Skip to content

Readme.md

Latest commit

 

History

History
184 lines (133 loc) · 6.19 KB

Readme.md

File metadata and controls

184 lines (133 loc) · 6.19 KB
Raw

DeepSpeed Windows Wheels with RTX 50 Support

Prebuilt DeepSpeed wheels for Windows with NVIDIA GPU support.
Supports Python 3.9 – 3.12 and GTX 10 - RTX 50 series. Compiled with pytorch 2.7, 2.8 and cuda 12.8

“The wheels already include CUDA 12.8 support – no separate CUDA toolkit installation required.”

📜 Notes

These wheels are provided to simplify DeepSpeed installation on Windows without manual building.
Official DeepSpeed repository: https://github.com/deepspeedai/DeepSpeed

⚠ Windows Notice
You may see warnings like:
LINK : fatal error LNK1181: cannot open file "aio.lib"
LINK : fatal error LNK1181: cannot open file 'cufile.lib'
This is normal on systems without the full CUDA Toolkit.
DeepSpeed will still work for all regular training and inference tasks.

🖥️ Requirements

  • Windows 10 / 11 (x64)
  • NVIDIA GPU with CUDA-capable drivers
  • Python 3.9 – 3.12
  • pip >= 21.0

🔧 Preparation

  • 1.) Create Project
    • Create new folder: mkdir MyProject
    • Navigate to new folder: cd MyProject
    • Optional install uv: conda install -y -c conda-forge uv
  • 2.) Python Environment
    • Create Environment: python -m venv env
    • Activate Environment: env\Scripts\activate
  • 3.) Pytorch 2.7.x+cu128 or 2.8.x+cu128

    • Install pytorch 2.7.0, pytorch 2.7.1 or pytorch 2.8.0 with pip:

      pytorch 2.7.0

      pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128

      pytorch 2.7.1

      pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128

      pytorch 2.8.0

      pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128

    • Or install with uv:

      pytorch 2.7.0

      uv pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128

      pytorch 2.7.1

      uv pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128

      pytorch 2.8.0

      uv pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128

📦 Installation

Search the appropriate .whl file for your pytorch and python version from the Releases page.

Install it with pip

pip install https://github.com/6Morpheus6/deepspeed-windows-wheels/releases/download/<tag>deepspeed‑<version>‑<torch version>torch+cu128-cp<python version>‑cp<python version>‑win_amd64.whl

For Example:

pip install https://github.com/6Morpheus6/deepspeed-windows-wheels/releases/download/v0.17.5/deepspeed-0.17.5+e1560d84-2.7torch_cu128-cp310-cp310-win_amd64.whl

Or install with uv

uv pip install https://github.com/6Morpheus6/deepspeed-windows-wheels/releases/download/<tag>deepspeed‑<version>‑<torch version>torch+cu128-cp<python version>‑cp<python version>‑win_amd64.whl

For Example:

uv pip install https://github.com/6Morpheus6/deepspeed-windows-wheels/releases/download/v0.17.5/deepspeed-0.17.5+e1560d84-2.7torch_cu128-cp310-cp310-win_amd64.whl

Build from source

Prerequistits

  • Git
  • Python
  • Miniconda
  • NVIDIA Toolkit
  • VS Buildtools 2019

Preparation

  • Open Anaconda Powershell in your Windows Start menu.
  • Create a new project folder mkdir C:\MyProject
  • Navigate to your project folder cd C:\MyProject
  • Create a python environment python -m venv env
  • Activate the environment env\Scripts\activate
  • Install torch, e.g.: pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128
  • Install build dependencies pip install setuptools wheel ninja packaging py-cpuinfo psutil
  • Optional: Install dependencies to test your wheel later on pip install tqdm pydantic msgpack hjson einops
  • Clone DeepSpeed git clone https://github.com/deepspeedai/DeepSpeed

Compile

  • Open x64 Native Tools Command Prompt for VS 2019 in your Windows Start menu.
  • Navigate to your project folder cd C:\MyProject
  • Activate the environment call env\Scripts\activate
  • Make Windows SDK accessible set DISTUTILS_USE_SDK=1
  • Set CUDA_HOME (modify the path and toolkit version accordingly)
    set CUDA_HOME=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8
  • Add CUDA_HOME to PATH set PATH=%CUDA_HOME%\bin;%PATH%
  • If git is installed globally it is still accessible. If git is installed as library in conda,
    you need to make it accessible (modify the path accordingly)
    set PATH=C:\Users<username>\AppData\Local\miniconda3\Library\bin;%PATH%
  • Specify the architecture(s) you want to build the wheel for. E.g. if you want to build for all architectures:
    set TORCH_CUDA_ARCH_LIST=6.1;7.5;8.6;8.9;12.0
    or if you only want to build for RTX 40 series (Ada Lovelace):
    set TORCH_CUDA_ARCH_LIST=8.9
  • Set DeepSpeed environmet variables.
set DS_BUILD_OPS=1
set DS_BUILD_GDS=0
set DS_BUILD_AIO=0
set DS_BUILD_CUTLASS_OPS=0
set DS_BUILD_SPARSE_ATTN=0
set DS_BUILD_SPARSE_UTILS=0
set DS_BUILD_FP_QUANTIZER=0
set DS_BUILD_DEEP_COMPILE=0
set DS_BUILD_EVOFORMER_ATTN=0
set DS_BUILD_RAGGED_DEVICE_OPS=0
  • Navigate into the DeepSpeed folder cd DeepSpeed
  • Compile DeepSpeed python setup.py bdist_wheel

The finished wheel can be found in C:\MyProject\DeepSpeed\dist

Test the wheel

  • Install DeepSpeed into your environment (modify the name accordingly to match your wheel if necessary)
    pip install .\dist\deepspeed-0.17.5+e1560d84-cp310-cp310-win_amd64.whl
  • Print Report python -m deepspeed.env_report

Successful build

alt text

⭐ Support

If this project is useful to you, please consider giving it a ⭐ on GitHub!