Function max_replicas 

pub fn max_replicas(model_bytes: u64, worker_vram: &[u64]) -> usize

Upper bound on replicas the pool could form for this model (VRAM-limited): ⌊ΣVRAM / model⌋. The device-count/layer cap can lower the achievable count below this; plan_replicas with ReplicaPolicy::Throughput returns the actual achievable number.