diff --git a/codes/models/gpt_voice/lucidrains_gpt.py b/codes/models/gpt_voice/lucidrains_gpt.py
index 1b314af2..b218f7fc 100644
--- a/codes/models/gpt_voice/lucidrains_gpt.py
+++ b/codes/models/gpt_voice/lucidrains_gpt.py
@@ -228,9 +228,9 @@ class Transformer(nn.Module):
         assert(len(prev_intermediates) == self.depth)
         new_intermediates = []
         for (attn, ff), (int_ff, int_out) in zip(self.layers.layers, prev_intermediates):
-            x = x + attn(x, only_last_two_elements=True)
+            x_ff = attn(x, only_last_two_elements=True)
             # Note that (x) is now only the last two element in the set. Conjoin it with the int_ff latent to compute the norm.
-            x_ff = torch.cat([int_ff[:,:-1], x], dim=1)
-            x = x + ff(x_ff, only_last_two_elements=True)
+            x_ff = x + torch.cat([int_ff[:,:-1], x_ff], dim=1)
+            x = x_ff + ff(x_ff, only_last_two_elements=True)
             new_intermediates.append((x_ff, x))
         return x, new_intermediates