Parcourir la source

Revert "saving the qk matrix in the attention module for convenience"

This reverts commit 68e44bd83ce6c3e352f74b266aa39d8b649af9e3.
Jong Wook Kim il y a 2 ans
Parent
commit
9323b2526c
1 fichiers modifiés avec 0 ajouts et 3 suppressions
  1. 0 3
      whisper/model.py

+ 0 - 3
whisper/model.py

@@ -62,7 +62,6 @@ class MultiHeadAttention(nn.Module):
         self.key = Linear(n_state, n_state, bias=False)
         self.value = Linear(n_state, n_state)
         self.out = Linear(n_state, n_state)
-        self.last_qk = None
 
     def forward(
         self,
@@ -97,8 +96,6 @@ class MultiHeadAttention(nn.Module):
         if mask is not None:
             qk = qk + mask[:n_ctx, :n_ctx]
 
-        self.last_qk = qk.detach()
-
         w = F.softmax(qk.float(), dim=-1).to(q.dtype)
         return (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2)