My pixi success story

[Bernd-Vogelgesang]

First, thanks a lot for this amazing project and thanks for using pixi!
I just recently stumbled upon pixi and as a complete noob (my friends would say otherwise, but what do they know) I am so happy to see the light at the end of python dependency hell.

Unfortunately the provided recipe for the toml did not work for me, so I had to fiddle around a lot with the configuration.
So, for others who may run as well into trouble as for myself as a reference after I forgot everything I did, I will post my steps:

create a folder geoai, open terminal in the folder
pixi init
open the newly created pixi.toml in an editor

Only insert the section for CUDA (in my case 13 and on Linux) and the python version:

[workspace]
channels = ["conda-forge"]
name = "geoai"
platforms = ["linux-64"]
version = "0.1.0"

[system-requirements]
cuda = "13.0"

[dependencies]
python = "3.12.*"

Save the toml

In the command line , I started with adding deepforest, because this refused to install beforhand under any circumstances:

pixi add "deepforest>=2.0.0"
then
pixi add "pytorch-gpu>=2.7.1,<3"

Now, that these steps went through without any hickups, I just added the other components without any version numbers:

pixi add qgis
pixi add geoai
pixi add segment-geospatial
pixi add sam3

Finally
pixi install

So, my toml now looks like this:

[workspace]
channels = ["conda-forge"]
name = "geoai"
platforms = ["linux-64"]
version = "0.1.0"

[system-requirements]
cuda = "13.0"

[dependencies]
python = "3.12.*"
deepforest = ">=2.0.0"
pytorch-gpu = ">=2.7.1,<3"
qgis = ">=3.44.7,<4"
geoai = ">=0.5.3,<0.6"
segment-geospatial = ">=0.15.2,<0.16"
sam3 = ">=0.1.0.20251211,<0.2"

In a nutshell: Only add really needed version requirements to the toml like CUDA and and python version, and then let pixi find the fitting packages by itself.

Maybe this is not how you should work with pixi, and I get the concept wrong, but it worked!

[giswqs]

@Bernd-Vogelgesang Thanks for sharing your success story. Installing QGIS, GeoAI, and Pytorch GPU is always a tricky issue due to the hundreds of dependencies involved. It gets even more tricky after adding deepforest due to its transformers version contraint weecology/DeepForest#1291. Hopefully, after the transformer constraint is lifted, it will be a bit easier to install.

[chameroy]

@Bernd-Vogelgesang I was also having issues with package conflicts when trying to enable GPU processing. Thank you for sharing! This was very helpful!

[giswqs]

I just released a new video tutorial on installing deepforest and the GeoAI QGIS Plugin: https://youtu.be/23Su8J6R4Cw