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
月光在发光
Paddle
提交
7b0c0273
P
Paddle
项目概览
月光在发光
/
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
7b0c0273
编写于
5月 14, 2018
作者:
T
typhoonzero
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
update by comments
上级
928418a9
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
17 addition
and
17 deletion
+17
-17
paddle/fluid/framework/parallel_executor.h
paddle/fluid/framework/parallel_executor.h
+1
-1
paddle/fluid/operators/gen_nccl_id_op.cc
paddle/fluid/operators/gen_nccl_id_op.cc
+12
-12
paddle/fluid/platform/nccl_helper.h
paddle/fluid/platform/nccl_helper.h
+2
-2
python/paddle/fluid/parallel_executor.py
python/paddle/fluid/parallel_executor.py
+2
-2
未找到文件。
paddle/fluid/framework/parallel_executor.h
浏览文件 @
7b0c0273
...
...
@@ -42,7 +42,7 @@ class ParallelExecutor {
const
std
::
vector
<
Scope
*>&
local_scopes
,
bool
allow_op_delay
,
bool
use_default_grad_scale
,
bool
balance_parameter_opt_between_cards
,
size_t
num_trainers
=
0
,
size_t
trainer_id
=
0
);
size_t
num_trainers
=
1
,
size_t
trainer_id
=
0
);
~
ParallelExecutor
();
...
...
paddle/fluid/operators/gen_nccl_id_op.cc
浏览文件 @
7b0c0273
...
...
@@ -75,29 +75,29 @@ class GenNCCLIdOp : public framework::OperatorBase {
// NOTE: Can not use unique_ptr here because the default
// deleter will call GRPC Server's base class's dtor and
// that will cause a wired crash.
rpc_service_
=
new
detail
::
AsyncGRPCServer
(
endpoint
,
true
);
detail
::
AsyncGRPCServer
rpc_service
(
endpoint
,
true
);
framework
::
ProgramDesc
empty_program
;
framework
::
Executor
executor
(
dev_ctx
.
GetPlace
());
rpc_service
_
->
SetScope
(
scope
);
rpc_service
_
->
SetDevCtx
(
&
dev_ctx
);
rpc_service
_
->
SetProgram
(
&
empty_program
);
rpc_service
_
->
SetExecutor
(
&
executor
);
rpc_service
.
SetScope
(
scope
);
rpc_service
.
SetDevCtx
(
&
dev_ctx
);
rpc_service
.
SetProgram
(
&
empty_program
);
rpc_service
.
SetExecutor
(
&
executor
);
std
::
thread
server_thread
(
std
::
bind
(
&
detail
::
AsyncGRPCServer
::
RunSyncUpdate
,
rpc_service_
));
rpc_service
_
->
SetCond
(
0
);
std
::
bind
(
&
detail
::
AsyncGRPCServer
::
RunSyncUpdate
,
&
rpc_service
));
rpc_service
.
SetCond
(
0
);
VLOG
(
3
)
<<
"start getting nccl id from trainer 0..."
;
auto
recv
=
rpc_service
_
->
Get
();
auto
recv
=
rpc_service
.
Get
();
VLOG
(
3
)
<<
"got nccl id and stop server..."
;
rpc_service
_
->
ShutDown
();
rpc_service
.
ShutDown
();
VLOG
(
3
)
<<
"rpc server stopped"
;
// TODO(wuyi): reinit nccl communicators
server_thread
.
join
();
delete
rpc_service_
;
}
protected:
mutable
detail
::
AsyncGRPCServer
*
rpc_service_
=
nullptr
;
//
protected:
//
mutable detail::AsyncGRPCServer* rpc_service_ = nullptr;
};
class
GenNCCLIdOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
...
...
paddle/fluid/platform/nccl_helper.h
浏览文件 @
7b0c0273
...
...
@@ -78,7 +78,7 @@ struct NCCLContextMap {
explicit
NCCLContextMap
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
ncclUniqueId
*
nccl_id
=
nullptr
,
size_t
num_trainers
=
0
,
size_t
trainer_id
=
0
)
{
size_t
num_trainers
=
1
,
size_t
trainer_id
=
0
)
{
PADDLE_ENFORCE
(
!
places
.
empty
());
order_
.
reserve
(
places
.
size
());
for
(
auto
&
p
:
places
)
{
...
...
@@ -100,7 +100,7 @@ struct NCCLContextMap {
PADDLE_ENFORCE
(
platform
::
dynload
::
ncclCommInitAll
(
comms
.
get
(),
static_cast
<
int
>
(
order_
.
size
()),
order_
.
data
()));
}
else
{
PADDLE_ENFORCE_GT
(
num_trainers
,
0
);
PADDLE_ENFORCE_GT
(
num_trainers
,
1
);
// TODO(wuyi): need to ensure each node have same number of GPUs
{
int
nranks
=
num_trainers
*
order_
.
size
();
...
...
python/paddle/fluid/parallel_executor.py
浏览文件 @
7b0c0273
...
...
@@ -32,7 +32,7 @@ class ParallelExecutor(object):
share_vars_from
=
None
,
use_default_grad_scale
=
True
,
balance_parameter_opt_between_cards
=
False
,
num_trainers
=
0
,
num_trainers
=
1
,
trainer_id
=
0
):
"""
ParallelExecutor can run program in parallel.
...
...
@@ -57,7 +57,7 @@ class ParallelExecutor(object):
balance_parameter_opt_between_cards(bool, default True): Whether
updating different gradients on different cards. Currently, it
is not recommended.
num_trainers(int, default
0): If greater than 0
, NCCL will be
num_trainers(int, default
1): If greater than 1
, NCCL will be
initialized with multpile rank of nodes, each node should have
same number of GPUs. Distributed training will be enabled then.
trainer_id(int, default 0): Must use together with num_trainers.
...
...
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