# Differential Transformer
An open source community implementation of the model from "DIFFERENTIAL TRANSFORMER" paper by Microsoft. [Paper Link](https://arxiv.org/abs/2410.05258). "Differential attention takes the difference between two softmax attention functions to eliminate attention noise. The idea is analogous to differential amplifiers [19] proposed in electrical engineering,where the difference between two signals is used as output, so that we can null out the common-mode noise of the input. In addition, the design of noise-canceling headphones is based on a similar idea. We can directly reuse FlashAttention [8] as described in Appendix A, which significantly improves model efficiency."
[![Join our Discord](https://img.shields.io/badge/Discord-Join%20our%20server-5865F2?style=for-the-badge&logo=discord&logoColor=white)](https://discord.gg/agora-999382051935506503) [![Subscribe on YouTube](https://img.shields.io/badge/YouTube-Subscribe-red?style=for-the-badge&logo=youtube&logoColor=white)](https://www.youtube.com/@kyegomez3242) [![Connect on LinkedIn](https://img.shields.io/badge/LinkedIn-Connect-blue?style=for-the-badge&logo=linkedin&logoColor=white)](https://www.linkedin.com/in/kye-g-38759a207/) [![Follow on X.com](https://img.shields.io/badge/X.com-Follow-1DA1F2?style=for-the-badge&logo=x&logoColor=white)](https://x.com/kyegomezb)
## Install
```bash
$ pip3 install differential-transformers
```
## Usage Transformer
```python
import torch
from differential_transformer.main import DifferentialTransformer
from loguru import logger
# Example usage:
# Example dimensions
batch_size = 32
seq_len = 128
embedding_dim = 64
h = 8
λ = 0.1
λinit = 0.05
# Create random input tensor
x = torch.randint(0, 256, (1, 1024))
# Instantiate and run the multi-head attention
multi_head = DifferentialTransformer(heads=h, dim=embedding_dim, λinit=λinit)
output = multi_head(x, λ=λ)
logger.info(f"Output shape: {output.shape}")
```
# License
MIT
## Citation
```bibtex
@misc{ye2024differentialtransformer,
title={Differential Transformer},
author={Tianzhu Ye and Li Dong and Yuqing Xia and Yutao Sun and Yi Zhu and Gao Huang and Furu Wei},
year={2024},
eprint={2410.05258},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2410.05258},
}
```
Raw data
{
"_id": null,
"home_page": "https://github.com/kyegomez/DifferentialTransformer",
"name": "differential-transformer",
"maintainer": null,
"docs_url": null,
"requires_python": "<4.0,>=3.10",
"maintainer_email": null,
"keywords": "artificial intelligence, deep learning, optimizers, Prompt Engineering",
"author": "Kye Gomez",
"author_email": "kye@apac.ai",
"download_url": "https://files.pythonhosted.org/packages/97/ee/2c9081c5b6cbf3a903d4d2f4f832fc87380bef8dff0c5e60bbd1cda753b5/differential_transformer-0.0.3.tar.gz",
"platform": null,
"description": "\n# Differential Transformer \n\nAn open source community implementation of the model from \"DIFFERENTIAL TRANSFORMER\" paper by Microsoft. [Paper Link](https://arxiv.org/abs/2410.05258). \"Differential attention takes the difference between two softmax attention functions to eliminate attention noise. The idea is analogous to differential amplifiers [19] proposed in electrical engineering,where the difference between two signals is used as output, so that we can null out the common-mode noise of the input. In addition, the design of noise-canceling headphones is based on a similar idea. We can directly reuse FlashAttention [8] as described in Appendix A, which significantly improves model efficiency.\"\n\n\n\n[![Join our Discord](https://img.shields.io/badge/Discord-Join%20our%20server-5865F2?style=for-the-badge&logo=discord&logoColor=white)](https://discord.gg/agora-999382051935506503) [![Subscribe on YouTube](https://img.shields.io/badge/YouTube-Subscribe-red?style=for-the-badge&logo=youtube&logoColor=white)](https://www.youtube.com/@kyegomez3242) [![Connect on LinkedIn](https://img.shields.io/badge/LinkedIn-Connect-blue?style=for-the-badge&logo=linkedin&logoColor=white)](https://www.linkedin.com/in/kye-g-38759a207/) [![Follow on X.com](https://img.shields.io/badge/X.com-Follow-1DA1F2?style=for-the-badge&logo=x&logoColor=white)](https://x.com/kyegomezb)\n\n\n## Install\n\n```bash\n$ pip3 install differential-transformers\n```\n\n## Usage Transformer\n\n```python\n\nimport torch\nfrom differential_transformer.main import DifferentialTransformer\nfrom loguru import logger\n\n# Example usage:\n# Example dimensions\nbatch_size = 32\nseq_len = 128\nembedding_dim = 64\nh = 8\n\u03bb = 0.1\n\u03bbinit = 0.05\n\n# Create random input tensor\nx = torch.randint(0, 256, (1, 1024))\n\n# Instantiate and run the multi-head attention\nmulti_head = DifferentialTransformer(heads=h, dim=embedding_dim, \u03bbinit=\u03bbinit)\noutput = multi_head(x, \u03bb=\u03bb)\n\nlogger.info(f\"Output shape: {output.shape}\")\n\n\n```\n\n# License\nMIT\n\n\n## Citation\n\n\n```bibtex\n@misc{ye2024differentialtransformer,\n title={Differential Transformer}, \n author={Tianzhu Ye and Li Dong and Yuqing Xia and Yutao Sun and Yi Zhu and Gao Huang and Furu Wei},\n year={2024},\n eprint={2410.05258},\n archivePrefix={arXiv},\n primaryClass={cs.CL},\n url={https://arxiv.org/abs/2410.05258}, \n}\n\n```",
"bugtrack_url": null,
"license": "MIT",
"summary": "differential-transformer - Pytorch",
"version": "0.0.3",
"project_urls": {
"Documentation": "https://github.com/kyegomez/DifferentialTransformer",
"Homepage": "https://github.com/kyegomez/DifferentialTransformer",
"Repository": "https://github.com/kyegomez/DifferentialTransformer"
},
"split_keywords": [
"artificial intelligence",
" deep learning",
" optimizers",
" prompt engineering"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "c9d41eed50fd3298090d975e821df436a4dc572352d1ef978c9d118a4b26edba",
"md5": "1fa3d1ec48797864b86eb050b52b42b5",
"sha256": "c99c054b086f6dd668d6dda0b1f66b499e1ad80ad8824a9ca99d11efbf4b5ab2"
},
"downloads": -1,
"filename": "differential_transformer-0.0.3-py3-none-any.whl",
"has_sig": false,
"md5_digest": "1fa3d1ec48797864b86eb050b52b42b5",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": "<4.0,>=3.10",
"size": 5811,
"upload_time": "2024-10-12T21:07:26",
"upload_time_iso_8601": "2024-10-12T21:07:26.006153Z",
"url": "https://files.pythonhosted.org/packages/c9/d4/1eed50fd3298090d975e821df436a4dc572352d1ef978c9d118a4b26edba/differential_transformer-0.0.3-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "97ee2c9081c5b6cbf3a903d4d2f4f832fc87380bef8dff0c5e60bbd1cda753b5",
"md5": "f713f28bd886136c55f691ac1e6278e5",
"sha256": "cb6acb67ad9ee80802c4e9a410de65c483e6bb25f7aa993ae255204b3de141f5"
},
"downloads": -1,
"filename": "differential_transformer-0.0.3.tar.gz",
"has_sig": false,
"md5_digest": "f713f28bd886136c55f691ac1e6278e5",
"packagetype": "sdist",
"python_version": "source",
"requires_python": "<4.0,>=3.10",
"size": 5202,
"upload_time": "2024-10-12T21:07:27",
"upload_time_iso_8601": "2024-10-12T21:07:27.540955Z",
"url": "https://files.pythonhosted.org/packages/97/ee/2c9081c5b6cbf3a903d4d2f4f832fc87380bef8dff0c5e60bbd1cda753b5/differential_transformer-0.0.3.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-10-12 21:07:27",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "kyegomez",
"github_project": "DifferentialTransformer",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"requirements": [
{
"name": "torch",
"specs": []
},
{
"name": "zetascale",
"specs": []
},
{
"name": "swarms",
"specs": []
}
],
"lcname": "differential-transformer"
}