# Copyright 2022 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""adasum"""
from __future__ import absolute_import
import copy
import hashlib
import math
import mindspore.nn as nn
import mindspore.log as logger
from mindspore import context
from mindspore._checkparam import Validator as validator
from mindspore.nn.cell import Cell
from mindspore.common.parameter import ParameterTuple, Parameter
from mindspore.parallel._utils import _get_global_rank, _get_stage_device_num
from mindspore.ops import composite as C
from mindspore.ops import functional as F
from mindspore.ops import operations as P
from mindspore.ops.operations._inner_ops import Send, Receive
from mindspore.common.tensor import Tensor
from mindspore.common import dtype as mstype
from mindspore.communication.management import create_group
__all__ = ["AdaSumByDeltaWeightWrapCell", "AdaSumByGradWrapCell"]
MAX_NUM_HASH = 2 ** 31
_update_parameters = C.MultitypeFuncGraph("update_parameters")
_reshape_grads = C.MultitypeFuncGraph("reshape_grads")
@_update_parameters.register("Tensor", "Tensor", "Tensor", "Tensor", "Function")
def _update_parameters_adasum(delta_weight, update_delta_weight, parameter, old_parameter, reshape):
shape = F.shape(delta_weight)
update_delta_weight = reshape(update_delta_weight, shape)
new_parameter = old_parameter - update_delta_weight
return P.Assign()(parameter, new_parameter)
@_reshape_grads.register("Tensor", "Tensor", "Function")
def reshape_grads_adasum(grads, update_grads, reshape):
"""
Reshape gradient.
"""
shape = F.shape(grads)
update_grads = reshape(update_grads, shape)
return update_grads
def _send_before_receive(send_part, send, recv):
send_ok = send(send_part)
return recv(send_ok)
def _receive_before_send(send_part, send, recv):
receive_ok = recv(send_part)
send_part = F.depend(send_part, receive_ok)
return F.depend(receive_ok, send(send_part))
def _send_recv_res(left_send, recv_part, local_part, allreduce, parameter_divisibility, allreduce_node_num):
"""send result and receive result."""
if parameter_divisibility:
recv_part = P.Squeeze()(recv_part)
if F.shape(recv_part) is None:
recv_part = Tensor([recv_part])
local_part = F.depend(local_part, recv_part)
eps = 1e-12
scale_value = P.ReduceMax()(local_part) + eps
local_part_scale = local_part / scale_value
recv_part_scale = recv_part / scale_value
recv_part_scale = F.depend(recv_part_scale, local_part_scale)
value_0 = P.ReduceSum()(local_part_scale * recv_part_scale) + eps
if left_send:
value_1 = P.ReduceSum()(local_part_scale * local_part_scale) + eps
value_2 = P.ReduceSum()(recv_part_scale * recv_part_scale) + eps
else:
value_1 = P.ReduceSum()(recv_part_scale * recv_part_scale) + eps
value_2 = P.ReduceSum()(local_part_scale * local_part_scale) + eps
value_0 = allreduce(value_0)
value_1 = F.depend(allreduce(value_1), value_0)
value_2 = F.depend(allreduce(value_2), value_1)
if left_send:
res = (1 - (value_0 / (2 * value_1))) * local_part + (1 - (value_0 / (2 * value_2))) * recv_part
else:
res = (1 - (value_0 / (2 * value_1))) * recv_part + (1 - (value_0 / (2 * value_2))) * local_part
else:
res = allreduce(local_part)
res = res / allreduce_node_num
return res
_adasum_opt_forward = C.MultitypeFuncGraph("adasum_opt_forward")
_adasum_opt_rollback = C.MultitypeFuncGraph("adasum_opt_rollback")
@_adasum_opt_forward.register("Bool", "Function", "Bool", "Int64", "Function", "Function", "Tensor")
def _adasum_opt_forward_process(left_send, allreduce, parameter_divisibility, allreduce_node_num, send, recv, delta_w):
"""adasum optimizer process."""
if parameter_divisibility:
delta_w = P.Squeeze()(delta_w)
ori_len = F.shape(delta_w)[0]
divide_len = ori_len / 2
left_part = delta_w[:divide_len]
right_part = delta_w[divide_len:]
else:
left_part = delta_w
right_part = delta_w
if left_send:
if parameter_divisibility:
recv_part = _send_before_receive(left_part, send, recv)
else:
recv_part = right_part
update_delta_w = _send_recv_res(left_send, recv_part, right_part, allreduce, parameter_divisibility,
allreduce_node_num)
else:
if parameter_divisibility:
recv_part = _receive_before_send(right_part, send, recv)
else:
recv_part = left_part
update_delta_w = _send_recv_res(left_send, recv_part, left_part, allreduce, parameter_divisibility,
allreduce_node_num)
return update_delta_w
@_adasum_opt_rollback.register("Bool", "Bool", "Tensor", "Function", "Function")
def _adasum_opt_rollback_process(left_send, parameter_divisibility, delta_w, send, recv):
"""adasum optimizer rollback process."""
if parameter_divisibility:
if left_send:
recv_part = _send_before_receive(delta_w, send, recv)
else:
recv_part = _receive_before_send(delta_w, send, recv)
recv_part = P.Squeeze()(recv_part)
if F.shape(recv_part) is None:
recv_part = Tensor([recv_part])
if F.shape(delta_w) is None:
delta_w = Tensor([delta_w])
recv_part = P.Reshape()(recv_part, (-1,))
delta_w = P.Reshape()(delta_w, (-1,))
if left_send:
res = P.Concat()((recv_part, delta_w))
else:
res = P.Concat()((delta_w, recv_part))
else:
res = delta_w
return res
class _AdaSum(Cell):
r"""
The Adaptive Summation, or AdaSum, is a novel algorithm for improving distributed data
parallel training of Deep Learning models.
Inputs:
- **delta_weights** (Tuple(Tensor)) - Tuple of gradients.
- **parameters** (Tuple(Parameter)) - Tuple of current parameters.
- **old_parameters** (Tuple(Parameter)) - Tuple of last parameters.
Outputs:
- **adasum_parameters** (Tuple(Tensor)) - Tuple of parameters after adasum process.
"""
def __init__(self, rank, device_number, group_number, parameter_tuple):
super(_AdaSum, self).__init__()
self.rank = rank
self.device_number = device_number
self.group_number = group_number
self.parameter_tuple = parameter_tuple
self.calc_times = int(math.log(self.group_number, 2))
self.send_node = []
self.send_list_forward = []
self.recv_list_forward = []
self.send_list_rollback = []
self.recv_list_rollback = []
self.allreduce_list = []
self.parameter_divisibility_list = []
self.allreduce_node_num_list = []
self._generate_communication_op()
self.hyper_map = C.HyperMap()
self.update_reshape_list = []
for parameter in self.parameter_tuple:
reshape = P.Reshape().add_prim_attr("target_param", "adasum_delta_weight." + parameter.name)
self.update_reshape_list.append(reshape)
@staticmethod
def _hash(step, target, weights_index):
target = "tag" + str(step) + str(target) + str(weights_index)
target_hash = hashlib.sha1(target.encode()).hexdigest()
hash_res = int(int(target_hash, 16) % MAX_NUM_HASH)
return hash_res
def construct(self, delta_weights, parameters, old_parameters):
forward_weights = [delta_weights]
for i in range(self.calc_times):
process_weights = self.hyper_map(F.partial(_adasum_opt_forward, self.send_node[i]), self.allreduce_list[i],
self.parameter_divisibility_list[i], self.allreduce_node_num_list[i],
self.send_list_forward[i], self.recv_list_forward[i], forward_weights[-1])
forward_weights.append(process_weights)
for i in range(self.calc_times):
j = self.calc_times - i - 1
process_weights = self.hyper_map(F.partial(_adasum_opt_rollback, self.send_node[j]),
self.parameter_divisibility_list[j], forward_weights[j + 1],
self.send_list_rollback[j], self.recv_list_rollback[j])
forward_weights[j] = process_weights
adasum_parameters = self.hyper_map(F.partial(_update_parameters), delta_weights, forward_weights[0],
parameters, old_parameters, self.update_reshape_list)
return adasum_parameters
def _generate_communication_op(self):
"""generate communication op."""
last_delta_weights = []
fusion_attr = "origin_fusion"
if context.get_auto_parallel_context("parallel_mode") in ["data_parallel", "hybrid_parallel"]:
fusion_attr = "fusion"
for step in range(self.calc_times):
current_group = self.device_number * (2 ** step)
if (self.rank // current_group) % 2 == 0:
dest_target = self.rank + current_group
self.send_node.append(True)
else:
dest_target = self.rank - current_group
self.send_node.append(False)
send_left = []
send_right = []
recv_left = []
recv_right = []
allreduce_node_num = ()
left_delta_weights, right_delta_weights, delta_weights_divisibility = \
self._get_delta_weights_info(last_delta_weights)
self.parameter_divisibility_list.append(delta_weights_divisibility)
weights_index = 0
fusion_id = (step + 1) * 3
for shape, dtype, name in left_delta_weights:
send_tag = self._hash(step, self.rank, weights_index)
send = Send(sr_tag=send_tag, dest_rank=dest_target, group="hccl_world_group")
send.add_prim_attr(fusion_attr, fusion_id)
send.add_prim_attr("opposite_rank", dest_target)
send.add_prim_attr("target_param", name)
recv_tag = self._hash(step, dest_target, weights_index)
recv = Receive(sr_tag=recv_tag, src_rank=dest_target, shape=shape, dtype=dtype,
group="hccl_world_group")
recv.add_prim_attr(fusion_attr, fusion_id)
recv.add_prim_attr("opposite_rank", dest_target)
recv.add_prim_attr("target_param", name)
send_left.append(send)
recv_left.append(recv)
weights_index += 1
for shape, dtype, name in right_delta_weights:
send_tag = self._hash(step, self.rank, weights_index)
send = Send(sr_tag=send_tag, dest_rank=dest_target, group="hccl_world_group")
send.add_prim_attr(fusion_attr, fusion_id + 1)
send.add_prim_attr("opposite_rank", dest_target)
send.add_prim_attr("target_param", name)
recv_tag = self._hash(step, dest_target, weights_index)
recv = Receive(sr_tag=recv_tag, src_rank=dest_target, shape=shape, dtype=dtype,
group="hccl_world_group")
recv.add_prim_attr(fusion_attr, fusion_id + 1)
recv.add_prim_attr("opposite_rank", dest_target)
recv.add_prim_attr("target_param", name)
send_right.append(send)
recv_right.append(recv)
weights_index += 1
if self.send_node and self.send_node[-1]:
self.send_list_forward.append(send_left)
self.send_list_rollback.append(send_right)
self.recv_list_forward.append(recv_right)
self.recv_list_rollback.append(recv_left)
last_delta_weights = right_delta_weights
else:
self.send_list_forward.append(send_right)
self.send_list_rollback.append(send_left)
self.recv_list_forward.append(recv_left)
self.recv_list_rollback.append(recv_right)
last_delta_weights = left_delta_weights
param_allreduce_list = []
neighbor_ids = []
rank_ids = []
for index in range(2 ** (step + 1)):
node_rank = self.rank // self.device_number
double_d = 2 ** (step + 1)
neighbor_id = (node_rank // double_d * double_d + index) * self.device_number + \
self.rank % self.device_number
neighbor_ids.append(str(neighbor_id))
rank_ids.append(neighbor_id)
group_name = "-".join(neighbor_ids)
if context.get_auto_parallel_context("parallel_mode") in ["data_parallel", "hybrid_parallel"]:
create_group(group_name, rank_ids)
for parameter in self.parameter_tuple:
allreduce = P.AllReduce("sum", group_name)
allreduce.add_prim_attr("target_param", "adasum_delta_weight." + parameter.name)
allreduce.add_prim_attr(fusion_attr, fusion_id + 2)
allreduce.add_prim_attr("step", step)
param_allreduce_list.append(allreduce)
self.allreduce_list.append(param_allreduce_list)
for param_divisibility in delta_weights_divisibility:
if param_divisibility:
allreduce_node_num += (0,)
else:
allreduce_node_num += (2 ** (step + 1),)
self.allreduce_node_num_list.append(allreduce_node_num)
def _get_delta_weights_info(self, last_delta_weights):
"""get delta weights info."""
half_delta_weights = []
if last_delta_weights:
half_delta_weights = last_delta_weights
else:
for parameter in self.parameter_tuple:
new_shape = [int(x) for x in parameter.shape]
half_delta_weights.append((new_shape, parameter.dtype, "adasum_delta_weight." + parameter.name))
left_delta_weights = []
right_delta_weights = []
delta_weights_divisibility = ()
for shape, dtype, name in half_delta_weights:
left_shape = copy.deepcopy(shape)
right_shape = copy.deepcopy(shape)
divisibility_flag = False
for i, value in enumerate(shape):
if value > 1:
left_shape[i] = int(value // 2)
right_shape[i] = value - int(value // 2)
divisibility_flag = True
break
left_delta_weights.append((left_shape, dtype, name))
right_delta_weights.append((right_shape, dtype, name))
delta_weights_divisibility += (divisibility_flag,)
return left_delta_weights, right_delta_weights, delta_weights_divisibility
class _AdaSumByGrad(_AdaSum):
"""Apply adasum by gradients"""
def construct(self, grads):
forward_grads = [grads]
for i in range(self.calc_times):
process_weights = self.hyper_map(F.partial(_adasum_opt_forward, self.send_node[i]), self.allreduce_list[i],
self.parameter_divisibility_list[i], self.allreduce_node_num_list[i],
self.send_list_forward[i], self.recv_list_forward[i], forward_grads[-1])
forward_grads.append(process_weights)
for i in range(self.calc_times):
j = self.calc_times - i - 1
process_weights = self.hyper_map(F.partial(_adasum_opt_rollback, self.send_node[j]),
self.parameter_divisibility_list[j], forward_grads[j + 1],
self.send_list_rollback[j], self.recv_list_rollback[j])
forward_grads[j] = process_weights
update_grads = self.hyper_map(F.partial(_reshape_grads), grads, forward_grads[0],
self.update_reshape_list)
return update_grads
_get_delta_weight = C.MultitypeFuncGraph("_get_delta_weight")
_save_weight = C.MultitypeFuncGraph("_save_weight")
scale_mul = P.Mul().add_prim_attr("keep_alive", True)
_clone_weight = C.MultitypeFuncGraph("_clone_weight")
@_get_delta_weight.register("Tensor", "Tensor")
def _get_delta_weight_process(new_parameter, old_parameter):
delta_w = old_parameter - new_parameter
return delta_w
@_save_weight.register("Tensor", "Tensor")
def _save_weight_process(new_parameter, old_parameter):
return P.Assign()(new_parameter, old_parameter)
@_clone_weight.register("Tensor", "Tensor")
def _clone_weight_process(scale, weight):
return scale_mul(weight, scale)
def _parallel_check():
"""Parallel infos checking"""
if context.get_auto_parallel_context("parallel_mode") == "stand_alone":
raise RuntimeError("Stand alone mode is not supported to apply adasum.")
if context.get_auto_parallel_context("parallel_mode") in ["data_parallel", "hybrid_parallel"]:
logger.warning("For data parallel mode or hybrid parallel mode, "
"it is recommended to using mindspore.boost to enable adasum.")
if context.get_auto_parallel_context("enable_parallel_optimizer"):
raise RuntimeError("Currently, the optimizer shard is not supported with applying adasum.")
if context.get_auto_parallel_context("pipeline_stages") > 1:
raise RuntimeError("Currently, the pipeline parallel is not supported with applying adasum.")
stage_device_num = _get_stage_device_num()
if stage_device_num < 16 or (stage_device_num & (stage_device_num - 1) != 0):
raise RuntimeError("The device_num must be at least 16 and must be the power of 2 when applying adasum.")
[docs]class AdaSumByGradWrapCell(Cell):
r"""
Enable the adasum in "auto_parallel/semi_auto_parallel" mode.
The implementation of the Adaptive Summation (AdaSum) algorithm is calculated by gradients.
See the paper `AdaSum: Scaling Distributed Training with Adaptive Summation <https://arxiv.org/abs/2006.02924>`_.
.. math::
\begin{array}{ll}
w_{t+1}=w_{t} - \alpha \cdot Adasum(g_{1}, g_{2}) \\
w_{t+1}=w_{t} - \alpha \cdot [(1 - \frac{g_2^{T}\cdot g_1}{2\cdot \left \| g_1 \right \|^2 })\cdot g_1 +
(1 - \frac{g_1^{T}\cdot g_2}{2\cdot \left \| g_2 \right \|^2 })\cdot g_2] \\
\end{array}
In this implementation, :math:`g` represents the gradient of the weights,
and the subscripts represent different devices in the data-parallel dimension.
Note:
When using AdaSum, the number of traning cards needs to be a power of 2 and at least 16 cards are required.
Currently, the optimizer sharding and pipeline parallel is not supported when using AdaSum.
It is recommended to using AdaSumByGradWrapCell in semi auto parallel/auto parallel mode. In data parallel
mode, we recommend to using mindspore.boost to applying AdaSum.
Args:
optimizer (Union[Cell]): Optimizer for updating the weights. The construct function of the optimizer
requires only one input.
Inputs:
- **grads** (Tuple(Tensor)) - Tuple of gradients, same with the input of passed optimizer.
Raises:
RuntimeError: If `parallel_mode` uses `stand_alone` mode, AdaSum only supports use in distributed scenarios.
RuntimeError: If the optimizer parallel is used when using AdaSum.
RuntimeError: If the pipeline parallel is used when using AdaSum.
RuntimeError: If `device_num` is not a power of 2, or less than 16.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> from mindspore import nn
>>> from mindspore.nn import AdaSumByGradWrapCell
>>> net = Net()
>>> optim = AdaSumByGradWrapCell(nn.Momentum(params=net.trainable_params(), learning_rate=0.1, momentum=0.9))
>>> loss = nn.SoftmaxCrossEntropyWithLogits()
>>> model = Model(net, loss_fn=loss, optimizer=optim, metrics=None)
"""
def __init__(self, optimizer):
super(AdaSumByGradWrapCell, self).__init__(auto_prefix=False)
_device_number = 8
_parallel_check()
self.optimizer = optimizer
validator.check_value_type('optimizer', optimizer, (nn.Optimizer,))
self.parameters = optimizer.parameters
self.hyper_map = C.HyperMap()
group_number = _get_stage_device_num() // _device_number
self.grad_clone = ParameterTuple(self.parameters)
self.adasum = _AdaSumByGrad(_get_global_rank(), _device_number, group_number, self.grad_clone)
self.sync_tensor = Parameter(Tensor(0, dtype=mstype.int32))
def construct(self, grads):
adasum_res = self.adasum(grads)
sync_tensor = F.depend(self.sync_tensor, adasum_res)
sync_flag = P.AllReduce()(sync_tensor)
return F.depend(self.optimizer(adasum_res), sync_flag)
[docs]class AdaSumByDeltaWeightWrapCell(Cell):
r"""
Enable the adasum in "auto_parallel/semi_auto_parallel" mode.
The implementation of the Adaptive Summation (AdaSum) algorithm is calculated based on the difference of weights
before and after the updating of optimizer.
See the paper `AdaSum: Scaling Distributed Training with Adaptive Summation <https://arxiv.org/abs/2006.02924>`_.
.. math::
\begin{array}{ll}
w_{t+1}=w_{t} - \alpha \cdot Adasum(g_{1}, g_{2}) \\
w_{t+1}=w_{t} - \alpha \cdot [(1 - \frac{g_2^{T}\cdot g_1}{2\cdot \left \| g_1 \right \|^2 })\cdot g_1 +
(1 - \frac{g_1^{T}\cdot g_2}{2\cdot \left \| g_2 \right \|^2 })\cdot g_2] \\
\end{array}
In this implementation, :math:`g` represents the weight difference before and after the updating of optimizer,
and the subscripts represent different devices in the data parallel dimension.
Note:
When using AdaSum, the number of traning cards needs to be a power of 2 and at least 16 cards are required.
Currently, the optimizer sharding and pipeline parallel is not supported when using AdaSum.
It is recommended to using AdaSumByDeltaWeightWrapCell in semi auto parallel/auto parallel mode.
In data parallel mode, we recommend to using mindspore.boost to applying AdaSum.
Args:
optimizer (Union[Cell]): Optimizer for updating the weights. The construct function of the optimizer
requires only one input.
Inputs:
- **grads** (Tuple(Tensor)) - Tuple of gradients, same with the input of passed optimizer.
Raises:
RuntimeError: If `parallel_mode` uses `stand_alone` mode, AdaSum only supports use in distributed scenarios.
RuntimeError: If the optimizer parallel is used when using AdaSum.
RuntimeError: If the pipeline parallel is used when using AdaSum.
RuntimeError: If `device_num` is not a power of 2, or less than 16.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> from mindspore import nn
>>> from mindspore.nn import AdaSumByDeltaWeightWrapCell
>>> net = Net()
>>> optim = AdaSumByDeltaWeightWrapCell(nn.Momentum(params=net.trainable_params(),
... learning_rate=0.1, momentum=0.9))
>>> loss = nn.SoftmaxCrossEntropyWithLogits()
>>> model = Model(net, loss_fn=loss, optimizer=optim, metrics=None)
"""
def __init__(self, optimizer):
super(AdaSumByDeltaWeightWrapCell, self).__init__(auto_prefix=False)
_parallel_check()
self.optimizer = optimizer
validator.check_value_type('optimizer', optimizer, (nn.Optimizer,))
self.parameters = optimizer.parameters
self.hyper_map = C.HyperMap()
_device_number = 8
group_number = _get_stage_device_num() // _device_number
self.grad_clone = ParameterTuple(self.parameters)
self.adasum = _AdaSum(_get_global_rank(), _device_number, group_number, self.grad_clone)
self.sync_tensor = Parameter(Tensor(0, dtype=mstype.int32))
self.scale = Tensor(1.0, dtype=mstype.float32)
def construct(self, grads):
grad_clone = self.hyper_map(F.partial(_clone_weight, self.scale), self.parameters)
grads = F.depend(grads, grad_clone)
opt_result = self.optimizer(grads)
parameters = F.depend(self.parameters, opt_result)
delta_w = self.hyper_map(F.partial(_get_delta_weight), parameters, grad_clone)
adasum_res = self.adasum(delta_w, parameters, grad_clone)
sync_tensor = F.depend(self.sync_tensor, adasum_res)
sync_flag = P.AllReduce()(sync_tensor)
updated_weights = F.depend(parameters, sync_flag)
return updated_weights