fix for bsz>1 because I forgot the old implementation implicitly handles this

vall_e/models/base_v2.py

@@ -1011,6 +1011,20 @@ class Base_V2(nn.Module):
 
 				break
 
+		# fill in gap to make dim=-1 equal by filling in with -infs
+		# the old implementation inherently does this through the master Classifiers class
+		# but it needs to explicitly be done here
+		dim_neg_1 = 0
+		for batch_index, logit in enumerate( loss_logits ):
+			dim_neg_1 = max( dim_neg_1, logit.shape[-1] )
+		
+		for batch_index, logit in enumerate( loss_logits ):
+			if dim_neg_1 == logit.shape[-1]:
+				continue
+
+			loss_logits[batch_index] = torch.cat([logit, torch.full( (logit.shape[0], dim_neg_1 - logit.shape[-1]), -float("inf"), device=logit.device, dtype=logit.dtype) ], dim=-1 )
+
+
 		loss_target = torch.cat( loss_targets )
 		loss_logit = torch.cat( loss_logits )