speergekit

This is a library for merging pretrained speech models (TTS & ASR). It is a refactor of the original speergekit with a more powerful and flexible YAML-based configuration system, inspired by mergekit.

Installation

To install the library, clone the repository and install it in editable mode:

git clone https://github.com/ehristoforu/speergekit.git
cd speergekit
pip install -e .

Usage

To perform a merge, you need to create a YAML configuration file that specifies the models to merge, the merge method, and other parameters. Then, you can run the merge script:

speergekit path/to/your/config.yml ./output-directory --cuda

This will merge the models according to your configuration and save the resulting model in the ./output-directory.

Merge Configuration

The YAML configuration file has the following structure:

• models: A list of models to merge. Each model can be specified by a local path or a HuggingFace repository ID (e.g., openai/whisper-tiny).
• base_model: The model to use as the base for the merge.
• merge_method: The merge method to use. Currently, linear and task_arithmetic are supported.
• parameters: A dictionary of parameters for the merge method.
• dtype: The data type to use for the merge. float16 is recommended.

Example Configurations

You can find example configurations in the examples/ directory:

• linear.yml: A simple weighted average of two models.
• task_arithmetic.yml: A task arithmetic merge. This method applies the formula (model_1 - base_model) * scaling_factor + model_2. The models list must contain exactly two models: the finetuned model first, and the target model second.