CLI lifecycle¶
The CLI exposes the main NexuML boundaries without requiring a project-specific training script. Use the generated CLI reference for the exact flags; this page explains when each command belongs in the workflow.
Inspect the environment¶
nexuml registry list scenarios
nexuml registry list layers
nexuml backend list
The registry shows discovered semantic components/scenarios. backend list shows runtime/export backends available in the current installation.
Resolve: Python → persisted configuration¶
nexuml resolve my-scenario
A registered Python scenario is evaluated and compiled, and its resolved configuration is written to configs/my-scenario.yaml by default.
Build: configuration → runtime pipeline¶
nexuml build configs/my-scenario.yaml
build restores typed definitions from YAML, materializes the pipeline, propagates shapes, and reports the compiled stages. When diagram output is enabled, the configured Mermaid file is written as part of the build.
Train: run the canonical lifecycle¶
nexuml train my-scenario
Training can also start from resolved YAML or a trusted scenario file; see Run scenarios.
The local NexuSession lifecycle is:
fit → validate → fit post-train pipeline layers → test
If the scenario contains ExportSpec(kind="train_package"), local training exports the trained package after the run. This is the simplest way to package the exact live trained model without locating a checkpoint manually.
Export an existing checkpoint¶
nexuml export my-scenario --checkpoint PATH -o packages/my-scenario
--checkpoint is optional syntactically, but omitting it does not search for the latest checkpoint. Supply the checkpoint explicitly when the command should package previously trained weights. See Model export.
Other workflows¶
nexuml export-dataset my-scenario -o exported-data
nexuml tune my-scenario --n-trials 20
nexuml smoke my-scenario --max-epochs 1
export-datasetpersists raw or partially processed dataset views.tuneruns Optuna search from a registered scenario or trusted scenario file.smokeexercises resolve → build → train → export → reload → inference for a registered scenario.