Designing a Resumable PyTorch Training Framework for Forecasting Models

Reported in interviews at Jane Street

Problem statement

How would you design a reusable PyTorch training framework for forecasting? The first team predicts retail sales from store histories, calendar features and planned promotions, with recent history held out for validation. Training uses mixed precision, clipping and early stopping on GPUs that can be interrupted at any time.

Explain reliable resume and sketch the framework interfaces and training loop.

Show hint

Treat the framework as a contract between teams: what does each component need to receive, and what must it hand back so a run can be reproduced or resumed exactly?

Relevant roles: ML Engineer, Deep Learning Engineer and MLOps Engineer.

Related free problems