llama.cpp b10232 brings DeepSeek V4 hyper-connections to Apple Metal

The llama.cpp project, led by ggml-org, released version b10232 on August 2nd with a significant new capability: Metal-based support for DeepSeek V4 hyper-connections. The release introduces three new operators—GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE, and GGML_OP_DSV4_HC_POST—that are implemented using SIMDgroup register and shuffle-optimized kernels. These low-level optimizations are designed to maximize throughput on Apple's GPU architecture, which is critical for handling the complex, multi-head attention patterns that hyper-connections introduce in DeepSeek V4.

Beyond the compute kernels, this release adds Metal dispatch plumbing so the new operators can be properly scheduled across Apple Silicon's unified memory GPU. It also tests the production Sinkhorn iteration count and embedding width, ensuring that the numerical results match what the reference implementation expects. For developers and researchers, this means they can now run DeepSeek V4 models locally on macOS and iOS devices without relying on cloud APIs, with performance that takes advantage of Metal's low-level GPU control.