Update handler.py

handler.py

@@ -21,7 +21,7 @@ class EndpointHandler():
 
         # tokenize the input
         encoded = tokenizer(text, return_tensors="pt")
-        encoded = encoded.to('cuda')
+        encoded = encoded.to(model.device)
         # run the model
         generated_tokens = model.generate(**encoded, forced_bos_token_id=tokenizer.get_lang_id(langId))
         result = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]