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
OPTHREE
Paddle
提交
c4ded17e
P
Paddle
项目概览
OPTHREE
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
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
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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 >>
提交
c4ded17e
编写于
2月 11, 2019
作者:
Q
Qiao Longfei
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
async mode support dist train
上级
84367cf8
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
10 addition
and
5 deletion
+10
-5
paddle/fluid/framework/details/build_strategy.cc
paddle/fluid/framework/details/build_strategy.cc
+3
-3
paddle/fluid/framework/details/multi_devices_graph_pass.cc
paddle/fluid/framework/details/multi_devices_graph_pass.cc
+6
-1
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+1
-1
未找到文件。
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @
c4ded17e
...
...
@@ -133,10 +133,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
void
AppendMultiDevPass
(
const
BuildStrategy
&
strategy
)
{
ir
::
Pass
*
multi_devices_pass
;
if
(
strategy_
.
async_mode_
)
{
multi_devices_pass
=
AppendPass
(
"async_multi_devices_pass"
).
get
();
}
else
if
(
strategy_
.
is_distribution_
)
{
if
(
strategy_
.
is_distribution_
)
{
multi_devices_pass
=
AppendPass
(
"dist_multi_devices_pass"
).
get
();
}
else
if
(
strategy_
.
async_mode_
)
{
multi_devices_pass
=
AppendPass
(
"async_multi_devices_pass"
).
get
();
}
else
{
if
(
strategy
.
reduce_
==
BuildStrategy
::
ReduceStrategy
::
kAllReduce
)
{
multi_devices_pass
=
...
...
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @
c4ded17e
...
...
@@ -756,6 +756,11 @@ bool DistSSAGraphBuilder::DealWithSpecialOp(ir::Graph *result,
insert_op
=
true
;
need_broadcast_var_
=
true
;
}
else
if
(
OpHaveRole
(
*
node
,
OpRole
::
kDist
))
{
// in async_mode, each graph will send it's own gradient, do not need to
// merge gradient.
if
(
strategy_
.
async_mode_
&&
node
->
Op
()
->
Type
()
!=
"concat"
)
{
return
false
;
}
int
op_dev_id
=
CreateDistTrainOp
(
result
,
node
);
if
(
node
->
Op
()
->
Type
()
==
"concat"
)
{
// the input(block of parameter) of concat is on different device,
...
...
@@ -827,7 +832,7 @@ int DistSSAGraphBuilder::CreateRPCOp(ir::Graph *result, ir::Node *node) const {
}
auto
recv_param_grad
=
boost
::
get
<
std
::
vector
<
std
::
string
>>
(
node
->
Op
()
->
GetAttr
(
OpProtoAndCheckerMaker
::
OpRoleVarAttrName
()));
if
(
recv_param_grad
.
size
()
==
2U
)
{
if
(
recv_param_grad
.
size
()
==
2U
&&
!
strategy_
.
async_mode_
)
{
op_dev_id
=
GetVarDeviceID
(
recv_param_grad
[
1
]);
VLOG
(
10
)
<<
"recv param "
<<
recv_param_grad
[
0
]
<<
" get grad place: "
<<
recv_param_grad
[
1
]
...
...
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
c4ded17e
...
...
@@ -283,7 +283,7 @@ ParallelExecutor::ParallelExecutor(
graphs
.
push_back
(
std
::
move
(
graph
));
}
#else
if
(
build_strategy
.
async_mode_
)
{
if
(
build_strategy
.
async_mode_
&&
!
build_strategy
.
is_distribution_
)
{
for
(
size_t
i
=
0
;
i
<
member_
->
places_
.
size
();
++
i
)
{
std
::
unique_ptr
<
ir
::
Graph
>
graph
=
build_strategy
.
Apply
(
main_program
,
{
member_
->
places_
[
i
]},
loss_var_name
,
...
...
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