KV 캐시 최적화

In a transformer, each new token attends to all previous tokens via attention's keys (K) and values (V). Recomputing K and V for the whole sequence at every step would be quadratic and wasteful, so models cache them: the KV cache. The downside is size. The cache grows linearly with sequence length, batch size, layers, and heads, so a long-context request can consume more GPU memory than the model weights themselves. Optimization tackles this from several angles: paged memory (vLLM's PagedAttention) stores the cache in non-contiguous blocks to reduce fragmentation and enable sharing; quantization stores K and V in 8-bit or 4-bit; and architectural changes like Grouped-Query Attention (GQA) and Multi-Query Attention (MQA) let many query heads share fewer key/value heads, slashing cache size at the source.