Skip to main content

Installation

Pie supports Linux (CUDA) and macOS (Metal). Choose the installation method that works best for you.

Quick Install (PyPI)​

The fastest way to get started:

# Linux/Windows (CUDA)
pip install "pie-server[cuda]"

# macOS (Apple Silicon)
pip install "pie-server[metal]"

Install from Source​

For development or to get the latest features:

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

# Use uv to sync dependencies (recommended)
# Options: cu126, cu128, cu130, metal
uv sync --extra cu128
tip

When building from source, prefix all commands with uv run (e.g., uv run pie serve).

Initial Setup​

After installation, initialize your configuration:

pie config init

This creates ~/.pie/config.toml with default settings.

Download a Model​

Download a model from HuggingFace:

pie model download meta-llama/Llama-3.2-1B-Instruct

Or use any compatible model:

pie model download Qwen/Qwen2.5-0.5B-Instruct

Verify Installation​

Run a test prompt:

pie run text-completion -- --prompt "Hello, world!"
note

The first run may take longer due to JIT compilation.

Configuration​

Edit ~/.pie/config.toml to customize:

[engine]
host = "127.0.0.1"
port = 8080
enable_auth = true

[[model]]
hf_repo = "meta-llama/Llama-3.2-1B-Instruct"
device = ["cuda:0"]

View your current configuration:

pie config show

Update specific values:

pie config update --hf-repo "Qwen/Qwen2.5-7B-Instruct"
pie config update --device "cuda:0,cuda:1"

Next Steps​