Mutation after tensor.expand returns wrong result.

[ysiraichi]

🐛 Bug

The following code mutates the value of the expanded tensor.

import torch
import torch_xla.core.xla_model as xm

device = "cpu"
x = torch.arange(6, device=device)
y = x.expand(2, *x.shape)
y[0, 0] = 5

print(y)

# CPU
tensor([[5, 1, 2, 3, 4, 5],
        [5, 1, 2, 3, 4, 5]])

# XLA Eager
tensor([[ 5,  2,  4,  6,  8, 10],
        [ 5,  2,  4,  6,  8, 10]], device='xla:0')

Environment

• PyTorch version: 63fc48257a02f8e28b79d13def7a7139589d4176 (Nov 2)
• PyTorch/XLA version: d5d023063bfa8ecb4629f621f9b5890bc8396f58 (Nov 9)

[JackCaoG]

yea... this is indeed incorrect. Hopefully if you dump the HLO it become obvious what happened

[ysiraichi]

I think this might be the same issue as the non-XLA version: pytorch/pytorch#114302

[JackCaoG]

oh nice!