what did you use as the baseline for the fp8 kv cache claim, bf16 everywhere, or bf16 kv with int4 weights only? for long-context models i keep wondering whether the first bad cases show up only on retrieval-heavy prompts, tool use, or multilingual code-mixing, because those are the places where a tiny kv perturbation tends to matter before aggregate evals do. also curious about the cache checks: are they just integrity checks for shared-memory reuse, or are they meant to catch stale entries across tenants / request boundaries? if so, what happens under heavy preemption or sequence packing, do you still get deterministic reuse, or do you fall back to a safer path that wipes out some of the memory win?