Running large language models (LLMs) often leads to a common frustration: the "CUDA Out of Memory" (OOM) error. Even with high-end A100 or H100 GPUs, standard inference engines waste a si…