For large models at the 70B scale, a single GPU's memory is often insufficient, requiring multi-GPU tensor parallelism or the use of quantization techniques.