Browsing: AttentionAware

Long-context inference makes the KV cache one of the main costs of serving LLMs. During autoregressive decoding, the cache grows with context length, batch size, and…