atft-training

Hyperparameter optimization automates tuning of learning rate, batch size, and other critical parameters across GPU resources. This Skill orchestrates production-grade training runs for the ATFT-GAT-FAN forecaster, leveraging Optuna for sweeps and 80GB GPU headroom for efficient exploration while monitoring real-time performance.

Yes. This Skill provides resume/stop controls, real-time monitoring dashboards via Tensorboard and Weights & Biases, GPU utilization thresholds, and metric aggregation to track and manage extended training sessions, ensuring optimal performance and timely job completion.

Dataset and version parity validation, environment health checks, and GPU utilization thresholds are automated before training begins. This Skill performs these checks to prevent training on stale or mismatched data, ensuring correctness and reproducibility across runs.

This Skill automatically compiles training runs with TorchInductor and FlashAttention2, optimizing GPU performance without manual configuration. It handles environment setup, logging aggregation, and post-training handoff procedures as part of the production pipeline.

This Skill integrates MLflow and Weights & Biases for experiment tracking, metric logging, and reproducibility. These tools aggregate results across hyperparameter sweeps and multiple runs, enabling systematic comparison and model selection without manual record-keeping.