b9075

The popular open-source LLM inference engine llama.cpp has released version b9075 with a key GPU optimization: a fused CUDA kernel for the snake activation function. Snake activation is commonly used in neural audio decoders such as BigVGAN and Vocos, where it is naively decomposed into five sequential operations: multiply, sin, square, multiply, and add. The new implementation recognizes this pattern and replaces it with a single elementwise CUDA kernel called `ggml_cuda_op_snake_fused`. This reduces kernel launch overhead, global memory traffic, and register pressure, leading to faster audio inference on NVIDIA GPUs.

The fusion works across F32, F16, and BF16 precision formats, with automatic type conversion handled via `ggml_cuda_cast`. The code also includes a fast division optimization (fastdiv) for tensor lengths. The release adds a comprehensive test (`test_snake_fuse`) that compares CPU naive execution against the fused CUDA path across all supported types. Contributors include community member am17an, who also helped refine the patch during review. The update is part of llama.cpp's ongoing effort to optimize non-transformer layers for real-time AI applications like speech synthesis and voice cloning.