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
兽拳
Paddle
提交
5d7e1c91
P
Paddle
项目概览
兽拳
/
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 >>
未验证
提交
5d7e1c91
编写于
9月 09, 2022
作者:
Y
Yuang Liu
提交者:
GitHub
9月 09, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix dygraph pp + mp nan after async send/recv (#45869)
上级
a5836222
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
40 addition
and
31 deletion
+40
-31
python/paddle/distributed/fleet/meta_parallel/pp_utils/p2p_communication.py
...ributed/fleet/meta_parallel/pp_utils/p2p_communication.py
+40
-31
未找到文件。
python/paddle/distributed/fleet/meta_parallel/pp_utils/p2p_communication.py
浏览文件 @
5d7e1c91
...
...
@@ -327,7 +327,8 @@ def _p2p_helper(tensor_send_next, tensor_send_prev, recv_prev, recv_next):
if
tensor_send_prev
is
not
None
:
if
isinstance
(
tensor_send_prev
,
tuple
):
for
d
in
tensor_send_prev
:
paddle
.
distributed
.
wait
(
d
,
use_calc_stream
=
True
)
if
_in_legacy_dygraph
():
paddle
.
distributed
.
wait
(
d
,
use_calc_stream
=
True
)
tasks
.
append
(
send_partial
(
d
,
dst
=
0
,
...
...
@@ -336,7 +337,8 @@ def _p2p_helper(tensor_send_next, tensor_send_prev, recv_prev, recv_next):
group
=
_hcg
.
send_prev_group
,
use_calc_stream
=
False
))
else
:
paddle
.
distributed
.
wait
(
tensor_send_prev
,
use_calc_stream
=
True
)
if
_in_legacy_dygraph
():
paddle
.
distributed
.
wait
(
tensor_send_prev
,
use_calc_stream
=
True
)
tasks
.
append
(
send_partial
(
tensor_send_prev
,
dst
=
0
,
...
...
@@ -355,12 +357,6 @@ def _p2p_helper(tensor_send_next, tensor_send_prev, recv_prev, recv_next):
rank_id
=
mp_rank
,
group
=
_hcg
.
recv_prev_group
,
use_calc_stream
=
True
))
tasks
.
append
(
allgather_partial
(
d
,
nranks
=
mp_degree
,
rank_id
=
mp_rank
,
group
=
mp_group
,
use_calc_stream
=
True
))
else
:
tasks
.
append
(
recv_partial
(
tensor_recv_prev
,
...
...
@@ -369,17 +365,12 @@ def _p2p_helper(tensor_send_next, tensor_send_prev, recv_prev, recv_next):
rank_id
=
mp_rank
,
group
=
_hcg
.
recv_prev_group
,
use_calc_stream
=
True
))
tasks
.
append
(
allgather_partial
(
tensor_recv_prev
,
nranks
=
mp_degree
,
rank_id
=
mp_rank
,
group
=
mp_group
,
use_calc_stream
=
True
))
if
tensor_send_next
is
not
None
:
if
isinstance
(
tensor_send_next
,
tuple
):
for
d
in
tensor_send_next
:
paddle
.
distributed
.
wait
(
d
,
use_calc_stream
=
True
)
if
_in_legacy_dygraph
():
paddle
.
distributed
.
wait
(
d
,
use_calc_stream
=
True
)
tasks
.
append
(
send_partial
(
d
,
dst
=
1
,
...
...
@@ -388,7 +379,8 @@ def _p2p_helper(tensor_send_next, tensor_send_prev, recv_prev, recv_next):
group
=
_hcg
.
send_next_group
,
use_calc_stream
=
False
))
else
:
paddle
.
distributed
.
wait
(
tensor_send_next
,
use_calc_stream
=
True
)
if
_in_legacy_dygraph
():
paddle
.
distributed
.
wait
(
tensor_send_next
,
use_calc_stream
=
True
)
tasks
.
append
(
send_partial
(
tensor_send_next
,
dst
=
1
,
...
...
@@ -407,12 +399,6 @@ def _p2p_helper(tensor_send_next, tensor_send_prev, recv_prev, recv_next):
rank_id
=
mp_rank
,
group
=
_hcg
.
recv_next_group
,
use_calc_stream
=
True
))
tasks
.
append
(
allgather_partial
(
d
,
nranks
=
mp_degree
,
rank_id
=
mp_rank
,
group
=
mp_group
,
use_calc_stream
=
True
))
else
:
tasks
.
append
(
...
...
@@ -423,17 +409,40 @@ def _p2p_helper(tensor_send_next, tensor_send_prev, recv_prev, recv_next):
group
=
_hcg
.
recv_next_group
,
use_calc_stream
=
True
))
tasks
.
append
(
allgather_partial
(
tensor_recv_next
,
nranks
=
mp_degree
,
rank_id
=
mp_rank
,
group
=
mp_group
,
use_calc_stream
=
True
))
if
in_dygraph_mode
():
# wait
tasks in new
dygraph mode with new comm library
# wait
isend/irecv tasks in eager
dygraph mode with new comm library
for
task
in
tasks
:
if
task
is
not
None
:
task
.
wait
()
assert
task
is
not
None
task
.
wait
()
tensors_for_all_gather
=
[]
if
tensor_recv_prev
is
not
None
:
if
isinstance
(
tensor_recv_prev
,
tuple
):
for
d
in
tensor_recv_prev
:
tensors_for_all_gather
.
append
(
d
)
else
:
tensors_for_all_gather
.
append
(
tensor_recv_prev
)
if
tensor_recv_next
is
not
None
:
if
isinstance
(
tensor_recv_next
,
tuple
):
for
d
in
tensor_recv_next
:
tensors_for_all_gather
.
append
(
d
)
else
:
tensors_for_all_gather
.
append
(
tensor_recv_next
)
tasks
=
[]
for
tensor
in
tensors_for_all_gather
:
tasks
.
append
(
allgather_partial
(
tensor
,
nranks
=
mp_degree
,
rank_id
=
mp_rank
,
group
=
mp_group
,
use_calc_stream
=
True
))
for
task
in
tasks
:
# wait partial all gather tasks
if
task
is
not
None
:
task
.
wait
()
return
tensor_recv_prev
,
tensor_recv_next
...
...
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