Skip to main content

Installation

Supported platforms​

  • Linux x86_64
  • NVIDIA GPU

GPU / CUDA​

Surogate requires a recent NVIDIA driver and CUDA libraries.

Commonly used CUDA versions:

  • CUDA 12.8
  • CUDA 12.9
  • CUDA 13.x

Multi-GPU training requires NCCL.

Python​

Python 3.12 is the default for the published wheels and the install script

curl -LsSf https://surogate.ai/install.sh | sh

What the script does (high level):

  • Installs uv if missing
  • Creates a local .venv using Python 3.12
  • Detects your CUDA version and installs a matching Surogate wheel
  • Downloads example configs into ./examples/ (if not already present)

Activate the environment:

source .venv/bin/activate

Option B: Build from source (developers)​

Prerequisites:

  • CUDA toolkit (12.8/12.9/13.x)
  • NCCL development libraries

From the repository root:

uv pip install -e .

Verify installation​

After install, these should work:

surogate --help
surogate sft --help
surogate pt --help

See also​