CrackML by @ml.with.umang
GenAI interview guides / KV Cache Interview Guide
GenAI / LLM

KV Cache Interview Guide

Explain KV caching, memory/latency trade-offs, long-context pressure, offloading, quantization, and prefix reuse.

Why it exists

Caching prior key/value tensors avoids recomputing them on every autoregressive decode step.

Memory trade-off

KV memory grows with context, concurrency, layers, heads, head dimension, and dtype; cache capacity can become the serving bottleneck.

Cache strategies

Dynamic, static, offloaded, quantized, and prefix-reused caches trade memory, bandwidth, latency, and implementation complexity.

Interview edge

Connect KV cache to prefill vs decode, continuous batching, eviction, isolation, and capacity planning.