Standardized Workflows

WarpRec abstracts complex workflows into five standardized execution pipelines, all controlled via declarative YAML configuration files. Each pipeline serves a distinct purpose in the experimentation lifecycle:

Pipeline	Command	Purpose
Design	-p design	Rapid prototyping and model debugging. Runs locally without Ray or HPO.
Training	-p train	Full-scale experiments with distributed HPO, cross-validation, and statistical testing via Ray.
Swarm	-p swarm	Aggressive full-scale experiment consuming all available resources in the cluster.
Evaluation	-p eval	Evaluate pre-trained checkpoints or external recommendation files without retraining.
Estimate	-p estimate	Estimate time and memory costs before full execution using lightweight profiling and analytical space estimates.

All pipelines are invoked with the same command structure:

python -m warprec.run -c <config_file>.yml -p <pipeline>

---

Choosing the Right Pipeline

Use Case	Pipeline	Ray Required?	Writer Required?
Debug a new model implementation	Design	No	No
Validate a configuration before full HPO	Design	No	No
Run a full benchmark with HPO	Training	Yes	Yes
Compare models with statistical testing	Training	Yes	Yes
Complete the training as fast as possible	Swarm	Yes	Yes
Evaluate a saved checkpoint on new metrics	Evaluation	No	Optional
Evaluate recommendations from another framework	Evaluation	No	Optional
Estimate RAM, VRAM, and runtime before a full run	Estimate	No	Yes