Getting Started

Command Line Tool

aicentre runs on your laptop and works with P1 compute in three modes:

Mode	Commands	Where it runs
Interactive, CPU	`aicentre shell`, `marimo`, `code`	A login node, at once
Interactive, GPU	the same, with `--gpu`	A GPU node, as an interactive job
Batch	`aicentre submit`, `list jobs`, `logs`, `cancel`	The batch queue, without you

aicentre check checks your setup. Commands cover every provider you are set up for, and --provider gefion (or -p gefion) picks one.

A provider that asks for more than your key, as DTU asks for a password, asks once on your terminal. The login then stays open for 30 minutes.

Install

bash

curl -fsSL https://hpc.aicentre.dk/install | sh

It installs the release for your machine in your home directory, with no admin rights. It runs on Linux and macOS, and in WSL on Windows. Set AICENTRE_VERSION for a particular version:

bash

curl -fsSL https://hpc.aicentre.dk/install | AICENTRE_VERSION=0.1.0 sh

With Nix, see With Nix.

Check your setup

bash

aicentre check              # the providers you are set up for
aicentre check --all        # every provider
aicentre check --connect    # also log in

It checks OpenSSH, ~/.ssh and ~/.ssh/config, then each provider:

Provider	What it checks
P1 DTU HPC	A `Host dtu` block for `login9.hpc.dtu.dk`, your username, an SSH key ssh will use, the host key
P1 Gefion HPC	A Host block for the login node, and whether your DCAI sign-in login is still valid
P1 NGC HPC	Nothing to set up here: you work on it through its remote desktop

A provider is set up when ~/.ssh/config has its Host block, or for Gefion when you have logged in to DCAI. --all also checks the rest and prints any missing Host block to paste in.

--connect tests the VPN, your key and your sign-in. It never prompts, so a login that would ask fails with a reason. The exit code is 0 when nothing failed and 1 otherwise, and --json prints results for scripts.

Settings

Settings are optional. Put names that differ from the defaults in ~/.config/aicentre/config.toml. A provider listed there is always checked.

~/.config/aicentre/config.toml

[dtu]
user = "s123456"

[gefion]
user = "jdoe"
alias = "gefion"       # the Host name in ~/.ssh/config
# account = "..."      # a project's own Slurm account; P1's iu_0096 is the default

The queues are P1's and need no setting: batch jobs go to p1 and --gpu sessions to p1i on DTU, and a script's own #BSUB -q wins.

Your jobs

bash

aicentre list jobs

It shows your running, waiting and suspended jobs across P1 in one table, batch and interactive. NGC is air-gapped, so its jobs show only in qstat on its remote desktop.

Your jobs on P1 compute  3 jobs
  PROVIDER  JOB    STATE    QUEUE  NAME   NODE       ELAPSED  LEFT
  DTU       123    running  p1     train  n-62-30-1  0:05:00  0:55:00
  Gefion    98765  running  p1     eval   dgx-01     2:03:04  21:56:56
  DTU       124    pending  p1     sweep  -          -        -        Not enough GPUs available

Interactive sessions

bash

aicentre shell                           # a shell on the login node
aicentre marimo project/notebook.py      # marimo in your browser
aicentre code project                    # VS Code in your browser
aicentre code project --gpu --time 2h    # the same, on a GPU node

Without --gpu, a session starts at once on your first provider's login node. The login node is shared, so keep to editing, installs and light runs.

With --gpu, it runs as an interactive job with one GPU on p1i, P1 DTU HPC's interactive queue. --gpu is for DTU only, since Gefion runs GPU work as batch jobs and NGC uses its remote desktop.

Paths are on the cluster, relative to your project directory (/dtu/p1/<username> on DTU, your home on Gefion) unless absolute or under ~. Nothing is copied from your machine.

Option	Default	What
`--gpu`		Run on a GPU node instead of the login node
`--time 2h`	1 hour	With `--gpu`: how long the job may run, as `90`, `2h` or `1:30`
`--cpus 4`	4	With `--gpu`: CPU cores
`--mem 8GB`	8GB	With `--gpu`: memory per core
`--server openvscode`	VS Code's own server	For `code`: openvscode-server, the open build, with extensions from Open VSX
`--no-browser`		Print the address without opening it

Ctrl-C ends the session and its job. The first marimo or VS Code session takes a minute to set up in .aicentre in your project directory.

Batch jobs

bash

aicentre submit train.sh
aicentre logs 4242 -f
aicentre cancel 4242

submit sends a #BSUB script to P1 DTU HPC and an #SBATCH script to P1 Gefion HPC. A local script is copied up first.

What the script leaves out gets P1's default: the p1 queue on DTU, the Slurm account iu_0096 on Gefion, and output in .aicentre/logs. Name another account in the script or in your settings.

logs and cancel need only the job id. logs -f follows the output and --err shows the error file.

Moving data

bash

aicentre push datasets              # datasets/ into your project directory
aicentre pull results               # results/ back to this folder

Both resume a stopped copy. The moving data example covers git, large transfers and NGC's staging area.

Examples

bash

aicentre gallery                    # the examples in the gallery
aicentre gallery get int-gpu        # one of them onto the cluster, in examples/int-gpu

gallery get downloads the example on the login node and prints what to run next.

Chat

bash

aicentre chat
aicentre chat --model <provider/id>
aicentre chat -c                    # pick up the last conversation

chat starts an agent with tools that read this site, your setup, your jobs and their output. It has no shell, and it submits or cancels a job only after you say yes. Sign in with /login in the chat, or set a key such as ANTHROPIC_API_KEY.

On the clusters

aicentre keeps itself on each provider you use, at your laptop's version, in ~/.local/bin/aicentre. On a login node, the same commands use the scheduler directly:

bash

aicentre list jobs
aicentre submit train.sh
aicentre logs 4242 -f
aicentre shell --gpu        # a GPU shell, from the login node

marimo and VS Code open in a browser, so start them from your laptop. Set sync = false under a provider in your settings to keep aicentre off it.

Upgrade and uninstall

bash

aicentre upgrade
aicentre uninstall

upgrade switches to the latest release, and keeps your version if anything fails. uninstall asks before it removes your settings, and --purge skips the question. It ends with the command that removes aicentre from each cluster.

With Nix

Run it from this site's flake without installing:

bash

nix run github:Pioneer-Centre-for-AI/p1-compute-documentation#aicentre -- check

To keep it, use the flake's Home Manager module. settings becomes ~/.config/aicentre/config.toml:

nix

# flake.nix
inputs.p1-compute.url = "github:Pioneer-Centre-for-AI/p1-compute-documentation";

# your Home Manager configuration
imports = [ inputs.p1-compute.homeManagerModules.default ];
programs.aicentre = {
  enable = true;
  settings = {
    dtu.user = "s123456";
    gefion.user = "jdoe";
  };
};

On NixOS without Home Manager, use nixosModules.default with programs.aicentre.enable = true, or overlays.default for pkgs.aicentre. Nix then handles updates, so upgrade and uninstall leave it alone.

Where it keeps files

Path	What
`~/.local/bin/aicentre`	A link to the program
`~/.local/share/aicentre/`	The program itself
`~/.config/aicentre/`	Your settings, and chat's model logins in `agent/`
`~/.local/state/aicentre/`	State it keeps between runs: job notes and chat conversations
`~/.cache/aicentre/`	Downloads during an upgrade
`$XDG_RUNTIME_DIR/aicentre/`	Open logins to the providers, and other files that only live while it runs

Source

The source is in pkgs/cli. Send ideas to #ask-compute or as a pull request.