fix

modeling_gptbert.py

@@ -306,7 +306,7 @@ class SelfAttention(nn.Module):
         self.d_qk = config.query_key_head_size
         self.d_v = config.value_head_size
         self.num_attention_heads = config.num_attention_heads
-        self.num_kv_heads = config.num_kv_heads
+        self.num_kv_heads = config.num_attention_heads
         self.hidden_size = config.hidden_size
 
         self.q_out_dim = self.d_qk * self.num_attention_heads