KV Cache Churn
kv_cache_churn targets the mutable memory-management path behind KV caching.
It performs sparse reads and updates over synthetic pages with changing offsets,
which is useful for investigating the cache pressure created by ragged requests.
| Area | What the workload exercises |
|---|---|
| Cache layout | Mutable synthetic pages sized from free VRAM. |
| Access pattern | Pseudo-random page selection, sparse reads, and token-like appends. |
| Verification | --verify detects invalid cache entries after the update phase. |
| Result | Cache updates per second, reported as cache-updates/s. |
This test does not emulate a specific runtime allocator, eviction policy, request queue, or model architecture. Use it as a controlled cache and memory stress test alongside an end-to-end serving benchmark.