llama.cpp b11067 Released with WebGPU Fused Operator Optimization
Open-source inference framework llama.cpp has released version b11067. This update introduces a fused GDN and CPY operation to enhance performance when running models on the WebGPU backend. WebGPU enables running machine learning models directly inside web browsers using local GPU hardware acceleration. Optimizing its computational graph reduces memory access overhead, facilitating faster edge AI inference on client devices. The patch combines the Gated Delta Net (GDN) calculation with the memory copy (CPY) pass to decrease roundtrips to memory during inference execution. Pre-compiled binaries for b11067 are available across macOS, Linux, Windows, Android, and iOS platforms.
## BACKGROUND
Operator fusion is an optimization technique in machine learning compilers where multiple consecutive operations are merged into a single GPU execution kernel to avoid writing intermediate results to VRAM. llama.cpp is a high-performance C/C++ library designed to run large language models on modern desktop, mobile, and browser environments.