I have tested with multiple GPUs with variable environments:
GPUs: A6000, A5000, RTX 2080Ti, RTX 3090
PyTorch: 2.1.2, 2.3.0, 2.3.1
CUDA version: 11.8 and 12.1.
In addition, I have compiled both cumm and spconv manually following the README, but it still shows same error.
import torch as th
from spconv.pytorch import SubMConv3d, SparseConvTensor
xyz = th.randint(0, 100, (100, 4), dtype=th.int32, device='cuda')
xyz[:, 0] = 0
feat = th.rand(100, 32, device='cuda')
sp = SparseConvTensor(feat, xyz, (200, 200, 200), 1)
SubMConv3d(32, 64, 3)(sp)
>>> cuda failed with error 700 an illegal memory access was encountered. use CUDA_LAUNCH_BLOCKING=1 to get correct traceback.
I have tested with multiple GPUs with variable environments: GPUs: A6000, A5000, RTX 2080Ti, RTX 3090 PyTorch: 2.1.2, 2.3.0, 2.3.1 CUDA version: 11.8 and 12.1.
In addition, I have compiled both cumm and spconv manually following the README, but it still shows same error.