INTERNAL ASSERT FAILED source.dtype() == self.dtype()

[creakseek]

🐛 Describe the bug

RuntimeError: source.dtype() == self.dtype() INTERNAL ASSERT FAILED , please report a bug to PyTorch.

import torch

input_ = [
    [
        [1, 2, 3],
        [4, 5, 6],
        [7, 8, 9]
    ],
    [
        [11, 12, 13],
        [14, 15, 16],
        [17, 18, 19]
    ]
]

input_ = torch.tensor(input_, dtype=torch.float32)

input_.set_(torch.tensor(
    [
        [
            [21, 22, 23],
            [24, 25, 26],
            [27, 28, 29]
        ],
        [
            [31, 32, 33],
            [34, 35, 36],
            [37, 38, 39]
        ]
    ]))

print(input_)

RuntimeError: source.dtype() == self.dtype() INTERNAL ASSERT FAILED at "/opt/conda/conda-bld/pytorch_1659484806139/work/torch/csrc/autograd/generated/python_variable_methods.cpp":16766, please report a bug to PyTorch. 

Versions

PyTorch 1.12.1

cc @ezyang @gchanan @zou3519

[bdhirsh]

Hi-pri for the internal assert failure.

In this repro, it looks like we're running float32_tensor.set_(int64_tensor). I'm not sure if that's actually supported - if it's not, we should make it a normal assert failure.

[malfet]

@ezyang : internal asserts should be high priority only if we don't know what is causing them.
We should change internal assert to a regular runtime error as this is not a supported behaviour.

[fool1280]

Hi, I'm interested in this issue, but not familiar with the codebase a lot (first time contributor --ever here). May I ask for some pointer/advice to start with? Thank you!

[ezyang]

A good start is to find the assert, change it into TORCH_CHECK and add a good error message

[fool1280]

Thank you, please let me take this issue

[PanosAntoniadis]

Hey! I submitted a PR where I replaced the internal error with a TORCH_CHECK! It's my first contribution.