TorchTL – A minimal training loop abstraction for PyTorch

I'm planning to expand this with more features, e.g. stochastic weight averaging, distributed training support and training anomaly detection. The idea is to stay minimal and dependency-free. Looking for feedback on what features would actually be useful vs just bloat—what do you find yourself rewriting in every project? What makes you reach for (or avoid) libraries like PyTorch Lightning?

Link: github.com/abdimoallim/torchtl