llama.cpp Release b10867 Disables Lazy Tensor Loading on iGPUs
llama.cpp version b10867 updates the default tensor loading behavior to disable lazy loading on integrated GPUs (iGPUs) when using automatic mode. This release includes pull request #28326 to fix a performance regression observed on iGPU hardware. Integrated GPUs share system memory with the CPU, where lazy tensor loading can introduce severe page faults or memory latency overhead. Disabling lazy mode by default on iGPUs ensures more stable performance and reliable throughput for users running local LLMs on integrated graphics hardware. The update specifically adjusts the automatic lazy tensor setting (`LLAMA_LAZY_MODE_AUTO`) so that lazy loading is disabled on iGPUs while preserving other memory-mapping capabilities. Updated binary builds were released across macOS, Windows, Linux, Android, and backends including CUDA 12/13, Vulkan, OpenVINO, ROCm, and SYCL.
## BACKGROUND
llama.cpp is a popular open-source C/C++ inference framework designed to run Large Language Models (LLMs) efficiently across varied hardware architectures. Lazy tensor loading (`--tensor-read-lazy`) is a feature designed to defer reading certain model weight tensors from memory-mapped files until execution, saving RAM at the expense of potential memory access overhead.