[VITS] Fix speaker_embed device mismatch (#26115)

* [VITS] Fix speaker_embed device mismatch

- pass device arg to speaker_id tensor

* [VITS] put speaker_embed on device when int

* [VITS] device=self.device
instead of self.embed_speaker.weight.device

* [VITS] make tensor directly on device
using torch.full()

src/transformers/models/vits/modeling_vits.py

@@ -1435,7 +1435,9 @@ class VitsModel(VitsPreTrainedModel):
         if self.config.num_speakers > 1 and speaker_id is not None:
             if not 0 <= speaker_id < self.config.num_speakers:
                 raise ValueError(f"Set `speaker_id` in the range 0-{self.config.num_speakers - 1}.")
-            speaker_embeddings = self.embed_speaker(torch.tensor([speaker_id])).unsqueeze(-1)
+            if isinstance(speaker_id, int):
+                speaker_id = torch.full(size=(1,), fill_value=speaker_id, device=self.device)
+            speaker_embeddings = self.embed_speaker(speaker_id).unsqueeze(-1)
         else:
             speaker_embeddings = None