Fix llamaindex AutoTokenizer bug (#10345)

* fix tokenizer

* fix AutoTokenizer bug

* modify code style

python/llm/example/CPU/LlamaIndex/README.md

@@ -64,6 +64,7 @@ python rag.py -m <path_to_model>
 - `-p PASSWORD`: password in the PostgreSQL database
 - `-q QUESTION`: question you want to ask
 - `-d DATA`: path to source data used for retrieval (in pdf format)
+- `-n N_PREDICT`: max predict tokens
 
 ### Example Output
 

python/llm/example/CPU/LlamaIndex/rag.py

@@ -14,6 +14,8 @@
 # limitations under the License.
 #
 
+
+import torch
 from llama_index.embeddings.huggingface import HuggingFaceEmbedding
 from sqlalchemy import make_url
 from llama_index.vector_stores.postgres import PGVectorStore
@@ -167,7 +169,7 @@ def main(args):
         model_name=args.model_path,
         tokenizer_name=args.model_path,
         context_window=512,
-        max_new_tokens=32,
+        max_new_tokens=args.n_predict,
         generate_kwargs={"temperature": 0.7, "do_sample": False},
         model_kwargs={},
         messages_to_prompt=messages_to_prompt,
@@ -242,6 +244,8 @@ if __name__ == "__main__":
                         help="the password of the user in the database")
     parser.add_argument('-e','--embedding-model-path',default="BAAI/bge-small-en",
                         help="the path to embedding model path")
+    parser.add_argument('-n','--n-predict', type=int, default=32,
+                        help='max number of predict tokens')
     args = parser.parse_args()
     
     main(args)

python/llm/example/GPU/LlamaIndex/README.md

@@ -154,6 +154,7 @@ python rag.py -m <path_to_model>
 - `-p PASSWORD`: password in the PostgreSQL database
 - `-q QUESTION`: question you want to ask
 - `-d DATA`: path to source data used for retrieval (in pdf format)
+- `-n N_PREDICT`: max predict tokens
 
 ### 5. Example Output
 

python/llm/example/GPU/LlamaIndex/rag.py

@@ -168,7 +168,7 @@ def main(args):
         model_name=args.model_path,
         tokenizer_name=args.model_path,
         context_window=512,
-        max_new_tokens=32,
+        max_new_tokens=args.n_predict,
         generate_kwargs={"temperature": 0.7, "do_sample": False},
         model_kwargs={},
         messages_to_prompt=messages_to_prompt,
@@ -243,6 +243,8 @@ if __name__ == "__main__":
                         help="the password of the user in the database")
     parser.add_argument('-e','--embedding-model-path',default="BAAI/bge-small-en",
                         help="the path to embedding model path")
+    parser.add_argument('-n','--n-predict', type=int, default=32,
+                        help='max number of predict tokens')
     args = parser.parse_args()
     
     main(args)

python/llm/src/bigdl/llm/llamaindex/llms/bigdlllm.py

@@ -235,9 +235,18 @@ class BigdlLLM(CustomLLM):
         """
         model_kwargs = model_kwargs or {}
         from bigdl.llm.transformers import AutoModelForCausalLM
-        self._model = model or AutoModelForCausalLM.from_pretrained(
+        if model:
-            model_name, load_in_4bit=True, **model_kwargs
+            self._model = model
+        else:
+            try:
+                self._model = AutoModelForCausalLM.from_pretrained(
+                    model_name, load_in_4bit=True, use_cache=True,
+                    trust_remote_code=True, **model_kwargs
                 )
+            except:
+                from bigdl.llm.transformers import AutoModel
+                self._model = AutoModel.from_pretrained(model_name,
+                                                        load_in_4bit=True, **model_kwargs)
 
         if 'xpu' in device_map:
             self._model = self._model.to(device_map)
@@ -259,9 +268,15 @@ class BigdlLLM(CustomLLM):
         if "max_length" not in tokenizer_kwargs:
             tokenizer_kwargs["max_length"] = context_window
 
-        self._tokenizer = tokenizer or AutoTokenizer.from_pretrained(
+        if tokenizer:
-            tokenizer_name, **tokenizer_kwargs
+            self._tokenizer = tokenizer
-        )
+        else:
+            print(f"load tokenizer: {tokenizer_name}")
+            try:
+                self._tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, **tokenizer_kwargs)
+            except:
+                self._tokenizer = LlamaTokenizer.from_pretrained(tokenizer_name,
+                                                                 trust_remote_code=True)
 
         if tokenizer_name != model_name:
             logger.warning(