llama.cpp SYCL Backend Documentation

SYCL backend primarily designed for Intel GPUs. Supports: Data Center Max Series (Max 1550, 1100), Flex Series (Flex 170), Arc A-Series (A770, A730M, A750), Arc B-Series (B580), built-in Arc GPU (Meteor Lake, Arrow Lake, Lunar Lake), iGPU in Core CPU (11th Gen+). Key performance data: oneAPI 2026.1 improves SYCL build performance 3.1x (Arc B570: pp 1331 vs 434 t/s vs 2025.3). 2025.2: MUL_MAT Q4_0 optimization increased llama-2-7b.Q4_0 performance 21-87% across Intel GPUs (PVC 1550: +87%, Flex 170: +28%, Arc A770: +30%, MTL: +23%, ARL-H: +21%). 2026.04-05: Optimize mul_mat by reorder for Q4_K, Q5_K, Q6_K, Q8_0. Fused MoE. 2026.03: Flash-Attention support. 2026.02: Removed Nvidia & AMD GPU support (oneAPI plugin unavailable). 2025.11: Support malloc memory on device >4GB. 2024.11: sycl_compat for oneAPI 2025.0+. 2024.8: oneDNN as default GEMM library. 2024.5: 34->37 t/s llama-2-7b.Q4_0 on Arc A770. Known issues: SYCL_CACHE_PERSISTENT=1 causes crashes (not recommended). Default context too big causes OOM (use -c 8192 or smaller). Level Zero relaxed allocation limits needed for >4GB allocations (UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1). FP16 recommended for better performance (GGML_SYCL_F16=ON). split-mode tensor requires flash attention. Tensor parallelism optimized for 2 GPUs.