llama.cpp b9995 tunes SYCL flash attention for Intel Battlemage GPUs, doubling the vector thread count from 128 to 256 on xe2-class hardware.
What changed in b9995:
- SYCL thread tuning:
fattn_vec_nthreadsset to 256 for Battlemage and Lunar Lake (xe2), default stays 128 for ARC Alchemist (#25205) - CUDA 13.3 DLLs: Windows x64 CUDA 13 build ships with CUDA 13.3 libraries; CUDA 12 build uses 12.4
- ROCm 7.2: Ubuntu x64 ROCm build updated to version 7.2
- KleidiAI macOS build: disabled this release
- openEuler default build: disabled, but Ascend 310p and 910b (ACL Graph) builds available for both x86 and aarch64
The Battlemage fix targets Intel’s xe2 GPU architecture — the same silicon family as Lunar Lake integrated graphics. ARC Alchemist (xe1) keeps the original 128-thread default, so existing Intel GPU users see no regression.
Build coverage stays broad: 30+ prebuilt artifacts across macOS, Linux, Windows, Android, and openEuler. Notable entries include Windows arm64 with OpenCL Adreno support (Snapdragon X laptops) and Ubuntu s390x CPU builds for mainframe Linux.
NOTE
Two build targets are disabled in b9995: macOS Apple Silicon with KleediAI and the default openEuler build. If you relied on either, stick with b9994 or build from source.
Available now on the b9995 release page. The SYCL Battlemage fix is PR #25205.