TensorIteratorBase::is_scalar return false for empty numpy tensors, but true for empty Torch ones

[malfet]

🐛 Describe the bug

Following example causes null pointer dereference at

pytorch/aten/src/ATen/native/cpu/BinaryOpsKernel.cpp

Line 229 in e396687

opmath_t b = iter.original_scalar_value<opmath_t>(2);

which was introduced by #98820

import torch
import numpy as np

args = {
    'input': torch.from_numpy(np.array(np.random.uniform(-1000_000, 1000_000, [8, 0])).astype('float16')),
    'other': torch.from_numpy(np.array(np.random.uniform(-1000_000, 1000_000, [0, 8, 0])).astype('float16')),
    'out': torch.from_numpy(np.array(np.random.uniform(-1000_000, 1000_000, [0])).astype('float32')),
    'rounding_mode': 'floor'
}
torch.div(**args)

Versions

2.1

cc @ezyang @gchanan @zou3519 @kadeng @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10

[malfet]

As @tringwald already commented is is due to a slightly odd definition of TensorIterator::is_scalar which returns true for empty numpy tensors

pytorch/aten/src/ATen/TensorIterator.cpp

Lines 797 to 805 in e396687

	bool TensorIteratorBase::is_scalar(int arg) const {
	const auto& stride = operands_[arg].stride_bytes;
	for (const auto i : c10::irange(ndim())) {
	if (stride[i] != 0 && shape_[i] != 1) {
	return false;
	}
	}
	return true;
	}

[mingfeima]

@jiayisunx could you please take a look at this one ?

As @tringwald already commented is is due to a slightly odd definition of TensorIterator::is_scalar which returns true for empty numpy tensors

pytorch/aten/src/ATen/TensorIterator.cpp

Lines 797 to 805 in e396687

	bool TensorIteratorBase::is_scalar(int arg) const {
	const auto& stride = operands_[arg].stride_bytes;
	for (const auto i : c10::irange(ndim())) {
	if (stride[i] != 0 && shape_[i] != 1) {
	return false;
	}
	}
	return true;
	}

might be ndim() return 0 on numpy empty tensor.

[tringwald]

might be ndim() return 0 on numpy empty tensor.

ndim does not seem to be the problem. However, when initializing a tensor from an empty numpy array, all strides seem to be 0. This is not the case when directly creating an empty torch tensor.

>>> a = torch.from_numpy(np.ones((0, 0, 0)))
>>> a.stride()
(0, 0, 0)
>>> a.shape
torch.Size([0, 0, 0])
>>> a.ndim
3
>>> b = torch.ones(0, 0, 0)
>>> b.stride()
(1, 1, 1)

Due to this behavior, the stride[i] != 0 && shape_[i] != 1 check will always we false, as stride[i] == 0.

[malfet]

Very easy fix is to add data_ptr(arg) != null:

diff --git a/aten/src/ATen/TensorIterator.cpp b/aten/src/ATen/TensorIterator.cpp
index b42188858f6..fb7edeb5cfd 100644
--- a/aten/src/ATen/TensorIterator.cpp
+++ b/aten/src/ATen/TensorIterator.cpp
@@ -801,7 +801,7 @@ bool TensorIteratorBase::is_scalar(int arg) const {
       return false;
     }
   }
-  return true;
+  return data_ptr(arg) != nullptr;
 }
 
 bool TensorIteratorBase::is_cpu_scalar(int arg) const {

[atalman]

closing this task since included in 2.1.2 and validated

[malfet]

I want to have a proper fix, when we don't even dispatch to the implementation if number of elements is 0

[atalman]

Validated with final rc. No failure

[atalman]

Removing from 2.2.0 milestone

[malfet]

Unassigning myself, changing title and removing crash and high priority, as underlying crash was fixed by #115183 but still question remains what is scalar from TensorIterator point of view and why native kernels are even called ones structured kernel execution on an empty tensors has finished