How lightning-ai/pytorch-lightning works
Lightning provides two training abstractions: PyTorch Lightning automates model training and evaluation, while Fabric offers more direct control over distributed execution. The documented workflow defines a LightningModule and data, then passes both to Trainer.fit. Trainer orchestrates training loops; model hooks, optimization, metrics, callbacks, logging, and checkpointing provide distinct capabilities. Fabric exposes a lower-level entry point with configurable strategies, precision, and device support. Several subsystems are represented from the README and tree, but their internal wiring was not sampled, so uncertain connections are omitted.
Main entry points: trainer.py Trainer · fabric.py Fabric API · training_epoch_loop.py Epoch Loop · module.py LightningModule
Components 19
PyTorch Lightning
- Trainer
src/lightning/pytorch/trainer/trainer.py
Model and Data
- LightningModule
src/lightning/pytorch/core/module.py - DataModule
src/lightning/pytorch/core/datamodule.py
Training Execution
- Fit Loop
src/lightning/pytorch/loops/fit_loop.py - Epoch Loop
src/lightning/pytorch/loops/training_epoch_loop.py - Evaluation Loop
src/lightning/pytorch/loops/evaluation_loop.py - Prediction Loop
src/lightning/pytorch/loops/prediction_loop.py - Optimization
src/lightning/pytorch/loops/optimization/automatic.py
Training Support
- Metric Results
src/lightning/pytorch/trainer/connectors/logger_connector/result.py - Callbacks
src/lightning/pytorch/callbacks/callback.py - Weight Averaging
src/lightning/pytorch/callbacks/weight_averaging.py - Checkpointing
src/lightning/pytorch/callbacks/model_checkpoint.py - Loggers
src/lightning/pytorch/loggers/logger.py
Lightning Fabric
- Fabric API
src/lightning/fabric/fabric.py - Device Connector
src/lightning/fabric/connector.py - Execution Strategy
src/lightning/fabric/strategies/strategy.py - Precision Plugins
src/lightning/fabric/plugins/precision/precision.py - Accelerators
src/lightning/fabric/accelerators/accelerator.py
Other components
- Engineer
Connections 21
- Engineer → LightningModule: defines model
- Engineer → DataModule: provides data
- Engineer → Trainer: starts training
- Trainer → LightningModule: fits model
- Trainer → DataModule: uses data
src/lightning/pytorch/trainer/trainer.py - Trainer → Fit Loop: dispatches fit
src/lightning/pytorch/trainer/trainer.py - Trainer → Evaluation Loop: dispatches evaluation
src/lightning/pytorch/trainer/trainer.py - Fit Loop → Epoch Loop: runs epochs
src/lightning/pytorch/loops/training_epoch_loop.py - Epoch Loop → Optimization: performs optimization
src/lightning/pytorch/loops/training_epoch_loop.py - Epoch Loop → Metric Results: updates results
src/lightning/pytorch/loops/training_epoch_loop.py - LightningModule → Metric Results: logs metrics
src/lightning/pytorch/core/module.py - Trainer → Prediction Loop: dispatches prediction
- Trainer → Callbacks: configures callbacks
src/lightning/pytorch/trainer/trainer.py - Callbacks → Weight Averaging: extends callbacks
src/lightning/pytorch/callbacks/weight_averaging.py - Trainer → Checkpointing: configures checkpointing
- Trainer → Loggers: configures logging
src/lightning/pytorch/trainer/trainer.py - Engineer → Fabric API: chooses expert control
- Fabric API → Device Connector: configures devices
src/lightning/fabric/fabric.py - Fabric API → Execution Strategy: uses strategy
src/lightning/fabric/fabric.py - Fabric API → Precision Plugins: uses precision
src/lightning/fabric/fabric.py - Device Connector → Accelerators: selects accelerator