Update modeling_llama.py

modeling_llama.py

@@ -1102,8 +1102,8 @@ class LlamaModel(LlamaPreTrainedModel):
                 return None
 
         dtype, device = input_tensor.dtype, input_tensor.device
-        #min_dtype = torch.finfo(dtype).min
-        min_dtype = -10000
+        min_dtype = torch.finfo(dtype).min
+        #min_dtype = -10000
         sequence_length = input_tensor.shape[1]
         if hasattr(
             getattr(self.layers[0], "self_attn", {}), "past_key_value"