llama.cpp b9315 release documents key on-device state limit

The open-source LLM inference engine llama.cpp has released version b9315, which includes a critical documentation update. The commit, authored by GitHub Actions and signed with a verified GPG key, adds a note to the codebase: 'Only one on-device state can be saved per sequence.' This addresses a common confusion among developers using llama.cpp on mobile or embedded devices where memory is scarce. Previously, there was ambiguity about how many cache states could be persisted across sequences, potentially leading to silent failures or resource exhaustion.

The b9315 release does not introduce new features but strengthens developer guidance for production deployments. llama.cpp, known for its efficient CPU and GPU inference of LLaMA-family models, is widely used in local AI assistants and privacy-focused chatbots. The new documentation helps developers avoid overwriting previous states unintentionally, which is especially important for scenarios like multi-turn conversations or stateful agents running on a single device. The release is available across all major platforms including macOS, iOS, Linux, Android, and Windows, with support for CPU, Vulkan, CUDA, ROCm, and other backends.