Speculative Decoding on Android

Implementing speculative decoding on-device using a small draft model (0.5B) paired with a larger target model (8B), covering the parallel verification algorithm, KV-cache sharing between models, rejection sampling mechanics, memory-mapped model loading to fit both models in RAM, and Android-specific NDK integration with llama.cpp's speculative decoding API — with real benchmarks showing tokens-per-second gains on Snapdragon 8 Gen 3