ipex-llm/python/llm/example/GPU/HF-Transformers-AutoModels/Save-Load/README.md

Save/Load Low-Bit Models with IPEX-LLM Optimizations

In this directory, you will find example on how you could save/load models with IPEX-LLM INT4 optimizations on Llama2 models on Intel GPUs. For illustration purposes, we utilize the meta-llama/Llama-2-7b-chat-hf and meta-llama/Llama-2-13b-chat-hf as reference Llama2 models.

0. Requirements

To run this example with IPEX-LLM, we have some recommended requirements for your machine, please refer to here for more information.

Example: Save/Load Model in Low-Bit Optimization

In the example generate.py, we show a basic use case of saving/loading model in low-bit optimizations to predict the next N tokens using generate() API. Also, saving and loading operations are platform-independent, so you could run it on different platforms.

1. Install

1.1 Installation on Linux

We suggest using conda to manage environment:

conda create -n llm 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/

1.2 Installation on Windows

We suggest using conda to manage environment:

conda create -n llm python=3.11 libuv
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/

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.

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

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

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

export SYCL_CACHE_PERSISTENT=1
export BIGDL_LLM_XMX_DISABLED=1

3.2 Configurations for Windows

For Intel iGPU

set SYCL_CACHE_PERSISTENT=1
set BIGDL_LLM_XMX_DISABLED=1

For Intel Arc™ A-Series Graphics

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

If you want to save the optimized low-bit model, run:

python ./generate.py --save-path path/to/save/model

If you want to load the optimized low-bit model, run:

python ./generate.py --load-path path/to/load/model

In the example, several arguments can be passed to satisfy your requirements:

• --repo-id-or-model-path REPO_ID_OR_MODEL_PATH: argument defining the huggingface repo id for the Llama2 model to be downloaded, or the path to the ModelScope checkpoint folder. It is default to be 'meta-llama/Llama-2-7b-chat-hf'.
• --save-path: argument defining the path to save the low-bit model. Then you can load the low-bit directly.
• --load-path: argument defining the path to load low-bit model.
• --prompt PROMPT: argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be 'What is AI?'.
• --n-predict N_PREDICT: argument defining the max number of tokens to predict. It is default to be 32.

Sample Output

meta-llama/Llama-2-7b-chat-hf

Inference time: xxxx s
-------------------- Output --------------------
### HUMAN:
What is AI?

### RESPONSE:

AI is a term used to describe the development of computer systems that can perform tasks that typically require human intelligence, such as understanding natural language, recognizing images