elyza

/

Llama-3-ELYZA-JP-8B

like 72

I understood that the reason for the faster inference of the 70B model is the Speculative Decoding mentioned in your blog post.
However, what is the reason why elyza/Llama-3-ELYZA-JP-8B is faster than the original LLama3-8B-Instruct?