According to @_avichawla, prefill is compute-bound and decode is memory-bound, shaping TTFT and ITL. Tackle KV cache growth with GQA, PagedAttention, quantization.