[QNN EP] Fix inconsistent inputs for graph

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

@@ -331,7 +331,15 @@ Status QnnModel::SetupTensors(std::vector<QnnTensorInfo>& qnn_tensor_infos,
                               bool is_input) {
   size_t tensor_count = tensor_wrappers.size();
   ORT_RETURN_IF(0 == tensor_count, "Zero tensor size!");
-  qnn_tensor_infos.resize(tensor_count);
+  if (is_input) {
+    // Resize qnn_tensor_infos according to the number of graph inputs.
+    auto input_count = GetGraphInputCount();
+    ORT_RETURN_IF(input_count < tensor_count,
+                  "The count of graph inputs should be at least the count of tensor_wrapper!");
+    qnn_tensor_infos.resize(input_count);
+  } else {
+    qnn_tensor_infos.resize(tensor_count);
+  }
 
   for (auto& tensor_wrapper : tensor_wrappers) {
     ORT_RETURN_IF(utils::QnnTensorHasDynamicShape(tensor_wrapper.GetQnnTensor()),
@@ -348,6 +356,18 @@ Status QnnModel::SetupTensors(std::vector<QnnTensorInfo>& qnn_tensor_infos,
     qnn_tensor_info.tensor_byte_size = static_cast<uint32_t>(length);
     qnn_tensor_info.ort_index = ort_index;
   }
+  // The number of graph inputs and the number of tensor wrappers may not match.
+  // - For example, for ResizeNearestNeighbor op, Qnn only cares about the 1st input,
+  //   so the rest of the inputs are not converted to tensor wrappers.
+  // - However, these remaining inputs still appear in the graph inputs, resulting in
+  //   a discrepancy in the input quantities.
+  // If not all inputs are used, erase the empty allocations in qnn_tensor_infos.
+  if (is_input) {
+    qnn_tensor_infos.erase(std::remove_if(qnn_tensor_infos.begin(),
+                                          qnn_tensor_infos.end(),
+                                          [](QnnTensorInfo qnn_tensor_info) { return qnn_tensor_info.tensor_wrapper == nullptr; }),
+                           qnn_tensor_infos.end());
+  }
   return Status::OK();
 }
 

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

@@ -77,6 +77,11 @@ class QnnModel {
     return it->second;
   }
 
+  // Return the number of graph inputs
+  size_t GetGraphInputCount() const {
+    return model_input_index_map_.size();
+  }
+
   size_t GetOutputIndex(const std::string& name) const {
     return GetInputOutputIndex(name, outputs_info_);
   }

onnxruntime/test/providers/qnn/qnn_basic_test.cc

@@ -6,6 +6,7 @@
 #include <thread>
 
 #include "core/graph/constants.h"
+#include "core/graph/node_attr_utils.h"
 #include "core/providers/cpu/cpu_provider_factory.h"  // For OrtSessionOptionsAppendExecutionProvider_CPU
 #if BUILD_QNN_EP_STATIC_LIB
 #include "core/providers/qnn/qnn_allocator.h"  // Used by QnnHTPBackendTests.UseHtpSharedMemoryAllocatorForInputs
@@ -1286,6 +1287,34 @@ TEST_F(QnnHTPBackendTests, AutoEp_PreferNpu) {
 }
 #endif  // defined(WIN32) && !BUILD_QNN_EP_STATIC_LIB
 
+// Test whether QNN EP can handle the case where the number of graph inputs and
+// the number of tensor wrappers do not match.
+// Take Resize op as an example.
+// - Qnn only cares about the 1st input, so the rest of the inputs are not converted
+//   to tensor wrappers.
+// - However, these remaining inputs still appear in the graph inputs,
+//   resulting in a discrepancy in the input quantities.
+TEST_F(QnnHTPBackendTests, TestMismatchedGraphInputAndTensorWrapperCount) {
+  onnxruntime::ProviderOptions provider_options;
+  provider_options["backend_type"] = "htp";
+
+  auto input_defs = {TestInputDef<float>({1, 3, 10, 10}, false, -10.0f, 10.0f),
+                     TestInputDef<float>({0}, false, {}),
+                     TestInputDef<float>({4}, true, {1.0f, 1.0f, 2.0f, 2.0f})};
+  auto attrs = {utils::MakeAttribute("mode", "nearest"),
+                utils::MakeAttribute("coordinate_transformation_mode", "asymmetric"),
+                utils::MakeAttribute("nearest_mode", "floor")};
+  RunQnnModelTest(BuildOpTestCase<float>("Resize",
+                                         input_defs,
+                                         {},
+                                         attrs,
+                                         kOnnxDomain),
+                  provider_options,
+                  11,
+                  ExpectedEPNodeAssignment::All,
+                  0.008f);
+}
+
 #endif  // defined(__aarch64__) || defined(_M_ARM64) || defined(__linux__)
 #endif  // !defined(ORT_MINIMAL_BUILD)