TL;DR: llama.cpp released an update fixing a critical WebGPU crash when tensor offsets were not multiples of 4, improving stability for browser-based AI inference.
Summary: ggerganov/llama.cpp released version b10721, which includes a fix for a WebGPU crash. The issue occurred when the offset in the ggml_backend_tensor_get() implementation was not a multiple of 4, leading to instability in WebGPU-based operations. This specific fix was contributed by Sigbjørn Skjæret and Stanisław Szymczyk.
Why it matters: This patch enhances the reliability of running llama.cpp models directly in web browsers via WebGPU, making it more robust for client-side AI applications. Developers should update to this version to ensure stable performance for WebGPU-accelerated inference.
Source: github_releases