[WebNN EP] Support MultiHeadAttention(MHA)

onnxruntime/core/providers/webnn/builders/helper.h

@@ -204,6 +204,7 @@ const std::map<std::string_view, std::vector<std::string_view>> decomposed_op_ma
       "softmax", "transpose", "where"}},
     {"LRN", {"add", "averagePool2d", "div", "mul", "pad", "pow", "transpose"}},
     {"MatMulNBits", {"add", "dequantizeLinear", "matmul", "reshape", "transpose"}},
+    {"MultiHeadAttention", {"add", "cast", "concat", "constant", "div", "matmul", "reshape", "softmax", "transpose"}},
     {"RotaryEmbedding", {"add", "concat", "gather", "mul", "reshape", "split"}},
     {"SimplifiedLayerNormalization", {"add", "div", "mul", "pow", "reduceMean", "sqrt"}},
     {"SkipSimplifiedLayerNormalization", {"add", "div", "mul", "pow", "reduceMean", "sqrt"}},

onnxruntime/core/providers/webnn/builders/impl/attention_helper.h

@@ -2,6 +2,8 @@
 // Copyright (c) Intel Corporation. All rights reserved.
 // Licensed under the MIT License.
 
+#pragma once
+
 namespace onnxruntime {
 namespace webnn {
 /*
@@ -26,10 +28,11 @@ namespace webnn {
                                  |
                                output
 */
-emscripten::val ScaledDotProductAttention(ModelBuilder& model_builder, const Node& node, const logging::Logger& logger,
-                                          emscripten::val query, emscripten::val key, emscripten::val value,
-                                          emscripten::val scale, emscripten::val attn_mask,
-                                          std::vector<uint32_t> reshape_output_shape) {
+inline emscripten::val ScaledDotProductAttention(ModelBuilder& model_builder, const Node& node,
+                                                 const logging::Logger& logger, emscripten::val query,
+                                                 emscripten::val key, emscripten::val value, emscripten::val scale,
+                                                 emscripten::val attn_mask,
+                                                 std::vector<uint32_t> reshape_output_shape) {
   emscripten::val common_options = emscripten::val::object();
   // B,H,S,N * B,H,kv_S,N = B,H,S,kv_S
   common_options.set("label", node.Name() + "_/Attention/qkv/matmul_1");

onnxruntime/core/providers/webnn/builders/impl/gqa_op_builder.cc

@@ -6,7 +6,7 @@
 #include "core/providers/webnn/builders/helper.h"
 #include "core/providers/webnn/builders/model_builder.h"
 #include "core/providers/webnn/builders/op_builder_factory.h"
-#include "cmath"
+#include <cmath>
 
 #include "base_op_builder.h"
 #include "attention_helper.h"