FLAGS.md

Modes

lc0 supports several operating modes, each of which has a different set of command line flags (although there are common ones).

Currently, lc0 has the following modes:

Mode	Description
uci (default)	Acts as a UCI chess engine
selfplay	Plays one or multiple games with itself and optionally generates training data
debug	Generates debug data for a position

To run lc0 in any of those modes, specify a mode name as the first argument (uci may be omitted). For example:

$ ./lc0 selfplay ...    # For selfplay mode
$ ./lc0 ...             # For UCI mode

In any of those modes, it's possible to get help using the --help command line argument:

$ ./lc0 --help            # Help for UCI mode
$ ./lc0 selfplay --help   # Help for selfplay mode

UCI Mode

In UCI engine mode, all of command line parameters can also be changed using a UCI parameter.

List of command line flags:

Flag	UCI Parameter	Description
-w PATH, --weights=PATH	Network weights file path	Path to load network weights from. Default is <autodiscover>, which makes it search for the latest (by file date) file in ./ and ./weights/ subdirectories which look like weights.
-t NUM, --threads=NUM	Number of worker threads	Number of (CPU) threads to use. Default is 2. There's currently no use of making it more than 3 as it's limited by mutex contention which is yet to be optimized.
--nncache=SIZE	NNCache size	Number of positions to store in cache. Default: 2000000
--backend=BACKEND --backend-opts=OPTS	NN backend to use NN backend parameters	Configuration of backend parameters. Described in details here. Default depends on particular build type (cuDNN, tensorflow, etc.).
--slowmover=NUM	Scale thinking time	Parameter value X means that the whole remaining time is split in such a way that the current move gets X × Y seconds, and next moves will get 1 × Y seconds. However, due to smart pruning, the engine usually doesn't use all allocated time. Default: 2.2
--move-overhead=NUM	Move time overhead in milliseconds	How much overhead should the engine allocate for every move (to counteract things like slow connection, interprocess communication, etc.). Default: 100 ms.
--minibatch-size=NUM	Minibatch size for NN inference	How many positions the engine tries to batch together for computation. Theoretically larger batches may reduce strengths a bit, especially on a small number of playouts. Default is 256. Every backend/hardware has different optimal value (e.g., 1 if batching is not supported).
--max-prefetch=NUM	Maximum prefetch nodes per NN call	When the engine can't gather a large enough batch for immediate use, try to prefetch up to X positions, which are likely to be useful soon, and put them in the cache. Default: 32.
--cpuct=NUM	Cpuct MCTS option	C_puct constant from Upper Confidence Tree search algorithm. Higher values promote more exploration/wider search, lower values promote more confidence/deeper search. Default: 1.2.
--temperature=NUM	Initial temperature	Tau value from softmax formula. If equal to 0, the engine also picks the best move to make. Larger values increase randomness while making the move. Default: 0
--tempdecay-moves=NUM	Moves with temperature decay	Reduce temperature for every move linearly from initial temperature to 0, during this number of moves since the game started. 0 disables temperature decay. Default: 0
-n, --[no-]noise	Add Dirichlet noise to root node	Add noise to root node prior probabilities. This allows the engine to explore moves which are known to be very bad, and this is useful to discover new ideas during training. Default: false
--[no-]verbose-move-stats	Display verbose move stats	Display Q, V, N, U and P values of every move candidate after each move. Default: false
--[no-]smart-pruning	Enable smart pruning	Default: true
--virtual-loss-bug=NUM	Virtual loss bug	Default: 0
--fpu-reduction=NUM	First Play Urgency reduction	Default: 0.2
--cache-history-length=NUM	The length of history to include in the cache	Default: 7
--extra-virtual-loss=NUM	Extra virtual loss	Default: 0
-l, --logfile=FILENAME	Do debug logging into a file	Default is off (empty string)

Configuration Files

lc0 supports using a configuration file instead of passing flags on the command line. The default configuration file is lc0.config, but it can be changed with the --config command line flag. lc0 configuration files only support the long flags that begin with --, and there must only be 1 flag per line. For example:

# Lines beginning with a # is a comment
--threads=1
--minibatch-size=32
--sticky-checkmate
# The -- is optional.  The following flags will work as well:
weights=10445.txt.gz
syzygy-paths=syzygy
logfile=lc0.log

You can tell lc0 to ignore the default configuration file by passing --config= on the command line. Command line arguments will override any arguments that also exist in the configuration file.

Backend Configuration

To be explained. That's the most interesting and undocumented!

Selfplay Mode

To be explained.

Debug Mode

To be explained.