Mirage — v0.1

A small template for training and serving DreamerV3 world models on Modal. Define a YAML, run one command, and DreamerV3 trains in a remote container — checkpoints persist, inference is one call away. The source lives on GitHub if you want to poke around.

§ workflow

~ /mirage — zsh

$ pip install -e .
$ mirage train --config configs/dreamerv3_cartpole.yaml
$ mirage play  --config configs/dreamerv3_cartpole.yaml --episodes 5
$ mirage dream --config configs/dreamerv3_cartpole.yaml --episodes 1

§ configs/dreamerv3_cartpole.yaml

configs/dreamerv3_cartpole.yaml

project:
  name: dreamerv3-mirage
  run_name: dreamerv3-cartpole
  seed: 42

modal:
  gpu: A10G
  timeout_seconds: 28800
  volume_name: worldmodel-checkpoints

environment:
  name: dmc-Cartpole-balance
  observation_type: image
  max_episode_steps: 1000
  image_size: 64

dreamer:
  repo_url: https://github.com/burchim/DreamerV3-PyTorch.git
  repo_ref: master
  repo_dir: /opt/DreamerV3-PyTorch
  config_path: configs/DreamerV3/dreamer_v3.py
  env_name: dmc-Cartpole-balance
  mode: training
  log_dreamer_metrics: true
  override_config:
    num_envs: 4
    epochs: 10
    epoch_length: 2000
    batch_size: 16 # was 16 — use the GPU
    L: 64 # was 32 — longer sequences

§ mirage play — dreamerv3-cartpole

One evaluation episode of the trained DreamerV3 agent on dmc-Cartpole-balance, recorded by mirage play on Modal and downloaded locally. Episode return: 969.13.