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
45d7a3ea
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
未验证
提交
45d7a3ea
编写于
5月 25, 2022
作者:
D
danleifeng
提交者:
GitHub
5月 25, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[GPUPS]fix gpups pscore (#42967)
上级
b6859054
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
55 addition
and
18 deletion
+55
-18
paddle/fluid/framework/hogwild_worker.cc
paddle/fluid/framework/hogwild_worker.cc
+8
-1
paddle/fluid/framework/multi_trainer.cc
paddle/fluid/framework/multi_trainer.cc
+30
-11
paddle/fluid/framework/trainer.h
paddle/fluid/framework/trainer.h
+1
-0
python/paddle/distributed/passes/ps_trainer_pass.py
python/paddle/distributed/passes/ps_trainer_pass.py
+11
-2
python/paddle/distributed/ps/the_one_ps.py
python/paddle/distributed/ps/the_one_ps.py
+5
-4
未找到文件。
paddle/fluid/framework/hogwild_worker.cc
浏览文件 @
45d7a3ea
...
@@ -219,6 +219,10 @@ void HogwildWorker::TrainFiles() {
...
@@ -219,6 +219,10 @@ void HogwildWorker::TrainFiles() {
device_reader_
->
Start
();
device_reader_
->
Start
();
int
cur_batch
;
int
cur_batch
;
int
batch_cnt
=
0
;
int
batch_cnt
=
0
;
#if defined(PADDLE_WITH_HETERPS) && defined(PADDLE_WITH_CUDA)
platform
::
SetDeviceId
(
thread_id_
);
#endif
while
((
cur_batch
=
device_reader_
->
Next
())
>
0
)
{
while
((
cur_batch
=
device_reader_
->
Next
())
>
0
)
{
for
(
auto
&
op
:
ops_
)
{
for
(
auto
&
op
:
ops_
)
{
bool
need_skip
=
false
;
bool
need_skip
=
false
;
...
@@ -244,9 +248,12 @@ void HogwildWorker::TrainFiles() {
...
@@ -244,9 +248,12 @@ void HogwildWorker::TrainFiles() {
++
batch_cnt
;
++
batch_cnt
;
PrintFetchVars
();
PrintFetchVars
();
thread_scope_
->
DropKids
();
thread_scope_
->
DropKids
();
#ifdef PADDLE_WITH_HETERPS
dev_ctx_
->
Wait
();
#endif
}
}
timeline
.
Pause
();
timeline
.
Pause
();
VLOG
(
3
)
<<
"worker "
<<
thread_id_
<<
" train cost "
<<
timeline
.
ElapsedSec
()
VLOG
(
1
)
<<
"worker "
<<
thread_id_
<<
" train cost "
<<
timeline
.
ElapsedSec
()
<<
" seconds, ins_num: "
<<
total_ins_num
;
<<
" seconds, ins_num: "
<<
total_ins_num
;
if
(
need_dump_field_
||
need_dump_param_
)
{
if
(
need_dump_field_
||
need_dump_param_
)
{
...
...
paddle/fluid/framework/multi_trainer.cc
浏览文件 @
45d7a3ea
...
@@ -148,6 +148,17 @@ void MultiTrainer::InitTrainerEnv(const ProgramDesc& main_program,
...
@@ -148,6 +148,17 @@ void MultiTrainer::InitTrainerEnv(const ProgramDesc& main_program,
}
}
}
}
#endif
#endif
for
(
auto
&
var
:
main_program
.
Block
(
0
).
AllVars
())
{
if
(
var
->
Persistable
())
{
auto
it
=
std
::
find
(
need_merge_var_names_
.
begin
(),
need_merge_var_names_
.
end
(),
var
->
Name
());
if
(
it
==
need_merge_var_names_
.
end
()
&&
var
->
GetType
()
!=
proto
::
VarType
::
SELECTED_ROWS
)
{
VLOG
(
2
)
<<
"train param: "
<<
var
->
Name
();
trainable_param_
.
push_back
(
var
->
Name
());
}
}
}
}
}
void
MultiTrainer
::
InitOtherEnv
(
const
ProgramDesc
&
main_program
)
{
void
MultiTrainer
::
InitOtherEnv
(
const
ProgramDesc
&
main_program
)
{
...
@@ -192,18 +203,30 @@ void MultiTrainer::Run() {
...
@@ -192,18 +203,30 @@ void MultiTrainer::Run() {
#ifdef PADDLE_WITH_HETERPS
#ifdef PADDLE_WITH_HETERPS
void
MultiTrainer
::
MergeDenseParam
()
{
void
MultiTrainer
::
MergeDenseParam
()
{
#ifdef PADDLE_W
TI
H_PSCORE
#ifdef PADDLE_W
IT
H_PSCORE
auto
communicator
=
paddle
::
distributed
::
Communicator
::
GetInstance
();
auto
communicator
=
paddle
::
distributed
::
Communicator
::
GetInstance
();
auto
&
recv_ctx
=
communicator
->
GetRecvCtxMap
();
auto
thread_scope
=
workers_
[
0
]
->
GetThreadScope
();
Scope
*
thread_scope
=
workers_
[
0
]
->
GetThreadScope
();
if
(
communicator
==
nullptr
)
{
for
(
auto
&
iter
:
recv_ctx
)
{
for
(
auto
&
name
:
trainable_param_
)
{
auto
&
varnames
=
iter
.
second
;
VLOG
(
2
)
<<
"merge var "
<<
name
<<
" to root scope"
;
for
(
auto
&
name
:
varnames
)
{
Variable
*
root_var
=
root_scope_
->
FindVar
(
name
);
Variable
*
root_var
=
root_scope_
->
FindVar
(
name
);
LoDTensor
*
root_tensor
=
root_var
->
GetMutable
<
LoDTensor
>
();
LoDTensor
*
root_tensor
=
root_var
->
GetMutable
<
LoDTensor
>
();
Variable
*
var
=
thread_scope
->
FindVar
(
name
);
Variable
*
var
=
thread_scope
->
FindVar
(
name
);
LoDTensor
*
tensor
=
var
->
GetMutable
<
LoDTensor
>
();
LoDTensor
*
tensor
=
var
->
GetMutable
<
LoDTensor
>
();
TensorCopy
((
*
tensor
),
root_tensor
->
place
(),
root_tensor
);
TensorCopySync
((
*
tensor
),
root_tensor
->
place
(),
root_tensor
);
}
}
else
{
auto
&
recv_ctx
=
communicator
->
GetRecvCtxMap
();
for
(
auto
&
iter
:
recv_ctx
)
{
auto
&
varnames
=
iter
.
second
;
for
(
auto
&
name
:
varnames
)
{
VLOG
(
2
)
<<
"merge var "
<<
name
<<
" to root scope"
;
Variable
*
root_var
=
root_scope_
->
FindVar
(
name
);
LoDTensor
*
root_tensor
=
root_var
->
GetMutable
<
LoDTensor
>
();
Variable
*
var
=
thread_scope
->
FindVar
(
name
);
LoDTensor
*
tensor
=
var
->
GetMutable
<
LoDTensor
>
();
TensorCopySync
((
*
tensor
),
root_tensor
->
place
(),
root_tensor
);
}
}
}
}
}
#endif
#endif
...
@@ -236,11 +259,7 @@ void MultiTrainer::Finalize() {
...
@@ -236,11 +259,7 @@ void MultiTrainer::Finalize() {
}
}
LoDTensor
*
root_tensor
=
root_var
->
GetMutable
<
LoDTensor
>
();
LoDTensor
*
root_tensor
=
root_var
->
GetMutable
<
LoDTensor
>
();
#ifdef PADDLE_WITH_HETERPS
for
(
size_t
j
=
0
;
j
<
places_
.
size
();
j
++
)
{
#else
for
(
int
j
=
1
;
j
<
thread_num_
;
j
++
)
{
for
(
int
j
=
1
;
j
<
thread_num_
;
j
++
)
{
#endif
Scope
*
cur_thread_scope
=
workers_
[
j
]
->
GetThreadScope
();
Scope
*
cur_thread_scope
=
workers_
[
j
]
->
GetThreadScope
();
Variable
*
thread_var
=
Variable
*
thread_var
=
cur_thread_scope
->
FindVar
(
need_merge_var_names_
[
i
]);
cur_thread_scope
->
FindVar
(
need_merge_var_names_
[
i
]);
...
...
paddle/fluid/framework/trainer.h
浏览文件 @
45d7a3ea
...
@@ -129,6 +129,7 @@ class MultiTrainer : public TrainerBase {
...
@@ -129,6 +129,7 @@ class MultiTrainer : public TrainerBase {
std
::
vector
<
DataFeed
*>
readers_
;
std
::
vector
<
DataFeed
*>
readers_
;
std
::
vector
<
std
::
shared_ptr
<
DeviceWorker
>>
workers_
;
std
::
vector
<
std
::
shared_ptr
<
DeviceWorker
>>
workers_
;
std
::
vector
<
std
::
string
>
need_merge_var_names_
;
std
::
vector
<
std
::
string
>
need_merge_var_names_
;
std
::
vector
<
std
::
string
>
trainable_param_
;
#ifdef PADDLE_WITH_HETERPS
#ifdef PADDLE_WITH_HETERPS
std
::
vector
<
platform
::
Place
>
places_
;
std
::
vector
<
platform
::
Place
>
places_
;
#endif
#endif
...
...
python/paddle/distributed/passes/ps_trainer_pass.py
浏览文件 @
45d7a3ea
...
@@ -614,15 +614,24 @@ class PsGpuPass(PassBase):
...
@@ -614,15 +614,24 @@ class PsGpuPass(PassBase):
return
True
return
True
def
_add_push_box_sparse_op
(
self
,
program
):
def
_add_push_box_sparse_op
(
self
,
program
):
insert_index
=
-
1
for
idx
,
op
in
list
(
enumerate
(
program
.
global_block
().
ops
)):
if
op
.
type
==
"lookup_table_grad"
:
insert_index
=
idx
for
op
in
program
.
global_block
().
ops
:
for
op
in
program
.
global_block
().
ops
:
if
op
.
type
!=
"pull_box_sparse"
:
if
op
.
type
!=
"pull_box_sparse"
and
op
.
type
!=
"pull_gpups_sparse"
:
continue
continue
grad_op_desc
,
op_grad_to_var
=
core
.
get_grad_op_desc
(
grad_op_desc
,
op_grad_to_var
=
core
.
get_grad_op_desc
(
op
.
desc
,
cpt
.
to_text
(
set
()),
[])
op
.
desc
,
cpt
.
to_text
(
set
()),
[])
for
op_desc
in
grad_op_desc
:
for
op_desc
in
grad_op_desc
:
new_op_desc
=
program
.
global_block
().
desc
.
append_op
()
new_op_desc
=
program
.
global_block
().
desc
.
_insert_op
(
insert_index
+
1
)
new_op_desc
.
copy_from
(
op_desc
)
new_op_desc
.
copy_from
(
op_desc
)
new_op_desc
.
_set_attr
(
op_role_attr_name
,
backward
)
new_op_desc
.
_set_attr
(
op_role_attr_name
,
backward
)
new_op
=
paddle
.
fluid
.
framework
.
Operator
(
program
.
global_block
(),
new_op_desc
)
program
.
global_block
().
ops
.
insert
(
insert_index
+
1
,
new_op
)
program
.
global_block
().
_sync_with_cpp
()
def
_remove_optimizer_var
(
self
,
program
):
def
_remove_optimizer_var
(
self
,
program
):
embedding_w
=
{}
embedding_w
=
{}
...
...
python/paddle/distributed/ps/the_one_ps.py
浏览文件 @
45d7a3ea
...
@@ -1013,12 +1013,13 @@ class TheOnePSRuntime(RuntimeBase):
...
@@ -1013,12 +1013,13 @@ class TheOnePSRuntime(RuntimeBase):
if
self
.
context
[
'ps_mode'
]
==
DistributedMode
.
GEO
:
if
self
.
context
[
'ps_mode'
]
==
DistributedMode
.
GEO
:
self
.
_communicator
.
init_params
(
init_params
)
self
.
_communicator
.
init_params
(
init_params
)
else
:
else
:
if
role_id
==
0
:
if
not
self
.
context
[
'use_ps_gpu'
]:
self
.
_init_all_params
(
scopes
,
send_ctx
,
dense_map
)
if
role_id
==
0
:
self
.
_init_all_params
(
scopes
,
send_ctx
,
dense_map
)
fleet
.
util
.
barrier
()
fleet
.
util
.
barrier
()
if
not
self
.
context
[
'use_ps_gpu'
]:
self
.
_pull_all_dense
(
scopes
,
send_ctx
,
dense_map
)
self
.
_pull_all_dense
(
scopes
,
send_ctx
,
dense_map
)
fleet
.
util
.
barrier
()
fleet
.
util
.
barrier
()
if
self
.
context
[
'ps_mode'
]
==
DistributedMode
.
GEO
:
if
self
.
context
[
'ps_mode'
]
==
DistributedMode
.
GEO
:
...
...
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