KV Cache Debugger

Configuration

Prompt tokens (1–32768)

Generated tokens (1–4096)

Batch size (1–64)

Transformer layers (1–160)

Query attention heads

KV heads

Head dimension

Data type

Current decode step

KV cache

Animation speed

1.00×

This tool models KV-cache memory and attention-score multiplications only. It does not simulate full transformer runtime, MLP, projections, softmax, memory transfer, or sampling.

Prefill & Decode Timeline

KV-Cache Shape

Charts

KV-cache memory vs sequence length

Per-step attention work ━━ cached━━ uncached

Cumulative attention work ━━ cached━━ uncached

Memory

KV caching reduces repeated compute but consumes additional memory. It never reduces KV-cache memory.

Attention-Work Comparison

Explanation

MHA vs GQA vs MQA

Holds all parameters constant except KV-head count. Attention work is unchanged; memory scales with KV heads.