Update llama.py

This commit is contained in:
Daniel Han 2025-03-21 17:26:20 -07:00
commit 5ba087859a

View file

@ -102,6 +102,7 @@ def _fast_prepare_inputs_for_generation(self, input_ids, **kwargs,):
if "past_key_values" in kwargs:
print("FIX", input_ids.shape)
input_ids = input_ids[:,[-1]]
print("FIX AFTER", input_ids.shape)
kwargs["attention_mask"] = kwargs["attention_mask"][:,[-1]]
if "cache_position" in kwargs:
kwargs["position_ids"] = kwargs["cache_position"]