From d9d496c8f3089d821b363430a731ee28edd70d5c Mon Sep 17 00:00:00 2001
From: Zhicun <59141989+ivy-lv11@users.noreply.github.com>
Date: Mon, 5 Feb 2024 11:12:47 +0800
Subject: [PATCH] add phixtral and optimize phi-moe (#10052)

---
 README.md                                     |   1 +
 python/llm/README.md                          |   1 +
 .../Model/phixtral/README.md                  |  73 +++++++++
 .../Model/phixtral/generate.py                |  72 +++++++++
 .../PyTorch-Models/Model/phixtral/README.md   |  64 ++++++++
 .../PyTorch-Models/Model/phixtral/generate.py |  66 ++++++++
 .../Model/phixtral/README.md                  | 119 +++++++++++++++
 .../Model/phixtral/generate.py                |  80 ++++++++++
 .../PyTorch-Models/Model/phixtral/README.md   | 123 +++++++++++++++
 .../PyTorch-Models/Model/phixtral/generate.py |  80 ++++++++++
 .../llm/src/bigdl/llm/transformers/convert.py |  11 ++
 .../bigdl/llm/transformers/models/phixtral.py | 144 ++++++++++++++++++
 12 files changed, 834 insertions(+)
 create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral/README.md
 create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral/generate.py
 create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/phixtral/README.md
 create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/phixtral/generate.py
 create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral/README.md
 create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral/generate.py
 create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/phixtral/README.md
 create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/phixtral/generate.py
 create mode 100644 python/llm/src/bigdl/llm/transformers/models/phixtral.py

diff --git a/README.md b/README.md
index b5746d9e5c7..86b6ebd20e6 100644
--- a/README.md
+++ b/README.md
@@ -177,6 +177,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
 | Yi | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/yi) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/yi) |
 | BlueLM | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/bluelm) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/bluelm) |
 | SOLAR | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/solar) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/solar) |
+| Phixtral | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral) |
 | InternLM2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/internlm2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/internlm2) |
 
 ***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 2a8369be05a..2389cc0b8c9 100644
--- a/python/llm/README.md
+++ b/python/llm/README.md
@@ -75,6 +75,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
 | Yi | [link](example/CPU/HF-Transformers-AutoModels/Model/yi) | [link](example/GPU/HF-Transformers-AutoModels/Model/yi) |
 | BlueLM | [link](example/CPU/HF-Transformers-AutoModels/Model/bluelm) | [link](example/GPU/HF-Transformers-AutoModels/Model/bluelm) |
 | SOLAR | [link](example/CPU/HF-Transformers-AutoModels/Model/solar) | [link](example/GPU/HF-Transformers-AutoModels/Model/solar) |
+| Phixtral | [link](example/CPU/HF-Transformers-AutoModels/Model/phixtral) | [link](example/GPU/HF-Transformers-AutoModels/Model/phixtral) |
 | InternLM2 | [link](example/CPU/HF-Transformers-AutoModels/Model/internlm2) | [link](example/GPU/HF-Transformers-AutoModels/Model/internlm2) |
 
 ### Working with `bigdl-llm`
diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral/README.md
new file mode 100644
index 00000000000..4382aec0ee9
--- /dev/null
+++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral/README.md
@@ -0,0 +1,73 @@
+# Phixtral-4x2_8
+
+In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on phi models. For illustration purposes, we utilize the [microsoft/phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8) as a reference phixtral 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 phixtral 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 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 phixtral 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 phixtral model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'mlabonne/phixtral-4x2_8'`.
+- `--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
+#### [mlabonne/phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+Question:What is AI?
+
+Answer:
+-------------------- Output --------------------
+Question:What is AI?
+
+Answer: AI, or artificial intelligence, is the simulation of human intelligence in machines that are programmed to think and learn like humans.
+```
\ No newline at end of file
diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral/generate.py b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral/generate.py
new file mode 100644
index 00000000000..276fa09e334
--- /dev/null
+++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phixtral/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 transformers import AutoTokenizer, GenerationConfig
+from bigdl.llm import optimize_model
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to  # TODO: https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py
+PHI1_5_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 phixtral model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="mlabonne/phixtral-4x2_8",
+                        help='The huggingface repo id for the phi 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
+    from bigdl.llm.transformers import AutoModelForCausalLM
+    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 = PHI1_5_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 phixtral 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/phixtral/README.md b/python/llm/example/CPU/PyTorch-Models/Model/phixtral/README.md
new file mode 100644
index 00000000000..f2bf2412190
--- /dev/null
+++ b/python/llm/example/CPU/PyTorch-Models/Model/phixtral/README.md
@@ -0,0 +1,64 @@
+# Phixtral
+In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate Qwen-VL models. For illustration purposes, we utilize the [mlabonne/phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8) as a reference Phixtral 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 phixtral 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 phixtral model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'mlabonne/phixtral-4x2_8'`.
+- `--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
+#### [mlabonne/phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+Question:What is AI?
+
+Answer:
+-------------------- Output --------------------
+Question:What is AI?
+
+Answer: AI, or artificial intelligence, is the simulation of human intelligence in machines that are programmed to think and learn like humans.
+```
\ No newline at end of file
diff --git a/python/llm/example/CPU/PyTorch-Models/Model/phixtral/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/phixtral/generate.py
new file mode 100644
index 00000000000..e66863adc67
--- /dev/null
+++ b/python/llm/example/CPU/PyTorch-Models/Model/phixtral/generate.py
@@ -0,0 +1,66 @@
+#
+# 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 transformers import AutoTokenizer, GenerationConfig
+from bigdl.llm import optimize_model
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to  # TODO: https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py
+PHI1_5_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 phixtral model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="mlabonne/phixtral-4x2_8",
+                        help='The huggingface repo id for the phi 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
+    
+    from transformers import AutoModelForCausalLM
+    model = AutoModelForCausalLM.from_pretrained(model_path,
+                                                 trust_remote_code=True)
+    model = optimize_model(model)
+
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path,
+                                              trust_remote_code=True)
+    
+    # Generate predicted tokens
+    with torch.inference_mode():
+        prompt = PHI1_5_PROMPT_FORMAT.format(prompt=args.prompt)
+        input_ids = tokenizer.encode(prompt, return_tensors="pt")
+        st = time.time()
+
+        # Note that phixtral 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/GPU/HF-Transformers-AutoModels/Model/phixtral/README.md b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral/README.md
new file mode 100644
index 00000000000..b271f5bada5
--- /dev/null
+++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral/README.md
@@ -0,0 +1,119 @@
+# Phixtral
+In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on phixtral models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [mlabonne/phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8) as a reference phixtral 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 InternLM 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
+```
+
+#### 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 --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 huggingface repo id for the phi model (e.g. `mlabonne/phixtral-4x2_8`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'mlabonne/phixtral-4x2_8'`.
+- `--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
+#### [mlabonne/phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+Question:What is AI?
+
+Answer:
+-------------------- Output --------------------
+Question:What is AI?
+
+Answer: AI, or artificial intelligence, is the simulation of human intelligence in machines that are programmed to think and learn like humans. It involves the development of computer systems that
+```
diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral/generate.py b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral/generate.py
new file mode 100644
index 00000000000..e806ba5431a
--- /dev/null
+++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral/generate.py
@@ -0,0 +1,80 @@
+#
+# 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 transformers import AutoTokenizer, GenerationConfig
+import intel_extension_for_pytorch as ipex
+
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to  # TODO: https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py
+PHI1_5_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 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="mlabonne/phixtral-4x2_8",
+                        help='The huggingface repo id for the phixtral 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.
+    from bigdl.llm.transformers import AutoModel,AutoModelForCausalLM
+    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
+    # for phi-moe
+    with torch.inference_mode():
+        prompt = PHI1_5_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 without profiling
+        st = time.time()
+        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/phixtral/README.md b/python/llm/example/GPU/PyTorch-Models/Model/phixtral/README.md
new file mode 100644
index 00000000000..29d0c8869a4
--- /dev/null
+++ b/python/llm/example/GPU/PyTorch-Models/Model/phixtral/README.md
@@ -0,0 +1,123 @@
+# phixtral
+In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate phi-1_5 models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [mlabonne/phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8) as a reference phixtral 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 phixtral 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 phixtral 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
+pip install einops # additional package required for phixtral to conduct generation
+```
+
+### 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 phixtral model (e.g. `mlabonne/phixtral-4x2_8`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'mlabonne/phixtral-4x2_8'`.
+- `--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
+#### [mlabonne/phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8)
+
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+Question:What is AI?
+
+Answer:
+-------------------- Output --------------------
+Question:What is AI?
+
+Answer: AI, or artificial intelligence, is the simulation of human intelligence in machines that are programmed to think and learn like humans. It involves the development of computer systems that
+```
\ No newline at end of file
diff --git a/python/llm/example/GPU/PyTorch-Models/Model/phixtral/generate.py b/python/llm/example/GPU/PyTorch-Models/Model/phixtral/generate.py
new file mode 100644
index 00000000000..991377fd2d5
--- /dev/null
+++ b/python/llm/example/GPU/PyTorch-Models/Model/phixtral/generate.py
@@ -0,0 +1,80 @@
+#
+# 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 transformers import AutoTokenizer, GenerationConfig
+import intel_extension_for_pytorch as ipex
+from bigdl.llm import optimize_model
+
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to  # TODO: https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py
+PHI1_5_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 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="mlabonne/phixtral-4x2_8",
+                        help='The huggingface repo id for the phixtral 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 huggingface model with optimize_model in BigDL
+    from transformers import AutoModelForCausalLM
+    model = AutoModelForCausalLM.from_pretrained(model_path,
+                                                 trust_remote_code=True)
+    model = optimize_model(model)
+
+    model = model.to('xpu')
+
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path,
+                                              trust_remote_code=True)
+    
+    # Generate predicted tokens
+    # for phi-moe
+    with torch.inference_mode():
+        prompt = PHI1_5_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 without profiling
+        st = time.time()
+        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/src/bigdl/llm/transformers/convert.py b/python/llm/src/bigdl/llm/transformers/convert.py
index 38910d3de99..518de262ae8 100644
--- a/python/llm/src/bigdl/llm/transformers/convert.py
+++ b/python/llm/src/bigdl/llm/transformers/convert.py
@@ -912,6 +912,17 @@ def _optimize_post(model, lightweight_bmm=False):
         convert_forward(model,
                         module.MixtralBLockSparseTop2MLP,
                         mixtral_mlp_forward)
+    elif model.config.model_type == "phi-msft":
+        modeling_module_name = model.__class__.__module__
+        module = importlib.import_module(modeling_module_name)
+        from bigdl.llm.transformers.models.phixtral import phixtral_moeblock_forward, \
+            phixtral_mlp_forward
+        convert_forward(model,
+                        module.MoE,
+                        phixtral_moeblock_forward)
+        convert_forward(model,
+                        module.MLP,
+                        phixtral_mlp_forward)
     elif model.config.model_type == "mistral":
         if model.config.architectures is not None and \
                 model.config.architectures[0] == "MixtralForCausalLM":
diff --git a/python/llm/src/bigdl/llm/transformers/models/phixtral.py b/python/llm/src/bigdl/llm/transformers/models/phixtral.py
new file mode 100644
index 00000000000..272ab53b5c6
--- /dev/null
+++ b/python/llm/src/bigdl/llm/transformers/models/phixtral.py
@@ -0,0 +1,144 @@
+#
+# 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.
+#
+# Some parts of this file is adapted from
+# https://github.com/huggingface/transformers/blob/main/src/transformers/models/mixtral/modeling_mixtral.py
+
+# coding=utf-8
+# Copyright 2023 Mistral AI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+
+""" PyTorch Phixtral model."""
+import math
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from bigdl.llm.ggml.quantize import ggml_tensor_qtype
+from bigdl.llm.utils.common import invalidInputError
+from bigdl.llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
+from bigdl.llm.transformers.models.utils import apply_rotary_pos_emb,\
+    apply_rotary_pos_emb_no_cache_xpu, is_enough_kv_cache_room_4_36
+from bigdl.llm.transformers.models.mistral import should_use_fuse_rope, use_decoding_fast_path
+from bigdl.llm.transformers.models.utils import use_flash_attention
+from bigdl.llm.transformers.models.utils import mlp_fusion_check
+
+
+KV_CACHE_ALLOC_BLOCK_LENGTH = 256
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep).
+    The hidden states go from (batch, num_key_value_heads, seqlen, head_dim)
+    to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads,
+                                                           n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def phixtral_moeblock_forward(self, hidden_states: torch.Tensor):
+    batch_size, sequence_length, hidden_dim = hidden_states.shape
+    hidden_states = hidden_states.view(-1, hidden_dim)
+    bs = hidden_states.shape[0]
+    # router_logits: (batch * sequence_length, n_experts)
+    router_logits = self.gate(hidden_states)
+
+    num_local_experts = len(self.mlp)
+
+    routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+    top_k = self.num_experts_per_tok
+    routing_weights, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+    routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+    # we cast back to the input dtype
+    routing_weights = routing_weights.to(hidden_states.dtype)
+
+    if bs > 1:
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim),
+            dtype=hidden_states.dtype,
+            device=hidden_states.device
+        )
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts,
+                                                  num_classes=num_local_experts).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(num_local_experts):
+            expert_layer = self.mlp[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            if top_x.shape[0] == 0:
+                continue
+
+            # in torch it is faster to index using lists than torch tensors
+            top_x_list = top_x.tolist()
+            idx_list = idx.tolist()
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state)
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+    else:
+        selected_experts = selected_experts[0].cpu().tolist()
+        for idx in range(top_k):
+            exp_id = selected_experts[idx]
+            expert_layer = self.mlp[exp_id]
+            weight = routing_weights[:, idx]
+            if idx == 0:
+                final_hidden_states = expert_layer(hidden_states)
+            else:
+                final_hidden_states = final_hidden_states + expert_layer(hidden_states)
+
+    final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+    return final_hidden_states
+
+
+def phixtral_mlp_forward(
+    self,
+    x: torch.Tensor,
+) -> torch.Tensor:
+    hidden_states = self.fc1(x)
+    hidden_states = self.act(hidden_states)
+    hidden_states = self.fc2(hidden_states)
+
+    return hidden_states