llama.cpp b10455 adds SYCL support for AdamW/SGD optimizers

The llama.cpp project, led by ggml-org, just pushed release b10455, a minor but notable update that expands its hardware reach. The headline change is SYCL backend support for OPT_STEP_ADAMW and OPT_STEP_SGD, two optimizer operations critical for training and fine-tuning neural networks. Previously, these operations were only available on backends like CUDA or ROCm, meaning Intel GPU users couldn't run training loops locally through llama.cpp. With this patch, co-authored with Intel's Neo Zhang Jianyu, developers can now execute AdamW and SGD update steps on Intel GPUs via SYCL, opening the door to on-device fine-tuning without requiring an NVIDIA card.

The release also includes a broad set of prebuilt binaries: macOS (Apple Silicon, Intel), iOS, Linux (x64, arm64, s390x, with Vulkan, ROCm, OpenVINO, and SYCL variants), Android (arm64, OpenCL Adreno), Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, ROCm), and even openEuler with Ascend ACL Graph support. This makes b10455 one of the most hardware-inclusive llama.cpp releases to date. While the change targets developers rather than end users, it's a sign that the open-source LLM runtime is steadily closing the gap with commercial frameworks by supporting training tasks across more accelerators. Expect further SYCL optimizations as Intel continues contributing to the project.