reflect use_parallel_residual to mlp_after_attn

deepspeed/module_inject/containers/gptneox.py

@@ -22,6 +22,7 @@ def __init__(self, **kwargs):
         super().__init__(**kwargs)
 
         # All model specific things should be defined here instead of the base class.
+        self.mlp_after_attn = not getattr(self.model_config, "use_parallel_residual", not self.mlp_after_attn)
 
     def create_module(self, config=None):
         _config = config if config is not None else self.ds_model_config