llama.cpp Release b10926 Fixes SYCL Handling for Unsupported tq1_0 Quantization
Open-source LLM inference engine llama.cpp has released version b10926, which incorporates pull request #28681. This patch ensures that the SYCL backend gracefully handles unsupported tq1_0 quantizations and fails safely instead of crashing unexpected runtime processes. This update enhances runtime stability for developers and users running llama.cpp on hardware accelerated via SYCL, such as Intel GPUs. Graceful failure prevention helps maintain application integrity when loading incompatible model weights. The patch specifically targets the `tq1_0` ternarization format (~1.69 bits per weight), which lacks native kernel support in SYCL. Pre-built binaries for release b10926 have been published across multiple platforms, including Linux, Windows, macOS, Android, and various hardware acceleration backends.
## BACKGROUND
SYCL is an open, cross-platform C++ programming model designed to enable code reuse across diverse hardware accelerators like GPUs and FPGAs. In llama.cpp, quantization formats such as `tq1_0` reduce model weights to extremely low precision to shrink memory usage and enable efficient LLM execution on consumer devices.