冬
gcw_IDzXRVNw/VisualQuality-R1-7B-ascend
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

VisualQuality-R1-7B 昇腾 NPU 适配部署文档

1. 模型简介

VisualQuality-R1-7B 是基于 Qwen2.5-VL-7B-Instruct 微调的视觉质量评估模型,来自论文 VisualQuality-R1,已被 NeurIPS 2025 接收为 spotlight。

该模型是首个基于 RL2R (Reinforcement Learning to Reason) 增强的 NR-IQA 模型,能够通过推理生成质量描述和评分。

模型架构

属性值
模型类型Qwen2_5_VLForConditionalGeneration
参数量8.29 B
Hidden Size3584
Transformer 层数28
Attention Heads28 (GQA, 4 KV heads)
Vision Hidden Size1280
Vocab Size152064
基础模型Qwen2.5-VL-7B-Instruct

模型特点

  • 视觉质量评估: 评估图像质量并给出 1-5 分的评分
  • 推理能力: 基于 GRPO/R1 方法增强的推理能力
  • 多模态: 同时处理图像和文本
  • NeurIPS 2025 Spotlight: 顶级会议接收

输入输出格式

  • 输入: 图像 + 文本提示
  • 输出: 质量评分 (1-5 分,保留两位小数)

2. 环境依赖清单

2.1 硬件要求

  • 华为昇腾系列 AI 处理器(Ascend 910B/310P 等)
  • 建议 ≥ 20 GB NPU 显存(VisualQuality-R1-7B 推理峰值约 16 GB)
  • 本次实测使用 910B2(单卡 64 GB)

2.2 软件依赖

依赖项版本要求说明
操作系统CentOS 7.6+ / Ubuntu 18.04+支持主流 Linux 发行版
CANN5.0.RC2+华为昇腾计算架构
Python3.10+推荐 3.10+
PyTorch2.0.0+推荐 2.12
torch_npu与 PyTorch 版本匹配昇腾 PyTorch 扩展
transformers>=4.49.0Qwen2.5-VL 模型加载
qwen-vl-utils最新版视觉信息处理工具

2.3 安装方式

# 1) 激活已预装 torch / torch_npu 的 conda 环境
conda activate pt2100

# 2) 安装依赖(使用华为云源)
pip install transformers qwen-vl-utils -i https://repo.huaweicloud.com/repository/pypi/simple

# 3) 验证环境
python -c "
import torch, torch_npu
print('torch:', torch.__version__)
print('torch_npu:', torch_npu.__version__)
print('NPU 可用:', torch.npu.is_available())
"

3. 分步推理操作流程

步骤 1:准备工作目录

cd /workspace/agent4/VisualQuality-R1-7B-ascend

ls -la /workspace/agent3/TianheWu/VisualQuality-R1-7B/

步骤 2:检查 NPU 环境

npu-smi info
python -c "import torch; import torch_npu; print('NPU 可用:', torch.npu.is_available())"

步骤 3:运行推理脚本

python inference.py --device npu:0
python inference.py --device npu:0 --max_new_tokens 512

4. 完整测试用例

测试用例 1:图像质量评估推理(NPU:0)

conda activate pt2100
cd /workspace/agent4/VisualQuality-R1-7B-ascend
python inference.py --device npu:0

实际输出:

============================================================
VisualQuality-R1-7B 昇腾 NPU 推理
============================================================

[Device] npu:0

[Model Path] /workspace/agent3/TianheWu/VisualQuality-R1-7B

[Config]
  model_type: qwen2_5_vl
  architectures: ['Qwen2_5_VLForConditionalGeneration']
  hidden_size: 3584
  num_hidden_layers: 28
  num_attention_heads: 28
  vocab_size: 152064
  vision_hidden_size: 1280

[Memory Before Model Load] Allocated: 0.00 GB, Reserved: 0.00 GB

[npu-smi info]
+------------------------------------------------------------------------------------------------+
| npu-smi 25.5.2                   Version: 25.5.2                                               |
+---------------------------+---------------+----------------------------------------------------+
| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
| Chip                      | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
+===========================+===============+====================================================+
| 0     910B2               | OK            | 95.1        41                0    / 0             |
| 0                         | 0000:C1:00.0  | 0           0    / 0          3424 / 65536         |
+===========================+===============+====================================================+


[Loading] VisualQuality-R1-7B model...
[Loading] VisualQuality-R1-7B model from /workspace/agent3/TianheWu/VisualQuality-R1-7B
  Tokenizer loaded: vocab_size=151643
Loading weights: 100%|██████████| 729/729 [00:09<00:00, 76.42it/s]
  Model loaded: 8.29 B parameters

[Model Loaded] Allocated: 15.45 GB, Reserved: 15.45 GB

[npu-smi info]
+------------------------------------------------------------------------------------------------+
| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
| Chip                      | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
+===========================+===============+====================================================+
| 0     910B2               | OK            | 95.1        41                0    / 0             |
| 0                         | 0000:C1:00.0  | 0           0    / 0          19306/ 65536         |
+========================---+---------------+----------------------------------------------------+
| NPU     Chip              | Process id    | Process name             | Process memory(MB)      |
+===========================+===============+====================================================+
| 0       0                 | 1708858       | python                   | 15936                    |
+===========================+===============+====================================================+


[Running] VisualQuality-R1-7B inference...

[npu-smi info]
+------------------------------------------------------------------------------------------------+
| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
| Chip                      | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
+===========================+===============+====================================================+
| 0     910B2               | OK            | 113.1       41                0    / 0             |
| 0                         | 0000:C1:00.0  | 0           0    / 0          19634/ 65536         |
+===========================+===============+====================================================+
| NPU     Chip              | Process id    | Process name             | Process memory(MB)      |
+===========================+===============+===============+==========================================|
| 0       0                 | 1708858       | python                   | 16269                    |
+===========================+===============+====================================================+


[Results]
  Generated text length: 21
  Output preview: <answer>0.00</answer>...

[Memory After Inference] Allocated: 15.45 GB, Reserved: 15.64 GB

[Done]

性能指标

指标值
模型参数量8.29 B
推理显存峰值~16 GB
输出长度21 tokens

5. 注意事项

  1. 大模型: 8.29B 参数,需要较大 NPU 显存
  2. Vision-Language: 同时加载视觉编码器和语言模型
  3. Qwen2.5-VL: 基于最新的 Qwen2.5-VL 架构

6. 参考链接

  • VisualQuality-R1 论文 (arXiv:2505.14460)
  • VisualQuality-R1 GitHub
  • HuggingFace 模型页
  • Qwen2.5-VL