Relax Inputs/Ouputs validation

* add RunOptions for qnn htp batch multiplier
* avoid validating inputs/outpus if this option is used

Author: chunghow-qti

include/onnxruntime/core/session/onnxruntime_run_options_config_keys.h

@@ -46,6 +46,9 @@ static const char* const kOrtRunOptionsConfigQnnRpcControlLatency = "qnn.rpc_con
 // Set QNN Lora Config File for apply Lora in QNN context binary
 static const char* const kOrtRunOptionsConfigQnnLoraConfig = "qnn.lora_config";
 
+// Set QNN enable batch mutiplier for htp backend
+static const char* const kOrtRunOptionsConfigQnnBatchMultiplier = "qnn.enable_htp_batch_multiplier";
+
 // Set graph annotation id for CUDA EP. Use with enable_cuda_graph=true.
 // The value should be an integer. If the value is not set, the default value is 0 and
 // ORT session only captures one cuda graph before another capture is requested.

onnxruntime/core/session/inference_session.cc

@@ -2606,29 +2606,10 @@ common::Status InferenceSession::CheckShapes(const std::string& input_output_nam
                            " Please fix either the inputs/outputs or the model.");
   }
 
-#ifdef USE_QNN
-  // Helper function to check whether QNN EP is used & all nodes are assigned to QNN EP,
-  // and relax the constraint to support batch multiplier on the first dimension.
-  // We will check whether only the Htp backend is used inside QnnModel::ExecuteGraph.
-  auto is_valid_qnn_batch_multiplier = [this](int64_t input_dim, int64_t expected_dim, const Graph& graph) -> bool {
-    if (!AreAllNodesInMainGraphAssignedToOneEp(graph, kQnnExecutionProvider)) {
-      LOGS_IF(input_dim != expected_dim, *session_logger_, WARNING) << "input batch size and expected batch size are different. "
-                                                                    << "Batch multiplier is only supported on the QNN EP, "
-                                                                    << "but some nodes in the graph are assigned to other EPs";
-      return false;
-    }
-    return expected_dim > 0 && input_dim % expected_dim == 0;
-  };
-#endif
-
   InlinedVector<size_t> invalid_dim_indices;
   for (size_t i = 0; i < shape_size; ++i) {
     if (expected_shape[i] < 0) {
       continue;  // this represents a symbolic shape dimension
-#ifdef USE_QNN
-    } else if (i == 0 && is_valid_qnn_batch_multiplier(input_output_shape[i], expected_shape[i], model_->MainGraph())) {
-      continue;  // Qnn API supports batch multiplier, but the running batch size must be divisible by the original batch size.
-#endif
     } else if (input_output_shape[i] != expected_shape[i]) {
       invalid_dim_indices.push_back(i);
     }
@@ -2999,10 +2980,17 @@ Status InferenceSession::Run(const RunOptions& run_options,
 
       // log evaluation start to trace logging provider
       env.GetTelemetryProvider().LogEvaluationStart(session_id_);
-
+#ifdef USE_QNN
+      const std::string& batch_multiplier = run_options.config_options.GetConfigOrDefault(kOrtRunOptionsConfigQnnBatchMultiplier, "");
+      if (batch_multiplier.empty()) {
+        LOGS(*session_logger_, INFO) << "Enable QNN HTP batch mutliplier. Don't validate the inputs/outptus";
+        ORT_RETURN_IF_ERROR_SESSIONID_(ValidateInputs(feed_names, feeds));
+        ORT_RETURN_IF_ERROR_SESSIONID_(ValidateOutputs(output_names, p_fetches));
+      }
+#else
       ORT_RETURN_IF_ERROR_SESSIONID_(ValidateInputs(feed_names, feeds));
       ORT_RETURN_IF_ERROR_SESSIONID_(ValidateOutputs(output_names, p_fetches));
-
+#endif
       // shrink certain default memory arenas if the user has requested for it
       const std::string& shrink_memory_arenas =
           run_options.config_options.GetConfigOrDefault(kOrtRunOptionsConfigEnableMemoryArenaShrinkage, "");