when using auto_insert_landmarks, it appears to be view problems

[EzTan1]

as mem_freq = 63, and seq_length 3535 can't divide 63, view[input_ids.shape[0], -1, mem_freq] would raise error?

[mkrima]

yes, when using auto_insert_landmarks and triton kernel during training, sequence length should be divisible by mem_freq.