ml-train

What problem does it solve?

This Skill helps you run PyTorch Lightning + Hydra training jobs that start correctly, save useful checkpoints, and surface progress through experiment tracking—so you can iterate faster instead of babysitting runs.

Core Features & Use Cases

• Execute training with Hydra configs: Runs training using configuration templates (single GPU, multi-GPU, distributed, FSDP) and supports CLI overrides and resuming from checkpoints.
• Operational monitoring and observability: Supports real-time visibility via Lightning metrics, GPU utilization checks, and W&B dashboards, including logging of losses and learning rate signals.
• Built-in resilience for common issues: Provides guidance for NaN/inf loss, OOM mitigation (mixed precision, gradient accumulation), overfitting prevention (early stopping), and profiling data-loading bottlenecks.

Quick Start

Run training with the default Hydra experiment template by executing: python src/train.py experiment=basic_training