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 <yuwen.hu@intel.com>
intel · Jan 13, 2025 · 350fae2 · 350fae2
1 parent a1da790
commit 350fae2
Show file tree

Hide file tree

Showing 6 changed files with 287 additions and 2 deletions.
diff --git 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
@@ -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
@@ -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
+<details>
+
+<summary>For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series</summary>
+
+```bash
+export USE_XETLA=OFF
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+export SYCL_CACHE_PERSISTENT=1
+```
+
+</details>
+
+<details>
+
+<summary>For Intel Data Center GPU Max Series</summary>
+
+```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`.
+</details>
+
+<details>
+
+<summary>For Intel iGPU</summary>
+
+```bash
+export SYCL_CACHE_PERSISTENT=1
+```
+
+</details>
+
+#### 3.2 Configurations for Windows
+<details>
+
+<summary>For Intel iGPU and Intel Arc™ A-Series Graphics</summary>
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+```
+
+</details>
+
+
+> [!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)):
+
+<a href="http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg"><img width=400px src="http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg" ></a>
diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/qwen2-vl/generate.py 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
@@ -21,5 +21,10 @@
         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
@@ -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