mirror of
https://github.com/Comfy-Org/ComfyUI-Manager.git
synced 2026-11-04 23:37:31 +00:00
[GH-ISSUE #2240] comfyui-manager should never upgrade torch, torchvision or torchaudio; numpy and opencv require complex logic #905
Labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
Comfy-Org/ComfyUI-Manager#905
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @doctorpangloss on GitHub (Oct 23, 2025).
Original GitHub issue: https://github.com/Comfy-Org/ComfyUI-Manager/issues/2240
torchThe overrides when using
piporuvshould basically freeze the currently installed torch, torchvision, torchaudionumpyanything that requires jax will require numpy >= 2
the numpy 1.x issue is related to opencv
it is basically always safe to upgrade numpy except in containers like nvidia's, where the packages are built against numpy 1. this means you will want to freeze numpy depending on the environment, which is complex. you can enumerate all the packages in the venv and determine which one is required.
opencvThere are four opencv packages:
the controlnet aux package is radioactive in that it installs an opencv package that requires numpy 1.x even though it doesn't have metadata as such. So you must always use
4.12.0.88or higher. If there is a currently installed non-headless opencv-python or opencv-contrib-python, you must get the path to the package (sometimes in dist-packages, not always in site-packages, due to packaging in images like Runpod), uninstall it, rm -rf the path, and then install the corresponding headless package. It's a completely stylistic choice whether or not to use the contrib packages, the most notable thing in them is tracking features.all the opencv packages specified by nodes should be ignored, and the installation must always firs tuninstall all currently installed opencv, then install exactly opencv-contrib-python-headless or opencv-python-headless >= 4.12.0.88
transformersIt should be >=4.54, the whole 4.53 series should have been yanked
@doctorpangloss commented on GitHub (Oct 23, 2025):
aiohttp must also be pinned to >=3.13
@doctorpangloss commented on GitHub (Oct 23, 2025):
onnxruntime-gpu>=1.23.0
@doctorpangloss commented on GitHub (Oct 23, 2025):
ultralytics>=8.3.206
it looks like a bunch of packages have been copying and pasting some particularly radioactive settings