Instructions to use Qwen/Qwen2-VL-7B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Qwen/Qwen2-VL-7B-Instruct with Transformers:

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

pipe = pipeline("image-text-to-text", model="Qwen/Qwen2-VL-7B-Instruct")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
model = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use Qwen/Qwen2-VL-7B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Qwen/Qwen2-VL-7B-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Qwen/Qwen2-VL-7B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Qwen/Qwen2-VL-7B-Instruct

SGLang

How to use Qwen/Qwen2-VL-7B-Instruct 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 "Qwen/Qwen2-VL-7B-Instruct" \
    --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": "Qwen/Qwen2-VL-7B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "Qwen/Qwen2-VL-7B-Instruct" \
        --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": "Qwen/Qwen2-VL-7B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Qwen/Qwen2-VL-7B-Instruct with Docker Model Runner:
```
docker model run hf.co/Qwen/Qwen2-VL-7B-Instruct
```

Questions about Naive Dynamic Resolution and the vision mask

#39

by YaYaGeGe - opened Oct 7, 2024

Discussion

YaYaGeGe

Oct 7, 2024

•

edited Oct 8, 2024

Hi Qwen team,
Great work and appreciate opensourcing it!
According to the code implementation in huggingface transformers, the image tokens firstly get reshaped into a token sequence and then being applied with conv3d which raises some disparities against to the paper. I shared the same question with this https://github.com/QwenLM/Qwen2-VL/issues/300

PatchEmbed doesn't count more than 2 frames in time dimension by that Conv3d Layer, given the sequence input; thus it doesn't involve the information over two-frames time domain at conv step, and multi frames will increase the conv computation times linearly.
I think it makes more sense that applying Conv3D on 4D image tensor first, then reshaping / flattening
Although, no matter flattening the tokens first or patchifying (conv) first, the sequence length will increase with more frames, as seq_len = grid_t * grid_h * grid_w.

Besides, when applying the vision masks to the vision attention weights by addition, can elaborate the rationality a bit? Is it an approach to emphasize the attention inside one frame pairs but mitigate the relationship between other frames?

    attention_mask = torch.zeros([1, seq_length, seq_length], device=q.device, dtype=torch.bool)
        for i in range(1, len(cu_seqlens)):
    attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
    attn_weights = attn_weights + attention_mask

YaYaGeGe

Oct 8, 2024

•

edited Feb 11, 2025

Updates:
it seems the code has been updated at the main branch, by adding -inf to 'intersection region' and zero at 'self mask' region, the mask when grid_t = 2 shows like
| 0 | 0 | -inf | -inf |
| 0 | 0 | -inf | -inf |
| -inf | -inf | 0 | 0 |
| -inf | -inf | 0 | 0 |
which mask out the attention between frames but only apply self attention of patches in each frame itself.
Does the mask setting really take effect for video understanding, without intersect attention across multiple frames? Given consecutive frames in a video clip share the same or semantically continuous contents.

        attention_mask = torch.full(
            [1, seq_length, seq_length], torch.finfo(q.dtype).min, device=q.device, dtype=q.dtype
        )
...
        attn_weights = attn_weights + attention_mask

https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L341-L343

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment