cupy in destructor is changed to None

Description

Hi,

I found out cupy in destructor would be None

To Reproduce

Here are the scripts to reproduce

c.py

import cupy
import numpy as np
import ctypes as C

class ManagedMemory():
    #  https://stackoverflow.com/questions/23930671/how-to-create-n-dim-numpy-array-from-a-pointer
    def __init__(self, shape, dtype):
        size = 1
        for s in shape:
            size *= s
        self.mem = cupy.cuda.runtime.mallocManaged(size * np.dtype(dtype).itemsize)
        self.ptr =C.cast(self.mem, C.POINTER(C.c_float))
        ctype = self.infer_from_dtype(dtype)
        self.tensor = np.ctypeslib.as_array(self.ptr, shape=shape)

    def infer_from_dtype(self, dtype):
        if dtype == np.float32:
            return C.c_float
        elif dtype == np.int32:
            return C.c_int
        else:
            raise NotImplementedError()
    def __del__(self):
        del self.tensor
        cupy.cuda.runtime.free(self.mem)

d.py

from c import ManagedMemory
import numpy as np
mem = ManagedMemory((5,), np.float32)

python d.py

would give

Exception ignored in: <bound method ManagedMemory.__del__ of <c.ManagedMemory object at 0x7f9669da58>>
Traceback (most recent call last):
  File "/home/acer/charmander/c.py", line 25, in __del__
AttributeError: 'NoneType' object has no attribute 'cuda'

Installation

Wheel (pip install cupy-***)

Environment

# Paste the output here
addict==2.4.0
-e git+ssh://[e-mail]/acer4smartanywhere/charmander.git@95646c34f2f333cdee8201caeae4b54197416f38#egg=charmander_v2
cupy @ file:///home/acer/cupy-8.2.0-cp36-cp36m-linux_aarch64.whl
cycler==0.11.0
Cython==0.29.26
dataclasses==0.8
decorator==4.4.2
easydict==1.9
fastrlock==0.8
imageio==2.13.5
joblib==1.1.0
kiwisolver==1.3.1
lmdb==1.3.0
matplotlib==3.3.4
-e git+https://github.com/open-mmlab/mmclassification.git@5232965b17b6c050f9b328b3740c631ed4034624#egg=mmcls
mmcv-full @ file:///home/acer/mmcv/dist/mmcv_full-1.4.3-cp36-cp36m-linux_aarch64.whl
-e git+https://github.com/open-mmlab/mmdetection.git@ff9bc39913cb3ff5dde79d3933add7dc2561bab7#egg=mmdet
networkx==2.5.1
ninja==1.10.2.3
numpy==1.19.4
onnx==1.10.2
opencv-python==4.5.5.62
packaging==21.3
Pillow==8.4.0
protobuf==3.19.3
pybind11==2.9.0
pycocotools==2.0.4
pyparsing==3.0.6
python-dateutil==2.8.2
PyWavelets==1.1.1
PyYAML==6.0
scikit-image==0.17.2
scikit-learn==0.24.2
scipy==1.5.4
setuptools-cythonize==1.0.6
six==1.16.0
terminaltables==3.1.10
threadpoolctl==3.0.0
tifffile==2020.9.3
torch==1.9.0
torchvision @ file:///home/acer/vision/dist/torchvision-0.10.0a0%2B300a8a4-cp36-cp36m-linux_aarch64.whl
tqdm==4.62.3
-e git+ssh://[e-mail]/acer4smartanywhere/treecko.git@9eb62aa97f27774d104950664b6610dd4041d5a2#egg=treecko
typing_extensions==4.0.1
yapf==0.32.0

Additional Information

By the way,

There is an cupy.cuda.ManagedMemory, Do I need to manually free its internal memory? I found out it would call https://github.com/cupy/cupy/blob/v10.0.0/cupy/cuda/memory.pyx#L245

[leofang]

1. There is no guarantee in Python that by the time any object’s __del__ is called other Python modules and objects are still alive. We hit (a few variants of) this issue before, and realized that even after Python 3.4 that implemented PEP 442 this could still happen. For a fix, you can either check if cupy.cuda.runtime.free is None, or better use the finalizer instead of __del__.
2. You don’t need to manually free the internal memory of ManagedMemory.

[twmht]

@leofang

yup. I am also referring ManagedMemory, which has cupy.cuda.ManagedMemory as its member.

when create ManagedMemory instance from d.py, it can’t destroy cupy.cuda.ManagedMemory automatically after program exits.

like

AttributeError: 'NoneType' object has no attribute '_declare_synchronize'
Exception ignored in: 'cupy.cuda.memory.ManagedMemory.__dealloc__'
AttributeError: 'NoneType' object has no attribute '_declare_synchronize