In this directory, you will find examples on how you could apply IPEX-LLM FP8 optimizations on GLM-4V models on Intel GPUs. For illustration purposes, we utilize the THUDM/glm-4v-9b (or ZhipuAI/glm-4v-9b for ModelScope) as a reference GLM-4V model.
To run these examples with IPEX-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to here for more information.
In the example generate.py, we show a basic use case for a GLM-4V model to predict the next N tokens using generate()
API, with IPEX-LLM FP8 optimizations on Intel GPUs.
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/
pip install tiktoken transformers==4.42.4 "trl<0.12.0"
# [optional] only needed if you would like to use ModelScope as model hub
pip install modelscope==1.11.0
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/
pip install tiktoken transformers==4.42.4 "trl<0.12.0"
# [optional] only needed if you would like to use ModelScope as model hub
pip install modelscope==1.11.0
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
For optimal performance, it is recommended to set several environment variables. Please check out the suggestions based on your device.
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 byconda install -c conda-forge -y gperftools=2.10
.
For Intel iGPU
export SYCL_CACHE_PERSISTENT=1
For Intel iGPU and 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.
# for Hugging Face model hub
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT --image-url-or-path IMAGE_URL_OR_PATH
# for ModelScope model hub
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT --image-url-or-path IMAGE_URL_OR_PATH --modelscope
Arguments info:
--repo-id-or-model-path REPO_ID_OR_MODEL_PATH
: argument defining the Hugging Face or ModelScope repo id for the GLM-4V model (e.g.THUDM/glm-4v-9b
) to be downloaded, or the path to the checkpoint folder. It is default to be'THUDM/glm-4v-9b'
for Hugging Face or'ZhipuAI/glm-4v-9b'
for ModelScope.--image-url-or-path IMAGE_URL_OR_PATH
: argument defining the image to be infered. It is default to be'http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg'
.--prompt PROMPT
: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be'What is in the image?'
.--n-predict N_PREDICT
: argument defining the max number of tokens to predict. It is default to be32
.--modelscope
: using ModelScope as model hub instead of Hugging Face.
Inference time: xxxx s
-------------------- Prompt --------------------
What is in the image?
-------------------- Output --------------------
The image shows a young child holding up a small white teddy bear dressed in a pink
The sample input image is (which is fetched from COCO dataset):