Fix: Skip CPU provider in SetProviderSessionOptions

src/config.cpp

@@ -18,7 +18,9 @@ namespace Generators {
 std::string_view NormalizeProviderName(std::string_view name) {
   std::string lower_name(name);
   std::transform(lower_name.begin(), lower_name.end(), lower_name.begin(), [](unsigned char c) { return static_cast<unsigned char>(std::tolower(c)); });
-  if (lower_name == "qnn") {
+  if (lower_name == "cpu" || lower_name == "cpuexecutionprovider") {
+    return "CPU";
+  } else if (lower_name == "qnn") {
     return "QNN";
   } else if (lower_name == "webgpu") {
     return "WebGPU";

src/models/session_options.cpp

@@ -3,6 +3,7 @@
 
 #include "session_options.h"
 
+#include <algorithm>
 #include <functional>
 #include <unordered_map>
 
@@ -145,9 +146,25 @@ DeviceInterface* SetProviderSessionOptions(OrtSessionOptions& session_options,
                                                          const Config&,
                                                          bool);
 
+  // CPU EP is always implicitly registered — no action needed.
+  // Throws if users attempt to set provider options (which would be silently lost).
+  static auto CPUAppendExecutionProvider = [](OrtSessionOptions&,
+                                              const Config::ProviderOptions& provider_options,
+                                              const Config&,
+                                              bool) -> DeviceInterface* {
+    if (!provider_options.options.empty()) {
+      throw std::runtime_error(
+          "CPU execution provider does not support provider options. "
+          "CPU is always available as the default fallback and does not need to be explicitly registered. "
+          "Remove the CPU provider entry and its options from your configuration.");
+    }
+    return nullptr;
+  };
+
   // Dispatch table: maps provider name (as it appears in genai_config.json) to
   // the corresponding provider-specific AppendExecutionProvider function.
   static const std::unordered_map<std::string, AppendExecutionProviderFn> append_execution_provider{
+      {"CPU", CPUAppendExecutionProvider},
       {"cuda", CUDAExecutionProvider::AppendExecutionProvider},
       {"DML", DMLExecutionProvider::AppendExecutionProvider},
       {"NvTensorRtRtx", NvTensorRtRtxExecutionProvider::AppendExecutionProvider},

test/c_api_tests.cpp

@@ -48,6 +48,31 @@ TEST(CAPITests, Config) {
 #endif
 }
 
+// Regression test: appending CPU provider should not throw.
+// See https://github.com/microsoft/onnxruntime-genai/pull/2179
+TEST(CAPITests, AppendCpuProvider) {
+#if TEST_PHI2
+  auto config = OgaConfig::Create(PHI2_PATH);
+  config->ClearProviders();
+  config->AppendProvider("cpu");
+  auto model = OgaModel::Create(*config);
+  ASSERT_NE(model.get(), nullptr);
+
+  // Also test other case variants
+  auto config2 = OgaConfig::Create(PHI2_PATH);
+  config2->ClearProviders();
+  config2->AppendProvider("CPU");
+  auto model2 = OgaModel::Create(*config2);
+  ASSERT_NE(model2.get(), nullptr);
+
+  auto config3 = OgaConfig::Create(PHI2_PATH);
+  config3->ClearProviders();
+  config3->AppendProvider("CPUExecutionProvider");
+  auto model3 = OgaModel::Create(*config3);
+  ASSERT_NE(model3.get(), nullptr);
+#endif
+}
+
 TEST(CAPITests, TokenizerCAPI) {
 #if TEST_PHI2
   auto config = OgaConfig::Create(PHI2_PATH);