I tried to inscrease batchsize but met some difficulties

[DengYizhe]

I inscreased batchsize by refill the shape of query(Instances in code) in different batches by 0 or -1, which made it possible to train in batchsize>1 during pre process, and keep the post process batch by batch. But I met the problem that loss is large, grad_norm is small and the loss descents slowly. I wonder if anyone else tried to achieve larger batchsize, and did you met the same case?