Sarathi

Question 1

What percent of compute time during LLM inference is spent on attention?

Answer 1

5-10%

Question 2

Ffn_ln1 stands for

Answer 2

Feed forward network layer normalization 1