diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index 833a28a757..834cdb422a 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -47,9 +47,6 @@ DECLARE_bool(benchmark);
 DECLARE_bool(check_nan_inf);
 DECLARE_bool(enable_unused_var_check);
 DEFINE_int32(inner_op_parallelism, 0, "number of threads for inner op");
-DEFINE_bool(fast_check_nan_inf, false,
-            "Fast checking NAN/INF after each operation. It will be a little"
-            "bit slow, much faster than check_nan_inf");
 
 namespace paddle {
 namespace framework {
@@ -1173,25 +1170,6 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
 #endif
   }
 
-  if (FLAGS_fast_check_nan_inf) {
-    for (auto& vname : OutputVars(true)) {
-      // only check inserted vars,
-      // please see executor.py for details of fast_check_nan_inf
-      if (vname.rfind("debug_var") == 0) {
-        VLOG(3) << "debugging nan/inf in var " << vname;
-
-        auto* var = exec_scope.FindVar(vname);
-        if (var == nullptr) continue;
-        if (var->IsType<framework::LoDTensor>()) {
-          CheckTensorNANOrInf(type_, vname, var->Get<framework::LoDTensor>());
-        } else if (var->IsType<framework::SelectedRows>()) {
-          CheckTensorNANOrInf(type_, vname,
-                              var->Get<framework::SelectedRows>().value());
-        }
-      }
-    }
-  }
-
   if (FLAGS_check_nan_inf) {
     framework::details::CheckOpHasNanOrInf(*this, exec_scope, place);
   }
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index 1a88d3512e..b24da29d0f 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -175,7 +175,6 @@ def __bootstrap__():
     sysstr = platform.system()
     read_env_flags = [
         'check_nan_inf',
-        'fast_check_nan_inf',
         'benchmark',
         'eager_delete_scope',
         'fraction_of_cpu_memory_to_use',
diff --git a/python/paddle/fluid/debugger.py b/python/paddle/fluid/debugger.py
index 9110b8daf3..75dc14a1d7 100644
--- a/python/paddle/fluid/debugger.py
+++ b/python/paddle/fluid/debugger.py
@@ -280,88 +280,3 @@ def draw_block_graphviz(block, highlights=None, path="./temp.dot"):
             add_op_link_var(opn, var, True)
 
     graph(path, show=False)
-
-
-def prepare_fast_nan_inf_debug(_program):
-    """
-    Given a program to run, insert a (reduce) sum op for every var in that program.
-    Instead of checking all vars originally defined in the program,
-    only those inserted ops will be checked in the c++ end, to detect if it contains NAN or INF.
-    Thereforce, the speed of nan/inf checking could be improved.
-    Please set ``FLAGS_fast_check_nan_inf" to open the fast nan/inf feature.
-    """
-
-    helper = LayerHelper('reduce_sum', **locals())
-
-    if _program is None:
-        _program = default_main_program()
-
-    for _block in _program.blocks:
-        # fetch vars in the current block
-        _vars_in_prog = []
-        for _var_name in _block.vars:
-            _vars_in_prog.append((_var_name, _block.vars[_var_name]))
-
-        # append sum_op in the current block
-        for _var_name, _var in _vars_in_prog:
-
-            try:
-
-                if _var.dtype == -1:
-                    continue
-
-                ## create a var for holding sum output
-                _output_var = _block.create_var(
-                    name=unique_name.generate("debug_var_" + _var_name),
-                    dtype=_var.dtype,
-                    type=core.VarDesc.VarType.LOD_TENSOR,
-                    persistable=False,
-                    stop_gradient=True)
-
-                ## create a sum op, input each existing var in the block
-                _block.append_op(
-                    type='sum',
-                    outputs={'Out': _output_var},
-                    inputs={'X': [_var]})
-            except Exception as e:
-                pass
-
-
-def run_fast_nan_inf_debug(executor,
-                           program=None,
-                           feed=None,
-                           fetch_list=None,
-                           feed_var_name='feed',
-                           fetch_var_name='fetch',
-                           scope=None,
-                           return_numpy=True,
-                           use_program_cache=False,
-                           dump_core=True):
-    """
-    Run a program by the given executor. Catch the exception of NAN and INF, and save persistables into the dumped core.
-    """
-
-    assert (executor is not None)
-
-    try:
-        output = executor.run(program=program,
-                              feed=feed,
-                              fetch_list=fetch_list,
-                              feed_var_name=feed_var_name,
-                              fetch_var_name=fetch_var_name,
-                              scope=scope,
-                              return_numpy=return_numpy,
-                              use_program_cache=use_program_cache)
-
-        return output
-
-    except Exception as e:
-
-        print("catch an exception:")
-        print(e)
-
-        core_filename = "core" + str(int(random.random() * 10000)) + ".pdckpt"
-        io.save_persistables(
-            executor, "./", main_program=program, filename=core_filename)
-
-        print("dumping a core into ./%s" % core_filename)