don't need attn weight in decoder

2022-12-06 18:31:17 +08:00 · 2022-12-06 18:31:17 +08:00 · 2005ab1f26
commit 2005ab1f26
parent be167b3dda
1 changed files with 1 additions and 1 deletions
--- a/torchscale/architecture/decoder.py
+++ b/torchscale/architecture/decoder.py
@ -462,7 +462,7 @@ class Decoder(nn.Module):
        return x, {
            "inner_states": inner_states,
            "l_aux": l_aux,
-            "attn": [layer_attn.mean(dim=0)],
+            "attn": None,
        }

    def output_layer(self, features):