At repo root:
pip install -e ".[oneforall]"For DGL-related tasks: pip install -e ".[oneforall-dgl]" (pick a DGL wheel that matches your CUDA).
From any working directory (child training processes cwd to repo root). Paths after -c resolve relative to CWD first, then repo root:
python scripts/oneforall/run.py -c configs/oneforall/smoke.yaml
# Skip pulling HuggingFace sentence-encoder weights (random placeholder encodings, smoke test only):
python scripts/oneforall/run.py -c configs/oneforall/smoke_stub.yaml
python scripts/oneforall/run.py -c pygfm/baseline_models/oneforall/e2e_all_config.yamlAppend key value overrides like run_cdm:
python scripts/oneforall/run.py -c configs/oneforall/smoke.yaml num_epochs 2 logger_backend noneExport merged YAML (no training):
python scripts/oneforall/run.py --export-run-yaml /tmp/merged.yaml -c configs/oneforall/smoke.yamlWithout YAML, from repo root:
python -m pygfm.baseline_models.oneforall.run_cdm task_names cora_node num_epochs 1(In-package run_cdm.py sets HF_ENDPOINT=https://hf-mirror.com before other imports.)
| Variable | Meaning |
|---|---|
ONEFORALL_DATA_ROOT |
Your PyG graph assets root (default <repo>/datasets/oneforall). Smoke Cora: place Cora.pt here for the cora_node task. |
ONEFORALL_CACHE_ROOT |
OFA preprocessing / text encoding cache (default <ONEFORALL_DATA_ROOT>/cache_data) |
ONEFORALL_EXP_ROOT |
Experiment outputs (default <repo>/ckpts/oneforall/runs) |
HF_ENDPOINT |
Hugging Face endpoint (if unset, package run_cdm defaults to the mirror) |
See docs/oneforall/ONEFORALL_INTEGRATION_PLAN.md.
Set in YAML or run.py trailing opts, e.g.:
logger_backend swanlab, offline_log False, swanlab_mode cloud (cloud needs swanlab login first).
Defaults: pygfm/baseline_models/oneforall/configs/default_config.yaml.