[2001zhaozhao]

I wonder whether it is much more cost-effective in terms of token throughput / hardware+power cost to get actual GPUs instead, given that the model size is only 27B.