The llama.cpp project released build b10017, which includes a specific update to the SYCL backend. The change increases the minimum buffer size threshold for Unified Shared Memory (USM) system allocations from 1 GiB to 4 GiB.

  • This adjustment is based on real-world experiments involving overcommitting device memory with model weights larger than available VRAM, such as running Qwen3.5-35B-A3B-Q8 on a B70.
  • A debug message was also added to better track USM system allocations.
  • The release provides binaries for macOS (Apple Silicon and Intel), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and openEuler.

The modification aims to improve stability when handling large models that exceed available VRAM on SYCL-compatible hardware.