duplicate cache: uv add cannot use cache from uv pip install

[xiejx5]

Summary

Duplicate packages from uv add and uv pip commands. Minimum steps to reproduce:

uv cache clean

# first download
uv venv test
uv pip install numpy --directory test

# second download, did not use cache
uv init --bare
uv add numpy

Platform

Linux 4.18.0

Version

uv 0.6.10

Python version

Python 3.13.2

[charliermarsh]

I can take a look.

[Aditya-PS-05]

@charliermarsh , can I take a look at this issue and solve it?

[charliermarsh]

I have an idea on what the problem is but I think it will be challenging to solve correctly. (There’s an “easy” fix that will regress a different, small bug.)

[charliermarsh]

Oh sorry. This is because in uv sync (or uv add), we require hashes to be present. But in uv pip install, we don't. So we have to re-download the wheel to hash it.

[xiejx5]

Thank you for the answer. Is there any setting to avoid these duplicates? As uv pip can use the cache from uv add.

[charliermarsh]

I think we could plausibly add --generate-hashes to uv pip install which would avoid this problem.

[xiejx5]

I think we could plausibly add --generate-hashes to uv pip install which would avoid this problem.

Thank you. Unfortunately, --generate-hashes only exists in uv pip compile while not uv pip install.

Anyway, thanks a lot for your development and reply.

[charliermarsh]

Ah yeah, I meant, we could add it to uv pip install.

[LoSealL]

What about add from an external index? Such as torch. For instance there are two projects both requires torch==2.7.1:

# pyproject.toml
dependencies = ["torch>=2.7.1"]

[tool.uv.sources]
torch = { index = "pytorch" }
torchaudio = { index = "pytorch" }
torchvision = { index = "pytorch" }

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

# project 1
uv add torch  # download and make a cache in `archive-v0`

# project 2
uv add torch  # download and make another cache in `archive-v0`

[LoSealL]

Can I expect uv to use identical cache for same version? Because duplicated torch takes too much disk spaces