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
5b737834
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2299
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
未验证
提交
5b737834
编写于
8月 24, 2021
作者:
G
gongweibao
提交者:
GitHub
8月 24, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add flags to control whether to check Nan value of hccl_allreduce_sum. (#35093)
上级
a8dee3bb
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
14 addition
and
3 deletion
+14
-3
paddle/fluid/operators/collective/c_allreduce_op.h
paddle/fluid/operators/collective/c_allreduce_op.h
+10
-3
paddle/fluid/platform/flags.cc
paddle/fluid/platform/flags.cc
+3
-0
python/paddle/fluid/__init__.py
python/paddle/fluid/__init__.py
+1
-0
未找到文件。
paddle/fluid/operators/collective/c_allreduce_op.h
浏览文件 @
5b737834
...
...
@@ -45,6 +45,10 @@ limitations under the License. */
#include "paddle/fluid/platform/hccl_helper.h"
#endif
#if defined(PADDLE_WITH_ASCEND_CL)
DECLARE_bool
(
hccl_check_nan
);
#endif
namespace
paddle
{
namespace
operators
{
...
...
@@ -140,6 +144,7 @@ inline bool ContainsNan(const paddle::platform::NPUDeviceContext& dev_ctx,
try
{
const
auto
&
runner_mean
=
paddle
::
operators
::
NpuOpRunner
(
"ReduceMeanD"
,
{
*
in
},
{
mean
},
{{
"axes"
,
axes
},
{
"keep_dims"
,
false
}});
runner_mean
.
Run
(
stream
);
TensorToVector
(
mean
,
dev_ctx
,
&
vec
);
}
catch
(...)
{
LOG
(
WARNING
)
<<
"ContainsNan catch exception"
;
...
...
@@ -233,9 +238,11 @@ class CAllReduceOpASCENDKernel : public framework::OpKernel<T> {
break
;
}
case
framework
::
proto
::
VarType
::
FP32
:
{
VLOG
(
4
)
<<
"prepare to FoundNanInf"
;
found_nan
=
ContainsNan
(
*
dev_ctx
,
dev_ctx
->
stream
(),
in
);
VLOG
(
4
)
<<
"check_numerics:"
<<
found_nan
;
if
(
FLAGS_hccl_check_nan
)
{
VLOG
(
3
)
<<
"prepare to FoundNanInf"
;
found_nan
=
ContainsNan
(
*
dev_ctx
,
dev_ctx
->
stream
(),
in
);
VLOG
(
3
)
<<
"check_numerics:"
<<
found_nan
;
}
break
;
}
default:
...
...
paddle/fluid/platform/flags.cc
浏览文件 @
5b737834
...
...
@@ -93,6 +93,9 @@ DEFINE_string(selected_npus, "",
"This option is useful when doing multi process training and "
"each process have only one device (NPU). If you want to use "
"all visible devices, set this to empty string."
);
DEFINE_bool
(
hccl_check_nan
,
false
,
"Check Nan in tensor before hccl_allreduce_sum otherwise it'll "
"core when meets Nan value"
);
DEFINE_string
(
npu_config_path
,
""
,
"The absolute path of configuration json file, like: /tmp/config.json. "
...
...
python/paddle/fluid/__init__.py
浏览文件 @
5b737834
...
...
@@ -248,6 +248,7 @@ def __bootstrap__():
'gpu_memory_limit_mb'
,
'npu_config_path'
,
'get_host_by_name_time'
,
'hccl_check_nan'
,
]
core
.
init_gflags
([
"--tryfromenv="
+
","
.
join
(
read_env_flags
)])
...
...
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