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

This reverts commit 68e44bd.
w7374520 · Dec 30, 2022 · 9323b25 · 9323b25
1 parent 68e44bd
commit 9323b25
Showing 1 changed file with 0 additions and 3 deletions.
diff --git a/whisper/model.py b/whisper/model.py
@@ -62,7 +62,6 @@ def __init__(self, n_state: int, n_head: int):
         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 @@ def qkv_attention(self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor]
         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)