From conda to uv

Contributing to a library

• Fork and clone, with the original repo as upstream: gh repo fork org/lib --clone
• If already cloned, add fork as origin: gh repo fork --remote
• Environment, if the project has a uv.lock: uv sync
• Otherwise: uv venv && uv pip install -e ".[dev]"
• Push to your fork and open a draft PR: git push -u origin my-branch && gh pr create --draft
• Sync with upstream: git fetch upstream && git rebase upstream/main

Using dev versions in an analysis

• New project: uv init my-analysis
• Local clone, editable (changes visible immediately): uv add --editable ~/src/lib
• Branch of a fork (commit pinned in uv.lock, could be shared with others): uv add "lib @ git+https://github.com/me/lib" --branch my-branch
• Get new commits from the branch: uv lock --upgrade-package lib
• Test against released versions, ignoring sources: uv sync --no-sources
• Back to a release: remove the line in [tool.uv.sources], then uv lock

Quick experiments

• Self-contained script (PEP 723): uv init --script exp.py
• Add deps, git branches included: uv add --script exp.py numpy "lib @ git+https://github.com/me/lib" --branch my-branch
• Run it, env created on the fly: uv run exp.py
• Named envs like conda activate: uve

CUDA

PyTorch wheels bundle CUDA. Declare the PyTorch index, Linux only so the project still installs on a Mac (see the uv guide):

[[tool.uv.index]]
name = "pytorch-cu126"
url = "https://download.pytorch.org/whl/cu126"
explicit = true

[tool.uv.sources]
torch = [{ index = "pytorch-cu126", marker = "sys_platform == 'linux'" }]

conda → uv