update lowbit path for baichuan2, qwen2, generate.py (#12051)

* update lowbit path for baichuan2, qwen2, `generate.py`

* update readme

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md

@@ -61,6 +61,7 @@ python ./generate.py
 
 Arguments info:
 - `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Llama2 model (e.g. `meta-llama/Llama-2-7b-chat-hf`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'meta-llama/Llama-2-7b-chat-hf'`, and more verified models please see the list in [Verified Models](#verified-models).
+- `--lowbit-path LOWBIT_MODEL_PATH`: argument defining the path to save/load lowbit version of the model. If it is an empty string, the original pretrained model specified by `REPO_ID_OR_MODEL_PATH` will be loaded. If it is an existing path, the lowbit model in `LOWBIT_MODEL_PATH` will be loaded. If it is a non-existing path, the original pretrained model specified by `REPO_ID_OR_MODEL_PATH` will be loaded, and the converted lowbit version will be saved into `LOWBIT_MODEL_PATH`. It is default to be `''`, i.e. an empty string.
 - `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'Once upon a time, there existed a little girl who liked to have adventures. She wanted to go to places and meet new people, and have fun'`.
 - `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
 - `--load_in_low_bit`: argument defining the `load_in_low_bit` format used. It is default to be `sym_int8`, `sym_int4` can also be used.
@@ -131,6 +132,9 @@ Arguments info:
 
 ### Troubleshooting
 
+#### `TypeError: can't convert meta device type tensor to numpy.` Error
+If you encounter `TypeError: can't convert meta device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.` error when loading lowbit model, please try re-saving the lowbit model with the example script you are currently using. Please note that lowbit models saved by `qwen2.py`, `llama.py`, etc. cannot be loaded by `generate.py`.
+
 #### Output Problem
 If you encounter output problem, please try to disable the optimization of transposing value cache with following command:
 ```bash

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

@@ -50,6 +50,12 @@ if __name__ == "__main__":
         help="The huggingface repo id for the Baichuan2 model to be downloaded"
         ", or the path to the huggingface checkpoint folder",
     )
+    parser.add_argument("--lowbit-path", type=str,
+        default="",
+        help="The path to the lowbit model folder, leave blank if you do not want to save. \
+            If path not exists, lowbit model will be saved there. \
+            Else, lowbit model will be loaded.",
+    )
     parser.add_argument('--prompt', type=str, default="What is AI?",
                         help='Prompt to infer')
     parser.add_argument("--n-predict", type=int, default=32, help="Max tokens to predict")
@@ -62,6 +68,7 @@ if __name__ == "__main__":
     args = parser.parse_args()
     model_path = args.repo_id_or_model_path
 
+    if not args.lowbit_path or not os.path.exists(args.lowbit_path):
         model = AutoModelForCausalLM.from_pretrained(
             model_path,
             torch_dtype=torch.bfloat16,
@@ -75,9 +82,25 @@ if __name__ == "__main__":
             inter_pp=args.inter_pp,
             transpose_value_cache=not args.disable_transpose_value_cache,
         )
+    else:
+        model = AutoModelForCausalLM.load_low_bit(
+            args.lowbit_path,
+            attn_implementation="eager",
+            torch_dtype=torch.bfloat16,
+            optimize_model=True,
+            max_output_len=args.max_output_len,
+            max_prompt_len=args.max_prompt_len,
+            intra_pp=args.intra_pp,
+            inter_pp=args.inter_pp,
+            transpose_value_cache=not args.disable_transpose_value_cache,
+            trust_remote_code=True,
+        )
 
     tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
+    if args.lowbit_path and not os.path.exists(args.lowbit_path):
+        model.save_low_bit(args.lowbit_path)
+
     DEFAULT_SYSTEM_PROMPT = """\
     """
 

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

@@ -17,6 +17,7 @@
 import torch
 import time
 import argparse
+import os
 
 from ipex_llm.transformers.npu_model import AutoModelForCausalLM
 from transformers import AutoTokenizer
@@ -27,6 +28,11 @@ if __name__ == '__main__':
     parser.add_argument('--repo-id-or-model-path', type=str, default="meta-llama/Llama-2-7b-chat-hf",
                         help='The huggingface repo id for the Llama2 model to be downloaded'
                              ', or the path to the huggingface checkpoint folder')
+    parser.add_argument("--lowbit-path", type=str,
+        default="",
+        help='The path to the lowbit model folder, leave blank if you do not want to save. \
+            If path not exists, lowbit model will be saved there. \
+            Else, lowbit model will be loaded.')
     parser.add_argument('--prompt', type=str, default="Once upon a time, there existed a little girl who liked to have adventures. She wanted to go to places and meet new people, and have fun",
                         help='Prompt to infer')
     parser.add_argument('--n-predict', type=int, default=32,
@@ -39,12 +45,26 @@ if __name__ == '__main__':
 
     tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
-    model = AutoModelForCausalLM.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(
+            model_path,
+            trust_remote_code=True,
             load_in_low_bit=args.load_in_low_bit,
-                                                 attn_implementation="eager")
+            attn_implementation="eager"
+        )
+    else:
+        model = AutoModelForCausalLM.load_low_bit(
+            args.lowbit_path,
+            trust_remote_code=True,
+            bigdl_transformers_low_bit=args.load_in_low_bit,
+            attn_implementation="eager"
+        )
     
     print(model)
 
+    if args.lowbit_path and not os.path.exists(args.lowbit_path):
+        model.save_low_bit(args.lowbit_path)
+
     with torch.inference_mode():
         prompt = "Once upon a time, there existed a little girl who liked to have adventures. She wanted to go to places and meet new people, and have fun"
         input_ids = tokenizer.encode(prompt, return_tensors="pt")

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/qwen2.py

@@ -37,6 +37,12 @@ if __name__ == "__main__":
         help="The huggingface repo id for the Qwen2 model to be downloaded"
         ", or the path to the huggingface checkpoint folder",
     )
+    parser.add_argument("--lowbit-path", type=str,
+        default="",
+        help="The path to the lowbit model folder, leave blank if you do not want to save. \
+            If path not exists, lowbit model will be saved there. \
+            Else, lowbit model will be loaded.",
+    )
     parser.add_argument('--prompt', type=str, default="What is AI?",
                         help='Prompt to infer')
     parser.add_argument("--n-predict", type=int, default=32, help="Max tokens to predict")
@@ -49,6 +55,7 @@ if __name__ == "__main__":
     args = parser.parse_args()
     model_path = args.repo_id_or_model_path
 
+    if not args.lowbit_path or not os.path.exists(args.lowbit_path):
         model = AutoModelForCausalLM.from_pretrained(
             model_path,
             torch_dtype=torch.float16,
@@ -62,9 +69,24 @@ if __name__ == "__main__":
             inter_pp=args.inter_pp,
             transpose_value_cache=not args.disable_transpose_value_cache,
         )
+    else:
+        model = AutoModelForCausalLM.load_low_bit(
+            args.lowbit_path,
+            attn_implementation="eager",
+            torch_dtype=torch.float16,
+            optimize_model=True,
+            max_output_len=args.max_output_len,
+            max_prompt_len=args.max_prompt_len,
+            intra_pp=args.intra_pp,
+            inter_pp=args.inter_pp,
+            transpose_value_cache=not args.disable_transpose_value_cache,
+        )
 
     tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
+    if args.lowbit_path and not os.path.exists(args.lowbit_path):
+        model.save_low_bit(args.lowbit_path)
+
     print("-" * 80)
     print("done")
     messages = [{"role": "system", "content": "You are a helpful assistant."},