Skip to main content
mshr-h.com

How to Build PyTorch from Source Using uv on Ubuntu 24.04 (February 2026)

PyTorch Build uv CUDA Ubuntu

Prerequisites

  • x86_64
  • uv
  • Ubuntu 24.04 LTS
  • CUDA toolkit + cuDNN (if you want to build with CUDA support)

Install dependencies

sudo apt update
sudo apt install -y \
    build-essential \
    git \
    ccache \
    curl \
    pkg-config \
    libjpeg-dev \
    libpng-dev \
    libopenblas-dev \
    libomp-dev

Create venv, clone source from GitHub

You can use any python interpreter by changing --python 3.12.0 part.

mkdir -p ~/src/pytorch-build
cd ~/src/pytorch-build
uv venv --python 3.12.0 .venv
source .venv/bin/activate
git clone https://github.com/pytorch/pytorch --recursive

Install python dependencies

cd pytorch
uv pip install setuptools wheel ninja cmake
uv pip install --group dev
uv pip install mkl-static mkl-include # if you need Intel MKL

Let CMake search packages in the virtual environment

This ensures CMake finds Python and dependencies inside the virtual environment.

export CMAKE_PREFIX_PATH="${VIRTUAL_ENV}${CMAKE_PREFIX_PATH:+:$CMAKE_PREFIX_PATH}"

Enable ccache

export USE_CCACHE=1

Set max parallel jobs to avoid system overload

CUDA kernel build takes too much memory so I recommend to set 16 parallel jobs if system memory is 64GB.

jobs=$(( $(nproc) - 1 ))
export MAX_JOBS="$jobs"
export CMAKE_BUILD_PARALLEL_LEVEL="$jobs"

(Optional) Disable CUDA support

CUDA support is enabled by default but to disable it, set USE_CUDA=0.

export USE_CUDA=0

Build and install editable

uv pip install --no-build-isolation --verbose --editable .

Verify that successfully installed

python -c "import torch; print(torch.__version__); print(torch.cuda.is_available())"

Output will be something like this:

2.12.0a0+git125edef
True

References