Update modeling_kphi3.py

modeling_kphi3.py

@@ -53,6 +53,9 @@ from transformers.utils import (
 )
 from .configuration_kphi3 import KPhi3Config
 
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
 def get_max_acceptable_common_divisor(a, b, max_acceptable=1000000):
   """
   This is an inefficient max acceptable common divisor implementation to be improved.