diff --git a/paddle/fluid/operators/gather.cu.h b/paddle/fluid/operators/gather.cu.h
index d74d4db92528d69492ab7b90a98d3775dadc35d1..e4df59c5d51c390cf593add0c5562665c91f33f6 100644
--- a/paddle/fluid/operators/gather.cu.h
+++ b/paddle/fluid/operators/gather.cu.h
@@ -50,7 +50,9 @@ void GPUGather(const platform::DeviceContext& ctx, const Tensor& src,
                const Tensor& index, Tensor* output) {
   // PADDLE_ENFORCE(platform::is_gpu_place(place));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index.dims().size() == 1);
+  PADDLE_ENFORCE(index.dims().size() == 1 ||
+                 (index.dims().size() == 2 && index.dims()[1] == 1));
+
   int index_size = index.dims()[0];
 
   auto src_dims = src.dims();
diff --git a/paddle/fluid/operators/gather.h b/paddle/fluid/operators/gather.h
index d72e07d76c97e9e455e54980207d7c02842cc04b..dc08ee5efacde5e232d751b13aaf11f51237634a 100644
--- a/paddle/fluid/operators/gather.h
+++ b/paddle/fluid/operators/gather.h
@@ -38,7 +38,8 @@ void CPUGather(const platform::DeviceContext& ctx, const Tensor& src,
                const Tensor& index, Tensor* output) {
   PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index.dims().size() == 1);
+  PADDLE_ENFORCE(index.dims().size() == 1 ||
+                 (index.dims().size() == 2 && index.dims()[1] == 1));
   int64_t index_size = index.dims()[0];
 
   auto src_dims = src.dims();
diff --git a/paddle/fluid/operators/gather_op.cc b/paddle/fluid/operators/gather_op.cc
index f84ff206fffddef1030b7ed439e887bdfef342a6..95aa9b573c795159079bdb5401b34d7a61252115 100644
--- a/paddle/fluid/operators/gather_op.cc
+++ b/paddle/fluid/operators/gather_op.cc
@@ -31,7 +31,8 @@ class GatherOp : public framework::OperatorWithKernel {
                    "Output(Out) of GatherOp should not be null.");
 
     auto index_dims = ctx->GetInputDim("Index");
-    PADDLE_ENFORCE(index_dims.size() == 1);
+    PADDLE_ENFORCE(index_dims.size() == 1 ||
+                   (index_dims.size() == 2 && index_dims[1] == 1));
     int batch_size = ctx->GetInputDim("Index")[0];
     framework::DDim output_dims(ctx->GetInputDim("X"));
     output_dims[0] = batch_size;
@@ -53,6 +54,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
 
   void InferShape(framework::InferShapeContext* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", /*-->*/ framework::GradVarName("X"));
   }
 
  protected:
@@ -75,7 +77,7 @@ Gather Operator.
 
 $Out = X[Index]$
 
-Out is obtained by gathering entries of the outer-most dimension 
+Out is obtained by gathering entries of the outer-most dimension
 of X indexed by Index and concatenate them together.
 
 Example:
diff --git a/paddle/fluid/operators/scatter.cu.h b/paddle/fluid/operators/scatter.cu.h
index ac7d69bfb549fd98c76fcf834e8d3ad9bec2ef23..b2e79f6c82bb748293f4219845e6798347c8c46e 100644
--- a/paddle/fluid/operators/scatter.cu.h
+++ b/paddle/fluid/operators/scatter.cu.h
@@ -51,7 +51,8 @@ void GPUScatterAssign(const platform::DeviceContext& ctx, const Tensor& src,
                       const Tensor& index, Tensor* output) {
   // PADDLE_ENFORCE(platform::is_gpu_place(place));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index.dims().size() == 1);
+  PADDLE_ENFORCE(index.dims().size() == 1 ||
+                 (index.dims().size() == 2 && index.dims()[1] == 1));
   int index_size = index.dims()[0];
 
   auto src_dims = src.dims();
diff --git a/paddle/fluid/operators/scatter.h b/paddle/fluid/operators/scatter.h
index 39af717615c01f5c121e32b176b74d05be738531..8bae6606c94620ab4fa8ae34f69236e7e87e9670 100644
--- a/paddle/fluid/operators/scatter.h
+++ b/paddle/fluid/operators/scatter.h
@@ -37,7 +37,8 @@ void ScatterAssign(const platform::DeviceContext& ctx, const Tensor& src,
                    const Tensor& index, Tensor* output) {
   PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index.dims().size() == 1);
+  PADDLE_ENFORCE(index.dims().size() == 1 ||
+                 (index.dims().size() == 2 && index.dims()[1] == 1));
   int index_size = index.dims()[0];
 
   auto src_dims = src.dims();
diff --git a/paddle/fluid/operators/stack_op.cc b/paddle/fluid/operators/stack_op.cc
index 3f4b48bc7391def082c82ed451fc5a752009a2f1..9345b495415d203728238c19621a20f446c40bf5 100644
--- a/paddle/fluid/operators/stack_op.cc
+++ b/paddle/fluid/operators/stack_op.cc
@@ -21,8 +21,12 @@ REGISTER_OPERATOR(stack, ops::StackOp, ops::StackOpMaker,
 REGISTER_OPERATOR(stack_grad, ops::StackOpGrad);
 
 REGISTER_OP_CPU_KERNEL(stack, ops::StackKernel<plat::CPUDeviceContext, float>,
-                       ops::StackKernel<plat::CPUDeviceContext, double>);
+                       ops::StackKernel<plat::CPUDeviceContext, double>,
+                       ops::StackKernel<plat::CPUDeviceContext, int>,
+                       ops::StackKernel<plat::CPUDeviceContext, int64_t>);
 
 REGISTER_OP_CPU_KERNEL(stack_grad,
                        ops::StackGradKernel<plat::CPUDeviceContext, float>,
-                       ops::StackGradKernel<plat::CPUDeviceContext, double>);
+                       ops::StackGradKernel<plat::CPUDeviceContext, double>,
+                       ops::StackGradKernel<plat::CPUDeviceContext, int>,
+                       ops::StackGradKernel<plat::CPUDeviceContext, int64_t>);
diff --git a/paddle/fluid/operators/stack_op.cu b/paddle/fluid/operators/stack_op.cu
index 92c1bde2bcf089e5c715e90e564408e6ad37ba17..bf2a9e5b3d22996e688621727cb280dc9aed7859 100644
--- a/paddle/fluid/operators/stack_op.cu
+++ b/paddle/fluid/operators/stack_op.cu
@@ -18,8 +18,12 @@ namespace plat = paddle::platform;
 namespace ops = paddle::operators;
 
 REGISTER_OP_CUDA_KERNEL(stack, ops::StackKernel<plat::CUDADeviceContext, float>,
-                        ops::StackKernel<plat::CUDADeviceContext, double>);
+                        ops::StackKernel<plat::CUDADeviceContext, double>,
+                        ops::StackKernel<plat::CUDADeviceContext, int>,
+                        ops::StackKernel<plat::CUDADeviceContext, int64_t>);
 
 REGISTER_OP_CUDA_KERNEL(stack_grad,
                         ops::StackGradKernel<plat::CUDADeviceContext, float>,
-                        ops::StackGradKernel<plat::CUDADeviceContext, double>);
+                        ops::StackGradKernel<plat::CUDADeviceContext, double>,
+                        ops::StackGradKernel<plat::CUDADeviceContext, int>,
+                        ops::StackGradKernel<plat::CUDADeviceContext, int64_t>);