[Feature] Please add spconv 2.x to mmcv #3285
Description
What is the feature?
Description:
Currently, the spconv operators used in MMCV are based on spconv 1.x, which was originally adopted from the mmdetection3d library. However, the spconv library has been updated to version 2.x, offering significant performance improvements, and mmdetection3d has already overwritten its implementation to support this new version.
Additionally, the current sparse_conv operator suffers from severe GPU memory consumption issues. Specifically, the allocated gridout matrix size depends on spatial_shape, leading to out-of-memory (OOM) errors in large-scale scenarios.
Request:
Integrate spconv 2.x into MMCV to leverage its enhanced performance and compatibility.
Add spconv 2.x as a dependency in requirements.txt.
This upgrade would resolve memory bottlenecks and align MMCV with the latest optimizations in mmdetection3d.
Any other context?
No response