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
一WILLPOWER一
Paddle
提交
2f3ad5ab
P
Paddle
项目概览
一WILLPOWER一
/
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
2f3ad5ab
编写于
10月 24, 2022
作者:
F
feng_shuai
提交者:
GitHub
10月 24, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
optimize: vit static shape (#47280)
上级
84273aaa
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
41 addition
and
6 deletion
+41
-6
paddle/fluid/inference/tensorrt/convert/multihead_matmul_op.cc
...e/fluid/inference/tensorrt/convert/multihead_matmul_op.cc
+14
-5
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_multihead_matmul.py
...ittests/ir/inference/test_trt_convert_multihead_matmul.py
+27
-1
未找到文件。
paddle/fluid/inference/tensorrt/convert/multihead_matmul_op.cc
浏览文件 @
2f3ad5ab
...
...
@@ -460,9 +460,10 @@ class MultiheadMatMulOpConverter : public OpConverter {
plugin_inputs
.
emplace_back
(
mask_tensor
);
// input_2 for plugin
std
::
vector
<
int
>
pos_id
=
{
0
};
int
max_batch
=
500
;
int
max_batch
=
512
;
int
length
=
(
input_dims
.
d
[
1
]
==
-
1
)
?
1
:
input_dims
.
d
[
1
];
for
(
int
i
=
1
;
i
<
max_batch
;
i
++
)
{
pos_id
.
push_back
(
i
);
pos_id
.
push_back
(
i
*
length
);
}
nvinfer1
::
ITensor
*
fake_pos_id_tensor
=
Add1DConstantLayer
(
pos_id
);
nvinfer1
::
ITensor
*
length_tensor
=
...
...
@@ -497,18 +498,26 @@ class MultiheadMatMulOpConverter : public OpConverter {
stride
.
d
[
0
]
=
1
;
size
.
d
[
0
]
=
1
;
nvinfer1
::
ITensor
*
pos_id_tensor
=
(
input_dims
.
d
[
1
]
==
-
1
)
?
pos_id_layer
->
getOutput
(
0
)
:
fake_pos_id_tensor
;
auto
*
slice_pos_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Slice
,
*
pos_id_
layer
->
getOutput
(
0
)
,
start
,
size
,
stride
);
engine_
,
Slice
,
*
pos_id_
tensor
,
start
,
size
,
stride
);
slice_pos_layer
->
setInput
(
2
,
*
size_layer
->
getOutput
(
0
));
plugin_inputs
.
emplace_back
(
slice_pos_layer
->
getOutput
(
0
));
// input_3 for plugin
std
::
vector
<
int
>
data
(
500
,
1
);
int
max_length
=
(
input_dims
.
d
[
1
]
==
-
1
)
?
512
:
input_dims
.
d
[
1
];
std
::
vector
<
int
>
data
(
max_length
,
1
);
nvinfer1
::
ITensor
*
fake_max_seqlen_tensor
=
Add1DConstantLayer
(
data
);
auto
*
slice_max_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Slice
,
*
fake_max_seqlen_tensor
,
start
,
size
,
stride
);
slice_max_layer
->
setInput
(
2
,
*
length_tensor
);
plugin_inputs
.
emplace_back
(
slice_max_layer
->
getOutput
(
0
));
nvinfer1
::
ITensor
*
max_seqlen_tensor
=
(
input_dims
.
d
[
1
]
==
-
1
)
?
slice_max_layer
->
getOutput
(
0
)
:
fake_max_seqlen_tensor
;
plugin_inputs
.
emplace_back
(
max_seqlen_tensor
);
// plugin_layer
auto
plugin_layer
=
engine_
->
network
()
->
addPluginV2
(
plugin_inputs
.
data
(),
plugin_inputs
.
size
(),
*
plugin
);
...
...
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_multihead_matmul.py
浏览文件 @
2f3ad5ab
...
...
@@ -808,7 +808,7 @@ class TrtConvertVitToMultiHeadMatmulTest(TrtLayerAutoScanTest):
for
batch
in
[
2
,
4
]:
self
.
batch
=
batch
for
length
in
[
64
,
384
]:
for
length
in
[
197
]:
self
.
length
=
length
ops_config
=
[
{
...
...
@@ -1006,6 +1006,17 @@ class TrtConvertVitToMultiHeadMatmulTest(TrtLayerAutoScanTest):
"input_data1"
:
[
1
,
197
,
768
],
}
def
generate_static_shape
(
attrs
):
self
.
dynamic_shape
.
min_input_shape
=
{
"input_data1"
:
[
1
,
197
,
768
],
}
self
.
dynamic_shape
.
max_input_shape
=
{
"input_data1"
:
[
16
,
197
,
768
],
}
self
.
dynamic_shape
.
opt_input_shape
=
{
"input_data1"
:
[
1
,
197
,
768
],
}
def
clear_dynamic_shape
():
self
.
dynamic_shape
.
max_input_shape
=
{}
self
.
dynamic_shape
.
min_input_shape
=
{}
...
...
@@ -1035,6 +1046,21 @@ class TrtConvertVitToMultiHeadMatmulTest(TrtLayerAutoScanTest):
1e-5
,
)
# for static_shape
clear_dynamic_shape
()
generate_static_shape
(
attrs
)
self
.
trt_param
.
workspace_size
=
2013265920
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(),
(
1e-3
,
1e-3
,
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(),
(
1e-5
,
1e-5
,
)
def
add_skip_trt_case
(
self
):
pass
...
...
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