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
a3790606
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 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
a3790606
编写于
9月 16, 2021
作者:
L
lilong12
提交者:
GitHub
9月 16, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
remove distributed attributes at the last stage for auto parallel (#35605)
* update
上级
2c70b844
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
22 addition
and
0 deletion
+22
-0
python/paddle/distributed/auto_parallel/parallelizer.py
python/paddle/distributed/auto_parallel/parallelizer.py
+15
-0
python/paddle/fluid/tests/unittests/test_auto_parallel_parallelizer.py
.../fluid/tests/unittests/test_auto_parallel_parallelizer.py
+7
-0
未找到文件。
python/paddle/distributed/auto_parallel/parallelizer.py
浏览文件 @
a3790606
...
...
@@ -14,6 +14,7 @@
import
paddle
from
paddle.distributed.fleet
import
cloud_utils
import
paddle.fluid.core
as
core
from
.context
import
DistributedContext
from
.context
import
get_default_distributed_context
from
.completion
import
complete_annotation
...
...
@@ -38,6 +39,16 @@ class AutoParallelizer:
# self._dist_context = DistributedContext()
self
.
_dist_context
=
get_default_distributed_context
()
def
_remove_distributed_attrs
(
self
,
main_program
):
suffix
=
core
.
kAutoParallelSuffix
()
# distributed attributes for variable have been removed
# in previous process.
for
block
in
main_program
.
blocks
:
for
op
in
block
.
ops
:
for
attr_name
in
op
.
attr_names
:
if
suffix
in
attr_name
:
op
.
_remove_attr
(
attr_name
)
def
parallelize
(
self
,
loss
,
startup_program
=
None
,
...
...
@@ -76,4 +87,8 @@ class AutoParallelizer:
for
process_group
in
all_process_groups
:
process_group
.
instantiate
()
# The last step: remove all distributed attributes to be compatiable
# with inference.
self
.
_remove_distributed_attrs
(
partitioned_main_prog
)
return
dist_optimize_ops
,
dist_params_grads
,
partitioned_startup_prog
,
partitioned_main_prog
python/paddle/fluid/tests/unittests/test_auto_parallel_parallelizer.py
浏览文件 @
a3790606
...
...
@@ -30,6 +30,7 @@ from paddle.fluid import layers
from
paddle.distributed
import
fleet
import
paddle.distributed.auto_parallel
as
auto
from
paddle.distributed.auto_parallel.utils
import
print_program_with_distributed_attr
import
paddle.fluid.core
as
core
paddle
.
enable_static
()
_global_parallel_strategy
=
None
...
...
@@ -83,6 +84,7 @@ def mlp_pretrain_forward(train_program, start_program):
name
=
"label"
,
shape
=
[
batch_size
,
sequence_len
,
1
],
dtype
=
'float32'
)
auto
.
shard_tensor
(
input
,
_global_process_mesh
,
dim_mapping
=
[
-
1
,
-
1
,
-
1
])
auto
.
set_pipeline_stage
(
1
)
mlp
=
MLPLayer
(
hidden_size
=
hidden_size
,
...
...
@@ -129,6 +131,11 @@ class TestMLPAutoParallelizer(unittest.TestCase):
optimizer
=
fleet
.
distributed_optimizer
(
optimizer
)
_
,
_
,
distributed_startup_program
,
distributed_main_program
=
optimizer
.
minimize
(
loss
,
start_program
)
suffix
=
core
.
kAutoParallelSuffix
()
for
block
in
distributed_main_program
.
blocks
:
for
op
in
block
.
ops
:
for
attr_name
in
op
.
attr_names
:
self
.
assertTrue
(
suffix
not
in
attr_name
)
# print_program_with_distributed_attr(distributed_main_program)
self
.
assertIsNotNone
(
distributed_startup_program
)
self
.
assertIsNotNone
(
distributed_main_program
)
...
...
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