Compatibility issues with mmcv, newer pytorch versions

[picagrad]

There seems to be a complicated compatibility issues caused by mmyolo's reliance on older versions of mmcv (<2.1.0).
I have not been able to solve this myself, hoping you can tackle this through updating compatibility of mmyolo, or perhaps through some changes to the available wheels of mmcv (going to post this issue there as well).

For example the issue comes up with the following torch/python/cuda combination (but also many others)

• python 3.11.9
• torch 2.5.1 (with CUDA12.4)

Under these conditions there is no way to successfully build a version mmcv that supports mmyolo
One can revert to older versions of torch and cuda, but that comes with its own set of issues elsewhere.

Any chance this compatibility issue can be solved? Or any idea for a workaround? perhaps using mmdet?

[Kenneth-X]

just change the version check or delete it

mmyolo/mmyolo/__init__.py

Line 10 in 8c4d9dc

mmcv_maximum_version = '2.1.0'

PS：
This project doesn't seem to be updated anymore; the development team may have abandoned it.
T.T sad

[collinmccarthy]

FWIW I switched to simply using MMDetection instead and have been very happy. It still has a few YOLO versions I think, including RTM-Det with support for training instance segmentation models, and mmdet is very popular and actively maintained. The data augmentation / loading pipelines might be slower for RTM-Det in mmdet versus mmyolo, but mmdet is plenty fast for many people and chances are it'll be fine for you too. If you really want to dig into the differences between the two libraries its relatively clear looking at identical RTM-Det config files in both repos.

I wish someone would have told me that a year ago, it would have saved me quite a bit of time.