Estimate model-serving memory from weights, key-value cache, activations, overhead, and per-GPU usable capacity.
