Multi-card GPUs can't train occupancy will be 100 card masters, single GPUs won't.

[Lxy811]

Reported an error during training, video card 100%, set batch-size high enough to run normally, error as follows:

First Precls training. At first set bachsize=16 will be stuck in the training before the start, and then set bachsize=32, two GPUs can be trained, but training 5000 and 7600 and so on halfway is stuck. But single card can train normally.