Hey Ryan, this is very interesting, if I’m understanding the shapes don’t even match for node 370 between python and c++?
So I’m guessing the input to node 370 is a tensor of shape [1, 2, 15, 4096] from the split operator. Very interestingly, the permutation of the transpose appears to be different between the two and the shape also. For each node do you also have the input nodes?
Alternatively you can just share the code and model you are running if possible and I can take a closer look.