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
cee7a3db
P
Paddle
项目概览
PaddlePaddle
/
Paddle
11 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
cee7a3db
编写于
12月 05, 2022
作者:
S
ShenLiang
提交者:
GitHub
12月 05, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix bug of reducer in best_fit (#48668)
上级
89f024e3
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
39 addition
and
13 deletion
+39
-13
paddle/fluid/distributed/collective/reducer.cc
paddle/fluid/distributed/collective/reducer.cc
+22
-11
paddle/fluid/distributed/collective/reducer.h
paddle/fluid/distributed/collective/reducer.h
+1
-1
python/paddle/fluid/tests/unittests/test_parallel_dygraph_dataparallel.py
...uid/tests/unittests/test_parallel_dygraph_dataparallel.py
+16
-1
未找到文件。
paddle/fluid/distributed/collective/reducer.cc
浏览文件 @
cee7a3db
...
...
@@ -17,10 +17,16 @@
#include "paddle/phi/backends/device_manager.h"
DECLARE_bool
(
use_stream_safe_cuda_allocator
);
DECLARE_string
(
allocator_strategy
);
namespace
paddle
{
namespace
distributed
{
static
bool
IsStreamSafeAllocator
()
{
return
FLAGS_allocator_strategy
==
"auto_growth"
&&
FLAGS_use_stream_safe_cuda_allocator
;
}
static
Backend
TransToBackend
(
platform
::
Place
place
)
{
static
const
std
::
map
<
phi
::
AllocationType
,
Backend
>
type_backend
=
{
{
phi
::
AllocationType
::
GPU
,
Backend
::
GPU
},
...
...
@@ -399,14 +405,14 @@ void EagerGroup::ConcatTensors(const platform::Place &place) {
}
}
void
EagerGroup
::
SplitTensors
Dev
(
const
platform
::
DeviceContext
&
context
)
{
void
EagerGroup
::
SplitTensors
(
const
platform
::
DeviceContext
&
context
)
{
auto
place
=
context
.
GetPlace
();
if
(
platform
::
is_gpu_place
(
place
))
{
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
auto
&
gpu_context
=
static_cast
<
const
phi
::
GPUContext
&>
(
context
);
SplitTensorsWithType
(
gpu_context
,
&
dense_contents_
,
&
dense_tensors_
,
dtype_
);
if
(
FLAGS_use_stream_safe_cuda_allocator
)
{
if
(
IsStreamSafeAllocator
()
)
{
auto
dense_tensor
=
std
::
dynamic_pointer_cast
<
phi
::
DenseTensor
>
(
dense_contents_
.
impl
());
VLOG
(
3
)
<<
"Free dense_contents_ "
<<
dense_contents_
.
numel
();
...
...
@@ -1011,12 +1017,11 @@ void EagerReducer::FinalizeBackward() {
for
(
auto
&
group
:
groups_
)
{
if
(
!
group
.
is_sparse_
)
{
group
.
task
->
Synchronize
();
}
}
for
(
auto
&
group
:
groups_
)
{
if
(
!
group
.
is_sparse_
)
{
group
.
dense_contents_
.
reset
();
if
(
!
IsStreamSafeAllocator
())
{
auto
*
default_ctx
=
platform
::
DeviceContextPool
::
Instance
().
Get
(
inner_place_
);
group
.
SplitTensors
(
*
default_ctx
);
}
}
}
...
...
@@ -1054,9 +1059,15 @@ void EagerReducer::FusedAllReduceSchedule(EagerGroup *group,
group
->
task
=
process_group_
->
AllReduce
(
in_out
,
in_out
,
opts
);
auto
*
context
=
process_group_
->
GetDeviceContext
(
inner_place_
);
group
->
SplitTensorsDev
(
*
context
);
group
->
task
->
UpdateWaitChain
(
*
context
);
// split in FinalizeBackward()
if
(
IsStreamSafeAllocator
())
{
// NOTE(shenliang03): The best_fit allocator strategy is multi-stream
// insecure. In the Split operator, additional memory will be applied for
// calculation, and if it is asynchronous, an illegal memory access may be
// encountered.
group
->
SplitTensors
(
*
context
);
group
->
task
->
UpdateWaitChain
(
*
context
);
}
}
void
EagerReducer
::
AllReduceSparse
(
EagerGroup
*
group
,
...
...
paddle/fluid/distributed/collective/reducer.h
浏览文件 @
cee7a3db
...
...
@@ -75,7 +75,7 @@ class EagerGroup {
// context is used to select the stream for split
void
SplitTensors
Dev
(
const
platform
::
DeviceContext
&
);
void
SplitTensors
(
const
platform
::
DeviceContext
&
);
friend
std
::
ostream
&
operator
<<
(
std
::
ostream
&
,
const
EagerGroup
&
);
};
...
...
python/paddle/fluid/tests/unittests/test_parallel_dygraph_dataparallel.py
浏览文件 @
cee7a3db
...
...
@@ -103,6 +103,7 @@ def start_local_trainers(
training_script
,
training_script_args
,
eager_mode
=
True
,
allocator_strategy
=
"auto_growth"
,
log_dir
=
None
,
):
current_env
=
copy
.
copy
(
os
.
environ
.
copy
())
...
...
@@ -126,6 +127,10 @@ def start_local_trainers(
if
not
eager_mode
:
proc_env
[
"FLAGS_enable_eager_mode"
]
=
"%d"
%
0
proc_env
[
"FLAGS_allocator_strategy"
]
=
allocator_strategy
if
allocator_strategy
==
"auto_growth"
:
proc_env
[
"FLAGS_fraction_of_gpu_memory_to_use"
]
=
"0.1"
current_env
.
update
(
proc_env
)
print
(
"trainer proc env:{}"
.
format
(
current_env
))
...
...
@@ -153,7 +158,12 @@ def start_local_trainers(
class
TestMultipleGpus
(
unittest
.
TestCase
):
def
run_mnist_2gpu
(
self
,
target_file_name
,
eager_mode
=
True
):
def
run_mnist_2gpu
(
self
,
target_file_name
,
eager_mode
=
True
,
allocator_strategy
=
"auto_growth"
,
):
if
(
not
fluid
.
core
.
is_compiled_with_cuda
()
or
fluid
.
core
.
get_cuda_device_count
()
==
0
...
...
@@ -170,6 +180,7 @@ class TestMultipleGpus(unittest.TestCase):
cluster
,
pod
,
eager_mode
=
eager_mode
,
allocator_strategy
=
allocator_strategy
,
training_script
=
target_file_name
,
training_script_args
=
[],
)
...
...
@@ -218,6 +229,10 @@ class TestDataParallelWithPyLayer(TestMultipleGpus):
self
.
run_mnist_2gpu
(
'parallel_dygraph_dataparallel_with_pylayer.py'
,
eager_mode
=
False
)
self
.
run_mnist_2gpu
(
'parallel_dygraph_dataparallel_with_pylayer.py'
,
allocator_strategy
=
"naive_best_fit"
,
)
class
TestGradientCheckInEagerMode
(
TestMultipleGpus
):
...
...
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