When kernel_size = (1,1,1) for SparseConv3d, the onnx inference result is different from the torch

[Icecream-blue-sky]

I change the last sparse conv of bevfusion from the first picture to the second (I don't want to downsample):

It will cause the final lidar feat to be incorrect. The onnx inference result is inaccurate while the torch inference is normal. I think this is probably a bug...

[hopef]

Please clean the argument "indice_key" because the rulebook needs to be recomputed.

[Icecream-blue-sky]

Please clean the argument "indice_key" because the rulebook needs to be recomputed.

This has nothing to do with indice_key. I tried or set indice_key="", indice_key="spconv_4", indice_key="spconv_down2", all do not work. However, after using 113 conv, I succeeded.

[hopef]

Could you share a minimal reproduced code with me?