diff --git a/python/llm/example/CPU/ModelScope-Models/README.md b/python/llm/example/CPU/ModelScope-Models/README.md
new file mode 100644
index 00000000..05c0e3b4
--- /dev/null
+++ b/python/llm/example/CPU/ModelScope-Models/README.md
@@ -0,0 +1,78 @@
+# Run ModelScope Model
+
+In this directory, you will find example on how you could apply BigDL-LLM INT4 optimizations on ModelScope models. For illustration purposes, we utilize the [ZhipuAI/chatglm3-6b](https://modelscope.cn/models/ZhipuAI/chatglm3-6b/summary) as a reference ModelScope model.
+
+## 0. Requirements
+To run these examples with BigDL-LLM, we have some recommended requirements for your machine, please refer to [here](../README.md#recommended-requirements) for more information.
+
+## Example: Predict Tokens using `generate()` API
+In the example [generate.py](./generate.py), we show a basic use case for a ChatGLM3 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations.
+### 1. Install
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.9
+conda activate llm
+
+pip install --pre --upgrade bigdl-llm[all] # install bigdl-llm with 'all' option
+pip install modelscope
+```
+
+### 2. Run
+```
+python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
+```
+
+Arguments info:
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the ModelScope repo id for the ModelScope ChatGLM3 model to be downloaded, or the path to the ModelScope checkpoint folder. It is default to be `'ZhipuAI/chatglm3-6b'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'AI是什么？'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+
+> **Note**: When loading the model in 4-bit, BigDL-LLM converts linear layers in the model into INT4 format. In theory, a *X*B model saved in 16-bit will requires approximately 2*X* GB of memory for loading, and ~0.5*X* GB memory for further inference.
+>
+> Please select the appropriate size of the ChatGLM3 model based on the capabilities of your machine.
+
+#### 2.1 Client
+On client Windows machine, it is recommended to run directly with full utilization of all cores:
+```powershell
+python ./generate.py
+```
+
+#### 2.2 Server
+For optimal performance on server, it is recommended to set several environment variables (refer to [here](../README.md#best-known-configuration-on-linux) for more information), and run the example with all the physical cores of a single socket.
+
+E.g. on Linux,
+```bash
+# set BigDL-LLM env variables
+source bigdl-llm-init
+
+# e.g. for a server with 48 cores per socket
+export OMP_NUM_THREADS=48
+numactl -C 0-47 -m 0 python ./generate.py
+```
+
+#### 2.3 Sample Output
+#### [ZhipuAI/chatglm3-6b](https://modelscope.cn/models/ZhipuAI/chatglm3-6b/summary)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+<|user|>
+AI是什么？
+<|assistant|>
+-------------------- Output --------------------
+[gMASK]sop <|user|>
+AI是什么？
+<|assistant|> AI是人工智能（Artificial Intelligence）的缩写，指的是通过计算机程序和算法模拟人类智能的技术。AI可以帮助我们解决各种问题，例如语音
+```
+
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+<|user|>
+What is AI?
+<|assistant|>
+-------------------- Output --------------------
+[gMASK]sop <|user|>
+What is AI?
+<|assistant|>
+AI stands for Artificial Intelligence. It refers to the development of computer systems that can perform tasks that would normally require human intelligence, such as recognizing speech or making
+```
diff --git a/python/llm/example/CPU/ModelScope-Models/generate.py b/python/llm/example/CPU/ModelScope-Models/generate.py
new file mode 100644
index 00000000..3fef46fd
--- /dev/null
+++ b/python/llm/example/CPU/ModelScope-Models/generate.py
@@ -0,0 +1,71 @@
+#
+# 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
+import numpy as np
+
+from bigdl.llm.transformers import AutoModel
+from modelscope import AutoTokenizer
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to https://github.com/THUDM/ChatGLM3/blob/main/PROMPT.md
+CHATGLM_V3_PROMPT_FORMAT = "<|user|>\n{prompt}\n<|assistant|>"
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ModelScope ChatGLM3 model')
+ parser.add_argument('--repo-id-or-model-path', type=str, default="ZhipuAI/chatglm3-6b",
+ help='The ModelScope repo id for the ChatGLM3 model to be downloaded'
+ ', or the path to the ModelScope checkpoint folder')
+ parser.add_argument('--prompt', type=str, default="AI是什么？",
+ help='Prompt to infer')
+ parser.add_argument('--n-predict', type=int, default=32,
+ help='Max tokens to predict')
+
+ args = parser.parse_args()
+ model_path = args.repo_id_or_model_path
+
+ # Load model in 4 bit,
+ # which convert the relevant layers in the model into INT4 format
+ # It is important to set `model_hub='modelscope'`, otherwise model hub is default to be huggingface
+ model = AutoModel.from_pretrained(model_path,
+ load_in_4bit=True,
+ trust_remote_code=True,
+ model_hub='modelscope')
+
+ # Load tokenizer
+ tokenizer = AutoTokenizer.from_pretrained(model_path,
+ trust_remote_code=True)
+
+ # Generate predicted tokens
+ with torch.inference_mode():
+ prompt = CHATGLM_V3_PROMPT_FORMAT.format(prompt=args.prompt)
+ input_ids = tokenizer.encode(prompt, return_tensors="pt")
+ st = time.time()
+ # if your selected model is capable of utilizing previous key/value attentions
+ # to enhance decoding speed, but has `"use_cache": false` in its model config,
+ # it is important to set `use_cache=True` explicitly in the `generate` function
+ # to obtain optimal performance with BigDL-LLM INT4 optimizations
+ output = model.generate(input_ids,
+ max_new_tokens=args.n_predict)
+ end = time.time()
+ output_str = tokenizer.decode(output[0], skip_special_tokens=True)
+ print(f'Inference time: {end-st} s')
+ print('-'*20, 'Prompt', '-'*20)
+ print(prompt)
+ print('-'*20, 'Output', '-'*20)
+ print(output_str)
diff --git a/python/llm/example/CPU/README.md b/python/llm/example/CPU/README.md
index a1ce8091..2f86dc82 100644
--- a/python/llm/example/CPU/README.md
+++ b/python/llm/example/CPU/README.md
@@ -5,11 +5,13 @@ This folder contains examples of running BigDL-LLM on Intel CPU:
- [HF-Transformers-AutoModels](HF-Transformers-AutoModels): running any ***Hugging Face Transformers*** model on BigDL-LLM (using the standard AutoModel APIs)
- [QLoRA-FineTuning](QLoRA-FineTuning): running ***QLoRA finetuning*** using BigDL-LLM on intel CPUs
- [vLLM-Serving](vLLM-Serving): running ***vLLM*** serving framework on intel CPUs (with BigDL-LLM low-bit optimized models)
-- [Deepspeed-AutoTP](https://github.com/intel-analytics/BigDL/tree/main/python/llm/example/CPU/Deepspeed-AutoTP): running distributed inference using ***DeepSpeed AutoTP*** (with BigDL-LLM low-bit optimized models)
+- [Deepspeed-AutoTP](Deepspeed-AutoTP): running distributed inference using ***DeepSpeed AutoTP*** (with BigDL-LLM low-bit optimized models)
- [LangChain](LangChain): running ***LangChain*** applications on BigDL-LLM
- [Applications](Applications): running LLM applications (such as agent, streaming-llm) on BigDl-LLM
- [PyTorch-Models](PyTorch-Models): running any PyTorch model on BigDL-LLM (with "one-line code change")
- [Native-Models](Native-Models): converting & running LLM in `llama`/`chatglm`/`bloom`/`gptneox`/`starcoder` model family using native (cpp) implementation
+- [Speculative-Decoding](Speculative-Decoding): running any ***Hugging Face Transformers*** model with ***self-speculative decoding*** on Intel CPUs
+- [ModelScope-Models](ModelScope-Models): running ***ModelScope*** model with BigDL-LLM on Intel CPUs
## System Support
diff --git a/python/llm/example/GPU/ModelScope-Models/README.md b/python/llm/example/GPU/ModelScope-Models/README.md
new file mode 100644
index 00000000..1ba31567
--- /dev/null
+++ b/python/llm/example/GPU/ModelScope-Models/README.md
@@ -0,0 +1,135 @@
+# Run ModelScope Model
+
+In this directory, you will find example on how you could apply BigDL-LLM INT4 optimizations on ModelScope models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [ZhipuAI/chatglm3-6b](https://modelscope.cn/models/ZhipuAI/chatglm3-6b/summary) as a reference ModelScope model.
+
+## 0. Requirements
+To run these examples with BigDL-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../../../README.md#requirements) for more information.
+
+## Example: Predict Tokens using `generate()` API
+In the example [generate.py](./generate.py), we show a basic use case for a ChatGLM3 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations on Intel GPUs.
+### 1. Install
+#### 1.1 Installation on Linux
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.9
+conda activate llm
+# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
+pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
+pip install modelscope
+```
+
+#### 1.2 Installation on Windows
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.9 libuv
+conda activate llm
+# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
+pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
+pip install modelscope
+```
+
+### 2. Configures OneAPI environment variables
+#### 2.1 Configurations for Linux
+```bash
+source /opt/intel/oneapi/setvars.sh
+```
+
+#### 2.2 Configurations for Windows
+```cmd
+call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
+```
+> Note: Please make sure you are using **CMD** (**Anaconda Prompt** if using conda) to run the command as PowerShell is not supported.
+
+### 3. Runtime Configurations
+For optimal performance, it is recommended to set several environment variables. Please check out the suggestions based on your device.
+#### 3.1 Configurations for Linux
+
+
+For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series
+```bash
+export USE_XETLA=OFF
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+```
+
+
+
+
+
+For Intel Data Center GPU Max Series
+
+```bash
+export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+export ENABLE_SDP_FUSION=1
+```
+> Note: Please note that `libtcmalloc.so` can be installed by `conda install -c conda-forge -y gperftools=2.10`.
+
+
+#### 3.2 Configurations for Windows
+
+
+For Intel iGPU
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+set BIGDL_LLM_XMX_DISABLED=1
+```
+
+
+
+
+
+For Intel Arc™ A300-Series or Pro A60
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+```
+
+
+
+
+
+For other Intel dGPU Series
+
+There is no need to set further environment variables.
+
+
+
+> Note: For the first time that each model runs on Intel iGPU/Intel Arc™ A300-Series or Pro A60, it may take several minutes to compile.
+### 4. Running examples
+
+```
+python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
+```
+
+Arguments info:
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the ModelScope repo id for the ModelScope ChatGLM3 model to be downloaded, or the path to the ModelScope checkpoint folder. It is default to be `'ZhipuAI/chatglm3-6b'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'AI是什么？'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+
+#### Sample Output
+#### [ZhipuAI/chatglm3-6b](https://modelscope.cn/models/ZhipuAI/chatglm3-6b/summary)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+<|user|>
+AI是什么？
+<|assistant|>
+-------------------- Output --------------------
+[gMASK]sop <|user|>
+AI是什么？
+<|assistant|> AI是人工智能（Artificial Intelligence）的缩写，指的是通过计算机程序和算法模拟人类智能的技术。AI可以帮助我们解决各种问题，例如语音
+```
+
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+<|user|>
+What is AI?
+<|assistant|>
+-------------------- Output --------------------
+[gMASK]sop <|user|>
+What is AI?
+<|assistant|>
+AI stands for Artificial Intelligence. It refers to the development of computer systems that can perform tasks that would normally require human intelligence, such as recognizing speech or making
+```
diff --git a/python/llm/example/GPU/ModelScope-Models/generate.py b/python/llm/example/GPU/ModelScope-Models/generate.py
new file mode 100644
index 00000000..48bce466
--- /dev/null
+++ b/python/llm/example/GPU/ModelScope-Models/generate.py
@@ -0,0 +1,82 @@
+#
+# 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
+import numpy as np
+
+from bigdl.llm.transformers import AutoModel
+from modelscope import AutoTokenizer
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to https://github.com/THUDM/ChatGLM3/blob/main/PROMPT.md
+CHATGLM_V3_PROMPT_FORMAT = "<|user|>\n{prompt}\n<|assistant|>"
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM3 model')
+ parser.add_argument('--repo-id-or-model-path', type=str, default="ZhipuAI/chatglm3-6b",
+ help='The ModelScope repo id for the ChatGLM3 model to be downloaded'
+ ', or the path to the ModelScope checkpoint folder')
+ parser.add_argument('--prompt', type=str, default="AI是什么？",
+ help='Prompt to infer')
+ parser.add_argument('--n-predict', type=int, default=32,
+ help='Max tokens to predict')
+
+ args = parser.parse_args()
+ model_path = args.repo_id_or_model_path
+
+ # Load model in 4 bit,
+ # which convert the relevant layers in the model into INT4 format
+ # When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the from_pretrained function.
+ # This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
+ # It is important to set `model_hub='modelscope'`, otherwise model hub is default to be huggingface
+ model = AutoModel.from_pretrained(model_path,
+ load_in_4bit=True,
+ optimize_model=True,
+ trust_remote_code=True,
+ use_cache=True,
+ model_hub='modelscope')
+ model = model.to('xpu')
+
+ # Load tokenizer
+ tokenizer = AutoTokenizer.from_pretrained(model_path,
+ trust_remote_code=True)
+
+ # Generate predicted tokens
+ with torch.inference_mode():
+ prompt = CHATGLM_V3_PROMPT_FORMAT.format(prompt=args.prompt)
+ input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
+ # ipex model needs a warmup, then inference time can be accurate
+ output = model.generate(input_ids,
+ max_new_tokens=args.n_predict)
+
+ # start inference
+ st = time.time()
+ # if your selected model is capable of utilizing previous key/value attentions
+ # to enhance decoding speed, but has `"use_cache": false` in its model config,
+ # it is important to set `use_cache=True` explicitly in the `generate` function
+ # to obtain optimal performance with BigDL-LLM INT4 optimizations
+ output = model.generate(input_ids,
+ max_new_tokens=args.n_predict)
+ torch.xpu.synchronize()
+ end = time.time()
+ output_str = tokenizer.decode(output[0], skip_special_tokens=True)
+ print(f'Inference time: {end-st} s')
+ print('-'*20, 'Prompt', '-'*20)
+ print(prompt)
+ print('-'*20, 'Output', '-'*20)
+ print(output_str)
diff --git a/python/llm/example/GPU/README.md b/python/llm/example/GPU/README.md
index f72dd80d..5e18af8a 100644
--- a/python/llm/example/GPU/README.md
+++ b/python/llm/example/GPU/README.md
@@ -2,11 +2,14 @@
This folder contains examples of running BigDL-LLM on Intel GPU:
+- [Applications](Applications): running LLM applications (such as autogen) on BigDL-LLM
- [HF-Transformers-AutoModels](HF-Transformers-AutoModels): running any ***Hugging Face Transformers*** model on BigDL-LLM (using the standard AutoModel APIs)
- [LLM-Finetuning](LLM-Finetuning): running ***finetuning*** (such as LoRA, QLoRA, QA-LoRA, etc) using BigDL-LLM on Intel GPUs
- [vLLM-Serving](vLLM-Serving): running ***vLLM*** serving framework on intel GPUs (with BigDL-LLM low-bit optimized models)
- [Deepspeed-AutoTP](Deepspeed-AutoTP): running distributed inference using ***DeepSpeed AutoTP*** (with BigDL-LLM low-bit optimized models) on Intel GPUs
- [PyTorch-Models](PyTorch-Models): running any PyTorch model on BigDL-LLM (with "one-line code change")
+- [Speculative-Decoding](Speculative-Decoding): running any ***Hugging Face Transformers*** model with ***self-speculative decoding*** on Intel GPUs
+- [ModelScope-Models](ModelScope-Models): running ***ModelScope*** model with BigDL-LLM on Intel GPUs
## System Support