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
bc3ca678
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2292
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
bc3ca678
编写于
2月 18, 2022
作者:
B
Baibaifan
提交者:
GitHub
2月 18, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix sharding group (#39668)
* fix_sharding_group * fix_sharding_group
上级
e674af23
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
9 addition
and
6 deletion
+9
-6
python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/sharding_optimizer_stage2.py
...optimizers/dygraph_optimizer/sharding_optimizer_stage2.py
+2
-2
python/paddle/distributed/fleet/meta_parallel/sharding/sharding_stage2.py
...stributed/fleet/meta_parallel/sharding/sharding_stage2.py
+5
-3
python/paddle/distributed/fleet/meta_parallel/sharding/sharding_stage3.py
...stributed/fleet/meta_parallel/sharding/sharding_stage3.py
+2
-1
未找到文件。
python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/sharding_optimizer_stage2.py
浏览文件 @
bc3ca678
...
...
@@ -98,7 +98,7 @@ class ShardingOptimizerStage2(Optimizer):
self
.
world_size
=
self
.
group
.
nranks
self
.
rank
=
self
.
group
.
rank
self
.
_global_root_rank
=
0
self
.
_global_root_rank
=
self
.
group
.
ranks
[
0
]
# Synchronous all ranks models
if
pertrain_sync_models
:
...
...
@@ -403,7 +403,7 @@ class ShardingOptimizerStage2(Optimizer):
for
dst_rank
,
internal_storage
in
dtype_per_rank
.
items
():
dist
.
broadcast
(
tensor
=
internal_storage
.
buffer
,
src
=
dst_rank
,
src
=
self
.
group
.
ranks
[
dst_rank
]
,
group
=
self
.
group
,
use_calc_stream
=
True
)
...
...
python/paddle/distributed/fleet/meta_parallel/sharding/sharding_stage2.py
浏览文件 @
bc3ca678
...
...
@@ -85,7 +85,8 @@ class ShardingStage2(nn.Layer):
self
.
_world_size_scaling
=
1.0
/
self
.
_group
.
nranks
assert
self
.
_group
.
nranks
>
1
,
"Training must be distributed, ranks must be greater than 1"
self
.
_rank
=
self
.
_group
.
rank
self
.
_global_root_rank
=
0
# picking rank 0 as the reference
self
.
_global_root_rank
=
self
.
_group
.
ranks
[
0
]
# picking rank 0 as the reference
self
.
_default_device
=
device
# Global statistical parameters
...
...
@@ -319,7 +320,7 @@ class ShardingStage2(nn.Layer):
Taskflow
(
task
=
dist
.
reduce
(
tensor
=
param
.
grad
,
dst
=
dst_rank
,
dst
=
self
.
_group
.
ranks
[
dst_rank
]
,
group
=
self
.
_group
,
use_calc_stream
=
True
),
callback
=
cleanup
))
...
...
@@ -377,7 +378,8 @@ class ShardingStage2(nn.Layer):
Taskflow
(
task
=
dist
.
reduce
(
tensor
=
grad_storage
.
buffer
,
dst
=
grad_storage
.
destination
,
dst
=
self
.
_group
.
ranks
[
grad_storage
.
destination
],
group
=
self
.
_group
,
use_calc_stream
=
True
),
callback
=
cleanup
))
...
...
python/paddle/distributed/fleet/meta_parallel/sharding/sharding_stage3.py
浏览文件 @
bc3ca678
...
...
@@ -101,7 +101,8 @@ class ShardingStage3(nn.Layer):
self
.
_world_size_scaling
=
1.0
/
self
.
_group
.
nranks
assert
self
.
_group
.
nranks
>
1
,
"Training must be distributed, ranks must be greater than 1."
self
.
_rank
=
self
.
_group
.
rank
self
.
_global_root_rank
=
0
# picking rank 0 as the reference
self
.
_global_root_rank
=
self
.
_group
.
ranks
[
0
]
# picking rank 0 as the reference
self
.
_global_ranks
=
self
.
_group
.
ranks
# Parameter segmentation for global ranks
...
...
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