From 0cff47965039460540366917799e9271a361421f Mon Sep 17 00:00:00 2001 From: ch1y0q Date: Fri, 20 Sep 2024 15:34:28 +0800 Subject: [PATCH 1/4] add internvl2 example --- .../HuggingFace/Multimodal/internvl2/chat.py | 99 +++++++++++++ .../Multimodal/internvl2/readme.md | 136 ++++++++++++++++++ 2 files changed, 235 insertions(+) create mode 100644 python/llm/example/GPU/HuggingFace/Multimodal/internvl2/chat.py create mode 100644 python/llm/example/GPU/HuggingFace/Multimodal/internvl2/readme.md diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/chat.py b/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/chat.py new file mode 100644 index 00000000000..31e4984ec89 --- /dev/null +++ b/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/chat.py @@ -0,0 +1,99 @@ +# +# 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 os +import time +import argparse +import requests +import torch +from PIL import Image +from ipex_llm.transformers import AutoModelForCausalLM +from transformers import AutoTokenizer, CLIPImageProcessor + + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `chat()` API for OpenGVLab/InternVL2-4B model') + parser.add_argument('--repo-id-or-model-path', type=str, default="OpenGVLab/InternVL2-4B", + help='The huggingface repo id for the OpenGVLab/InternVL2-4B model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--image-url-or-path', type=str, + default='https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg', + help='The URL or path to the image to infer') + parser.add_argument('--prompt', type=str, default="What is in the image?", + help='Prompt to infer') + + args = parser.parse_args() + model_path = args.repo_id_or_model_path + image_path = args.image_url_or_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, trust_remote_code=True, + load_in_low_bit="sym_int4", + modules_to_not_convert=["vision_model"]) + model = model.half().to('xpu') + tokenizer = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + model.eval() + + query = args.prompt + image_processor = CLIPImageProcessor.from_pretrained(model_path) + + if os.path.exists(image_path): + image = Image.open(image_path).convert('RGB') + else: + image = Image.open(requests.get(image_path, stream=True).raw).convert('RGB') + + pixel_values = image_processor(images=[image], return_tensors='pt').pixel_values + pixel_values = pixel_values.to('xpu') + + question = "" + query + + generation_config = { + "max_new_tokens": 64, + "do_sample": False, + } + + # ipex_llm model needs a warmup, then inference time can be accurate + model.chat( + pixel_values=None, + question=question, + generation_config=generation_config, + tokenizer=tokenizer, + ) + + + st = time.time() + res = model.chat( + tokenizer=tokenizer, + pixel_values=pixel_values, + question=question, + generation_config=generation_config, + history=[] + ) + torch.xpu.synchronize() + end = time.time() + + print(f'Inference time: {end-st} s') + print('-'*20, 'Input Image', '-'*20) + print(image_path) + print('-'*20, 'Input Prompt', '-'*20) + print(args.prompt) + print('-'*20, 'Chat Output', '-'*20) + print(res) diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/readme.md b/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/readme.md new file mode 100644 index 00000000000..53f46b234e4 --- /dev/null +++ b/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/readme.md @@ -0,0 +1,136 @@ +# InternVL2 +In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on InternVL2 model on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [OpenGVLab/InternVL2-4B](https://huggingface.co/OpenGVLab/InternVL2-4B) as a reference InternVL2 model. + +## 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. + +## Example: Predict Tokens using `chat()` API +In the example [chat.py](./chat.py), we show a basic use case for an InternVL2-4B model to predict the next N tokens using `chat()` API, with IPEX-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=3.11 +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 einops timm + +``` + +#### 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 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 einops timm + +``` + +### 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 + +- chat with specified prompt: + ``` + python ./chat.py --prompt 'What is in the image?' + ``` + +Arguments info: +- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the InternVL2 (e.g. `OpenGVLab/InternVL2-4B`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'OpenGVLab/InternVL2-4B'`. +- `--image-url-or-path IMAGE_URL_OR_PATH`: argument defining the image to be infered. It is default to be `'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg'`. +- `--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?'`. + +#### Sample Output + +#### [OpenGVLab/InternVL2-4B](https://huggingface.co/OpenGVLab/InternVL2-4B) + +```log +-------------------- Input Image -------------------- +https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg +-------------------- Input Prompt -------------------- +What is in the image? +-------------------- Chat Output -------------------- +The image shows a tiger lying on the grass. +``` + +The sample input image is: + + From 496cc049ac10207598e241a74295721931041b87 Mon Sep 17 00:00:00 2001 From: ch1y0q Date: Fri, 20 Sep 2024 15:55:48 +0800 Subject: [PATCH 2/4] add to README.md --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 3d059d6bb96..f7a8390ab67 100644 --- a/README.md +++ b/README.md @@ -272,6 +272,7 @@ Over 50 models have been optimized/verified on `ipex-llm`, including *LLaMA/LLaM | Baichuan | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/baichuan) | [link](python/llm/example/GPU/HuggingFace/LLM/baichuan) | | Baichuan2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/baichuan2) | [link](python/llm/example/GPU/HuggingFace/LLM/baichuan2) | | InternLM | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/internlm) | [link](python/llm/example/GPU/HuggingFace/LLM/internlm) | +| InternVL2 | | [link](python/llm/example/GPU/HuggingFace/Multimodal/internvl2) | | Qwen | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen) | [link](python/llm/example/GPU/HuggingFace/LLM/qwen) | | Qwen1.5 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen1.5) | [link](python/llm/example/GPU/HuggingFace/LLM/qwen1.5) | | Qwen2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen2) | [link](python/llm/example/GPU/HuggingFace/LLM/qwen2) | From dc627d2cadc8f4f5c9d358e645278abf4eff25b0 Mon Sep 17 00:00:00 2001 From: ch1y0q Date: Fri, 20 Sep 2024 16:14:50 +0800 Subject: [PATCH 3/4] update --- .../HuggingFace/Multimodal/internvl2/chat.py | 44 ++++++++++--------- .../Multimodal/internvl2/readme.md | 1 + 2 files changed, 24 insertions(+), 21 deletions(-) diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/chat.py b/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/chat.py index 31e4984ec89..22517d1bdd8 100644 --- a/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/chat.py +++ b/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/chat.py @@ -35,18 +35,20 @@ help='The URL or path to the image to infer') parser.add_argument('--prompt', type=str, default="What is in the image?", help='Prompt to infer') + parser.add_argument('--n-predict', type=int, default=64, help='Max tokens to predict') args = parser.parse_args() model_path = args.repo_id_or_model_path image_path = args.image_url_or_path + n_predict = args.n_predict # 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, trust_remote_code=True, - load_in_low_bit="sym_int4", - modules_to_not_convert=["vision_model"]) + load_in_low_bit="sym_int4", + modules_to_not_convert=["vision_model"]) model = model.half().to('xpu') tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) @@ -66,29 +68,29 @@ question = "" + query generation_config = { - "max_new_tokens": 64, + "max_new_tokens": n_predict, "do_sample": False, } - # ipex_llm model needs a warmup, then inference time can be accurate - model.chat( - pixel_values=None, - question=question, - generation_config=generation_config, - tokenizer=tokenizer, - ) + with torch.inference_mode(): + # ipex_llm model needs a warmup, then inference time can be accurate + model.chat( + pixel_values=None, + question=question, + generation_config=generation_config, + tokenizer=tokenizer, + ) - - st = time.time() - res = model.chat( - tokenizer=tokenizer, - pixel_values=pixel_values, - question=question, - generation_config=generation_config, - history=[] - ) - torch.xpu.synchronize() - end = time.time() + st = time.time() + res = model.chat( + tokenizer=tokenizer, + pixel_values=pixel_values, + question=question, + generation_config=generation_config, + history=[] + ) + torch.xpu.synchronize() + end = time.time() print(f'Inference time: {end-st} s') print('-'*20, 'Input Image', '-'*20) diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/readme.md b/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/readme.md index 53f46b234e4..4183d309f0d 100644 --- a/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/readme.md +++ b/python/llm/example/GPU/HuggingFace/Multimodal/internvl2/readme.md @@ -117,6 +117,7 @@ Arguments info: - `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the InternVL2 (e.g. `OpenGVLab/InternVL2-4B`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'OpenGVLab/InternVL2-4B'`. - `--image-url-or-path IMAGE_URL_OR_PATH`: argument defining the image to be infered. It is default to be `'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg'`. - `--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 be `64`. #### Sample Output From b9ddef12731baa860ae99aaa27cbd788a2fdca52 Mon Sep 17 00:00:00 2001 From: ch1y0q Date: Fri, 20 Sep 2024 16:17:49 +0800 Subject: [PATCH 4/4] add link to zh-CN readme --- README.zh-CN.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.zh-CN.md b/README.zh-CN.md index b5365508d26..671358b817c 100644 --- a/README.zh-CN.md +++ b/README.zh-CN.md @@ -276,6 +276,7 @@ See the demo of running [*Text-Generation-WebUI*](https://ipex-llm.readthedocs.i | Baichuan | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/baichuan) | [link](python/llm/example/GPU/HuggingFace/LLM/baichuan) | | Baichuan2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/baichuan2) | [link](python/llm/example/GPU/HuggingFace/LLM/baichuan2) | | InternLM | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/internlm) | [link](python/llm/example/GPU/HuggingFace/LLM/internlm) | +| InternVL2 | | [link](python/llm/example/GPU/HuggingFace/Multimodal/internvl2) | | Qwen | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen) | [link](python/llm/example/GPU/HuggingFace/LLM/qwen) | | Qwen1.5 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen1.5) | [link](python/llm/example/GPU/HuggingFace/LLM/qwen1.5) | | Qwen2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen2) | [link](python/llm/example/GPU/HuggingFace/LLM/qwen2) |