LC0 V0.17 UCI options

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

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
User avatar
Rubinus
Posts: 953
Joined: Thu Jan 18, 2007 3:05 pm
Location: Prague
Contact:

LC0 V0.17 UCI options

Post by Rubinus » Mon Aug 20, 2018 8:21 am

Code: Select all

[OPTIONS]
Network weights file path=<autodiscover>
Number of worker threads=2
NNCache size=200000
NN backend to use=opencl
Scale thinking time=2.660000
Move time overhead in milliseconds=100
Time weight curve peak ply=26.200001
Time weight curve width left of peak=82.000000
Time weight curve width right of peak=74.000000
Ponder=false
Minibatch size for NN inference=256
Max prefetch nodes, per NN call=32
Cpuct MCTS option=3.400000
Initial temperature=0.000000
Moves with temperature decay=0
Add Dirichlet noise at root node=false
Display verbose move stats=false
Aversion to search if change unlikely=1.470000
First Play Urgency Reduction=0.900000
Length of history to include in cache=1
Policy softmax temperature=2.200000
Allowed node collisions, per batch=32
Out-of-order cache backpropagation=false
Ignore alternatives to checkmate=false
Configuration file path=lc0.config
Which value to edit if the engine exceeds the limit? Scale thinking time? More or less?

crem
Posts: 116
Joined: Wed May 23, 2018 7:29 pm

Re: LC0 V0.17 UCI options

Post by crem » Mon Aug 20, 2018 8:23 am

Rubinus wrote:
Mon Aug 20, 2018 8:21 am

Code: Select all

[OPTIONS]
Network weights file path=<autodiscover>
Number of worker threads=2
NNCache size=200000
NN backend to use=opencl
Scale thinking time=2.660000
Move time overhead in milliseconds=100
Time weight curve peak ply=26.200001
Time weight curve width left of peak=82.000000
Time weight curve width right of peak=74.000000
Ponder=false
Minibatch size for NN inference=256
Max prefetch nodes, per NN call=32
Cpuct MCTS option=3.400000
Initial temperature=0.000000
Moves with temperature decay=0
Add Dirichlet noise at root node=false
Display verbose move stats=false
Aversion to search if change unlikely=1.470000
First Play Urgency Reduction=0.900000
Length of history to include in cache=1
Policy softmax temperature=2.200000
Allowed node collisions, per batch=32
Out-of-order cache backpropagation=false
Ignore alternatives to checkmate=false
Configuration file path=lc0.config
Which value to edit if the engine exceeds the limit? Scale thinking time? More or less?
If it's not CUDA version but opencl / blas, then "Minibatch size for NN inference=16" should help.
If it's CUDA, it's worth looking further and fixing a bug.

Post Reply