
The open-source inference project llama.cpp has published pre-release b12470 on GitHub, delivering a targeted backend fix for Intel SYCL execution alongside its multi-platform binary build matrix. Tagged from commit dde9de8 and published on November 17, 2026, the automated release centers on pull request #31536, which resolves a SYCL Flash Attention error involving oneAPI Math Kernel Library (MKL) mask broadcasting over attention heads or batches.
What Changed in b12470
According to the b12470 release notes, the primary code change fixes an issue in the SYCL backend where Flash Attention operations using MKL failed when broadcasting masks across heads or batches. During development of the pull request, temporary changes adding a cache-ram option to the helper scripts start-svr.sh and win-start-svr.bat were removed and deferred to a separate pull request, keeping this build focused on the SYCL Flash Attention fix.
Supported Platforms and Build Status
As a rolling pre-release build, b12470 ships with build provenance attestations and precompiled packages for Linux, Windows, macOS, iOS, and Android. Listed targets span x64 and arm64 CPUs, CUDA 12 and CUDA 13, Vulkan, ROCm 10.0, OpenVINO, SYCL FP32 and FP16 on Linux, Windows SYCL, IBM s390x including zDNN on Ubuntu 26.04, and Snapdragon CPU, Adreno GPU, and Hexagon NPU configurations. The release page notes that macOS Apple Silicon builds with KleidiAI enabled and openEuler builds for Ascend 310p and 910b hardware are marked disabled in this build.
Why It Matters and Limitations
Because Flash Attention is widely used to reduce memory overhead and improve throughput during long-context inference, mask-broadcasting bugs in specific backends such as SYCL can break model execution or produce runtime errors on Intel GPU and accelerator setups. However, b12470 is marked as a GitHub pre-release, and the release notes do not include throughput benchmarks, accuracy evaluations, or hardware validation reports. Users relying on SYCL FP16 or FP32 builds should verify behavior on their own hardware before deploying the build in production workflows.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- ggml-org/llama.cpp ↗View project details
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-17.



