Fit the large model

Use tensor parallelism when one complete weight set cannot fit on one GPU.