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.
Features:
- 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
π Call to Evolve - to solve evolutionary merge methods as a community - please see #207.
π GUI Launch Alert π€ - We are excited to announce the launch of a graphical user interface for mergekit in Hugging Face Spaces! This GUI simplifies the merging process, making it more accessible to a broader audience. Check it out and contribute at Hugging Face Spaces - mergekit-community.
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 withmodels
.models
: Defines entire models to be used for merging. This field is mutually exclusive withslices
.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_source
: Determines how to construct a tokenizer 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:
slices.*.sources.parameters
- applying to a specific input sliceslices.*.parameters
- applying to a specific output slicemodels.*.parameters
orinput_model_parameters
- applying to any tensors coming from specific input modelsparameters
- catchall
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.
This functionality is still experimental and may break. Please file an issue if you encounter any issues with it.
Valid values:
base
: use the tokenizer from the base modelunion
: construct a tokenizer with all tokens from all modelsmodel:<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.
Several examples of merge configurations are available in examples/
.
A quick overview of the currently supported merge methods:
Method | merge_method value |
Multi-Model | Uses base model |
---|---|---|---|
Linear (Model Soups) | linear |
β | β |
SLERP | slerp |
β | β |
Task Arithmetic | task_arithmetic |
β | β |
TIES | ties |
β | β |
DARE TIES | dare_ties |
β | β |
DARE Task Arithmetic | dare_linear |
β | β |
Passthrough | passthrough |
β | β |
Model Breadcrumbs | breadcrumbs |
β | β |
Model Breadcrumbs + TIES | breadcrumbs_ties |
β | β |
Model Stock | model_stock |
β | β |
DELLA | della |
β | β |
DELLA Task Arithmetic | della_linear |
β | β |
The classic merge method - a simple weighted average.
Parameters:
weight
- relative (or absolute ifnormalize=False
) weighting of a given tensornormalize
- 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. Att=0
will returnbase_model
, att=1
will return the other one.
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
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
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 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 retaingamma
- fraction of largest magnitude differences to remove
Note that gamma
corresponds with the parameter Ξ²
described in the paper, while density
is the final density of the sparsified tensors (related to Ξ³
and Ξ²
by density = 1 - Ξ³ - Ξ²
). 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.
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 retainepsilon
- maximum change in drop probability based on magnitude. Drop probabilities assigned will range fromdensity - epsilon
todensity + epsilon
. (When selecting values fordensity
andepsilon
, 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.
Mergekit allows extracting PEFT-compatible low-rank approximations of finetuned models.
mergekit-extract-lora finetuned_model_id_or_path base_model_id_or_path output_path [--no-lazy-unpickle] --rank=desired_rank
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.
We now have a paper you can cite for the MergeKit library:
@article{goddard2024arcee,
title={Arcee's MergeKit: A Toolkit for Merging Large Language Models},
author={Goddard, Charles and Siriwardhana, Shamane and Ehghaghi, Malikeh and Meyers, Luke and Karpukhin, Vlad and Benedict, Brian and McQuade, Mark and Solawetz, Jacob},
journal={arXiv preprint arXiv:2403.13257},
year={2024}
}