KataGo Docker

This guide is meant to run KataGo with TensorRT in a container.

It may also work with other version of KataGo (OpenCL, CUDA, Eigen), but you will have to edit the Dockerfile.

Build KataGo in a Docker container

1. Clone the repository.

2. Build the docker image or use darkness4/katago:latest :

   docker build -t katago:tensorrt .

You can customize the Dockerfile by setting --build-arg parameters like CUDA_VERSION, OS_VERSION and TRT_VERSION. Note that TRT_VERSION must be available in the container.

Run KataGo in Docker

1. Install Docker and the NVIDIA Container Toolkit.

2. Create an executable shell script to run KataGo:

   #!/bin/sh
   # katago.sh
   
   set -e
   
   SCRIPTPATH="$(dirname "$(realpath "$0")")"
   
   # Change this if you want to use an another model, see https://katagotraining.org
   RELEASE=kata1-b18c384nbt-s7192213760-d3579182099
   
   if [ ! -f "$SCRIPTPATH/default_model.bin.gz" ]; then
      curl -fsSL https://media.katagotraining.org/uploaded/networks/models/kata1/$RELEASE.bin.gz -o "$SCRIPTPATH/default_model.bin.gz"
   fi
   
   docker run --rm --gpus all -it \
     -v "$SCRIPTPATH/default_gtp.cfg:/app/default_gtp.cfg:ro" \
     -v "$SCRIPTPATH/default_model.bin.gz:/app/default_model.bin.gz" \
     katago:tensorrt \
     $@
   

3. Use katago.sh as the main entrypoint.

   chmod +x katago.sh
   ./katago.sh gtp
   # KataGo v1.13.2
   # Using TrompTaylor rules initially, unless GTP/GUI overrides this
   # Initializing board with boardXSize 19 boardYSize 19
   # Loaded config /app/default_gtp.cfg
   # Loaded model /app/default_model.bin.gz
   # Model name: kata1-b18c384nbt-s5832081920-d3223508649
   # GTP ready, beginning main protocol loop

Run KataGo remotely with Docker and SSH

On the remote machine:

1. Install the SSH server and push your public SSH key to your user authorized_keys file. The authentication must not use a password.

2. Install Docker and the NVIDIA Container Toolkit.

On your local machine:

1. Prepare the default_gtp.cfg.

2. Alongside the default_gtp.cfg, create an executable shell script to run KataGo:

   #!/bin/sh
   # katago-remote.sh
   
   set -e
   
   SCRIPTPATH="$(dirname "$(realpath "$0")")"
   
   # Change this if you want to use an another model, see https://katagotraining.org
   RELEASE=kata1-b18c384nbt-s7192213760-d3579182099
   
   scp "$SCRIPTPATH/default_gtp.cfg" remote-user@remote-machine:/tmp/default_gtp.cfg
   
   ssh remote-user@remote-machine "set -e
   if [ ! -f /tmp/default_model.bin.gz ]; then
      curl -fsSL https://media.katagotraining.org/uploaded/networks/models/kata1/$RELEASE.bin.gz -o /tmp/default_model.bin.gz
   fi
   
   docker run --rm -it \
     --gpus all \
     -v /tmp/default_gtp.cfg:/app/default_gtp.cfg:ro \
     -v /tmp/default_model.bin.gz:/app/default_model.bin.gz:ro \
     docker.io/darkness4/katago:latest \
     $@"
   

   Edit the script to match your configuration (ssh parameters, etc.).

3. Make it executable and test it.

   chmod +x katago-remote.sh
   ./katago-remote.sh gtp
   # KataGo v1.13.2
   # Using TrompTaylor rules initially, unless GTP/GUI overrides this
   # Initializing board with boardXSize 19 boardYSize 19
   # Loaded config /app/default_gtp.cfg
   # Loaded model /app/default_model.bin.gz
   # Model name: kata1-b18c384nbt-s5832081920-d3223508649
   # GTP ready, beginning main protocol loop

Run KataGo remotely with Enroot and SSH

On the remote machine:

1. Install the SSH server and push your public SSH key to your user authorized_keys file. The authentication must not use a password.

2. Install Enroot and the NVIDIA Container Toolkit.

On your local machine:

1. Prepare the default_gtp.cfg.

2. Alongside the default_gtp.cfg, create an executable shell script to run KataGo with Enroot:

   #!/bin/sh
   # katago-remote.sh
   
   set -e
   
   SCRIPTPATH="$(dirname "$(realpath "$0")")"
   
   # Change this if you want to use an another model, see https://katagotraining.org
   RELEASE=kata1-b18c384nbt-s7192213760-d3579182099
   
   # Copy config file to remote server
   scp "$SCRIPTPATH/default_gtp.cfg" remote-user@remote-machine:/tmp/default_gtp.cfg
   
   ssh remote-user@remote-machine "set -e
   rm -f /tmp/katago.sqsh
   enroot remove -f -- katago || true
   enroot import -o /tmp/katago.sqsh -- docker://registry-1.docker.io#darkness4/katago:latest
   enroot create -n katago -- /tmp/katago.sqsh
   if [ ! -f /tmp/default_model.bin.gz ]; then
   curl -fsSL https://media.katagotraining.org/uploaded/networks/models/kata1/$RELEASE.bin.gz -o /tmp/default_model.bin.gz
   fi
   
   enroot start  \
   --mount /tmp/default_gtp.cfg:/app/default_gtp.cfg:ro,x-create=file,bind \
   --mount /tmp/default_model.bin.gz:/app/default_model.bin.gz:ro,x-create=file,bind \
   katago \
   $@"
   

   Edit the script to match your configuration (ssh parameters, etc.).

3. Make it executable and test it.

   chmod +x katago-remote.sh
   ./katago-remote.sh gtp
   # KataGo v1.13.2
   # Using TrompTaylor rules initially, unless GTP/GUI overrides this
   # Initializing board with boardXSize 19 boardYSize 19
   # Loaded config /app/default_gtp.cfg
   # Loaded model /app/default_model.bin.gz
   # Model name: kata1-b18c384nbt-s5832081920-d3223508649
   # GTP ready, beginning main protocol loop

Running on DeepSquare

DeepSquare is a decentralized HPC as a service. You can read more about it on DeepSquare's official website and try the portal.

How to use with Neko

Install the dps, the DeepSquare CLI.

Using with Neko is the easiest way to use Sabaki and Katago.

Just submit the job:

dps submit -w -e --no-ts --job-name katago --credits 10000 ./job.katago.yaml

A bore URL will appear in the logs, just connect to it (credentials: user: admin and pass: neko).

You can start the engine by opening the Engines Sidebar (Engines > Show Engines Sidebar).

Then, you can toggle the Analysis mode with F4.

How to use with SSH

1. Edit the job, and replace the SSH public key.

2. You need to start Sabaki locally.

3. Then, you can submit the job:

   dps submit -w -e --no-ts --job-name katago --credits 10000 ./job.yaml

   The user ID and bore URL will appear. This is the URL to the SSH server. You need to use the TCP one.

4. Write a script on your computer to connect to the server via SSH and launch katago with Sabaki:

   /projects/katago/script

   #!/bin/sh
   set -e
   
   SCRIPTPATH="$(dirname "$(realpath "$0")")"
   
   # Copy config file to remote server
   scp -P <port> "$SCRIPTPATH/default_gtp.cfg" <user>@bore.deepsquare.run:/tmp/default_gtp.cfg
   
   ssh -p <port> <user>@bore.deepsquare.run "set -e
   
   # Change this if you want to use an another model, see https://katagotraining.org
   RELEASE=kata1-b18c384nbt-s8493331456-d3920571699
   
   if [ ! -f \$DEEPSQUARE_DISK_TMP/default_model.bin.gz ]; then
     curl -fsSL https://media.katagotraining.org/uploaded/networks/models/kata1/\$RELEASE.bin.gz -o \$DEEPSQUARE_DISK_TMP/default_model.bin.gz
   fi
   
   /app/katago $@ -model \$DEEPSQUARE_DISK_TMP/default_model.bin.gz -config /tmp/default_gtp.cfg"
   

   Replace <user> and <port> with the user ID and bore port.

5. Add an Engine from the Engines Sidebar (Engines > Show Engines Sidebar) with:

   Path: The path to the script (/projects/katago/script)

   Arguments: gtp

   

6. Then start the engine:

   

7. Finally, you can toggle the Analysis mode with F4.