[API Compatiblity] Add more coverage

paddle/phi/core/compat/convert_utils.cc

@@ -44,6 +44,13 @@ Backend TransToPhiBackend(const phi::Place& place) {
     }
     case AllocationType::XPU:
       return Backend::XPU;
+    case AllocationType::XPUPINNED: {
+      if (FLAGS_pinned_memory_as_cpu_backend) {
+        return Backend::CPU;
+      } else {
+        return Backend::XPU;
+      }
+    }
     case AllocationType::IPU:
       return Backend::IPU;
     case AllocationType::UNDEFINED:

python/paddle/base/dygraph/math_op_patch.py

@@ -20,6 +20,9 @@
 
 import paddle
 from paddle import _C_ops
+from paddle.utils.decorator_utils import (
+    size_args_decorator_patch,
+)
 
 from .. import core
 from ..framework import convert_np_dtype_to_dtype_
@@ -312,6 +315,7 @@ def _new_full_(
             pin_memory=pin_memory,
         )
 
+    @size_args_decorator_patch
     def _new_empty_(
         var: Tensor,
         size: ShapeLike,
@@ -334,6 +338,7 @@ def _new_empty_(
             pin_memory=pin_memory,
         )
 
+    @size_args_decorator_patch
     def _new_ones_(
         var: Tensor,
         size: ShapeLike,
@@ -357,6 +362,7 @@ def _new_ones_(
             pin_memory=pin_memory,
         )
 
+    @size_args_decorator_patch
     def _new_zeros_(
         var: Tensor,
         size: ShapeLike,

python/paddle/base/dygraph/tensor_patch_methods.py

@@ -1157,10 +1157,16 @@ def cuda(
 
     @framework.dygraph_only
     def pin_memory(self: Tensor, blocking: bool = True) -> Tensor:
-        if self.place.is_cuda_pinned_place():
+        if (
+            self.place.is_cuda_pinned_place()
+            or self.place.is_xpu_pinned_place()
+        ):
             return self
         else:
-            res = self._copy_to(core.CUDAPinnedPlace(), blocking)
+            if paddle.device.is_compiled_with_xpu():
+                res = self._copy_to(core.XPUPinnedPlace(), blocking)
+            else:
+                res = self._copy_to(core.CUDAPinnedPlace(), blocking)
             res.stop_gradient = self.stop_gradient
             res.persistable = self.persistable
             return res

python/paddle/pir/math_op_patch.py

@@ -25,6 +25,9 @@
 from paddle import _C_ops
 from paddle.base.libpaddle import DataType
 from paddle.base.wrapped_decorator import wrap_decorator
+from paddle.utils.decorator_utils import (
+    size_args_decorator_patch,
+)
 
 from . import Value
 
@@ -686,6 +689,7 @@ def _new_full_(
             pin_memory=pin_memory,
         )
 
+    @size_args_decorator_patch
     def _new_empty_(
         self,
         size: ShapeLike,
@@ -731,6 +735,7 @@ def _new_empty_(
             pin_memory=pin_memory,
         )
 
+    @size_args_decorator_patch
     def _new_ones_(
         self,
         size: ShapeLike,
@@ -777,6 +782,7 @@ def _new_ones_(
             pin_memory=pin_memory,
         )
 
+    @size_args_decorator_patch
     def _new_zeros_(
         self,
         size: ShapeLike,

python/paddle/tensor/creation.py

@@ -1792,6 +1792,8 @@ def _check_attr(attr, message):
         )
         if requires_grad is True:
             tensor.stop_gradient = False
+            if out is not None:
+                out.stop_gradient = False
         if pin_memory and in_dynamic_mode():
             tensor = tensor.pin_memory()
         return tensor
@@ -1960,6 +1962,8 @@ def full(
     )
     if requires_grad is True:
         tensor.stop_gradient = False
+        if out is not None:
+            out.stop_gradient = False
     if pin_memory and in_dynamic_mode():
         tensor = tensor.pin_memory()
     return tensor
@@ -2109,6 +2113,8 @@ def arange(
             out=out,
         )
         tensor.stop_gradient = not requires_grad
+        if out is not None:
+            out.stop_gradient = not requires_grad
         if pin_memory and in_dynamic_mode():
             tensor = tensor.pin_memory()
         return tensor
@@ -2161,6 +2167,8 @@ def arange(
             out=out,
         )
         tensor.stop_gradient = not requires_grad
+        if out is not None:
+            out.stop_gradient = not requires_grad
         if pin_memory and in_dynamic_mode():
             tensor = tensor.pin_memory()
         return tensor
@@ -2299,6 +2307,8 @@ def range(
             out=out,
         )
         tensor.stop_gradient = not requires_grad
+        if out is not None:
+            out.stop_gradient = not requires_grad
         return tensor
 
     if not isinstance(start, (Variable, paddle.pir.Value)):
@@ -2332,6 +2342,8 @@ def range(
         out=out,
     )
     tensor.stop_gradient = not requires_grad
+    if out is not None:
+        out.stop_gradient = not requires_grad
     return tensor
 
 
@@ -3013,6 +3025,8 @@ def empty(
             tensor = tensor.pin_memory()
         if requires_grad is True:
             tensor.stop_gradient = False
+            if out is not None:
+                out.stop_gradient = False
         return tensor
     else:
         helper = LayerHelper("empty", **locals())

python/paddle/utils/decorator_utils.py

@@ -295,6 +295,35 @@ def wrapped_func(*args: Any, **kwargs: Any) -> Any:
     return wrapped_func
 
 
+def size_args_decorator_patch(method: Callable) -> Callable:
+    """
+    A decorator that allow *size for patching method to Tensor.
+    e.g. Tensor.method(*size, *, ...).
+
+    Usage Example:
+
+    paddle.randn([]).new_ones(1, dtype=paddle.float32)
+    paddle.randn([]).new_ones(1, 2, 3, dtype=paddle.float32)
+    paddle.randn([]).new_ones([1, 2, 3], dtype=paddle.float32)
+    paddle.randn([]).new_ones(size=[1, 2, 3], dtype=paddle.float32)
+    paddle.randn([]).new_ones([1, 2, 3], paddle.float32)
+    """
+
+    @functools.wraps(method)
+    def wrapped_func(*args: Any, **kwargs: Any) -> Any:
+        if len(args) >= 2 and isinstance(args[1], int):
+            # args[0]: Tensor
+            # args[1:]: *size
+            kwargs['size'] = list(args[1:])
+            args = (args[0],)
+
+        return method(*args, **kwargs)
+
+    wrapped_func.__signature__ = inspect.signature(method)
+
+    return wrapped_func
+
+
 class VariableArgsDecorator(DecoratorBase):
     def __init__(self, var: str) -> None:
         super().__init__()