This course uses uv for all Python dependency management. Written in Rust, uv is 10-100x faster than pip and replaces multiple tools — pip, pip-tools, pipx, poetry, pyenv, and virtualenv — with a single binary.
Installing uv
macOS / Linux
Windows
Homebrew
curl -LsSf https://astral.sh/uv/install.sh | sh
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
Inside the course Docker containers, uv is already installed. You only need to install it locally if you are working outside Docker.
Managing Python versions
uv can install and manage Python versions directly — no need for pyenv or system package managers:
# Install a specific Python version
uv python install 3.12
# Pin the Python version for your project
uv python pin 3.12
# List installed versions
uv python list
Virtual environments
uv creates and manages virtual environments:
# Create a virtual environment
uv venv
# Create with a specific Python version
uv venv --python 3.12
# Create with access to system packages (used in course containers)
uv venv --system-site-packages
.venv/ by default. Activate it with:
source .venv/bin/activate
Do not commit the .venv folder to your GitHub repo. The repository’s .gitignore already excludes it.
Installing packages
uv provides a pip-compatible interface for installing packages:
# Install a package
uv pip install numpy
# Install from pyproject.toml
uv pip install -e .
# Install with extras (e.g., notebooks, dev)
uv pip install -e ".[notebooks]"
# Install with a constraints file
uv pip install -e . --constraint constraints.txt
Project management
For projects with a pyproject.toml (like the course repository), uv provides higher-level commands:
# Add a dependency to the project
uv add pandas
# Add a dev dependency
uv add --dev pytest
# Lock dependencies (creates uv.lock)
uv lock
# Sync environment to match the lockfile
uv sync
# Run a script in the project environment
uv run python train.py
uv.lock file is a cross-platform lockfile that ensures reproducible installs across machines and operating systems.
Course workflow
Inside the course Docker container, dependencies are pre-installed. The typical workflow is:
# Set up the environment (first time)
make start
# Install the package with notebook extras
make install-notebooks
# Or manually:
uv pip install -e ".[notebooks]"
# Inside the container
uv pip install <package>
Why uv over pip or conda
| uv | pip | conda |
|---|
| Speed | 10-100x faster | Baseline | Slower |
| Lockfiles | Cross-platform uv.lock | No built-in lockfile | environment.yml (not locked) |
| Python management | Built-in (uv python install) | Requires pyenv or system packages | Built-in |
| Virtual envs | Built-in (uv venv) | Requires virtualenv | Built-in |
| Disk usage | Global cache, deduplication | Per-environment copies | Large per-environment copies |
| NVIDIA containers | Compatible | Compatible | Conflicts with pre-installed packages |
Avoid conda environments in this course. They conflict with the optimized NVIDIA containers and create dependency resolution issues with pre-installed PyTorch packages.
Useful commands reference
| Command | Description |
|---|
uv pip install <pkg> | Install a package |
uv pip install -e ".[extras]" | Install project with extras |
uv pip compile pyproject.toml | Generate locked requirements |
uv pip sync requirements.txt | Sync environment to requirements |
uv add <pkg> | Add a dependency to pyproject.toml |
uv lock | Create/update the lockfile |
uv sync | Sync environment to lockfile |
uv run <command> | Run a command in the project environment |
uv venv | Create a virtual environment |
uv python install <version> | Install a Python version |
uv python pin <version> | Pin Python version for the project |
uv tool run <tool> | Run a tool in an ephemeral environment |
