Update.

diff --git a/mygpt.py b/mygpt.py
index 3bb3519..7117e76 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -304,7 +304,7 @@ class MyGPT(nn.Module):
                 t_next = dist.sample()
                 sum_logits += logits.log_softmax(dim=-1)[
                     torch.arange(t_next.size(0)), t_next
-                ]
+                ].sum()
             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 
         return sum_logits