WARNING

[cooperdk]

Just a warning.

This tool destroys a modern working environment, downloading torch 2.6 etc. - do NOT use!

[PawelPeczek-Roboflow]

hi there - yes, we download specific version of torch. Could you elaborate a little bit more on the specific issue you experience - maybe we could help?

[PawelPeczek-Roboflow]

Made series of changes this week to improve dependency management, also clarified docs for inference-gpu - hope this helps - let us know. With more details, chances are we will make it work for you.

[PawelPeczek-Roboflow]

Changes will be released soon

[JiwaniZakir]

The issue is that installing this package pulls in large, potentially version-conflicting dependencies (like PyTorch 2.6) that can overwrite or break an existing environment. This typically stems from broad or unpinned dependency declarations in setup.py or pyproject.toml — if torch is listed as a top-level dependency without extras-based gating (e.g., extras_require), pip will eagerly resolve and install it regardless of whether the user needs GPU inference. The root cause is likely that heavy ML dependencies aren't made optional, so a simple pip install inference triggers a full torch installation rather than a lightweight core install.

[PawelPeczek-Roboflow]

Correct, we had versions handling torch deps quite specifically, but this has changed, yet now torch is made required dependency. It's declaration is
torch>=2.0.1,<3.0.0
so we changed install instructions to guide users towards selecting their own version.

That being said - what is the desired state of installation from your perspective?