pub const RESIDENCY_UTILISATION: f64 = 0.9;Expand description
Fraction of free VRAM a co-resident candidate is allowed to claim.
The remainder absorbs what this estimate deliberately does not model: the
compute buffer llama-server allocates per batch, allocator fragmentation,
and whatever the desktop compositor takes while the process is starting.
The same 0.9 that
recommendation sizes first-run model
suggestions against — one convention for “do not fill the card to the brim”,
not two.