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
ec8e8782
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
ec8e8782
编写于
2月 23, 2019
作者:
乔
乔龙飞 Qiao Longfei
提交者:
GitHub
2月 23, 2019
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #15840 from jacquesqiao/revert-15684-revert-15661-fix-cpu-broadcast
fix cpu broadcast
上级
8a7efc78
2b7931d5
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
31 addition
and
13 deletion
+31
-13
paddle/fluid/framework/details/build_strategy.cc
paddle/fluid/framework/details/build_strategy.cc
+3
-0
paddle/fluid/framework/details/multi_devices_graph_pass.cc
paddle/fluid/framework/details/multi_devices_graph_pass.cc
+15
-3
python/paddle/fluid/compiler.py
python/paddle/fluid/compiler.py
+3
-0
python/paddle/fluid/framework.py
python/paddle/fluid/framework.py
+9
-0
python/paddle/fluid/parallel_executor.py
python/paddle/fluid/parallel_executor.py
+1
-10
未找到文件。
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @
ec8e8782
...
...
@@ -135,12 +135,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
void
AppendMultiDevPass
(
const
BuildStrategy
&
strategy
)
{
ir
::
Pass
*
multi_devices_pass
;
if
(
strategy_
.
is_distribution_
)
{
VLOG
(
3
)
<<
"multi device parameter server mode"
;
multi_devices_pass
=
AppendPass
(
"dist_multi_devices_pass"
).
get
();
}
else
{
if
(
strategy
.
reduce_
==
BuildStrategy
::
ReduceStrategy
::
kAllReduce
)
{
VLOG
(
3
)
<<
"multi devices collective mode with allreduce"
;
multi_devices_pass
=
AppendPass
(
"allreduce_mode_multi_devices_pass"
).
get
();
}
else
if
(
strategy
.
reduce_
==
BuildStrategy
::
ReduceStrategy
::
kReduce
)
{
VLOG
(
3
)
<<
"multi deivces collective mode with reduce"
;
multi_devices_pass
=
AppendPass
(
"reduce_mode_multi_devices_pass"
).
get
();
}
else
{
PADDLE_THROW
(
"Unknown reduce strategy."
);
...
...
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @
ec8e8782
...
...
@@ -937,9 +937,21 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
}
void
DistSSAGraphBuilder
::
InsertPostprocessOps
(
ir
::
Graph
*
result
)
const
{
if
(
need_broadcast_var_
||
(
UseGPU
()
&&
strategy_
.
reduce_
==
BuildStrategy
::
ReduceStrategy
::
kReduce
))
{
// broad cast received parameters when training in parameter server mode.
if
(
need_broadcast_var_
)
{
// There are 4 conditions:
// 1. GPU && Reduce: Reduce gradient then broadcast gradient to other GPUS.
// Need to broadcast received parameters to other GPU.
// 2. GPU && AllReduce: AllReduce all graident to each GPU. Need to
// broadcast received parameters to other GPU.
// 3. CPU && AllReduce: AllReduce all gradient to each thread. Need to
// broadcast received parameters to other scope.
// 4. CPU && Reduce: because all parameters share the same memory, did not
// broadcast received parameters.
if
(
!
UseGPU
()
&&
strategy_
.
reduce_
==
BuildStrategy
::
ReduceStrategy
::
kReduce
)
{
return
;
}
if
(
strategy_
.
fuse_broadcast_op_
)
{
CreateFusedBroadcastOp
(
result
,
bcast_var_name_set_
);
}
else
{
...
...
python/paddle/fluid/compiler.py
浏览文件 @
ec8e8782
...
...
@@ -19,6 +19,7 @@ import sys
from
..
import
compat
as
cpt
from
.
import
core
from
.
import
framework
__all__
=
[
'CompiledProgram'
,
'ExecutionStrategy'
,
'BuildStrategy'
]
...
...
@@ -110,6 +111,8 @@ class CompiledProgram(object):
self
.
_exec_strategy
=
ExecutionStrategy
()
if
self
.
_build_strategy
is
None
:
self
.
_build_strategy
=
BuildStrategy
()
self
.
_build_strategy
.
is_distribution
=
framework
.
is_pserver_mode
(
self
.
_program
)
return
self
def
with_inference_optimize
(
self
,
config
):
...
...
python/paddle/fluid/framework.py
浏览文件 @
ec8e8782
...
...
@@ -87,6 +87,15 @@ def _current_expected_place():
return
_imperative_current_expected_place_
def
is_pserver_mode
(
main_program
):
main
=
main_program
if
main_program
\
else
default_main_program
()
for
op
in
main
.
global_block
().
ops
:
if
op
.
type
in
[
"send"
,
"recv"
]:
return
True
return
False
class
NameScope
(
object
):
def
__init__
(
self
,
name
=
""
,
parent
=
None
):
self
.
_children
=
dict
()
...
...
python/paddle/fluid/parallel_executor.py
浏览文件 @
ec8e8782
...
...
@@ -29,15 +29,6 @@ ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy
BuildStrategy
=
core
.
ParallelExecutor
.
BuildStrategy
def
_is_pserver_mode
(
main_program
):
main
=
main_program
if
main_program
\
else
framework
.
default_main_program
()
for
op
in
main
.
global_block
().
ops
:
if
op
.
type
in
[
"send"
,
"recv"
]:
return
True
return
False
class
ParallelExecutor
(
object
):
"""
ParallelExecutor is designed for data parallelism, which focuses on distributing
...
...
@@ -140,7 +131,7 @@ class ParallelExecutor(object):
# FIXME(zcd): is_distribution_ is a temporary field, because in pserver mode,
# num_trainers is 1, so the current fields of build_strategy doesn't tell if
# it's distributed model.
build_strategy
.
is_distribution
=
_
is_pserver_mode
(
build_strategy
.
is_distribution
=
framework
.
is_pserver_mode
(
main_program
)
or
num_trainers
>
1
# step4: get main_program, scope, local_scopes
...
...
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