llama.cpp b10425 fuses gated-delta-net SYCL, boosts Arc Pro B70 by 1.2%

llama.cpp, the widely adopted C++ inference engine for LLMs, shipped release b10425 on August 14 with a targeted optimization for SYCL—Intel's cross-platform programming model. The key change, "sycl: fuse the gated-delta-net state writeback cpy," is a port of upstream PR #23940. It eliminates a separate memory-copy kernel by fusing the state writeback into the main gated-delta-net operation, reducing launch overhead on Intel GPUs. The benchmark uses a Qwen 3.6 27B model quantized to Q4_K, running on an Arc Pro B70 with 48 of 64 blocks using the gated-delta-net architecture, and flags like -ngl 99, -fa 1, and f16 cache tensors.

Measured results show a modest but consistent lift: single-sequence token generation (tg128) rose from 23.91 to 24.19 tokens/s (+1.2%), confirmed across interleaved A/B passes and a rebuild. A two-sequence workload also improved, from 32.73 to 33.11 tokens/s (+1.1%). Prompt processing (pp2048) stayed flat at around 1050–1055 tokens/s, indicating the fusion specifically targets decode-phase memory traffic. For developers deploying hybrid linear-attention models on Intel hardware, this release delivers a free, low-risk performance gain. The release also includes prebuilt binaries for macOS, Linux (CPU/Vulkan/ROCm/SYCL), Windows (CUDA/Vulkan/SYCL), and Android, making the update easy to adopt.