[OPT] Fix attention scaling (#38290)

* fix opt attention scaling

* add comment to why we do this

Author: vasqu

src/transformers/models/opt/modeling_opt.py

@@ -154,7 +154,11 @@ def forward(
         """Input shape: Batch x Time x Channel"""
         bsz, tgt_len, _ = hidden_states.size()
 
-        # get query proj
+        # Scaling is susceptible to floating point arithmetics' inprecisions
+        # which can lead to different results (this is dependent from model
+        # to model, e.g. whisper is one such case). We therefore keep the
+        # original order of scaling to follow the original implementation
+        # and enforce no scaling (1.0) in the attention call below.
         query_states = self.q_proj(hidden_states) * self.scaling
         query_states = query_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
 
@@ -187,7 +191,7 @@ def forward(
             value_states,
             attention_mask,
             dropout=0.0 if not self.training else self.dropout,
-            scaling=self.scaling,
+            scaling=1.0,
             **kwargs,
         )