llama.cpp b9279 fuses snake activation for faster audio inference on Vulkan

The llama.cpp project (GitHub stars: 112k) dropped release b9279 with a headline feature: fused snake activation on Vulkan. Previously, audio decoders like BigVGAN and Vocos decomposed the snake activation function into five separate operations (mul, sin, sqr, mul, add). The new ggml_vk_snake_dispatch_fused kernel recognizes that pattern and collapses it into one elementwise shader (snake.comp). This not only reduces kernel launch overhead but also allows better memory locality across F32, F16, and BF16 precision. On the C++ side, the matcher was tightened — it now requires contiguous tensors, F32 broadcast multipliers, and rejects cases with batch dimensions >1 (ne[2] or ne[3] > 1). The shader bindings were also renamed to follow the standard ggml naming convention (data_a, data_b, data_c, data_d) with explicit type documentation.

Beyond the core feature, this release bundles pre-built binaries for ten platforms — including macOS (ARM and Intel), Linux (x64/arm64/s390x, Vulkan, ROCm 7.2, SYCL), Windows (x64/arm64, CUDA 12/13, Vulkan, HIP), Android (arm64 CPU), and even openEuler. Each build is signed with a verified GPG key. The commit history also reveals multiple review rounds from contributors jeffbolznv and 0cc4m, ensuring the fusion path is correct and robust. For developers running LLM inference with audio output (e.g., text-to-speech or voice cloning), this update can meaningfully reduce latency on Vulkan-compatible GPUs.