From 9323b2526c2fe3fa7c139f4998c6f27bd71b50bb Mon Sep 17 00:00:00 2001
From: Jong Wook Kim <jongwook@nyu.edu>
Date: Thu, 29 Dec 2022 23:53:31 -0700
Subject: [PATCH] Revert "saving the qk matrix in the attention module for
 convenience"

This reverts commit 68e44bd83ce6c3e352f74b266aa39d8b649af9e3.
---
 whisper/model.py | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/whisper/model.py b/whisper/model.py
index 820d3c105..ca3928eca 100644
--- 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)