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
fead5631
P
Paddle
项目概览
PaddlePaddle
/
Paddle
1 年多 前同步成功
通知
2302
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
fead5631
编写于
3月 24, 2021
作者:
L
Leo Chen
提交者:
GitHub
3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
[NPU] fix bug of lookup_table_v2_grad (#31834)
上级
149f76e6
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
38 addition
and
77 deletion
+38
-77
paddle/fluid/operators/controlflow/fetch_op.cc
paddle/fluid/operators/controlflow/fetch_op.cc
+5
-0
paddle/fluid/operators/lookup_table_v2_op_npu.cc
paddle/fluid/operators/lookup_table_v2_op_npu.cc
+14
-5
python/paddle/fluid/tests/unittests/npu/test_lookup_table_v2_op_npu.py
.../fluid/tests/unittests/npu/test_lookup_table_v2_op_npu.py
+19
-72
未找到文件。
paddle/fluid/operators/controlflow/fetch_op.cc
浏览文件 @
fead5631
...
@@ -44,6 +44,11 @@ static void DataCopy(const framework::LoDTensor &src_item,
...
@@ -44,6 +44,11 @@ static void DataCopy(const framework::LoDTensor &src_item,
TensorCopySync
(
src_item
,
platform
::
CPUPlace
(),
dst_item
);
TensorCopySync
(
src_item
,
platform
::
CPUPlace
(),
dst_item
);
}
}
#else
#else
#ifdef PADDLE_WITH_ASCEND_CL
if
(
platform
::
is_npu_place
(
src_item
.
place
()))
{
platform
::
DeviceContextPool
::
Instance
().
Get
(
src_item
.
place
())
->
Wait
();
}
#endif
TensorCopySync
(
src_item
,
platform
::
CPUPlace
(),
dst_item
);
TensorCopySync
(
src_item
,
platform
::
CPUPlace
(),
dst_item
);
#endif
#endif
}
else
{
}
else
{
...
...
paddle/fluid/operators/lookup_table_v2_op_npu.cc
浏览文件 @
fead5631
...
@@ -51,18 +51,27 @@ class LookupTableV2GradNPUKernel : public framework::OpKernel<T> {
...
@@ -51,18 +51,27 @@ class LookupTableV2GradNPUKernel : public framework::OpKernel<T> {
auto
*
ids_t
=
ctx
.
Input
<
framework
::
LoDTensor
>
(
"Ids"
);
auto
*
ids_t
=
ctx
.
Input
<
framework
::
LoDTensor
>
(
"Ids"
);
auto
*
output_grad_t
=
auto
*
output_grad_t
=
ctx
.
Input
<
framework
::
LoDTensor
>
(
framework
::
GradVarName
(
"Out"
));
ctx
.
Input
<
framework
::
LoDTensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
table_t
=
ctx
.
Input
<
framework
::
LoDTensor
>
(
"W"
);
auto
*
table_grad_t
=
auto
*
table_grad_t
=
ctx
.
Output
<
framework
::
LoDTensor
>
(
framework
::
GradVarName
(
"W"
));
ctx
.
Output
<
framework
::
LoDTensor
>
(
framework
::
GradVarName
(
"W"
));
table_grad_t
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
table_grad_t
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
framework
::
NPUAttributeMap
attr_input
=
{{
"use_locking"
,
true
}};
auto
runner
=
NpuOpRunner
(
"ScatterAdd"
,
{
*
table_t
,
*
ids_t
,
*
output_grad_t
},
{
*
table_grad_t
},
attr_input
);
auto
stream
=
auto
stream
=
ctx
.
template
device_context
<
paddle
::
platform
::
NPUDeviceContext
>()
ctx
.
template
device_context
<
paddle
::
platform
::
NPUDeviceContext
>()
.
stream
();
.
stream
();
runner
.
Run
(
stream
);
// step2: ZerosLike x in device
Tensor
zeroslike_w
(
table_grad_t
->
type
());
zeroslike_w
.
Resize
(
table_grad_t
->
dims
());
auto
p
=
zeroslike_w
.
mutable_data
<
T
>
(
ctx
.
GetPlace
());
platform
::
NPUMemsetAsync
(
static_cast
<
void
*>
(
p
),
0
,
zeroslike_w
.
numel
()
*
sizeof
(
T
),
stream
);
table_grad_t
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
runner_scatter
=
NpuOpRunner
(
"ScatterAdd"
,
{
zeroslike_w
,
*
ids_t
,
*
output_grad_t
},
{
*
table_grad_t
},
{});
runner_scatter
.
Run
(
stream
);
}
}
};
};
}
// namespace operators
}
// namespace operators
...
...
python/paddle/fluid/tests/unittests/npu/test_lookup_table_v2_op_npu.py
浏览文件 @
fead5631
...
@@ -36,19 +36,22 @@ class TestLookupTableV2(OpTest):
...
@@ -36,19 +36,22 @@ class TestLookupTableV2(OpTest):
self
.
init_dtype
()
self
.
init_dtype
()
np
.
random
.
seed
(
SEED
)
np
.
random
.
seed
(
SEED
)
bsz
=
2
bsz
=
6
seqlen
=
2
seqlen
=
8
vocab
=
3
vocab
=
10
dim
=
2
dim
=
20
w
=
np
.
ones
([
vocab
,
dim
]).
astype
(
self
.
dtype
)
w
=
np
.
ones
([
vocab
,
dim
]).
astype
(
self
.
dtype
)
x
=
np
.
random
.
randint
(
0
,
vocab
,
size
=
(
bsz
,
seqlen
)).
astype
(
np
.
int64
)
x
=
np
.
random
.
randint
(
0
,
vocab
,
size
=
(
bsz
,
seqlen
)).
astype
(
np
.
int64
)
out
=
np
.
ones
([
bsz
,
seqlen
,
dim
]).
astype
(
self
.
dtype
)
out
=
np
.
ones
([
bsz
,
seqlen
,
dim
]).
astype
(
self
.
dtype
)
self
.
inputs
=
{
'W'
:
OpTest
.
np_dtype_to_fluid_dtype
(
w
),
'Ids'
:
OpTest
.
np_dtype_to_fluid_dtype
(
x
)}
self
.
inputs
=
{
'W'
:
OpTest
.
np_dtype_to_fluid_dtype
(
w
),
'Ids'
:
OpTest
.
np_dtype_to_fluid_dtype
(
x
)
}
self
.
attrs
=
{
self
.
attrs
=
{
'is_sparse'
:
False
,
'is_sparse'
:
False
,
'is_distributed'
:
False
,
'is_distributed'
:
False
,
'remote_prefetch'
:
False
,
'remote_prefetch'
:
False
,
'padding_idx'
:
-
1
'padding_idx'
:
-
1
}
}
self
.
outputs
=
{
'Out'
:
out
}
self
.
outputs
=
{
'Out'
:
out
}
...
@@ -62,81 +65,25 @@ class TestLookupTableV2(OpTest):
...
@@ -62,81 +65,25 @@ class TestLookupTableV2(OpTest):
def
test_check_output
(
self
):
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
,
check_dygraph
=
False
)
self
.
check_output_with_place
(
self
.
place
,
check_dygraph
=
False
)
# TODO(ascendrc): Add grad test
def
test_check_grad
(
self
):
# def test_check_grad(self):
if
self
.
dtype
==
np
.
float16
:
# if self.dtype == np.float16:
return
# return
self
.
check_grad_with_place
(
# self.check_grad(['X'], 'Out')
self
.
place
,
[
'W'
],
'Out'
,
check_dygraph
=
False
)
@
unittest
.
skipIf
(
not
paddle
.
is_compiled_with_npu
(),
@
unittest
.
skipIf
(
not
paddle
.
is_compiled_with_npu
(),
"core is not compiled with NPU"
)
"core is not compiled with NPU"
)
class
TestLookupTableV2FP16
(
TestLookupTableV2
):
class
TestLookupTableV2FP16
(
TestLookupTableV2
):
no_need_check_grad
=
True
no_need_check_grad
=
True
def
init_dtype
(
self
):
def
init_dtype
(
self
):
self
.
dtype
=
np
.
float16
self
.
dtype
=
np
.
float16
#@unittest.skipIf(not paddle.is_compiled_with_npu(),
def
set_npu
(
self
):
# "core is not compiled with NPU")
self
.
__class__
.
use_npu
=
True
#class TestLookupTableV2Int8(TestLookupTableV2):
self
.
__class__
.
no_need_check_grad
=
True
# def init_dtype(self):
# self.dtype = np.int8
#
#@unittest.skipIf(not paddle.is_compiled_with_npu(),
# "core is not compiled with NPU")
#class TestLookupTableV2UInt8(TestLookupTableV2):
# def init_dtype(self):
# self.dtype = np.uint8
@
unittest
.
skipIf
(
not
paddle
.
is_compiled_with_npu
(),
"core is not compiled with NPU"
)
class
TestLookupTableV2Net
(
unittest
.
TestCase
):
def
_test
(
self
,
run_npu
=
True
):
main_prog
=
paddle
.
static
.
Program
()
startup_prog
=
paddle
.
static
.
Program
()
main_prog
.
random_seed
=
SEED
startup_prog
.
random_seed
=
SEED
np
.
random
.
seed
(
SEED
)
bsz
=
3
seqlen
=
2
vocab
=
3
dim
=
2
ids_np
=
np
.
random
.
randint
(
0
,
vocab
,
size
=
(
bsz
,
seqlen
)).
astype
(
'int64'
)
with
paddle
.
static
.
program_guard
(
main_prog
,
startup_prog
):
emb
=
paddle
.
nn
.
Embedding
(
vocab
,
dim
)
ids
=
paddle
.
static
.
data
(
name
=
"ids"
,
shape
=
[
bsz
,
seqlen
],
dtype
=
'int64'
)
res
=
emb
(
ids
)
loss
=
res
.
sum
()
if
run_npu
:
place
=
paddle
.
NPUPlace
(
0
)
else
:
place
=
paddle
.
CPUPlace
()
exe
=
paddle
.
static
.
Executor
(
place
)
exe
.
run
(
startup_prog
)
for
epoch
in
range
(
1
):
loss_res
,
w
=
exe
.
run
(
main_prog
,
feed
=
{
"ids"
:
ids_np
},
fetch_list
=
[
loss
,
emb
.
weight
])
if
epoch
%
10
==
0
:
print
(
w
)
print
(
"Epoch {} | Loss: {}"
.
format
(
epoch
,
loss
))
return
loss_res
def
test_npu
(
self
):
cpu_loss
=
self
.
_test
(
False
)
npu_loss
=
self
.
_test
(
True
)
self
.
assertTrue
(
np
.
allclose
(
npu_loss
,
cpu_loss
))
if
__name__
==
'__main__'
:
if
__name__
==
'__main__'
:
unittest
.
main
()
unittest
.
main
()
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