diff --git a/docs/ContribOperators.md b/docs/ContribOperators.md
index 44d6c8ae4b..4ef770568c 100644
--- a/docs/ContribOperators.md
+++ b/docs/ContribOperators.md
@@ -359,7 +359,7 @@ This version of the operator has been available since version 1 of the 'com.micr
 <dt><tt>data</tt> : T</dt>
 <dd>The input data as Tensor.</dd>
 <dt><tt>bias</tt> : T</dt>
-<dd>The bias input, a vector with the same shape as last dim of data</dd>
+<dd>The bias input, a vector with the same shape as last dim of data OR same shape with data</dd>
 <dt><tt>residual</tt> (optional) : T</dt>
 <dd>The residual input, must have the same shape as data</dd>
 <dt><tt>ratio</tt> (optional) : T1</dt>
diff --git a/onnxruntime/contrib_ops/cuda/math/bias_dropout.cc b/onnxruntime/contrib_ops/cuda/math/bias_dropout.cc
index 950106d879..192c28c86d 100644
--- a/onnxruntime/contrib_ops/cuda/math/bias_dropout.cc
+++ b/onnxruntime/contrib_ops/cuda/math/bias_dropout.cc
@@ -36,7 +36,8 @@ struct BiasDropoutComputeImpl {
                     const Tensor& bias,
                     const Tensor* residual,
                     Tensor& Y,
-                    bool* mask_data) const {
+                    bool* mask_data,
+                    bool has_same_shape_bias) const {
     typedef typename ToCudaType<T>::MappedType CudaT;
 
     const CudaT* X_data = reinterpret_cast<const CudaT*>(X.template Data<T>());
@@ -52,7 +53,7 @@ struct BiasDropoutComputeImpl {
 
     CudaT* Y_data = reinterpret_cast<CudaT*>(Y.template MutableData<T>());
 
-    BiasDropoutKernelImpl<CudaT>(prop, stream, N, fdm_dim, ratio_data, generator, X_data, bias_data, residual_data, Y_data, mask_data);
+    BiasDropoutKernelImpl<CudaT>(prop, stream, N, fdm_dim, ratio_data, generator, X_data, bias_data, residual_data, Y_data, mask_data, has_same_shape_bias);
 
     return Status::OK();
   }
@@ -70,12 +71,16 @@ Status BiasDropout::ComputeInternal(OpKernelContext* context) const {
   const Tensor* bias = context->Input<Tensor>(1);
   if (bias == nullptr) return Status(common::ONNXRUNTIME, common::FAIL, "Bias input of BiasDropout is not available.");
   const TensorShape& bias_shape = bias->Shape();
-  if (bias_shape.NumDimensions() != 1) {
-    return Status(common::ONNXRUNTIME, common::FAIL, "Bias input is not a 1D tensor.");
-  }
-  const int64_t dim = bias_shape[0];
-  if (dim != x_shape.GetDims().back()) {
-    return Status(common::ONNXRUNTIME, common::FAIL, "Bias' dimension doesn't match input's last dimension.");
+  const int64_t dim = bias_shape.GetDims().back();
+  bool has_same_shape_bias = (bias_shape == x_shape);
+  if (!has_same_shape_bias) {
+    if (bias_shape.NumDimensions() != 1) {
+      return Status(common::ONNXRUNTIME, common::FAIL, "Bias input is not a 1D tensor.");
+    }
+
+    if (dim != x_shape.GetDims().back()) {
+      return Status(common::ONNXRUNTIME, common::FAIL, "Bias' dimension doesn't match input's last dimension.");
+    }
   }
 
   //Get residual_data
@@ -114,9 +119,9 @@ Status BiasDropout::ComputeInternal(OpKernelContext* context) const {
 
   utils::MLTypeCallDispatcher<ALL_IEEE_FLOAT_DATA_TYPES> t_disp(X->GetElementType());
   return t_disp.InvokeRet<Status, BiasDropoutComputeImpl>(
-      GetDeviceProp(), Stream(), N, fdm_dim, ratio_data, generator, *X, *bias, residual, *Y, mask_data);
+      GetDeviceProp(), Stream(), N, fdm_dim, ratio_data, generator, *X, *bias, residual, *Y, mask_data, has_same_shape_bias);
 }
 
 }  // namespace cuda
 }  // namespace contrib
-}  // namespace onnxruntime
\ No newline at end of file
+}  // namespace onnxruntime
diff --git a/onnxruntime/contrib_ops/cuda/math/bias_dropout.h b/onnxruntime/contrib_ops/cuda/math/bias_dropout.h
index 97c294e641..eb931dbd54 100644
--- a/onnxruntime/contrib_ops/cuda/math/bias_dropout.h
+++ b/onnxruntime/contrib_ops/cuda/math/bias_dropout.h
@@ -26,7 +26,8 @@ void BiasDropoutKernelImpl(
     const T* bias_data,
     const T* residual_data,
     T* Y_data,
-    bool* mask_data);
+    bool* mask_data,
+    bool has_same_shape_bias);
 
 class BiasDropout final : public CudaKernel {
  public:
diff --git a/onnxruntime/contrib_ops/cuda/math/bias_dropout_impl.cu b/onnxruntime/contrib_ops/cuda/math/bias_dropout_impl.cu
index 27fb09c969..ad2ff05a58 100644
--- a/onnxruntime/contrib_ops/cuda/math/bias_dropout_impl.cu
+++ b/onnxruntime/contrib_ops/cuda/math/bias_dropout_impl.cu
@@ -28,7 +28,7 @@ namespace cuda {
 
 constexpr int UNROLL = 4;
 
-template <typename T, bool has_residual>
+template <typename T, bool has_same_shape_bias, bool has_residual>
 __global__ void BiasDropoutKernel(
     const int64_t N,
     const fast_divmod fdm_dim,
@@ -58,14 +58,19 @@ __global__ void BiasDropoutKernel(
   //   use of Philox_4x32_10 is to generate a multiple of 4 times number of threads.
   for (CUDA_LONG id = idx * UNROLL; id < N; id += step_size) {
     rand = curand_uniform4(&state);
-  
+
     // actual computation
     #pragma unroll
     for (int i = 0; i < UNROLL; i++) {
       CUDA_LONG li = id + i;
       if (li < N) {
-        int offset = fdm_dim.mod(li);
-        float bias = float(bias_data[offset]);
+        float bias;
+        if (has_same_shape_bias) {
+          bias = float(bias_data[li]);
+        } else {
+          int offset = fdm_dim.mod(li);
+          bias = float(bias_data[offset]);
+        }
 
         mask_data[li] = (&rand.x)[i] < p;
         float output_data = (float(X_data[li]) + bias) * mask_data[li] * scale;
@@ -83,7 +88,7 @@ __global__ void BiasDropoutKernel(
 }
 
 
-template <typename T, bool has_residual>
+template <typename T, bool has_same_shape_bias, bool has_residual>
 __global__ void BiasDropoutVectorizedKernel(
     const int64_t N,
     const fast_divmod fdm_dim,
@@ -105,7 +110,7 @@ __global__ void BiasDropoutVectorizedKernel(
 
   float4 rand;
 
-  // using vectorized data load/store approach when N % 4 == 0 
+  // using vectorized data load/store approach when N % 4 == 0
   // since this is typical case for input shape size
   using LoadT = aligned_vector<T, UNROLL>;
   using MaskLoadT = aligned_vector<bool, UNROLL>;
@@ -113,8 +118,14 @@ __global__ void BiasDropoutVectorizedKernel(
 
   for (CUDA_LONG id = idx * UNROLL; id < N; id += step_size) {
     rand = curand_uniform4(&state);
-  
+
     // vectorized load into storage
+    T bias_vec[UNROLL];
+    if (has_same_shape_bias) {
+      LoadT *value0 = reinterpret_cast<LoadT*>(&bias_vec);
+      *value0 = *reinterpret_cast<const LoadT*>(&bias_data[id]);
+    }
+
     T src[UNROLL];
     LoadT *value1 = reinterpret_cast<LoadT*>(&src);
     *value1 = *reinterpret_cast<const LoadT*>(&X_data[id]);
@@ -131,8 +142,13 @@ __global__ void BiasDropoutVectorizedKernel(
     // actual computation
     #pragma unroll
     for (int ii = 0; ii < UNROLL; ii++) {
-      int offset = fdm_dim.mod(id + ii);
-      float bias = float(bias_data[offset]);
+      float bias;
+      if (has_same_shape_bias) {
+        bias = float(bias_vec[ii]);
+      } else {
+        int offset = fdm_dim.mod(id + ii);
+        bias = float(bias_data[offset]);
+      }
 
       mask[ii] = (&rand.x)[ii] < p;
       float output_data = (float(src[ii]) + bias) * mask[ii] * scale;
@@ -162,7 +178,8 @@ void BiasDropoutKernelImpl(
     const T* bias_data,
     const T* residual_data,
     T* Y_data,
-    bool* mask_data) {
+    bool* mask_data,
+    bool has_same_shape_bias) {
   const int block_size = 256;
   const int blocks_per_sm = prop.maxThreadsPerMultiProcessor / block_size;
   const int grid_size = std::min(prop.multiProcessorCount * blocks_per_sm, static_cast<int>(CeilDiv(N, block_size * UNROLL)));
@@ -172,16 +189,32 @@ void BiasDropoutKernelImpl(
   auto seeds = generator.NextPhiloxSeeds(counter_offset);
 
   if (N % UNROLL != 0) {
-    if (residual_data == nullptr) {
-      BiasDropoutKernel<T, false><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+    if (has_same_shape_bias) {
+      if (residual_data == nullptr) {
+        BiasDropoutKernel<T, true, false><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      } else {
+        BiasDropoutKernel<T, true, true><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      }
     } else {
-      BiasDropoutKernel<T, true><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      if (residual_data == nullptr) {
+        BiasDropoutKernel<T, false, false><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      } else {
+        BiasDropoutKernel<T, false, true><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      }
     }
   } else {
-    if (residual_data == nullptr) {
-      BiasDropoutVectorizedKernel<T, false><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+    if (has_same_shape_bias) {
+      if (residual_data == nullptr) {
+        BiasDropoutVectorizedKernel<T, true, false><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      } else {
+        BiasDropoutVectorizedKernel<T, true, true><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      }
     } else {
-      BiasDropoutVectorizedKernel<T, true><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      if (residual_data == nullptr) {
+        BiasDropoutVectorizedKernel<T, false, false><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      } else {
+        BiasDropoutVectorizedKernel<T, false, true><<<grid_size, block_size, 0, stream>>>(N, fdm_dim, ratio, seeds, X_data, bias_data, residual_data, Y_data, mask_data);
+      }
     }
   }
 }
@@ -198,7 +231,9 @@ void BiasDropoutKernelImpl(
       const T* bias_data,           \
       const T* residual_data,       \
       T* Y_data,                    \
-      bool* mask_data);
+      bool* mask_data,              \
+      bool has_same_shape_bias);
+
 
 SPECIALIZED_BIAS_DROPOUT_IMPL(float)
 SPECIALIZED_BIAS_DROPOUT_IMPL(double)
diff --git a/onnxruntime/core/graph/contrib_ops/contrib_defs.cc b/onnxruntime/core/graph/contrib_ops/contrib_defs.cc
index dfaf089979..375f45eb5e 100644
--- a/onnxruntime/core/graph/contrib_ops/contrib_defs.cc
+++ b/onnxruntime/core/graph/contrib_ops/contrib_defs.cc
@@ -2856,7 +2856,7 @@ It's an extension of Gelu. It takes the sum of input A and bias input B as the i
       .Attr("seed", "(Optional) Seed to the random generator, if not specified we will auto generate one.", AttributeProto::INT, OPTIONAL_VALUE)
       .AllowUncheckedAttributes()
       .Input(0, "data", "The input data as Tensor.", "T")
-      .Input(1, "bias", "The bias input, a vector with the same shape as last dim of data", "T")
+      .Input(1, "bias", "The bias input, a vector with the same shape as last dim of data OR same shape with data", "T")
       .Input(2, "residual", "The residual input, must have the same shape as data", "T", OpSchema::Optional)
       .Input(3, "ratio",
              "The ratio of random dropout, with value in [0, 1). If this input was not set, "
diff --git a/onnxruntime/core/optimizer/bias_dropout_fusion.cc b/onnxruntime/core/optimizer/bias_dropout_fusion.cc
index 0095f94ddc..2e662ee570 100644
--- a/onnxruntime/core/optimizer/bias_dropout_fusion.cc
+++ b/onnxruntime/core/optimizer/bias_dropout_fusion.cc
@@ -10,6 +10,18 @@ using namespace ONNX_NAMESPACE;
 using namespace ::onnxruntime::common;
 namespace onnxruntime {
 
+static bool IsSameShape(const TensorShapeProto& shape1, const TensorShapeProto& shape2) {
+  int rank1 = shape1.dim_size();
+  if (rank1 != shape2.dim_size()) {
+    return false;
+  }
+  bool same_shape = true;
+  for (int i = 0; i < rank1; ++i) {
+    same_shape &= ONNX_NAMESPACE::operator==(shape1.dim(i), shape2.dim(i));
+  }
+  return same_shape;
+}
+
 void FuseResidualAddIfAny(Graph& graph, const Node& dropout_node,
                           std::vector<NodeArg*>& dropout_input,
                           std::vector<NodeArg*>& dropout_output,
@@ -37,17 +49,12 @@ void FuseResidualAddIfAny(Graph& graph, const Node& dropout_node,
         if (input1_shape == nullptr ||
             input2_shape == nullptr ||
             input1_shape->dim_size() < 1 ||
-            input2_shape->dim_size() < 1 ||
-            input1_shape->dim_size() != input2_shape->dim_size()) {
+            input2_shape->dim_size() < 1) {
           continue;
         }
 
         // Inputs of Residual Add must match in shape
-        bool match = true;
-        for (int i = 0; i < input1_shape->dim_size(); ++i) {
-          match &= ONNX_NAMESPACE::operator==(input1_shape->dim(i), input2_shape->dim(i));
-        }
-        if (!match) {
+        if (!IsSameShape(*input1_shape, *input2_shape)) {
           continue;
         }
 
@@ -107,22 +114,29 @@ Status BiasDropoutFusion::ApplyImpl(Graph& graph, bool& modified, int graph_leve
       continue;
     }
 
-    int last_dim_shape1 = input1_shape->dim_size() - 1;
-    int last_dim_shape2 = input2_shape->dim_size() - 1;
-    if (!utils::HasDimValue(input1_shape->dim(last_dim_shape1)) ||
-        !utils::HasDimValue(input2_shape->dim(last_dim_shape2)) ||
-        input1_shape->dim(last_dim_shape1).dim_value() != input2_shape->dim(last_dim_shape2).dim_value()) {
-      continue;
-    }
-
-    if (input1_shape->dim_size() == 1) {
-      dropout_input.push_back(node.MutableInputDefs()[1]);  // dropout input
-      dropout_input.push_back(node.MutableInputDefs()[0]);  // bias
-    } else if (input2_shape->dim_size() == 1) {
+    if (IsSameShape(*input1_shape, *input2_shape)) {
       dropout_input.push_back(node.MutableInputDefs()[0]);  // dropout input
       dropout_input.push_back(node.MutableInputDefs()[1]);  // bias
     } else {
-      continue;
+      const int last_dim_shape1 = input1_shape->dim_size() - 1;
+      const int last_dim_shape2 = input2_shape->dim_size() - 1;
+      if (!(utils::HasDimValue(input1_shape->dim(last_dim_shape1)) &&
+            utils::HasDimValue(input2_shape->dim(last_dim_shape2)) &&
+            input1_shape->dim(last_dim_shape1).dim_value() == input2_shape->dim(last_dim_shape2).dim_value()) &&
+          !(utils::HasDimParam(input1_shape->dim(last_dim_shape1)) &&
+            utils::HasDimParam(input2_shape->dim(last_dim_shape2)) &&
+            input1_shape->dim(last_dim_shape1).dim_param() == input2_shape->dim(last_dim_shape2).dim_param())) {
+        continue;  // continue if no same DimValue && no same DimParam
+      }
+      if (input1_shape->dim_size() == 1) {
+        dropout_input.push_back(node.MutableInputDefs()[1]);  // dropout input
+        dropout_input.push_back(node.MutableInputDefs()[0]);  // bias
+      } else if (input2_shape->dim_size() == 1) {
+        dropout_input.push_back(node.MutableInputDefs()[0]);  // dropout input
+        dropout_input.push_back(node.MutableInputDefs()[1]);  // bias
+      } else {
+        continue;
+      }
     }
     Node& add_node = node;
     nodes_to_fuse.push_back(add_node);
@@ -163,7 +177,7 @@ Status BiasDropoutFusion::ApplyImpl(Graph& graph, bool& modified, int graph_leve
     const std::string op_type = "BiasDropout";
     Node& dropout_add_fusion_node = graph.AddNode(graph.GenerateNodeName(op_type),
                                                   op_type,
-                                                  "fused Add and Dropout",
+                                                  "fused Add-Dropout-(Add) for " + dropout_node.Name(),
                                                   dropout_input,
                                                   dropout_output,
                                                   {},
diff --git a/onnxruntime/test/contrib_ops/bias_dropout_op_test.cc b/onnxruntime/test/contrib_ops/bias_dropout_op_test.cc
index b16b3f190c..eaa583a633 100644
--- a/onnxruntime/test/contrib_ops/bias_dropout_op_test.cc
+++ b/onnxruntime/test/contrib_ops/bias_dropout_op_test.cc
@@ -30,7 +30,8 @@ enum TrainingMode { TrainingFalse,
 #if defined(USE_CUDA) || defined(USE_ROCM)
 namespace {
 void RunBiasDropoutTest(const bool use_mask, const std::vector<int64_t>& input_shape, float ratio = -1.0f,
-                        TrainingMode training_mode = TrainingTrue, bool use_float16_ratio = false, bool has_residual = true) {
+                        TrainingMode training_mode = TrainingTrue, bool use_float16_ratio = false,
+                        bool has_residual = true, bool has_same_shape_bias = false) {
   OpTester t{"BiasDropout", 1, kMSDomain};
   const int64_t seed = 42;
   t.AddAttribute("seed", seed);
@@ -40,8 +41,13 @@ void RunBiasDropoutTest(const bool use_mask, const std::vector<int64_t>& input_s
   const std::vector<float> input = ValueRange(input_size, 1.0f, 1.0f);
   t.AddInput("data", input_shape, input);
 
-  std::vector<int64_t> bias_shape{input_shape.back()};
-  const auto bias_size = input_shape.back();
+  std::vector<int64_t> bias_shape;
+  if (has_same_shape_bias) {
+    bias_shape = input_shape;
+  } else {
+    bias_shape.push_back(input_shape.back());
+  }
+  const auto bias_size = has_same_shape_bias ? input_size : input_shape.back();
   const std::vector<float> bias = ValueRange(bias_size, 2.0f, 1.0f);
   t.AddInput("bias", bias_shape, bias);
 
@@ -143,7 +149,7 @@ TEST(BiasDropoutTest, BasicWithoutResidualAndNotVectorized) {
   RunBiasDropoutTest(false, {10, 5, 5}, 0.75f, TrainingTrue, false, false);
 }
 TEST(BiasDropoutTest, MaskAndNotVectorized) {
-  RunBiasDropoutTest(true, {3, 5, 100}, 0.25f);
+  RunBiasDropoutTest(true, {3, 5, 10}, 0.25f);
 }
 
 // N % 4 == 0
@@ -178,6 +184,15 @@ TEST(BiasDropoutTest, RatioLimit) {
 TEST(BiasDropoutTest, EmptyRatio) {
   RunBiasDropoutTest(true, {2, 7, 1024});
 }
+
+// has_same_bias_shape == true
+TEST(BiasDropoutTest, BasicBiasSameShape) {
+  RunBiasDropoutTest(false, {10, 10, 10}, 0.75f, TrainingTrue, false, true, true);
+}
+
+TEST(BiasDropoutTest, BasicBiasSameShapeNotVectorized) {
+  RunBiasDropoutTest(false, {10, 5, 5}, 0.75f, TrainingTrue, false, true, true);
+}
 #endif
 
 }  // namespace test
diff --git a/onnxruntime/test/optimizer/graph_transform_test.cc b/onnxruntime/test/optimizer/graph_transform_test.cc
index cebd850cff..876e3caf6c 100644
--- a/onnxruntime/test/optimizer/graph_transform_test.cc
+++ b/onnxruntime/test/optimizer/graph_transform_test.cc
@@ -3053,6 +3053,12 @@ TEST_F(GraphTransformationTests, BiasDropoutFusionTest) {
   TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_residual_fusion_mismatch.onnx", *logger_, 1);
   TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_residual_fusion_multiple_consumers1.onnx", *logger_, 1);
   TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_residual_fusion_multiple_consumers2.onnx", *logger_, 1);
+  TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_same_shape_fusion.onnx", *logger_);
+  TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_residual_same_shape_fusion.onnx", *logger_);
+  TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_fusion_dim_is_param.onnx", *logger_);
+  TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_residual_fusion_dim_is_param.onnx", *logger_);
+  TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_same_shape_fusion_dim_is_param.onnx", *logger_);
+  TestBiasDropoutFusion(MODEL_FOLDER "fusion/bias_dropout_residual_same_shape_fusion_dim_is_param.onnx", *logger_);
 }
 
 TEST_F(GraphTransformationTests, LayerNormFusionTest) {
diff --git a/onnxruntime/test/testdata/transform/fusion/bias_dropout_fusion_dim_is_param.onnx b/onnxruntime/test/testdata/transform/fusion/bias_dropout_fusion_dim_is_param.onnx
new file mode 100644
index 0000000000..f61f761c2a
Binary files /dev/null and b/onnxruntime/test/testdata/transform/fusion/bias_dropout_fusion_dim_is_param.onnx differ
diff --git a/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_fusion_dim_is_param.onnx b/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_fusion_dim_is_param.onnx
new file mode 100644
index 0000000000..47898128f6
Binary files /dev/null and b/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_fusion_dim_is_param.onnx differ
diff --git a/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_gen.py b/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_gen.py
index b0642194c6..aa010fb409 100644
--- a/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_gen.py
+++ b/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_gen.py
@@ -133,3 +133,102 @@ graph = helper.make_graph(
 
 model = helper.make_model(graph, producer_name='onnx-example', **kwargs)
 onnx.save(model, 'bias_dropout_residual_fusion_multiple_consumers2.onnx')
+
+
+# Create the model (ModelProto)
+A2 = helper.make_tensor_value_info('A2', TensorProto.FLOAT, ['unk_1', 'unk_2', 3072])
+
+bias = helper.make_node("Add", ["A", "A2"], ["add0_out"], "add0")
+dropout_12 = helper.make_node("Dropout", ["add0_out", "ratio_const", "training_mode"], ["C", "mask"], "dropout0")
+
+graph = helper.make_graph(
+    [bias, dropout_12],
+    "Bias_Dropout_Fusion",  #name
+    [A, A2],
+    [C],
+    [ratio, training_mode])
+
+model = helper.make_model(graph, producer_name='onnx-example', **kwargs)
+onnx.save(model, 'bias_dropout_same_shape_fusion.onnx')
+
+# Create the model (ModelProto)
+bias = helper.make_node("Add", ["A", "A2"], ["add0_out"], "add0")
+dropout_12 = helper.make_node("Dropout", ["add0_out", "ratio_const", "training_mode"], ["dropout_out", "mask"], "dropout0")
+residual = helper.make_node("Add", ["dropout_out", "R"], ["C"], "add1")
+
+graph = helper.make_graph(
+    [bias, dropout_12, residual],
+    "Bias_Dropout_Fusion",  #name
+    [A, A2, R],
+    [C],
+    [ratio, training_mode])
+
+model = helper.make_model(graph, producer_name='onnx-example', **kwargs)
+onnx.save(model, 'bias_dropout_residual_same_shape_fusion.onnx')
+
+
+# Create the model (ModelProto)
+A_unk = helper.make_tensor_value_info('A_unk', TensorProto.FLOAT, ['unk_1', 'unk_2', 'unk_3'])
+B_unk = helper.make_tensor_value_info('B_unk', TensorProto.FLOAT, ['unk_3'])
+C_unk = helper.make_tensor_value_info('C_unk', TensorProto.FLOAT, ['unk_1', 'unk_2', 'unk_3'])
+
+bias = helper.make_node("Add", ["A_unk", "B_unk"], ["add0_out"], "add0")
+dropout_12 = helper.make_node("Dropout", ["add0_out", "ratio_const", "training_mode"], ["C_unk", "mask"], "dropout0")
+
+graph = helper.make_graph(
+    [bias, dropout_12],
+    "Bias_Dropout_Fusion",  #name
+    [A_unk, B_unk],
+    [C_unk],
+    [ratio, training_mode])
+
+model = helper.make_model(graph, producer_name='onnx-example', **kwargs)
+onnx.save(model, 'bias_dropout_fusion_dim_is_param.onnx')
+
+# Create the model (ModelProto)
+R_unk = helper.make_tensor_value_info('R_unk', TensorProto.FLOAT, ['unk_1', 'unk_2', 'unk_3'])
+
+bias = helper.make_node("Add", ["A_unk", "B_unk"], ["add0_out"], "add0")
+dropout_12 = helper.make_node("Dropout", ["add0_out", "ratio_const", "training_mode"], ["dropout_out", "mask"], "dropout0")
+residual = helper.make_node("Add", ["dropout_out", "R_unk"], ["C_unk"], "add1")
+
+graph = helper.make_graph(
+    [bias, dropout_12, residual],
+    "Bias_Dropout_Fusion",  #name
+    [A_unk, B_unk, R_unk],
+    [C_unk],
+    [ratio, training_mode])
+
+model = helper.make_model(graph, producer_name='onnx-example', **kwargs)
+onnx.save(model, 'bias_dropout_residual_fusion_dim_is_param.onnx')
+
+# Create the model (ModelProto)
+A_unk2 = helper.make_tensor_value_info('A_unk2', TensorProto.FLOAT, ['unk_1', 'unk_2', 'unk_3'])
+
+bias = helper.make_node("Add", ["A_unk", "A_unk2"], ["add0_out"], "add0")
+dropout_12 = helper.make_node("Dropout", ["add0_out", "ratio_const", "training_mode"], ["C_unk", "mask"], "dropout0")
+
+graph = helper.make_graph(
+    [bias, dropout_12],
+    "Bias_Dropout_Fusion",  #name
+    [A_unk, A_unk2],
+    [C_unk],
+    [ratio, training_mode])
+
+model = helper.make_model(graph, producer_name='onnx-example', **kwargs)
+onnx.save(model, 'bias_dropout_same_shape_fusion_dim_is_param.onnx')
+
+# Create the model (ModelProto)
+bias = helper.make_node("Add", ["A_unk", "A_unk2"], ["add0_out"], "add0")
+dropout_12 = helper.make_node("Dropout", ["add0_out", "ratio_const", "training_mode"], ["dropout_out", "mask"], "dropout0")
+residual = helper.make_node("Add", ["dropout_out", "R_unk"], ["C_unk"], "add1")
+
+graph = helper.make_graph(
+    [bias, dropout_12, residual],
+    "Bias_Dropout_Fusion",  #name
+    [A_unk, A_unk2, R_unk],
+    [C_unk],
+    [ratio, training_mode])
+
+model = helper.make_model(graph, producer_name='onnx-example', **kwargs)
+onnx.save(model, 'bias_dropout_residual_same_shape_fusion_dim_is_param.onnx')
diff --git a/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_same_shape_fusion.onnx b/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_same_shape_fusion.onnx
new file mode 100644
index 0000000000..56488c30a1
Binary files /dev/null and b/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_same_shape_fusion.onnx differ
diff --git a/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_same_shape_fusion_dim_is_param.onnx b/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_same_shape_fusion_dim_is_param.onnx
new file mode 100644
index 0000000000..caaedef826
Binary files /dev/null and b/onnxruntime/test/testdata/transform/fusion/bias_dropout_residual_same_shape_fusion_dim_is_param.onnx differ
diff --git a/onnxruntime/test/testdata/transform/fusion/bias_dropout_same_shape_fusion.onnx b/onnxruntime/test/testdata/transform/fusion/bias_dropout_same_shape_fusion.onnx
new file mode 100644
index 0000000000..51a0b64b98
Binary files /dev/null and b/onnxruntime/test/testdata/transform/fusion/bias_dropout_same_shape_fusion.onnx differ
diff --git a/onnxruntime/test/testdata/transform/fusion/bias_dropout_same_shape_fusion_dim_is_param.onnx b/onnxruntime/test/testdata/transform/fusion/bias_dropout_same_shape_fusion_dim_is_param.onnx
new file mode 100644
index 0000000000..7bf25a713b
Binary files /dev/null and b/onnxruntime/test/testdata/transform/fusion/bias_dropout_same_shape_fusion_dim_is_param.onnx differ