From f1f4094a09be523a43652f1c292f673d37636e98 Mon Sep 17 00:00:00 2001
From: yb-peng <75617475+pengyb2001@users.noreply.github.com>
Date: Fri, 23 Feb 2024 14:05:53 +0800
Subject: [PATCH] Add CPU and GPU examples of phi-2 (#10014)

* Add CPU and GPU examples of phi-2

* In GPU hf example, updated the readme for Windows GPU supports

* In GPU torch example, updated the readme for Windows GPU supports

* update the table in BigDL/README.md

* update the table in BigDL/python/llm/README.md
---
 README.md                                     |   1 +
 python/llm/README.md                          |   1 +
 .../Model/phi-2/README.md                     |  74 +++++++++++
 .../Model/phi-2/generate.py                   |  72 +++++++++++
 .../CPU/PyTorch-Models/Model/phi-2/README.md  |  60 +++++++++
 .../PyTorch-Models/Model/phi-2/generate.py    |  61 +++++++++
 .../Model/phi-2/README.md                     | 118 ++++++++++++++++++
 .../Model/phi-2/generate.py                   |  83 ++++++++++++
 .../GPU/PyTorch-Models/Model/phi-2/README.md  | 116 +++++++++++++++++
 .../PyTorch-Models/Model/phi-2/generate.py    |  73 +++++++++++
 10 files changed, 659 insertions(+)
 create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2/README.md
 create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2/generate.py
 create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/phi-2/README.md
 create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/phi-2/generate.py
 create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2/README.md
 create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2/generate.py
 create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/phi-2/README.md
 create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/phi-2/generate.py

diff --git a/README.md b/README.md
index d2ab19b9..cc63e481 100644
--- a/README.md
+++ b/README.md
@@ -190,6 +190,7 @@ Over 40 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
 | Bark | [link](python/llm/example/CPU/PyTorch-Models/Model/bark) | [link](python/llm/example/GPU/PyTorch-Models/Model/bark) |
 | SpeechT5 |  | [link](python/llm/example/GPU/PyTorch-Models/Model/speech-t5) |
 | Ziya-Coding-34B-v1.0 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya) | |
+| Phi-2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2) |
 
 ***For more details, please refer to the `bigdl-llm` [Document](https://test-bigdl-llm.readthedocs.io/en/main/doc/LLM/index.html), [Readme](python/llm), [Tutorial](https://github.com/intel-analytics/bigdl-llm-tutorial) and [API Doc](https://bigdl.readthedocs.io/en/latest/doc/PythonAPI/LLM/index.html).***
 
diff --git a/python/llm/README.md b/python/llm/README.md
index 334296b5..55ea146a 100644
--- a/python/llm/README.md
+++ b/python/llm/README.md
@@ -82,6 +82,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
 | Bark | [link](example/CPU/PyTorch-Models/Model/bark) | [link](example/GPU/PyTorch-Models/Model/bark) |
 | SpeechT5 |  | [link](example/GPU/PyTorch-Models/Model/speech-t5) |
 | Ziya-Coding-34B-v1.0 | [link](example/CPU/HF-Transformers-AutoModels/Model/ziya) | |
+| Phi-2 | [link](example/CPU/HF-Transformers-AutoModels/Model/phi-2) | [link](example/GPU/HF-Transformers-AutoModels/Model/phi-2) |
 
 ### Working with `bigdl-llm`
 
diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2/README.md
new file mode 100644
index 00000000..4400e0e5
--- /dev/null
+++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2/README.md
@@ -0,0 +1,74 @@
+# phi-2
+
+In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on phi-2 models. For illustration purposes, we utilize the [microsoft/phi-2](https://huggingface.co/microsoft/phi-2 as a reference phi-2 model.
+
+> **Note**: If you want to download the Hugging Face *Transformers* model, please refer to [here](https://huggingface.co/docs/hub/models-downloading#using-git).
+>
+> BigDL-LLM optimizes the *Transformers* model in INT4 precision at runtime, and thus no explicit conversion is needed.
+
+## 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 phi-2 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations.
+### 1. Install
+We suggest using conda to manage the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
+
+After installing conda, create a Python environment for BigDL-LLM:
+```bash
+conda create -n llm python=3.9 # recommend to use Python 3.9
+conda activate llm
+
+pip install --pre --upgrade bigdl-llm[all] # install the latest bigdl-llm nightly build with 'all' option
+pip install einops  # additional package required for phi-2 to conduct generation
+```
+
+### 2. Run
+After setting up the Python environment, you could run the example by following steps.
+
+> **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 phi-2 model based on the capabilities of your machine.
+
+#### 2.1 Client
+On client Windows machines, it is recommended to run directly with full utilization of all cores:
+```powershell
+python ./generate.py --prompt 'What is AI?'
+```
+More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
+
+#### 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 --prompt 'What is AI？'
+```
+More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
+
+#### 2.3 Arguments Info
+In the example, several arguments can be passed to satisfy your requirements:
+
+- `--repo-id-or-model-path`: str, argument defining the huggingface repo id for the phi-2 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/phi-2'`.
+- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `What is AI?`.
+- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `32`.
+
+#### 2.4 Sample Output
+#### [microsoft/phi-2](https://huggingface.co/microsoft/phi-2)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+ Question:What is AI?
+
+ Answer:
+-------------------- Output --------------------
+ Question:What is AI?
+
+ Answer: AI stands for Artificial Intelligence. It is a field of computer science that focuses on creating intelligent machines that can perform tasks that would typically require human intelligence.
+
+```
\ No newline at end of file
diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2/generate.py b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2/generate.py
new file mode 100644
index 00000000..e33ec220
--- /dev/null
+++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2/generate.py
@@ -0,0 +1,72 @@
+#
+# 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,AutoModelForCausalLM
+from transformers import AutoTokenizer, GenerationConfig
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to  https://huggingface.co/microsoft/phi-2
+PHI2_PROMPT_FORMAT = " Question:{prompt}\n\n Answer:"
+generation_config = GenerationConfig(use_cache = True)
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-2 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-2",
+                        help='The huggingface repo id for the phi-2 model to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    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')
+
+    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
+    model = AutoModelForCausalLM.from_pretrained(model_path,
+                                                 load_in_4bit=True,
+                                                 trust_remote_code=True)
+
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path,
+                                              trust_remote_code=True)
+    
+    # Generate predicted tokens
+    with torch.inference_mode():
+        prompt = PHI2_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
+
+        # Note that phi-2 uses GenerationConfig to enable 'use_cache'
+        output = model.generate(input_ids, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config)
+
+        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/PyTorch-Models/Model/phi-2/README.md b/python/llm/example/CPU/PyTorch-Models/Model/phi-2/README.md
new file mode 100644
index 00000000..2d3dd257
--- /dev/null
+++ b/python/llm/example/CPU/PyTorch-Models/Model/phi-2/README.md
@@ -0,0 +1,60 @@
+# phi-2
+In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate phi-2 models. For illustration purposes, we utilize the [microsoft/phi-2](https://huggingface.co/microsoft/phi-2) as a reference phi-2 model.
+
+## 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 phi-2 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations.
+### 1. Install
+We suggest using conda to manage the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
+
+After installing conda, create a Python environment for BigDL-LLM:
+```bash
+conda create -n llm python=3.9 # recommend to use Python 3.9
+conda activate llm
+
+pip install --pre --upgrade bigdl-llm[all] # install the latest bigdl-llm nightly build with 'all' option
+pip install einops 
+```
+
+### 2. Run
+After setting up the Python environment, you could run the example by following steps.
+
+#### 2.1 Client
+On client Windows machines, it is recommended to run directly with full utilization of all cores:
+```powershell
+python ./generate.py --prompt 'What is AI?'
+```
+More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
+
+#### 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 --prompt 'What is AI?'
+```
+More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
+
+#### 2.3 Arguments Info
+In the example, several arguments can be passed to satisfy your requirements:
+
+- `--repo-id-or-model-path`: str, argument defining the huggingface repo id for the phi-2 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/phi-2'`.
+- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `'What is AI?'`.
+- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `32`.
+
+#### 2.4 Sample Output
+#### [microsoft/phi-2](https://huggingface.co/microsoft/phi-2)
+```log
+Inference time: xxxx s
+-------------------- Output --------------------
+Question: What is AI?
+
+Answer: AI stands for Artificial Intelligence. It is a field of computer science that focuses on creating intelligent machines that can perform tasks that typically require human intelligence.
+```
diff --git a/python/llm/example/CPU/PyTorch-Models/Model/phi-2/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/phi-2/generate.py
new file mode 100644
index 00000000..b66a3b2b
--- /dev/null
+++ b/python/llm/example/CPU/PyTorch-Models/Model/phi-2/generate.py
@@ -0,0 +1,61 @@
+#
+# 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 transformers import AutoModel, AutoTokenizer, AutoModelForCausalLM, GenerationConfig
+from bigdl.llm import optimize_model
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to https://huggingface.co/microsoft/phi-2
+PHI_2_V1_PROMPT_FORMAT = "Question: {prompt}\n\n Answer:"
+generation_config = GenerationConfig(use_cache = True)
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-2 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-2",
+                        help='The huggingface repo id for the phi-2 model to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    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')
+
+    args = parser.parse_args()
+    model_path = args.repo_id_or_model_path
+
+    # Load model
+    model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
+
+    # With only one line to enable BigDL-LLM optimization on model
+    model = optimize_model(model)
+    
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    
+    # Generate predicted tokens
+    with torch.inference_mode():
+        prompt = PHI_2_V1_PROMPT_FORMAT.format(prompt=args.prompt)
+        input_ids = tokenizer.encode(prompt, return_tensors="pt")
+        st = time.time()
+        output = model.generate(input_ids, max_new_tokens=args.n_predict, generation_config = generation_config)
+        end = time.time()
+        output_str = tokenizer.decode(output[0], skip_special_tokens=True)
+        print(f'Inference time: {end-st} s')
+        print('-'*20, 'Output', '-'*20)
+        print(output_str)
diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2/README.md b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2/README.md
new file mode 100644
index 00000000..a251f4da
--- /dev/null
+++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2/README.md
@@ -0,0 +1,118 @@
+# phi-2
+In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on phi-2 models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [microsoft/phi-2](https://huggingface.co/microsoft/phi-2) as a reference phi-2 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 phi-2 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 einops # additional package required for phi-2 to conduct generation
+```
+#### 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
+```
+
+### 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
+<details>
+
+<summary>For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series</summary>
+
+```bash
+export USE_XETLA=OFF
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+```
+</details>
+
+<details>
+
+<summary>For Intel Data Center GPU Max Series</summary>
+
+```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`.
+</details>
+#### 3.2 Configurations for Windows
+<details>
+
+<summary>For Intel iGPU</summary>
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+set BIGDL_LLM_XMX_DISABLED=1
+```
+
+</details>
+
+<details>
+
+<summary>For Intel Arc™ A300-Series or Pro A60</summary>
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+```
+
+</details>
+
+<details>
+
+<summary>For other Intel dGPU Series</summary>
+
+There is no need to set further environment variables.
+
+</details>
+
+> 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 --prompt 'What is AI?'
+```
+
+Arguments info:
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the phi-2 model (e.g. `microsoft/phi-2`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/phi-2'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (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`.
+
+#### Sample Output
+#### [microsoft/phi-2](https://huggingface.co/microsoft/phi-2)
+
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+Question: What is AI?
+
+ Answer:
+-------------------- Output --------------------
+Question: What is AI?
+
+ Answer: AI stands for Artificial Intelligence. It is a field of computer science that focuses on creating intelligent machines that can perform tasks that would typically require human intelligence.
+```
diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2/generate.py b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2/generate.py
new file mode 100644
index 00000000..79c2fbfd
--- /dev/null
+++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2/generate.py
@@ -0,0 +1,83 @@
+#
+# 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,AutoModelForCausalLM
+from transformers import AutoTokenizer, GenerationConfig
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to  https://huggingface.co/microsoft/phi-2
+PHI2_PROMPT_FORMAT = " Question:{prompt}\n\n Answer:"
+generation_config = GenerationConfig(use_cache = True)
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-2 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-2",
+                        help='The huggingface repo id for the phi-2 model to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    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')
+
+    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.
+    model = AutoModelForCausalLM.from_pretrained(model_path,
+                                                 load_in_4bit=True,
+                                                 trust_remote_code=True)
+
+    model = model.to('xpu')
+
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path,
+                                              trust_remote_code=True)
+    
+    # Generate predicted tokens
+    with torch.inference_mode():
+        prompt = PHI2_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,
+                                generation_config = generation_config)
+        # 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
+
+        # Note that phi-2 uses GenerationConfig to enable 'use_cache'
+        output = model.generate(input_ids, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config)
+        torch.xpu.synchronize()
+        end = time.time()
+        output = output.cpu()
+        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/PyTorch-Models/Model/phi-2/README.md b/python/llm/example/GPU/PyTorch-Models/Model/phi-2/README.md
new file mode 100644
index 00000000..c9614ab4
--- /dev/null
+++ b/python/llm/example/GPU/PyTorch-Models/Model/phi-2/README.md
@@ -0,0 +1,116 @@
+# phi-2
+In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate phi-2 models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [microsoft/phi-2](https://huggingface.co/microsoft/phi-2) as a reference phi-2 model.
+
+## Requirements
+To run these examples with BigDL-LLM, 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 phi-2 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 the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
+
+After installing conda, create a Python environment for BigDL-LLM:
+```bash
+conda create -n llm python=3.9 # recommend to use Python 3.9
+conda activate llm
+
+pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
+pip install einops # additional package required for phi-2 to conduct generation
+```
+#### 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
+```
+
+### 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
+<details>
+
+<summary>For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series</summary>
+
+```bash
+export USE_XETLA=OFF
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+```
+</details>
+
+<details>
+
+<summary>For Intel Data Center GPU Max Series</summary>
+
+```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`.
+</details>
+#### 3.2 Configurations for Windows
+<details>
+
+<summary>For Intel iGPU</summary>
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+set BIGDL_LLM_XMX_DISABLED=1
+```
+
+</details>
+
+<details>
+
+<summary>For Intel Arc™ A300-Series or Pro A60</summary>
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+```
+
+</details>
+
+<details>
+
+<summary>For other Intel dGPU Series</summary>
+
+There is no need to set further environment variables.
+
+</details>
+
+> 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 --prompt 'What is AI?'
+```
+
+Arguments info:
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the phi-2 model (e.g. `microsoft/phi-2`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/phi-2'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (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`.
+
+#### Sample Output
+#### [microsoft/phi-2](https://huggingface.co/microsoft/phi-2)
+
+```log
+Inference time: xxxx s
+-------------------- Output --------------------
+Question: What is AI?
+
+Answer: AI stands for Artificial Intelligence. It is a field of computer science that focuses on creating intelligent machines that can perform tasks that would normally require human intelligence.
+```
\ No newline at end of file
diff --git a/python/llm/example/GPU/PyTorch-Models/Model/phi-2/generate.py b/python/llm/example/GPU/PyTorch-Models/Model/phi-2/generate.py
new file mode 100644
index 00000000..242f815d
--- /dev/null
+++ b/python/llm/example/GPU/PyTorch-Models/Model/phi-2/generate.py
@@ -0,0 +1,73 @@
+#
+# 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 transformers import AutoModel, AutoTokenizer, AutoModelForCausalLM, GenerationConfig
+from bigdl.llm import optimize_model
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to https://huggingface.co/microsoft/phi-2
+PHI_2_V1_PROMPT_FORMAT = "Question: {prompt}\n\n Answer:"
+generation_config = GenerationConfig(use_cache = True)
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-2 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-2",
+                        help='The huggingface repo id for the phi-2 model to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    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')
+
+    args = parser.parse_args()
+    model_path = args.repo_id_or_model_path
+
+    # Load model
+    model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
+
+    # With only one line to enable BigDL-LLM optimization on model
+    # When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the optimize_model function.
+    # This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
+    model = optimize_model(model)
+    model = model.to('xpu')
+    
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    
+    # Generate predicted tokens
+    with torch.inference_mode():
+        prompt = PHI_2_V1_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, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config)
+        # start inference
+        st = time.time()
+        # Note that phi-2 uses GenerationConfig to enable 'use_cache'
+        output = model.generate(input_ids, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config)
+        torch.xpu.synchronize()
+        end = time.time()
+        output = output.cpu()
+        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)