提交 76119020 编写于 作者: M Megvii Engine Team

docs(imperative): add dlpack docstring

GitOrigin-RevId: 4898d4a4db9053d70633ae8ec8cc4c10ca9dc0c9
上级 15637898
from typing import Any
from ..core._imperative_rt.core2 import _from_dlpack
__all__ = [
"to_dlpack",
"from_dlpack",
]
def to_dlpack(tensor, stream=None):
"""
Encodes a megengine tensor to DLPack.
Args:
tensor (Tensor): The input tensor, and the data type can be `float16`, `float32`,
`int8`, `int16`, `int32`, `uint8`, `uint16`, `complex64`.
stream (Integer or None): An optional Python integer representing a CUDA stream,
The current stream is synchronized with this stream before the capsule is created.
If None or -1 is passed then no synchronization is performed.
Returns:
dltensor, and the data type is PyCapsule.
Examples:
.. code-block:: python
import megengine as mge
# x is a tensor with shape [2, 3]
x = mge.tensor([[0.2, 0.3, 0.5],
[0.1, 0.2, 0.6]])
dlpack = mge.utils.dlpack.to_dlpack(x)
print(dlpack)
# <capsule object "dltensor" at 0x7ff04b69cc00>
"""
if stream is not None and stream != -1:
return tensor.__dlpack__(stream)
else:
return tensor.__dlpack__()
def from_dlpack(ext_tensor: Any, stream=None):
def from_dlpack(ext_tensor, stream=None):
"""
Decodes a DLPack to a megengine tensor.
Args:
ext_tensor (PyCapsule): a PyCapsule object with the dltensor.
stream (Integer or None): An optional Python integer representing a CUDA stream,
user needs to know which stream the dlpack is generated.
If None then represent producers and consumers on the same stream.
Returns:
tensor (Tensor): a megengine tensor decoded from DLPack.
Examples:
.. code-block:: python
import megengine as mge
# x is a tensor with shape [2, 3]
x = mge.tensor([[0.2, 0.3, 0.5],
[0.1, 0.2, 0.6]])
dlpack = mge.utils.dlpack.to_dlpack(x)
x = mge.utils.dlpack.from_dlpack(dlpack)
print(x)
# Tensor([[0.2, 0.3, 0.5],
# [0.1, 0.2, 0.6]], device=gpu0:0)
"""
if isinstance(stream, int):
assert stream >= 0, "device stream should be a positive value"
stream = 0 if stream is None else stream
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册