move the option from provider option to session option

microsoft · Jul 25, 2024 · 1896559 · 1896559
1 parent 3f9d550
commit 1896559
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 12 deletions.
diff --git a/include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h b/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]

diff --git a/onnxruntime/core/providers/qnn/qnn_execution_provider.cc b/onnxruntime/core/providers/qnn/qnn_execution_provider.cc
@@ -199,6 +199,13 @@ QNNExecutionProvider::QNNExecutionProvider(const ProviderOptions& provider_optio
 
     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";
@@ -378,17 +385,6 @@ QNNExecutionProvider::QNNExecutionProvider(const ProviderOptions& provider_optio
     LOGS_DEFAULT(VERBOSE) << "User specified enable_htp_fp16_precision: " << enable_HTP_FP16_precision_;
   }
 
-  // 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.
-  static const std::string QNN_CONTEXT_NODE_NAME_PREFIX = "context_node_name_prefix";
-  auto context_node_name_prefix_pos = provider_options_map.find(QNN_CONTEXT_NODE_NAME_PREFIX);
-  if (context_node_name_prefix_pos != provider_options_map.end()) {
-    context_node_name_prefix_ = context_node_name_prefix_pos->second;
-    LOGS_DEFAULT(VERBOSE) << "User specified QNN context node name prefix: " << context_node_name_prefix_;
-  }
-
   qnn_backend_manager_ = std::make_unique<qnn::QnnBackendManager>(
       std::move(backend_path),
       profiling_level_etw,

diff --git a/onnxruntime/test/providers/qnn/qnn_ep_context_test.cc b/onnxruntime/test/providers/qnn/qnn_ep_context_test.cc
@@ -287,7 +287,6 @@ TEST_F(QnnHTPBackendTests, QnnContextGenerationNodeNamePrefix) {
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
   std::string node_name_prefix = "node_name_prefix_test";
-  provider_options["context_node_name_prefix"] = node_name_prefix;
 
   // Add kMSDomain to cover contrib op like Gelu
   const std::unordered_map<std::string, int> domain_to_version = {{"", 13}, {kMSDomain, 1}};
@@ -299,6 +298,7 @@ TEST_F(QnnHTPBackendTests, QnnContextGenerationNodeNamePrefix) {
   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);