X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=0cf70e0f674317b0c5c4884d248eb55a18ef6232;hb=545b06257dd69285c6321d7bd713d819da74953b;hp=ac1c55e84d91fecb06b453533f1800aead640ed7;hpb=00b2d5ed01fb523fbc4e699f0419329efbee0ea8;p=picoclvr.git

diff --git a/mygpt.py b/mygpt.py
index ac1c55e..0cf70e0 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -46,7 +46,7 @@ class BracketedSequence:
         return self.x[:, self.first : self.first + self.nb]
 
     def complete(self):
-        return self.first == 0 and self.nb == x.size(1)
+        return self.first == 0 and self.nb == self.x.size(1)
 
 
 ######################################################################
@@ -169,9 +169,6 @@ class QKVAttention(nn.Module):
             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs_q.first + bs_q.nb]
         ) / math.sqrt(self.w_q.size(1))
 
-        if self.record_attention:
-            self.a = a
-
         if self.causal:
             if bs_q.first == 0:
                 self.cache_attzero = (
@@ -186,6 +183,10 @@ class QKVAttention(nn.Module):
             )
 
         a = a.softmax(dim=3)
+
+        if self.record_attention:
+            self.a = a
+
         a = F.dropout(a, self.attention_dropout, self.training)
 
         y = torch.einsum(