llama.cpp b10291 fixes Vulkan DeviceLost crashes with new debug tools

ggml-org's llama.cpp, the popular C++ library for running large language models locally, has released b10291, a maintenance-focused update that improves stability on Vulkan-based GPUs. The release targets 'DeviceLost' errors—a frustrating crash where the GPU disconnects from the driver, often during long inference runs. This version fixes a submission batching size issue and corrects a threshold that was applied too late, which could cause the GPU to be overwhelmed. Additionally, the team added debug tooling to gather more information when DeviceLost occurs, making it easier for developers to pinpoint the root cause rather than guessing. The update also replaces abort() calls with proper exceptions and cleans up a circular dependency, improving code robustness.

With over 123k GitHub stars and 21.4k forks, llama.cpp remains a cornerstone of local AI development. This release isn't flashy—no new model support or performance benchmarks—but it's essential for developers deploying LLMs on Vulkan across platforms. Prebuilt binaries are available for macOS (Apple Silicon and Intel, with optional KleidiAI), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CUDA 12/13, Vulkan, OpenCL Adreno), and Android (arm64 CPU). By stabilizing the Vulkan backend and giving developers better debugging tools, b10291 reduces the guesswork in GPU-related crashes, making local AI inference more dependable for production workloads.