llama.cpp b10256 speeds up SYCL prompt processing 9.4% on Intel GPUs

The llama.cpp maintainers (ggml-org) have shipped release b10256, an incremental but impactful update for users running LLMs on Intel GPUs via the SYCL backend. The headline change comes from a commit that parallelizes the non-contiguous concat kernel. Previously, this kernel launched a single-lane work-group of (1,1,1), severely underutilizing the GPU's compute resources. The fix now launches a (1, 1, SYCL_CONCAT_BLOCK_SIZE) work-group, where SYCL_CONCAT_BLOCK_SIZE is defined in ggml/src/ggml-sycl/presets.hpp. This simple geometry change unlocks significant parallel throughput for concatenation operations, which are common in transformer models with concatenated KV caches or embeddings.

Benchmarks show the impact clearly: using llama-bench with an Intel Arc Pro B70 and a Qwen3.6-27B-UD-Q4_K_XL model (quantized to 4-bit), with flash attention enabled and a q8_0 KV cache, prompt processing at 2048 tokens (pp2048) improved from 920 to 1006 tokens per second — a 9.4% gain. The developers also capped the concat block size at ne0 and briefly introduced an environment variable (GGML_SYCL_CONCAT_BLOCK_SIZE) for tuning, but that configuration knob was later reverted to keep the change simple. The release includes prebuilt binaries across macOS (Apple Silicon, Intel), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CUDA 12/13, Vulkan, SYCL, OpenCL), and Android. For developers running local LLM inference on non-NVIDIA hardware, this update makes prompt evaluation noticeably faster without changing model compatibility — a practical win in llama.cpp's ongoing effort to optimize every backend it supports.