ggerganov/llama.cpp Released b12114 with WebGPU Fixes
Llama.cpp release b12114 introduces important WebGPU fixes, specifically addressing buffer size management and shader improvements for memset operations. These updates resolve issues related to handling buffers larger than the maximum binding size. These fixes improve the stability and performance of running large language models via WebGPU, making browser-based and cross-platform GPU acceleration more robust. This enhancement benefits developers utilizing web-based or native WebGPU backends for AI inference. The update fixes get_max_size to correctly return maxBufferSize and avoids whole-buffer bindings in memset or clear operations. This allows buffer_memset to seamlessly handle buffer sizes exceeding the maximum binding size.
## BACKGROUND
llama.cpp is an open-source library that enables efficient inference for large language models locally across various hardware backends. WebGPU is a modern web standard API that grants web applications high-performance access to system graphics processing units for compute and graphics tasks.