bmm doesn't seem to work with pytorch autograd at the moment. Here is a simple code

x = Variable(torch.FloatTensor(1, 2, 3).fill_(0.1))
y = Variable(torch.FloatTensor(1, 3, 4).fill_(0.3))

z = x.bmm(y)

and the error

RuntimeError: output tensor of incorrect size at /Users/soumith/anaconda/conda-bld/pytorch-0.1.6_1480182266313/work/torch/lib/TH/generic/THTensorMath.c:993