convert data type

cosyvoice_rodis/llm/llm.py

@@ -162,11 +162,15 @@ class TransformerLM(torch.nn.Module):
         text_token, text_token_len = self.encode(text_token, text_token_len)
 
         embedding = F.normalize(embedding, dim=1)
+        if embedding.dtype != self.spk_embed_affine_layer.weight.dtype:
+            embedding = embedding.to(self.spk_embed_affine_layer.weight.dtype)
         embedding = self.spk_embed_affine_layer(embedding)
         embedding = embedding.unsqueeze(1) 
 
         if self.emotion_embedding and emotion_embedding is not None:
             emotion_embedding = F.normalize(emotion_embedding, dim=1)
+            if emotion_embedding.dtype != self.spk_embed_affine_layer.weight.dtype:
+                emotion_embedding = emotion_embedding.to(self.spk_embed_affine_layer.weight.dtype)
             emotion_embedding = self.spk_embed_affine_layer(emotion_embedding) 
             emotion_embedding = emotion_embedding.unsqueeze(1) 
             embedding += emotion_embedding