pub fn cache_thrashing() -> boolExpand description
Whether the cache is THRASHING: the f32 working set does not fit the budget, so entries are evicted before they are ever reused and every lookup pays a full re-dequant.
True once cumulative evictions exceed the budget itself, i.e. we have churned more than one whole cache — a model whose f32 form fits never trips this.
Callers use it to route around the cache. Measured on Muse-Glimmer-30B
UD-Q4_K_XL (27.85B params ⇒ ~111 GB of f32 against a 15 GB budget): a
decode-shaped m = 1 forward took 43.1 s cached-and-thrashing vs 0.76 s on
the fused int8 path — 57x. Repeat passes showed 43.1 / 43.2 / 43.5 s, i.e.
the cache never paid for itself at all.