Skills pytorch-lightning
⚡

pytorch-lightning

Safe ⚙️ External commands⚡ Contains scripts🌐 Network access

Build neural networks with PyTorch Lightning

Also available from: davila7

This skill helps you organize PyTorch code into reusable LightningModules. It provides templates and documentation for configuring multi-GPU training, implementing data pipelines, and setting up experiment tracking with popular tools like W&B and TensorBoard.

Supports: Claude Codex Code(CC)
🥉 76 Bronze
1

Download the skill ZIP

2

Upload in Claude

Go to Settings → Capabilities → Skills → Upload skill

3

Toggle on and start using

Test it

Using "pytorch-lightning". Create a simple CNN LightningModule for image classification

Expected outcome:

  • A LightningModule class with __init__, training_step, validation_step, and configure_optimizers
  • Example CNN architecture using torch.nn layers
  • Training loop that returns loss and logs metrics with self.log()
  • Optimizer configuration with Adam and learning rate scheduler

Using "pytorch-lightning". Configure Trainer for GPU training with checkpointing

Expected outcome:

  • Trainer configuration with accelerator='gpu', devices=2
  • ModelCheckpoint callback to save best model based on validation loss
  • EarlyStopping callback to halt training when metrics plateau
  • Progress bar and logger configuration

Security Audit

Safe
v4 • 1/17/2026

All 843 static findings are false positives. The 'Ruby/shell backtick execution' alerts are markdown code blocks, 'weak cryptographic algorithm' alerts flag normal text like 'DDP/FSDP', and 'eval()' refers to PyTorch's model.eval() method. This is legitimate deep learning documentation with no malicious code.

12
Files scanned
9,738
Lines analyzed
3
findings
4
Total audits

Risk Factors

⚙️ External commands (4)
SKILL.md:31-40 references/best_practices.md:8-25 references/logging.md:14-35 references/callbacks.md:18-50
⚡ Contains scripts (2)
references/lightning_module.md:443-447 scripts/quick_trainer_setup.py:8-10
🌐 Network access (2)
references/logging.md:80-88 references/distributed_training.md:437-446
Audited by: claude View Audit History →

Quality Score

68
Architecture
100
Maintainability
87
Content
22
Community
100
Security
83
Spec Compliance

What You Can Build

Organize research experiments

Structure PyTorch code into reusable LightningModules for cleaner experimentation and faster iteration.

Scale training to multiple GPUs

Configure distributed training across clusters with DDP, FSDP, or DeepSpeed for large model training.

Track experiments automatically

Integrate with W&B, TensorBoard, or MLflow to log metrics, hyperparameters, and model checkpoints.

Try These Prompts

Basic model setup 
Show me how to create a LightningModule for an image classifier with training_step, validation_step, and configure_optimizers methods.
Multi-GPU training 
How do I configure a Trainer for multi-GPU training using DDP strategy with 4 GPUs on a single node?
Data pipeline 
Create a LightningDataModule for loading image data with custom transforms for training, validation, and test sets.
Experiment tracking 
Set up Weights & Biases logging with WandbLogger in PyTorch Lightning to track training metrics and hyperparameters.

Best Practices

  • Use self.device instead of .cuda() for device-agnostic code that works on GPU and CPU
  • Call self.save_hyperparameters() in __init__() to save configuration for reproducibility
  • Use self.log() with sync_dist=True when logging metrics in distributed training

Avoid

  • Do not manually call loss.backward() or optimizer.step() - let the Trainer handle optimization
  • Avoid mixing research code (model architecture, loss computation) with engineering code (device management, checkpointing)
  • Do not use .cuda() directly - use self.to(device) or rely on Lightning automatic device placement

Frequently Asked Questions

How do I install PyTorch Lightning?
Run pip install lightning. The skill provides templates and documentation after installation.
What is the difference between DDP, FSDP, and DeepSpeed?
DDP for models under 500M parameters. FSDP shards model across GPUs for larger models. DeepSpeed offers advanced features like CPU offloading.
How do I debug my model quickly?
Use Trainer(fast_dev_run=True) to run one batch through train, validation, and test loops for rapid debugging.
Can I use this skill for inference only?
Yes, use model.eval() mode and trainer.predict() method for inference on new data without training.
How do I resume training from a checkpoint?
Pass ckpt_path='path/to/checkpoint.ckpt' to trainer.fit(), trainer.validate(), or trainer.test() methods.
What loggers are supported?
TensorBoard (default), Weights & Biases, MLflow, Neptune, Comet, and CSVLogger for local files.

Developer Details

Author

K-Dense-AI

License

Apache-2.0 license

Repository

https://github.com/K-Dense-AI/claude-scientific-skills/tree/main/scientific-skills/pytorch-lightning

Ref

main

File structure

📁 references/

📄 best_practices.md

📄 callbacks.md

📄 data_module.md

📄 distributed_training.md

📄 lightning_module.md

📄 logging.md

📄 trainer.md

📁 scripts/

📄 quick_trainer_setup.py

📄 template_datamodule.py

📄 template_lightning_module.py

📄 SKILL.md