Attempt to speed up

handler.py

@@ -18,12 +18,14 @@ class EndpointHandler:
         checkpoint = "Dupaja/speecht5_tts"
         vocoder_id = "Dupaja/speecht5_hifigan"
         dataset_id = "Dupaja/cmu-arctic-xvectors"
-        
-        self.model= SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
-        self.processor = SpeechT5Processor.from_pretrained(checkpoint)
-        self.vocoder = SpeechT5HifiGan.from_pretrained(vocoder_id)
-        embeddings_dataset = load_dataset(dataset_id, split="validation", trust_remote_code=True)
-        self.embeddings_dataset = embeddings_dataset
+
+        with torch.device("cuda"):
+            self.model= SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
+            self.processor = SpeechT5Processor.from_pretrained(checkpoint)
+            self.vocoder = SpeechT5HifiGan.from_pretrained(vocoder_id)
+            embeddings_dataset = load_dataset(dataset_id, split="validation", trust_remote_code=True)
+            self.embeddings_dataset = embeddings_dataset
+            
         self.speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)