Where and how are the models being loaded?

[amannsonii3]

When I do pip install mlipx[<mlip model name>], is the model being loaded in my local machine (what I understand is that a new env is created within my mlipx env)?
Also, if I run into memory issues, can just doing pip uninstall mlipx[<mlip model name>] do the trick (remove all cache and everything of that model)?

[PythonFZ]

Hi Aman,

The model code is then installed alongside mlipx, e.g. if you do pip install mlipx[mace] it will also install mace-torch.
You can see this here

mlipx/pyproject.toml

Lines 98 to 100 in b284dd2

	mace = [
	"mace-torch>=0.3.12",
	]

How the actual model is handled, is different between the models. We do provide a download script here https://github.com/basf/mlipx-hub/blob/main/download.py but it would be difficult to provide all models and installations because of different dependencies and licenses.

For the pip uninstall I don't know exactly what pip will be doing. What I've been using is uv (https://docs.astral.sh/uv/getting-started/installation/) and there I know the install / uninstall works as you'd expect. A typical workflow for me looks like

mkdir ex01
cd ex01
uv init
uv add "mlipx[mace]"
# build the workflow for the mace model
uv add "mlipx[sevenn]"
# extend with the svenn model 
...

where uv add will replace the previous installation with the new one and thus there should be no complications with dependency interactions between the different models.

[amannsonii3]

Hello Fabian,

Thank you for replying so quickly!
To be specific, I wanted to know what would be the exact path where mlipx would store the models .pt files. It does download and cache it somewhere, right?

A curious follow-up (I'd be happy to start a new discussion for this):

• So when I uv add "mlipx[sevenn]" it would replace the previous mlipx[mace] installation with this new one. Is it the same as running pip install mace-torch>=0.3.12 and then pip install sevenn>=0.11.0?
• How is mlipx able to make sure that, despite mace-torch and sevenn requiring different dependencies are running within the same virtual env? (Unless it can only work with uMLIPs having similar dependencies)

Btw, exciting project! Really liked this work!

[PythonFZ]

Happy to hear, you like the project.

To be specific, I wanted to know what would be the exact path where mlipx would store the models .pt files

This is handled by the respective mlip package, not by mlipx itself. You can use the download script I mentioned above, which will download the models to a model directory that will be generated for you. MLIPX itself interfaces with the models through their ase interface which makes it very flexibel but leaves these parts of the model management to the mlip package developers.

So when I uv add "mlipx[sevenn]" it would replace the previous mlipx[mace] installation with this new one. Is it the same as running pip install mace-torch>=0.3.12 and then pip install sevenn>=0.11.0?

It is not the same, using uv add ... will replace the previous installation and uninstall all packages solely required by e.g. mace-torch. If you do pip uninstall mace-torch, only mace-torch will be uninstalled, but e.g. the torch package will remain.
But uv can basically replace pip install and there is also uv pip install. But of course, yo can use pip install, but it could lead in some rare cases to installation warnings / inconsistencies.

How is mlipx able to make sure that, despite mace-torch and sevenn requiring different dependencies are running within the same virtual env?

MLIPX can't do that. I change from one model to the other, running first all tests with uv add mlipx[mlip-1] and then change to uv add mlipx[mlip-2]