No-Code Tuning
Run hyperparameter tuning from the command line without writing Python scripts. Simply specify a predefined preset for the target command, and hyperparameter search, command execution, and performance evaluation are performed automatically. You can start tuning within minutes.
- optimize: Specify an existing preset and tune the target command's hyperparameters
- analyze: Visualize tuning results with graphs
- apply: Apply the best parameters found during tuning and run the command
Supported Frameworks
The following framework presets are currently available:
| Framework | Optimization Target | Search Strategy |
|---|---|---|
| Megatron family (Megatron-LM, ms-swift) | Training throughput TFLOP/s/GPU | Heuristic search based on parallelization strategy knowledge |
Guide
- Learn how to use no-code tuning
- Component reference
- General-Purpose Strategies - Description of random, grid, and optuna strategies
- Customize no-code tuning for specific use cases
- Customization - Creating custom presets and evaluators