llama.cpp fuses DeepSeek-V4 hyper-connections on Vulkan

llama.cpp has landed fused hyper-connection operators for DeepSeek-V4 on the Vulkan backend, cutting graph overhead that had limited multi-device inference.

Antoine Zambelli contributed the combined, pre, and post fused paths. On two AMD Strix Halo boxes linked over USB4 with a 50/50 layer split, both ends running Vulkan/RADV, tester kh0pper measured clear gains against the unfused graph for DeepSeek-V4-Flash at 262k context with f16 KV. A 1.75k-token copy task rose from 7.74 to 14.06 tokens/s decode and 106 to 195 tokens/s prefill. Short chat decode moved from 7.83 to 14.54 tokens/s. A 28k-token prompt saw prefill climb from 82 to 126 tokens/s and decode from 7.20 to 12.33 tokens/s. Worker GTT fell from 88 to 78 GiB because fewer intermediate tensors were materialized.

With the fused path, two-box decode now matches or beats single-box results on a smaller IQ3_XXS quant. The work matters for anyone splitting large DeepSeek-V4 runs across Vulkan devices, where hyper-connection traffic previously left throughput on the table.