LLM inference

A synthetic model has 40 layers, 40 query heads, 8 KV heads, and $d_{\text{head}} = 128$. The KV cache is stored in BF16 (2 bytes). How many bytes of KV cache does one token of one sequence need?