llama.cpp Vulkan backend adds Intel Xe flash attention kernels

llama.cpp has gained Intel Xe-specific flash attention kernels in its Vulkan backend, aimed at speeding up long-context inference on newer Intel integrated and discrete GPUs.

The change covers Xe-LPG Plus, Xe2, and Xe3 hardware. Contributor virajwad reported that the work should improve long-context token generation substantially, sharing benchmarks at a 16,384-token context depth that place the new kernels ahead of both current master and an earlier multi-shader approach.

Flash attention is a core path for efficient transformer inference at long sequence lengths. Hardware-tuned kernels matter because generic Vulkan paths often leave performance on the table on vendor-specific GPU architectures. The addition continues llama.cpp's pattern of layering backend-specific optimizations so local LLM runs can better exploit the GPU already in a machine, rather than depending only on portable defaults.