katago-docker

A Docker image to run KataGo in GCP

Important notice

Please don't forget to stop your GCP instance to avoid incurring charges to your account when you don't use it.

Quick Start

Launch a new instance

See instructions in Google Cloud Platform to launch an instance with GPU.

This Docker image is tested with the "GPU Optimized Debian m32 (with CUDA 10.0)" image with NVIDIA Tesla K80 GPU.

SSH to the instance

Set up NVIDIA Container Toolkit

You need to install NVIDIA Container Toolkit to build run GPU accelerated Docker containers.

See Quick Start section in the nvidia-docker page.

You will need to run following commands for Debian Stretch in the instance.

# Add the package repositories
$ distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
$ curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
$ curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list

$ sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit
$ sudo systemctl restart docker

Prepare a model

Download one of the model files from the release page and extract it.

For example,

$ wget https://github.com/lightvector/KataGo/releases/download/v1.1/g104-b15c192-s297383936-d140330251.zip
$ unzip g104-b15c192-s297383936-d140330251.zip

Prepare a config file

If you would like to run KataGo as an engine for Go client such as Lizzie, you will want to use gtp_example.cfg in the KataGo repository as a base line.

For example,

$ wget https://raw.githubusercontent.com/lightvector/KataGo/master/cpp/configs/gtp_example.cfg

Running KataGo from the Docker image

Now, you are ready to run KataGo Docker image. Run the follwoing command.

$ docker run --rm -it \
    -v $(pwd):/data \
    --runtime nvidia \
    hoshir/katago-v1.2-cuda10.0-linux-x64 gtp \
    -model /data/g104-b15c192-s297383936-d140330251/model.txt.gz \
    -config /data/gtp_example.cfg \
    -override-version 0.17

If KataGo works successfully, it will show its version and some outputs like this.

KataGo v1.2
Loaded model /data/g104-b15c192-s297383936-d140330251/model.txt.gz
GTP ready, beginning main protocol loop

Try some showboard GTP command, and quit.

KataGo v1.2
Loaded model /data/g104-b15c192-s297383936-d140330251/model.txt.gz
GTP ready, beginning main protocol loop
showboard
= MoveNum: 0 HASH: 7BF12F3F24903F0C225CD6C55BA8BA1F
   A B C D E F G H J K L M N O P Q R S T
19 . . . . . . . . . . . . . . . . . . .
18 . . . . . . . . . . . . . . . . . . .
17 . . . . . . . . . . . . . . . . . . .
16 . . . . . . . . . . . . . . . . . . .
15 . . . . . . . . . . . . . . . . . . .
14 . . . . . . . . . . . . . . . . . . .
13 . . . . . . . . . . . . . . . . . . .
12 . . . . . . . . . . . . . . . . . . .
11 . . . . . . . . . . . . . . . . . . .
10 . . . . . . . . . . . . . . . . . . .
 9 . . . . . . . . . . . . . . . . . . .
 8 . . . . . . . . . . . . . . . . . . .
 7 . . . . . . . . . . . . . . . . . . .
 6 . . . . . . . . . . . . . . . . . . .
 5 . . . . . . . . . . . . . . . . . . .
 4 . . . . . . . . . . . . . . . . . . .
 3 . . . . . . . . . . . . . . . . . . .
 2 . . . . . . . . . . . . . . . . . . .
 1 . . . . . . . . . . . . . . . . . . .

quit
= 

$

Communicating with the engine from your local machine

You can communicate with the engine from your local machine with SSH.

Register you SSH key to the instance

Create a launcher script in your local machine

Create a small shell script named katago.sh by adding GCP_USER and GCP_IP in it.

#!/bin/bash

set -eu

# User name of the GCP insance
GCP_USER=

# External IP address of the GCP instance
GCP_IP=

ssh -o 'StrictHostKeyChecking no' $GCP_USER@$IP \
  docker run \
    -i \
    --rm \
    --runtime nvidia \
    -v /home/$GCP_USER:/data \
    hoshir/katago-v1.2-cuda10.0-linux-x64 gtp \
    -model /data/g104-b15c192-s297383936-d140330251/model.txt.gz \
    -config /data/gtp_example.cfg \
    -override-version 0.17

The script assumes

• The model file is stored as /home/$GCP_USER/g104-b15c192-s297383936-d140330251/model.txt.gz in the GCP instance.
• The config file is stored as /home/$GCP_USER/gtp_example.cfg in the GCP instance.

The command mounts /home/$GCP_USER directory in the GCP instance as /data when running the image.

Register the script to Go client

Register the script to your Go client. If you use Lizzie, the entry of the engine will be ./katago.sh --weights %network-file. You will need to add --weights %network-file argument since Lizzie checks it.