Update multipurpose_chatbot/engines/transformers_engine.py

multipurpose_chatbot/engines/transformers_engine.py

@@ -429,7 +429,8 @@ class TransformersEngine(BaseEngine):
         
         # ! MUST PUT INSIDE torch.no_grad() otherwise it will overflow OOM
         import sys
-        self._model._sample = types.MethodType(NewGenerationMixin.sample_stream, self._model)
+        # self._model._sample = types.MethodType(NewGenerationMixin.sample_stream, self._model)
+        self._model.sample = types.MethodType(NewGenerationMixin.sample_stream, self._model)
         with torch.no_grad():
             inputs = self.tokenizer(prompt, return_tensors='pt')
             num_tokens = inputs.input_ids.size(1)