Developer reduces RAM usage by 74% while running SmolLM2 on Samsung Galaxy Watch 4 by fixing llama.cpp's duplicate model loading issue

SmolLM2 360M model was being loaded twice in RAM simultaneously — once via APK mmap page cache and again through tensor allocations — peaking at 524MB for a 270MB model

Fix involved adding host_ptr to llama_model_params so CPU tensors point directly to mmap region while only Vulkan tensors get copied separately

Peak RAM usage dropped from 524MB to 142MB (74% reduction) and first boot time improved from 19s to 11s on real hardware

Second boot now achieves ~2.5s startup time thanks to mmap and KV cache optimization

Code available on GitHub at Perinban/llama.cpp with performance metrics documented via VmRSS measurements