⏱ 7 min read The transformer internals interviewers push on, from why attention scales by sqrt(d_k) to KV cache memory and grouped-query attention. Read article