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
1fbd4440
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
1fbd4440
编写于
6月 08, 2022
作者:
W
Wangzheee
提交者:
GitHub
6月 08, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[Paddle-Inference]support matmulv2 in multihead (#43269)
* support matmulv2 in multihead
上级
e1a34bc4
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
42 addition
and
22 deletion
+42
-22
paddle/fluid/framework/ir/trt_multihead_matmul_fuse_pass.cc
paddle/fluid/framework/ir/trt_multihead_matmul_fuse_pass.cc
+42
-22
未找到文件。
paddle/fluid/framework/ir/trt_multihead_matmul_fuse_pass.cc
浏览文件 @
1fbd4440
...
...
@@ -235,16 +235,18 @@ static int BuildFusion(Graph* graph, const std::string& name_scope) {
}
PDNode
*
TrtMultiHeadMatmulPattern
::
operator
()()
{
std
::
unordered_set
<
std
::
string
>
mul_ops
{
"mul"
,
"matmul_v2"
};
std
::
unordered_set
<
std
::
string
>
matmul_ops
{
"matmul"
,
"matmul_v2"
};
auto
*
input0
=
pattern
->
NewNode
(
input0_repr
());
input0
->
assert_is_op
_input
(
"mul"
);
input0
->
assert_is_op
s_input
(
mul_ops
);
// First path with scale
auto
*
mul0
=
pattern
->
NewNode
(
mul0_repr
())
->
assert_is_op
(
"mul"
);
auto
*
mul0
=
pattern
->
NewNode
(
mul0_repr
())
->
assert_is_op
s
(
mul_ops
);
auto
*
mul0_w_var
=
pattern
->
NewNode
(
mul0_w_repr
())
->
AsInput
()
->
assert_is_op
_input
(
"mul"
,
"Y"
);
->
assert_is_op
s_input
(
mul_ops
,
"Y"
);
auto
*
mul0_out_var
=
pattern
->
NewNode
(
mul0_out_repr
())
->
assert_is_op
_output
(
"mul"
);
pattern
->
NewNode
(
mul0_out_repr
())
->
assert_is_op
s_output
(
mul_ops
);
decltype
(
mul0
)
eltadd0
;
decltype
(
mul0
)
eltadd0_b_var
;
...
...
@@ -277,11 +279,12 @@ PDNode* TrtMultiHeadMatmulPattern::operator()() {
auto
*
scale
=
pattern
->
NewNode
(
scale_repr
())
->
assert_is_op
(
"scale"
);
auto
*
scale_out_var
=
pattern
->
NewNode
(
scale_out_repr
())
->
assert_is_op_output
(
"scale"
);
scale_out_var
->
AsIntermediate
()
->
assert_is_op
_input
(
"matmul"
);
scale_out_var
->
AsIntermediate
()
->
assert_is_op
s_input
(
matmul_ops
);
auto
*
matmul_qk
=
pattern
->
NewNode
(
matmul_qk_repr
())
->
assert_is_op
(
"matmul"
);
auto
*
matmul_qk
=
pattern
->
NewNode
(
matmul_qk_repr
())
->
assert_is_ops
(
matmul_ops
);
auto
*
matmul_qk_out_var
=
pattern
->
NewNode
(
matmul_qk_out_repr
())
->
assert_is_op
_output
(
"matmul"
);
pattern
->
NewNode
(
matmul_qk_out_repr
())
->
assert_is_op
s_output
(
matmul_ops
);
matmul_qk_out_var
->
AsIntermediate
()
->
assert_is_op_input
(
"elementwise_add"
);
auto
*
eltadd_qk
=
...
...
@@ -297,12 +300,12 @@ PDNode* TrtMultiHeadMatmulPattern::operator()() {
pattern
->
NewNode
(
softmax_qk_repr
())
->
assert_is_op
(
"softmax"
);
auto
*
softmax_qk_out_var
=
pattern
->
NewNode
(
softmax_qk_out_repr
())
->
assert_is_op_output
(
"softmax"
);
softmax_qk_out_var
->
AsIntermediate
()
->
assert_is_op
_input
(
"matmul"
);
softmax_qk_out_var
->
AsIntermediate
()
->
assert_is_op
s_input
(
matmul_ops
);
auto
*
matmul_qkv
=
pattern
->
NewNode
(
matmul_qkv_repr
())
->
assert_is_op
(
"matmul"
);
pattern
->
NewNode
(
matmul_qkv_repr
())
->
assert_is_op
s
(
matmul_ops
);
auto
*
matmul_qkv_out_var
=
pattern
->
NewNode
(
matmul_qkv_out_repr
())
->
assert_is_op
_output
(
"matmul"
);
pattern
->
NewNode
(
matmul_qkv_out_repr
())
->
assert_is_op
s_output
(
matmul_ops
);
matmul_qkv_out_var
->
AsIntermediate
()
->
assert_is_op_input
(
"transpose2"
);
auto
*
transpose2_qkv
=
...
...
@@ -315,15 +318,15 @@ PDNode* TrtMultiHeadMatmulPattern::operator()() {
pattern
->
NewNode
(
reshape2_qkv_repr
())
->
assert_is_op
(
"reshape2"
);
auto
*
reshape2_qkv_out_var
=
pattern
->
NewNode
(
reshape2_qkv_out_repr
())
->
assert_is_op_output
(
"reshape2"
);
reshape2_qkv_out_var
->
assert_is_op
_input
(
"mul"
);
reshape2_qkv_out_var
->
assert_is_op
s_input
(
mul_ops
);
// Second path to matmul
auto
*
mul1
=
pattern
->
NewNode
(
mul1_repr
())
->
assert_is_op
(
"mul"
);
auto
*
mul1
=
pattern
->
NewNode
(
mul1_repr
())
->
assert_is_op
s
(
mul_ops
);
auto
*
mul1_w_var
=
pattern
->
NewNode
(
mul1_w_repr
())
->
AsInput
()
->
assert_is_op
_input
(
"mul"
,
"Y"
);
->
assert_is_op
s_input
(
mul_ops
,
"Y"
);
auto
*
mul1_out_var
=
pattern
->
NewNode
(
mul1_out_repr
())
->
assert_is_op
_output
(
"mul"
);
pattern
->
NewNode
(
mul1_out_repr
())
->
assert_is_op
s_output
(
mul_ops
);
decltype
(
mul1
)
eltadd1
;
decltype
(
mul1
)
eltadd1_b_var
;
...
...
@@ -350,16 +353,16 @@ PDNode* TrtMultiHeadMatmulPattern::operator()() {
pattern
->
NewNode
(
transpose2_1_repr
())
->
assert_is_op
(
"transpose2"
);
auto
*
transpose2_1_out_var
=
pattern
->
NewNode
(
transpose2_1_out_repr
())
->
assert_is_op_output
(
"transpose2"
);
transpose2_1_out_var
->
AsIntermediate
()
->
assert_is_op_input
(
"matmul"
);
// link to matmul qk
transpose2_1_out_var
->
AsIntermediate
()
->
assert_is_op
s
_input
(
matmul_ops
);
// link to matmul qk
// Third path to matmul
auto
*
mul2
=
pattern
->
NewNode
(
mul2_repr
())
->
assert_is_op
(
"mul"
);
auto
*
mul2
=
pattern
->
NewNode
(
mul2_repr
())
->
assert_is_op
s
(
mul_ops
);
auto
*
mul2_w_var
=
pattern
->
NewNode
(
mul2_w_repr
())
->
AsInput
()
->
assert_is_op
_input
(
"mul"
,
"Y"
);
->
assert_is_op
s_input
(
mul_ops
,
"Y"
);
auto
*
mul2_out_var
=
pattern
->
NewNode
(
mul2_out_repr
())
->
assert_is_op
_output
(
"mul"
);
pattern
->
NewNode
(
mul2_out_repr
())
->
assert_is_op
s_output
(
mul_ops
);
decltype
(
mul2
)
eltadd2
;
decltype
(
mul2
)
eltadd2_b_var
;
...
...
@@ -386,8 +389,8 @@ PDNode* TrtMultiHeadMatmulPattern::operator()() {
pattern
->
NewNode
(
transpose2_2_repr
())
->
assert_is_op
(
"transpose2"
);
auto
*
transpose2_2_out_var
=
pattern
->
NewNode
(
transpose2_2_out_repr
())
->
assert_is_op_output
(
"transpose2"
);
transpose2_2_out_var
->
AsIntermediate
()
->
assert_is_op_input
(
"matmul"
);
// link to matmul qkv
transpose2_2_out_var
->
AsIntermediate
()
->
assert_is_op
s
_input
(
matmul_ops
);
// link to matmul qkv
// Q path
mul0
->
LinksFrom
({
input0
,
mul0_w_var
}).
LinksTo
({
mul0_out_var
});
...
...
@@ -734,6 +737,23 @@ TrtMultiHeadMatmulV2FusePass::TrtMultiHeadMatmulV2FusePass() {
.
IsType
<
bool
>
()
.
End
();
AddOpCompat
(
OpCompat
(
"matmul_v2"
))
.
AddInput
(
"X"
)
.
IsTensor
()
.
End
()
.
AddInput
(
"Y"
)
.
IsTensor
()
.
End
()
.
AddOutput
(
"Out"
)
.
IsTensor
()
.
End
()
.
AddAttr
(
"trans_x"
)
.
IsType
<
bool
>
()
.
End
()
.
AddAttr
(
"trans_y"
)
.
IsType
<
bool
>
()
.
End
();
AddOpCompat
(
OpCompat
(
"softmax"
))
.
AddInput
(
"X"
)
.
IsTensor
()
...
...
@@ -866,7 +886,7 @@ int TrtMultiHeadMatmulV2FusePass::BuildFusionV2(Graph* graph,
auto
*
mul0_op_desc
=
mul0
->
Op
();
// all mul op has same input.
if
(
mul
tihead_op_desc
.
HasAttr
(
"Input_scale"
))
{
if
(
mul
0_op_desc
->
HasAttr
(
"Input_scale"
))
{
multihead_op_desc
.
SetAttr
(
"Input_scale"
,
mul0_op_desc
->
GetAttr
(
"Input_scale"
));
}
...
...
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