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
极客李华
Paddle
提交
695dd371
P
Paddle
项目概览
极客李华
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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 >>
未验证
提交
695dd371
编写于
3月 31, 2021
作者:
L
lilong12
提交者:
GitHub
3月 31, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Adjust pipeline optimizer for 3d parallelism (#31939)
* update, test=develop
上级
6f85e241
变更
4
展开全部
隐藏空白更改
内联
并排
Showing
4 changed file
with
168 addition
and
161 deletion
+168
-161
paddle/fluid/framework/pipeline_trainer.cc
paddle/fluid/framework/pipeline_trainer.cc
+3
-24
python/paddle/distributed/fleet/meta_optimizers/common.py
python/paddle/distributed/fleet/meta_optimizers/common.py
+5
-0
python/paddle/distributed/fleet/meta_optimizers/pipeline_optimizer.py
...e/distributed/fleet/meta_optimizers/pipeline_optimizer.py
+1
-10
python/paddle/fluid/optimizer.py
python/paddle/fluid/optimizer.py
+159
-127
未找到文件。
paddle/fluid/framework/pipeline_trainer.cc
浏览文件 @
695dd371
...
...
@@ -71,37 +71,16 @@ void PipelineTrainer::CopyParameters(int microbatch_id,
const
ProgramDesc
&
program
,
const
platform
::
Place
&
place
)
{
auto
&
global_block
=
program
.
Block
(
0
);
std
::
map
<
std
::
string
,
int
>
param_map
;
for
(
auto
&
var
:
global_block
.
AllVars
())
{
if
(
var
->
Persistable
())
{
param_map
[
var
->
Name
()]
=
1
;
}
}
for
(
auto
&
var
:
global_block
.
AllVars
())
{
bool
is_param_grad
=
false
;
size_t
pos
=
0
;
// A magic suffix to indicate the merged gradient
std
::
string
magicSuffix
=
std
::
string
(
kGradVarSuffix
)
+
"@MERGED"
;
if
((
pos
=
var
->
Name
().
find
(
magicSuffix
))
!=
std
::
string
::
npos
)
{
auto
prefix_name
=
var
->
Name
().
substr
(
0
,
pos
);
if
(
param_map
.
find
(
prefix_name
)
!=
param_map
.
end
())
{
is_param_grad
=
true
;
}
}
if
(
var
->
Persistable
()
&&
microbatch_id
==
0
)
{
auto
*
ptr
=
root_scope_
->
Var
(
var
->
Name
());
InitializeVariable
(
ptr
,
var
->
GetType
());
VLOG
(
3
)
<<
"Create persistable var: "
<<
var
->
Name
()
<<
", which pointer is "
<<
ptr
;
}
else
if
(
is_param_grad
&&
microbatch_id
==
0
)
{
auto
*
ptr
=
minibatch_scope_
->
Var
(
var
->
Name
());
InitializeVariable
(
ptr
,
var
->
GetType
());
VLOG
(
3
)
<<
"Create grad for persistable var: "
<<
var
->
Name
()
VLOG
(
5
)
<<
"Create persistable var: "
<<
var
->
Name
()
<<
", which pointer is "
<<
ptr
;
}
else
if
(
!
var
->
Persistable
()
&&
!
is_param_grad
)
{
}
else
if
(
!
var
->
Persistable
())
{
auto
*
ptr
=
microbatch_scopes_
[
microbatch_id
]
->
Var
(
var
->
Name
());
VLOG
(
3
)
<<
"Create variable "
<<
var
->
Name
()
<<
" for microbatch "
VLOG
(
5
)
<<
"Create variable "
<<
var
->
Name
()
<<
" for microbatch "
<<
microbatch_id
<<
", which pointer is "
<<
ptr
;
InitializeVariable
(
ptr
,
var
->
GetType
());
}
...
...
python/paddle/distributed/fleet/meta_optimizers/common.py
浏览文件 @
695dd371
...
...
@@ -106,6 +106,11 @@ class CollectiveHelper(object):
'use_calc_stream'
:
True
,
OP_ROLE_KEY
:
OpRole
.
Forward
})
block
.
append_op
(
type
=
'c_sync_calc_stream'
,
inputs
=
{
'X'
:
sync_var
},
outputs
=
{
'Out'
:
sync_var
},
attrs
=
{
OP_ROLE_KEY
:
OpRole
.
Forward
})
block
=
program
.
global_block
()
if
core
.
is_compiled_with_cuda
():
...
...
python/paddle/distributed/fleet/meta_optimizers/pipeline_optimizer.py
浏览文件 @
695dd371
...
...
@@ -171,6 +171,7 @@ class PipelineOptimizer(MetaOptimizerBase):
program
.
_pipeline_opt
[
'ring_id'
]
=
self
.
start_pipeline_ring_id
program
.
_pipeline_opt
[
'micro_batch_size'
]
=
self
.
micro_batch_size
program
.
_pipeline_opt
[
'schedule_mode'
]
=
self
.
schedule_mode
program
.
_pipeline_opt
[
'use_sharding'
]
=
False
optimize_ops
,
params_grads
,
prog_list
,
pp_pair
,
ring_map
=
self
.
wrapped_opt
.
minimize
(
loss
,
startup_program
,
parameter_list
,
no_grad_set
)
self
.
startup_program
=
orig_startup_program
.
_pipeline_opt
[
...
...
@@ -218,7 +219,6 @@ class PipelineOptimizer(MetaOptimizerBase):
grad
=
None
processed_param_name
=
set
()
first_optimize_op_idx
=
None
add_sync_calc_stream
=
False
for
idx
,
op
in
reversed
(
list
(
enumerate
(
block
.
ops
))):
if
is_backward_op
(
op
)
and
not
first_optimize_op_idx
:
first_optimize_op_idx
=
idx
+
1
...
...
@@ -242,15 +242,6 @@ class PipelineOptimizer(MetaOptimizerBase):
origin_param
=
origin_block
.
vars
[
op_role_var
[
i
]]
if
origin_param
.
is_distributed
:
continue
if
not
add_sync_calc_stream
:
add_sync_calc_stream
=
True
block
.
_insert_op
(
first_optimize_op_idx
+
offset
,
type
=
'c_sync_calc_stream'
,
inputs
=
{
'X'
:
grad
},
outputs
=
{
'Out'
:
grad
},
attrs
=
{
OP_ROLE_KEY
:
OpRole
.
Optimize
})
offset
+=
1
block
.
_insert_op
(
first_optimize_op_idx
+
offset
,
...
...
python/paddle/fluid/optimizer.py
浏览文件 @
695dd371
此差异已折叠。
点击以展开。
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