June 2026 performance snapshot for ggml-org/llama.cpp focused on SYCL backend enhancements with an emphasis on documentation clarity, defaults stabilization, and FP16 performance. Implemented default FP16 behavior to improve runtime performance and user guidance, and tightened build options for FP16/FP32 operations. Documentation and onboarding were improved by clarifying SYCL behavior, correcting the GGML_SYCL_GRAPH default to ON, and relocating Docker-specific guidance to the Docker documentation. Updated backend references to reflect Intel SYCL support and the 2026.02 state, removing outdated Nvidia references. These changes reduce onboarding friction, improve performance consistency across SYCL-enabled runs, and streamline maintenance for the project.

May 2026 monthly summary for vllm-gaudi focused on stabilizing large-model loading under expert parallelism and improving deployment reliability for GPT-OSS-120B. The month's primary accomplishment was fixing a tensor-size handling bug in the GPT-OSS-120B weight loading path when using expert parallelism, preventing runtime load failures and ensuring correct intermediate-size copying instead of TP-sized slicing. This work reduces risk in production deployments and enables scalable usage of 120B models under expert parallelism with --enable-expert-parallel.