primus-multi-node.json

{
  "_comment": "Primus pretrain across MULTIPLE pods (Indexed Job). Each pod uses k8s.gpu_count GPUs; nnodes matches pod count. Same file as single-node for any Primus backend: set primus.config_path; BACKEND inferred from model name (primus_pretrain/<launcher>_...) unless primus.backend is set. Usage: madengine build/run ... --additional-context-file examples/k8s-configs/basic/primus-multi-node.json",
  "gpu_vendor": "AMD",
  "guest_os": "UBUNTU",
  "docker_env_vars": {"HF_TOKEN": "YOUR_HF_TOKEN_OR_USE_RUNTIME_SECRET"},
  "k8s": {
    "namespace": "default",
    "gpu_resource_name": "amd.com/gpu",
    "gpu_count": 8
  },
  "distributed": {
    "enabled": true,
    "launcher": "primus",
    "nnodes": 2,
    "nproc_per_node": 8,
    "master_port": 1234,
    "primus": {
      "config_path": "examples/torchtitan/configs/MI300X/qwen3_1.7B-pretrain.yaml",
      "cli_extra": ""
    }
  }
}