What is the feature?
Currently, older versions of MMCV (e.g., mmcv-full 1.7.2) exhibit poor compatibility with PyTorch 2.x series. This creates significant barriers for users who wish to leverage the advanced features (e.g., TorchInductor compilation engine, FlashAttention) and performance improvements of PyTorch 2.x while using MMCV-based computer vision workflows. Users are forced to either downgrade PyTorch to outdated versions (e.g., 1.13) or abandon MMCV, which hinders productivity and technology adoption.
Any other context?
No response