Group convolution error when group=2 in pytorch 0.4.0

i use torch.nn.DataParallel for 4 gpu
when featuremap size is [128, 106, 32, 32], kernel_size=3, group=2, padding=1 in_channels=106, out_channels= 106
check for stride=2 and 1

splited featuremap size on each gpu is
torch.Size([32, 106, 32, 32])

misaligned address error is printed

i don’t know why this happens
other size with group=2 works well

Could you post an executable code snippet reproducing the error?
What kind of error message do you get?

misaligned address error