diff --git a/paddle/fluid/operators/math/sequence_padding.cu b/paddle/fluid/operators/math/sequence_padding.cu
index 035e10dcbe4e2083723e47d7dda75ce267a9f141..1b433067900af71bb8a6833cef019d41f9c76858 100644
--- a/paddle/fluid/operators/math/sequence_padding.cu
+++ b/paddle/fluid/operators/math/sequence_padding.cu
@@ -78,12 +78,6 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
                    "The numel of 'pad_value' can only be 1 or be equal to the "
                    "'step_width'.");
 
-    if (!norm_by_times && seq_num == 1UL && pad_seq_len == max_seq_len) {
-      TensorCopy(seq_tensor, context.GetPlace(), context, pad_tensor);
-      pad_tensor->Resize(pad_tensor_dims);
-      return;
-    }
-
     const int kBlockSize = 512;
 
     /* At least use 32 threads to copy sequence_width elements,
@@ -129,12 +123,13 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
 
     CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len,
               step_width, layout);
-
+    /*
     if (!norm_by_times && seq_num == 1UL && pad_seq_len == max_seq_len) {
       TensorCopy(pad_tensor, context.GetPlace(), context, seq_tensor);
       seq_tensor->Resize(seq_tensor_dims);
       return;
     }
+    */
 
     const int kBlockSize = 512;
 
diff --git a/paddle/fluid/operators/matmul_op.cc b/paddle/fluid/operators/matmul_op.cc
index 242a1b9ae92ade0caf1b0f1fcb5458b8b7070d84..f18282745200cc8ef9460e60728d777112f2b798 100644
--- a/paddle/fluid/operators/matmul_op.cc
+++ b/paddle/fluid/operators/matmul_op.cc
@@ -290,8 +290,10 @@ class MatMulOp : public framework::OperatorWithKernel {
                                      context->Attrs().Get<bool>("transpose_Y"));
 
     PADDLE_ENFORCE_EQ(mat_dim_x.width_, mat_dim_y.height_);
-    PADDLE_ENFORCE(mat_dim_x.batch_size_ == mat_dim_y.batch_size_ ||
-                   mat_dim_x.batch_size_ == 0 || mat_dim_y.batch_size_ == 0);
+    if (context->IsRuntime()) {
+      PADDLE_ENFORCE(mat_dim_x.batch_size_ == mat_dim_y.batch_size_ ||
+                     mat_dim_x.batch_size_ == 0 || mat_dim_y.batch_size_ == 0);
+    }
     std::vector<int64_t> dim_out;
     if (mat_dim_x.batch_size_ != 0) {
       dim_out = framework::vectorize(dim_x);
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index 7b72b382624f43ca8c9ddb1c511dc88977c6ebbe..31e12f9b27feb134a027d8e150348c1d1da058b3 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -4901,6 +4901,9 @@ def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
 
         if len(y_shape) > 2 and len(x_shape) > 2:
             for i, dim_x in enumerate(x_shape[:-2]):
+                # don't check neg shape
+                if dim_x < 0 or y_shape[i] < 0:
+                    continue
                 if dim_x != y_shape[i]:
                     raise ValueError("Invalid inputs for matmul. x(%s), y(%s)" %
                                      (x.shape, y.shape))