mergekitty 0.2.3

Tools for merging pre-trained large language models

mergekitty

mergekitty is a toolkit for merging pre-trained language models. mergekitty 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.

Contents

• Reasons for the fork
• Breaking changes
• Why Merge Models?
• Features
• Installation
• Usage
• Merge Configuration
  • Parameter Specification
  • Tokenizer Configuration
  • Chat Template Configuration
  • Examples
• Merge Methods
• LoRA extraction
• Mixture of Experts merging
• Development
• Citation

Reasons for the fork

This project is a fork of mergekit by Arcee.ai, and originally created by Charles Goddard. This fork was created from the last LGPL licensed commit from the original mergekit repository, mainly due to it's anti-community switch to the source-available BSL license (in light of how much work the community did to make mergekit what it is today).

Update: Mergekit itself has switched back to the LGPL, however it now has a CLA attached for all contributions, meaning they are given ownership of your contributions and allowed to change the license back at any time.

Breaking changes

This fork is a work in progress. Here are some of the breaking changes we've made so far:

• support for VLMs has been added; the architecture files are incompatible with mergekit's and are implemented differently. (more specifically, templating support for pre- and post-weights has been added)
• there is a lora merging script based on the task graph system that can be ran with mergekitty-merge-lora
• ALL SCRIPTS, LIBRARY NAMES, ETC HAVE BEEN RENAMED TO mergekitty. THIS ONE IS BIG (but you can just replace mergekit with mergekitty in your codebase/scripts and it will work)
• legacy tokenizer copying has been removed -- tokenizer_source now defaults to "base", which is the same as the legacy functionality
• bakllama and mergekit-legacy have been removed. if someone can give me a real (not fake) use for them then they will be added back.
• mergekit-evolve has been removed. if anyone can give me a real reason why it should remain a wart tacked onto the rest of mergekit instead of an out-of-tree tool that depends on mergekit, I'll reconsider.
• nuslerp has been renamed to slerp (and the original slerp has been removed), and it now supports using the parameter t (SLERP behavior) OR tensor parameter weight (NuSLERP behavior) to specify the weighting of a given tensor.
• moved from black and isort to ruff for formatting, as well as enabling linting for the whole package.
• moved from setuptools to hatch for building and distributing the package.

Why Merge Models?

Model merging is chaos magick that has been proven multiple times to create models that are better than the sum of their parts.

Features

Key features of mergekitty include:

• Supports new (Llama 3, Qwen 3 (Dense and MoE), Mistral, GLM4) and old (GPT-NeoX, BERT) model architectures
• Many merge methods. An excessive amount even
• 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 (i.e. "Frankenmerging", ala Goliath or Midnight Miqu)
• Mixture of Experts merging
• LORA extraction

Installation

mergekitty is available on PyPI.

Using uv to install in a separate tool environment (recommended):

uv tool install mergekitty

Using pip:

pip install mergekitty

From source:

git clone https://github.com/allura-org/mergekitty.git
cd mergekitty

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

Usage

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

mergekitty-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 mergekitty-yaml run the command mergekitty-yaml --help.

Uploading to Huggingface

When you have a merged model you're happy with, you may want to share it on Huggingface. mergekitty 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 HF.

Once you're happy with your model card and merged model, you can upload it to Huggingface 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 Configuration

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.

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

Tokenizer Configuration

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.

Modern Configuration (tokenizer)

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

Tokenizer Source

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

Token Embedding Handling

When merging models with different vocabularies, mergekitty 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

Practical Example

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

Legacy Configuration (tokenizer_source)

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.

Chat Template Configuration

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

Examples

Several examples of merge configurations are available in examples/.

Merge Methods

A quick overview of the currently supported merge methods:

Method	merge_method value	Multi-Model	Uses base model
Linear (Model Soups)	linear	✅	❌
SLERP	slerp	✅*	✅
Nearswap	nearswap	❌	✅
Task Arithmetic	task_arithmetic	✅	✅
TIES	ties	✅	✅
DARE TIES	dare_ties	✅	✅
DARE Task Arithmetic	dare_linear	✅	✅
Passthrough	passthrough	❌	❌
Model Breadcrumbs	breadcrumbs	✅	✅
Model Breadcrumbs + TIES	breadcrumbs_ties	✅	✅
Model Stock	model_stock	✅	✅
DELLA	della	✅	✅
DELLA Task Arithmetic	della_linear	✅	✅
SCE	sce	✅	✅
* only supports two to three models

Linear

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.

SLERP

Spherically interpolate model parameters with support for both traditional interpolation (t) and tensor-specific weighting (weight). Works with two models or two task vectors and an optional base model.

Parameters:

One of the following parameters must be specified:

• t - interpolation factor. At t=0 will return base_model, at t=1 will return the other model. if specified, weight will be ignored. (only supports one base model and one non-base model)
• weight - relative weighting of a given tensor. (supports one base model, and either one or two non-base models)

Additionally, the following parameters are supported:

• 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

Nearswap

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

Parameters:

• t - similarity threshold

Task Arithmetic

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

TIES

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 Linear, plus:

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

DARE

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

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.

Model Breadcrumbs

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 Linear, 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.

Model Stock

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.

DELLA

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 Linear, 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)
• lambda - scaling factor for the final merged delta parameters before merging with the base parameters.

SCE

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:

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

LoRA extraction

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

Usage

mergekitty-extract-lora finetuned_model_id_or_path base_model_id_or_path output_path [--no-lazy-unpickle] --rank=desired_rank

Mixture of Experts merging

The mergekitty-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.

Development

mergekitty is developed using Hatch and UV. My recommended setup:

# after installing uv:
uv tool install hatch
hatch test # run tests
hatch run lint # run ruff
hatch run format # run ruff format
hatch run mergekitty-yaml examples/bio-merge.yml ./bio-merge --cuda # run a test merge

Citation

If you find mergekitty useful in your research, please consider citing the original mergekit 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.",
}