From 245c7348bc8654d15dda67c24e4f6843625c93a7 Mon Sep 17 00:00:00 2001
From: hxsz1997 <45651968+hxsz1997@users.noreply.github.com>
Date: Tue, 7 May 2024 13:35:42 +0800
Subject: [PATCH] Add codegemma example (#10884)
* add codegemma example in GPU/HF-Transformers-AutoModels/
* add README of codegemma example in GPU/HF-Transformers-AutoModels/
* add codegemma example in GPU/PyTorch-Models/
* add readme of codegemma example in GPU/PyTorch-Models/
* add codegemma example in CPU/HF-Transformers-AutoModels/
* add readme of codegemma example in CPU/HF-Transformers-AutoModels/
* add codegemma example in CPU/PyTorch-Models/
* add readme of codegemma example in CPU/PyTorch-Models/
* fix typos
* fix filename typo
* add codegemma in tables
* add comments of lm_head
* remove comments of use_cache
---
README.md | 1 +
docs/readthedocs/source/index.rst | 7 +
.../Model/codegemma/README.md | 75 +++++++++
.../Model/codegemma/generate.py | 71 +++++++++
.../PyTorch-Models/Model/codegemma/README.md | 73 +++++++++
.../Model/codegemma/generate.py | 68 ++++++++
.../Model/codegemma/README.md | 144 +++++++++++++++++
.../Model/codegemma/generate.py | 80 ++++++++++
.../PyTorch-Models/Model/codegemma/README.md | 145 ++++++++++++++++++
.../Model/codegemma/generate.py | 82 ++++++++++
10 files changed, 746 insertions(+)
create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma/README.md
create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma/generate.py
create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/codegemma/README.md
create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/codegemma/generate.py
create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/codegemma/README.md
create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/codegemma/generate.py
create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/codegemma/README.md
create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/codegemma/generate.py
diff --git a/README.md b/README.md
index 61eef911..07bc914f 100644
--- a/README.md
+++ b/README.md
@@ -183,6 +183,7 @@ Over 50 models have been optimized/verified on `ipex-llm`, including *LLaMA/LLaM
| DeciLM-7B | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/deciLM-7b) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/deciLM-7b) |
| Deepseek | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/deepseek) |
| StableLM | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/stablelm) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/stablelm) |
+| CodeGemma | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/codegemma) |
## Get Support
- Please report a bug or raise a feature request by opening a [Github Issue](https://github.com/intel-analytics/ipex-llm/issues)
diff --git a/docs/readthedocs/source/index.rst b/docs/readthedocs/source/index.rst
index 5a307f3f..2a261ce4 100644
--- a/docs/readthedocs/source/index.rst
+++ b/docs/readthedocs/source/index.rst
@@ -580,6 +580,13 @@ Verified Models
link |
+
+ | CodeGemma |
+
+ link |
+
+ link |
+
diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma/README.md
new file mode 100644
index 00000000..a959d278
--- /dev/null
+++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma/README.md
@@ -0,0 +1,75 @@
+# CodeGemma
+In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on CodeGemma models. For illustration purposes, we utilize the [google/codegemma-7b-it](https://huggingface.co/google/codegemma-7b-it) as reference CodeGemma models.
+
+## 0. Requirements
+To run these examples with IPEX-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 CodeGemma model to predict the next N tokens using `generate()` API, with IPEX-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 IPEX-LLM:
+```bash
+conda create -n llm python=3.11 # recommend to use Python 3.11
+conda activate llm
+
+# install ipex-llm with 'all' option
+pip install ipex-llm[all]
+
+# According to CodeGemma's requirement, please make sure you are using a stable version of Transformers, 4.38.1 or newer.
+pip install transformers==4.38.1
+```
+
+### 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 huggingface repo id for the CodeGemma model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'google/codegemma-7b-it'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'Write a hello world program'`.
+- `--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, IPEX-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 CodeLlama 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 IPEX-LLM env variables
+source ipex-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
+#### [google/codegemma-7b-it](https://huggingface.co/google/codegemma-7b-it)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+user
+Write a hello world program
+model
+
+-------------------- Output --------------------
+user
+Write a hello world program
+model
+```python
+print("Hello, world!")
+```
+
+This program will print the message "Hello, world!" to the console.
+```
diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma/generate.py b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma/generate.py
new file mode 100644
index 00000000..8e370f75
--- /dev/null
+++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegemma/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
+
+from ipex_llm.transformers import AutoModelForCausalLM
+from transformers import AutoTokenizer
+
+# The instruction-tuned models use a chat template that must be adhered to for conversational use.
+# see https://huggingface.co/google/codegemma-7b-it#chat-template.
+chat = [
+ { "role": "user", "content": "Write a hello world program" },
+]
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for CodeGemma model')
+ parser.add_argument('--repo-id-or-model-path', type=str, default="google/codegemma-7b-it",
+ help='The huggingface repo id for the CodeGemma to be downloaded'
+ ', or the path to the huggingface checkpoint folder')
+ parser.add_argument('--prompt', type=str, default="Write a hello world program",
+ 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
+ # To fix the issue that the output of codegemma-7b-it is abnormal, skip the 'lm_head' module during optimization
+ model = AutoModelForCausalLM.from_pretrained(model_path,
+ load_in_4bit=True,
+ trust_remote_code=True,
+ use_cache=True,
+ modules_to_not_convert=["lm_head"])
+
+ # Load tokenizer
+ tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+ # Generate predicted tokens
+ with torch.inference_mode():
+ chat[0]['content'] = args.prompt
+ prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+ input_ids = tokenizer.encode(prompt, return_tensors="pt")
+
+ # start inference
+ st = time.time()
+ 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/PyTorch-Models/Model/codegemma/README.md b/python/llm/example/CPU/PyTorch-Models/Model/codegemma/README.md
new file mode 100644
index 00000000..22bdabc5
--- /dev/null
+++ b/python/llm/example/CPU/PyTorch-Models/Model/codegemma/README.md
@@ -0,0 +1,73 @@
+# CodeGemma
+In this directory, you will find examples on how you could use IPEX-LLM `optimize_model` API to accelerate CodeGemma models. For illustration purposes, we utilize the [google/codegemma-7b-it](https://huggingface.co/google/codegemma-7b-it) as reference CodeGemma models.
+
+## 0. Requirements
+To run these examples with IPEX-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 CodeGemma model to predict the next N tokens using `generate()` API, with IPEX-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 IPEX-LLM:
+```bash
+conda create -n llm python=3.11 # recommend to use Python 3.11
+conda activate llm
+
+# install ipex-llm with 'all' option
+pip install --pre --upgrade ipex-llm[all]
+
+# According to CodeGemma's requirement, please make sure you are using a stable version of Transformers, 4.38.1 or newer.
+pip install transformers==4.38.1
+```
+
+### 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 --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
+```
+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 IPEX-LLM env variables
+source ipex-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 --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
+```
+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 REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the CodeGemma model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'google/codegemma-7b-it'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'Write a hello world program'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+
+#### 2.4 Sample Output
+#### [google/codegemma-7b-it](https://huggingface.co/google/codegemma-7b-it)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+user
+Write a hello world program
+model
+
+-------------------- Output --------------------
+user
+Write a hello world program
+model
+```python
+print("Hello, world!")
+```
+
+This program will print the message "Hello, world!" to the console.
+```
diff --git a/python/llm/example/CPU/PyTorch-Models/Model/codegemma/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/codegemma/generate.py
new file mode 100644
index 00000000..e64d842b
--- /dev/null
+++ b/python/llm/example/CPU/PyTorch-Models/Model/codegemma/generate.py
@@ -0,0 +1,68 @@
+#
+# 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 AutoModelForCausalLM, AutoTokenizer
+from ipex_llm import optimize_model
+
+# The instruction-tuned models use a chat template that must be adhered to for conversational use.
+# see https://huggingface.co/google/codegemma-7b-it#chat-template.
+chat = [
+ { "role": "user", "content": "Write a hello world program" },
+]
+
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for CodeGemma model')
+ parser.add_argument('--repo-id-or-model-path', type=str, default="google/codegemma-7b-it",
+ help='The huggingface repo id for the CodeGemma model to be downloaded'
+ ', or the path to the huggingface checkpoint folder')
+ parser.add_argument('--prompt', type=str, default="Write a hello world program",
+ 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 IPEX-LLM optimization on model
+ # To fix the issue that the output of codegemma-7b-it is abnormal, skip the 'lm_head' module during optimization
+ model = optimize_model(model, modules_to_not_convert=["lm_head"])
+
+ # Load tokenizer
+ tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+ # Generate predicted tokens
+ with torch.inference_mode():
+ chat[0]['content'] = args.prompt
+ prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+ input_ids = tokenizer.encode(prompt, return_tensors="pt")
+ st = time.time()
+ 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/GPU/HF-Transformers-AutoModels/Model/codegemma/README.md b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/codegemma/README.md
new file mode 100644
index 00000000..606284da
--- /dev/null
+++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/codegemma/README.md
@@ -0,0 +1,144 @@
+# CodeGemma
+In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on CodeGemma models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [google/codegemma-7b-it](https://huggingface.co/google/codegemma-7b-it) as reference CodeGemma models.
+
+## 0. Requirements
+To run these examples with IPEX-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../../../README.md#requirements) for more information.
+
+**Important: According to CodeGemma's requirement, please make sure you have installed `transformers==4.38.1` to run the example.**
+
+## Example: Predict Tokens using `generate()` API
+In the example [generate.py](./generate.py), we show a basic use case for a CodeGemma model to predict the next N tokens using `generate()` API, with IPEX-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 IPEX-LLM:
+```bash
+conda create -n llm python=3.11 # recommend to use Python 3.11
+conda activate llm
+# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
+pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+
+# According to CodeGemma's requirement, please make sure you are using a stable version of Transformers, 4.38.1 or newer.
+pip install transformers==4.38.1
+```
+
+#### 1.2 Installation on Windows
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.11 libuv
+conda activate llm
+# below command will use pip to install the Intel oneAPI Base Toolkit 2024.0
+pip install dpcpp-cpp-rt==2024.0.2 mkl-dpcpp==2024.0.0 onednn==2024.0.0
+
+# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
+pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+
+# According to CodeGemma's requirement, please make sure you are using a stable version of Transformers, 4.38.1 or newer.
+pip install transformers==4.38.1
+```
+
+### 2. Configures OneAPI environment variables for Linux
+
+> [!NOTE]
+> Skip this step if you are running on Windows.
+
+This is a required step on Linux for APT or offline installed oneAPI. Skip this step for PIP-installed oneAPI.
+
+```bash
+source /opt/intel/oneapi/setvars.sh
+```
+
+### 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
+export SYCL_CACHE_PERSISTENT=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 SYCL_CACHE_PERSISTENT=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`.
+
+
+
+
+For Intel iGPU
+
+```bash
+export SYCL_CACHE_PERSISTENT=1
+export BIGDL_LLM_XMX_DISABLED=1
+```
+
+
+
+#### 3.2 Configurations for Windows
+
+
+For Intel iGPU
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+set BIGDL_LLM_XMX_DISABLED=1
+```
+
+
+
+
+
+For Intel Arc™ A-Series Graphics
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+```
+
+
+
+> [!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 CodeGemma model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'google/codegemma-7b-it'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'Write a hello world program'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+
+##### Sample Output
+##### [google/codegemma-7b-it](https://huggingface.co/google/codegemma-7b-it)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+user
+Write a hello world program
+model
+
+-------------------- Output --------------------
+user
+Write a hello world program
+model
+```python
+print("Hello, world!")
+```
+
+This program will print the message "Hello, world!" to the console.
+```
diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/codegemma/generate.py b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/codegemma/generate.py
new file mode 100644
index 00000000..9add373b
--- /dev/null
+++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/codegemma/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
+
+from ipex_llm.transformers import AutoModelForCausalLM
+from transformers import AutoTokenizer
+
+# The instruction-tuned models use a chat template that must be adhered to for conversational use.
+# see https://huggingface.co/google/codegemma-7b-it#chat-template.
+chat = [
+ { "role": "user", "content": "Write a hello world program" },
+]
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for CodeGemma model')
+ parser.add_argument('--repo-id-or-model-path', type=str, default="google/codegemma-7b-it",
+ help='The huggingface repo id for the CodeGemma to be downloaded'
+ ', or the path to the huggingface checkpoint folder')
+ parser.add_argument('--prompt', type=str, default="Write a hello world program",
+ 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
+ # To fix the issue that the output of codegemma-7b-it is abnormal, skip the 'lm_head' module during optimization
+ # 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,
+ optimize_model=True,
+ trust_remote_code=True,
+ use_cache=True,
+ modules_to_not_convert=["lm_head"])
+ model = model.to('xpu')
+
+ # Load tokenizer
+ tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+ # Generate predicted tokens
+ with torch.inference_mode():
+ chat[0]['content'] = args.prompt
+ prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+ input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
+ # ipex_llm 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()
+ output = model.generate(input_ids,
+ max_new_tokens=args.n_predict)
+ 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/codegemma/README.md b/python/llm/example/GPU/PyTorch-Models/Model/codegemma/README.md
new file mode 100644
index 00000000..fa7363b3
--- /dev/null
+++ b/python/llm/example/GPU/PyTorch-Models/Model/codegemma/README.md
@@ -0,0 +1,145 @@
+# CodeGemma
+In this directory, you will find examples on how you could use IPEX-LLM `optimize_model` API to accelerate CodeGemma models. For illustration purposes, we utilize the [google/codegemma-7b-it](https://huggingface.co/google/codegemma-7b-it) as reference CodeGemma models.
+
+## 0. Requirements
+To run these examples with IPEX-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../../../README.md#requirements) for more information.
+
+**Important: According to CodeGemma's requirement, please make sure you have installed `transformers==4.38.1` to run the example.**
+
+## Example: Predict Tokens using `generate()` API
+In the example [generate.py](./generate.py), we show a basic use case for a CodeGemma model to predict the next N tokens using `generate()` API, with IPEX-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 IPEX-LLM:
+```bash
+conda create -n llm python=3.11 # recommend to use Python 3.11
+conda activate llm
+# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
+pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+
+# According to CodeGemma's requirement, please make sure you are using a stable version of Transformers, 4.38.1 or newer.
+pip install transformers==4.38.1
+```
+
+#### 1.2 Installation on Windows
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.11 libuv
+conda activate llm
+# below command will use pip to install the Intel oneAPI Base Toolkit 2024.0
+pip install dpcpp-cpp-rt==2024.0.2 mkl-dpcpp==2024.0.0 onednn==2024.0.0
+
+# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
+pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+
+# According to CodeGemma's requirement, please make sure you are using a stable version of Transformers, 4.38.1 or newer.
+pip install transformers==4.38.1
+```
+
+### 2. Configures OneAPI environment variables for Linux
+
+> [!NOTE]
+> Skip this step if you are running on Windows.
+
+This is a required step on Linux for APT or offline installed oneAPI. Skip this step for PIP-installed oneAPI.
+
+```bash
+source /opt/intel/oneapi/setvars.sh
+```
+
+### 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
+export SYCL_CACHE_PERSISTENT=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 SYCL_CACHE_PERSISTENT=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`.
+
+
+
+
+For Intel iGPU
+
+```bash
+export SYCL_CACHE_PERSISTENT=1
+export BIGDL_LLM_XMX_DISABLED=1
+```
+
+
+
+#### 3.2 Configurations for Windows
+
+
+For Intel iGPU
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+set BIGDL_LLM_XMX_DISABLED=1
+```
+
+
+
+
+
+For Intel Arc™ A-Series Graphics
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+```
+
+
+
+> [!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
+
+```bash
+python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
+```
+
+In the example, several arguments can be passed to satisfy your requirements:
+
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the CodeGemma model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'google/codegemma-7b-it'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `Write a hello world program'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+
+#### 4.1 Sample Output
+#### [google/codegemma-7b-it](https://huggingface.co/google/codegemma-7b-it)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+user
+Write a hello world program
+model
+
+-------------------- Output --------------------
+user
+Write a hello world program
+model
+```python
+print("Hello, world!")
+```
+
+This program will print the message "Hello, world!" to the console.
+```
diff --git a/python/llm/example/GPU/PyTorch-Models/Model/codegemma/generate.py b/python/llm/example/GPU/PyTorch-Models/Model/codegemma/generate.py
new file mode 100644
index 00000000..ce3ec819
--- /dev/null
+++ b/python/llm/example/GPU/PyTorch-Models/Model/codegemma/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
+
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from ipex_llm import optimize_model
+
+# The instruction-tuned models use a chat template that must be adhered to for conversational use.
+# see https://huggingface.co/google/codegemma-7b-it#chat-template.
+chat = [
+ { "role": "user", "content": "Write a hello world program" },
+]
+
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for CodeGemma model')
+ parser.add_argument('--repo-id-or-model-path', type=str, default="google/codegemma-7b-it",
+ help='The huggingface repo id for the CodeGemma model to be downloaded'
+ ', or the path to the huggingface checkpoint folder')
+ parser.add_argument('--prompt', type=str, default="Write a hello world program",
+ 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,
+ torch_dtype='auto',
+ low_cpu_mem_usage=True)
+
+ # With only one line to enable IPEX-LLM optimization on model
+ # To fix the issue that the output of codegemma-7b-it is abnormal, skip the 'lm_head' module during optimization
+ # 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, modules_to_not_convert=["lm_head"])
+
+ model = model.to('xpu')
+
+ # Load tokenizer
+ tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+ # Generate predicted tokens
+ with torch.inference_mode():
+ chat[0]['content'] = args.prompt
+ prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+ input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
+ # ipex_llm 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()
+ output = model.generate(input_ids,
+ max_new_tokens=args.n_predict)
+ 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)