It is easy to crash MXNet when tensor goes larger

[classicsong]

Description

When I use large tensor, it is easy to crash the MXNet kernel.
Using following python code to reproduce:

>>> import mxnet.ndarray as nd

>>> a = nd.random.randn(4, 256, 1, 100, 100)
>>> b = nd.broadcast_axis(a, axis=2, size=256)
>>> b.size
2621440000
>>> b.asnumpy()
CRASH HERE

The error looks like an int32 overflow on shape.size.
Any easy way to fix this out? The only way I found out is to compile MXNet with USE_INT64_TENSOR_SIZE = ON, which is slower than the default one.

Environment info (Required)

mxnet 1.5.1 (pip3 install)

Package used (Python/R/Scala/Julia):
Python

Error Message:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.5/dist-packages/mxnet/ndarray/ndarray.py", line 1996, in asnumpy
    ctypes.c_size_t(data.size)))
  File "/usr/local/lib/python3.5/dist-packages/mxnet/base.py", line 253, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [07:26:09] include/mxnet/././tensor_blob.h:290: Check failed: this->shape_.Size() == static_cast<size_t>(shape.Size()) (2621440000 vs. 18446744072036024320) : TBlob.get_with_shape: new and old shape do not match total elements

[ddavydenko]

@mxnet-label-bot Add [Bug, Large Tensor Support]

[roywei]

cc @access2rohit

[sxjscience]

We should raise an error message in the C++ side when we are going to create a large NDArray.

[ChaiBapchya]

Yes it is being tracked here #16570

[zachgk]

Is this resolved now that #16570 is merged?

[samskalicky]

@lanking520 assign @ChaiBapchya

[ChaiBapchya]

We can close this ticket. As the solution is to build with large tensor as the issue author pointed it out. An error message is already raised as part of #16570 if large array is created when large tensor support isn't enabled.