llama.cpp b9194 fuses SSM kernels for faster Vulkan inference

llama.cpp has shipped version b9194, focusing on a Vulkan backend optimization that fuses three consecutive operations — SSM_CONV (state-space model convolution), BIAS, and SILU (sigmoid linear unit activation) — into a single GPU kernel. This fusion reduces kernel launch overhead and memory bandwidth pressure, resulting in measurable speedups for state-space models (SSMs) such as Mamba and its variants. SSMs are gaining traction as efficient alternatives to transformers for long-context tasks, and this optimization makes them more viable on consumer GPUs.

The release provides pre-compiled binaries across an extensive range of platforms: macOS (Apple Silicon, Intel), Linux (multiple architectures with Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Windows (x64/arm64 with CUDA 12/13, Vulkan, SYCL, HIP), Android arm64, iOS XCFramework, and openEuler (with Ascend NPU support). This broad support underscores llama.cpp's role as the go-to open-source engine for running large language models locally. The fusion optimization is part of ongoing efforts to reduce inference latency and make high-quality LLMs accessible without cloud dependencies.