[Eager] Adapt faster tokenizer op (#42718)

* adapt faster tokenizer op

* add eager test

* add unittest

paddle/fluid/eager/eager_tensor.h

@@ -21,24 +21,176 @@
 #include "paddle/phi/api/include/tensor.h"
 #include "paddle/phi/api/lib/utils/tensor_utils.h"
 #include "paddle/phi/core/compat/convert_utils.h"
+
+namespace egr {
+
+/**
+ * VariableCompatTensor class is used by Eager mode for now. It's painful to
+ * do this in Eager Mode, the better choice is to design the special Tensor
+ * directly in phi and use it in paddle::experimental::Tensor.
+ * However, we have some special operators, and they use special input variable
+ * type, such as vector<string>, unordered_map<wstring, int>, these type cannot
+ * cover by DenseTensor or SparseTensor. So, we have to provide a compatible
+ * Tensor type like variable to support these special input type. We should
+ * remove this as soon as we finish the ResourceTensor in phi.
+ *
+ * Note: Keep this class as clean as possible.
+ * This class should only support method declared in framework::Variable and
+ * necessary overridden methods.
+ *
+ * Note: This class is only used to support types that cannot be supported by
+ * the phi Tensor system temporarily. You CANNOT use this class to handle types
+ * such as DenseTensor, SelectedRows, etc.
+ **/
+class VariableCompatTensor
+    : public phi::TensorBase,
+      public phi::TypeInfoTraits<phi::TensorBase, VariableCompatTensor> {
+ public:
+  template <typename T>
+  const T& Get() const {
+    static_assert(
+        paddle::framework::IsRegisteredVarType<T>(),
+        "Not registered type. Please register T inside var_type_traits.h");
+    PADDLE_ENFORCE_NOT_NULL(holder_, paddle::platform::errors::NotFound(
+                                         "Variable is not initialized."));
+    PADDLE_ENFORCE_EQ(
+        holder_->Type(), paddle::framework::VarTypeTrait<T>::kId,
+        paddle::platform::errors::InvalidArgument(
+            "The Variable type must be %s, but the type it holds is %s.",
+            paddle::framework::ToTypeName(
+                paddle::framework::VarTypeTrait<T>::kId),
+            paddle::framework::ToTypeName(holder_->Type())));
+    return *static_cast<const T*>(holder_->Ptr());
+  }
+
+  bool IsInitialized() const { return holder_ != nullptr; }
+
+  template <typename T>
+  T* GetMutable() {
+    if (!holder_) {
+      holder_.reset(new PlaceholderImpl<T>());
+    } else {
+      PADDLE_ENFORCE_EQ(
+          holder_->Type(), paddle::framework::VarTypeTrait<T>::kId,
+          paddle::platform::errors::InvalidArgument(
+              "The Variable type must be %s, but the type it holds is %s.",
+              paddle::framework::ToTypeName(
+                  paddle::framework::VarTypeTrait<T>::kId),
+              paddle::framework::ToTypeName(holder_->Type())));
+    }
+    return static_cast<T*>(holder_->Ptr());
+  }
+
+  template <typename T>
+  bool IsType() const {
+    return holder_ &&
+           holder_->Type() == paddle::framework::VarTypeTrait<T>::kId;
+  }
+
+  void Clear() { holder_.reset(); }
+
+  int Type() const {
+    PADDLE_ENFORCE_NOT_NULL(holder_, paddle::platform::errors::NotFound(
+                                         "Variable is not initialized."));
+    return holder_->Type();
+  }
+
+  // necessary overridden methods
+
+  static const char* name() { return "VariableCompatTensor"; }
+
+  ~VariableCompatTensor() override = default;
+
+  int64_t numel() const override {
+    PADDLE_THROW(paddle::platform::errors::Unavailable(
+        "VariableCompatTensor does not support `numel` method."));
+  }
+
+  const phi::DDim& dims() const override {
+    PADDLE_THROW(paddle::platform::errors::Unavailable(
+        "VariableCompatTensor does not support `dims` method."));
+  }
+
+  phi::DataType dtype() const override {
+    PADDLE_THROW(paddle::platform::errors::Unavailable(
+        "VariableCompatTensor does not support `dtype` method."));
+  }
+
+  phi::DataLayout layout() const override {
+    PADDLE_THROW(paddle::platform::errors::Unavailable(
+        "VariableCompatTensor does not support `layout` method."));
+  }
+
+  const phi::Place& place() const override {
+    PADDLE_THROW(paddle::platform::errors::Unavailable(
+        "VariableCompatTensor does not support `place` method."));
+  }
+
+  bool valid() const override { return IsInitialized(); }
+
+  bool initialized() const override { return IsInitialized(); }
+
+  void* AllocateFrom(phi::Allocator* allocator, phi::DataType dtype,
+                     size_t requested_size = 0) override {
+    PADDLE_THROW(paddle::platform::errors::Unavailable(
+        "VariableCompatTensor does not support `AllocateFrom` method."));
+  }
+
+ private:
+  struct Placeholder {
+    virtual ~Placeholder() PADDLE_MAY_THROW {}
+
+    inline int Type() const { return type_; }
+    inline const void* Ptr() const { return ptr_; }
+    inline void* Ptr() { return ptr_; }
+
+   protected:
+    inline void Init(void* p, int type) {
+      ptr_ = p;
+      type_ = type;
+    }
+
+    void* ptr_;
+    int type_;
+  };
+
+  // Placeholder hides type T, so it doesn't appear as a template
+  // parameter of Variable.
+  template <typename T>
+  struct PlaceholderImpl : public Placeholder {
+    static_assert(
+        paddle::framework::IsRegisteredVarType<T>(),
+        "Not registered type. Please register T inside var_type_traits.h");
+    PlaceholderImpl() {
+      this->Init(&obj_, paddle::framework::VarTypeTrait<T>::kId);
+    }
+
+   private:
+    T obj_;
+  };
+
+  // pointers to a PlaceholderImpl object indeed.
+  std::shared_ptr<Placeholder> holder_;
+};
+
+inline bool IsVariableCompatTensor(const paddle::experimental::Tensor& tensor) {
+  return VariableCompatTensor::classof(tensor.impl().get());
+}
+
 /**
  * This class is used by Eager mode for now. It's painful to do this in Eager
- * Mode, the better
- * choice is to use paddle::experimental::Tensor directly. However, we have a
- * punch of nested kernel code, and
- * they use paddle::framework::Variable in inner logic code. So, we have to
- * provide variable in
- * paddle::framework::ExecutionContext to support it. We should remove this as
- * soon as we finish our latest
- * Phi Lib, and use paddle::experimental::Tensor instead.
+ * Mode, the better choice is to use paddle::experimental::Tensor directly.
+ * However, we have a punch of nested kernel code, and they use
+ * paddle::framework::Variable in inner logic code. So, we have to provide
+ * variable in paddle::framework::ExecutionContext to support it. We should
+ * remove this as soon as we finish our latest Phi Lib, and use
+ * paddle::experimental::Tensor instead.
  *
  * Note: Keep this class as clean as possible.
  * This class should only support method declared in
  * paddle::experimental::Tensor with access method of
  * paddle::framework::Variable no more members are acceptable.
  * **/
-
-namespace egr {
 class EagerVariable final {
  public:
   /* Default constructor and name constructor should only be used for contruct
@@ -54,6 +206,14 @@ class EagerVariable final {
         ConstructVariableFromTensor<phi::DenseTensor>(tensor);
       } else if (tensor.is_selected_rows()) {
         ConstructVariableFromTensor<phi::SelectedRows>(tensor);
+      } else if (IsVariableCompatTensor(tensor) &&
+                 static_cast<const VariableCompatTensor*>(tensor.impl().get())
+                     ->IsType<paddle::framework::Vocab>()) {
+        ConstructVariableFromCompatTensor<paddle::framework::Vocab>(tensor);
+      } else if (IsVariableCompatTensor(tensor) &&
+                 static_cast<const VariableCompatTensor*>(tensor.impl().get())
+                     ->IsType<paddle::framework::Strings>()) {
+        ConstructVariableFromCompatTensor<paddle::framework::Strings>(tensor);
       } else {
         PADDLE_THROW(paddle::platform::errors::Fatal(
             "Unrecognized egr::EagerVariable type, only "
@@ -119,6 +279,22 @@ class EagerVariable final {
     *framework_tensor = *tensor_dense;
   }
 
+  template <typename VarType>
+  void ConstructVariableFromCompatTensor(
+      const paddle::experimental::Tensor& tensor) {
+    auto* framework_holder = var_.GetMutable<VarType>();
+    // Contruct framework::Tensor from egr::EagerVariable
+    auto* compat_tensor =
+        static_cast<VariableCompatTensor*>(tensor.impl().get());
+    PADDLE_ENFORCE_NOT_NULL(compat_tensor,
+                            paddle::platform::errors::Fatal(
+                                "Tensor %s holds empty impl, this should not "
+                                "happend since we should "
+                                "treat all kinds of tensor as what they are.",
+                                tensor.name()));
+    *framework_holder = compat_tensor->Get<VarType>();
+  }
+
  private:
   std::string name_{""};
   paddle::framework::Variable var_;

paddle/fluid/eager/tests/data_structure_tests/eager_tensor_test.cc

@@ -233,3 +233,88 @@ TEST(EagerVariable, DataLayout) {
   layout = paddle::imperative::GetDataLayout(eager_var);
   CHECK_EQ(layout, paddle::experimental::DataLayout::NCHW);
 }
+
+TEST(VariableCompatTensor, MemberFunction) {
+  egr::VariableCompatTensor var_tensor;
+  // test GetMutable and Get
+  var_tensor.GetMutable<paddle::framework::Vocab>();
+  auto& vocab = var_tensor.Get<paddle::framework::Vocab>();
+  EXPECT_EQ(vocab.size(), 0UL);
+  bool caught_exception = false;
+  try {
+    var_tensor.GetMutable<paddle::framework::Strings>();
+  } catch (paddle::platform::EnforceNotMet& error) {
+    caught_exception = true;
+    std::string ex_msg = error.what();
+    EXPECT_TRUE(ex_msg.find("The Variable type must be") != std::string::npos);
+  }
+  EXPECT_TRUE(caught_exception);
+  // test Type and IsType
+  EXPECT_TRUE(var_tensor.IsType<paddle::framework::Vocab>());
+  EXPECT_EQ(var_tensor.Type(),
+            static_cast<int>(paddle::framework::proto::VarType::VOCAB));
+  // test valid and initialized
+  EXPECT_TRUE(var_tensor.IsInitialized());
+  EXPECT_TRUE(var_tensor.valid());
+  EXPECT_TRUE(var_tensor.initialized());
+  // test name
+  EXPECT_EQ(var_tensor.name(), "VariableCompatTensor");
+  // test other throw error methods
+  caught_exception = false;
+  try {
+    var_tensor.numel();
+  } catch (paddle::platform::EnforceNotMet& error) {
+    caught_exception = true;
+    std::string ex_msg = error.what();
+    EXPECT_TRUE(ex_msg.find("numel") != std::string::npos);
+  }
+  EXPECT_TRUE(caught_exception);
+  caught_exception = false;
+  try {
+    var_tensor.dims();
+  } catch (paddle::platform::EnforceNotMet& error) {
+    caught_exception = true;
+    std::string ex_msg = error.what();
+    EXPECT_TRUE(ex_msg.find("dims") != std::string::npos);
+  }
+  EXPECT_TRUE(caught_exception);
+  caught_exception = false;
+  try {
+    var_tensor.dtype();
+  } catch (paddle::platform::EnforceNotMet& error) {
+    caught_exception = true;
+    std::string ex_msg = error.what();
+    EXPECT_TRUE(ex_msg.find("dtype") != std::string::npos);
+  }
+  EXPECT_TRUE(caught_exception);
+  caught_exception = false;
+  try {
+    var_tensor.layout();
+  } catch (paddle::platform::EnforceNotMet& error) {
+    caught_exception = true;
+    std::string ex_msg = error.what();
+    EXPECT_TRUE(ex_msg.find("layout") != std::string::npos);
+  }
+  EXPECT_TRUE(caught_exception);
+  caught_exception = false;
+  try {
+    var_tensor.place();
+  } catch (paddle::platform::EnforceNotMet& error) {
+    caught_exception = true;
+    std::string ex_msg = error.what();
+    EXPECT_TRUE(ex_msg.find("place") != std::string::npos);
+  }
+  EXPECT_TRUE(caught_exception);
+  caught_exception = false;
+  try {
+    var_tensor.AllocateFrom(nullptr, phi::DataType::UNDEFINED);
+  } catch (paddle::platform::EnforceNotMet& error) {
+    caught_exception = true;
+    std::string ex_msg = error.what();
+    EXPECT_TRUE(ex_msg.find("AllocateFrom") != std::string::npos);
+  }
+  EXPECT_TRUE(caught_exception);
+  // test Clear
+  var_tensor.Clear();
+  EXPECT_FALSE(var_tensor.IsInitialized());
+}

paddle/fluid/pybind/eager_method.cc

@@ -18,6 +18,7 @@ typedef SSIZE_T ssize_t;
 #include <Python.h>
 
 #include <string>
+#include <unordered_map>
 #include <vector>
 
 #include "pybind11/numpy.h"
@@ -675,7 +676,9 @@ static PyObject* tensor_method_get_underline_tensor(TensorObject* self,
                                                     PyObject* kwargs) {
   EAGER_TRY
   if (!self->tensor.defined()) {
-    RETURN_PY_NONE
+    // The original `get_tensor` method of Variable will create a empty tensor
+    phi::DenseTensor empty_tensor;
+    return ToPyObject(&empty_tensor);
   }
   if (self->tensor.is_dense_tensor()) {
     auto* tensor =
@@ -1275,6 +1278,47 @@ static PyObject* tensor__copy_gradient_from(TensorObject* self, PyObject* args,
 
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
+
+static PyObject* tensor_method_set_vocab(TensorObject* self, PyObject* args,
+                                         PyObject* kwargs) {
+  EAGER_TRY
+  using Vocab = std::unordered_map<std::wstring, int>;
+  auto vocab = CastPyArg2Vocab(PyTuple_GET_ITEM(args, 0), 0);
+  auto var_tensor = std::make_shared<egr::VariableCompatTensor>();
+  *var_tensor->GetMutable<Vocab>() = vocab;
+  self->tensor.set_impl(var_tensor);
+  RETURN_PY_NONE
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
+static PyObject* tensor_method_set_string_list(TensorObject* self,
+                                               PyObject* args,
+                                               PyObject* kwargs) {
+  EAGER_TRY
+  using Strings = std::vector<std::string>;
+  auto strings = CastPyArg2Strings(PyTuple_GET_ITEM(args, 0), 0);
+  auto var_tensor = std::make_shared<egr::VariableCompatTensor>();
+  *var_tensor->GetMutable<Strings>() = strings;
+  self->tensor.set_impl(var_tensor);
+  RETURN_PY_NONE
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
+static PyObject* tensor_method_get_map_tensor(TensorObject* self,
+                                              PyObject* args,
+                                              PyObject* kwargs) {
+  EAGER_TRY
+  PADDLE_ENFORCE_EQ(
+      egr::IsVariableCompatTensor(self->tensor), true,
+      paddle::platform::errors::Fatal(
+          "this method is only effective for VariableCompatTensor"));
+  using Vocab = std::unordered_map<std::wstring, int>;
+  auto* var_tensor =
+      static_cast<const egr::VariableCompatTensor*>(self->tensor.impl().get());
+  return ToPyObject(var_tensor->Get<Vocab>());
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
 static PyObject* tensor_method_get_non_zero_indices(TensorObject* self,
                                                     PyObject* args,
                                                     PyObject* kwargs) {
@@ -1655,6 +1699,15 @@ PyMethodDef variable_methods[] = {
     {"_copy_gradient_from",
      (PyCFunction)(void (*)(void))tensor__copy_gradient_from,
      METH_VARARGS | METH_KEYWORDS, NULL},
+    /** the methods to adapt old dygraph, will be removed in the future **/
+    {"set_string_list",
+     (PyCFunction)(void (*)(void))tensor_method_set_string_list,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"set_vocab", (PyCFunction)(void (*)(void))tensor_method_set_vocab,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"get_map_tensor",
+     (PyCFunction)(void (*)(void))tensor_method_get_map_tensor,
+     METH_VARARGS | METH_KEYWORDS, NULL},
     /***the method of sparse tensor****/
     {"indices", (PyCFunction)(void (*)(void))tensor_method_get_non_zero_indices,
      METH_VARARGS | METH_KEYWORDS, NULL},

paddle/fluid/pybind/eager_properties.cc

@@ -58,6 +58,10 @@ PyObject* tensor_properties_get_type(TensorObject* self, void* closure) {
     return ToPyObject(paddle::framework::proto::VarType::LOD_TENSOR);
   } else if (self->tensor.is_selected_rows()) {
     return ToPyObject(paddle::framework::proto::VarType::SELECTED_ROWS);
+  } else if (egr::IsVariableCompatTensor(self->tensor)) {
+    return ToPyObject(static_cast<paddle::framework::proto::VarType::Type>(
+        static_cast<const egr::VariableCompatTensor*>(self->tensor.impl().get())
+            ->Type()));
   } else {
     RETURN_PY_NONE
   }
@@ -152,12 +156,28 @@ PyObject* tensor_properties_get_shape(TensorObject* self, void* closure) {
   if (!self->tensor.defined()) {
     return ToPyObject(value);
   }
+  if (egr::IsVariableCompatTensor(self->tensor)) {
+    auto* var_tensor = static_cast<const egr::VariableCompatTensor*>(
+        self->tensor.impl().get());
+    if (var_tensor->IsType<paddle::framework::Vocab>()) {
+      value.emplace_back(static_cast<int64_t>(
+          var_tensor->Get<paddle::framework::Vocab>().size()));
+    } else if (var_tensor->IsType<paddle::framework::Strings>()) {
+      value.emplace_back(static_cast<int64_t>(
+          var_tensor->Get<paddle::framework::Strings>().size()));
+    } else {
+      PADDLE_THROW(paddle::platform::errors::Unavailable(
+          "VariableCompatTensor only support get shape from Vocab or "
+          "Strings."));
+    }
+  } else {
     auto ddim = self->tensor.shape();
     size_t rank = static_cast<size_t>(ddim.size());
     value.resize(rank);
     for (size_t i = 0; i < rank; i++) {
       value[i] = ddim[i];
     }
+  }
 
   return ToPyObject(value);
   EAGER_CATCH_AND_THROW_RETURN_NULL
@@ -183,8 +203,22 @@ PyObject* tensor_properties_get_dtype(TensorObject* self, void* closure) {
     // be same to old dygraph
     return ToPyObject(framework::proto::VarType::FP32);
   }
+  if (egr::IsVariableCompatTensor(self->tensor)) {
+    auto* var_tensor = static_cast<const egr::VariableCompatTensor*>(
+        self->tensor.impl().get());
+    if (var_tensor->IsType<paddle::framework::Vocab>()) {
+      return ToPyObject(framework::proto::VarType::RAW);
+    } else if (var_tensor->IsType<paddle::framework::Strings>()) {
+      return ToPyObject(framework::proto::VarType::STRING);
+    } else {
+      PADDLE_THROW(paddle::platform::errors::Unavailable(
+          "VariableCompatTensor only support get shape from Vocab or "
+          "Strings."));
+    }
+  } else {
     return ToPyObject(
         paddle::framework::TransToProtoVarType(self->tensor.type()));
+  }
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 

paddle/fluid/pybind/eager_utils.cc

@@ -472,6 +472,28 @@ paddle::framework::proto::VarType::Type CastPyArg2ProtoType(PyObject* obj,
   return dtype;
 }
 
+std::unordered_map<std::wstring, int> CastPyArg2Vocab(PyObject* obj,
+                                                      ssize_t arg_pos) {
+  if (PyDict_Check(obj)) {
+    return ::pybind11::handle(obj)
+        .cast<std::unordered_map<std::wstring, int>>();
+  } else {
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "argument (position %d) must be dict, but got %s", arg_pos + 1,
+        reinterpret_cast<PyTypeObject*>(obj->ob_type)->tp_name));
+  }
+}
+
+std::vector<std::string> CastPyArg2Strings(PyObject* obj, ssize_t arg_pos) {
+  if (PyList_Check(obj)) {
+    return ::pybind11::handle(obj).cast<std::vector<std::string>>();
+  } else {
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "argument (position %d) must be list, but got %s", arg_pos + 1,
+        reinterpret_cast<PyTypeObject*>(obj->ob_type)->tp_name));
+  }
+}
+
 paddle::CustomOpKernelContext CastPyArg2CustomOpKernelContext(PyObject* obj,
                                                               ssize_t arg_pos) {
   if (PyObject_IsInstance(
@@ -719,6 +741,28 @@ PyObject* ToPyObject(
   return dict;
 }
 
+PyObject* ToPyObject(const std::unordered_map<std::wstring, int>& value) {
+  PyObject* dict = PyDict_New();
+  for (const auto map_iter : value) {
+    // Convert Key
+    PyObject* key_string =
+        PyUnicode_FromWideChar(map_iter.first.c_str(), map_iter.first.size());
+    if (!key_string) {
+      PADDLE_THROW(platform::errors::Fatal(
+          "Unable to convert std::wstring to PyObject"));
+    }
+
+    // Convert Val
+    PyObject* py_int = PyLong_FromLong(map_iter.second);
+
+    if (PyDict_SetItem(dict, key_string, py_int) != 0) {
+      PADDLE_THROW(
+          platform::errors::Fatal("Unable to set key:value for py_dict"));
+    }
+  }
+  return dict;
+}
+
 // For Final State Dygraph,
 // We directly use paddle::optional(Tensor) as dispensable Tensor
 paddle::optional<const paddle::experimental::Tensor&> GetOptionalTensorFromArgs(

paddle/fluid/pybind/eager_utils.h

@@ -65,6 +65,9 @@ std::vector<std::vector<size_t>> CastPyArg2VectorOfVectorOfSize_t(
     PyObject* obj, size_t arg_pos);
 framework::proto::VarType::Type CastPyArg2ProtoType(PyObject* obj,
                                                     ssize_t arg_pos);
+std::unordered_map<std::wstring, int> CastPyArg2Vocab(PyObject* obj,
+                                                      ssize_t arg_pos);
+std::vector<std::string> CastPyArg2Strings(PyObject* obj, ssize_t arg_pos);
 
 PyObject* ToPyObject(int value);
 PyObject* ToPyObject(uint32_t value);
@@ -96,6 +99,7 @@ PyObject* ToPyObject(const paddle::framework::proto::VarType& type);
 PyObject* ToPyObject(const void* value);
 PyObject* ToPyObject(
     const std::unordered_map<std::string, std::vector<std::string>>& value);
+PyObject* ToPyObject(const std::unordered_map<std::wstring, int>& value);
 
 template <typename Tuple, size_t N>
 struct TupleTensorResult {

paddle/phi/api/lib/tensor.cc

@@ -394,8 +394,8 @@ uint32_t Tensor::current_inplace_version() {
         static_cast<phi::DenseTensor *>(impl_.get())->InplaceVersionCounter();
     return inplace_version_counter.CurrentVersion();
   } else {
-    PADDLE_THROW(phi::errors::Unimplemented(
-        "current_inplace_version is only supported on DenseTensor now."));
+    LOG_FIRST_N(WARNING, 1)
+        << "current_inplace_version is only supported on DenseTensor now.";
   }
   return 0;
 }

python/paddle/fluid/tests/unittests/test_faster_tokenizer_op.py

@@ -22,8 +22,7 @@ import numpy as np
 import paddle
 import paddle.nn as nn
 from paddle.dataset.common import DATA_HOME
-from paddle.fluid.framework import core, _non_static_mode, _enable_legacy_dygraph
-_enable_legacy_dygraph()
+from paddle.fluid.framework import core, _non_static_mode, _test_eager_guard
 from paddle.fluid.layer_helper import LayerHelper
 from paddle import _C_ops
 
@@ -151,13 +150,12 @@ class Predictor(object):
 class TestBertTokenizerOp(unittest.TestCase):
     def setUp(self):
         self.bert_tokenizer = BertTokenizer.from_pretrained("bert-base-chinese")
-        self.faster_tokenizer = FasterTokenizer(self.bert_tokenizer.vocab)
-        self.init_data()
         self.save_path = os.path.join(DATA_HOME, "fast_tokenizer")
         self.param_path = os.path.join(self.save_path, "model.pdparams")
         self.inference_path = os.path.join(self.save_path, "inference")
 
     def init_data(self):
+        self.faster_tokenizer = FasterTokenizer(self.bert_tokenizer.vocab)
         self.text = [
             '选择珠江花园的原因就是方便，有电动扶梯直接到达海边，周围餐馆、食廊、商场、超市、摊位一应俱全。'
             '酒店装修一般，但还算整洁。 泳池在大堂的屋顶，因此很小，不过女儿倒是喜欢。 包的早餐是西式的，'
@@ -179,8 +177,8 @@ class TestBertTokenizerOp(unittest.TestCase):
         self.texts_tensor = to_string_tensor(self.texts, "texts")
         self.text_pairs_tensor = to_string_tensor(self.text_pairs, "text_pairs")
 
-    def test_padding(self):
-
+    def run_padding(self):
+        self.init_data()
         self.max_seq_len = 128
         self.pad_to_max_seq_len = True
         self.is_split_into_words = False
@@ -283,7 +281,13 @@ class TestBertTokenizerOp(unittest.TestCase):
             np.allclose(
                 token_type_ids, py_token_type_ids, rtol=0, atol=0.01))
 
-    def test_no_padding(self):
+    def test_padding(self):
+        with _test_eager_guard():
+            self.run_padding()
+        self.run_padding()
+
+    def run_no_padding(self):
+        self.init_data()
         self.max_seq_len = 128
         self.pad_to_max_seq_len = False
         self.is_split_into_words = False
@@ -336,7 +340,13 @@ class TestBertTokenizerOp(unittest.TestCase):
             np.allclose(
                 token_type_ids, py_token_type_ids, rtol=0, atol=0.01))
 
-    def test_is_split_into_words(self):
+    def test_no_padding(self):
+        with _test_eager_guard():
+            self.run_no_padding()
+        self.run_no_padding()
+
+    def run_is_split_into_words(self):
+        self.init_data()
         self.is_split_into_words = True
 
         input_ids, token_type_ids = self.faster_tokenizer(
@@ -355,7 +365,13 @@ class TestBertTokenizerOp(unittest.TestCase):
             np.allclose(
                 token_type_ids, py_token_type_ids, rtol=0, atol=0.01))
 
+    def test_is_split_into_words(self):
+        with _test_eager_guard():
+            self.run_is_split_into_words()
+        self.run_is_split_into_words()
+
     def test_inference(self):
+        self.init_data()
         if not os.path.exists(self.save_path):
             os.makedirs(self.save_path, exist_ok=True)
         paddle.save(self.faster_tokenizer.state_dict(), self.param_path)
@@ -383,6 +399,7 @@ class TestBertTokenizerOp(unittest.TestCase):
                 token_type_ids, py_token_type_ids, rtol=0, atol=0.01))
 
     def test_feed_string_var(self):
+        self.init_data()
         paddle.enable_static()
         x = paddle.static.data(
             name="x", shape=[-1], dtype=core.VarDesc.VarType.STRINGS)