TL;DR: llama.cpp release b10754 patches out-of-bounds reads in Adreno OpenCL kernels, improving stability for on-device inference.
Summary: The llama.cpp project released b10754 with OpenCL fixes targeting Adreno image kernels, addressing out-of-bounds reads that could cause crashes or corruption. Key changes include clamping the q4_K decode GEMV fetch row on padded x-grids and enforcing the tiling contract for KQ/KQV image GEMMs. The update is part of the ongoing maintenance of the OpenCL backend for Qualcomm GPUs.
Why it matters: Builders deploying local LLMs on Android or Adreno-based hardware gain more reliable inference after this release. Update llama.cpp and re-run OpenCL workloads to avoid potential memory safety issues on these GPUs.
Source: github_releases