Zum Inhalt springen
Kitbay
Skill

pytorch-lightning

MITAktualisiert am 11. Nov. 2026262.606GitHub-Sterne

An optional skill in Nous Research's Hermes Agent, credited to Orchestra Research, for writing model training code with PyTorch Lightning.

Was der Skill vermittelt

An optional skill in Nous Research's Hermes Agent, credited to Orchestra Research, for writing model training code with PyTorch Lightning. It shows how to move a plain PyTorch loop into a LightningModule run by a Trainer, add validation and test steps, schedule the learning rate and use callbacks for checkpoints, early stopping and learning-rate logging. It covers distributed training with DDP, FSDP and DeepSpeed, devices from CPUs and multi-node GPUs to TPUs and Apple chips, and precision settings down to FP8. It also compares Lightning with Accelerate, Ray Train and plain PyTorch, lists fixes for common problems, and has reference files on callbacks, distributed strategies and hyperparameter search.

Funktionen

  • Convert a PyTorch loop to Lightning
  • Set up DDP, FSDP or DeepSpeed training
  • Add checkpoint and early-stopping callbacks
  • Troubleshoot common training problems