Update configuration.py

configuration.py

@@ -12,6 +12,10 @@ class CustomConfig(PretrainedConfig):
         self.n_heads = 64
         self.n_vocab = 50257
         self.d_head = 128
+
+        # for vllm
         self.num_attention_heads = 64
         self.hidden_size = 8192
+        self.vocab_size = 50257
+        
         super().__init__(**kwargs)