hamgnn-training

name	hamgnn-training
description	Prepare HamGNN training configs and run locally or submit through the canonical HamGNN Inspire flow.

HamGNN Training

1. Start from <skill_root>/templates/HamGNN/train_config.yaml.template, then render a run-local config under <target_dir>/HamGNN/.
2. Use inspire-hamgnn-submit for the actual remote submission. inspire-hamgnn-submit is the sole authority for HamGNN workspace, GPU resource, image, and helper selection.
3. Use <skill_root>/scripts/train.sh --workdir <hamgnn_workdir> --config <rendered_train_config> as the required runnable entrypoint for this stage.
4. Do not edit a shared global HamGNN/train_config.yaml in place.
5. When the upstream pipeline already produced a dedicated training .npz, render the config so it only creates an internal train/validation split inside that training dataset. Do not allocate an extra internal test split from the training .npz.
6. This stage consumes the training .npz generated from the selected Stage 2 DFT route (OpenMX or ABACUS).
7. Render __HAM_TYPE__ from the upstream route:

• OpenMX route -> openmx
• ABACUS route -> abacus

8. Apply the same rendered __HAM_TYPE__ value to both output_nets.HamGNN_out.ham_type and representation_nets.HamGNN_pre.radius_type.
9. Local runtime env policy: prefer conda env HamGNN; if it is unavailable, fall back to base. Do not create new envs unless explicitly requested.
10. For ordinary HamGNN training, render uni_model_pkl_path: null by default. Only populate uni_model_pkl_path when the user explicitly asks to load a Uni model.
11. This stage must emit the trained model or checkpoint path required by the inference stage.
12. Continue monitoring until training is complete. Since this stage is long-running, poll every 60 minutes.