Falcon support for generation about alpaca_eval HOT 1 CLOSED

that's a known warning that has no impact on generations (see https://huggingface.co/tiiuae/falcon-40b/discussions/8 huggingface/transformers#24598 ), and the way we generate for falcon is actually the one recommended by the authors on HF. A PR was merged to avoid this warning (huggingface/transformers#24603 ) so it will disappear once transformers is updated on pypi

Besides if really necessary I would rather not reinvent the wheel for inference and use what HF has. This decreases maintenance on our side (e.g. if libraries or models change) , makes the code more readable, and ensures that we have any new feature that HF incorporates. If someone has a different inference pipeline, they can always use theirs and then use evaluate instead of evaluate_from_model

from alpaca_eval.