Hi,
Usually with different sequence length you can pad all inputs to become the same length. After padding a sequence, if you are using an torch.nn RNN block such as LSTM() or GRU(), you can use pack_padded_sequence to feed in a padded input.
Otherwise you could create batches according to the length of the 0th dimension, like you said, but that might be inefficient. And using torch.utils.data.Dataset and torch.utils.data.DataLoader won’t be possible.