diff --git a/paddle/fluid/eager/to_static/run_program_op_node.h b/paddle/fluid/eager/to_static/run_program_op_node.h
index 8d2dd38113c94cb7c26231bc63ed58aec304d544..781d6616e38ae305125c5da380800bd5db105fe2 100644
--- a/paddle/fluid/eager/to_static/run_program_op_node.h
+++ b/paddle/fluid/eager/to_static/run_program_op_node.h
@@ -382,7 +382,9 @@ class GradNodeRunProgram : public egr::GradNodeBase {
       x_grad_ptr.emplace_back(&i);
     }
     for (auto &i : params_grad) {
-      params_grad_ptr.emplace_back(&i);
+      if (i.defined()) {
+        params_grad_ptr.emplace_back(&i);
+      }
     }
 
     PADDLE_ENFORCE_EQ(hooked_grads[0].size(), fwd_out_names_.size(),
diff --git a/python/paddle/fluid/dygraph/io.py b/python/paddle/fluid/dygraph/io.py
index cf3602a7898ca0e8607ac37a0bc1f630d23afeeb..f10b65222021457d1e892430b69f42f32e0107bc 100644
--- a/python/paddle/fluid/dygraph/io.py
+++ b/python/paddle/fluid/dygraph/io.py
@@ -883,7 +883,7 @@ def _run_dygraph(instance, input, program_holder):
     # transform SelectedRows to LoDTensor forcibly, it may not
     # be user wanted result.
     for persistable_var in persistable_vars:
-        grad_var_name = var.name + core.grad_var_suffix()
+        grad_var_name = persistable_var.name + core.grad_var_suffix()
         grad_var = trace_program.block(0).find_var(cpt.to_bytes(grad_var_name))
         # NOTE: cannot find var desc maybe not problem, 
         # such as in batch_norm
diff --git a/python/paddle/fluid/dygraph/jit.py b/python/paddle/fluid/dygraph/jit.py
index 31af26cb13713618652016de736a0aef4456c85e..7957b33bf1dcee18b80fbd9df8d05324688fba1c 100644
--- a/python/paddle/fluid/dygraph/jit.py
+++ b/python/paddle/fluid/dygraph/jit.py
@@ -37,7 +37,7 @@ from paddle.fluid.dygraph.dygraph_to_static.program_translator import ProgramTra
 from paddle.fluid.dygraph.io import TranslatedLayer, INFER_MODEL_SUFFIX, INFER_PARAMS_SUFFIX, INFER_PARAMS_INFO_SUFFIX
 from paddle.fluid.dygraph.layers import Layer
 from paddle.fluid.executor import Executor, scope_guard
-from paddle.fluid.framework import Block, ParamBase, Program, Variable, Parameter
+from paddle.fluid.framework import Block, ParamBase, Program, Variable, Parameter, EagerParamBase
 from paddle.fluid.framework import _current_expected_place, _dygraph_guard, _dygraph_tracer
 from paddle.fluid.framework import dygraph_only, _non_static_mode
 from paddle.fluid.wrapped_decorator import wrap_decorator
@@ -921,7 +921,8 @@ def save(layer, path, input_spec=None, **configs):
                                     param_or_buffer.name]
                         extra_info_dict[
                             'stop_gradient'] = param_or_buffer.stop_gradient
-                        if isinstance(param_or_buffer, ParamBase):
+                        if isinstance(param_or_buffer,
+                                      (ParamBase, EagerParamBase)):
                             extra_info_dict[
                                 'trainable'] = param_or_buffer.trainable
                         extra_var_info[param_or_buffer.name] = extra_info_dict
diff --git a/python/paddle/fluid/dygraph/tracer.py b/python/paddle/fluid/dygraph/tracer.py
index 4375bae6787139acf4dd074135e17e410d4c3826..e1fabf9aeda1004337caf9dcacc94c2264dd4919 100644
--- a/python/paddle/fluid/dygraph/tracer.py
+++ b/python/paddle/fluid/dygraph/tracer.py
@@ -147,6 +147,13 @@ class Tracer(core.Tracer):
             attrs_list.append(v)
         returns = function_ptr(*arg_list, *attrs_list)
 
+        if type == 'load_combine':
+            assert len(outputs.keys()) == 1
+            key = list(outputs.keys())[0]
+            for j in range(len(returns)):
+                returns[j]._share_underline_tensor_to(outputs[key][j])
+            return
+
         if isinstance(returns, tuple):
             for i in range(len(op_returns)):
                 retname = op_returns[i]