Add guide for save-load usage (#12498)

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

@@ -0,0 +1,70 @@
+# Save/Load Low-Bit Models with IPEX-LLM Optimizations
+
+In this directory, you will find example on how you could save/load models with IPEX-LLM optimizations on Intel NPU.
+
+## Example: Save/Load Optimized Models
+In the example [generate.py](./generate.py), we show a basic use case of saving/loading model in low-bit optimizations to predict the next N tokens using `generate()` API.
+
+## 0. Prerequisites
+For `ipex-llm` NPU support, please refer to [Quick Start](../../../../../../../docs/mddocs/Quickstart/npu_quickstart.md#install-prerequisites) for details about the required preparations.
+
+## 1. Install & Runtime Configurations
+### 1.1 Installation on Windows
+We suggest using conda to manage environment:
+```cmd
+conda create -n llm python=3.11
+conda activate llm
+
+:: install ipex-llm with 'npu' option
+pip install --pre --upgrade ipex-llm[npu]
+
+:: [optional] for Llama-3.2-1B-Instruct & Llama-3.2-3B-Instruct
+pip install transformers==4.45.0 accelerate==0.33.0
+```
+Please refer to [Quick Start](../../../../../../../docs/mddocs/Quickstart/npu_quickstart.md#install-prerequisites) for more details about `ipex-llm` installation on Intel NPU.
+
+### 1.2 Runtime Configurations
+Please refer to [Quick Start](../../../../../../../docs/mddocs/Quickstart/npu_quickstart.md#runtime-configurations) for environment variables setting based on your device.
+
+## 3. Running examples
+
+If you want to save the optimized model, run:
+```
+python ./generate.py --repo-id-or-model-path "meta-llama/Llama-2-7b-chat-hf" --save-directory path/to/save/model
+```
+
+If you want to load the optimized low-bit model, run:
+```
+python ./generate.py --load-directory path/to/load/model
+```
+
+In the example, several arguments can be passed to satisfy your requirements:
+
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Llama2 model to be downloaded, or the path to the ModelScope checkpoint folder. It is default to be `'meta-llama/Llama-2-7b-chat-hf'`.
+- `--save-directory`: argument defining the path to save the low-bit model. Then you can load the low-bit directly.
+- `--load-directory`: argument defining the path to load low-bit model.
+- `--prompt PROMPT`: argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `'What is AI?'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+- `--max-context-len MAX_CONTEXT_LEN`: Defines the maximum sequence length for both input and output tokens. It is default to be `1024`.
+- `--max-prompt-len MAX_PROMPT_LEN`: Defines the maximum number of tokens that the input prompt can contain. It is default to be `512`.
+
+### Sample Output
+#### [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf)
+```log
+Inference time: xxxx s
+-------------------- Input --------------------
+<s><s>  [INST] <<SYS>>
+
+<</SYS>>
+
+What is AI? [/INST]
+
+-------------------- Output --------------------
+<s><s>  [INST] <<SYS>>
+
+<</SYS>>
+
+What is AI? [/INST]
+
+Artificial Intelligence (AI) is a field of computer science and technology that focuses on the development of intelligent machines that can perform tasks that
+```

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

@@ -0,0 +1,104 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import torch
+import time
+import argparse
+from ipex_llm.transformers.npu_model import AutoModelForCausalLM
+from transformers import AutoTokenizer
+from ipex_llm.utils.common.log4Error import invalidInputError
+
+
+# you could tune the prompt based on your own model,
+LLAMA2_PROMPT_FORMAT = """<s> [INST] <<SYS>>
+
+<</SYS>>
+
+{prompt} [/INST]
+"""
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Example of saving and loading the optimized model')
+    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 (e.g. `meta-llama/Llama-2-7b-chat-hf` and `meta-llama/Llama-2-13b-chat-hf`) to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--save-directory', type=str, default=None,
+                        help='The path to save the low-bit model.')
+    parser.add_argument('--load-directory', type=str, default=None,
+                        help='The path to load the low-bit model.')
+    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')
+    parser.add_argument("--max-context-len", type=int, default=1024)
+    parser.add_argument("--max-prompt-len", type=int, default=512)
+    
+    args = parser.parse_args()
+    model_path = args.repo_id_or_model_path
+    save_directory = args.save_directory
+    load_directory = args.load_directory
+
+    if save_directory:
+        # first time to load and save
+        model = AutoModelForCausalLM.from_pretrained(
+            model_path,
+            torch_dtype=torch.float16,
+            trust_remote_code=True,
+            attn_implementation="eager",
+            load_in_low_bit="sym_int4",
+            optimize_model=True,
+            max_context_len=args.max_context_len,
+            max_prompt_len=args.max_prompt_len,
+            save_directory=save_directory
+        )
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer.save_pretrained(save_directory)
+        print(f"Finish to load model from {model_path} and save to {save_directory}")
+    elif load_directory:
+        # load low-bit model
+        model = AutoModelForCausalLM.load_low_bit(
+            load_directory,
+            attn_implementation="eager",
+            torch_dtype=torch.float16,
+            optimize_model=True,
+            max_context_len=args.max_context_len,
+            max_prompt_len=args.max_prompt_len
+        )
+        tokenizer = AutoTokenizer.from_pretrained(load_directory, trust_remote_code=True)
+        print(f"Finish to load model from {load_directory}")
+    else:
+        invalidInputError(False,
+                          "Both `--save-directory` and `--load-directory` are None, please provide one of this.")
+
+    # Generate predicted tokens
+    with torch.inference_mode():
+        for i in range(3):
+            prompt = LLAMA2_PROMPT_FORMAT.format(prompt=args.prompt)
+            _input_ids = tokenizer.encode(prompt, return_tensors="pt")
+
+            st = time.time()
+            output = model.generate(
+                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict
+            )
+            end = time.time()
+
+            print(f"Inference time: {end-st} s")
+            input_str = tokenizer.decode(_input_ids[0], skip_special_tokens=False)
+            print("-" * 20, "Input", "-" * 20)
+            print(input_str)
+            output_str = tokenizer.decode(output[0], skip_special_tokens=False)
+            print("-" * 20, "Output", "-" * 20)
+            print(output_str)