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
SummerGao.
Paddle
提交
86bbb0f2
P
Paddle
项目概览
SummerGao.
/
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
86bbb0f2
编写于
4月 15, 2022
作者:
L
lilong12
提交者:
GitHub
4月 15, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add backend for heter training (#41526) (#41651)
上级
9f2ae360
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
26 addition
and
1 deletion
+26
-1
python/paddle/distributed/collective.py
python/paddle/distributed/collective.py
+26
-1
未找到文件。
python/paddle/distributed/collective.py
浏览文件 @
86bbb0f2
...
...
@@ -138,7 +138,7 @@ _group_map_by_name = {}
# Name of the default group for init_parallel_env
_default_group_name
=
"_default_pg"
_valid_backend_list
=
[
'nccl'
,
'gloo'
,
'hccl'
]
_valid_backend_list
=
[
'nccl'
,
'gloo'
,
'hccl'
,
'heter'
]
_default_store
=
None
# the default tcp store
_default_backend
=
None
...
...
@@ -234,6 +234,31 @@ def _new_process_group_impl(backend,
pg
=
core
.
ProcessGroupNCCL
(
store
,
rank
,
world_size
,
group_id
)
elif
backend
==
"hccl"
:
pg
=
core
.
ProcessGroupHCCL
(
store
,
rank
,
world_size
,
group_id
)
elif
backend
==
"heter"
:
cluster_id
=
int
(
os
.
getenv
(
"CLUSTER_ID"
,
"-1"
))
assert
cluster_id
>=
0
,
"please set the CLUSTER_ID variable."
cluster_size
=
os
.
getenv
(
"CLUSTER_SIZE"
,
None
)
assert
cluster_size
,
"please set the CLUSTER_SIZE variable."
cluster_size
=
cluster_size
.
split
(
","
)
cluster_size
=
[
int
(
s
)
for
s
in
cluster_size
]
switch_ep
=
os
.
getenv
(
"CLUSTER_SWITCH"
,
None
)
assert
switch_ep
,
"please set the CLUSTER_SWITCH variable."
cluster_size_cumsum
=
np
.
cumsum
(
cluster_size
)
cluster_offset
=
0
if
cluster_id
==
0
else
cluster_size_cumsum
[
cluster_id
-
1
]
global_rank
=
cluster_offset
+
rank
global_world_size
=
cluster_size_cumsum
[
-
1
]
pg
=
core
.
ProcessGroupHeter
(
store
,
rank
=
global_rank
,
world_size
=
global_world_size
,
gid
=
0
,
local_rank
=
rank
,
local_size
=
world_size
,
gloo_rank
=
cluster_id
,
gloo_size
=
len
(
cluster_size
),
with_switch
=
True
,
switch_endpoint
=
switch_ep
)
return
pg
...
...
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