[WebGPU] Fix SkipSimplifiedLayerNormalization bias

onnxruntime/contrib_ops/webgpu/bert/skip_layer_norm.cc

@@ -146,8 +146,8 @@ Status SkipLayerNorm<simplified>::ComputeInternal(onnxruntime::webgpu::ComputeCo
   const Tensor* skip = context.Input(1);
   const Tensor* gamma = context.Input(2);
   // optional
-  const Tensor* beta = context.Input(3);
-  const Tensor* bias = context.Input(4);
+  const Tensor* beta = simplified ? nullptr : context.Input(3);
+  const Tensor* bias = context.Input(simplified ? 3 : 4);
 
   const auto x_shape = x->Shape();
 
@@ -168,9 +168,10 @@ Status SkipLayerNorm<simplified>::ComputeInternal(onnxruntime::webgpu::ComputeCo
   const auto skip_shape = skip->Shape();
   const uint32_t skip_size = onnxruntime::narrow<uint32_t>(skip_shape.Size());
 
-  SkipLayerNormProgram program{beta != nullptr, bias != nullptr, epsilon_, hidden_size, has_input_skip_bias_sum, simplified, split_hidden_dim};
+  SkipLayerNormProgram program{
+      beta != nullptr, bias != nullptr, epsilon_, hidden_size, has_input_skip_bias_sum, simplified, split_hidden_dim};
   program
-      .CacheHint(simplified, has_input_skip_bias_sum, split_hidden_dim)
+      .CacheHint(simplified, beta != nullptr, bias != nullptr, has_input_skip_bias_sum, split_hidden_dim)
       .AddInputs({{x, ProgramTensorMetadataDependency::Type, components}})
       .AddInputs({{skip, ProgramTensorMetadataDependency::Type, components}})
       .AddInputs({{gamma, ProgramTensorMetadataDependency::Type, components}})

onnxruntime/test/contrib_ops/skiplayernorm_op_test.cc

@@ -875,6 +875,61 @@ TEST(SkipLayerNormTest, SkipSimplifiedLayerNormBatch1_Float16) {
           simplified);
 }
 
+TEST(SkipLayerNormTest, SkipSimplifiedLayerNormBatch1_Bias_Float16) {
+  int batch_size = 1;
+  int sequence_length = 1;
+  int hidden_size = 8;
+
+  std::vector<float> input_data = {
+      0.12573242f, -0.13208008f, 0.640625f, 0.10491943f,
+      -0.53564453f, 0.36157227f, 1.3037109f, 0.94726562f};
+
+  std::vector<float> skip_data = {
+      -0.70361328f, -1.265625f, -0.62304688f, 0.041320801f,
+      -2.3242188f, -0.21875f, -1.2460938f, -0.73242188f};
+
+  std::vector<float> gamma_data = {
+      0.94580078f, 0.96826172f, 1.0410156f, 1.1044922f,
+      0.98730469f, 1.1367188f, 0.93359375f, 1.0351562f};
+
+  std::vector<float> bias_data = {
+      0.45166016f, 0.04699707f, -0.37182617f, -0.4609375f,
+      -0.22888184f, 0.11010742f, -0.50488281f, -0.10461426f};
+
+  std::vector<float> output_data = {
+      -0.098144531f, -1.0732422f, -0.30273438f, -0.28515625f,
+      -2.5019531f, 0.23596191f, -0.34277344f, 0.09362793f};
+
+  std::vector<float> sum_output_data = {
+      -0.12646484f, -1.3505859f, -0.35424805f, -0.31469727f,
+      -3.0878906f, 0.25292969f, -0.44726562f, 0.11022949f};
+
+  bool use_float16 = true;
+  bool use_bfloat16 = false;
+  bool no_beta = true;
+  bool simplified = true;
+
+  if (DefaultWebGpuExecutionProvider().get() == nullptr && DefaultDmlExecutionProvider().get() != nullptr) {
+    GTEST_SKIP() << "DirectML does not support this test case.";
+  }
+
+  RunTest(input_data,
+          skip_data,
+          gamma_data,
+          std::vector<float>(),
+          bias_data,
+          output_data,
+          sum_output_data,
+          1e-5f,
+          batch_size,
+          sequence_length,
+          hidden_size,
+          use_float16,
+          use_bfloat16,
+          no_beta,
+          simplified);
+}
+
 TEST(SkipLayerNormTest, SkipLayerNormBatch2_Skip_Broadcast_No_Batch_Size) {
   int batch_size = 2;
   int sequence_length = 2;