Update convert_chatglm.py (#185)

neural_speed/convert/convert_chatglm.py

@@ -973,7 +973,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
     # ChatGLM3 shares the same architecture and model config with ChatGLM2
     # but its tokenizer further supports system prompts,
     # so we can check system token to discriminate ChatGLM3 from ChatGLM2.
-    if "<|system|>" in tokenizer.tokenizer.special_tokens:
+    if hasattr(tokenizer, "tokenizer") and "<|system|>" in tokenizer.tokenizer.special_tokens:
         if args.format == "GGUF":
             chatglm3_convert_gguf(model, tokenizer, dir_model, fname_out, ftype, hparams)
         else: