diff --git a/exllama.py b/exllama.py index c9b7c05..4e2f96e 100644 --- a/exllama.py +++ b/exllama.py @@ -112,12 +112,12 @@ class Generator: input = model.tokenizer.encode(text) stop_conditions = [model.tokenizer.eos_token_id] - if stop_on_newline: - stop_conditions.append(model.tokenizer.newline_token_id) - if not max_new_tokens: max_new_tokens = model.config.max_seq_len - input.shape[-1] + if stop_on_newline: + stop_conditions.append(model.tokenizer.newline_token_id) + model.generator.set_stop_conditions(stop_conditions) random.seed(seed)