Instructions to use LinkRur/Qwen_2.5_1.5B_Instruct_Vision with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LinkRur/Qwen_2.5_1.5B_Instruct_Vision with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="LinkRur/Qwen_2.5_1.5B_Instruct_Vision", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("LinkRur/Qwen_2.5_1.5B_Instruct_Vision", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("LinkRur/Qwen_2.5_1.5B_Instruct_Vision", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use LinkRur/Qwen_2.5_1.5B_Instruct_Vision with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "LinkRur/Qwen_2.5_1.5B_Instruct_Vision"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LinkRur/Qwen_2.5_1.5B_Instruct_Vision",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/LinkRur/Qwen_2.5_1.5B_Instruct_Vision

SGLang

How to use LinkRur/Qwen_2.5_1.5B_Instruct_Vision with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "LinkRur/Qwen_2.5_1.5B_Instruct_Vision" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LinkRur/Qwen_2.5_1.5B_Instruct_Vision",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "LinkRur/Qwen_2.5_1.5B_Instruct_Vision" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LinkRur/Qwen_2.5_1.5B_Instruct_Vision",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use LinkRur/Qwen_2.5_1.5B_Instruct_Vision with Docker Model Runner:
```
docker model run hf.co/LinkRur/Qwen_2.5_1.5B_Instruct_Vision
```

Qwen_2.5_1.5B_Instruct_Vision

基于 DINOv2 与 NSD (Normalized State Decomposition) 的视觉语言模型，通过超球面归一化实现视觉-语言特征对齐。

方法

NSD 定义

给定向量 v ∈ R^D，NSD 将其投影到单位超球面 S^{D-1}：

$\text{NSD}(\mathbf{v}) = \frac{\mathbf{v}}{\|\mathbf{v}\|_2}$

归一化后，内积等价于余弦相似度：

$\langle \text{NSD}(\mathbf{v}_1), \text{NSD}(\mathbf{v}_2) \rangle = \cos(\theta)$

视觉-语言对齐

DINOv2 输出 h ∈ R^{384}，经投影层映射到 Qwen 的 1536 维嵌入空间，再通过 NSD 归一化到 S^{1535}：

$\mathbf{z} = \text{NSD}\left(W_2 \cdot \text{ReLU}(W_1 \mathbf{h} + \mathbf{b}_1) + \mathbf{b}_2\right)$

其中 W_1 ∈ R^{768×384}, W_2 ∈ R^{1536×768} 为投影层参数。

Qwen 词表嵌入 E ∈ R^{V×1536} 同样归一化：

$\hat{\mathbf{E}} = \text{NSD}(\mathbf{E})$

每个视觉 token 在词表中检索最近邻：

$t^* = \arg\max_{t \in \mathcal{V}} \ \mathbf{z}^\top \hat{\mathbf{e}}_t$

训练

投影层与 DINOv2 后 4 层使用 InfoNCE 对比损失微调：

$\mathcal{L} = -\log \frac{\exp(\mathbf{z}_i^\top \mathbf{y}_i / \tau)}{\sum_{j=1}^{B} \exp(\mathbf{z}_i^\top \mathbf{y}_j / \tau)}$

其中 y_i 为类别 i 的 Qwen 嵌入，τ = 0.07 为温度系数。

架构

组件	说明	参数量
DINOv2 ViT-S/14	视觉编码器 (微调后 4 层)	22M
MLP Projection	384 → 768 → 1536	1.18M
Qwen2.5-1.5B-Instruct	语言模型 (冻结)	1543.7M

性能

在 tiny-imagenet-200 验证集 (10000 张) 上的视觉 token 准确率：

Top-K	命中率
Top-1	72.43%
Top-3	80.02%
Top-5	81.18%
Top-10	81.67%

使用

from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image

# 加载模型
model = AutoModelForCausalLM.from_pretrained(
    "LinkRur/Qwen_2.5_1.5B_Instruct_Vision",  # 替换为你的模型名
    trust_remote_code=True,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(
    "LinkRur/Qwen_2.5_1.5B_Instruct_Vision",  # 替换为你的模型名
    trust_remote_code=True
)

# 推理
image = Image.open("photo.jpg").convert("RGB")
vis_desc, response = model.generate_with_image(image, "描述这张图片", tokenizer)
print(response)

License

MIT

Downloads last month: 53

Safetensors

Model size

2B params

Tensor type

F32

BF16