Updated handler.py

handler.py

@@ -21,4 +21,5 @@ class EndpointHandler:
         parameters = data.pop("parameters", {})
         with torch.autocast(self.pipeline.device.type, dtype=torch.bfloat16):
             outputs = self.pipeline(inputs, **parameters, use_cache=True)
+            torch.cuda.empty_cache()
             return outputs