I'm curious that why MobileLLaMA 2.7B is faster then OpenLLaMA reported in v1 paper?

[XiaohuJoshua]

Thanks for the impressive work.
Have I overlooked any modifications to MobileLLaMA's architecture in the v1 paper that could enhance its inference speed? It appears that LDP is not applied to the LLM component.

[huyiming2018]

You can take a look at Section 3.3 of the MobileVLM paper; in fact, the model size of MobileLLaMa is smaller than that of OpenLLaMa.

[XiaohuJoshua]

So, it is because MobileVLM 2.7B has 0.3B fewer parameters than OpenLLaMA 3B? Thanks for your reply.