llama.cpp b10210 fixes speculative decoding token replay bug

llama.cpp, the community-driven C++ implementation of LLM inference from ggml-org, continues its rapid release cadence with build b10210. With 122,000 GitHub stars and 21,200 forks, it remains the go-to solution for running large language models locally on consumer hardware. This latest release focuses on a subtle but important correctness fix: "server: correct accepted tokens when need draft token replay" (commit #26320), alongside a related spec refinement for naming consistency. The change addresses a bug in speculative decoding where the server would incorrectly count accepted tokens during draft token replay, potentially leading to sampling errors and degraded output quality.

Speculative decoding is a key optimization technique that uses a smaller draft model to predict multiple tokens in parallel, while the larger target model verifies them. When a draft needs to be replayed, the token acceptance count must be exact—otherwise, the generation loop can diverge, producing incorrect or inconsistent text. This patch ensures that acceptance counts are correctly tracked, which is especially critical for server deployments that handle many concurrent requests. The release also includes prebuilt binaries for an extensive array of platforms and accelerators: macOS Apple Silicon (with optional KleidiAI), Intel, iOS, Linux, Android, and Windows. Backends range from CPU-only to Vulkan, ROCm 7.2, OpenVINO, SYCL, CUDA 12/13, and even Ascend NPUs on openEuler. For developers running llama.cpp as a self-hosted inference service, this update is a low-risk but meaningful quality-of-life improvement.