ZenithTune
ZenithTune is a library that brings out the full potential of applications and tunes their performance to the zenith. By using the tuners provided by ZenithTune according to your needs, you can achieve the best performance while minimizing user coding.
Features
ZenithTune offers a variety of optimization methods aimed at improving overall application performance, including reducing execution time and optimizing command output. In contrast to common tools that primarily focus on improving model accuracy, it takes a more comprehensive approach.
It is also designed to run seamlessly in a wide variety of distributed environments. While many existing tools are designed for use on a single computer, ZenithTune supports performance improvements in distributed environments.
ZenithTune is a library specialized in improving application performance and has the ability to fine-tune hardware behavior such as CPU core allocation. This is a feature not found in other similar tools.
Finally, users can focus solely on optimization strategies. While many existing tools require users to write non-essential code outside of optimization themselves, ZenithTune is designed with the goal of minimizing the effort required from users.
Setup
Choose a Tuning Method
ZenithTune provides two tuning methods.
| No-Code Tuning | Python Script Tuning | |
|---|---|---|
| Python Code | Not required (CLI only) | Required (define your own objective function) |
| Usage | Specify a preset name and command | Configure tuner and objective function via Python API |
| Best for | Frameworks with available presets | Custom evaluation logic or fine-grained control |
| Supported Frameworks | Megatron-LM / ms-swift | Any (write your own objective function) |
If a preset is available for the framework you are using, No-Code Tuning is the easiest option. For frameworks without presets, or when you need custom evaluation metrics, use Python Script Tuning.
No-Code Tuning
Simply specify a preset and command to automatically perform parameter search, application, and evaluation.
zenithtune optimize --preset megatron --args n_gpus=8,gbs=64 --n-trials 50 -- bash train.sh
Python Script Tuning
Define your own objective function using the Python API for flexible tuning.