Merge Language Models with Ease
Getting Started
-
Contributing
-
Issues
-
Website
-
flow-merge-UI
Model merging is an innovative technique that allows you to combine pre-trained and fine-tuned language models (LMs) into new models with unique capabilities.
By merging existing LMs, you can potentially create a new model that inherits the strengths and capabilities of its constituent models. This way, you can explore new model variations and experiment with different combinations without the need for expensive GPU resources or extensive training from scratch.
flow-merge is a fully open-source library written in Python that implements some of the most popular merge methods such as model soups, SLERP, ties-MERGING or DARE. The library is built on top of the Hugging Face transformers library and the deep learning framework Pytorch, and provides a simple and easy-to-use interface to merge models and upload them to the Hugging Face Hub.
flow-merge has been designed to serve both beginners and experts in merging transformer-based language models (LMs). You don't need prior experience with merge methods or advanced knowledge of LMs; a basic understanding of LMs and the command-line interface (CLI) is sufficient.
The library walks you through the merging process, so you can focus on finding the best possible merges without getting bogged down in details of the complex merge methods. Our ultimate goal is to make language model merging simple, flexible, and customizable to your specific needs.
The key features of the library consists of:
- Default parameter settings: Sane default values for the most important parameters based on the experiments in the papers.
- Input validations:
flow-mergevalidates all the user inputs before starting the merge and provides helpful error messages if something is wrong. - CLI and Library: A command-line interface (CLI) for easy merging and uploading of models to the Hugging Face Hub. Also a library that you can use in your own projects.
- Memory efficient:
flow-mergeis designed to be memory efficient, so you can merge large models without running out of memory or without a GPU.
Clone the repository and navigate to the root directory:
# via ssh
git clone [email protected]:flowritecom/flow-merge.git
cd flow-mergeCreate a new python environment and activate it. For example, with conda:
Note
flow-mergerequirespython>=3.10
conda create -n flow-merge python>=3.10 && conda activate flow-mergeflow-merge can be installed with running pip inside the project directory (-e for editable install):
pip install -e .A merge config is a YAML file that defines the models you want to merge and how you want to merge them.
Below is an example of a merge config that merges three models using the addition-task-arithmetic method and saves the merged model to the ./merged_model directory:
method: addition-task-arithmetic
method_global_parameters:
scaling_coefficient: 0.7
normalize: False
base_model: Qwen/Qwen1.5-0.5B
models:
- model: Qwen/Qwen1.5-0.5B
- model: Qwen/Qwen1.5-0.5B-chat
- model: minghaowu/Qwen1.5-0.5B-OpenHermes-2.5
tokenizer:
mode: base
interpolation_method: linear
directory_settings:
cache_dir: null
local_dir: ./models
output_dir: ./merged_model
hf_token:
token: null
trust_remote_code: False
device: cpuThe only required fields are method, and models. The method field specifies the merge method you want to use, and models is a list of models you want to merge. The rest of the fields are optionally and flow-merge will use the default values if they are not provided. For a complete list of the default values, see the config file documentation.
Save the config to a file, for example my_first_merge.yaml.
Merging models with flow-merge is as simple as choosing a YAML template from the examples folder, modifying the paths to the models you want to merge, and running the following command:
flow-merge run --config my_first_merge.yaml --model_name qwen_mergeAfter the merge is complete, you can easily upload the merged model to the Hugging Face Hub by running the following command:
flow-merge upload --model_dir ./merged_model --username <hf_user_id> --model_name qwen_merge --token <hf_token> --private <True/False>You can check the available commands and options by running:
flow-merge --helpYou can display the config yaml schema and the default values by running:
flow-merge schema
# extra tip: pipe to highlighted json with 'flow-merge schema | jq' or 'flow-merge schema | fx'
# where you require either 'jq' or 'fx' installed beforehandYou can optionally validate your config file before running the merge:
flow-merge validate --config my_first_merge.yamlCurrently flow-merge supports most of the popular and proven merge methods.
| Method | Identifier | Paper |
|---|---|---|
| Linear or Model Soups | model-soup |
Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time |
| SLERP | slerp |
- |
| Addition Task Arithmetic | addition-task-arithmetic |
Editing Models with Task Arithmetic |
| Ties-MERGING | ties-merging |
TIES-Merging: Resolving Interference When Merging Models |
| DARE Ties-MERGING | dare-ties |
Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch |
π’ We are working hard on adding more methods to the library.
| Method | Description | Uses a Base Model | Can Merge Multiple Models | Supports Weighted Merge |
|---|---|---|---|---|
| Linear or Model Soups | Averages the weights of the models | No | Yes | Yes |
| SLERP | Smoothly interpolates between the weights of two models using spherical linear interpolation | No | No | No |
| Addition Task Arithmetic | Obtains task vectors or deltas and applies them to the base model | Yes | Yes | Yes |
| Ties-MERGING | It addresses the problem of interference between parameters from different models before merging with addition task arithmetic | Yes | Yes | Yes |
| DARE Ties-MERGING | Similar to Ties-MERGING but it uses a different approach that prunes the task vectors and rescale them. | Yes | Yes | Yes |
flow-merge currently supports merging models that are based on the following architectures:
| Model type | Architecture |
|---|---|
qwen |
QwenForCausalLM |
mistral |
MistralForCausalLM |
llama |
LlamaForCausalLM |
π’ We plan to support many models and architectures more, including encoder models such as BERT-Family models too.
When merging language models, it's crucial to consider the tokenizers involved, as they convert text into tokens that the models can process.
flow-merge currently supports two modes for constructing the tokenizer that is used by the resulting merged model:
base: Default mode. The merged model utilizes the tokenizer of the base model. If no base model is specified in the merged configuration, the first model in the models list is used as the base model.merged: If the tokenizers of the models use different vocabularies, a common vocabulary is created, and a new tokenizer is constructed based on this vocabulary.
If the tokenizers of the models use different vocabularies, flow-merge creates input_ids mappings for the models and linearly interpolates the embedding and language modeling layers.
Currently, only linear interpolation is supported.
Conflicts can arise from special tokens used by different models' tokenizers, such as differing eos_token tokens. In such cases, flow-merge uses the special token of the last model in the list.
Here we have prepared some additional resources to help developers understand the supported merge methods better.
Coming soon..
Coming soon..
Wanna pitch in? We're totally open to contributions for the core flow-merge library as well as any cool integrations built on top of it! Check out our Contribution Guide for all the details on how to get started.
Install conda (refer here for instructions) and make sure it's initialized for your shell. Git clone the repository and spawn the environment from environment.yml.
git clone [email protected]:flowritecom/flow-merge.git; cd flow-merge
conda env create # creates conda env with name flow-merge, python ~3.10 and installs the listed dependencies
conda activate flow-merge
pip install -e . # install flow-merge in editable mode
code . # open your editor, for example vscodeTo easily jump into PRs you can use for example the (Github CLI)[https://cli.github.com/] client gh pr checkout <insert_pr_number>.
Special thanks to these amazing projects that helped us build flow-merge:
Also, a big shoutout to the authors of the papers of the merge methods implemented in flow-merge, and to Charles O. Goddard, creator of mergekit, who inspired us to create our own merging toolkit.
Finally, thanks to Derrick Schultz for the pytorch-tensor-slerp.py gist that helped us implement the SLERP method.
@misc{flowrite_2024_flow_merge,
author = {The Flowrite Team},
title = {flow-merge},
howpublished = {\url{https://https://github.com/flowritecom/flow-merge}},
year = {2024}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent