/mergekit

Tools for merging pretrained large language models.

Primary LanguagePythonGNU Lesser General Public License v3.0LGPL-3.0

mergekit

mergekit is a toolkit for merging pre-trained language models, using a variety of merge methods including TIES, linear, and slerp merging. The toolkit also enables piecewise assembly of a language model from layers.

Run pip install -e . to install the package and make the scripts available.

The script mergekit-yaml takes a YAML configuration file defining the operations to perform.

Configuration

Below are the primary elements of a configuration file:

  • merge_method: Specifies the method to use for merging models. Can be one of 'ties', 'linear', 'slerp', or 'passthrough'.
  • slices: Defines slices of layers from different models to be used. This field is mutually exclusive with models.
  • models: Defines entire models to be used for merging. This field is mutually exclusive with slices.
  • base_model: Specifies the base model used in some merging methods.
  • parameters: Holds various parameters such as weights and densities, which can also be specified at different levels of the configuration.
  • dtype: Specifies the data type for the merging operation.
  • tokenizer_source: Determines how to construct a tokenizer for the merged model.

Parameter Specification

Parameters are flexible and can be set with varying precedence. They can be specified conditionally using tensor name filters, which allows finer control such as differentiating between attention heads and fully connected layers.

Parameters can be specified as:

  • Scalars: Single floating-point values.
  • Gradients: List of floating-point values, specifying an interpolated gradient.

The parameters can be set at different levels, with decreasing precedence as follows:

  1. slices.*.sources.parameters - applying to a specific input slice
  2. slices.*.parameters - applying to a specific output slice
  3. input_model_parameters - applying to any tensors coming from specific input models
  4. parameters - catchall

Merge Methods

Requires a base model. Parameters:

  • density - fraction of weights in differences from the base model to retain
  • weight - relative (or absolute if normalize=False) weighting of a given tensor
  • normalize - if true, the weights of all models contributing to a tensor will be normalized. Default behavior.

Linear

Does not require a base model. Takes parameters weight and normalize, with same definition as above.

SLERP

Requires exactly two models, one of which must be the base model. Takes one parameter - t - the interpolation factor from the base model to the secondary model.

Tokenizer Source

The tokenizer_source field of a configuration file determines what tokenizer is used by the merged model. This also effects how embeddings and language model heads are merged.

Valid values:

  • base: use the tokenizer from the base model
  • union: construct a tokenizer with all tokens from all models
  • model:<model_path>: use the tokenizer from a specific model

If set, mergekit will find a mapping between each model's vocabulary and the output tokenizer. This allows models with different vocabularies or added tokens to be meaningfully merged.

tokenizer_source is compatible with all merge methods, but when used lm_head/embed_tokens will be merged linearly. For two-model merges, the embed_slerp parameter can be set to true to use SLERP instead.

If the tokenizer_source field is not set, mergekit will fall back to its legacy default behavior. The tokenizer for the base model (or first model in the merge, if no base model is specified) will be copied to the output directory. The parameter matrices for lm_head/embed_tokens will be truncated to the smallest size present in the merge. In most cases this corresponds to using the tokenizer for the base model.

Examples

  • Simple linear merge of multiple models:

    models:
      - model: psmathur/orca_mini_v3_13b
        parameters:
          weight: 1.0
      - model: WizardLM/WizardLM-13B-V1.2
        parameters:
          weight: 0.3
      - model: garage-bAInd/Platypus2-13B
        parameters:
          weight: 0.5
    merge_method: linear
    dtype: float16
  • bakllama.py style layer recombination:

    slices:
      - sources:
        - model: psmathur/orca_mini_v3_13b
          layer_range: [0, 24]
      - sources:
        - model: garage-bAInd/Platypus2-13B
          layer_range: [20, 40]
    merge_method: passthrough
    dtype: float16
  • Gradient SLERP with different weights for mlp/self attention:

    slices:
      - sources:
          - model: psmathur/orca_mini_v3_13b
            layer_range: [0, 40]
          - model: garage-bAInd/Platypus2-13B
            layer_range: [0, 40]
    merge_method: slerp
    base_model: psmathur/orca_mini_v3_13b
    parameters:
      t:
        - filter: self_attn
          value: [0, 0.5, 0.3, 0.7, 1]
        - filter: mlp
          value: [1, 0.5, 0.7, 0.3, 0]
        - value: 0.5 # fallback for rest of tensors
    dtype: float16

Usage

Once you have created the YAML configuration file, run mergekit-yaml with the config file and output path as arguments:

mergekit-yaml path/to/your/config.yml ./output-model-directory [--cuda]

Legacy Wrappers

Mergekit originally featured two separate scripts with different inputs. The functionality of these is maintained in the mergekit-legacy and bakllama wrappers. Example usage:

mergekit-legacy ./output-model --base-model TheBloke/Llama-2-13B-fp16 --cuda \
    --merge WizardLM/WizardLM-13B-V1.2 --weight 0.3 --density 0.5 \
    --merge garage-bAInd/Platypus2-13B --weight 0.5 --density 0.5

mergekit-legacy can output a YAML configuration for easy migration with the --print-yaml option.