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
89951472
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
89951472
编写于
4月 27, 2022
作者:
L
lilong12
提交者:
GitHub
4月 27, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add the support for allreduce_prod for new dygraph (#42284)
上级
8395d660
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
62 addition
and
12 deletion
+62
-12
python/paddle/distributed/collective.py
python/paddle/distributed/collective.py
+15
-10
python/paddle/distributed/parallel.py
python/paddle/distributed/parallel.py
+3
-2
python/paddle/fluid/tests/unittests/process_group_nccl.py
python/paddle/fluid/tests/unittests/process_group_nccl.py
+44
-0
未找到文件。
python/paddle/distributed/collective.py
浏览文件 @
89951472
...
...
@@ -350,18 +350,19 @@ def new_group(ranks=None, backend=None):
global
_default_group_name
gid
=
_new_ring_id
()
group_name
=
_default_group_name
+
str
(
gid
)
global_group
=
_get_default_group
()
global_rank
=
global_group
.
rank
global_ranks
=
global_group
.
ranks
backend
=
_default_backend
if
backend
is
None
else
backend
if
ranks
is
None
:
ranks
=
global_ranks
assert
len
(
ranks
)
<=
len
(
global_ranks
),
(
"Size of new group must be less than or "
"equal to that of the default global group."
)
if
ranks
is
None
or
len
(
ranks
)
>
1
:
global_group
=
_get_default_group
()
global_rank
=
global_group
.
rank
global_ranks
=
global_group
.
ranks
backend
=
_default_backend
if
backend
is
None
else
backend
if
ranks
is
None
:
ranks
=
global_ranks
assert
len
(
ranks
)
<=
len
(
global_ranks
),
(
"Size of new group must be less than or "
"equal to that of the default global group."
)
size
=
len
(
ranks
)
ranks
=
sorted
(
ranks
)
if
global_rank
in
ranks
and
size
>
1
:
if
size
>
1
and
global_rank
in
ranks
:
rank
=
ranks
.
index
(
global_rank
)
pg
=
_new_process_group_impl
(
backend
,
...
...
@@ -642,6 +643,8 @@ def all_reduce(tensor, op=ReduceOp.SUM, group=None, use_calc_stream=True):
op_type
=
core
.
ReduceOp
.
MAX
elif
op
==
ReduceOp
.
MIN
:
op_type
=
core
.
ReduceOp
.
MIN
elif
op
==
ReduceOp
.
PROD
:
op_type
=
core
.
ReduceOp
.
PRODUCT
else
:
raise
ValueError
(
"Unknown reduce_op type for allreduce."
)
group
=
_get_default_group
()
if
group
is
None
else
group
...
...
@@ -744,6 +747,8 @@ def reduce(tensor, dst, op=ReduceOp.SUM, group=None, use_calc_stream=True):
op_type
=
core
.
ReduceOp
.
MAX
elif
op
==
ReduceOp
.
MIN
:
op_type
=
core
.
ReduceOp
.
MIN
elif
op
==
ReduceOp
.
PROD
:
op_type
=
core
.
ReduceOp
.
PRODUCT
else
:
raise
ValueError
(
"Unknown reduce_op type for reduce."
)
group
=
_get_default_group
()
if
group
is
None
else
group
...
...
python/paddle/distributed/parallel.py
浏览文件 @
89951472
...
...
@@ -219,8 +219,9 @@ def init_parallel_env():
"required to create a process group."
)
master_addr
=
os
.
getenv
(
"MASTER_ADDR"
,
None
)
master_port
=
os
.
getenv
(
"MASTER_PORT"
,
None
)
endpoints
=
None
if
not
master_addr
or
not
master_port
:
endpoints
=
":"
.
join
(
[
master_addr
,
master_port
])
if
master_addr
and
master_port
else
None
if
endpoints
is
None
:
endpoints
=
os
.
getenv
(
"PADDLE_MASTER"
,
None
)
if
endpoints
is
None
:
endpoints
=
os
.
getenv
(
"PADDLE_TRAINER_ENDPOINTS"
).
split
(
','
)[
0
]
...
...
python/paddle/fluid/tests/unittests/process_group_nccl.py
浏览文件 @
89951472
...
...
@@ -122,6 +122,29 @@ class TestProcessGroupFp32(unittest.TestCase):
print
(
"test allreduce min api ok"
)
# test allreduce prod
# rank 0
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_x
=
paddle
.
to_tensor
(
x
)
# rank 1
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_y
=
paddle
.
to_tensor
(
y
)
prod_result
=
np
.
multiply
(
x
,
y
)
if
pg
.
rank
()
==
0
:
task
=
dist
.
all_reduce
(
tensor_x
,
dist
.
ReduceOp
.
PROD
,
use_calc_stream
=
False
)
task
.
wait
()
assert
np
.
array_equal
(
tensor_x
,
prod_result
)
else
:
task
=
dist
.
all_reduce
(
tensor_y
,
dist
.
ReduceOp
.
PROD
,
use_calc_stream
=
False
)
task
.
wait
()
assert
np
.
array_equal
(
tensor_y
,
prod_result
)
print
(
"test allreduce prod api ok"
)
# test broadcast
# rank 0
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
...
...
@@ -332,6 +355,27 @@ class TestProcessGroupFp32(unittest.TestCase):
print
(
"test reduce min api ok"
)
# test reduce product
# rank 0
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_x
=
paddle
.
to_tensor
(
x
)
# rank 1
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_y
=
paddle
.
to_tensor
(
y
)
prod_result
=
np
.
multiply
(
x
,
y
)
if
pg
.
rank
()
==
0
:
task
=
dist
.
reduce
(
tensor_x
,
0
,
dist
.
ReduceOp
.
PROD
,
use_calc_stream
=
False
)
task
.
wait
()
assert
np
.
array_equal
(
tensor_x
,
prod_result
)
else
:
task
=
dist
.
reduce
(
tensor_y
,
0
,
dist
.
ReduceOp
.
PROD
,
use_calc_stream
=
False
)
task
.
wait
()
print
(
"test reduce prod api ok"
)
# test Scatter
# rank 0
in_shape
=
list
(
self
.
shape
)
...
...
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