diff --git a/paddle/phi/infermeta/multiary.cc b/paddle/phi/infermeta/multiary.cc
index 29c2436c8e8c0b6dd098a8dffd11bf0bc64f747b..70111217ad9a571b9e39d9c1e45746bf48834924 100644
--- a/paddle/phi/infermeta/multiary.cc
+++ b/paddle/phi/infermeta/multiary.cc
@@ -911,13 +911,14 @@ void ConcatInferMeta(const std::vector<const MetaTensor*>& x,
   // 1. calculate axis
   int rank = x.at(0)->dims().size();
   PADDLE_ENFORCE_EQ(
-      axis >= -rank && axis < rank,
+      !rank || (axis >= -rank && axis < rank),
       true,
       phi::errors::InvalidArgument(
           "The axis is expected to be in range of [%d, %d), but got %d",
           -rank,
           rank,
           axis));
+  axis = rank ? axis : 0;
   if (axis < 0) {
     axis = axis + rank;
   }
diff --git a/paddle/phi/kernels/funcs/concat_funcs.h b/paddle/phi/kernels/funcs/concat_funcs.h
index db965c2ef9b65413aa8faf2b0dd7813cd8487dd8..61a0e6ad7e59ee94710375515650f47a4a75bbd5 100644
--- a/paddle/phi/kernels/funcs/concat_funcs.h
+++ b/paddle/phi/kernels/funcs/concat_funcs.h
@@ -21,13 +21,14 @@ namespace funcs {
 
 static inline int64_t ComputeAxis(int64_t axis, int64_t rank) {
   PADDLE_ENFORCE_EQ(
-      axis >= -rank && axis < rank,
+      !rank || (axis >= -rank && axis < rank),
       true,
       phi::errors::InvalidArgument(
           "The axis is expected to be in range of [%d, %d), but got %d",
           -rank,
           rank,
           axis));
+  axis = rank ? axis : 0;
   if (axis < 0) {
     axis = axis + rank;
   }
diff --git a/paddle/phi/kernels/gpu/concat_kernel.cu b/paddle/phi/kernels/gpu/concat_kernel.cu
index 80ff71b2158241370cde44dc581e7b3f388877fa..497f78ca9fc8fa17f4bbf145c4fedb60d8bf5edb 100644
--- a/paddle/phi/kernels/gpu/concat_kernel.cu
+++ b/paddle/phi/kernels/gpu/concat_kernel.cu
@@ -34,6 +34,35 @@ void ConcatKernel(const Context& dev_ctx,
                   DenseTensor* out) {
   int64_t axis = axis_scalar.to<int64_t>();
 
+  if (UNLIKELY(x[0]->dims().size() == 0)) {
+    // for dims is 0 specially
+    phi::DDim tmp_1dim, out_dims;
+    out_dims[0] = x.size();
+    tmp_1dim[0] = 1;
+
+    out->Resize(out_dims);
+    dev_ctx.template Alloc<T>(out);
+
+    size_t output_offset = 0;
+    for (auto* in : x) {
+      if (in->numel() == 0UL) {
+        continue;
+      }
+      auto in_stride = phi::stride_numel(tmp_1dim);
+      auto out_stride = phi::stride_numel(out->dims());
+      paddle::operators::StridedNumelCopyWithAxis<T>(
+          dev_ctx,
+          axis,
+          out->data<T>() + output_offset,
+          out_stride,
+          in->data<T>(),
+          in_stride,
+          in_stride[axis]);
+      output_offset += in_stride[axis];
+    }
+    return;
+  }
+
   axis = phi::funcs::ComputeAxis(axis, x[0]->dims().size());
 
   std::vector<phi::DDim> x_dims;
diff --git a/python/paddle/fluid/tests/unittests/collective/process_group_nccl.py b/python/paddle/fluid/tests/unittests/collective/process_group_nccl.py
index 4bee60642b5c85cb3c8076ca3ec9f029c18567ad..2e585e1cc6d716b1c9bc987e4a7ea5d4d969f7a2 100644
--- a/python/paddle/fluid/tests/unittests/collective/process_group_nccl.py
+++ b/python/paddle/fluid/tests/unittests/collective/process_group_nccl.py
@@ -167,6 +167,29 @@ class TestProcessGroupFp32(unittest.TestCase):
 
             print("test broadcast api ok")
 
+            # test broadcast with shape=[]
+            # rank 0
+            x = np.random.random([]).astype(self.dtype)
+            tensor_x = paddle.to_tensor(x)
+            # rank 1
+            y = np.random.random([]).astype(self.dtype)
+            tensor_y = paddle.to_tensor(y)
+
+            broadcast_result = paddle.assign(tensor_x)
+            if pg.rank() == 0:
+                task = dist.broadcast(tensor_x, 0, sync_op=False)
+                task.synchronize()
+                paddle.device.cuda.synchronize()
+                assert task.is_completed()
+                assert np.array_equal(broadcast_result, tensor_x)
+            else:
+                task = dist.broadcast(tensor_y, 0)
+                paddle.device.cuda.synchronize()
+                assert np.array_equal(broadcast_result, tensor_y)
+            assert tensor_y.shape == []
+
+            print("test broadcast api with shape=[] ok")
+
             # test barrier
             # rank 0
             if pg.rank() == 0:
@@ -417,6 +440,30 @@ class TestProcessGroupFp32(unittest.TestCase):
                 assert np.array_equal(tensor_y, out2)
             print("test scatter api ok\n")
 
+            # test Scatter with shape=[]
+            # rank 0
+            x = np.random.random([]).astype(self.dtype)
+            y = np.random.random([]).astype(self.dtype)
+            tensor_x = paddle.to_tensor(x)
+            tensor_y = paddle.to_tensor(y)
+            if pg.rank() == 0:
+                in_1, in_2 = tensor_x, tensor_x + 1
+                task = dist.scatter(tensor_y, [in_1, in_2], 0, sync_op=True)
+                paddle.device.cuda.synchronize()
+            # rank 1
+            else:
+                task = dist.scatter(tensor_y, [], 0, sync_op=True)
+                task.wait()
+                paddle.device.cuda.synchronize()
+            out1 = paddle.assign(tensor_x)
+            out2 = paddle.assign(tensor_x + 1)
+            if pg.rank() == 0:
+                assert np.array_equal(tensor_y, out1)
+            else:
+                assert np.array_equal(tensor_y, out2), f"{tensor_y}, {out2}"
+            assert tensor_y.shape == []
+            print("test scatter api with shape=[] ok\n")
+
             # test send min
             # rank 0
             x = np.random.random(self.shape).astype(self.dtype)