Skip to main content

Installation

This page installs the pie CLI. After this, see Initial setup to create a config and run your first prompt.

Install​

On Linux and macOS:

curl -fsSL https://pie-project.org/install.sh | bash

On Windows (PowerShell):

irm https://pie-project.org/install.ps1 | iex

The installer downloads a release archive and installs pie into ~/.local/bin on Unix or %LOCALAPPDATA%\Pie\bin on Windows. It picks a build for your platform automatically:

  • Linux (x86_64 or arm64) with an NVIDIA GPU: cuda13.0 (driver 580+) or cuda12.8 (driver 525+). The installer also reads your GPU's compute capability (nvidia-smi) and downloads the matching per-compute-capability binary — e.g. sm90 for H100, sm100 for B200/GB200, sm120 for RTX 50 — so the download is small and architecture-optimal. It falls back to an all-architecture build if your exact capability isn't published.
  • Other Linux, macOS Apple Silicon: portable. GPU via Vulkan (Linux) or Metal (macOS), with a built-in CPU fallback.
  • Windows: vulkan (CPU + any Vulkan-capable GPU), or cuda for NVIDIA.

If pie is not on your PATH, add the install directory:

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc

Linux CUDA prerequisites​

The CUDA flavors require a matching NVIDIA driver, the CUDA toolkit runtime libraries, and NCCL on the host. On Debian or Ubuntu with NVIDIA's package repositories configured:

sudo apt install libnccl2 libnccl-dev

If pie doctor reports missing libraries on a CUDA build, install the runtime that matches the flavor (CUDA 12.x for cuda12.8, CUDA 13.x for cuda13.0).

Verify​

pie --version
pie doctor

pie doctor checks platform readiness, lists compiled-in drivers, probes GPU visibility, and confirms that ~/.pie/config.toml and any subprocess driver venvs resolve.

Optional: inferlet authoring​

The release installer provides the pie engine CLI. To scaffold, build, and publish inferlets, install Bakery from PyPI:

pip install pie-bakery

This provides the standalone bakery command and the Python package used by pie new / pie build.

Optional: Python-backed drivers​

The pie binary runs the embedded cuda_native and portable drivers without Python. Python is required only for the subprocess drivers: vllm, sglang, and tensorrt_llm. Install one with:

Note: The vllm and sglang drivers are currently not supported on Windows. Please use the portable driver on Windows, or run these drivers in WSL/Linux.

pie driver vllm install ~/.pie/venvs/vllm --run
pie driver vllm set venv ~/.pie/venvs/vllm

Replace vllm with sglang or tensorrt-llm for the other drivers. Each driver gets a Python 3.12 virtual environment with its pinned wheels. See the driver reference for what each driver does and when to pick it.

Advanced installation​

Most users do not need anything in this section.

Choose a build manually​

Override the auto-picked flavor with PIE_FLAVOR, and (for CUDA) the compute capability with PIE_CC:

curl -fsSL https://pie-project.org/install.sh | PIE_FLAVOR=portable bash
curl -fsSL https://pie-project.org/install.sh | PIE_FLAVOR=cuda12.8 bash
# Force a specific compute capability (e.g. on a build host without the GPU):
curl -fsSL https://pie-project.org/install.sh | PIE_FLAVOR=cuda13.0 PIE_CC=90 bash

On Windows:

$env:PIE_FLAVOR = "cuda"; irm https://pie-project.org/install.ps1 | iex
PlatformAvailable flavors
Linux x86_64portable, cuda12.8, cuda13.0 (per-CC: sm80/86/89/90/100/120)
Linux aarch64portable, cuda12.8, cuda13.0 (per-CC: sm90/100)
macOS Apple Siliconportable
Windows x86_64vulkan, cuda

Intel Macs are not supported.

The Linux cuda* flavors bundle the embedded cuda_native driver (the recommended CUDA path on Linux). Windows cuda uses the portable driver with a CUDA backend, as there is no native Windows driver.

Override the install directory​

curl -fsSL https://pie-project.org/install.sh | PIE_INSTALL_DIR=/usr/local/bin bash

Install a specific version​

curl -fsSL https://pie-project.org/install.sh | PIE_VERSION=0.4.0 bash

Install from source​

For development:

git clone https://github.com/pie-project/pie.git
cd pie

# Default features: embedded portable driver plus the always-linked dummy.
cargo install --path server

# Include the embedded CUDA driver too.
CUDACXX=/usr/local/cuda/bin/nvcc \
  cargo install --path server --features driver-portable,driver-cuda

# Build the release binary for Linux/macOS.
cargo install --path server --features driver-portable

Re-run pie doctor after building to confirm which drivers are compiled in.

Windows CMD​

curl -fsSL https://pie-project.org/install.cmd -o install.cmd && install.cmd && del install.cmd

Next steps​