[QNN EP] Add support for Attention Op

docs/execution_providers/QNN-ExecutionProvider.md

@@ -509,6 +509,7 @@ ort.unregister_execution_provider_library(ep_registration_name)
 |ai.onnx:ArgMin||
 |ai.onnx:Asin||
 |ai.onnx:Atan||
+|ai.onnx:Attention|Opsets 23–24. 3D \[B,S,n·hs\] and 4D \[B,n,S,hs\] inputs. HTP: full decomposition for MHA, GQA/MQA, is\_causal, attn\_mask, softcap, KV cache, qk\_matmul\_output modes 0–3. GPU: native QNN\_OP\_GROUP\_QUERY\_ATTENTION when is\_causal=1 and no softcap/attn\_mask/qk\_output (covers MHA and GQA/MQA); decomposition otherwise. Static shapes only.|
 |ai.onnx:AveragePool||
 |ai.onnx:Bernoulli|CPU and HTP backends supported.|
 |ai.onnx:BatchNormalization|fp16 supported since 1.18.0|

onnxruntime/core/providers/qnn/builder/op_builder_factory.cc

@@ -13,6 +13,7 @@ namespace qnn {
 OpBuilderRegistrations::OpBuilderRegistrations() {
   CreateArgMaxMinOpBuilder("ArgMax", *this);
   CreateArgMaxMinOpBuilder("ArgMin", *this);
+  CreateAttentionOpBuilder("Attention", *this);
   CreateBatchNormalizationOpBuilder("BatchNormalization", *this);
   CreateBernoulliOpBuilder("Bernoulli", *this);
   CreateCastOpBuilder("Cast", *this);

onnxruntime/core/providers/qnn/builder/op_builder_factory.h

@@ -58,6 +58,7 @@ class OpBuilderRegistrations {
 const IOpBuilder* GetOpBuilder(const std::string& onnx_op_type);
 
 void CreateArgMaxMinOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);
+void CreateAttentionOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);
 void CreateBatchNormalizationOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);
 void CreateBernoulliOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);
 void CreateCastOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);