chore: import upstream snapshot with attribution

python/paddle/cuda/graphs.py

@@ -0,0 +1,111 @@
+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# 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.
+"""CUDA graph APIs."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from paddle import core, device as _paddle_device
+from paddle.device.cuda.graphs import (
+    CUDAGraph,
+    is_cuda_graph_supported,
+)
+
+if TYPE_CHECKING:
+    from typing_extensions import Self
+
+    from paddle.device import Stream
+
+
+__all__ = [
+    "CUDAGraph",
+    "graph",
+    "graph_pool_handle",
+    "is_cuda_graph_supported",
+]
+
+
+def graph_pool_handle() -> int:
+    """Return an opaque token usable as the ``pool`` argument of
+    :meth:`CUDAGraph.capture_begin` or :class:`graph`. Graphs sharing the
+    same token share an underlying memory pool.
+    """
+    if not is_cuda_graph_supported():
+        raise RuntimeError(
+            "CUDA Graph is only supported on PaddlePaddle compiled with "
+            "NVIDIA GPU."
+        )
+    return core.CUDAGraph.gen_new_memory_pool_id()
+
+
+class graph:
+    """Context manager that wraps a CUDA graph capture.
+
+    Args:
+        cuda_graph (CUDAGraph): The :class:`CUDAGraph` instance to capture into.
+        pool (int, optional): Memory pool token from :func:`graph_pool_handle`
+            or another graph's :meth:`CUDAGraph.pool`.
+        stream (paddle.cuda.Stream, optional): CUDA stream to capture on.
+            When ``None``, capture happens on the current stream.
+        capture_error_mode (str, optional): One of ``'global'``,
+            ``'thread_local'``, ``'relaxed'``. When ``None`` (default) the
+            underlying :class:`CUDAGraph`'s constructor ``mode`` is used.
+
+    Examples:
+        .. code-block:: pycon
+
+            >>> # doctest: +REQUIRES(env:GPU)
+            >>> import paddle
+            >>> paddle.device.set_device('gpu')
+            >>> g = paddle.cuda.CUDAGraph()
+            >>> x = paddle.zeros([2, 3])
+            >>> with paddle.cuda.graph(g):
+            ...     y = x + 1
+            >>> g.replay()
+    """
+
+    def __init__(
+        self,
+        cuda_graph: CUDAGraph,
+        pool: int | None = None,
+        stream: Stream | None = None,
+        capture_error_mode: str | None = None,
+    ) -> None:
+        self.cuda_graph = cuda_graph
+        self.pool = pool
+        self.capture_stream = stream
+        self.capture_error_mode = capture_error_mode
+        self.stream_ctx = _paddle_device.stream(stream)
+
+    def __enter__(self) -> Self:
+        # Synchronize the graph's own device, not the process-wide current
+        # device which may be CPU (synchronize rejects non-accelerator places).
+        _paddle_device.synchronize(self.cuda_graph._place)
+        _paddle_device.empty_cache()
+        self.stream_ctx.__enter__()
+        try:
+            self.cuda_graph.capture_begin(
+                pool=self.pool, capture_error_mode=self.capture_error_mode
+            )
+        except BaseException:
+            self.stream_ctx.__exit__(None, None, None)
+            raise
+        return self
+
+    def __exit__(self, *args: object) -> None:
+        try:
+            self.cuda_graph.capture_end()
+        finally:
+            self.stream_ctx.__exit__(*args)