Update llama.py

This commit is contained in:
Daniel Han-Chen 2024-02-02 20:24:44 +11:00
commit e748078095

View file

@ -752,7 +752,7 @@ def LlamaForCausalLM_fast_forward(
if past_key_value is not None and \
hasattr(self.model.layers[0].self_attn, "paged_attention"):
print(1)
outputs = LlamaModel_fast_forward_inference(
self.model,
input_ids,