Command line

GPULayerCLI (also glt.LightningCLI) builds a command line from the __init__ signatures of the Trainer, your model and your DataModule, like Lightning's LightningCLI.

# train.py
import gpulayer.pytorch as glt
from my_project import MyModel, MyData

if __name__ == "__main__":
    glt.LightningCLI(MyModel, MyData, trainer_defaults={"accelerator": "auto"})

python train.py fit --trainer.max_epochs 10 --model.lr 0.01 --data.batch_size 64
python train.py fit --trainer.devices 4 --trainer.num_nodes 2 --trainer.accelerator cuda
python train.py fit --print_config > experiment.yaml     # the full configuration
python train.py fit --config experiment.yaml --trainer.max_epochs 20
python train.py test --ckpt_path gpulayer_runs/my_project-train/checkpoints/best.ckpt

• Subcommands: fit, validate, test, predict.
• Values follow the type annotation; lists and dicts are JSON.
• Config files are JSON, or YAML with PyYAML, with sections trainer, model, data, seed_everything and ckpt_path. Several --config files merge in order; flags win.
• Objects such as callbacks use class_path and init_args:

trainer:
  max_epochs: 30
  callbacks:
    - class_path: gpulayer.pytorch.EarlyStopping
      init_args: {monitor: val_loss, patience: 3}
model:
  lr: 0.003

• seed_everything defaults to true (a recorded random seed); give a number or false.

• run=False skips subcommands and builds cli.model, cli.datamodule and cli.trainer.

• --optimizer AdamW --optimizer.lr 1e-3 and --lr_scheduler define configure_optimizers from the command line; --model MyModel picks a subclass (subclass_mode_model=True).

• The resolved configuration is saved as config.yaml in the run's log folder (save_config_callback).

• Subclass the CLI and override Lightning's hooks (add_arguments_to_parser, before_instantiate_classes, ...); parser.link_arguments("data.num_classes", "model.num_classes", apply_on="instantiate") links sections.

It uses the standard library's argparse, not jsonargparse: values are converted from the type annotations, and there are no environment-variable configs. cli.config is flat, not nested under the subcommand.