torch._foreach_mul_ segmentation fault

[alanhe151220037]

🐛 Describe the bug

The following example will cause segmentation fault.

import torch
import numpy as np
x_array = np.array([],dtype=np.float16)
x_tensor = torch.from_numpy(x_array)
torch._foreach_mul_((x_tensor,),1.0)

Segmentation fault (core dumped)

Strangely, only torch._foreach_mul_ will have this problem and torch._foreach_div_ is fine. Although I haven't quite understand the direct segmentation fault reason, but I think this may caused by the following code that mul_.Scalar calls mul.out but div_.scalar calls div_.Tensor.

pytorch/aten/src/ATen/native/BinaryOps.cpp

Lines 997 to 999 in c6f435b

	Tensor& mul_(Tensor& self, const Scalar& other) {
	return at::mul_out(self, wrapped_scalar_tensor(other), self); // redispatch!
	}

pytorch/aten/src/ATen/native/BinaryOps.cpp

Lines 897 to 899 in c6f435b

	Tensor& div_(Tensor& self, const Scalar& other) {
	return self.div_(wrapped_scalar_tensor(other)); // redispatch!
	}

And, after I rewrite the mul_.Scalar like div_.scalar, it works!

Versions

PyTorch version: 2.2.0a0+git0d669f0
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.5 LTS (aarch64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: Could not collect
CMake version: version 3.20.5
Libc version: glibc-2.31

Python version: 3.9.17 (main, Jul 5 2023, 20:44:37) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-4.19.90-2107.6.0.0098.oe1.bclinux.aarch64-aarch64-with-glibc2.31
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

cc @malfet @mruberry @rgommers

[jon-chuang]

Does this occur without np?

[malfet]

I suspect root of the problem is the same as #113037 ....

[malfet]

Does this occur without np?

I suspect that it would not, because some of the assumption we made in PyTorch about what constitute a scalar (or tensor-wrapped-scalar) are wrong for empty tensors created from numpy

[alanhe151220037]

Does this occur without np?

No，only with np

[malfet]

Yes, this might not look the same, but this is a duplicate of #113037 as it crashes because iter.is_scalar(2) returns true on an empty tensor

[malfet]

Closing as duplicate of #113037 (also it is already fixed in nightlies)