diff --git a/paddle/fluid/operators/affine_channel_op.cu b/paddle/fluid/operators/affine_channel_op.cu
index cddc288c24c2bdbe12c553a6266f960acb6860f8..5fa1e18553bd53231691f3fda9de251d646f5b29 100644
--- a/paddle/fluid/operators/affine_channel_op.cu
+++ b/paddle/fluid/operators/affine_channel_op.cu
@@ -71,7 +71,11 @@ class AffineChannelCUDAKernel : public framework::OpKernel<T> {
     const T* bias_d = bias->data<T>();
     T* y_d = y->data<T>();
 
+#ifdef PADDLE_WITH_HIP
+    int block = 256;
+#else
     int block = 1024;
+#endif  // PADDLE_WITH_HIP
     int grid = (num + block - 1) / block;
 
     int max_threads = dev_ctx.GetMaxPhysicalThreadCount();
@@ -153,7 +157,11 @@ class AffineChannelGradCUDAKernel : public framework::OpKernel<T> {
     T* ds_d = dscale ? dscale->mutable_data<T>(ctx.GetPlace()) : nullptr;
     T* db_d = dbias ? dbias->mutable_data<T>(ctx.GetPlace()) : nullptr;
 
+#ifdef PADDLE_WITH_HIP
+    const int block = 256;
+#else
     const int block = 1024;
+#endif  // PADDLE_WITH_HIP
     int max_threads = dev_ctx.GetMaxPhysicalThreadCount();
     const int max_blocks = std::max(max_threads / block, 1);
     int grid1 = (num + block - 1) / block;
diff --git a/paddle/fluid/operators/lookup_table_op.cu b/paddle/fluid/operators/lookup_table_op.cu
index 3e06e5caed31797b191dabc9846105d7a2480ad5..6985b9167571733a3116e2485cf81b3a217f536c 100644
--- a/paddle/fluid/operators/lookup_table_op.cu
+++ b/paddle/fluid/operators/lookup_table_op.cu
@@ -105,9 +105,24 @@ class LookupTableCUDAKernel : public framework::OpKernel<T> {
     auto *table = table_t->data<T>();
     auto *output = output_t->mutable_data<T>(context.GetPlace());
 
+#ifdef PADDLE_WITH_HIP
+    dim3 threads(64, 4);
+#else
     dim3 threads(128, 8);
+#endif  // PADDLE_WITH_HIP
     dim3 grids(8, 1);
-
+#ifdef PADDLE_WITH_HIP
+    if (padding_idx == -1)
+      LookupTable<
+          T, 64, 4, 8,
+          false><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
+          output, table, ids, N, K, D, padding_idx);
+    else
+      LookupTable<
+          T, 64, 4, 8,
+          true><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
+          output, table, ids, N, K, D, padding_idx);
+#else
     if (padding_idx == -1)
       LookupTable<
           T, 128, 8, 8,
@@ -118,6 +133,7 @@ class LookupTableCUDAKernel : public framework::OpKernel<T> {
           T, 128, 8, 8,
           true><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
           output, table, ids, N, K, D, padding_idx);
+#endif  // PADDLE_WITH_HIP
   }
 };
 
@@ -185,10 +201,20 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
       auto t = framework::EigenVector<T>::Flatten(*d_table_t);
       t.device(*dev_ctx.eigen_device()) = t.constant(static_cast<T>(0));
 
+#ifdef PADDLE_WITH_HIP
+      dim3 threads(64, 4);
+#else
       dim3 threads(128, 8);
+#endif  // PADDLE_WITH_HIP
       dim3 grids(8, 1);
+
+#ifdef PADDLE_WITH_HIP
+      LookupTableGrad<T, 64, 4, 8><<<grids, threads, 0, dev_ctx.stream()>>>(
+          d_table, d_output, ids, N, K, D);
+#else
       LookupTableGrad<T, 128, 8, 8><<<grids, threads, 0, dev_ctx.stream()>>>(
           d_table, d_output, ids, N, K, D);
+#endif  // PADDLE_WITH_HIP
     }
   }
 };
diff --git a/python/paddle/fluid/tests/unittests/test_cholesky_op.py b/python/paddle/fluid/tests/unittests/test_cholesky_op.py
index 93f62b20f2997447e453988709a0f26fe2e47181..633aa2cd613b690ccbc6f2a87e904d61b99f548a 100644
--- a/python/paddle/fluid/tests/unittests/test_cholesky_op.py
+++ b/python/paddle/fluid/tests/unittests/test_cholesky_op.py
@@ -58,7 +58,7 @@ class TestCholeskyOp(OpTest):
 
     def test_check_grad(self):
         places = [fluid.CPUPlace()]
-        if core.is_compiled_with_cuda():
+        if core.is_compiled_with_cuda() and (not core.is_compiled_with_rocm()):
             places.append(fluid.CUDAPlace(0))
         for p in places:
             self.func(p)
@@ -92,7 +92,10 @@ class TestCholeskyOp2D(TestCholeskyOp):
 
 class TestDygraph(unittest.TestCase):
     def test_dygraph(self):
-        paddle.disable_static()
+        if core.is_compiled_with_rocm():
+            paddle.disable_static(place=fluid.CPUPlace())
+        else:
+            paddle.disable_static()
         a = np.random.rand(3, 3)
         a_t = np.transpose(a, [1, 0])
         x_data = np.matmul(a, a_t) + 1e-03
@@ -103,7 +106,7 @@ class TestDygraph(unittest.TestCase):
 class TestCholeskySingularAPI(unittest.TestCase):
     def setUp(self):
         self.places = [fluid.CPUPlace()]
-        if core.is_compiled_with_cuda():
+        if core.is_compiled_with_cuda() and (not core.is_compiled_with_rocm()):
             self.places.append(fluid.CUDAPlace(0))
 
     def check_static_result(self, place, with_out=False):
diff --git a/python/paddle/fluid/tests/unittests/test_compare_op.py b/python/paddle/fluid/tests/unittests/test_compare_op.py
index 63a43432b4e555542f395a69819c385be0c090e4..fbf7384b86bc1c844dee09c5b439a523026044e5 100644
--- a/python/paddle/fluid/tests/unittests/test_compare_op.py
+++ b/python/paddle/fluid/tests/unittests/test_compare_op.py
@@ -61,6 +61,9 @@ def create_test_class(op_type, typename, callback):
 
 
 for _type_name in {'float32', 'float64', 'int32', 'int64'}:
+    if _type_name == 'float64' and core.is_compiled_with_rocm():
+        _type_name = 'float32'
+
     create_test_class('less_than', _type_name, lambda _a, _b: _a < _b)
     create_test_class('less_equal', _type_name, lambda _a, _b: _a <= _b)
     create_test_class('greater_than', _type_name, lambda _a, _b: _a > _b)