RuntimeError when trying to reuse a buffer #21469
Description
🐛 Bug
The code snippet below throws error. I'm trying to see if I could reuse a buffer.
To Reproduce
Steps to reproduce the behavior:
b = torch.randn(3, requires_grad=True)
c = torch.zeros(3)
c[[1,2]] = b[[1,1]]
c.sum().backward() # ok
c.sum().backward() # fail
RuntimeError: ntensor >= 3 INTERNAL ASSERT FAILED at caffe2/aten/src/ATen/native/cpu/IndexKernel.cpp:51, please report a bug to PyTorch.