b8611

The open-source project llama.cpp, maintained by ggml-org, has rolled out a substantial new release tagged b8611. This update primarily addresses a bug in thread assignment for RWKV operations (referenced as issue #21226), which should improve performance and stability for users running RWKV architecture models through the efficient C++ inference framework. The fix ensures computational workloads are properly distributed across available CPU cores.

The most notable aspect of release b8611 is its massive expansion of pre-built binary support for different hardware accelerators. The team now provides builds for Vulkan graphics APIs, AMD's ROCm 7.2 platform, Intel's OpenVINO toolkit, and SYCL for cross-architecture programming. This support spans operating systems including Windows, various Linux distributions (Ubuntu), and the openEuler OS, covering both x64 and arm64 architectures. For Windows users specifically, there are now dedicated builds for CUDA 12.4 and 13.1, Vulkan, SYCL, and even HIP for AMD GPUs.

This release underscores llama.cpp's commitment to being the most versatile and hardware-agnostic engine for local LLM inference. By broadening official support to include these backend technologies, the project reduces the friction for developers and enthusiasts who want to deploy models on their existing hardware, whether it's an Intel integrated GPU via OpenVINO, an AMD card via ROCm, or any Vulkan-compatible device. The continued iteration on performance fixes, like the RWKV thread patch, ensures the engine remains a robust and reliable choice for running state-of-the-art open-weight models efficiently on consumer hardware.