Same model Llama 7B, why does zero3 initialize different parameter sizes？

When I use Zero3, in initializing the network, if my Llama is rewritten by inheritance as follows:

class FlashLlamaModel(LlamaModel):
   def __init__(self, config: LlamaConfig):
       super().__init__(config) 

class FlashLlamaForCausalLM(LlamaForCausalLM):
   def __init__(self, config):
       super().__init__(config)
       self.model = FlashLlamaModel(config) 

actor_model = create_hf_model(
           model_class= FlashLlamaForCausalLM,  
           model_name_or_path=actor_model_name_or_path,
           tokenizer=self.tokenizer,
           ds_config=ds_config, 
           disable_dropout=self.args.disable_actor_dropout,
           debug=debug)

Log output: initialize about 14B parameters.
'[2023-09-27 01:48:00,149] [INFO] [partition_parameters.py:454:__exit__] finished initializing model with 13.63B parameters'

But when not overriding the parent self.model, the log output goes back to 7B.

class FlashLlamaModel(LlamaModel):
    def __init__(self, config: LlamaConfig):
        super().__init__(config) 

class FlashLlamaForCausalLM(LlamaForCausalLM):
    def __init__(self, config):
        super().__init__(config) 

actor_model = create_hf_model(
            model_class= FlashLlamaForCausalLM,  
            model_name_or_path=actor_model_name_or_path,
            tokenizer=self.tokenizer,
            ds_config=ds_config, 
            disable_dropout=self.args.disable_actor_dropout,
            debug=debug)

Log output: initialize about 7B parameters.
[2023-09-27 01:52:23,475] [INFO] [partition_parameters.py:454:__exit__] finished initializing model with 6.93B parameters

Start at create_hf_model and compare the two FlashLlamaForCausalLM constructors, focusing on the explicit self.model assignment and the Zero3 initialization logs. Trace how each model instance is registered and verify what explains the change from about 7B to 14B parameters; done means the cause is documented or a focused fix is identified.