!4404 update field split

Merge pull request !4404 from yangzhenzhang/update-field-split-in-forward-communication
This commit is contained in:
mindspore-ci-bot 2020-08-14 10:16:24 +08:00 committed by Gitee
commit 0f637fdc58
2 changed files with 199 additions and 0 deletions

View File

@ -550,6 +550,10 @@ RankList GetRankFromGroup(const Group &group) {
}
Status GatherV2PInfo::InferForwardCommunication() {
if (manual_split_) {
return SUCCESS;
}
forward_op_.clear();
auto param_strategy = strategy_->GetInputDim().at(0);
// don't split axis or target is not CPU, no need forward communication

View File

@ -0,0 +1,195 @@
# Copyright 2020 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.
# ============================================================================
import numpy as np
import pytest
import mindspore as ms
from mindspore import context, Tensor, Parameter
from mindspore.common.api import _executor
from mindspore.nn import Cell, TrainOneStepCell, LazyAdam
from mindspore.ops import operations as P
from mindspore.common.initializer import initializer
context.set_context(enable_sparse=True)
class Net(Cell):
def __init__(self,
strategy1=None,
strategy2=None,
strategy3=None,
axis=0,
init_flag=True,
split_tuple=(4, 4),
split_string="manual_split",
param_shape=(8, 8)):
super().__init__()
self.gatherv2 = P.EmbeddingLookup().set_strategy(strategy1)
self.gatherv2.add_prim_attr(split_string, split_tuple)
self.gatherv2.add_prim_attr("primitive_target", "CPU")
self.mul = P.Mul().set_strategy(strategy2)
self.reshape = P.Reshape()
self.matmul = P.MatMul().set_strategy(strategy3)
self.matmul.add_prim_attr("forward_reduce_scatter", True)
if init_flag:
self.param = Parameter(initializer("ones", param_shape, ms.float32), name="gatherv2_param")
else:
self.param = Parameter(Tensor(np.ones(param_shape), dtype=ms.float32), name="gatherv2_param")
self.mul_weight = Parameter(initializer("ones", (8, 8, 8), ms.float32), name="mul_weight")
self.matmul_weight = Parameter(initializer("ones", (64, 16), ms.float32), name="matmul_weight")
self.axis = axis
def construct(self, x, b):
out = self.gatherv2(self.param, x, self.axis)
out = self.mul(out, b)
return out
_x = Tensor(np.ones([8, 8]), dtype=ms.int32)
_b = Tensor(np.ones([8, 8, 8]), dtype=ms.float32)
def compile_net(net):
context.set_context(save_graphs=True)
optimizer = LazyAdam(net.trainable_params(), learning_rate=0.1)
optimizer.sparse_opt.add_prim_attr("primitive_target", "CPU")
train_net = TrainOneStepCell(net, optimizer)
train_net.set_auto_parallel()
_executor.compile(train_net, _x, _b, auto_parallel_mode=True)
context.reset_auto_parallel_context()
def test_normal_split():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
strategy1 = ((2, 1), (1, 2))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3)
compile_net(net)
def test_normal_split2():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=4, global_rank=0)
strategy1 = ((4, 1), (1, 4))
strategy2 = ((1, 4, 1), (1, 4, 1))
strategy3 = ((1, 4), (4, 1))
net = Net(strategy1, strategy2, strategy3, split_tuple=(10, 20, 30, 4), param_shape=(64, 8))
compile_net(net)
def test_normal_split3():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=32, global_rank=17)
strategy1 = ((4, 8), (1, 4))
strategy2 = ((1, 4, 8), (1, 4, 8))
strategy3 = ((1, 32), (32, 1))
net = Net(strategy1, strategy2, strategy3, split_tuple=(10, 20, 30, 4), param_shape=(64, 8))
compile_net(net)
def test_normal_split_with_offset():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
strategy1 = ((2, 1), (1, 2))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3, split_string="manual_split_with_offset", split_tuple=((4, 0), (4, 4)))
compile_net(net)
def test_auto_parallel_error():
context.set_context(save_graphs=True)
context.set_auto_parallel_context(parallel_mode="auto_parallel", device_num=2, global_rank=0)
net = Net()
with pytest.raises(RuntimeError):
compile_net(net)
def test_axis_error():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
strategy1 = ((2, 1), (1, 2))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3, axis=1)
with pytest.raises(RuntimeError):
compile_net(net)
def test_strategy_error():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
strategy1 = ((4, 1), (8, 1))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3)
with pytest.raises(RuntimeError):
compile_net(net)
def test_strategy_error2():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
strategy1 = ((4, 1), (1, 8))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3)
with pytest.raises(RuntimeError):
compile_net(net)
def test_strategy_error3():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
strategy1 = ((2, 1), (1, 2))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3)
with pytest.raises(RuntimeError):
compile_net(net)
def test_strategy_error4():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
strategy1 = ((2, 8), (1, 2))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3)
with pytest.raises(RuntimeError):
compile_net(net)
def test_strategy_error5():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=4, global_rank=0)
strategy1 = ((4, 1), (1, 4))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3)
with pytest.raises(RuntimeError):
compile_net(net)
def test_split_tuple_error():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
strategy1 = ((2, 1), (1, 2))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3, split_tuple=((5, 0), (5, 5)))
with pytest.raises(RuntimeError):
compile_net(net)
def test_parameter_use_tensor_error():
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
strategy1 = ((2, 1), (1, 2))
strategy2 = ((1, 2, 1), (1, 2, 1))
strategy3 = ((1, 2), (2, 1))
net = Net(strategy1, strategy2, strategy3, init_flag=False)
with pytest.raises(RuntimeError):
compile_net(net)