How to use xlstm when the context length is not fixed? Or must it be fixed?

[alexdemartos]

The context_length is used for the causal mask in the mLSTM block. By looking at the code here:

    if lower_triangular_matrix is None or S < lower_triangular_matrix.size(-1):
        ltr = torch.tril(torch.ones((S, S), dtype=torch.bool, device=_device))

you can just set context_length to your maximum sequence length.