How does the shared_experts work when using EPMoE in DeepSeekV2?

[ly338]

Is the shared_expert configured with TP when Using EPMoE in DeepSeekV2?
In the code, The all_reduce operation in DeepseekV2MLP is disabled with reduce_results flag set to False like below,
self.shared_experts = DeepseekV2MLP(
hidden_size=config.hidden_size,
intermediate_size=intermediate_size,
hidden_act=config.hidden_act,
quant_config=quant_config,
reduce_results=False,
)
So how to get the input data for EPMoE layer?