Wiring iOS CoreML's Stateful Models to a Streaming Inference Pipeline

Deep dive into CoreML's MLState API for stateful LLM inference — managing key-value cache across prediction calls, handling memory warnings with graceful cache eviction, and benchmarking token throughput across A16/A17/A18 chips with different quantization tiers (4-bit vs 8-bit) to find the model size ceiling that doesn't trigger jetsam on iPhone 14 vs 15 Pro