Fixed dumb shit 2

This commit is contained in:
Андреев Григорий 2026-03-29 14:13:20 +03:00
parent a0edbb6659
commit 11b8483b63

View File

@ -141,7 +141,7 @@ def generate_llm_pieces(bundle: ModelBundle, messages: list) -> Iterable[str]:
token_text = bundle.tokenizer.decode([token_id], skip_special_tokens=True)
if token_text:
yield token_text
-
past_key_values = outputs.past_key_values
input_ids = next_token_id.unsqueeze(1)
if attention_mask is not None: