by the name of the op. Reductions are
performed across tensors in the same list position. The tensor type and
shape must be the same on all Horovod processes for tensors sharing
positions in the input tensor list. The reduction will not start until all
processes are ready to send and receive the tensors.
Returns:
A list of tensors of the same shape and type as those in `tensors`,
summed across all processes.
"""
if name is None and not _executing_eagerly():
name = _normalize_name('HorovodGroupedAllreduce_%s_%s' % (tensors[0].name, tensors[-1].name))
return MPI_LIB.horovod_grouped_allreduce(tensors, name=name, reduce_op=op,
prescale_factor=prescale_factor,
postscale_factor=postscale_factor,
ignore_name_scope=ignore_name_scope,
process_set_id=process_set.process_set_id)
@ops.RegisterGradient('HorovodGroupedAllreduce')
def _grouped_allreduce_grad(op, *grads):
"""Gradient for the grouped allreduce op.
Args:
op: An operation.
grads: List of `Tensor` gradients with respect to the outputs of the op.
Returns:
The gradients with respect to the inputs of the op.
"""
reduce_op = op.get_attr('reduce_op')
prescale_factor = op.get_attr('prescale_factor')
postscale_factor = op.get_attr('postscale_factor')
ignore_name_scope = op.get_attr('ignore_name_scope')
process_set_id = op.get_attr('process_set_id')
# TODO(joshr): should this be done as separate allreduce ops?
return _grouped_allreduce(list(grads), op=reduce_op, prescale_factor=prescale_factor,
postscale_factor=postscale_factor,
ignore_name_scope=ignore_name_scope,
process_set=_temp_process_set_object(process_set_id))
[docs]def allgather(tensor, name=None, ignore_name_scope=False, process_set=global_process_set):
"""An op which concatenates the input tensor with the same input tensor on
all other Horovod processes.
The concatenation is done on the first dimension, so the input tensors on the
different processes must have the same rank and shape, except for the first
dimension, which is allowed to be different.
Returns:
A tensor of the same type as `tensor`, concatenated on dimension zero
across all processes. The shape is identical to the input shape, except for
the first dimension, which may be greater and is the sum of all first
dimensions of the tensors in different Horovod processes.
"""
if name is None and not _executing_eagerly():
name = 'HorovodAllgather_%s' % _normalize_name(tensor.name)
return MPI_LIB.horovod_allgather(tensor, name=name,
ignore_name_scope=ignore_name_scope,
process_set_id=process_set.process_set_id)
@ops.RegisterGradient('HorovodAllgather')
def _allgather_grad(op, grad):
"""Gradient for allgather op.
Args:
op: An operation.
grad: `Tensor` gradient with respect to the output of the op.
Returns:
The gradient with respect to the input of the op.
"""
ignore_name_scope = op.get_attr('ignore_name_scope')
process_set_id = op.get_attr('process_set_id')
temp_process_set_object = _temp_process_set_object(process_set_id)
grad = _allreduce(grad, op=Average, ignore_name_scope=ignore_name_scope,
process_set=temp_process_set_object)
with tf.device('/cpu:0'):
# Keep the tensor of split sizes on CPU.
x = op.inputs[0]
d = tf.shape(x)
d = tf.reshape(d[0], [1])
s = temp_process_set_object.size()
d = tf.reshape(
allgather(d, ignore_name_scope=ignore_name_scope, process_set=temp_process_set_object),
[s])
splits = tf.split(grad, num_or_size_splits=d, axis=0)
return splits[temp_process_set_object.rank()]
[docs]def grouped_allgather(tensors, name=None, ignore_name_scope=False, process_set=global_process_set):
"""An op which concatenates the input tensor with the same input tensor on
all other Horovod processes.
The concatenation is done on the first dimension, so the input tensors on the
different processes must have the same rank and shape, except for the first
dimension, which is allowed to be different.
Returns:
A list of tensors of the same rank and type as `tensor`, concatenated on
dimension zero across all processes. For each returned tensor the shape
is identical to the input shape, except for the first dimension, which
may be greater and is the sum of all first dimensions of the tensors
in different Horovod processes.
"""
if name is None and not _executing_eagerly():
name = _normalize_name('HorovodGroupedAllgather_%s_%s' % (tensors[0].name, tensors[-1].name))
return MPI_LIB.horovod_grouped_allgather(tensors, name=name,
ignore_name_scope=ignore_name_scope,
process_set_id=process_set.process_set_id)
@ops.RegisterGradient('HorovodGroupedAllgather')
def _grouped_allgather_grad(op, *grads):
"""Gradient for the grouped allgather op.
Args:
op: An operation.
grad: List of `Tensor` gradients with respect to the outputs of the op.
Returns:
The gradients with respect to the inputs of the op.
"""
ignore_name_scope = op.get_attr('ignore_name_scope')
process_set_id = op.get_attr('process_set_id')
temp_process_set_object = _temp_process_set_object(process_set_id)
# Not using _grouped_allreduce here because all its input tensors need to be of the same dtype, but TensorFlow may
# give us float32 zero gradients here despite non-zero gradients being, e.g, float64.
grads = [_allreduce(g, op=Average, ignore_name_scope=ignore_name_scope, process_set=temp_process_set_object)
for g in grads]
split_sizes = []
with tf.device('/cpu:0'):
# Keep the split size tensors on CPU.
for x in op.inputs:
d = tf.shape(x)
d = tf.reshape(d[0], [1])
s = temp_process_set_object.size()
d = tf.reshape(
allgather(d, ignore_name_scope=ignore_name_scope, process_set=temp_process_set_object),
[s])
split_sizes.append(d)
splits_list = [tf.split(grad, num_or_size_splits=d, axis=0)
for grad, d in zip(grads, split_sizes)]
return [splits[temp_process_set_object.rank()] for splits in splits_list]
[docs]def broadcast(tensor, root_rank, name=None, ignore_name_scope=False, process_set=global_process_set):
"""An op which broadcasts the input tensor on root rank to the same input tensor
on all other Horovod processes.
The broadcast operation is keyed by the name of the op. The tensor type and
shape must be the same on all Horovod processes for a given name. The broadcast
will not start until all processes are ready to send and receive the tensor.
Returns:
A tensor of the same shape and type as `tensor`, with the value broadcasted
from root rank.
"""
if name is None and not _executing_eagerly():
name = 'HorovodBroadcast_%s' % _normalize_name(tensor.name)
return MPI_LIB.horovod_broadcast(tensor, name=name, root_rank=root_rank,
ignore_name_scope=ignore_name_scope,
process_set_id=process_set.process_set_id)
@ops.RegisterGradient('HorovodBroadcast')
def _broadcast_grad(op, grad):
"""Gradient for broadcast op.
Args:
op: An operation.
grad: `Tensor` gradient with respect to the output of the op.
Returns:
The gradient with respect to the input of the op.
"""
root_rank = op.get_attr('root_rank')
ignore_name_scope = op.get_attr('ignore_name_scope')
process_set_id = op.get_attr('process_set_id')
grad_reduced = _allreduce(grad, op=Average,
ignore_name_scope=ignore_name_scope,
process_set=_temp_process_set_object(process_set_id))
if rank() != root_rank:
return grad_reduced * 0
return grad_reduced
[docs]def broadcast_(variables, root_rank, name=None, process_set=global_process_set):
"""An op which broadcasts the input variables from the root rank to the same
input variables on all other Horovod processes. The operation is performed
in-place.
The broadcast operation is keyed by the name of the op combined with the names
of the variables. The variable type and shape must be the same on all Horovod
processes for any given name. The broadcast will not start until all processes
are ready to send and receive all variables. In each process all variables need
to be located on the same device (CPU or GPU).
Note: This is only supported with TensorFlow 2.6 or later.
Returns:
The tensor values of the updated `variables` as broadcasted from root rank.
"""
from packaging import version
if version.parse(tf.__version__) < version.parse('2.6.0'):
raise NotImplementedError("In-place broadcasts are only supported with TensorFlow 2.6 or later")
from tensorflow.python.ops import resource_variable_ops
if all(resource_variable_ops.is_resource_variable(var) for var in variables):
with tf.control_dependencies(
[MPI_LIB.horovod_broadcast_inplace_resource([var.handle for var in variables],
variable_names=[var.name for var in variables],
name=name,
root_rank=root_rank,
process_set_id=process_set.process_set_id)]):
return [var.read_value() for var in variables]
elif all(not resource_variable_ops.is_resource_variable(var) for var in variables):
return MPI_LIB.horovod_broadcast_inplace(variables, variable_names=[var.name for var in variables],
name=name, root_rank=root_rank,
process_set_id=process_set.process_set_id)
else:
raise ValueError("All variables passed to broadcast_() should be of the same kind, resource or reference")
[docs]def alltoall(tensor, splits=None, name=None, ignore_name_scope=False, process_set=global_process_set):
"""An op that scatters slices of the input tensor to all other Horovod processes
and returns a tensor of gathered slices from all other Horovod processes.
The slicing is done on the first dimension, so the input tensors on the
different processes must have the same rank and shape, except for the first
dimension, which is allowed to be different.
Arguments:
tensor: A tensor to distribute with alltoall.
splits: A tensor of integers in rank order describing how many
elements in `tensor` to send to each worker. Splitting is
applied along the first dimension of `tensor`. If `splits` is
not provided, the first dimension is split equally by the
number of Horovod processes.
name: A name of the alltoall operation.
ignore_name_scope: If True, ignores any outer name scope applied by
TensorFlow in the name used by the Horovod operation.
Returns:
1) A tensor of the same type as `tensor`, concatenated on dimension zero
across all processes. The shape is identical to the input shape, except for
the first dimension, which may be greater and is the sum of all first
dimensions of the gathered tensor slices from different Horovod processes.
2) If `splits` has been provided: A tensor of integers in rank order
describing how many elements in the output tensor have been received
from each worker.
"""
# If splits not provided, create empty tensor as placeholder
splits_ = tf.convert_to_tensor(splits) if splits is not None else tf.constant([], dtype=tf.int32)
if name is None and not _executing_eagerly():
name = 'HorovodAlltoall_%s' % _normalize_name(tensor.name)
output, rsplits = MPI_LIB.horovod_alltoall(tensor, splits=splits_, name=name,
ignore_name_scope=ignore_name_scope,
process_set_id=process_set.process_set_id)
return (output, rsplits) if splits is not None else output
@ops.RegisterGradient('HorovodAlltoall')
def _alltoall_grad(op, grad_wrt_output, grad_wrt_received_splits):
"""Gradient for alltoall op.
Args:
op: Original operation.
grad_wrt_output: `Tensor` gradient with respect to the output of the op.
grad_wrt_received_splits: dead argument (integer output)
Returns:
The gradient with respect to the input of the op.
"""
ignore_name_scope = op.get_attr('ignore_name_scope')
process_set_id = op.get_attr('process_set_id')
recvsplits = op.outputs[1]
grad_wrt_tensor, _ = alltoall(grad_wrt_output, splits=recvsplits, ignore_name_scope=ignore_name_scope,
process_set=_temp_process_set_object(process_set_id))
grad_wrt_splits = None # not differentiable (integer variable)
return [grad_wrt_tensor, grad_wrt_splits]
def _reducescatter(tensor, name=None, op=Sum, ignore_name_scope=False,
process_set=global_process_set, prescale_factor=1.0,
postscale_factor=1.0):
"""An op which reduces an input tensor over all the Horovod processes, then
scatters the result across all the Horovod processes. The default reduction
is a sum.
The reduction operation is keyed by the name of the op. The tensor type and
shape must be the same on all Horovod processes for a given name. The reduction
will not start until all processes are ready to send and receive the tensor.
Returns:
A tensor of the same rank and type as `tensor`. The shape is identical to the
input shape, except for the first dimension, which will be divided across
the different Horovod processes.
"""
if name is None and not _executing_eagerly():
name = 'HorovodReducescatter_%s' % _normalize_name(tensor.name)
return MPI_LIB.horovod_reducescatter(tensor, name=name, reduce_op=op,
ignore_name_scope=ignore_name_scope,
process_set_id=process_set.process_set_id,
prescale_factor=prescale_factor,
postscale_factor=postscale_factor)
@ops.RegisterGradient('HorovodReducescatter')
def _reducescatter_grad(op, grad):
"""Gradient for reducescatter op.
Args:
op: An operation.
grad: `Tensor` gradient with respect to the output of the op.
Returns:
The gradient with respect to the input of the op.
"""
ignore_name_scope = op.get_attr('ignore_name_scope')
process_set_id = op.get_attr('process_set_id')
reduce_op = op.get_attr('reduce_op')
prescale_factor = op.get_attr('prescale_factor')
postscale_factor = op.get_attr('postscale_factor')
process_set = _temp_process_set_object(process_set_id)
if reduce_op == Sum:
grad *= process_set.size()
if prescale_factor != 1.0:
grad *= prescale_factor
if postscale_factor != 1.0:
grad *= postscale_factor
return allgather(grad, ignore_name_scope=ignore_name_scope,
process_set=process_set)
def _grouped_reducescatter(tensors, name=None, op=Sum, ignore_name_scope=False,
process_set=global_process_set,
prescale_factor=1.0, postscale_factor=1.0):
"""An op which sums an input tensor over all the Horovod processes, then
scatters the result across all the Horovod processes.
The reduction operations are keyed by the name of the op. Reductions are
performed across tensors in the same list position. The tensor type and
shape must be the same on all Horovod processes for tensors sharing
positions in the input tensor list. The reduction will not start until all
processes are ready to send and receive the tensors.
Returns:
A list of tensors of the same rank and type as those in `tensors`.
For each returned tensor the shape is identical to the corresponding
input shape, except for the first dimension, which will be divided across
the different Horovod processes.
"""
if name is None and not _executing_eagerly():
name = _normalize_name('HorovodGroupedReducescatter_%s_%s' % (tensors[0].name, tensors[-1].name))
return MPI_LIB.horovod_grouped_reducescatter(tensors, name=name, reduce_op=op,
ignore_name_scope=ignore_name_scope,
process_set_id=process_set.process_set_id,
prescale_factor=prescale_factor,
postscale_factor=postscale_factor)
@ops.RegisterGradient('HorovodGroupedReducescatter')
def _grouped_reducescatter_grad(op, *grads):
"""Gradient for the grouped reducescatter op.
Args:
op: An operation.
grads: List of `Tensor` gradients with respect to the outputs of the op.
Returns:
The gradients with respect to the inputs of the op.
"""
ignore_name_scope = op.get_attr('ignore_name_scope')
process_set_id = op.get_attr('process_set_id')
reduce_op = op.get_attr('reduce_op')
prescale_factor = op.get_attr('prescale_factor')
postscale_factor = op.get_attr('postscale_factor')
process_set = _temp_process_set_object(process_set_id)
if reduce_op == Sum:
grads = [grad * process_set.size() for grad in grads]
if prescale_factor != 1.0:
grads = [grad * prescale_factor for grad in grads]
if postscale_factor != 1.0:
grads = [grad * postscale_factor for grad in grads]
# Not using grouped_allgather here because all its input tensors need to be of the same dtype, but TensorFlow may
# give us float32 zero gradients here despite non-zero gradients being, e.g, float64.
return [allgather(g, ignore_name_scope=ignore_name_scope, process_set=process_set) for g in grads]
[docs]def join():
"""An op to indicate that the rank finished processing data.
All ranks that did not call join() continue to process allreduce
operations. This op is not done before all ranks have joined.
Returns:
An integer scalar containing the the last rank that joined.
"""
return MPI_LIB.horovod_join()
[docs]def size_op(process_set_id=0, name=None):
"""An op that returns the number of Horovod processes.
This operation determines the return value at the graph execution time,
rather than at the graph construction time, and so allows for a graph to be
constructed in a different environment than where it will be executed.
Returns:
An integer scalar containing the number of Horovod processes.
"""
return MPI_LIB.horovod_size(process_set_id=process_set_id, name=name)
ops.NotDifferentiable('HorovodSize')
[docs]def process_set_included_op(process_set_id=0, name=None):
"""An op that 0 or 1 depending on whether the current process is
included in the specified process set or an error code:
HOROVOD_PROCESS_SET_ERROR_INIT if Horovod is not initialized,
HOROVOD_PROCESS_SET_ERROR_UNKNOWN_SET if the process set is unknown.
This operation determines the return value at the graph execution time,
rather than at the graph construction time, and so allows for a graph to be
constructed in a different environment than where it will be executed.
Returns:
An integer scalar with value 0, 1, or an error code.
"""
return MPI_LIB.horovod_process_set_included(process_set_id=process_set_id, name=name)
ops.NotDifferentiable('HorovodProcessSetIncluded')
[docs]def local_size_op(name=None):
"""An op that returns the number of Horovod processes within the
node the current process is running on.
This operation determines the return value at the graph execution time,
rather than at the graph construction time, and so allows for a graph to be
constructed in a different environment than where it will be executed.
Returns:
An integer scalar containing the number of local Horovod processes.
"""
return MPI_LIB.horovod_local_size(name=name)
ops.NotDifferentiable('HorovodLocalSize')
[docs]def rank_op(name=None):
"""An op that returns the Horovod rank of the calling process.
This operation determines the return value at the graph execution time,
rather than at the graph construction time, and so allows for a graph to be
constructed in a different environment than where it will be executed.
Returns:
An integer scalar with the Horovod rank of the calling process.
"""
return MPI_LIB.horovod_rank(name=name)
ops.NotDifferentiable('HorovodRank')
[docs]def local_rank_op(name=None):
"""An op that returns the local Horovod rank of the calling process, within the
node that it is running on. For example, if there are seven processes running
on a node, their local ranks will be zero through six, inclusive.
This operation determines the return value at the graph execution time,
rather than at the graph construction time, and so allows for a graph to be
constructed in a different environment than where it will be executed.
Returns:
An integer scalar with the local Horovod rank of the calling process.
"""
return MPI_LIB.horovod_rank(name=name)
ops.NotDifferentiable('HorovodLocalRank')
