llama.cpp b9263 merges HunyuanOCR into HunyuanVL for improved vision precision

The latest release (b9263) of the popular local LLM inference engine llama.cpp introduces a significant architectural consolidation: HunyuanOCR is now merged directly into HunyuanVL. Previously, HunyuanOCR shared the same Hugging Face architecture and vision layout as HunyuanVL but was implemented as a separate code path that omitted the +0.1 bilinear sampler used by the original reference implementation. This oversight led to reduced OCR precision in vision tasks. By collapsing OCR into the HUNYUANVL projector and HUNYUAN_VL text architecture, the fix ensures consistent application of the bilinear sampler, aligning output quality with the upstream model.

This release also reflects llama.cpp's broad platform support: it's distributed as pre-built binaries for macOS (Apple Silicon with optional KleidiAI acceleration, Intel x64, iOS XCFramework), Linux (x64/ARM/s390x with Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16 backends), Windows (x64/ARM64 with CPU, CUDA 12/13, Vulkan, SYCL, HIP), and Android ARM64. For enterprise users, openEuler builds are also available. The consolidation reduces code complexity and improves maintainability, making it easier for developers to deploy vision-language models with accurate OCR capabilities on local hardware.