fix

algorithmic_efficiency/workloads/librispeech_conformer/librispeech_pytorch/models.py

@@ -297,7 +297,7 @@ def forward(self, inputs, key_padding_mask=None):
         attn_mask=~key_padding_mask[:, None, None],
         dropout_p=self.dropout,
     ).transpose(1, 2).reshape(batch_size, seq_len, embed_dim)
-    out = out * attention_temperature
+    out = out * self.attention_temperature
     out = self.out_proj(out)
     return out