LLM - 使用 vLLM 部署 Qwen2-VL 多模态大语言模型 (配置 FlashAttention) 教程

news2024/11/17 21:21:50

欢迎关注我的CSDN：https://spike.blog.csdn.net/
本文地址：https://spike.blog.csdn.net/article/details/142528967

免责声明：本文来源于个人知识与公开资料，仅用于学术交流，欢迎讨论，不支持转载。

vLLM
vLLM 用于大语言模型(LLM) 的推理和服务，具有多项优化技术，包括先进的服务吞吐量、高效的内存管理、连续批处理请求、优化 CUDA 内核以及支持量化技术，如GPTQ、AWQ等。FlashAttention 是先进的注意力机制优化工具，通过减少内存访问和优化计算过程，显著提高大型语言模型的推理速度。

GitHub：

FlashAttention: https://github.com/Dao-AILab/flash-attention
Transformers: https://github.com/huggingface/transformers
vLLM: https://github.com/vllm-project/vllm

1. 配置 vLLM

准备 Qwen2-VL 模型，包括 7B 和 72B，即：

modelscope --token [your token] download --model Qwen/Qwen2-VL-7B-Instruct
modelscope --token [your token] download --model Qwen/Qwen2-VL-72B-Instruct-GPTQ-Int4

注意：Qwen2-VL 暂时不支持 GGUF 转换，因此不能使用 Ollama 提供服务。

配置 vLLM：

pip install vllm==0.6.1 -i https://pypi.tuna.tsinghua.edu.cn/simple

参考：vLLM - Using VLMs

注意：当前(2024.9.26)最新 Transformers 版本不支持 Qwen2-VL，需要使用固定 commit 版本，参考：

pip install git+https://github.com/huggingface/transformers.git@21fac7abba2a37fae86106f87fcf9974fd1e3830

Transformers 的 Commit ID (21fac7abba2a37fae86106f87fcf9974fd1e3830) 内容，以更新 Qwen2-VL 为主，即：

commit 21fac7abba2a37fae86106f87fcf9974fd1e3830 (HEAD)
Author: Shijie <821898965@qq.com>
Date:   Fri Sep 6 00:19:30 2024 +0800
    simple align qwen2vl kv_seq_len calculation with qwen2 (#33161)
    * qwen2vl_align_kv_seqlen_to_qwen2
    * flash att test
    * [run-slow] qwen2_vl
    * [run-slow] qwen2_vl fix OOM
    * [run-slow] qwen2_vl
    * Update tests/models/qwen2_vl/test_modeling_qwen2_vl.py
    Co-authored-by: Raushan Turganbay <raushan.turganbay@alumni.nu.edu.kz>
    * Update tests/models/qwen2_vl/test_modeling_qwen2_vl.py
    Co-authored-by: Raushan Turganbay <raushan.turganbay@alumni.nu.edu.kz>
    * code quality
    ---------

vLLM 的视觉文本测试代码，如下：

通过 SamplingParams 设置最大的 Tokens 数量。
注意，不同的模型 Image Token 也不同，Qwen2-VL 是 <|image_pad|>，而 InternVL2-2B 是 <image>

即：

from vllm import LLM, SamplingParams
import PIL
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "1"


def main():
    # Qwen2-VL
    llm = LLM(model="llm/Qwen/Qwen2-VL-7B-Instruct/")
    prompt = "USER: <|image_pad|>\nWhat is the content of this image?\nASSISTANT:"
    
    # InternVL2-2B
    # llm = LLM(model="llm/InternVL2-2B/", trust_remote_code=True)
    # Refer to the HuggingFace repo for the correct format to use
    # prompt = "USER: <image>\nWhat is the content of this image?\nASSISTANT:"
    
    # 设置最大输出 Token 数量
    sampling_params = SamplingParams(max_tokens=8172)

    # Load the image using PIL.Image
    image = PIL.Image.open("llm/img_test.jpg")

    # Single prompt inference
    outputs = llm.generate({
        "prompt": prompt,
        "multi_modal_data": {"image": image},
    }, sampling_params)

    for o in outputs:
        generated_text = o.outputs[0].text
        print(generated_text)


if __name__ == '__main__':
    main()

Qwen2-VL 的输出：

The image shows a person standing on a road near a sidewalk. The person is dressed in a light-colored outfit consisting of a short-sleeved blouse and a skirt. The background features greenery, including trees and some buildings, with a few cars parked along the street. The overall scene appears to be a sunny day with good visibility.

BugFix1：

  File "miniconda3/envs/torch-llm/lib/python3.9/site-packages/vllm/transformers_utils/configs/__init__.py", line 13, in <module>
    from vllm.transformers_utils.configs.mllama import MllamaConfig
  File "miniconda3/envs/torch-llm/lib/python3.9/site-packages/vllm/transformers_utils/configs/mllama.py", line 1, in <module>
    from transformers.models.mllama import configuration_mllama as mllama_hf_config
ModuleNotFoundError: No module named 'transformers.models.mllama'

原因：降级 vLLM 版本至 0.6.1，vllm/transformers_utils/configs/mllama.py 是 0.6.2 版本加入，即：

pip install vllm==0.6.1 -i https://pypi.tuna.tsinghua.edu.cn/simple

BugFix2：

[rank0]:   File "miniconda3/envs/torch-llm/lib/python3.9/site-packages/vllm/inputs/registry.py", line 256, in process_input
[rank0]:     return processor(InputContext(model_config), inputs)
[rank0]:   File "miniconda3/envs/torch-llm/lib/python3.9/site-packages/vllm/model_executor/models/qwen2_vl.py", line 770, in input_processor_for_qwen2_vl
[rank0]:     assert len(image_indices) == len(image_inputs)
[rank0]: AssertionError

原因，参考 vllm/model_executor/models/qwen2_vl.py，hf_config.image_token_id 与当前 Prompt 的 Image Token (<image>)，不一致，即：

prompt_token_ids = llm_inputs.get("prompt_token_ids", None)
if prompt_token_ids is None:
    prompt = llm_inputs["prompt"]
    prompt_token_ids = processor.tokenizer(
        prompt,
        padding=True,
        return_tensors=None,
    )["input_ids"]
print(f"[Info] decode prompt: \n{processor.decode(prompt_token_ids)}\n")
print(f"[Info] decode image_token_id (151655): {processor.decode([151655])}")

# Expand image pad tokens.
if image_inputs is not None:
    image_indices = [
        idx for idx, token in enumerate(prompt_token_ids)
        if token == hf_config.image_token_id
    ]
    print(f"[Info] hf_config.image_token_id: {hf_config.image_token_id}, prompt_token_ids: {prompt_token_ids}")
    image_inputs = make_batched_images(image_inputs)
    print(f"[Info] image_indices: {len(image_indices)} and image_inputs: {len(image_inputs)}")
    assert len(image_indices) == len(image_inputs)

经过分析，确定 Qwen2-VL 的 Image Token 是 <|image_pad|>，而不是 <image>，替换 Prompt 即可。

输出：

[Info] decode prompt: 
USER: <|image_pad|>
What is the content of this image?
ASSISTANT:
[Info] decode image_token_id (151655): <|image_pad|>
[Info] hf_config.image_token_id: 151655, prompt_token_ids: [6448, 25, 220, 151655, 198, 3838, 374, 279, 2213, 315, 419, 2168, 5267, 4939, 3846, 2821, 25]
[Info] image_indices: 1 and image_inputs: 1

2. 配置 FlashAttention

FlashAttention 可以加速大模型的推理过程，配置 FlashAttention，参考，安装依赖的 Python 包：

pip install packaging
pip install ninja

测试 ninja 包是否可用，即：

ninja --version  # 1.11.1.git.kitware.jobserver-1
echo $?  # 0

Ninja 类似于 Makefile，语法简单，但是比 Makefile 更加简洁。

不推荐 直接安装 flash-attn，建议使用源码安装，安装过程可控，请耐心等待，即：

pip install flash-attn --no-build-isolation

# log
Building wheels for collected packages: flash-attn
  Building wheel for flash-attn (setup.py) ... |

检测 Python 版本：

python --version # Python 3.9.19
nvidia-smi  # CUDA Version: 12.0

python

import torch
print(torch.__version__)  # 2.4.0+cu121
print(torch.cuda.is_available())  
exit()

建议通过直接源码进行安装，即：

git clone git@github.com:Dao-AILab/flash-attention.git
python setup.py install

整体的编译过程，包括 85 步，耐心等待，即：

Using envvar MAX_JOBS (64) as the number of workers...
[1/85] c++ -MMD -MF ...
# ...
Using miniconda3/envs/torch-llm/lib/python3.9/site-packages
Finished processing dependencies for flash-attn==2.6.3