TL;DR: llama.cpp now supports Deepseek V4 operations on Vulkan backends via LIGHTNING_INDEXER, enhancing performance for specific models on Vulkan-compatible hardware.
Summary: ggerganov/llama.cpp release b10655 introduces LIGHTNING_INDEXER support for Deepseek V4 operations on the Vulkan backend. This update includes optimizations like moving 128-lane dot-product reduction from shared memory to subgroupAdd within Vulkan compute shaders.
Why it matters: This improves efficiency for running Deepseek V4 models on devices with Vulkan GPUs, potentially enabling faster inference on a wider range of hardware. AI developers should explore how this impacts performance for their Deepseek V4-based applications on Vulkan-enabled systems.
Source: github_releases