ml-train

To run PyTorch Lightning training with Hydra, execute your config-driven entrypoint and pass CLI overrides like experiment=basic_training. This Skill applies Hydra templates to manage single-GPU, multi-GPU, and distributed training settings.

Yes, you can resume PyTorch Lightning training from a checkpoint using Hydra configuration overrides. This Skill manages checkpoint loading and continuation so you can recover from interruptions without restarting runs.

Yes, this approach supports distributed multi-node and large-model FSDP training. It uses Lightning trainer settings for device strategy, allowing you to scale training across multiple GPUs and nodes through Hydra configuration templates.

To track PyTorch Lightning metrics and GPU utilization, this Skill integrates Weights and Biases (W&B) for experiment tracking. It logs losses and learning rate signals to W&B dashboards, providing real-time visibility into training progress.

The best way to handle NaN loss and OOM errors is using built-in resilience guidance. This Skill provides strategies for mixed precision, gradient accumulation, and early stopping to mitigate memory issues and prevent unstable training runs.

To profile data-loading bottlenecks, this Skill provides operational monitoring and observability features. It checks GPU utilization and uses Lightning metrics to identify inefficiencies in your distributed training pipeline.