avoiding conditional indexing in positionalencoding to avoid possibil… (#42090)

avoiding conditional indexing in positionalencoding to avoid possibility of empty tensors

Author: ppadjinTT

src/transformers/models/speecht5/modeling_speecht5.py

@@ -434,8 +434,8 @@ def forward(self, hidden_states):
         pos_seq = torch.arange(0, seq_len).to(device=hidden_states.device, dtype=torch.long)
         pos_seq = pos_seq[:, None] - pos_seq[None, :]
 
-        pos_seq[pos_seq < -self.max_length] = -self.max_length
-        pos_seq[pos_seq >= self.max_length] = self.max_length - 1
+        pos_seq = torch.where(pos_seq < -self.max_length, -self.max_length, pos_seq)
+        pos_seq = torch.where(pos_seq >= self.max_length, self.max_length - 1, pos_seq)
         pos_seq = pos_seq + self.max_length
 
         return self.pe_k(pos_seq)