From 350fae285ddaea2c84043d332cbf2a0950459528 Mon Sep 17 00:00:00 2001 From: "Xu, Shuo" <100334393+ATMxsp01@users.noreply.github.com> Date: Mon, 13 Jan 2025 15:42:04 +0800 Subject: [PATCH] Add Qwen2-VL HF GPU example with ModelScope Support (#12606) * Add qwen2-vl example * complete generate.py & readme * improve lint style * update 1-6 * update main readme * Format and other small fixes --------- Co-authored-by: Yuwen Hu --- README.md | 2 +- README.zh-CN.md | 2 +- .../HuggingFace/Multimodal/qwen2-vl/README.md | 149 ++++++++++++++++++ .../Multimodal/qwen2-vl/generate.py | 126 +++++++++++++++ .../llm/src/ipex_llm/transformers/__init__.py | 5 + python/llm/src/ipex_llm/transformers/model.py | 5 + 6 files changed, 287 insertions(+), 2 deletions(-) create mode 100644 python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/README.md create mode 100644 python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/generate.py diff --git a/README.md b/README.md index 9a016e22..5d0f61d2 100644 --- a/README.md +++ b/README.md @@ -292,7 +292,7 @@ Over 70 models have been optimized/verified on `ipex-llm`, including *LLaMA/LLaM | Qwen2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen2) | [link](python/llm/example/GPU/HuggingFace/LLM/qwen2) | [Python link](python/llm/example/NPU/HF-Transformers-AutoModels/LLM), [C++ link](python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) | | Qwen2.5 | | [link](python/llm/example/GPU/HuggingFace/LLM/qwen2.5) | [Python link](python/llm/example/NPU/HF-Transformers-AutoModels/LLM), [C++ link](python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) | | Qwen-VL | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen-vl) | [link](python/llm/example/GPU/HuggingFace/Multimodal/qwen-vl) | -| Qwen2-VL || [link](python/llm/example/GPU/PyTorch-Models/Model/qwen2-vl) | +| Qwen2-VL || [link](python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl) | | Qwen2-Audio | | [link](python/llm/example/GPU/HuggingFace/Multimodal/qwen2-audio) | | Aquila | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/aquila) | [link](python/llm/example/GPU/HuggingFace/LLM/aquila) | | Aquila2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/aquila2) | [link](python/llm/example/GPU/HuggingFace/LLM/aquila2) | diff --git a/README.zh-CN.md b/README.zh-CN.md index 01c477aa..2df483e0 100644 --- a/README.zh-CN.md +++ b/README.zh-CN.md @@ -292,7 +292,7 @@ See the demo of running [*Text-Generation-WebUI*](https://ipex-llm.readthedocs.i | Qwen2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen2) | [link](python/llm/example/GPU/HuggingFace/LLM/qwen2) | [Python link](python/llm/example/NPU/HF-Transformers-AutoModels/LLM), [C++ link](python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) | | Qwen2.5 | | [link](python/llm/example/GPU/HuggingFace/LLM/qwen2.5) | [Python link](python/llm/example/NPU/HF-Transformers-AutoModels/LLM), [C++ link](python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) | | Qwen-VL | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen-vl) | [link](python/llm/example/GPU/HuggingFace/Multimodal/qwen-vl) | -| Qwen2-VL || [link](python/llm/example/GPU/PyTorch-Models/Model/qwen2-vl) | +| Qwen2-VL || [link](python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl) | | Qwen2-Audio | | [link](python/llm/example/GPU/HuggingFace/Multimodal/qwen2-audio) | | Aquila | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/aquila) | [link](python/llm/example/GPU/HuggingFace/LLM/aquila) | | Aquila2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/aquila2) | [link](python/llm/example/GPU/HuggingFace/LLM/aquila2) | diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/README.md b/python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/README.md new file mode 100644 index 00000000..0a0cea48 --- /dev/null +++ b/python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/README.md @@ -0,0 +1,149 @@ +# Qwen2-VL +In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on Qwen2-VL models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [Qwen/Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) (or [Qwen/Qwen2-VL-7B-Instruct](https://www.modelscope.cn/models/Qwen/Qwen2-VL-7B-Instruct) for ModelScope) as a reference Qwen2-VL 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 `generate()` API +In the example [generate.py](./generate.py), we show a basic use case for a Qwen2-VL 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 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 transformers==4.45.0 # install transformers which supports Qwen2-VL +pip install accelerate==0.33.0 +pip install qwen_vl_utils +pip install "trl<0.12.0" + +# [optional] only needed if you would like to use ModelScope as model hub +pip install modelscope[datasets]==1.21.1 +``` + +#### 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 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 transformers==4.45.0 # install transformers which supports Qwen2-VL +pip install accelerate==0.33.0 +pip install qwen_vl_utils +pip install "trl<0.12.0" + +# [optional] only needed if you would like to use ModelScope as model hub +pip install modelscope[datasets]==1.21.1 +``` + +### 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 +``` + +
+ +#### 3.2 Configurations for Windows +
+ +For Intel iGPU and 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 +# 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 Qwen2-VL model (e.g. `Qwen/Qwen2-VL-7B-Instruct`) to be downloaded, or the path to the checkpoint folder. It is default to be `'Qwen/Qwen2-VL-7B-Instruct'`. +- `--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 `'Describe this image.'`. +- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`. +- `--modelscope`: using **ModelScope** as model hub instead of **Hugging Face**. + +#### Sample Output +##### [Qwen/Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) +```log +Inference time: xxxx s +-------------------- Input Image -------------------- +http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg +-------------------- Prompt -------------------- +图片里有什么？ +-------------------- Output -------------------- +图片里有一个小女孩，她穿着粉红色的条纹连衣裙，手里拿着一个白色的毛绒玩具。背景中有一堵石墙和一些 +``` + +```log +Inference time: xxxx s +-------------------- Input Image -------------------- +http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg +-------------------- Prompt -------------------- +What is in the image? +-------------------- Output -------------------- +The image shows a young child holding a small white teddy bear dressed in a pink outfit. The child is standing in front of a stone wall with red flowers +``` + +The sample input image is (which is fetched from [COCO dataset](https://cocodataset.org/#explore?id=264959)): + + diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/generate.py b/python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/generate.py new file mode 100644 index 00000000..39373ef4 --- /dev/null +++ b/python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/generate.py @@ -0,0 +1,126 @@ +# +# 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 ipex_llm.transformers import Qwen2VLForConditionalGeneration +from qwen_vl_utils import process_vision_info + + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using generate() API for Qwen2-VL model') + parser.add_argument('--repo-id-or-model-path', type=str, default="Qwen/Qwen2-VL-7B-Instruct", + help='The huggingface repo id for the Qwen2-VL model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="图片里有什么？", + help='Prompt to infer') + parser.add_argument('--image-url-or-path', type=str, + default='http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg' , + help='The URL or path to the image to infer') + + parser.add_argument('--n-predict', type=int, default=32, + help='Max tokens to predict') + parser.add_argument('--modelscope', action="store_true", default=False, + help="Use models from modelscope") + + args = parser.parse_args() + if args.modelscope: + from modelscope import AutoProcessor + model_hub = 'modelscope' + else: + from transformers import AutoProcessor + model_hub = 'huggingface' + + model_path = args.repo_id_or_model_path + + model = Qwen2VLForConditionalGeneration.from_pretrained(model_path, + load_in_4bit=True, + optimize_model=True, + trust_remote_code=True, + modules_to_not_convert=["vision"], + use_cache=True, + model_hub=model_hub) + + # Use .float() for better output, and use .half() for better speed + model = model.half().to("xpu") + + # The following code for generation is adapted from https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct#quickstart + + # The default range for the number of visual tokens per image in the model is 4-16384. + # You can set min_pixels and max_pixels according to your needs, such as a token count range of 256-1280, + # to balance speed and memory usage. + min_pixels = 256*28*28 + max_pixels = 1280*28*28 + processor = AutoProcessor.from_pretrained(model_path, min_pixels=min_pixels, max_pixels=max_pixels) + + prompt = args.prompt + image_path = args.image_url_or_path + + with torch.inference_mode(): + messages = [ + { + "role": "user", + "content": [ + { + "type": "image", + "image": image_path, + }, + {"type": "text", "text": prompt}, + ], + } + ] + text = processor.apply_chat_template( + messages, tokenize=False, add_generation_prompt=True + ) + image_inputs, video_inputs = process_vision_info(messages) + inputs = processor( + text=[text], + images=image_inputs, + videos=video_inputs, + padding=True, + return_tensors="pt", + ) + inputs = inputs.to('xpu') + + # ipex_llm model needs a warmup, then inference time can be accurate + generated_ids = model.generate( + **inputs, + max_new_tokens=args.n_predict + ) + + st = time.time() + generated_ids = model.generate( + **inputs, + max_new_tokens=args.n_predict + ) + torch.xpu.synchronize() + end = time.time() + generated_ids = generated_ids.cpu() + generated_ids = [ + output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, generated_ids) + ] + + response = processor.batch_decode(generated_ids, skip_special_tokens=True)[0] + print(f'Inference time: {end-st} s') + print('-'*20, 'Input Image', '-'*20) + print(image_path) + print('-'*20, 'Prompt', '-'*20) + print(prompt) + print('-'*20, 'Output', '-'*20) + print(response) diff --git a/python/llm/src/ipex_llm/transformers/__init__.py b/python/llm/src/ipex_llm/transformers/__init__.py index 6904e897..fe5a5bfb 100644 --- a/python/llm/src/ipex_llm/transformers/__init__.py +++ b/python/llm/src/ipex_llm/transformers/__init__.py @@ -21,5 +21,10 @@ from .model import AutoModelForCausalLM, AutoModel, AutoModelForSeq2SeqLM, \ AutoModelForSequenceClassification, AutoModelForMaskedLM, \ AutoModelForNextSentencePrediction, AutoModelForMultipleChoice, \ AutoModelForTokenClassification + +import transformers +if transformers.__version__ >= '4.45.0': + from .model import Qwen2VLForConditionalGeneration + from .modelling_bigdl import * from .pipeline_parallel import init_pipeline_parallel, PPModelWorker diff --git a/python/llm/src/ipex_llm/transformers/model.py b/python/llm/src/ipex_llm/transformers/model.py index 182b1a83..5459056b 100644 --- a/python/llm/src/ipex_llm/transformers/model.py +++ b/python/llm/src/ipex_llm/transformers/model.py @@ -826,3 +826,8 @@ class AutoModelForMultipleChoice(_BaseAutoModelClass): class AutoModelForTokenClassification(_BaseAutoModelClass): HF_Model = transformers.AutoModelForTokenClassification + + +if transformers.__version__ >= '4.45.0': + class Qwen2VLForConditionalGeneration(_BaseAutoModelClass): + HF_Model = transformers.Qwen2VLForConditionalGeneration