Add QNN EP option context_node_name_prefix to set EPContext node name prefix

include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h

@@ -265,6 +265,10 @@ static const char* const kOrtSessionOptionEpContextFilePath = "ep.context_file_p
 // "1": dump the EP context into the Onnx model. (default).
 static const char* const kOrtSessionOptionEpContextEmbedMode = "ep.context_embed_mode";
 
+// Specify the EPContext node name prefix to make it unique
+// in case user need to merge/connect multiple EPContext nodes in one model
+static const char* const kOrtSessionOptionEpContextNodeNamePrefix = "ep.context_node_name_prefix";
+
 // Gemm fastmath mode provides fp32 gemm acceleration with bfloat16 based matmul.
 // Option values:
 // - "0": Gemm FastMath mode is not enabled. [DEFAULT]

onnxruntime/core/providers/qnn/qnn_execution_provider.cc

@@ -199,6 +199,13 @@
 
     context_cache_path_cfg_ = session_options->config_options.GetConfigOrDefault(kOrtSessionOptionEpContextFilePath, "");
     LOGS_DEFAULT(VERBOSE) << "User specified context cache path: " << context_cache_path_cfg_;
+
+    // For the case that workaround QNN context PD memory limit, user need split the model into pieces and
+    // generate the QNN context model separately.
+    // It could happen that the generated EPContext node in separate graph has same node name.
+    // User can set this context_node_name_prefix for each split pieces to avoid that happens.
+    context_node_name_prefix_ = session_options->config_options.GetConfigOrDefault(kOrtSessionOptionEpContextNodeNamePrefix, "");
+    LOGS_DEFAULT(VERBOSE) << "User specified QNN context node name prefix: " << context_node_name_prefix_;
   }
 
   static const std::string BACKEND_PATH = "backend_path";
@@ -613,7 +620,7 @@
   const auto gen_metadef_name = [&]() {
     uint64_t model_hash;
     int metadef_id = metadef_id_generator_.GenerateId(graph_viewer, model_hash);
-    return MakeString(QNN, "_", model_hash, "_", metadef_id);
+    return MakeString(QNN, context_node_name_prefix_, "_", model_hash, "_", metadef_id);
   };
 
   // For model with EPContext, make sure each partition only has one single EPContext node

onnxruntime/core/providers/qnn/qnn_execution_provider.h

@@ -80,6 +80,7 @@ class QNNExecutionProvider : public IExecutionProvider {
   std::unordered_map<std::string, std::unique_ptr<qnn::QnnModel>> qnn_models_;
   bool context_cache_enabled_ = false;
   std::string context_cache_path_cfg_ = "";
+  std::string context_node_name_prefix_ = "";
   bool disable_cpu_ep_fallback_ = false;  // True if CPU EP fallback has been disabled for this session.
   bool qnn_context_embed_mode_ = true;
   int32_t vtcm_size_in_mb_ = 0;

onnxruntime/test/perftest/command_args_parser.cc

@@ -97,6 +97,7 @@ namespace perftest {
       "\t    [QNN only] [device_id]: The ID of the device to use when setting 'htp_arch'. Defaults to '0' (for single device). \n"
       "\t    [QNN only] [enable_htp_fp16_precision]: Enable the HTP_FP16 precision so that the float32 model will be inferenced with fp16 precision. \n"
       "\t    Otherwise, it will be fp32 precision. Only works for float32 model. Defaults to '0' (with FP32 precision.). \n"
+      "\t    [QNN only] [context_node_name_prefix]: QNN EPContext node name prefix to make sure the generated node name unique for split models. e.g 'phi3_part1'.\n"
       "\t    [Example] [For QNN EP] -e qnn -i \"backend_path|/folderpath/libQnnCpu.so\" \n"
       "\n"
       "\t    [TensorRT only] [trt_max_partition_iterations]: Maximum iterations for TensorRT parser to get capability.\n"

onnxruntime/test/perftest/ort_test_session.cc

@@ -269,7 +269,7 @@ OnnxRuntimeTestSession::OnnxRuntimeTestSession(Ort::Env& env, std::random_device
           std::string str = str_stream.str();
           ORT_THROW("Supported htp_performance_mode: " + str);
         }
-      } else if (key == "qnn_saver_path") {
+      } else if (key == "qnn_saver_path" || key == "context_node_name_prefix") {
         // no validation
       } else if (key == "htp_graph_finalization_optimization_mode") {
         std::unordered_set<std::string> supported_htp_graph_final_opt_modes = {"0", "1", "2", "3"};
@@ -307,7 +307,7 @@ OnnxRuntimeTestSession::OnnxRuntimeTestSession(Ort::Env& env, std::random_device
         ORT_THROW(R"(Wrong key type entered. Choose from options: ['backend_path',
 'profiling_level', 'profiling_file_path', 'rpc_control_latency', 'vtcm_mb', 'htp_performance_mode',
 'qnn_saver_path', 'htp_graph_finalization_optimization_mode', 'qnn_context_priority', 'soc_model',
-'htp_arch', 'device_id', 'enable_htp_fp16_precision'])");
+'htp_arch', 'device_id', 'enable_htp_fp16_precision', 'context_node_name_prefix'])");
       }
 
       qnn_options[key] = value;

onnxruntime/test/providers/qnn/qnn_ep_context_test.cc

@@ -279,6 +279,45 @@ TEST_F(QnnHTPBackendTests, QnnContextGeneration2InputsOrderIssue) {
   ASSERT_EQ(std::remove(context_binary_file.c_str()), 0);
 }
 
+TEST_F(QnnHTPBackendTests, QnnContextGenerationNodeNamePrefix) {
+  ProviderOptions provider_options;
+#if defined(_WIN32)
+  provider_options["backend_path"] = "QnnHtp.dll";
+#else
+  provider_options["backend_path"] = "libQnnHtp.so";
+#endif
+  std::string node_name_prefix = "node_name_prefix_test";
+
+  // Add kMSDomain to cover contrib op like Gelu
+  const std::unordered_map<std::string, int> domain_to_version = {{"", 13}, {kMSDomain, 1}};
+
+  auto& logging_manager = DefaultLoggingManager();
+  logging_manager.SetDefaultLoggerSeverity(logging::Severity::kERROR);
+
+  const std::string context_binary_file = "./qnn_ctx_2_inputs_order_test_gen.onnx";
+  Ort::SessionOptions so;
+  so.AddConfigEntry(kOrtSessionOptionEpContextEnable, "1");
+  so.AddConfigEntry(kOrtSessionOptionEpContextFilePath, context_binary_file.c_str());
+  so.AddConfigEntry(kOrtSessionOptionEpContextNodeNamePrefix, node_name_prefix.c_str());
+  so.AppendExecutionProvider("QNN", provider_options);
+
+  Ort::Session session(*ort_env, ORT_TSTR("testdata/qnn_ctx_2_inputs_order_test.onnx"), so);
+
+  // Make sure the Qnn context cache binary file is generated
+  EXPECT_TRUE(std::filesystem::exists(context_binary_file.c_str()));
+
+  std::shared_ptr<Model> model;
+  ASSERT_STATUS_OK(Model::Load(ToPathString(context_binary_file), model, nullptr, DefaultLoggingManager().DefaultLogger()));
+  for (auto& node : model->MainGraph().Nodes()) {
+    if (node.OpType() == "EPContext") {
+      EXPECT_TRUE(node.Name().find(node_name_prefix) != std::string::npos);
+    }
+  }
+
+  // clean up
+  ASSERT_EQ(std::remove(context_binary_file.c_str()), 0);
+}
+
 // Run QDQ model on HTP 3 times
 // 1st run will generate the Qnn context cache onnx file
 // 2nd run directly loads and run from Qnn context cache model