Remove unnecessary includings for pstring.h (#43752)

* Remove unnecessary including for pstring.h

* Fix typos

paddle/fluid/framework/convert_utils.cc

@@ -11,10 +11,12 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+
 #include "paddle/fluid/framework/convert_utils.h"
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/platform/device/gpu/gpu_info.h"
 
+#include "paddle/phi/common/pstring.h"
 namespace paddle {
 namespace framework {
 

paddle/fluid/pybind/tensor_py.h

@@ -38,6 +38,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/float16.h"
 #include "paddle/fluid/platform/profiler/event_tracing.h"
+#include "paddle/phi/common/pstring.h"
 #include "paddle/phi/core/string_tensor.h"
 #include "paddle/phi/kernels/strings/unicode.h"
 #include "pybind11/numpy.h"
@@ -180,14 +181,17 @@ template <typename T>
 class PYBIND11_HIDDEN NumpyAllocation : public memory::Allocation {
  public:
   explicit NumpyAllocation(const py::array &arr)
-      : Allocation(const_cast<void *>(arr.data()), sizeof(T) * (arr.size()),
+      : Allocation(const_cast<void *>(arr.data()),
+                   sizeof(T) * (arr.size()),
                    paddle::platform::CPUPlace()),
         arr_(arr.ptr()) {
-    PADDLE_ENFORCE_NOT_NULL(arr_, platform::errors::InvalidArgument(
-                                      "The underlying PyObject pointer of "
-                                      "numpy array cannot be nullptr"));
+    PADDLE_ENFORCE_NOT_NULL(
+        arr_,
+        platform::errors::InvalidArgument("The underlying PyObject pointer of "
+                                          "numpy array cannot be nullptr"));
     PADDLE_ENFORCE_NE(
-        arr_, Py_None,
+        arr_,
+        Py_None,
         platform::errors::PreconditionNotMet(
             "The underlying PyObject pointer of numpy array cannot be None"));
     Py_INCREF(arr_);
@@ -241,7 +245,8 @@ inline std::string TensorDTypeToPyDTypeStr(
     } else {                                                                \
       constexpr auto kIsValidDType = ValidDTypeToPyArrayChecker<T>::kValue; \
       PADDLE_ENFORCE_EQ(                                                    \
-          kIsValidDType, true,                                              \
+          kIsValidDType,                                                    \
+          true,                                                             \
           platform::errors::Unimplemented(                                  \
               "This type [%s] of tensor cannot be expose to Python",        \
               typeid(T).name()));                                           \
@@ -259,7 +264,8 @@ inline std::string TensorDTypeToPyDTypeStr(
 
 template <typename T>
 T TensorGetElement(const framework::Tensor &self, size_t offset) {
-  PADDLE_ENFORCE_LT(offset, self.numel(),
+  PADDLE_ENFORCE_LT(offset,
+                    self.numel(),
                     platform::errors::InvalidArgument(
                         "The offset exceeds the size of tensor."));
 
@@ -276,29 +282,29 @@ T TensorGetElement(const framework::Tensor &self, size_t offset) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
     const T *a = self.data<T>();
     auto p = self.place();
-    paddle::memory::Copy(platform::CPUPlace(), &b, p, a + offset, sizeof(T),
-                         nullptr);
+    paddle::memory::Copy(
+        platform::CPUPlace(), &b, p, a + offset, sizeof(T), nullptr);
 #endif
   } else if (platform::is_mlu_place(self.place())) {
 #ifdef PADDLE_WITH_MLU
     const T *a = self.data<T>();
     auto p = self.place();
-    paddle::memory::Copy(platform::CPUPlace(), &b, p, a + offset, sizeof(T),
-                         nullptr);
+    paddle::memory::Copy(
+        platform::CPUPlace(), &b, p, a + offset, sizeof(T), nullptr);
 #endif
   } else if (platform::is_npu_place(self.place())) {
 #if defined(PADDLE_WITH_ASCEND_CL)
     const T *a = self.data<T>();
     auto p = self.place();
-    paddle::memory::Copy(platform::CPUPlace(), &b, p, a + offset, sizeof(T),
-                         nullptr);
+    paddle::memory::Copy(
+        platform::CPUPlace(), &b, p, a + offset, sizeof(T), nullptr);
 #endif
   } else if (platform::is_custom_place(self.place())) {
 #if defined(PADDLE_WITH_CUSTOM_DEVICE)
     const T *a = self.data<T>();
     auto p = self.place();
-    paddle::memory::Copy(platform::CPUPlace(), &b, p, a + offset, sizeof(T),
-                         nullptr);
+    paddle::memory::Copy(
+        platform::CPUPlace(), &b, p, a + offset, sizeof(T), nullptr);
 #endif
   }
   VLOG(10) << "TensorGetElement, place: " << self.place()
@@ -308,7 +314,8 @@ T TensorGetElement(const framework::Tensor &self, size_t offset) {
 
 template <typename T>
 void TensorSetElement(framework::Tensor *self, size_t offset, T elem) {
-  PADDLE_ENFORCE_LT(offset, self->numel(),
+  PADDLE_ENFORCE_LT(offset,
+                    self->numel(),
                     platform::errors::InvalidArgument(
                         "The offset exceeds the size of tensor."));
   VLOG(10) << "TensorSetElement, place: " << self->place()
@@ -325,29 +332,29 @@ void TensorSetElement(framework::Tensor *self, size_t offset, T elem) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
     auto p = self->place();
     T *a = self->mutable_data<T>(p);
-    paddle::memory::Copy(p, a + offset, platform::CPUPlace(), &elem, sizeof(T),
-                         nullptr);
+    paddle::memory::Copy(
+        p, a + offset, platform::CPUPlace(), &elem, sizeof(T), nullptr);
 #endif
   } else if (platform::is_mlu_place(self->place())) {
 #ifdef PADDLE_WITH_MLU
     auto p = self->place();
     T *a = self->mutable_data<T>(p);
-    paddle::memory::Copy(p, a + offset, platform::CPUPlace(), &elem, sizeof(T),
-                         nullptr);
+    paddle::memory::Copy(
+        p, a + offset, platform::CPUPlace(), &elem, sizeof(T), nullptr);
 #endif
   } else if (platform::is_npu_place(self->place())) {
 #if defined(PADDLE_WITH_ASCEND_CL)
     auto p = self->place();
     T *a = self->mutable_data<T>(p);
-    paddle::memory::Copy(p, a + offset, platform::CPUPlace(), &elem, sizeof(T),
-                         nullptr);
+    paddle::memory::Copy(
+        p, a + offset, platform::CPUPlace(), &elem, sizeof(T), nullptr);
 #endif
   } else if (platform::is_custom_place(self->place())) {
 #if defined(PADDLE_WITH_CUSTOM_DEVICE)
     auto p = self->place();
     T *a = self->mutable_data<T>(p);
-    paddle::memory::Copy(p, a + offset, platform::CPUPlace(), &elem, sizeof(T),
-                         nullptr);
+    paddle::memory::Copy(
+        p, a + offset, platform::CPUPlace(), &elem, sizeof(T), nullptr);
 #endif
   }
 }
@@ -356,7 +363,8 @@ template <typename T, typename P>
 void SetTensorFromPyArrayT(
     framework::Tensor *self,
     const py::array_t<T, py::array::c_style | py::array::forcecast> &array,
-    const P &place, bool zero_copy) {
+    const P &place,
+    bool zero_copy) {
   std::vector<int64_t> dims;
   dims.reserve(array.ndim());
   for (decltype(array.ndim()) i = 0; i < array.ndim(); ++i) {
@@ -380,8 +388,11 @@ void SetTensorFromPyArrayT(
     platform::Place tmp_place = place;
     platform::XPUDeviceGuard guard(tmp_place.device);
     auto dst = self->mutable_data<T>(place);
-    memory::Copy(tmp_place, static_cast<void *>(dst), platform::CPUPlace(),
-                 static_cast<const void *>(array.data()), array.nbytes());
+    memory::Copy(tmp_place,
+                 static_cast<void *>(dst),
+                 platform::CPUPlace(),
+                 static_cast<const void *>(array.data()),
+                 array.nbytes());
 #else
     PADDLE_THROW(platform::errors::PermissionDenied(
         "Cannot use XPUPlace in CPU/GPU version, "
@@ -413,8 +424,8 @@ void SetTensorFromPyArrayT(
     platform::Place tmp_place = place;
     platform::NPUDeviceGuard guard(tmp_place.device);
     auto dst = self->mutable_data<T>(place);
-    platform::NPUMemcpySync(dst, array.data(), array.nbytes(),
-                            ACL_MEMCPY_HOST_TO_DEVICE);
+    platform::NPUMemcpySync(
+        dst, array.data(), array.nbytes(), ACL_MEMCPY_HOST_TO_DEVICE);
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &ctx = *pool.Get(place);
     ctx.Wait();
@@ -460,11 +471,11 @@ void SetTensorFromPyArrayT(
       platform::CUDADeviceGuard guard(place.device);
       auto dst = self->mutable_data<T>(place);
 #ifdef PADDLE_WITH_HIP
-      paddle::platform::GpuMemcpySync(dst, array.data(), array.nbytes(),
-                                      hipMemcpyHostToDevice);
+      paddle::platform::GpuMemcpySync(
+          dst, array.data(), array.nbytes(), hipMemcpyHostToDevice);
 #else
-      paddle::platform::GpuMemcpySync(dst, array.data(), array.nbytes(),
-                                      cudaMemcpyHostToDevice);
+      paddle::platform::GpuMemcpySync(
+          dst, array.data(), array.nbytes(), cudaMemcpyHostToDevice);
 #endif
 
     } else if (paddle::platform::is_cuda_pinned_place(place)) {
@@ -486,8 +497,10 @@ void SetTensorFromPyArrayT(
 }
 
 template <typename P>
-void SetTensorFromPyArray(framework::Tensor *self, const py::object &obj,
-                          const P &place, bool zero_copy) {
+void SetTensorFromPyArray(framework::Tensor *self,
+                          const py::object &obj,
+                          const P &place,
+                          bool zero_copy) {
   auto array = obj.cast<py::array>();
   if (py::isinstance<py::array_t<float>>(array)) {
     SetTensorFromPyArrayT<float, P>(self, array, place, zero_copy);
@@ -504,8 +517,8 @@ void SetTensorFromPyArray(framework::Tensor *self, const py::object &obj,
   } else if (py::isinstance<py::array_t<uint8_t>>(array)) {
     SetTensorFromPyArrayT<uint8_t, P>(self, array, place, zero_copy);
   } else if (py::isinstance<py::array_t<paddle::platform::float16>>(array)) {
-    SetTensorFromPyArrayT<paddle::platform::float16, P>(self, array, place,
-                                                        zero_copy);
+    SetTensorFromPyArrayT<paddle::platform::float16, P>(
+        self, array, place, zero_copy);
   } else if (py::isinstance<py::array_t<paddle::platform::complex<float>>>(
                  array)) {
     SetTensorFromPyArrayT<paddle::platform::complex<float>, P>(
@@ -517,8 +530,8 @@ void SetTensorFromPyArray(framework::Tensor *self, const py::object &obj,
   } else if (py::isinstance<py::array_t<uint16_t>>(array)) {
     // since there is still no support for bfloat16 in NumPy,
     // uint16 is used for casting bfloat16
-    SetTensorFromPyArrayT<paddle::platform::bfloat16, P>(self, array, place,
-                                                         zero_copy);
+    SetTensorFromPyArrayT<paddle::platform::bfloat16, P>(
+        self, array, place, zero_copy);
   } else if (py::isinstance<py::array_t<bool>>(array)) {
     SetTensorFromPyArrayT<bool, P>(self, array, place, zero_copy);
   } else {
@@ -533,11 +546,13 @@ void SetTensorFromPyArray(framework::Tensor *self, const py::object &obj,
 }
 
 template <typename P>
-void SetStringTensorFromPyArray(phi::StringTensor *self, const py::array &array,
+void SetStringTensorFromPyArray(phi::StringTensor *self,
+                                const py::array &array,
                                 const P &place) {
   bool is_string_pyarray =
       array.dtype().kind() == 'S' || array.dtype().kind() == 'U';
-  PADDLE_ENFORCE_EQ(is_string_pyarray, true,
+  PADDLE_ENFORCE_EQ(is_string_pyarray,
+                    true,
                     platform::errors::InvalidArgument(
                         "Expect the dtype of numpy array is string or "
                         "unicode, but recevie dtype %s",
@@ -575,7 +590,8 @@ void SetStringTensorFromPyArray(phi::StringTensor *self, const py::array &array,
         pstring pstr(utf8_len - 1, 0);
         phi::strings::GetUTF8Str(
             reinterpret_cast<const uint32_t *>(array.data()) + unicode_len * i,
-            pstr.mdata(), unicode_len);
+            pstr.mdata(),
+            unicode_len);
         dst[i] = pstr;
       }
     }
@@ -588,7 +604,8 @@ void SetStringTensorFromPyArray(phi::StringTensor *self, const py::array &array,
 
 template <typename T>
 void SetUVATensorFromPyArrayImpl(framework::LoDTensor *self_tensor,
-                                 const py::array_t<T> &array, int device_id) {
+                                 const py::array_t<T> &array,
+                                 int device_id) {
 #if defined(PADDLE_WITH_CUDA)
   VLOG(4) << "Running in SetUVATensorFromPyArrayImpl.";
   std::vector<int64_t> dims;
@@ -603,16 +620,19 @@ void SetUVATensorFromPyArrayImpl(framework::LoDTensor *self_tensor,
   auto data_type = framework::ToDataType(std::type_index(typeid(T)));
   const auto &need_allocate_size = numel * framework::SizeOfType(data_type);
   T *data_ptr;
-  cudaHostAlloc(reinterpret_cast<void **>(&data_ptr), need_allocate_size,
+  cudaHostAlloc(reinterpret_cast<void **>(&data_ptr),
+                need_allocate_size,
                 cudaHostAllocWriteCombined | cudaHostAllocMapped);
   std::memcpy(data_ptr, array.data(), array.nbytes());
 
   void *cuda_device_pointer = nullptr;
   cudaHostGetDevicePointer(reinterpret_cast<void **>(&cuda_device_pointer),
-                           reinterpret_cast<void *>(data_ptr), 0);
+                           reinterpret_cast<void *>(data_ptr),
+                           0);
   std::shared_ptr<memory::allocation::Allocation> holder =
       std::make_shared<memory::allocation::Allocation>(
-          cuda_device_pointer, need_allocate_size,
+          cuda_device_pointer,
+          need_allocate_size,
           platform::CUDAPlace(device_id));
   self_tensor->ResetHolderWithType(holder,
                                    framework::TransToPhiDataType(data_type));
@@ -622,7 +642,8 @@ void SetUVATensorFromPyArrayImpl(framework::LoDTensor *self_tensor,
 template <typename T>
 void SetUVATensorFromPyArray(
     const std::shared_ptr<paddle::imperative::VarBase> &self,
-    const py::array_t<T> &array, int device_id) {
+    const py::array_t<T> &array,
+    int device_id) {
 #if defined(PADDLE_WITH_CUDA)
   VLOG(4) << "Running in SetUVATensorFromPyArray for VarBase.";
   auto *self_tensor = self->MutableVar()->GetMutable<framework::LoDTensor>();
@@ -633,7 +654,8 @@ void SetUVATensorFromPyArray(
 template <typename T>
 void SetUVATensorFromPyArray(
     const std::shared_ptr<paddle::experimental::Tensor> &self,
-    const py::array_t<T> &array, int device_id) {
+    const py::array_t<T> &array,
+    int device_id) {
 #if defined(PADDLE_WITH_CUDA)
   VLOG(4) << "Running in SetUVATensorFromPyArray for Phi::Tensor.";
   phi::DenseTensorMeta meta =
@@ -652,7 +674,8 @@ void SetUVATensorFromPyArray(
 }
 
 template <typename T, size_t D>
-void _sliceCompute(const framework::Tensor *in, framework::Tensor *out,
+void _sliceCompute(const framework::Tensor *in,
+                   framework::Tensor *out,
                    const platform::CPUDeviceContext &ctx,
                    const std::vector<int> &axes,
                    const std::vector<int> &starts) {
@@ -688,15 +711,21 @@ void _sliceCompute(const framework::Tensor *in, framework::Tensor *out,
 template <typename T>
 void _concatCompute(const std::vector<paddle::framework::Tensor> &ins,
                     paddle::framework::Tensor *out,
-                    const platform::CPUDeviceContext &ctx, int64_t axis) {
+                    const platform::CPUDeviceContext &ctx,
+                    int64_t axis) {
   if (axis == 0 && ins.size() < 10) {
     size_t output_offset = 0;
     for (auto &in : ins) {
       auto in_stride = phi::stride_numel(in.dims());
       auto out_stride = phi::stride_numel(out->dims());
       paddle::operators::StridedNumelCopyWithAxis<T>(
-          ctx, axis, out->data<T>() + output_offset, out_stride, in.data<T>(),
-          in_stride, in_stride[axis]);
+          ctx,
+          axis,
+          out->data<T>() + output_offset,
+          out_stride,
+          in.data<T>(),
+          in_stride,
+          in_stride[axis]);
       output_offset += in_stride[axis];
     }
   } else {
@@ -706,9 +735,13 @@ void _concatCompute(const std::vector<paddle::framework::Tensor> &ins,
   }
 }
 
-inline void _getSliceinfo(const framework::Tensor &self, py::object obj,
-                          const int64_t dim, int64_t *pstart, int64_t *pstop,
-                          int64_t *pstep, int64_t *pslicelength) {
+inline void _getSliceinfo(const framework::Tensor &self,
+                          py::object obj,
+                          const int64_t dim,
+                          int64_t *pstart,
+                          int64_t *pstop,
+                          int64_t *pstep,
+                          int64_t *pslicelength) {
   auto &start = *pstart;
   auto &stop = *pstop;
   auto &step = *pstep;
@@ -718,7 +751,8 @@ inline void _getSliceinfo(const framework::Tensor &self, py::object obj,
       0 <= dim && dim < srcDDim.size(),
       platform::errors::OutOfRange("The dim %d of slice is out of bounds, it "
                                    "shound be in the range of [0, %d).",
-                                   dim, srcDDim.size()));
+                                   dim,
+                                   srcDDim.size()));
 
   if (py::isinstance<py::slice>(obj)) {
     size_t lstart, lstop, lstep, lslicelength;
@@ -739,7 +773,9 @@ inline void _getSliceinfo(const framework::Tensor &self, py::object obj,
         std::abs(start) < srcDDim[dim],
         platform::errors::OutOfRange("The start %d of slice is out of bounds, "
                                      "it shound be in the range of (%d, %d).",
-                                     start, -srcDDim[dim], srcDDim[dim]));
+                                     start,
+                                     -srcDDim[dim],
+                                     srcDDim[dim]));
     start = (start >= 0) ? start : srcDDim[dim] - start;
     stop = start + 1;
     step = 1;
@@ -779,9 +815,11 @@ inline framework::Tensor *_getTensor(const framework::Tensor &self,
 }
 
 template <typename T>
-void _sliceDapper(const framework::Tensor *in, framework::Tensor *out,
+void _sliceDapper(const framework::Tensor *in,
+                  framework::Tensor *out,
                   const platform::CPUDeviceContext &ctx,
-                  const std::vector<int> &axes, const std::vector<int> &starts,
+                  const std::vector<int> &axes,
+                  const std::vector<int> &starts,
                   int size) {
   switch (size) {
     case 1:
@@ -821,7 +859,9 @@ void _sliceDapper(const framework::Tensor *in, framework::Tensor *out,
 template <typename T>
 inline framework::Tensor *_sliceWrapper(const framework::Tensor &self,
                                         const platform::CPUDeviceContext &ctx,
-                                        py::object obj, int dim, int64_t start,
+                                        py::object obj,
+                                        int dim,
+                                        int64_t start,
                                         int64_t slicelength) {
   framework::DDim dstDDim = self.dims();
   dstDDim[dim] = static_cast<int64_t>(slicelength);
@@ -834,7 +874,8 @@ inline framework::Tensor *_sliceWrapper(const framework::Tensor &self,
 
 template <typename T>
 inline framework::Tensor *_sliceAndConcat(const framework::Tensor &self,
-                                          py::object obj, int dim) {
+                                          py::object obj,
+                                          int dim) {
   platform::CPUDeviceContext ctx;
   int64_t start, stop, step, slicelength;
   _getSliceinfo(self, obj, dim, &start, &stop, &step, &slicelength);
@@ -856,7 +897,8 @@ inline framework::Tensor *_sliceAndConcat(const framework::Tensor &self,
 }
 
 inline framework::Tensor *_sliceTensor(const framework::Tensor &self,
-                                       py::object obj, int dim) {
+                                       py::object obj,
+                                       int dim) {
   auto src_type = framework::TransToProtoVarType(self.dtype());
   switch (src_type) {
     case framework::proto::VarType::FP16:
@@ -959,43 +1001,53 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
       !is_custom_device_tensor) {
     if (!need_deep_copy) {
       auto base = py::cast(std::move(tensor));
-      return py::array(py::dtype(py_dtype_str.c_str()), py_dims, py_strides,
-                       const_cast<void *>(tensor_buf_ptr), base);
+      return py::array(py::dtype(py_dtype_str.c_str()),
+                       py_dims,
+                       py_strides,
+                       const_cast<void *>(tensor_buf_ptr),
+                       base);
     } else {
       py::array py_arr(py::dtype(py_dtype_str.c_str()), py_dims, py_strides);
       PADDLE_ENFORCE_EQ(
-          py_arr.writeable(), true,
+          py_arr.writeable(),
+          true,
           platform::errors::InvalidArgument(
               "PyArray is not writable, in which case memory leak "
               "or double free would occur"));
       PADDLE_ENFORCE_EQ(
-          py_arr.owndata(), true,
+          py_arr.owndata(),
+          true,
           platform::errors::InvalidArgument(
               "PyArray does not own data, in which case  memory leak "
               "or double free would occur"));
       platform::CPUPlace place;
       size_t copy_bytes = sizeof_dtype * numel;
-      paddle::memory::Copy(place, py_arr.mutable_data(), place, tensor_buf_ptr,
-                           copy_bytes);
+      paddle::memory::Copy(
+          place, py_arr.mutable_data(), place, tensor_buf_ptr, copy_bytes);
       return py_arr;
     }
   } else if (is_xpu_tensor) {
 #ifdef PADDLE_WITH_XPU
     py::array py_arr(py::dtype(py_dtype_str.c_str()), py_dims, py_strides);
-    PADDLE_ENFORCE_EQ(py_arr.writeable(), true,
+    PADDLE_ENFORCE_EQ(py_arr.writeable(),
+                      true,
                       platform::errors::InvalidArgument(
                           "PyArray is not writable, in which case memory leak "
                           "or double free would occur"));
     PADDLE_ENFORCE_EQ(
-        py_arr.owndata(), true,
+        py_arr.owndata(),
+        true,
         platform::errors::InvalidArgument(
             "PyArray does not own data, in which case  memory leak "
             "or double free would occur"));
 
     size_t copy_bytes = sizeof_dtype * numel;
     auto p = tensor.place();
-    paddle::memory::Copy(platform::CPUPlace(), py_arr.mutable_data(), p,
-                         tensor_buf_ptr, copy_bytes);
+    paddle::memory::Copy(platform::CPUPlace(),
+                         py_arr.mutable_data(),
+                         p,
+                         tensor_buf_ptr,
+                         copy_bytes);
     return py_arr;
 #else
     PADDLE_THROW(platform::errors::PermissionDenied(
@@ -1005,20 +1057,26 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
   } else if (is_gpu_tensor) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
     py::array py_arr(py::dtype(py_dtype_str.c_str()), py_dims, py_strides);
-    PADDLE_ENFORCE_EQ(py_arr.writeable(), true,
+    PADDLE_ENFORCE_EQ(py_arr.writeable(),
+                      true,
                       platform::errors::InvalidArgument(
                           "PyArray is not writable, in which case memory leak "
                           "or double free would occur"));
     PADDLE_ENFORCE_EQ(
-        py_arr.owndata(), true,
+        py_arr.owndata(),
+        true,
         platform::errors::InvalidArgument(
             "PyArray does not own data, in which case  memory leak "
             "or double free would occur"));
 
     size_t copy_bytes = sizeof_dtype * numel;
     auto p = tensor.place();
-    paddle::memory::Copy(platform::CPUPlace(), py_arr.mutable_data(), p,
-                         tensor_buf_ptr, copy_bytes, nullptr);
+    paddle::memory::Copy(platform::CPUPlace(),
+                         py_arr.mutable_data(),
+                         p,
+                         tensor_buf_ptr,
+                         copy_bytes,
+                         nullptr);
     return py_arr;
 #else
     PADDLE_THROW(platform::errors::PermissionDenied(
@@ -1028,12 +1086,14 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
   } else if (is_npu_tensor) {
 #ifdef PADDLE_WITH_ASCEND_CL
     py::array py_arr(py::dtype(py_dtype_str.c_str()), py_dims, py_strides);
-    PADDLE_ENFORCE_EQ(py_arr.writeable(), true,
+    PADDLE_ENFORCE_EQ(py_arr.writeable(),
+                      true,
                       platform::errors::InvalidArgument(
                           "PyArray is not writable, in which case memory leak "
                           "or double free would occur"));
     PADDLE_ENFORCE_EQ(
-        py_arr.owndata(), true,
+        py_arr.owndata(),
+        true,
         platform::errors::InvalidArgument(
             "PyArray does not own data, in which case  memory leak "
             "or double free would occur"));
@@ -1043,7 +1103,10 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &ctx = *pool.Get(tensor.place());
     paddle::memory::Copy(
-        platform::CPUPlace(), py_arr.mutable_data(), p, tensor_buf_ptr,
+        platform::CPUPlace(),
+        py_arr.mutable_data(),
+        p,
+        tensor_buf_ptr,
         copy_bytes,
         reinterpret_cast<const platform::NPUDeviceContext &>(ctx).stream());
     ctx.Wait();
@@ -1056,12 +1119,14 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
   } else if (is_mlu_tensor) {
 #ifdef PADDLE_WITH_MLU
     py::array py_arr(py::dtype(py_dtype_str.c_str()), py_dims, py_strides);
-    PADDLE_ENFORCE_EQ(py_arr.writeable(), true,
+    PADDLE_ENFORCE_EQ(py_arr.writeable(),
+                      true,
                       platform::errors::InvalidArgument(
                           "PyArray is not writable, in which case memory leak "
                           "or double free would occur"));
     PADDLE_ENFORCE_EQ(
-        py_arr.owndata(), true,
+        py_arr.owndata(),
+        true,
         platform::errors::InvalidArgument(
             "PyArray does not own data, in which case  memory leak "
             "or double free would occur"));
@@ -1071,7 +1136,10 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &ctx = *pool.Get(tensor.place());
     paddle::memory::Copy(
-        platform::CPUPlace(), py_arr.mutable_data(), p, tensor_buf_ptr,
+        platform::CPUPlace(),
+        py_arr.mutable_data(),
+        p,
+        tensor_buf_ptr,
         copy_bytes,
         reinterpret_cast<const platform::MLUDeviceContext &>(ctx).stream());
     ctx.Wait();
@@ -1084,12 +1152,14 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
   } else if (is_custom_device_tensor) {
 #ifdef PADDLE_WITH_CUSTOM_DEVICE
     py::array py_arr(py::dtype(py_dtype_str.c_str()), py_dims, py_strides);
-    PADDLE_ENFORCE_EQ(py_arr.writeable(), true,
+    PADDLE_ENFORCE_EQ(py_arr.writeable(),
+                      true,
                       platform::errors::InvalidArgument(
                           "PyArray is not writable, in which case memory leak "
                           "or double free would occur"));
     PADDLE_ENFORCE_EQ(
-        py_arr.owndata(), true,
+        py_arr.owndata(),
+        true,
         platform::errors::InvalidArgument(
             "PyArray does not own data, in which case  memory leak "
             "or double free would occur"));
@@ -1098,8 +1168,11 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &ctx = *pool.Get(tensor.place());
     paddle::memory::Copy(
-        platform::CPUPlace(), py_arr.mutable_data(), tensor.place(),
-        tensor_buf_ptr, copy_bytes,
+        platform::CPUPlace(),
+        py_arr.mutable_data(),
+        tensor.place(),
+        tensor_buf_ptr,
+        copy_bytes,
         reinterpret_cast<const platform::CustomDeviceContext &>(ctx).stream());
     ctx.Wait();
     return py_arr;

paddle/phi/common/data_type.h

@@ -18,7 +18,12 @@ limitations under the License. */
 #include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/complex.h"
 #include "paddle/phi/common/float16.h"
-#include "paddle/phi/common/pstring.h"
+
+namespace phi {
+namespace dtype {
+class pstring;
+}  // namespace dtype
+}  // namespace phi
 
 namespace paddle {
 namespace experimental {

paddle/phi/core/string_tensor.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/phi/core/string_tensor.h"
 
 #include "paddle/fluid/memory/malloc.h"
+#include "paddle/phi/common/pstring.h"
 
 namespace phi {
 

paddle/phi/core/string_tensor.h

@@ -14,13 +14,16 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/phi/common/pstring.h"
 #include "paddle/phi/core/allocator.h"
 #include "paddle/phi/core/tensor_base.h"
 #include "paddle/phi/core/tensor_meta.h"
 
 namespace phi {
 
+namespace dtype {
+class pstring;
+}  // namespace dtype
+
 /// \brief In Paddle 2.3, we add a new type of Tensor, StringTensor,
 /// which is designed for string data management.
 /// During the entire life cycle of a StringTensor, its device type and key

paddle/phi/kernels/strings/cpu/strings_copy_kernel.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/phi/kernels/strings/strings_copy_kernel.h"
 
 #include "glog/logging.h"
+#include "paddle/phi/common/pstring.h"
 #include "paddle/phi/core/kernel_registry.h"
 
 namespace phi {

paddle/phi/tests/api/test_strings_lower_upper_api.cc

@@ -19,6 +19,7 @@ limitations under the License. */
 #include "paddle/phi/api/include/strings_api.h"
 #include "paddle/phi/api/lib/utils/allocator.h"
 #include "paddle/phi/backends/all_context.h"
+#include "paddle/phi/common/pstring.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/core/string_tensor.h"