b8957
Reverting to -lm linking solves CUDA compilation break for GGML_STATIC users.
The latest b8957 release of llama.cpp, a popular open-source project for running LLMs locally, addresses a critical build regression. Authored by Hugging Face's Adrien Gallouët, the commit reverts a recent change from using find_library(MATH_LIBRARY m) back to the standard -lm linking approach. This revert fixes a CUDA compilation break when GGML_STATIC is enabled, which occurred after the find_library method was introduced. The commit notes that no valid use case was found for preferring find_library over -lm, and it invites discussion if there's a legitimate reason to keep the former.
The release also includes updated binary builds for multiple platforms: macOS (Apple Silicon and Intel, with KleidiAI-enabled variants), Linux (Ubuntu x64, arm64, s390x, with Vulkan, ROCm 7.2, OpenVINO, SYCL support), Windows (x64 and arm64 CPU, CUDA 12 and 13, Vulkan, SYCL, HIP), Android (arm64 CPU), iOS (XCFramework), and openEuler (x86 and aarch64 with ACL Graph). This ensures broad compatibility for developers running LLMs like Llama, Mistral, and GPT-2 locally on diverse hardware.
- Reverts find_library(MATH_LIBRARY m) to standard -lm linking to fix CUDA static compilation
- Authored by Hugging Face's Adrien Gallouët, includes signed commit with GPG verification
- Updates builds for 20+ platforms including macOS, Linux, Windows, Android, iOS, and openEuler
Why It Matters
Fixes CUDA static linking for local LLM inference, ensuring developers can compile llama.cpp on GPU systems.