KV cache

Knows KV cache avoids recomputing past tokens during generation; understands it uses memory.

Explains what keys and values are cached; calculates memory footprint for a given model and context length.

Applies paged attention (vLLM), prefix caching, and sliding-window cache; tunes batch size vs cache tradeoff.

Implements custom caching policies; understands GQA/MQA impact on cache size; analyses decode throughput.

Contributes to KV compression, speculative decoding, or sub-quadratic attention with persistent state.