Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
Paddle
提交
3ec289a6
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2299
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
提交
3ec289a6
编写于
12月 31, 2019
作者:
W
WangXi
提交者:
gongweibao
12月 31, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix sync_batch_norm hang in fleet (#21838)
上级
a0b53376
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
65 addition
and
6 deletion
+65
-6
python/paddle/fluid/incubate/fleet/collective/__init__.py
python/paddle/fluid/incubate/fleet/collective/__init__.py
+11
-0
python/paddle/fluid/tests/unittests/CMakeLists.txt
python/paddle/fluid/tests/unittests/CMakeLists.txt
+14
-5
python/paddle/fluid/tests/unittests/dist_test.sh
python/paddle/fluid/tests/unittests/dist_test.sh
+9
-1
python/paddle/fluid/tests/unittests/test_dist_base.py
python/paddle/fluid/tests/unittests/test_dist_base.py
+5
-0
python/paddle/fluid/tests/unittests/test_dist_mnist_fleetapi.py
.../paddle/fluid/tests/unittests/test_dist_mnist_fleetapi.py
+26
-0
未找到文件。
python/paddle/fluid/incubate/fleet/collective/__init__.py
浏览文件 @
3ec289a6
...
...
@@ -337,6 +337,17 @@ class CollectiveOptimizer(DistributedOptimizer):
"with multi nccl comm, please export FLAGS_sync_nccl_allreduce = 0"
)
# NOTE. open sync_batch_norm will hang when use multi num_threads
sync_batch_norm
=
self
.
_strategy
.
sync_batch_norm
if
sync_batch_norm
is
not
None
and
sync_batch_norm
is
True
:
self
.
_strategy
.
nccl_comm_num
=
1
self
.
_strategy
.
use_hierarchical_allreduce
=
False
exec_strategy
.
num_threads
=
1
logging
.
warn
(
"use sync_batch_norm will hang when set num_threads > 1, so "
"set num_threads=1, nccl_comm_num=1, use_hierarchical_allreduce=False."
)
if
self
.
print_config
:
print
(
"node_num:"
,
node_num
,
"num_threads:"
,
exec_strategy
.
num_threads
,
"use_hierarchical_allreduce:"
,
...
...
python/paddle/fluid/tests/unittests/CMakeLists.txt
浏览文件 @
3ec289a6
...
...
@@ -128,11 +128,20 @@ function(bash_test_modules TARGET_NAME)
set
(
timeout
${
bash_test_modules_TIMEOUT
}
)
endif
()
add_test
(
NAME
${
TARGET_NAME
}
COMMAND
${
CMAKE_COMMAND
}
-E env PYTHONPATH=
${
PADDLE_BINARY_DIR
}
/python
TEST_TARGET_NAME=
${
TARGET_NAME
}
TEST_TIMEOUT=
${
timeout
}
${
bash_test_modules_ENVS
}
bash
${
CMAKE_CURRENT_BINARY_DIR
}
/
${
bash_test_modules_MODULES
}
WORKING_DIRECTORY
${
CMAKE_CURRENT_BINARY_DIR
}
)
if
(
WITH_COVERAGE
)
add_test
(
NAME
${
TARGET_NAME
}
COMMAND
${
CMAKE_COMMAND
}
-E env PYTHONPATH=
${
PADDLE_BINARY_DIR
}
/python
TEST_TARGET_NAME=
${
TARGET_NAME
}
TEST_TIMEOUT=
${
timeout
}
${
bash_test_modules_ENVS
}
WITH_COVERAGE=ON COVERAGE_FILE=
${
PADDLE_BINARY_DIR
}
/python-coverage.data
bash
${
CMAKE_CURRENT_BINARY_DIR
}
/
${
bash_test_modules_MODULES
}
WORKING_DIRECTORY
${
CMAKE_CURRENT_BINARY_DIR
}
)
else
()
add_test
(
NAME
${
TARGET_NAME
}
COMMAND
${
CMAKE_COMMAND
}
-E env PYTHONPATH=
${
PADDLE_BINARY_DIR
}
/python
TEST_TARGET_NAME=
${
TARGET_NAME
}
TEST_TIMEOUT=
${
timeout
}
${
bash_test_modules_ENVS
}
bash
${
CMAKE_CURRENT_BINARY_DIR
}
/
${
bash_test_modules_MODULES
}
WORKING_DIRECTORY
${
CMAKE_CURRENT_BINARY_DIR
}
)
endif
()
if
(
bash_test_modules_SERIAL
)
set_property
(
TEST
${
TARGET_NAME
}
PROPERTY RUN_SERIAL 1
)
...
...
python/paddle/fluid/tests/unittests/dist_test.sh
浏览文件 @
3ec289a6
...
...
@@ -22,7 +22,15 @@ rm -f ${name}_*.log
# start the unit test
run_time
=
$((
$TEST_TIMEOUT
-
10
))
echo
"run_time:
${
run_time
}
"
timeout
-s
SIGKILL
${
run_time
}
python
-u
${
name
}
.py
>
${
name
}
_run.log 2>&1
if
[[
${
WITH_COVERAGE
}
==
"ON"
]]
;
then
PYTHON_EXEC
=
"python -u -m coverage run --branch -p "
else
PYTHON_EXEC
=
"python -u "
fi
timeout
-s
SIGKILL
${
run_time
}
${
PYTHON_EXEC
}
${
name
}
.py
>
${
name
}
_run.log 2>&1
exit_code
=
$?
if
[[
$exit_code
-eq
0
]]
;
then
exit
0
...
...
python/paddle/fluid/tests/unittests/test_dist_base.py
浏览文件 @
3ec289a6
...
...
@@ -137,6 +137,8 @@ class TestDistRunnerBase(object):
dist_strategy
.
use_local_sgd
=
True
if
args
.
ut4grad_allreduce
:
dist_strategy
.
_ut4grad_allreduce
=
True
if
args
.
sync_batch_norm
:
dist_strategy
.
sync_batch_norm
=
True
role
=
role_maker
.
PaddleCloudRoleMaker
(
is_collective
=
True
)
fleet
.
init
(
role
)
...
...
@@ -487,6 +489,7 @@ def runtime_main(test_class):
required
=
False
,
type
=
bool
,
default
=
False
)
parser
.
add_argument
(
'--sync_batch_norm'
,
action
=
'store_true'
)
args
=
parser
.
parse_args
()
...
...
@@ -837,6 +840,8 @@ class TestDistBase(unittest.TestCase):
tr_cmd
+=
" --use_local_sgd"
if
self
.
_ut4grad_allreduce
:
tr_cmd
+=
" --ut4grad_allreduce"
if
hasattr
(
self
,
'_sync_batch_norm'
)
and
self
.
_sync_batch_norm
:
tr_cmd
+=
" --sync_batch_norm"
if
os
.
getenv
(
'WITH_COVERAGE'
,
'OFF'
)
==
'ON'
:
env
[
'COVERAGE_FILE'
]
=
os
.
getenv
(
'COVERAGE_FILE'
,
''
)
...
...
python/paddle/fluid/tests/unittests/test_dist_mnist_fleetapi.py
浏览文件 @
3ec289a6
...
...
@@ -24,6 +24,7 @@ class TestDistMnistNCCL2FleetApi(TestDistBase):
self
.
_use_reader_alloc
=
False
self
.
_nccl2_mode
=
True
self
.
_gpu_fleet_api
=
True
self
.
_sync_batch_norm
=
True
def
test_dist_train
(
self
):
import
paddle.fluid
as
fluid
...
...
@@ -31,5 +32,30 @@ class TestDistMnistNCCL2FleetApi(TestDistBase):
self
.
check_with_place
(
"dist_mnist.py"
,
delta
=
1e-5
)
class
FleetCollectiveTest
(
unittest
.
TestCase
):
def
test_open_sync_batch_norm
(
self
):
import
paddle.fluid
as
fluid
import
paddle.fluid.incubate.fleet.base.role_maker
as
role_maker
from
paddle.fluid.incubate.fleet.collective
import
fleet
,
DistributedStrategy
data
=
fluid
.
layers
.
data
(
name
=
'X'
,
shape
=
[
1
],
dtype
=
'float32'
)
hidden
=
fluid
.
layers
.
fc
(
input
=
data
,
size
=
10
)
loss
=
fluid
.
layers
.
mean
(
hidden
)
optimizer
=
fluid
.
optimizer
.
AdamOptimizer
()
role
=
role_maker
.
UserDefinedCollectiveRoleMaker
(
0
,
[
'127.0.0.1:6170'
])
fleet
.
init
(
role
)
dist_strategy
=
DistributedStrategy
()
dist_strategy
.
sync_batch_norm
=
True
dist_optimizer
=
fleet
.
distributed_optimizer
(
optimizer
,
strategy
=
dist_strategy
)
dist_optimizer
.
minimize
(
loss
)
self
.
assertEqual
(
dist_strategy
.
exec_strategy
.
num_threads
,
1
)
if
__name__
==
"__main__"
:
unittest
.
main
()
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录