Allowing setting timeout in DeepSpeedStrategy

[jedyang97]

Outline & Motivation

In DDPStrategy / FSDPStrategy, the timeout=datetime.timedelta(seconds=1800) flag is exposed and thus allowing user to tune. However, in DeepSpeedStrategy, which is a subclass of DDPStrategy, this flag is not exposed, which makes it hard to change the timeout behavior.

Is there any workaround? Otherwise, I think it might be worth adding kwargs to the __init__() function of DeepSpeedStrategy, and pass along those parameters to the parent class DDPStrategy.

Pitch

No response

Additional context

No response

cc @Borda @awaelchli @justusschock

[lantiga]

Hey @jedyang97 definitely, do you want to send in a quick PR?