Fix NPU LLM example save/load tokenizer (#12485)

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/Pipeline-Models/baichuan2.py

@@ -79,6 +79,8 @@ if __name__ == "__main__":
                                                      transpose_value_cache=not args.disable_transpose_value_cache,
                                                      trust_remote_code=True,
                                                      save_directory=args.save_directory)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -90,8 +92,8 @@ if __name__ == "__main__":
             transpose_value_cache=not args.disable_transpose_value_cache,
             trust_remote_code=True
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     if args.disable_streaming:
         streamer = None

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/Pipeline-Models/llama2.py

@@ -78,6 +78,8 @@ if __name__ == "__main__":
                                                      attn_implementation="eager",
                                                      transpose_value_cache=not args.disable_transpose_value_cache,
                                                      save_directory=args.save_directory)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -88,8 +90,8 @@ if __name__ == "__main__":
             pipeline=True,
             transpose_value_cache=not args.disable_transpose_value_cache,
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     if args.disable_streaming:
         streamer = None

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/Pipeline-Models/llama3.py

@@ -84,6 +84,8 @@ if __name__ == "__main__":
                                                     attn_implementation="eager",
                                                     transpose_value_cache=not args.disable_transpose_value_cache,
                                                     save_directory=args.save_directory)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -94,8 +96,8 @@ if __name__ == "__main__":
             pipeline=True,
             transpose_value_cache=not args.disable_transpose_value_cache,
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     if args.disable_streaming:
         streamer = None

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/Pipeline-Models/minicpm.py

@@ -66,6 +66,8 @@ if __name__ == "__main__":
                                                      transpose_value_cache=not args.disable_transpose_value_cache,
                                                      trust_remote_code=True,
                                                      save_directory=args.save_directory)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -77,8 +79,8 @@ if __name__ == "__main__":
             transpose_value_cache=not args.disable_transpose_value_cache,
             trust_remote_code=True
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     if args.disable_streaming:
         streamer = None

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/Pipeline-Models/qwen.py

@@ -70,6 +70,8 @@ if __name__ == "__main__":
                                                      mixed_precision=True,
                                                      trust_remote_code=True,
                                                      save_directory=args.save_directory)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -79,8 +81,8 @@ if __name__ == "__main__":
             max_prompt_len=args.max_prompt_len,
             pipeline=True,
             transpose_value_cache=not args.disable_transpose_value_cache)
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     if args.disable_streaming:
         streamer = None

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/baichuan2.py

@@ -79,6 +79,8 @@ if __name__ == "__main__":
             transpose_value_cache=not args.disable_transpose_value_cache,
             save_directory=args.save_directory
         )
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -90,8 +92,8 @@ if __name__ == "__main__":
             transpose_value_cache=not args.disable_transpose_value_cache,
             trust_remote_code=True,
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     DEFAULT_SYSTEM_PROMPT = """\
     """

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/generate.py

@@ -43,7 +43,6 @@ if __name__ == '__main__':
     args = parser.parse_args()
     model_path = args.repo_id_or_model_path
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     if not args.lowbit_path or not os.path.exists(args.lowbit_path):
         model = AutoModelForCausalLM.from_pretrained(
@@ -52,6 +51,8 @@ if __name__ == '__main__':
             load_in_low_bit=args.load_in_low_bit,
             attn_implementation="eager"
         )
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.lowbit_path)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.lowbit_path,
@@ -59,6 +60,7 @@ if __name__ == '__main__':
             bigdl_transformers_low_bit=args.load_in_low_bit,
             attn_implementation="eager"
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.lowbit_path, trust_remote_code=True)        
     
     print(model)
 

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/llama2.py

@@ -79,6 +79,8 @@ if __name__ == "__main__":
             transpose_value_cache=not args.disable_transpose_value_cache,
             save_directory=args.save_directory
         )
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -89,8 +91,8 @@ if __name__ == "__main__":
             max_prompt_len=args.max_prompt_len,
             transpose_value_cache=not args.disable_transpose_value_cache,
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     DEFAULT_SYSTEM_PROMPT = """\
     """

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/llama3.py

@@ -80,6 +80,8 @@ if __name__ == "__main__":
             transpose_value_cache=not args.disable_transpose_value_cache,
             save_directory=args.save_directory
         )
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -90,8 +92,8 @@ if __name__ == "__main__":
             max_prompt_len=args.max_prompt_len,
             transpose_value_cache=not args.disable_transpose_value_cache,
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     DEFAULT_SYSTEM_PROMPT = """\
     """

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/minicpm.py

@@ -65,6 +65,8 @@ if __name__ == "__main__":
             transpose_value_cache=not args.disable_transpose_value_cache,
             save_directory=args.save_directory
         )
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -76,7 +78,7 @@ if __name__ == "__main__":
             transpose_value_cache=not args.disable_transpose_value_cache,
             trust_remote_code=True,
         )
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
     print("-" * 80)
     print("done")

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/qwen.py

@@ -71,6 +71,8 @@ if __name__ == "__main__":
             quantization_group_size=args.quantization_group_size,
             save_directory=args.save_directory
         )
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(args.save_directory)
     else:
         model = AutoModelForCausalLM.load_low_bit(
             args.save_directory,
@@ -81,8 +83,8 @@ if __name__ == "__main__":
             max_prompt_len=args.max_prompt_len,
             transpose_value_cache=not args.disable_transpose_value_cache,
         )
+        tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
     print("-" * 80)
     print("done")