llama.cpp b9663 adds EXPM1, FLOOR, TRUNC, ROUND operations

llama.cpp's latest release, b9663, focuses on expanding mathematical operation support for local AI inference. The key addition is the EXPM1 operation (exponential minus one) along with full unit test coverage for FLOOR, TRUNC, and ROUND functions. These operations are critical for numerical transformations in large language model pipelines, enabling more precise control over tensor computations and activation functions. Developers can now leverage these operations in their local inference workflows without relying on external libraries.

The release also emphasizes broad platform compatibility. Pre-built binaries and build targets span macOS (Apple Silicon with optional KleidiAI acceleration, Intel x64, and iOS XCFramework), Linux (x64/arm64 across CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Android (arm64), and Windows (x64 CPU, arm64 CPU, CUDA 12/13, Vulkan, SYCL, HIP). The openEuler distribution is disabled in this release. All artifacts are signed with a verified GPG key (B5690EEEBB952194). This update underscores llama.cpp's commitment to running AI models efficiently on consumer hardware, making advanced math ops accessible to a wide developer audience.