Unable to build duckdb-python

[jakkes]

HI,

Am trying to compile from source, but seem to be running into some issues when resolving the torch dependency. I ran

git clone https://github.com/duckdb/duckdb-python.git
cd duckdb-python/
git submodule update --init
uv sync -p 3.11 --no-install-project

and it results in

Using CPython 3.12.11
Removed virtual environment at: .venv
Creating virtual environment at: .venv
  × No solution found when resolving dependencies for split (python_full_version >= '3.12' and platform_machine == 'AMD64' and sys_platform == 'win32'):
  ╰─▶ Because only the following versions of torch{(python_full_version < '3.14' and platform_machine != 'x86_64') or (python_full_version < '3.13' and platform_machine == 'x86_64' and sys_platform
      == 'darwin') or (python_full_version < '3.14' and sys_platform != 'darwin')} are available:
          torch{(python_full_version < '3.14' and platform_machine != 'x86_64') or (python_full_version < '3.13' and platform_machine == 'x86_64' and sys_platform == 'darwin') or
      (python_full_version < '3.14' and sys_platform != 'darwin')}<=2.2.2
...

I can install torch just fine on its own.

I'm on linux x86_64. Have tried various versions of Python, none work. :/

[paultiq]

Here's what I do . Let me know if this works / doesn't work, I've been meaning to throw a blog up for it. This creates a separate .venv for each version of PYTHON you're using. Change UV_PYTHON to whatever you want.

* cccache is very helpful for reducing rebuilds, especially of the external duckdb submodule, I'd suggest uncommenting it and setting the BASEIDR.

** Also, run "uv self update" first... you might be running an older version of UV.

** I also re-enable UNITY (DISABLE_UNITY=0) in the pyproject.toml, this reduces build times for me.

export UV_PYTHON=3.13
export UV_PROJECT_ENVIRONMENT=.venv_$UV_PYTHON

if [ ! -d $UV_PROJECT_ENVIRONMENT ]; then
    uv venv
else
    echo $UV_PROJECT_ENVIRONMENT exists
fi 

source $UV_PROJECT_ENVIRONMENT/bin/activate
uv sync --no-install-project

# export CCACHE_PROGRAM=ccache 
# export CCACHE_BASEDIR=...

uv sync --no-build-isolation -vv  --reinstall

[paultiq]

I just build this on an Ubuntu server, using 3.12, didn't see the dependency issue.

[evertlammerts]

× No solution found when resolving dependencies for split (python_full_version >= '3.12' and platform_machine == 'AMD64' and sys_platform == 'win32'):

Looking at a recent CI build with py312 + win32 / amd64 I see that torch=2.8.0+cpu is resolved and installed. This should work for you as well.

Using CPython 3.12.11
Removed virtual environment at: .venv
Creating virtual environment at: .venv

This is also not supposed to happen when you run uv sync -p 3.11 --no-install-project. You're asking for Python 3.11 and uv removes your venv.

I suggest you:

• remove your venv
• clean up the uv cache with uv cache clean
• re-create a venv with uv venv -p 3.11 (or whatever python version you want)
• run uv sync --only-group dev --no-install-project --verbose
  ... and see what happens. If it's still not working, can you put the full output of all your commands into a gist and link it here?

[jakkes]

Same error. I did manage to build it if I commented out the torch and tensorflow dependencies.

Below are logs.

https://gist.github.com/jakkes/fd51951b64fcad4e13406ebe673d9288

[paultiq]

Did not look closely at the log, but - you're running an old version of uv. Run "uv self update" first.

[evertlammerts]

Yes, do update your uv please. As you can see from here I've had my share of fun with env markers and uv. We're using specific resolution environments for torch because it is such a pain to get it right, and not all features we use for that might be in your uv version.

If that's not it then I have no idea why this isn't working for you. The requirement specifier that was generated is python_full_version >= '3.9' and platform_machine == 'AMD64' and sys_platform == 'win32' should be fine... something else is off, and I can't reproduce this on the few windows machines I have access to, nor on CI.

If you're just looking to get things working, just remove torch as a dependency from pyproject.toml. It's used for a few tests, so unless you're specifically looking to do something with torch just ignore this.

[jakkes]

Yup, updating uv did it. Thanks all!