LC0 autoconfigure would be nice.

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

Dann Corbit
Posts: 12537
Joined: Wed Mar 08, 2006 8:57 pm
Location: Redmond, WA USA

LC0 autoconfigure would be nice.

Post by Dann Corbit »

It would be very nice if LC0 could look at your system and configure itself for optimal performance, or its best guess as to what that is.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
MOBMAT
Posts: 385
Joined: Sat Feb 04, 2017 11:57 pm
Location: USA

Re: LC0 autoconfigure would be nice.

Post by MOBMAT »

It has so many, I'm afraid to touch them.
i7-6700K @ 4.00Ghz 32Gb, Win 10 Home, EGTBs on PCI SSD
Benchmark: Stockfish15.1 NNUE x64 bmi2 (nps): 1277K
corres
Posts: 3657
Joined: Wed Nov 18, 2015 11:41 am
Location: hungary

Re: LC0 autoconfigure would be nice.

Post by corres »

Dann Corbit wrote: Fri Mar 27, 2020 12:53 am It would be very nice if LC0 could look at your system and configure itself for optimal performance, or its best guess as to what that is.
You can expect this from a big video game but from a some kB chess engine this is an absolute unreal expectation.
Btw. There is only one AB engine what can do it.
I do not know only one either.
User avatar
Ovyron
Posts: 4556
Joined: Tue Jul 03, 2007 4:30 am

Re: LC0 autoconfigure would be nice.

Post by Ovyron »

corres wrote: Fri Mar 27, 2020 8:24 pm You can expect this from a big video game but from a some kB chess engine this is an absolute unreal expectation.
Why? It looks straight-forward. The engine could also take a look at the time control and adjust her own settings to perform better.

What may be unreal is expecting someone to devote time to it, but that's a different subject.
the_real_greco
Posts: 70
Joined: Thu Sep 19, 2019 5:55 am
Full name: Andy!

Re: LC0 autoconfigure would be nice.

Post by the_real_greco »

I've just given up and am going for defaults at home. I don't think it would ever be worth the time to figure them all out, and optimize them, rather than just running my analyses longer.

For the CCC (where they're trying to squeeze out maximum strength), we have

"options" : [
{"name" : "WeightsFile","value" : "/data/engines/weights/384x30-t60-3180.pb.gz"},
{"name" : "Backend","value" : "demux"},
{"name" : "BackendOptions","value" : "(backend=cudnn-fp16,gpu=0),(backend=cudnn-fp16,gpu=1),(backend=cudnn-fp16,gpu=2),(backend=cudnn-fp16,gpu=3)"},
{"name" : "CPuctFactor","value" : "2.815"},
{"name" : "CPuctBase","value" : "18368"},
{"name" : "MaxPrefetch","value" : "160"},
{"name" : "MinibatchSize","value" : "352"},
{"name" : "NNCacheSize","value" : "20000000"},
{"name" : "Threads","value" : "2"},
{"name" : "CPuct","value" : "2.147"},
{"name" : "FpuValue","value" : "0.443"},
{"name" : "MoveOverheadMs","value" : "1000"},
{"name" : "MaxCollisionEvents","value" : "160"},
{"name" : "SyzygyPath","value" : "/data/tmptablebases/tablebases"},
{"name" : "Ponder","value" : "false"},
{"name" : "PolicyTemperature","value" : "1.607"},
{"name" : "MaxOutOfOrderEvalsFactor","value" : "2.0"},
{"name" : "SmartPruningMinimumBatches","value" : "160"},
{"name" : "ScoreType","value" : "centipawn"}
],

Which, what user will even know what MaxOutOfOrderEvalsFactor is? I sure don't.