[Compat] Compatible with API development (PaddlePaddle#76247)

fxyfxy777 · web-flow · commit 6fd172397a11 · 2025-11-13T21:15:05.000+08:00
diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py
@@ -271,6 +271,7 @@ def new_init(self, *args, **kwargs):
     set_default_dtype,
 )
 from .framework.random import (
+    Generator,
     get_cuda_rng_state,
     get_rng_state,
     seed,
@@ -1484,6 +1485,7 @@ def __dir__(self):
     'conv3d',
     'manual_seed',
     'softmax',
+    'Generator',
     'adaptive_avg_pool1d',
     'autocast',
 ]
diff --git a/python/paddle/compat/__init__.py b/python/paddle/compat/__init__.py
@@ -955,11 +955,7 @@ def __getattr__(self, name: str) -> Any:
         return getattr(self._original_module, name)
 
 
-GLOBAL_OVERRIDES = {
-    "torch.Generator": create_fake_class(
-        "Generator", {"manual_seed": create_fake_function("manual_seed")}
-    ),
-}
+GLOBAL_OVERRIDES = {}
 
 
 def _is_torch_module(name: str) -> bool:
diff --git a/python/paddle/cuda/__init__.py b/python/paddle/cuda/__init__.py
@@ -23,8 +23,10 @@
 from paddle.device import (
     Event,
     Stream,
+    StreamContext,
     _device_to_paddle as _device_to_paddle,
     amp,  # noqa: F401
+    current_device,
     device,
     is_available as _device_is_available,
     is_bf16_supported,
@@ -33,7 +35,7 @@
     manual_seed_all as device_manual_seed_all,
     reset_peak_memory_stats,
     set_stream,
-    stream_guard as _PaddleStreamGuard,
+    stream,
 )
 from paddle.tensor.creation import (
     BFloat16Tensor,
@@ -284,39 +286,6 @@ def manual_seed_all(seed: int) -> None:
     device_manual_seed_all(seed)
 
 
-class StreamContext(_PaddleStreamGuard):
-    """
-    Notes:
-        This API only supports dynamic graph mode currently.
-    A context manager that specifies the current stream context by the given stream.
-
-    Args:
-        stream(Stream, optional): the selected stream. If stream is None, just yield.
-
-    Returns:
-        None.
-
-    Examples:
-        .. code-block:: python
-
-            >>> # doctest: +REQUIRES(env:CUSTOM_DEVICE)
-            >>> import paddle
-
-            >>> paddle.set_device('cuda')
-            >>> s = paddle.cuda.Stream()
-            >>> data1 = paddle.ones(shape=[20])
-            >>> data2 = paddle.ones(shape=[20])
-            >>> data3 = data1 + data2
-            >>> with paddle.cuda.StreamContext(s):
-            ...     s.wait_stream(paddle.cuda.current_stream()) # type: ignore[attr-defined]
-            ...     data4 = data1 + data3
-
-    """
-
-    def __init__(self, stream: paddle_device.Stream):
-        super().__init__(stream)
-
-
 def get_rng_state(device: DeviceLike | None = None) -> core.GeneratorState:
     """
     Return the random number generator state of the specified device.
@@ -369,40 +338,6 @@ def set_rng_state(
     paddle_device.set_rng_state(new_state, device)
 
 
-def stream(stream_obj: paddle_device.Stream | None) -> StreamContext:
-    '''
-
-    Notes:
-        This API only supports dynamic graph mode currently.
-    A context manager that specifies the current stream context by the given stream.
-
-    Args:
-        stream(Stream, optional): the selected stream. If stream is None, just yield.
-
-    Returns:
-        None.
-
-    Examples:
-        .. code-block:: python
-
-            >>> # doctest: +REQUIRES(env:CUSTOM_DEVICE)
-            >>> import paddle
-
-            >>> paddle.set_device('cuda')
-            >>> s = paddle.cuda.Stream()
-            >>> data1 = paddle.ones(shape=[20])
-            >>> data2 = paddle.ones(shape=[20])
-            >>> data3 = data1 + data2
-
-            >>> with paddle.cuda.stream(s):
-            ...     s.wait_stream(paddle.cuda.current_stream())
-            ...     data4 = data1 + data3
-            >>> print(data4)
-
-    '''
-    return StreamContext(stream_obj)
-
-
 class nvtx:
     """Namespace for NVTX marker operations."""
 
@@ -559,35 +494,6 @@ def mem_get_info(device: DeviceLike = None) -> tuple[int, int]:
     return cudart().cudaMemGetInfo(device_id)
 
 
-def current_device() -> int:
-    """
-    Return the index of a currently selected device.
-
-    Returns:
-        int: The index of the currently selected device.
-
-    Examples:
-        .. code-block:: python
-
-            >>> # doctest: +REQUIRES(env:GPU)
-            >>> import paddle
-            >>> device_id = paddle.cuda.current_device()
-            >>> print(f"Current device index: {device_id}")
-    """
-    # Use paddle.device.get_device() to get the current device string
-    device_str = paddle_device.get_device()
-
-    # Parse the device string to extract the device index
-    # Format examples: 'gpu:0', 'xpu:0', 'custom_device:0'
-    if ':' in device_str:
-        device_id = int(device_str.split(':')[1])
-    else:
-        # If no device index is specified, default to 0
-        device_id = 0
-
-    return device_id
-
-
 def device_count() -> int:
     """
     Return the number of devices available.
@@ -972,4 +878,5 @@ def get_stream_from_external(
     "max_memory_allocated",
     "reset_peak_memory_stats",
     "Event",
+    "StreamContext",
 ]
diff --git a/python/paddle/device/__init__.py b/python/paddle/device/__init__.py
@@ -516,6 +516,35 @@ def __exit__(
         return False
 
 
+def current_device() -> int:
+    """
+    Return the index of a currently selected device.
+
+    Returns:
+        int: The index of the currently selected device.
+
+    Examples:
+        .. code-block:: python
+
+            >>> # doctest: +REQUIRES(env:GPU)
+            >>> import paddle
+            >>> device_id = paddle.device.current_device() # this is equivalent to paddle.cuda.current_device()
+            >>> print(f"Current device index: {device_id}")
+    """
+    # Use paddle.device.get_device() to get the current device string
+    device_str = get_device()
+
+    # Parse the device string to extract the device index
+    # Format examples: 'gpu:0', 'xpu:0', 'custom_device:0'
+    if ':' in device_str:
+        device_id = int(device_str.split(':')[1])
+    else:
+        # If no device index is specified, default to 0
+        device_id = 0
+
+    return device_id
+
+
 def is_bf16_supported(including_emulation: bool = True) -> bool:
     """
     Return a bool indicating if the current CUDA/ROCm device supports dtype bfloat16.
@@ -538,9 +567,26 @@ def is_bf16_supported(including_emulation: bool = True) -> bool:
 
     """
     # including_emulation is not used here, but kept for compatibility with the original implementation
-    return core.is_bfloat16_supported(
-        paddle.framework._current_expected_place()
-    )
+    if core.is_bfloat16_supported(paddle.framework._current_expected_place()):
+        return True
+
+    # If CUDA is not available, than it does not support bf16 either
+    if not is_available():
+        return False
+
+    device = get_device()
+
+    # Check for CUDA version and device compute capability.
+    # This is a fast way to check for it.
+    if not including_emulation:
+        return False
+
+    # Finally try to create a bfloat16 device.
+    try:
+        paddle.tensor([1.0], dtype=paddle.bfloat16, device=device)
+        return True
+    except:
+        return False
 
 
 def set_device(device: PlaceLike | int) -> PlaceLike:
@@ -1584,7 +1630,7 @@ class stream_guard:
             >>> data1 = paddle.ones(shape=[20])
             >>> data2 = paddle.ones(shape=[20])
             >>> data3 = data1 + data2
-            >>> with paddle.device.stream_guard(s):
+            >>> with paddle.device.stream_guard(s):# this is equivalent to paddle.cuda.StreamContext(s) and paddle.device.StreamContext(s)
             ...     s.wait_stream(paddle.device.default_stream()) # type: ignore[attr-defined]
             ...     data4 = data1 + data3
 
@@ -1627,6 +1673,43 @@ def __exit__(
             set_stream(self.src_prev_stream)
 
 
+StreamContext = stream_guard
+
+
+def stream(stream: Stream | None) -> stream_guard:
+    '''
+
+    Notes:
+        This API only supports dynamic graph mode currently.
+    A context manager that specifies the current stream context by the given stream.
+
+    Args:
+        stream(Stream, optional): the selected stream. If stream is None, just yield.
+
+    Returns:
+        None.
+
+    Examples:
+        .. code-block:: python
+
+            >>> # doctest: +REQUIRES(env:CUSTOM_DEVICE)
+            >>> import paddle
+
+            >>> paddle.set_device('cuda')
+            >>> s = paddle.device.Stream()
+            >>> data1 = paddle.ones(shape=[20])
+            >>> data2 = paddle.ones(shape=[20])
+            >>> data3 = data1 + data2
+
+            >>> with paddle.device.stream(s): # this is equivalent to paddle.cuda.stream(s)
+            ...     s.wait_stream(paddle.cuda.current_stream())
+            ...     data4 = data1 + data3
+            >>> print(data4)
+
+    '''
+    return StreamContext(stream)
+
+
 class device_guard:
     '''
 
@@ -1900,15 +1983,16 @@ def reset_peak_memory_stats(device: PlaceLike | int | None = None) -> None:
     It sets the peak memory usage back to zero for all devices.
 
     Example:
-        >>> # doctest: +REQUIRES(env:GPU)
-        >>> import paddle
-        >>> paddle.device.set_device('gpu')  # or '<custom_device>'
+        .. code-block:: python
+            >>> # doctest: +REQUIRES(env:GPU)
+            >>> import paddle
+            >>> paddle.device.set_device('gpu')  # or '<custom_device>'
 
-        >>> # paddle.cuda.reset_max_memory_allocated() is equivalent to paddle.device.reset_max_memory_allocated()
+            >>> # paddle.cuda.reset_max_memory_allocated() is equivalent to paddle.device.reset_max_memory_allocated()
 
-        >>> paddle.device.reset_max_memory_allocated(paddle.CUDAPlace(0))
-        >>> paddle.device.reset_max_memory_allocated(0)
-        >>> paddle.device.reset_max_memory_allocated("gpu:0")
+            >>> paddle.device.reset_max_memory_allocated(paddle.CUDAPlace(0))
+            >>> paddle.device.reset_max_memory_allocated(0)
+            >>> paddle.device.reset_max_memory_allocated("gpu:0")
     """
     reset_max_memory_allocated()
 
diff --git a/python/paddle/framework/random.py b/python/paddle/framework/random.py
@@ -271,3 +271,32 @@ def set_random_seed_generator(name: str, seed: int) -> None:
 
 def get_random_seed_generator(name: str) -> paddle.base.core.Generator:
     return core.get_random_seed_generator(name)
+
+
+class Generator:
+    def __new__(
+        cls, device: str | int | paddle.core.Place = None
+    ) -> core.Generator:
+        """
+        Generator is a random number generator.
+
+        Args:
+            device(str|int|paddle.core.Place): The device type to create the generator on.
+                It can be ``cpu``, ``gpu``, ``xpu``, or a paddle.core.Place instance.
+                default is None, which means using current device.
+
+        Examples:
+            .. code-block:: python
+
+                >>> import paddle
+                >>> g_cpu = paddle.Generator()
+        """
+        place = paddle.device.device_to_place(device)
+        if isinstance(place, core.CPUPlace):
+            return core.default_cpu_generator()
+        elif isinstance(place, core.CUDAPlace):
+            return core.default_cuda_generator(place.gpu_device_id())
+        elif isinstance(place, core.XPUPlace):
+            return core.default_xpu_generator(place.gpu_device_id())
+        elif isinstance(place, core.CustomPlace):
+            return core.default_custom_device_generator(place)
diff --git a/test/compat/test_torch_proxy.py b/test/compat/test_torch_proxy.py
@@ -86,14 +86,5 @@ def test_use_torch_inside_inner_function(self):
         )
 
 
-class TestTorchOverriddenClass(unittest.TestCase):
-    def test_overridden_class(self):
-        self.assertRaises(AttributeError, lambda: paddle.Generator)
-        with paddle.compat.use_torch_proxy_guard():
-            import torch
-
-            gen = torch.Generator()
-
-
 if __name__ == "__main__":
     unittest.main()
diff --git a/test/legacy_test/test_cuda_unittest.py b/test/legacy_test/test_cuda_unittest.py
diff --git a/test/legacy_test/test_random_generator_set_get_state.py b/test/legacy_test/test_random_generator_set_get_state.py
