can‘t ’get 51.8 AP on coco use the source code

[UcanSee]

Thanks for your great work！
I have used the source code to reproduce the performance of mask_rcnn_vitbase，I used the original
"mimdet_vit_base_mask_rcnn_fpn_sr_0p5_800_1333_4xdec_coco_3x.py" config with 64 4090D GPU, But I have just got 50.0 AP with test sample_ratio=1.0. Is there anything special when you training?

[UcanSee]

I tried training a second time and achieved 50.8，but it was still not up to the level of the paper