How to define the pytorch-cuda version in the toml file for the same platform with different cuda versions

[datalw]

Hi,

I run a python project on two linux-64 machines, one with cuda 11.4 and the other one with cuda 12.4. Pytorch does not seem to support 11.4 anymore and my other dependencies need newer versions of pytorch anyway. It has been suggested on the pytorch forum that pytorch-cuda 11.8 will work with cuda 11.4.
So I would like to define the dependencies as following:
(1) for linux-64 with cuda 11.4: use pytorch-cuda 11.8.*
(2) for linux-64 with cuda 12.4: use pytorch-cuda 12.4.*
And of course, they should have the same version of torch, torchvision and torchaudio.

As far as I understand, system requirements should be the minimal requirements. But for my case, setting any of the three CUDA versions would be wrong. Is it possible to define these in my situation?
Thanks for helping!

Btw, I am new to pixi and I think I saw a similar thread but for different os versions here.

[kusutori]

I'm not sure if I understood you correctly, but I think just using the pixi's feature can create multiple cuda environments, like this

[environments]
cu124 = ["cu124"]
cu118 = ["cu118"]

[feature.cu124.system-requirements]
cuda = "12.4"

[feature.cu124.dependencies]
cuda-version = "12.4.*"
pytorch-gpu = "*"
torchaudio = "*"
torchvision = "*"

[feature.cu118.system-requirements]
cuda = "11.8"

[feature.cu118.dependencies]
cuda-version = "11.8.*"
pytorch-gpu = "*"
torchaudio = "*"
torchvision = "*"