The accuracy on the val set cannot be reproduced

[Spongebobbbbbbbb]

I used the joint training dancetrack and crowdhuman datasets. The proposal directly used the detection results of yolox downloaded from the readme. The configurations were 8 GPU A100 and 4 GPU A100 for training, 5 echpos, and the other configurations were the same as the article, but the training results were:
8 GPU A100:
HOTA DetA AssA
54.681 69.666 43.073
4 GPU A100:
HOTA DetA AssA
59.053 74.578 46.963
This is the result of the article:
CrowdHuman YOLOX HOTA DetA AssA
57.1 66.2 49.5
✓ 60.7 74.8 49.6
✓ 56.7 73.7 43.9
✓ ✓ 63.7 76.6 53.2

1. Why is there such a big difference with different GPU numbers
2. The results of multiple reproductions are very different and the results are unstable
3. The reproduction accuracy is very different