Let's say I set m=8 and batch_size=32, and I have eight classes (A-H). That should mean that each batch is comprised of samples from only four classes, and each class is only represented in a single batch. That would then mean that samples from Class A could be in negative pairs with samples from Classes B-D, but never with samples from Classes E-H. Is that all correct?

How then does this i

It seems that you used different criteria during training and testing as the code below shows:
IN TEST:
scores = np.dot(vecs.T, qvecs)
IN TRAIN:
dif = x1 - x2
D = torch.pow(dif+eps, 2).sum(dim=0).sqrt()

   y = 0.5*lbl*torch.pow(D,2) + 0.5*(1-lbl)*torch.pow(torch.clamp(margin-D, min=0),2)
   y = torch.sum(y)

I did not get it why you do so?