fix aoti model generation output types

Author: whoisj

qa/common/gen_qa_models.py

@@ -1389,6 +1389,9 @@ def create_torch_aoti_modelfile(
     ):
         return
 
+    torch_output0_dtype = np_to_torch_dtype(output0_dtype)
+    torch_output1_dtype = np_to_torch_dtype(output1_dtype)
+
     model_name = tu.get_model_name(
         "torch_aoti",
         input_dtype,
@@ -1415,7 +1418,7 @@ def __init__(self, swap):
         def forward(self, INPUT0, INPUT1):
             op0 = (INPUT0 - INPUT1) if self.swap else (INPUT0 + INPUT1)
             op1 = (INPUT0 + INPUT1) if self.swap else (INPUT0 - INPUT1)
-            return op0, op1
+            return op0.to(self.output0_dtype), op1.to(self.output1_dtype)
 
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     model = AddSubNet(swap)