The "mask" array actually specifies what attention to discard.

diff --git a/mygpt.py b/mygpt.py
index ebc9a83..f954797 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -66,9 +66,9 @@ class QKVAttention(nn.Module):
         a = torch.einsum('nhtd,nhsd->nhts', q, k) / math.sqrt(q.size(3))
 
         if self.causal:
-            mask = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
-                   < torch.arange(a.size(3), device = q.device)[None, None, None, :]
-            a = a.masked_fill(mask, float('-inf'))
+            forbidden_attention = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
+                                < torch.arange(a.size(3), device = q.device)[None, None, None, :]
+            a = a.masked_fill(forbidden_attention, float('-inf'))
 
         a = a.softmax(dim = 3)
         a = F.dropout(a, self.attention_dropout, self.training)