Problems in parallel training on GPU

[qiurong13]

With the help of Horovod, the real batch size is 8 when batch_size is set to 2 in the input file and we launch 4 workers. According to the Linear Scaling Rule in the attached paper Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour: When the minibatch size is multiplied by k, multiply the learning rate by k.
Then how should we manually change the value of decay steps?
Why the decay steps should be reduced to 1/2 of steps in the above case?
In this relationship lr(t) = start_lr * decay_rate ^ ( t / decay_steps ), if t / decay_steps always equals to 200, then we could not possibly multiply the learning rate by k without changing decay_rate...

[wanghan-iapcm]

In the latest deepmd-kit, one does not need to set decay rate, but to set stop_lr instead.

[qiurong13]

Then how about the start_lr? do we have to set it manually? or Horovod will multiply it by k atomatically?

And, what is the rule to change decay steps? as in the manual, when batch size multiplies 4, decay steps decrease by half. but i think we can reduce it to 1/4 (1/k), as well as stop batch