Changing default `n_chains` on GPUs and `n_discard_per_chain` to something more reasonable about netket

[jwnys]

Lowering the n_discard_per_chain sounds like a great idea (I default to 4). It sounds a bit strange to me, though, to change the number of chains based on the backend (this is what will effectively happen, right?). I would rather take the number of chains constant. If you run on CPU, the number of ranks will be high, and so the number of chains per rank low, while the opposite then happens on GPU.

[gcarleo]

Yes I would use a constant number of discarded (not a fraction) but I wouldn't be in favor of having backend dependent defaults,this might be dangerous and give different results unexpectedly

[alleSini99]

I super agree about the constant and small number of discards per chain. I would keep the number of chains the same on the different backends, though, since the opposite could generate unexpected behaviours.

[PhilipVinc]

Ok, so it seems that there is agreement on a lower n_discard_per_chain. Do we all agree on 10 or is this too large?

About changing n_chains_per_rank dynamically...

• I am proposing to set the default value of n_chains_per_rank based on the backend (@jwnys , not the n_chains)

Right now the default is 16 chains per rank, which is ok for CPUs (even a bit high in some cases when running under MPI) but completely not ok for GPUs.

The reason I'd like to set this is because the current default is completely useless on GPUs, and I see the defaults as something that newcomers can learn from.

What do you fear might break if we change the number of chains dynamically?