trying to use torch compile with Tacotron model, but I always get that error (‘lengths’ argument should be a 1D CPU int64 tensor, but got 1D meta Long tensor), that fired in forward function when calling x = nn.utils.rnn.pack_padded_sequence()
even when converting input_lengths to CPU.
1 Like
I guess it is related to fake tensor tracing support. Would you please file a bug to PyTorch github with a simple reproducer?