-
Notifications
You must be signed in to change notification settings - Fork 1.6k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* [Cherry-Pick][XPU] fixed inplace op mem reuse issue when the previous op is an invalid op (#9562) (#9564) * [XPU] support roformer relative embedding (#9536) * fix sampling_id, fix xpu python whl, fix quant_dequant pass (#9636) * [XPU] support ffn intermediate size M!=4 (#9646) * [xpu] fix scope new tensor, max weight is unchanged (#9641) * [XPU] Fixed the bug in op calib. (#9700) * [XPU] support skip ffn quant in K200 (#9704)
- Loading branch information
Showing
16 changed files
with
673 additions
and
56 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
208 changes: 174 additions & 34 deletions
208
lite/core/optimizer/mir/fusion/__xpu__multi_encoder_fuse_pass.cc
Large diffs are not rendered by default.
Oops, something went wrong.
195 changes: 195 additions & 0 deletions
195
lite/core/optimizer/mir/fusion/__xpu__roformer_relative_pos_fuse_pass.cc
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,195 @@ | ||
| // Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. | ||
| // | ||
| // Licensed under the Apache License, Version 2.0 (the "License"); | ||
| // you may not use this file except in compliance with the License. | ||
| // You may obtain a copy of the License at | ||
| // | ||
| // http://www.apache.org/licenses/LICENSE-2.0 | ||
| // | ||
| // Unless required by applicable law or agreed to in writing, software | ||
| // distributed under the License is distributed on an "AS IS" BASIS, | ||
| // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| // See the License for the specific language governing permissions and | ||
| // limitations under the License. | ||
|
|
||
| #include <memory> | ||
| #include <string> | ||
| #include "lite/backends/xpu/math.h" | ||
| #include "lite/core/optimizer/mir/pass_registry.h" | ||
| #include "lite/core/optimizer/mir/pattern_matcher_high_api.h" | ||
|
|
||
| namespace paddle { | ||
| namespace lite { | ||
| namespace mir { | ||
| namespace fusion { | ||
|
|
||
| /* support xpu roformer relative pos */ | ||
| /* in_Input --------------- */ | ||
| /* | \ | */ | ||
| /* | \ | */ | ||
| /* split shape | */ | ||
| /* / | \ | */ | ||
| /* / | \ | */ | ||
| /* | scale slice | */ | ||
| /* \ | / \ | */ | ||
| /* \ | / \ | */ | ||
| /* concat slice slice | */ | ||
| /* | / \ | */ | ||
| /* | / \ | */ | ||
| /* elementwise_mul elementwise_mul */ | ||
| /* | / */ | ||
| /* | / */ | ||
| /* elementwise_add */ | ||
| /* | */ | ||
| /* | */ | ||
| /* out_Output */ | ||
| /*-------------------------------------------*/ | ||
| /* After the pass apply: */ | ||
| /* in_Input */ | ||
| /* cos_emb | sin_emb */ | ||
| /* \ | / */ | ||
| /* xpu_roformer_relative */ | ||
| /* | */ | ||
| /* | */ | ||
| /* out_Output */ | ||
| /*-------------------------------------------*/ | ||
|
|
||
| class XPURoformerRelativePosFuser : public FuseBase { | ||
| public: | ||
| void BuildPattern() override { | ||
| auto* input = VarNode("input") | ||
| ->assert_is_op_input("split", "X") | ||
| ->assert_is_op_input("elementwise_mul", "X") | ||
| ->assert_is_op_input("shape", "Input") | ||
| ->AsInput(); | ||
| auto* split = | ||
| OpNode("split", "split") | ||
| ->assert_op_attr<int32_t>("axis", 3) | ||
| ->assert_op_attr<int32_t>("num", 2) // do we really need it | ||
| ->AsIntermediate(); | ||
| auto* split_out0 = VarNode("split_out0") | ||
| ->assert_is_op_nth_input("concat", "X", 1) | ||
| ->assert_is_op_nth_output("split", "Out", 0) | ||
| ->AsIntermediate(); | ||
| auto* split_out1 = VarNode("split_out1") | ||
| ->assert_is_op_input("scale", "X") | ||
| ->assert_is_op_nth_output("split", "Out", 1) | ||
| ->AsIntermediate(); | ||
| auto* scale = | ||
| OpNode("scale", "scale") | ||
| ->assert_op_attr_satisfied<float>( | ||
| "scale", | ||
| [](float attr) { return (std::fabs(attr + 1.0) < 1e-5); }) | ||
| ->AsIntermediate(); | ||
| auto* scale_out = VarNode("scale_out") | ||
| ->assert_is_op_input("concat", "X") | ||
| ->assert_is_op_output("scale", "Out") | ||
| ->AsIntermediate(); | ||
| auto* concat = OpNode("concat", "concat")->AsIntermediate(); | ||
| auto* concat_out = VarNode("concat_out") | ||
| ->assert_is_op_input("elementwise_mul", "X") | ||
| ->assert_is_op_output("concat", "Out") | ||
| ->AsIntermediate(); | ||
| auto* shape = OpNode("shape", "shape")->AsIntermediate(); | ||
| auto* shape_out = VarNode("shape_out") | ||
| ->assert_is_op_input("slice", "Input") | ||
| ->assert_is_op_output("shape", "Out") | ||
| ->AsIntermediate(); | ||
| auto* slice1 = OpNode("slice1", "slice")->AsIntermediate(); | ||
| auto* slice1_out = VarNode("slice1_out") | ||
| ->assert_is_op_input("slice", "EndsTensorList") | ||
| ->assert_is_op_output("slice", "Out") | ||
| ->AsIntermediate(); | ||
| auto* sin_emb = | ||
| VarNode("sin_emb")->assert_is_op_input("slice", "Input")->AsInput(); | ||
| auto* cos_emb = | ||
| VarNode("cos_emb")->assert_is_op_input("slice", "Input")->AsInput(); | ||
| auto* slice_sin = OpNode("slice_sin", "slice")->AsIntermediate(); | ||
| auto* slice_sin_out = VarNode("slice_sin_out") | ||
| ->assert_is_op_input("elementwise_mul", "Y") | ||
| ->assert_is_op_output("slice", "Out") | ||
| ->AsIntermediate(); | ||
| auto* ew_mul_sin = | ||
| OpNode("ew_mul_sin", "elementwise_mul")->AsIntermediate(); | ||
| auto* ew_mul_sin_out = VarNode("ew_mul_sin_out") | ||
| ->assert_is_op_input("elementwise_add", "Y") | ||
| ->assert_is_op_output("elementwise_mul", "Out") | ||
| ->AsIntermediate(); | ||
| auto* ew_add = OpNode("ew_add", "elementwise_add")->AsIntermediate(); | ||
| auto* ew_add_out = VarNode("ew_add_out") | ||
| ->assert_is_op_output("elementwise_add", "Out") | ||
| ->AsOutput(); | ||
| auto* slice_cos = OpNode("slice_cos", "slice")->AsIntermediate(); | ||
| auto* slice_cos_out = VarNode("slice_cos_out") | ||
| ->assert_is_op_input("elementwise_mul", "Y") | ||
| ->assert_is_op_output("slice", "Out") | ||
| ->AsIntermediate(); | ||
| auto* ew_mul_cos = | ||
| OpNode("ew_mul_cos", "elementwise_mul")->AsIntermediate(); | ||
| auto* ew_mul_cos_out = VarNode("ew_mul_cos_out") | ||
| ->assert_is_op_input("elementwise_add", "X") | ||
| ->assert_is_op_output("elementwise_mul", "Out") | ||
| ->AsIntermediate(); | ||
| *input >> *split >> *split_out1 >> *scale >> *scale_out >> *concat >> | ||
| *concat_out >> *ew_mul_sin >> *ew_mul_sin_out >> *ew_add >> *ew_add_out; | ||
| *input >> *ew_mul_cos >> *ew_mul_cos_out >> *ew_add; | ||
| *input >> *shape >> *shape_out >> *slice1 >> *slice1_out >> *slice_sin >> | ||
| *slice_sin_out >> *ew_mul_sin; | ||
| *slice1_out >> *slice_cos >> *slice_cos_out >> *ew_mul_cos; | ||
| *sin_emb >> *slice_sin; | ||
| *cos_emb >> *slice_cos; | ||
| *split >> *split_out0 >> *concat; | ||
| } | ||
|
|
||
| void InsertNewNode(SSAGraph* graph, const key2nodes_t& matched) override { | ||
| cpp::OpDesc op_desc; | ||
| op_desc.SetType("__xpu__roformer_relative_embedding"); | ||
| // use "X", be consistent with target_op_type_ in multiencoder pass | ||
| op_desc.SetInput("X", {matched.at("input")->arg()->name}); | ||
| op_desc.SetInput("CosEmbbeding", {matched.at("cos_emb")->arg()->name}); | ||
| op_desc.SetInput("SinEmbbeding", {matched.at("sin_emb")->arg()->name}); | ||
| op_desc.SetOutput("Out", {matched.at("ew_add_out")->arg()->name}); | ||
| auto* scope = matched.at("split")->stmt()->op()->scope(); | ||
|
|
||
| auto cos_emb_name = matched.at("cos_emb")->arg()->name; | ||
| auto cos_emb_shape = scope->FindMutableTensor(cos_emb_name)->dims(); | ||
| auto sin_emb_name = matched.at("sin_emb")->arg()->name; | ||
| auto sin_emb_shape = scope->FindMutableTensor(sin_emb_name)->dims(); | ||
| CHECK_EQ(cos_emb_shape.size(), 4) << cos_emb_shape.size(); | ||
| CHECK_GT(cos_emb_shape[2], 0) << cos_emb_shape[2]; | ||
| CHECK_EQ(sin_emb_shape.size(), 4) << sin_emb_shape.size(); | ||
| for (int i = 0; i < sin_emb_shape.size(); ++i) { | ||
| CHECK_EQ(sin_emb_shape[i], cos_emb_shape[i]) | ||
| << i << " th dim: " << sin_emb_shape[i] << ", " << cos_emb_shape[i]; | ||
| } | ||
| op_desc.SetAttr<int>("max_pos_len", cos_emb_shape[2]); | ||
|
|
||
| auto& valid_places = matched.at("split")->stmt()->op()->valid_places(); | ||
| auto new_op = LiteOpRegistry::Global().Create(op_desc.Type()); | ||
| new_op->Attach(op_desc, scope); | ||
| auto* new_op_node = graph->GraphCreateInstructNode(new_op, valid_places); | ||
|
|
||
| DirectedLink(matched.at("input"), new_op_node); | ||
| DirectedLink(matched.at("cos_emb"), new_op_node); | ||
| DirectedLink(matched.at("sin_emb"), new_op_node); | ||
| DirectedLink(new_op_node, matched.at("ew_add_out")); | ||
| } | ||
| }; | ||
|
|
||
| } // namespace fusion | ||
|
|
||
| class XPURoformerRelativePosFusePass : public ProgramPass { | ||
| public: | ||
| void Apply(const std::unique_ptr<SSAGraph>& graph) override { | ||
| fusion::XPURoformerRelativePosFuser fuser; | ||
| fuser(graph.get()); | ||
| } | ||
| }; | ||
|
|
||
| } // namespace mir | ||
| } // namespace lite | ||
| } // namespace paddle | ||
|
|
||
| REGISTER_MIR_PASS(__xpu__roformer_relative_pos_fuse_pass, | ||
| paddle::lite::mir::XPURoformerRelativePosFusePass) | ||
| .BindTargets({TARGET(kXPU)}); |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
73 changes: 73 additions & 0 deletions
73
lite/kernels/xpu/__xpu__roformer_relative_embedding_compute.cc
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,73 @@ | ||
| // Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. | ||
| // | ||
| // Licensed under the Apache License, Version 2.0 (the "License"); | ||
| // you may not use this file except in compliance with the License. | ||
| // You may obtain a copy of the License at | ||
| // | ||
| // http://www.apache.org/licenses/LICENSE-2.0 | ||
| // | ||
| // Unless required by applicable law or agreed to in writing, software | ||
| // distributed under the License is distributed on an "AS IS" BASIS, | ||
| // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| // See the License for the specific language governing permissions and | ||
| // limitations under the License. | ||
|
|
||
| #include "lite/kernels/xpu/__xpu__roformer_relative_embedding_compute.h" | ||
| #include <vector> | ||
| #include "lite/backends/xpu/xpu_header_sitter.h" | ||
| #include "lite/core/op_registry.h" | ||
|
|
||
| namespace paddle { | ||
| namespace lite { | ||
| namespace kernels { | ||
| namespace xpu { | ||
|
|
||
| void RoformerRelativeEmbeddingCompute::Run() { | ||
| auto& param = this->template Param<param_t>(); | ||
| auto& ctx = this->ctx_->template As<XPUContext>(); | ||
| auto input_dim = param.input->dims(); | ||
| CHECK_EQ(input_dim.size(), 4); | ||
| int batch = input_dim[0]; | ||
| int head_num = param.input->dims()[1]; | ||
| int seqlen = param.input->dims()[2]; | ||
| int head_dim = param.input->dims()[3]; | ||
| CHECK_LE(seqlen, param.max_pos_len); | ||
| std::vector<int> lod; | ||
| lod.resize(batch + 1); | ||
| for (int i = 0; i < batch + 1; i++) { | ||
| lod[i] = i * seqlen; | ||
| } | ||
| int r = | ||
| xdnn::rope<float>(ctx.GetRawContext(), | ||
| param.input->data<float>(), | ||
| param.output->mutable_data<float>(TARGET(kXPU)), | ||
| param.cos_embedding->data<float>(), | ||
| param.sin_embedding->data<float>(), | ||
| batch, | ||
| head_num, | ||
| head_dim, | ||
| head_num * head_dim, | ||
| lod, | ||
| param.max_pos_len, | ||
| false, // no vsl | ||
| true); // transpose to [n, seql, head_num, head_dim] | ||
| CHECK_EQ(r, 0) << "call RoformerRelativeEmbeddingCompute failed"; | ||
| } | ||
|
|
||
| } // namespace xpu | ||
| } // namespace kernels | ||
| } // namespace lite | ||
| } // namespace paddle | ||
|
|
||
| REGISTER_LITE_KERNEL( | ||
| __xpu__roformer_relative_embedding, | ||
| kXPU, | ||
| kFloat, | ||
| kNCHW, | ||
| paddle::lite::kernels::xpu::RoformerRelativeEmbeddingCompute, | ||
| def) | ||
| .BindInput("X", {LiteType::GetTensorTy(TARGET(kXPU))}) | ||
| .BindInput("CosEmbbeding", {LiteType::GetTensorTy(TARGET(kXPU))}) | ||
| .BindInput("SinEmbbeding", {LiteType::GetTensorTy(TARGET(kXPU))}) | ||
| .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kXPU))}) | ||
| .Finalize(); |
Oops, something went wrong.