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
Crayon鑫
Paddle
提交
0d3d4ae7
P
Paddle
项目概览
Crayon鑫
/
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
提交
0d3d4ae7
编写于
6月 11, 2018
作者:
Q
qiaolongfei
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
refine prefetch logic
上级
17b42fc2
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
64 addition
and
52 deletion
+64
-52
paddle/fluid/operators/listen_and_serv_op.cc
paddle/fluid/operators/listen_and_serv_op.cc
+4
-3
paddle/fluid/operators/listen_and_serv_op.h
paddle/fluid/operators/listen_and_serv_op.h
+1
-1
python/paddle/fluid/transpiler/distribute_transpiler.py
python/paddle/fluid/transpiler/distribute_transpiler.py
+59
-48
未找到文件。
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @
0d3d4ae7
...
...
@@ -248,7 +248,8 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
request_prefetch_handler_
.
get
());
auto
*
optimize_block
=
Attr
<
framework
::
BlockDesc
*>
(
kOptimizeBlock
);
auto
*
prefetch_block
=
Attr
<
framework
::
BlockDesc
*>
(
kPrefetchBlock
);
auto
grad_to_block_id_str
=
Attr
<
std
::
vector
<
std
::
string
>>
(
kPrefetchBlock
);
framework
::
BlockDesc
*
prefetch_block
=
nullptr
;
auto
*
program
=
optimize_block
->
Program
();
framework
::
Executor
executor
(
dev_place
);
...
...
@@ -302,8 +303,8 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker {
AddAttr
<
bool
>
(
"sync_mode"
,
"if works at sync_mode or not"
).
SetDefault
(
true
);
AddAttr
<
framework
::
BlockDesc
*>
(
kOptimizeBlock
,
"BlockID to run on server side."
);
AddAttr
<
framework
::
BlockDesc
*
>
(
kPrefetchBlock
,
"prefetch block to run on server side."
);
AddAttr
<
std
::
vector
<
std
::
string
>
>
(
kPrefetchBlock
,
"prefetch block to run on server side."
);
AddAttr
<
int
>
(
"Fanin"
,
"How many clients send to this server."
)
.
SetDefault
(
1
);
}
...
...
paddle/fluid/operators/listen_and_serv_op.h
浏览文件 @
0d3d4ae7
...
...
@@ -30,7 +30,7 @@ namespace paddle {
namespace
operators
{
constexpr
char
kOptimizeBlock
[]
=
"OptimizeBlock"
;
constexpr
char
kPrefetchBlock
[]
=
"
PrefetchBlock
"
;
constexpr
char
kPrefetchBlock
[]
=
"
prefetch_var_name_to_block_id
"
;
void
RunServer
(
std
::
shared_ptr
<
detail
::
RPCServer
>
service
);
...
...
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @
0d3d4ae7
...
...
@@ -515,21 +515,20 @@ class DistributeTranspiler:
grad_to_block_id
,
None
)
# process distributed lookup_table
prefetch_
block
=
None
prefetch_
var_name_to_block_id
=
[]
if
self
.
has_distributed_lookup_table
:
pserver_index
=
self
.
pserver_endpoints
.
index
(
endpoint
)
table_opt_block
=
self
.
_create_table_optimize_block
(
pserver_index
,
pserver_program
,
pre_block_idx
,
grad_to_block_id
)
prefetch_
block
=
self
.
_create_prefetch_block
(
prefetch_
var_name_to_block_id
=
self
.
_create_prefetch_block
(
pserver_index
,
pserver_program
,
table_opt_block
)
# NOTE: if has_distributed_lookup_table is False, then prefetch_block will
# not be executed, so it's safe to use optimize_block to hold the place
if
self
.
has_distributed_lookup_table
:
assert
prefetch_block
is
not
None
assert
len
(
prefetch_var_name_to_block_id
)
>
0
else
:
assert
prefetch_block
is
None
prefetch_block
=
pserver_program
.
global_block
()
assert
len
(
prefetch_var_name_to_block_id
)
==
0
# step5 append the listen_and_serv op
pserver_program
.
global_block
().
append_op
(
...
...
@@ -540,7 +539,7 @@ class DistributeTranspiler:
"OptimizeBlock"
:
pserver_program
.
block
(
1
),
"endpoint"
:
endpoint
,
"Fanin"
:
self
.
trainer_num
,
"
PrefetchBlock"
:
prefetch_block
,
"
prefetch_var_name_to_block_id"
:
prefetch_var_name_to_block_id
,
"sync_mode"
:
self
.
sync_mode
,
"grad_to_block_id"
:
grad_to_block_id
})
...
...
@@ -608,8 +607,15 @@ class DistributeTranspiler:
def
_replace_lookup_table_op_with_prefetch
(
self
,
program
,
pserver_endpoints
):
# 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op
self
.
prefetch_input_vars
=
None
self
.
prefetch_output_vars
=
None
# self.all_prefetch_input_vars =
# [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
# [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
self
.
all_prefetch_input_vars
=
[]
# self.all_prefetch_input_vars =
# [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
# [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
self
.
all_prefetch_output_vars
=
[]
continue_search_lookup_table_op
=
True
while
continue_search_lookup_table_op
:
...
...
@@ -623,18 +629,19 @@ class DistributeTranspiler:
ids_name
=
op
.
input
(
"Ids"
)
out_name
=
op
.
output
(
"Out"
)
if
self
.
prefetch_input_vars
is
None
:
ids_var
=
program
.
global_block
().
vars
[
ids_name
[
0
]]
self
.
prefetch_input_vars
=
self
.
create_splited_vars
(
source_var
=
ids_var
,
block
=
program
.
global_block
(),
tag
=
"_prefetch_in_"
)
if
self
.
prefetch_output_vars
is
None
:
out_var
=
program
.
global_block
().
vars
[
out_name
[
0
]]
self
.
prefetch_output_vars
=
self
.
create_splited_vars
(
source_var
=
out_var
,
block
=
program
.
global_block
(),
tag
=
"_prefetch_out_"
)
ids_var
=
program
.
global_block
().
vars
[
ids_name
[
0
]]
prefetch_input_vars
=
self
.
create_splited_vars
(
source_var
=
ids_var
,
block
=
program
.
global_block
(),
tag
=
"_prefetch_in_"
)
self
.
all_prefetch_input_vars
.
append
(
prefetch_input_vars
)
out_var
=
program
.
global_block
().
vars
[
out_name
[
0
]]
prefetch_output_vars
=
self
.
create_splited_vars
(
source_var
=
out_var
,
block
=
program
.
global_block
(),
tag
=
"_prefetch_out_"
)
self
.
all_prefetch_output_vars
.
append
(
prefetch_output_vars
)
# insert split_ids_op
program
.
global_block
().
insert_op
(
...
...
@@ -646,14 +653,14 @@ class DistributeTranspiler:
for
varname
in
ids_name
]
},
outputs
=
{
"Out"
:
self
.
prefetch_input_vars
})
outputs
=
{
"Out"
:
prefetch_input_vars
})
# insert prefetch_op
program
.
global_block
().
insert_op
(
index
=
op_index
+
1
,
type
=
"prefetch"
,
inputs
=
{
'X'
:
self
.
prefetch_input_vars
},
outputs
=
{
"Out"
:
self
.
prefetch_output_vars
},
inputs
=
{
'X'
:
prefetch_input_vars
},
outputs
=
{
"Out"
:
prefetch_output_vars
},
attrs
=
{
"epmap"
:
pserver_endpoints
,
RPC_OP_ROLE_ATTR_NAME
:
RPC_OP_ROLE_ATTR_VALUE
...
...
@@ -663,7 +670,7 @@ class DistributeTranspiler:
program
.
global_block
().
insert_op
(
index
=
op_index
+
2
,
type
=
"concat"
,
inputs
=
{
'X'
:
self
.
prefetch_output_vars
},
inputs
=
{
'X'
:
prefetch_output_vars
},
outputs
=
{
"Out"
:
[
program
.
global_block
().
vars
[
varname
]
...
...
@@ -709,30 +716,34 @@ class DistributeTranspiler:
optimize_block
):
# STEP: create prefetch block
table_var
=
pserver_program
.
global_block
().
vars
[
self
.
table_name
]
prefetch_block
=
pserver_program
.
create_block
(
optimize_block
.
idx
)
trainer_ids
=
self
.
prefetch_input_vars
[
pserver_index
]
pserver_ids
=
pserver_program
.
global_block
().
create_var
(
name
=
trainer_ids
.
name
,
type
=
trainer_ids
.
type
,
shape
=
trainer_ids
.
shape
,
dtype
=
trainer_ids
.
dtype
)
trainer_out
=
self
.
prefetch_output_vars
[
pserver_index
]
pserver_out
=
pserver_program
.
global_block
().
create_var
(
name
=
trainer_out
.
name
,
type
=
trainer_out
.
type
,
shape
=
trainer_out
.
shape
,
dtype
=
trainer_out
.
dtype
)
prefetch_block
.
append_op
(
type
=
"lookup_sparse_table"
,
inputs
=
{
'Ids'
:
pserver_ids
,
"W"
:
table_var
},
outputs
=
{
"Out"
:
pserver_out
},
attrs
=
{
"is_sparse"
:
True
,
# has no effect on lookup_table op
"is_distributed"
:
True
,
"padding_idx"
:
-
1
})
return
prefetch_block
prefetch_var_name_to_block_id
=
[]
for
index
in
range
(
len
(
self
.
all_prefetch_input_vars
)):
prefetch_block
=
pserver_program
.
create_block
(
optimize_block
.
idx
)
trainer_ids
=
self
.
all_prefetch_input_vars
[
index
][
pserver_index
]
pserver_ids
=
pserver_program
.
global_block
().
create_var
(
name
=
trainer_ids
.
name
,
type
=
trainer_ids
.
type
,
shape
=
trainer_ids
.
shape
,
dtype
=
trainer_ids
.
dtype
)
trainer_out
=
self
.
all_prefetch_output_vars
[
index
][
pserver_index
]
pserver_out
=
pserver_program
.
global_block
().
create_var
(
name
=
trainer_out
.
name
,
type
=
trainer_out
.
type
,
shape
=
trainer_out
.
shape
,
dtype
=
trainer_out
.
dtype
)
prefetch_block
.
append_op
(
type
=
"lookup_sparse_table"
,
inputs
=
{
'Ids'
:
pserver_ids
,
"W"
:
table_var
},
outputs
=
{
"Out"
:
pserver_out
},
attrs
=
{
"is_sparse"
:
True
,
# has no effect on lookup_table op
"is_distributed"
:
True
,
"padding_idx"
:
-
1
})
prefetch_var_name_to_block_id
.
append
(
trainer_ids
.
name
+
":"
+
str
(
prefetch_block
.
idx
))
return
prefetch_var_name_to_block_id
def
_create_table_optimize_block
(
self
,
pserver_index
,
pserver_program
,
pre_block_idx
,
grad_to_block_id
):
...
...
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