Fix num_key_value_groups

pull/499/head
Aleksandr Borzunov 9 months ago
parent cac654a16f
commit cda4fe8cef

@ -21,8 +21,10 @@ class DistributedFalconConfig(DefaultRevisionMixin, FalconConfig, ClientConfig,
@property
def num_key_value_groups(self) -> int:
if self.new_decoder_architecture or not self.multi_query:
if self.new_decoder_architecture:
return self.num_attention_heads // self.num_kv_heads
if self.multi_query:
return self.num_attention_heads
return 1
@classmethod

Loading…
Cancel
Save