多GPU加载后，运行 Image Understanding， GPU不能满负荷运行。

使用本项目的代码， Image Understanding 一个3M的表格，8卡 H20。使用到了6张卡，每张卡的使用率在9-10%，速度并不快。
nvtop 如下：
3784743 root 1 Compute 9% 5372MiB 4% 0% 2224MiB
3784743 root 2 Compute 9% 5372MiB 4% 0% 2224MiB
3784743 root 3 Compute 9% 5372MiB 4% 0% 2224MiB
3784743 root 4 Compute 9% 5372MiB 4% 0% 2224MiB
3784743 root 5 Compute 10% 5370MiB 4% 0% 2224MiB
3784743 root 0 Compute 6% 4804MiB 3% 100% 2224MiB

是否正常？如何配置能够提升GPU使用率？

Start from the Image Understanding entry point and its multi-GPU configuration, then reproduce the reported 3M-table workload on eight H20 GPUs while monitoring utilization with nvtop. Done means determining whether the low utilization and six-GPU usage are expected, and identifying a documented configuration or a confirmed implementation issue.

Written by the indexing model from the issue text.