Llama.cpp b10056 adds OpenCL ABS op for cross-GPU acceleration
New release supports AMD, Intel GPUs with ABS operation for better performance
ggml-org has released version b10056 of llama.cpp, the popular open-source C++ library for running LLMs locally on consumer hardware. The key change in this release is the addition of the ABS operation to the OpenCL backend, as detailed in PR #25115. This operation (absolute value) is a fundamental building block for many neural network layers, and its support means that models using abs activations or other abs-based operations can now be offloaded to OpenCL-compatible GPUs without falling back to CPU. While a small addition, it broadens the range of models and operators that can benefit from GPU acceleration on non-NVIDIA hardware.
This release ships with an extensive set of prebuilt binaries across platforms. For Windows, users can grab builds for CPU, OpenCL (Adreno), CUDA 12/13, Vulkan, OpenVINO, SYCL, and HIP. Linux users get Ubuntu packages for x64 and arm64 with CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL. macOS and iOS binaries support Apple Silicon (arm64) and Intel (x64), with a KleidiAI-enabled variant. Android arm64 and even openEuler builds are included. This makes llama.cpp one of the most accessible tools for running LLMs on virtually any system, from mobile phones to high-end servers.
- Adds absolute value (ABS) operation to OpenCL backend for broader GPU compatibility
- Prebuilt binaries for Windows, Linux, macOS, iOS, Android, and openEuler across multiple backends
- Enables local LLM inference on AMD, Intel, and Qualcomm GPUs via OpenCL, Vulkan, and ROCm
Why It Matters
Expands local LLM deployment beyond NVIDIA GPUs, democratizing AI for AMD and Intel hardware users.