llama.cpp b9871 fixes CPU concat bug for quantized models

ggml-org has released llama.cpp version b9871, a maintenance update that addresses a critical bug in the CPU concatenation implementation for quantized types. The issue affected ggml's tensor concatenation operation when applied to quantized data types—commonly used to reduce memory footprint and speed up inference of large language models on CPU. Without this fix, concatenation could produce incorrect results, potentially breaking model inference pipelines that rely on tensor merging. The patch was co-authored by Stanisław Szymczyk and includes new test coverage to prevent regressions.

The release is accompanied by pre-built binaries for a wide range of platforms and backends: macOS (Apple Silicon and Intel), Linux (x64, arm64, s390x) with Vulkan, ROCm 7.2, OpenVINO, and SYCL support; Windows (x64, arm64) with CUDA 12/13, Vulkan, OpenCL, and HIP; plus Android arm64 and iOS XCFramework. This broad support underscores llama.cpp's role as the go-to library for running LLMs locally. The fix ensures that users leveraging quantized models (e.g., Q4_K_M, Q8_0) on CPU will experience reliable tensor operations, which is vital for applications like chat, code generation, and RAG pipelines running entirely on-device.