API Reference

Most interaction with Forgather happens through the CLI or configuration files, but the Python API is available for programmatic use — loading projects in notebooks, building custom training loops, or integrating with external tooling.

Modules

Module	Key classes
Project System	Project, MetaConfig, ConfigEnvironment
Trainers	Trainer, DDPTrainer, FSDP2Trainer, PipelineTrainer
Callbacks	TrainerCallback, JsonLogger, DivergenceDetector, TrainerControlCallback
Checkpoints	CheckpointMeta, save_checkpoint, load_checkpoint
Optimizers	AdamW, Adafactor, Apollo, schedulers
Datasets	FastDatasetLoaderSimple, InterleavedDataset, dataset utilities
Analysis	TrainingLog, plotting, summary statistics

Quick Example

from forgather.project import Project

# Load a configuration and materialize components
proj = Project("train_config.yaml", "/path/to/project_dir")
training_script = proj()               # Full training script
model_factory = proj("model")          # Just the model factory
model = model_factory()                # Instantiate the model