Leelenscorer

Purpose

Collection of tools to assist in the development of the Leelenstein nets

Setup

Server

All you should need to run the server is a collection of games in the .gz format that you wish to score, as well as python3.7. Then, install requirements.txt globally or in a virtualenv using pip install -r requirements.txt

Client

If you're pro python, just set up a virtualenv and pip install the requirements in there. Client uses asyncio features that are new as of python3.7, so python3.7 is a requirement

git clone [email protected]:kmorrison/leelenscorer.git
cd leelenscorer/
pip install -r requirements.txt
python3 multi_client.py \
  --clients-per-gpu=3 \
  --engine-path=<path to lc0> \
  --weights-path=<path to LS 11.1 network> \
  --host=173.67.18.127 \
  --port=8889 \
  --backend=cudnn-fp16 \
  --num-nodes=128 \
  --minibatchsize=16 \
  --client-name=RAFmb16N128 \
  --chunk-size=5

If you like docker instead, the client code should be all ready to go in the kmorrison64/leelenscorer dockerhub repo. If you're using vast.ai, just switch the container to the latest version at that repo and you're half-way there.

Usage

Server

The server can be run like python game_server.py --input-folder=<example folder> --output-folder=<different folder>

Client

The client can be run in two forms, either the single client which will pull games from the server and give to one engine instance for scoring, or the multi-client.py which uses nvidia-smi to detect the number of available GPUs and attempts to use them all.

Examples

python multi_client.py --clients-per-gpu=1 --chunk-size=10 --engine-path=<where to engine> --weights-path=<where to weights> --host=<route to server> --port=<server port>

python rescore_client.py --chunk-size=10 --engine-path=<where to engine> --weights-path=<where to weights> --host=<route to server> --port=<server port>

#To just parrot back the input files and not score anything, use --dry-run option
python rescore_client.py --dry-run=True --chunk-size=10 --engine-path=<where to engine> --weights-path=<where to weights> --host=<route to server> --port=<server port>

python3 multi_client.py --host=localhost --port=8888 --backend=cudnn-fp16 --clients-per-gpu=5 --engine-path=/root/binaries/lc0 --weights-path=/root/binaries/ls-n11-1.pb.gz --chunk-size=10

Gotchas

• the client-server protocol is custom (seemed like a good idea at the time :P) and uses four newlines as a separator between messages. If any files you're transmitting have b'\n\n\n\n' in them, we're gonna have a bad time
• if a client fails to score a set of games it is handed for some reason, there is no mechanism for requeueing them and handing them to another client.
• when running locally via docker you will have to set --network="host" as an arg to docker run, and pass --host="host.docker.internal" to your client script