q yao

Thanks for reporting and this is a potential bug. In the models that adopts MaskedConv2d, input and output sizes are the same, and current implementation does not support a smaller...

> In addition, the existence of data coverage will lead to uncertainty. For example, height = 2, width =2, when c_im=0, h_im=2, w_im=0, data_im[4] will be writed, but c_im=1, h_im=0,...

Thanks for the notification! You are right, we should check the value of `x_low ...` before atomicAdd like [this](https://github.com/open-mmlab/mmcv/blob/73066430bed3dfe229ed0de8265d785da37f343e/mmcv/ops/csrc/common/cuda/roi_align_cuda_kernel.cuh#L196). We will fix it ASAP. Or would you like to send...

Try replace all `AT_DISPATCH_FLOATING_TYPES` in [ms_deform_attn_cuda.cu](https://github.com/open-mmlab/mmcv/blob/master/mmcv/ops/csrc/pytorch/cuda/ms_deform_attn_cuda.cu) with `AT_DISPATCH_FLOATING_TYPES_AND_HALF` and recompile MMCV

Can you pass all the unit test? ```bash pytest tests/test_ops ``` `tests/test_ops/test_convex_iou.py`, `tests/test_ops/test_roiaware_pool3d.py`, `tests/test_ops/test_voxelization.py` would give `Fatal Python error: Aborted` in the test. I am not sure if this has...

Sure, I see. Maybe there are something wrong with my env. I will let someone else have a try.

Sorry, my old ROCm environment is not available anymore. I will find a new one and continue this review ASAP.

Please update the unit test.

Yes, TensorRT plugins in MMCV do not support TensorRT8+, please try the plugin in [MMDeploy](https://github.com/open-mmlab/mmdeploy).

Try edit `mmcv/ops/csrc/pytorch/cuda/ms_deform_attn_cuda.cu`, replace `AT_DISPATCH_FLOATING_TYPES` with `AT_DISPATCH_FLOATING_TYPES_AND_HALF`.