llama.cpp Release b10921 Fixes WebGPU Alignment for Quantized Tensors
The llama.cpp project released build b10921 featuring a targeted fix for WebGPU tensor bindings. The update adjusts binding offsets to align with block sizes, ensuring block-quantized views receive valid element offsets within GPU shaders. This bug fix prevents memory alignment errors and shader miscalculations when running quantized LLMs using WebGPU backends. It enhances the reliability and stability of in-browser open-source AI inference. The fix walks back the binding offset until the distance to the tensor is an integer number of blocks, aligning data structures with WebGPU storage buffer restrictions. Updated binary builds were published across macOS, Windows, Linux, and Android platforms.
## BACKGROUND
llama.cpp is a widely used C/C++ library designed for efficient local inference of Large Language Models on diverse hardware, including WebGPU for web browsers. WebGPU requires strict memory buffer alignment, which can conflict with block quantization techniques that compress model weights into fixed-size block groups to reduce memory usage.