LLM: Add qwen-vl gpu example (#9290)
* create qwen-vl gpu example. * add readme. * fix. * change input figure and update outputs. * add qwen-vl pytorch model gpu example. * fix. * add readme.
This commit is contained in:
Cengguang Zhang
GitHub
parent
7e73c354a6
commit
9f3d4676c6
6 changed files with 356 additions and 0 deletions
|
|
@ -149,6 +149,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
|
|||
| Baichuan2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/baichuan2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/baichuan2) |
|
||||
| InternLM | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/internlm) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/internlm) |
|
||||
| Qwen | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/qwen) |
|
||||
| Qwen-VL | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen-vl) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/qwen-vl) |
|
||||
| Aquila | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/aquila) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/aquila) |
|
||||
| MOSS | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/moss) | |
|
||||
| Whisper | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/whisper) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/whisper) |
|
||||
|
|
|
|||
|
|
@ -56,6 +56,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
|
|||
| Baichuan2 | [link](example/CPU/HF-Transformers-AutoModels/Model/baichuan2) | [link](example/GPU/HF-Transformers-AutoModels/Model/baichuan2) |
|
||||
| InternLM | [link](example/CPU/HF-Transformers-AutoModels/Model/internlm) | [link](example/GPU/HF-Transformers-AutoModels/Model/internlm) |
|
||||
| Qwen | [link](example/CPU/HF-Transformers-AutoModels/Model/qwen) | [link](example/GPU/HF-Transformers-AutoModels/Model/qwen) |
|
||||
| Qwen-VL | [link](example/CPU/HF-Transformers-AutoModels/Model/qwen-vl) | [link](example/GPU/HF-Transformers-AutoModels/Model/qwen-vl) |
|
||||
| Aquila | [link](example/CPU/HF-Transformers-AutoModels/Model/aquila) | [link](example/GPU/HF-Transformers-AutoModels/Model/aquila) |
|
||||
| MOSS | [link](example/CPU/HF-Transformers-AutoModels/Model/moss) | |
|
||||
| Whisper | [link](example/CPU/HF-Transformers-AutoModels/Model/whisper) | [link](example/GPU/HF-Transformers-AutoModels/Model/whisper) |
|
||||
|
|
|
|||
|
|
@ -0,0 +1,78 @@
|
|||
# Qwen-VL
|
||||
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Qwen-VL models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [Qwen/Qwen-VL-Chat](https://huggingface.co/Qwen/Qwen-VL-Chat) as a reference Qwen-VL model.
|
||||
|
||||
## Requirements
|
||||
To run these examples with BigDL-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../README.md#recommended-requirements) for more information.
|
||||
|
||||
## Example: Multimodal chat using `chat()` API
|
||||
In the example [chat.py](./chat.py), we show a basic use case for a Qwen-VL model to start a multimodal chat using `chat()` API, with BigDL-LLM INT4 optimizations on Intel GPUs.
|
||||
### 1. Install
|
||||
We suggest using conda to manage the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
|
||||
|
||||
After installing conda, create a Python environment for BigDL-LLM:
|
||||
```bash
|
||||
conda create -n llm python=3.9 # recommend to use Python 3.9
|
||||
conda activate llm
|
||||
# below command will install intel_extension_for_pytorch==2.0.110+xpu as default
|
||||
# you can install specific ipex/torch version for your need
|
||||
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
|
||||
pip install accelerate tiktoken einops transformers_stream_generator==0.0.4 scipy torchvision pillow tensorboard matplotlib # additional package required for Qwen-VL-Chat to conduct generation
|
||||
```
|
||||
|
||||
### 2. Configures OneAPI environment variables
|
||||
```bash
|
||||
source /opt/intel/oneapi/setvars.sh
|
||||
```
|
||||
|
||||
### 3. Run
|
||||
|
||||
For optimal performance on Arc, it is recommended to set several environment variables.
|
||||
|
||||
```bash
|
||||
export USE_XETLA=OFF
|
||||
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
|
||||
```
|
||||
```
|
||||
python ./chat.py
|
||||
```
|
||||
|
||||
Arguments info:
|
||||
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Qwen-VL model (e.g `Qwen/Qwen-VL-Chat`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'Qwen/Qwen-VL-Chat'`.
|
||||
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
|
||||
|
||||
In every session, image and text can be entered into cmd (user can skip the input by type **'Enter'**) ; please type **'exit'** anytime you want to quit the dialouge.
|
||||
|
||||
Every image output will be named as the round of session and placed under the current directory.
|
||||
|
||||
#### Sample Chat
|
||||
#### [Qwen/Qwen-VL-Chat](https://huggingface.co/Qwen/Qwen-VL-Chat)
|
||||
|
||||
```log
|
||||
-------------------- Session 1 --------------------
|
||||
Please input a picture: http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg
|
||||
Please enter the text: 这是什么？
|
||||
---------- Response ----------
|
||||
这是一张图片，展现了一个穿着粉色条纹连衣裙的小女孩，她手持一只穿粉色裙子的小熊。这个场景发生在一个户外环境，有砖块背景墙和花朵。
|
||||
|
||||
-------------------- Session 2 --------------------
|
||||
Please input a picture:
|
||||
Please enter the text: 这个小女孩多大了？
|
||||
---------- Response ----------
|
||||
根据图片中的描述，这个小女孩应该是年龄较小的孩子，但具体年龄难以确定。从她的外表来看，可能是在5岁左右。。
|
||||
|
||||
-------------------- Session 3 --------------------
|
||||
Please input a picture:
|
||||
Please enter the text: 在图中检测框出玩具熊
|
||||
---------- Response ----------
|
||||
<ref>玩具熊</ref><box>(330,267),(603,869)</box>
|
||||
|
||||
-------------------- Session 4 --------------------
|
||||
Please input a picture: exit
|
||||
```
|
||||
The sample input image in Session 1 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>
|
||||
|
||||
The sample output image in Session 3 is:
|
||||
|
||||
<a href="https://llm-assets.readthedocs.io/en/latest/_images/qwen-vl-example-output-gpu.png"><img width=400px src="https://llm-assets.readthedocs.io/en/latest/_images/qwen-vl-example-output-gpu.png" ></a>
|
||||
|
|
@ -0,0 +1,98 @@
|
|||
#
|
||||
# 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 argparse
|
||||
import os
|
||||
|
||||
import torch
|
||||
from transformers import AutoTokenizer
|
||||
from transformers.generation import GenerationConfig
|
||||
|
||||
from bigdl.llm.transformers import AutoModelForCausalLM
|
||||
import intel_extension_for_pytorch as ipex
|
||||
|
||||
torch.manual_seed(1234)
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `chat()` API for Qwen-VL model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="Qwen/Qwen-VL-Chat",
|
||||
help='The huggingface repo id for the Qwen-VL model to be downloaded'
|
||||
', or the path to the huggingface checkpoint folder')
|
||||
parser.add_argument('--n-predict', type=int, default=32, help='Max tokens to predict')
|
||||
|
||||
current_path = os.path.dirname(os.path.abspath(__file__))
|
||||
args = parser.parse_args()
|
||||
model_path = args.repo_id_or_model_path
|
||||
|
||||
# Load model
|
||||
# For successful BigDL-LLM optimization on Qwen-VL-Chat, skip the 'c_fc' and 'out_proj' modules during optimization
|
||||
model = AutoModelForCausalLM.from_pretrained(model_path,
|
||||
load_in_4bit=True,
|
||||
trust_remote_code=True,
|
||||
modules_to_not_convert=['c_fc', 'out_proj'])
|
||||
model = model.to('xpu')
|
||||
# Due to issue https://github.com/intel/intel-extension-for-pytorch/issues/454,
|
||||
# currently put interpolation execution into cpu
|
||||
def to_cpu(module, input, output):
|
||||
return output.to("cpu")
|
||||
|
||||
def to_xpu(module, input):
|
||||
return (input[0].to("xpu"),)
|
||||
|
||||
model.transformer.visual.ln_pre.register_forward_hook(to_cpu)
|
||||
model.transformer.visual.transformer.register_forward_pre_hook(to_xpu)
|
||||
|
||||
# Specify hyperparameters for generation (No need to do this if you are using transformers>=4.32.0)
|
||||
model.generation_config = GenerationConfig.from_pretrained(model_path, trust_remote_code=True)
|
||||
|
||||
# Load tokenizer
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
||||
|
||||
# Session ID
|
||||
session_id = 1
|
||||
|
||||
while True:
|
||||
print('-'*20, 'Session %d' % session_id, '-'*20)
|
||||
image_input = input(f' Please input a picture: ')
|
||||
if image_input.lower() == 'exit' : # type 'exit' to quit the dialouge
|
||||
break
|
||||
|
||||
text_input = input(f' Please enter the text: ')
|
||||
if text_input.lower() == 'exit' : # type 'exit' to quit the dialouge
|
||||
break
|
||||
|
||||
if session_id == 1:
|
||||
history = None
|
||||
|
||||
all_input = [{'image': image_input}, {'text': text_input}]
|
||||
input_list = [_input for _input in all_input if list(_input.values())[0] != '']
|
||||
|
||||
if len(input_list) == 0:
|
||||
print("Input list should not be empty. Please try again with valid input.")
|
||||
continue
|
||||
|
||||
query = tokenizer.from_list_format(input_list)
|
||||
response, history = model.chat(tokenizer, query = query, history = history)
|
||||
torch.xpu.synchronize()
|
||||
|
||||
print('-'*10, 'Response', '-'*10)
|
||||
print(response, '\n')
|
||||
|
||||
image = tokenizer.draw_bbox_on_latest_picture(response, history)
|
||||
if image is not None:
|
||||
image.save(os.path.join(current_path, f'Session_{session_id}.png'), )
|
||||
|
||||
session_id += 1
|
||||
|
|
@ -0,0 +1,78 @@
|
|||
# Qwen-VL
|
||||
In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate Qwen-VL models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [Qwen/Qwen-VL-Chat](https://huggingface.co/Qwen/Qwen-VL-Chat) as a reference Qwen-VL model.
|
||||
|
||||
## Requirements
|
||||
To run these examples with BigDL-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../README.md#recommended-requirements) for more information.
|
||||
|
||||
## Example: Multimodal chat using `chat()` API
|
||||
In the example [chat.py](./chat.py), we show a basic use case for a Qwen-VL model to start a multimodal chat using `chat()` API, with BigDL-LLM 'optimize_model' API on Intel GPUs.
|
||||
### 1. Install
|
||||
We suggest using conda to manage the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
|
||||
|
||||
After installing conda, create a Python environment for BigDL-LLM:
|
||||
```bash
|
||||
conda create -n llm python=3.9 # recommend to use Python 3.9
|
||||
conda activate llm
|
||||
# below command will install intel_extension_for_pytorch==2.0.110+xpu as default
|
||||
# you can install specific ipex/torch version for your need
|
||||
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
|
||||
pip install accelerate tiktoken einops transformers_stream_generator==0.0.4 scipy torchvision pillow tensorboard matplotlib # additional package required for Qwen-VL-Chat to conduct generation
|
||||
```
|
||||
|
||||
### 2. Configures OneAPI environment variables
|
||||
```bash
|
||||
source /opt/intel/oneapi/setvars.sh
|
||||
```
|
||||
|
||||
### 3. Run
|
||||
|
||||
For optimal performance on Arc, it is recommended to set several environment variables.
|
||||
|
||||
```bash
|
||||
export USE_XETLA=OFF
|
||||
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
|
||||
```
|
||||
```
|
||||
python ./chat.py
|
||||
```
|
||||
|
||||
Arguments info:
|
||||
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Qwen-VL model (e.g `Qwen/Qwen-VL-Chat`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'Qwen/Qwen-VL-Chat'`.
|
||||
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
|
||||
|
||||
In every session, image and text can be entered into cmd (user can skip the input by type **'Enter'**) ; please type **'exit'** anytime you want to quit the dialouge.
|
||||
|
||||
Every image output will be named as the round of session and placed under the current directory.
|
||||
|
||||
#### Sample Chat
|
||||
#### [Qwen/Qwen-VL-Chat](https://huggingface.co/Qwen/Qwen-VL-Chat)
|
||||
|
||||
```log
|
||||
-------------------- Session 1 --------------------
|
||||
Please input a picture: http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg
|
||||
Please enter the text: 这是什么？
|
||||
---------- Response ----------
|
||||
这是一张图片，展现了一个穿着粉色条纹连衣裙的小女孩，她手持一只穿粉色裙子的小熊。这个场景发生在一个户外环境，有砖块背景墙和花朵。
|
||||
|
||||
-------------------- Session 2 --------------------
|
||||
Please input a picture:
|
||||
Please enter the text: 这个小女孩多大了？
|
||||
---------- Response ----------
|
||||
根据图片中的描述，这个小女孩应该是年龄较小的孩子，但具体年龄难以确定。从她的外表来看，可能是在5岁左右。。
|
||||
|
||||
-------------------- Session 3 --------------------
|
||||
Please input a picture:
|
||||
Please enter the text: 在图中检测框出玩具熊
|
||||
---------- Response ----------
|
||||
<ref>玩具熊</ref><box>(330,267),(603,869)</box>
|
||||
|
||||
-------------------- Session 4 --------------------
|
||||
Please input a picture: exit
|
||||
```
|
||||
The sample input image in Session 1 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>
|
||||
|
||||
The sample output image in Session 3 is:
|
||||
|
||||
<a href="https://llm-assets.readthedocs.io/en/latest/_images/qwen-vl-example-output-gpu.png"><img width=400px src="https://llm-assets.readthedocs.io/en/latest/_images/qwen-vl-example-output-gpu.png" ></a>
|
||||
100
python/llm/example/GPU/PyTorch-Models/Model/qwen-vl/chat.py
Normal file
100
python/llm/example/GPU/PyTorch-Models/Model/qwen-vl/chat.py
Normal file
|
|
@ -0,0 +1,100 @@
|
|||
#
|
||||
# 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 argparse
|
||||
import os
|
||||
|
||||
import torch
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
from transformers.generation import GenerationConfig
|
||||
|
||||
from bigdl.llm import optimize_model
|
||||
import intel_extension_for_pytorch as ipex
|
||||
|
||||
torch.manual_seed(1234)
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `chat()` API for Qwen-VL model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="Qwen/Qwen-VL-Chat",
|
||||
help='The huggingface repo id for the Qwen-VL model to be downloaded'
|
||||
', or the path to the huggingface checkpoint folder')
|
||||
parser.add_argument('--n-predict', type=int, default=32, help='Max tokens to predict')
|
||||
|
||||
current_path = os.path.dirname(os.path.abspath(__file__))
|
||||
args = parser.parse_args()
|
||||
model_path = args.repo_id_or_model_path
|
||||
|
||||
# Load model
|
||||
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="cpu", trust_remote_code=True)
|
||||
|
||||
# With only one line to enable BigDL-LLM optimization on model
|
||||
# For successful BigDL-LLM optimization on Qwen-VL-Chat, skip the 'c_fc' and 'out_proj' modules during optimization
|
||||
model = optimize_model(model,
|
||||
low_bit='sym_int4',
|
||||
modules_to_not_convert=['c_fc', 'out_proj'])
|
||||
model = model.to('xpu')
|
||||
# Due to issue https://github.com/intel/intel-extension-for-pytorch/issues/454,
|
||||
# currently put interpolation execution into cpu
|
||||
def to_cpu(module, input, output):
|
||||
return output.to("cpu")
|
||||
|
||||
def to_xpu(module, input):
|
||||
return (input[0].to("xpu"),)
|
||||
|
||||
model.transformer.visual.ln_pre.register_forward_hook(to_cpu)
|
||||
model.transformer.visual.transformer.register_forward_pre_hook(to_xpu)
|
||||
|
||||
# Specify hyperparameters for generation (No need to do this if you are using transformers>=4.32.0)
|
||||
model.generation_config = GenerationConfig.from_pretrained(model_path, trust_remote_code=True)
|
||||
|
||||
# Load tokenizer
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
||||
|
||||
# Session ID
|
||||
session_id = 1
|
||||
|
||||
while True:
|
||||
print('-'*20, 'Session %d' % session_id, '-'*20)
|
||||
image_input = input(f' Please input a picture: ')
|
||||
if image_input.lower() == 'exit' : # type 'exit' to quit the dialouge
|
||||
break
|
||||
|
||||
text_input = input(f' Please enter the text: ')
|
||||
if text_input.lower() == 'exit' : # type 'exit' to quit the dialouge
|
||||
break
|
||||
|
||||
if session_id == 1:
|
||||
history = None
|
||||
|
||||
all_input = [{'image': image_input}, {'text': text_input}]
|
||||
input_list = [_input for _input in all_input if list(_input.values())[0] != '']
|
||||
|
||||
if len(input_list) == 0:
|
||||
print("Input list should not be empty. Please try again with valid input.")
|
||||
continue
|
||||
|
||||
query = tokenizer.from_list_format(input_list)
|
||||
response, history = model.chat(tokenizer, query = query, history = history)
|
||||
torch.xpu.synchronize()
|
||||
|
||||
print('-'*10, 'Response', '-'*10)
|
||||
print(response, '\n')
|
||||
|
||||
image = tokenizer.draw_bbox_on_latest_picture(response, history)
|
||||
if image is not None:
|
||||
image.save(os.path.join(current_path, f'Session_{session_id}.png'), )
|
||||
|
||||
session_id += 1
|
||||
Loading…
Reference in a new issue