Unispac/Llama2-7B-Chat-Augmented is very slow in generation with `generate_ragged_batched` - would be nice to understand why.