cleaned the scaled-dot attention path

OpenNMT · May 28, 2024 · 506b355 · 506b355
1 parent e9edb12
commit 506b355
Showing 1 changed file with 0 additions and 4 deletions.
diff --git a/onmt/modules/multi_headed_attn.py b/onmt/modules/multi_headed_attn.py
@@ -733,10 +733,6 @@ def forward(
                 attn_output.add_(relative_matmul(drop_attn, relations_values, False))
 
         context = unshape(attn_output)
-        if key_pad_mask is not None:
-            if key_pad_mask.size(0) > 1 and context.size(1) > 1:
-                x = key_pad_mask.squeeze(1).unsqueeze(2).expand(-1, -1, context.size(2))
-                context = context.masked_fill(x, 0)
 
         if self.layer_cache[0]:
             attn_output = self.final_linear(context)