Quickstart

Note

The quickstart will be written once the M2 MVP lands. See the Linear Orion M2 milestone for status.

The intended shape:

# Validate the config without launching a single GPU
orion train configs/cognitive_load_v3_baseline.py --dry-run

# Train on Polaris (Slurm)
orion train configs/cognitive_load_v3_baseline.py

# Burst to cloud (SkyPilot)
orion train configs/cognitive_load_v3_baseline.py --infra.skypilot.cloud auto --infra.nodes 4

# Resume from any failure
orion train configs/cognitive_load_v3_baseline.py --resume latest

# Run a benchmark on an existing checkpoint
orion bench --suite cognitive_load_eval --version 1 --checkpoint <ckpt-id>