mergekit is a toolkit for merging pre-trained language models. mergekit uses an out-of-core approach to perform unreasonably elaborate merges in resource-constrained situations. Merges can be run entirely on CPU or accelerated with as little as 8 GB of VRAM. Many merging algorithms are supported, with more coming as they catch my attention.

• Why Merge Models?
• Features
• Installation
• Usage
• Merge Configuration
  • Parameter Specification
  • Tokenizer Configuration
  • Chat Template Configuration
  • Examples
• Merge Methods
• LoRA extraction
• Mixture of Experts merging
• Evolutionary merge methods
• Merge in the Cloud
• Citation

Model merging is a powerful technique that allows combining the strengths of different models without the computational overhead of ensembling or the need for additional training. By operating directly in the weight space of models, merging can:

• Combine multiple specialized models into a single versatile model
• Transfer capabilities between models without access to training data
• Find optimal trade-offs between different model behaviors
• Improve performance while maintaining inference costs
• Create new capabilities through creative model combinations

Unlike traditional ensembling which requires running multiple models, merged models maintain the same inference cost as a single model while often achieving comparable or superior performance.

Key features of mergekit include:

• Supports Llama, Mistral, GPT-NeoX, StableLM, and more
• Many merge methods
• GPU or CPU execution
• Lazy loading of tensors for low memory use
• Interpolated gradients for parameter values (inspired by Gryphe's BlockMerge_Gradient script)
• Piecewise assembly of language models from layers ("Frankenmerging")
• Mixture of Experts merging
• LORA extraction
• Evolutionary merge methods

🌐 GUI Launch Alert 🤗 - We are excited to announce the launch of a mega-GPU backed graphical user interface for mergekit in Arcee! This GUI simplifies the merging process, making it more accessible to a broader audience. Check it out and contribute at the Arcee App. There is also a Hugging Face Space with limited amounts of GPUs.

git clone https://github.com/arcee-ai/mergekit.git
cd mergekit

pip install -e .  # install the package and make scripts available

If the above fails with the error of:

ERROR: File "setup.py" or "setup.cfg" not found. Directory cannot be installed in editable mode:
(A "pyproject.toml" file was found, but editable mode currently requires a setuptools-based build.)

You may need to upgrade pip to > 21.3 with the command python3 -m pip install --upgrade pip

The script mergekit-yaml is the main entry point for mergekit. It takes a YAML configuration file and an output path, like so:

mergekit-yaml path/to/your/config.yml ./output-model-directory [--cuda] [--lazy-unpickle] [--allow-crimes] [... other options]

This will run the merge and write your merged model to ./output-model-directory.

For more information on the arguments accepted by mergekit-yaml run the command mergekit-yaml --help.

When you have a merged model you're happy with, you may want to share it on the Hugging Face Hub. mergekit generates a README.md for your merge with some basic information for a model card. You can edit it to include more details about your merge, like giving it a good name or explaining what it's good at; rewrite it entirely; or use the generated README.md as-is. It is also possible to edit your README.md online once it has been uploaded to the Hub.

Once you're happy with your model card and merged model, you can upload it to the Hugging Face Hub using the huggingface_hub Python library.

# log in to huggingface with an access token (must have write permission)
huggingface-cli login
# upload your model
huggingface-cli upload your_hf_username/my-cool-model ./output-model-directory .

The documentation for huggingface_hub goes into more detail about other options for uploading.

Merge configurations are YAML documents specifying the operations to perform in order to produce your merged model. Below are the primary elements of a configuration file:

• merge_method: Specifies the method to use for merging models. See Merge Methods for a list.
• 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 used for the merging operation.
• tokenizer or tokenizer_source: Determines how to construct a tokenizer for the merged model.
• chat_template: Specifies a chat template for the merged model.

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. models.*.parameters or input_model_parameters - applying to any tensors coming from specific input models
4. parameters - catchall

The tokenizer behavior can be configured in two ways: using the new tokenizer field (recommended) or the legacy tokenizer_source field (maintained for backward compatibility). These fields are mutually exclusive - you should use one or the other, not both.

The tokenizer field provides fine-grained control over vocabulary and embeddings:

tokenizer:
  source: "union"  # or "base" or a specific model path
  tokens:          # Optional: configure specific tokens
    <token_name>:
      source: ...  # Specify embedding source
      force: false # Optional: force this embedding for all models
  pad_to_multiple_of: null  # Optional: pad vocabulary size

The source field determines the vocabulary of the output model:

• union: Combine vocabularies from all input models (default)
• base: Use vocabulary from the base model
• "path/to/model": Use vocabulary from a specific model

When merging models with different vocabularies, mergekit uses smart defaults to handle token embeddings:

• If a token exists in the base model, its embedding is used as the default
• If only one model has the token, that model's embedding is used
• Otherwise, an average of all available embeddings is used

You can override these defaults for specific tokens:

tokenizer:
  source: union
  tokens:
    # Use embedding from a specific model
    <|im_start|>:
      source: "path/to/chatml/model"

    # Force a specific embedding for all models
    <|special|>:
      source: "path/to/model"
      force: true

    # Map a token to another model's token embedding
    <|renamed_token|>:
      source:
        kind: "model_token"
        model: "path/to/model"
        token: "<|original_token|>"  # or use token_id: 1234

Here's how you might preserve both Llama 3 Instruct and ChatML prompt formats when merging models:

tokenizer:
  source: union
  tokens:
    # ChatML tokens
    <|im_start|>:
      source: "chatml_model"
    <|im_end|>:
      source: "chatml_model"

    # Llama 3 tokens - force original embeddings
    <|start_header_id|>:
      source: "llama3_model"
      force: true
    <|end_header_id|>:
      source: "llama3_model"
      force: true
    <|eot_id|>:
      source: "llama3_model"
      force: true

For backward compatibility, the tokenizer_source field is still supported:

tokenizer_source: "union"  # or "base" or a model path

This provides basic tokenizer selection but lacks the fine-grained control of the modern tokenizer field.

The optional chat_template field allows overriding the chat template used for the merged model.

chat_template: "auto"  # or a template name or Jinja2 template

Options include:

• "auto": Automatically select the most common template among input models
• Built-in templates: "alpaca", "chatml", "llama3", "mistral", "exaone"
• A Jinja2 template string for custom formatting

Several examples of merge configurations are available in examples/.

A quick overview of the currently supported merge methods:

The classic merge method - a simple weighted average.

Parameters:

• 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.

Spherically interpolate the parameters of two models. One must be set as base_model.

Parameters:

• t - interpolation factor. At t=0 will return base_model, at t=1 will return the other one.

Interpolates base model with secondary model if similarity is below t. Accepts two models.

Parameters:

• t - similarity threshold

Computes "task vectors" for each model by subtracting a base model. Merges the task vectors linearly and adds back the base. Works great for models that were fine tuned from a common ancestor. Also a super useful mental framework for several of the more involved merge methods.

Parameters: same as Linear, plus:

• lambda - scaling factor applied after weighted sum of task vectors

Builds on the task arithmetic framework. Resolves interference between models by sparsifying the task vectors and applying a sign consensus algorithm. Allows you to merge a larger number of models and retain more of their strengths.

Parameters: same as Task Arithmetic, plus:

• density - fraction of weights in differences from the base model to retain

In the same vein as TIES, sparsifies task vectors to reduce interference. Differs in that DARE uses random pruning with a novel rescaling to better match performance of the original models. DARE can be used either with the sign consensus algorithm of TIES (dare_ties) or without (dare_linear).

Parameters: same as TIES for dare_ties, or Linear for dare_linear

passthrough is a no-op that simply passes input tensors through unmodified. It is meant to be used for layer-stacking type merges where you have only one input model. Useful for frankenmerging.

An extension of task arithmetic that discards both small and extremely large differences from the base model. As with DARE, the Model Breadcrumbs algorithm can be used with (breadcrumbs_ties) or without (breadcrumbs) the sign consensus algorithm of TIES.

Parameters: same as Task Arithmetic, plus:

• density - fraction of weights in differences from the base model to retain
• gamma - fraction of largest magnitude differences to remove

Note that gamma corresponds with the parameter β described in the paper, while density is the final density of the sparsified tensors (related to γ and β by density = 1 - γ - β). For good default values, try density: 0.9 and gamma: 0.01.

Uses some neat geometric properties of fine tuned models to compute good weights for linear interpolation. Requires at least three models, including a base model.

Parameters:

• filter_wise: if true, weight calculation will be per-row rather than per-tensor. Not recommended.

Spherically interpolate between parameters, but with more options and more sensical configuration! Does not require a base model, but can use one to do spherical interpolation of task vectors. Only works with either two models or two plus a base model.

Parameters:

• weight: relative weighting of a given tensor
• nuslerp_flatten: set to false to do row-wise/column-wise interpolation instead of treating tensors as vectors
• nuslerp_row_wise: SLERP row vectors instead of column vectors

To replicate the behavior of the original slerp method, set weight to 1-t and t for your first and second model respectively.

Building upon DARE, DELLA uses adaptive pruning based on parameter magnitudes. DELLA first ranks parameters in each row of delta parameters and assigns drop probabilities inversely proportional to their magnitudes. This allows it to retain more important changes while reducing interference. After pruning, it rescales the remaining parameters similar to DARE. DELLA can be used with (della) or without (della_linear) the sign elect step of TIES

Parameters: same as Task Arithmetic, plus:

• density - fraction of weights in differences from the base model to retain
• epsilon - maximum change in drop probability based on magnitude. Drop probabilities assigned will range from density - epsilon to density + epsilon. (When selecting values for density and epsilon, ensure that the range of probabilities falls within 0 to 1)

SCE introduces adaptive matrix-level merging weights based on parameter variances. SCE first selects the top-k% elements from each parameter matrix that exhibit high variance across all delta parameters. Following this selection, SCE calculates matrix-level merging weights based on the sum of squares of elements in the delta parameters. Finally, it erases minority elements, a step similar to the sign election process in TIES.

Parameters: same as TIES, plus:

• select_topk - fraction of elements with the highest variance in the delta parameters to retain.

Mergekit allows extracting PEFT-compatible low-rank approximations of finetuned models.

mergekit-extract-lora --model finetuned_model_id_or_path --base-model base_model_id_or_path --out-path output_path [--no-lazy-unpickle] [--cuda] [--max-rank=desired_rank] [--sv-epsilon=tol]

The mergekit-moe script supports merging multiple dense models into a mixture of experts, either for direct use or for further training. For more details see the mergekit-moe documentation.

See docs/evolve.md for details.

We host merging on Arcee's cloud GPUs - you can launch a cloud merge in the Arcee App. Or through python - grab an ARCEE_API_KEY:

export ARCEE_API_KEY=<your-api-key> pip install -q arcee-py

import arcee
arcee.merge_yaml("bio-merge","./examples/bio-merge.yml")

Check your merge status at the Arcee App

When complete, either deploy your merge:

arcee.start_deployment("bio-merge", merging="bio-merge")

Or download your merge:

!arcee merging download bio-merge

If you find mergekit useful in your research, please consider citing the paper:

@inproceedings{goddard-etal-2024-arcees,
    title = "Arcee{'}s {M}erge{K}it: A Toolkit for Merging Large Language Models",
    author = "Goddard, Charles  and
      Siriwardhana, Shamane  and
      Ehghaghi, Malikeh  and
      Meyers, Luke  and
      Karpukhin, Vladimir  and
      Benedict, Brian  and
      McQuade, Mark  and
      Solawetz, Jacob",
    editor = "Dernoncourt, Franck  and
      Preo{\c{t}}iuc-Pietro, Daniel  and
      Shimorina, Anastasia",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track",
    month = nov,
    year = "2024",
    address = "Miami, Florida, US",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.emnlp-industry.36",
    doi = "10.18653/v1/2024.emnlp-industry.36",
    pages = "477--485",
    abstract = "The rapid growth of open-source language models provides the opportunity to merge model checkpoints, combining their parameters to improve performance and versatility. Advances in transfer learning have led to numerous task-specific models, which model merging can integrate into powerful multitask models without additional training. MergeKit is an open-source library designed to support this process with an efficient and extensible framework suitable for any hardware. It has facilitated the merging of thousands of models, contributing to some of the world{'}s most powerful open-source model checkpoints. The library is accessible at: https://github.com/arcee-ai/mergekit.",
}