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
weixin_46846685
Paddle
提交
687219fe
P
Paddle
项目概览
weixin_46846685
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
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
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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
687219fe
编写于
4月 28, 2022
作者:
W
WangXi
提交者:
GitHub
4月 28, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix FusedResidualDropoutBias nan in v100 (#42344)
上级
8ad38701
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
30 addition
and
3 deletion
+30
-3
paddle/fluid/operators/fused/fused_dropout_common.h
paddle/fluid/operators/fused/fused_dropout_common.h
+11
-3
paddle/fluid/operators/fused/fused_residual_dropout_bias_test.cu
...fluid/operators/fused/fused_residual_dropout_bias_test.cu
+19
-0
未找到文件。
paddle/fluid/operators/fused/fused_dropout_common.h
浏览文件 @
687219fe
...
@@ -43,9 +43,17 @@ inline platform::GpuLaunchConfig Get1DBlocksAnd2DGrids(
...
@@ -43,9 +43,17 @@ inline platform::GpuLaunchConfig Get1DBlocksAnd2DGrids(
const
platform
::
CUDADeviceContext
&
ctx
,
const
uint32_t
rows
,
const
platform
::
CUDADeviceContext
&
ctx
,
const
uint32_t
rows
,
const
uint32_t
cols
,
const
int
vec_size
)
{
const
uint32_t
cols
,
const
int
vec_size
)
{
const
uint32_t
tmp_cols
=
cols
/
vec_size
;
const
uint32_t
tmp_cols
=
cols
/
vec_size
;
int
threads
=
std
::
max
(
// NOTE(wangxi): We set max_block_size to 512, for `FusedResidualDropoutBias`
static_cast
<
uint32_t
>
(
32
),
// needs too many register resources. If data_type is float16, CUDA
std
::
min
(
tmp_cols
,
static_cast
<
uint32_t
>
(
ctx
.
GetMaxThreadsPerBlock
())));
// error(701) will occur when block_size is 1024. Which error is
// 'cudaErrorLaunchOutOfResources', this indicates that a launch did not
// occur because it did not have appropriate resources.
// Of course, this kernel can be optimized later to reduce the use
// of registers.
int
threads
=
std
::
max
(
static_cast
<
uint32_t
>
(
32
),
std
::
min
(
tmp_cols
,
static_cast
<
uint32_t
>
(
std
::
min
(
ctx
.
GetMaxThreadsPerBlock
(),
512
))));
const
auto
blocks_x
=
const
auto
blocks_x
=
std
::
max
(
static_cast
<
uint32_t
>
(
1
),
(
tmp_cols
+
threads
-
1
)
/
threads
);
std
::
max
(
static_cast
<
uint32_t
>
(
1
),
(
tmp_cols
+
threads
-
1
)
/
threads
);
const
auto
blocks_y
=
std
::
max
(
static_cast
<
uint32_t
>
(
1
),
rows
);
const
auto
blocks_y
=
std
::
max
(
static_cast
<
uint32_t
>
(
1
),
rows
);
...
...
paddle/fluid/operators/fused/fused_residual_dropout_bias_test.cu
浏览文件 @
687219fe
...
@@ -147,6 +147,7 @@ struct TestFusedResidualDropoutBias {
...
@@ -147,6 +147,7 @@ struct TestFusedResidualDropoutBias {
dropout_prob
,
is_upscale_in_train
,
is_test
);
dropout_prob
,
is_upscale_in_train
,
is_test
);
}
}
ctx
->
Wait
();
ctx
->
Wait
();
PADDLE_ENFORCE_GPU_SUCCESS
(
platform
::
GpuGetLastError
());
// add residual
// add residual
for
(
int
i
=
0
;
i
<
rows
;
i
++
)
{
for
(
int
i
=
0
;
i
<
rows
;
i
++
)
{
for
(
int
j
=
0
;
j
<
cols
;
j
++
)
{
for
(
int
j
=
0
;
j
<
cols
;
j
++
)
{
...
@@ -186,6 +187,7 @@ struct TestFusedResidualDropoutBias {
...
@@ -186,6 +187,7 @@ struct TestFusedResidualDropoutBias {
src
.
data
<
T
>
(),
residual
.
data
<
T
>
(),
bias_ptr
,
mask
.
data
<
uint8_t
>
(),
src
.
data
<
T
>
(),
residual
.
data
<
T
>
(),
bias_ptr
,
mask
.
data
<
uint8_t
>
(),
out
.
data
<
T
>
(),
*
ctx
);
out
.
data
<
T
>
(),
*
ctx
);
ctx
->
Wait
();
ctx
->
Wait
();
PADDLE_ENFORCE_GPU_SUCCESS
(
platform
::
GpuGetLastError
());
}
}
void
FusedBackward
()
{
void
FusedBackward
()
{
...
@@ -313,3 +315,20 @@ TEST(FusedDropout, GPUFusedResidualDropoutBiasLargeShape) {
...
@@ -313,3 +315,20 @@ TEST(FusedDropout, GPUFusedResidualDropoutBiasLargeShape) {
test
.
CheckOut
(
static_cast
<
float
>
(
1e-5
));
test
.
CheckOut
(
static_cast
<
float
>
(
1e-5
));
test
.
CheckGrad
(
static_cast
<
float
>
(
1e-3
));
test
.
CheckGrad
(
static_cast
<
float
>
(
1e-3
));
}
}
TEST
(
FusedDropout
,
GPUFusedResidualDropoutBiasLargeShapeFp16
)
{
// Used to test that `cudaErrorLaunchOutOfResources` will not occur
int
rows
=
1
;
int
cols
=
12288
;
if
(
std
::
getenv
(
"_rows"
)
!=
nullptr
)
{
rows
=
atoi
(
std
::
getenv
(
"_rows"
));
}
if
(
std
::
getenv
(
"_cols"
)
!=
nullptr
)
{
cols
=
atoi
(
std
::
getenv
(
"_cols"
));
}
TestFusedResidualDropoutBias
<
platform
::
float16
>
test
(
rows
,
cols
,
0
,
0.0
,
true
,
true
);
test
.
Run
();
test
.
CheckOut
(
static_cast
<
platform
::
float16
>
(
1e-1
));
test
.
CheckGrad
(
static_cast
<
platform
::
float16
>
(
1e-1
));
}
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