llama.cpp

732707df - quantize: cap working memory size to avoid loading big tensors onto RAM (#27795)