Variables needed for gradient computation should be modified by an inplace operation.

[aleversn]

In the README.md, it suggests using the torch of version 1.3.0, but there seems no that version in the previous version of PyTorch, link.

So, I use the latest version (1.7.1) of the torch, and when I start training, I got this Runtime Error. And then I found that the error was caused in the prophetnet/ngram_multihead_attention.py line 255.

q = q * self.scaling

It looks like this operation is not allowed anymore, then I fixed the problem by the following:

q_ = q * self.scaling

if self.bias_k is not None:
    assert self.bias_v is not None
    k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
    v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
    q = q_.contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim).transpose(0, 1)