Orchestration — graphids/orchestrate/

Status: implemented | Last refactor: 2026-04-15 (pipeline route deleted — one route, ablation presets own their own run_dir)

A training run is a jsonnet preset rendered to a dict, validated, then fed through build → train → evaluate. No planner, no cross-stage driver. Multi-stage chains are built in bash by submitting each preset with SBATCH_DEP=afterok:<jid> between them.

Layout

Module	Role
config.py	ResolvedConfig, InstantiatedRun — boundary types.
instantiate.py	build_run(rendered) — class_path + signature-filtered kwargs, callback/logger wiring.
stage.py	build(resolved), train(artifacts, resolved), evaluate(artifacts, resolved).

Execution flow

fit | test  (cli/training.py)
|
+-- render(config_path, tla=...)              [config/jsonnet.py]
+-- apply_overrides(rendered, --set ...)      [cli/app.py]
+-- ResolvedConfig.from_rendered(rendered)    [orchestrate/config.py]
|     -> validate_config(...)                 [config/schemas.py]
|     -> run_dir = trainer.default_root_dir
+-- build(resolved)                           [stage.py]
|     -> gc + torch.cuda reset
|     -> build_run(rendered, validated)       [instantiate.py]
+-- train(artifacts, resolved, resume_from)   [stage.py]
|     -> wire_file_exporters(run_dir)
|     -> trainer.fit(...)
|     -> touch .train_complete
+-- evaluate(artifacts, resolved)             [stage.py]
      -> trainer.test(...)
      -> touch .test_complete + save predictions

Key decisions

Decision	Rationale
Jsonnet preset owns run_dir	Every configs/ablations/*.jsonnet computes run_dir from (lake_root, dataset, seed) via _paths.libsonnet. Path logic lives next to the config, not in a Python planner.
No in-process multi-stage driver	A bash loop over scripts/run <preset> with afterok deps does this without a parallel Python declaration.
build / train / evaluate are dumb primitives	No ResolvedConfig knowledge, no cache knowledge. Same primitives used by fit and test.