Fix ConfigEps execution provider ordering in test files

When BaseTester::ConfigEps() is called with multiple EPs, it runs the
first available EP for the given operator. CPU EP should be added last
to ensure other EPs get tested first.

Files fixed:
- matmul_2bits_test.cc: Move CPU EP after WebGPU EP
- skiplayernorm_op_test.cc: Move CPU EP after WebGPU EP
- gather_op_test.cc: Move CPU EP after CUDA EP (4 test cases)

Co-authored-by: fs-eire <7679871+fs-eire@users.noreply.github.com>

Author: Copilot

onnxruntime/test/contrib_ops/matmul_2bits_test.cc

@@ -197,10 +197,11 @@ void RunTest2Bits(const TestOptions2Bits& opts) {
 
   std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
   if constexpr (std::is_same<T1, float>::value) {
-    execution_providers.emplace_back(DefaultCpuExecutionProvider());
 #ifdef USE_WEBGPU
     execution_providers.push_back(DefaultWebGpuExecutionProvider());
 #endif
+    // CPU EP should be added last so that other EPs get tested first
+    execution_providers.emplace_back(DefaultCpuExecutionProvider());
     test.ConfigEps(std::move(execution_providers));
     test.RunWithConfig();
   }

onnxruntime/test/contrib_ops/skiplayernorm_op_test.cc

@@ -94,12 +94,14 @@ static void RunOneTest(
                             sum_output_data);
     }
 
-    if (cpu_ep != nullptr) {
-      execution_providers.push_back(DefaultCpuExecutionProvider());
-    }
+    // Add WebGPU EP first so it gets tested before CPU EP
+    // (ConfigEps runs the first available EP for the operator)
     if (webgpu_ep != nullptr) {
       execution_providers.push_back(DefaultWebGpuExecutionProvider());
     }
+    if (cpu_ep != nullptr) {
+      execution_providers.push_back(DefaultCpuExecutionProvider());
+    }
 
     test.Run(OpTester::ExpectResult::kExpectSuccess, "", {}, nullptr, &execution_providers);
   } else if (CudaHasBF16Support() && use_bfloat16) {

onnxruntime/test/providers/cpu/tensor/gather_op_test.cc

@@ -434,10 +434,12 @@ TEST(GatherOpTest, Gather_axis1_scalar_indices) {
 
 TEST(ShrunkenGatherOpTest, ShrunkenGather_PositiveAxis) {
   std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
-  execution_providers.emplace_back(DefaultCpuExecutionProvider());
+  // Add CUDA EP first so it gets tested before CPU EP
+  // (ConfigEps runs the first available EP for the operator)
 #ifdef USE_CUDA
   execution_providers.emplace_back(DefaultCudaExecutionProvider());
 #endif
+  execution_providers.emplace_back(DefaultCpuExecutionProvider());
 
   OpTester test("ShrunkenGather", 1, onnxruntime::kMSDomain);
   test.AddAttribute<int64_t>("axis", 0LL);
@@ -455,10 +457,12 @@ TEST(ShrunkenGatherOpTest, ShrunkenGather_PositiveAxis) {
 
 TEST(ShrunkenGatherOpTest, ShrunkenGather_NegativeAxis) {
   std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
-  execution_providers.emplace_back(DefaultCpuExecutionProvider());
+  // Add CUDA EP first so it gets tested before CPU EP
+  // (ConfigEps runs the first available EP for the operator)
 #ifdef USE_CUDA
   execution_providers.emplace_back(DefaultCudaExecutionProvider());
 #endif
+  execution_providers.emplace_back(DefaultCpuExecutionProvider());
 
   OpTester test("ShrunkenGather", 1, onnxruntime::kMSDomain);
   test.AddAttribute<int64_t>("axis", -1LL);
@@ -476,10 +480,12 @@ TEST(ShrunkenGatherOpTest, ShrunkenGather_NegativeAxis) {
 
 TEST(ShrunkenGatherOpTest, ShrunkenGather_InvalidIndicesRank) {
   std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
-  execution_providers.emplace_back(DefaultCpuExecutionProvider());
+  // Add CUDA EP first so it gets tested before CPU EP
+  // (ConfigEps runs the first available EP for the operator)
 #ifdef USE_CUDA
   execution_providers.emplace_back(DefaultCudaExecutionProvider());
 #endif
+  execution_providers.emplace_back(DefaultCpuExecutionProvider());
 
   OpTester test("ShrunkenGather", 1, onnxruntime::kMSDomain);
   test.AddAttribute<int64_t>("axis", 0LL);
@@ -497,10 +503,12 @@ TEST(ShrunkenGatherOpTest, ShrunkenGather_InvalidIndicesRank) {
 
 TEST(ShrunkenGatherOpTest, ShrunkenGather_InvalidInputRank) {
   std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
-  execution_providers.emplace_back(DefaultCpuExecutionProvider());
+  // Add CUDA EP first so it gets tested before CPU EP
+  // (ConfigEps runs the first available EP for the operator)
 #ifdef USE_CUDA
   execution_providers.emplace_back(DefaultCudaExecutionProvider());
 #endif
+  execution_providers.emplace_back(DefaultCpuExecutionProvider());
 
   OpTester test("ShrunkenGather", 1, onnxruntime::kMSDomain);
   test.AddAttribute<int64_t>("axis", 0LL);