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
6833ecfe
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
6833ecfe
编写于
9月 14, 2022
作者:
S
sneaxiy
提交者:
GitHub
9月 14, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix DistributedFusedLAMB NaN problem (#46011)
* fix distributed_fused_lamb nan * remove CUDA_ASSERT
上级
65dd828e
变更
1
显示空白变更内容
内联
并排
Showing
1 changed file
with
59 addition
and
10 deletion
+59
-10
paddle/fluid/operators/optimizers/distributed_fused_lamb_op.cu
...e/fluid/operators/optimizers/distributed_fused_lamb_op.cu
+59
-10
未找到文件。
paddle/fluid/operators/optimizers/distributed_fused_lamb_op.cu
浏览文件 @
6833ecfe
...
...
@@ -1193,6 +1193,38 @@ static void PrintAllMinMaxRange(const framework::ExecutionContext &ctx,
}
}
template
<
typename
T
>
static
bool
HasNanInf
(
const
phi
::
GPUContext
&
dev_ctx
,
const
T
*
x
,
int
numel
)
{
if
(
numel
<=
0
)
return
false
;
cub
::
TransformInputIterator
<
bool
,
IsNanInfFunctor
<
T
>
,
const
T
*>
iter
(
x
,
IsNanInfFunctor
<
T
>
());
memory
::
Buffer
buffer
(
dev_ctx
.
GetPlace
());
memory
::
Buffer
out
(
dev_ctx
.
GetPlace
());
CubDeviceReduce
(
iter
,
out
.
Alloc
<
bool
>
(
1
),
numel
,
OrFunctor
(),
false
,
dev_ctx
.
stream
(),
&
buffer
);
bool
flag
;
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_GPU_SUCCESS
(
hipMemcpyAsync
(
&
flag
,
out
.
Get
<
bool
>
(),
sizeof
(
flag
),
hipMemcpyDeviceToHost
,
dev_ctx
.
stream
()));
#else
PADDLE_ENFORCE_GPU_SUCCESS
(
cudaMemcpyAsync
(
&
flag
,
out
.
Get
<
bool
>
(),
sizeof
(
flag
),
cudaMemcpyDeviceToHost
,
dev_ctx
.
stream
()));
#endif
dev_ctx
.
Wait
();
return
flag
;
}
static
void
CheckHasNanInfGrad
(
const
float
*
fp32_grad
,
int
fp32_numel
,
const
platform
::
float16
*
fp16_grad
,
...
...
@@ -1830,17 +1862,11 @@ class DistributedFusedLambOpKernel<phi::GPUContext, T>
}
else
{
VLOG
(
1
)
<<
"Grad scale: "
<<
FlattenToString
(
fp16_scale
,
1
,
place
);
}
if
(
nranks
>
1
)
{
PADDLE_ENFORCE_GPU_SUCCESS
(
platform
::
dynload
::
ncclAllReduce
(
fp32_square_grad_norm
,
fp32_square_grad_norm
,
1
,
ncclFloat32
,
ncclSum
,
global_comm
,
stream
));
}
// (3) Do ReduceScatter with scale
VLOG
(
1
)
<<
"FP32 HasNanInf before all reduce: "
<<
HasNanInf
(
dev_ctx
,
fp32_grad
,
fp32_numel
);
VLOG
(
1
)
<<
"FP16 HasNanInf before all reduce: "
<<
HasNanInf
(
dev_ctx
,
fp16_grad
,
fp16_numel
);
if
(
local_shard
)
{
if
(
use_hierarchical_allreduce
)
{
NCCLReduceScatterWithScale
(
...
...
@@ -1916,6 +1942,29 @@ class DistributedFusedLambOpKernel<phi::GPUContext, T>
dev_ctx
,
fp16_scale
);
}
VLOG
(
1
)
<<
"FP32 HasNanInf after all reduce: "
<<
HasNanInf
(
dev_ctx
,
fp32_sum_grad
,
fp32_numel_each_device
);
VLOG
(
1
)
<<
"FP16 HasNanInf after all reduce: "
<<
HasNanInf
(
dev_ctx
,
fp16_sum_grad
,
fp16_numel_each_device
);
CheckHasNanInfGrad
(
fp32_sum_grad
,
fp32_numel_each_device
,
fp16_sum_grad
,
fp16_numel_each_device
,
fp32_square_grad_norm
,
stream
,
&
cub_tmp_buffer
);
if
(
num_devices
>
1
)
{
PADDLE_ENFORCE_GPU_SUCCESS
(
platform
::
dynload
::
ncclAllReduce
(
fp32_square_grad_norm
,
fp32_square_grad_norm
,
1
,
ncclFloat32
,
ncclSum
,
local_comm
,
stream
));
VLOG
(
1
)
<<
"Grad square norm after all reduce: "
<<
FlattenToString
(
fp32_square_grad_norm
,
1
,
place
);
}
// (4) mark max_global_grad_norm as 0, meaning that clip has been
// already performed
max_global_grad_norm
=
0
;
...
...
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