Qwen VL Sample

This document describes running the upstream Qwen VL multimodal vision-language sample with QAI AppBuilder on Linux aarch64 (Gradio UI or in-script Inference).

Supported devices

Device	SoC
Fogwise® AIRbox Q900	QCS9075

Install QAI AppBuilder

tip

1. Install QAI AppBuilder by following the QAI AppBuilder installation guide.

2. Configure ADSP environment variables as described in Create ADSP environment variables.

Run the sample

Install dependencies

Install sample dependencies in the activated virtual environment:

Device

pip3 install "transformers==4.57.0" qwen-vl-utils gradio Pillow torch torchvision opencv-python-headless

tip

qwen_vl.py imports Qwen3-related modules, so you need transformers ≥ 4.57 (includes transformers.models.qwen3_vl). Also install qwen-vl-utils.

Prepare the model

This sample does not auto-download Qwen weights via --chipset / AI Hub. Follow the upstream README for a pre-quantized package, for example Qwen2-VL 2B:

Device

cd qai-appbuilder/samples/Multimodal/qwen_vl
wget -c https://www.aidevhome.com/data/adh2/models/suggested/qwen2vl2b.zip
unzip -o qwen2vl2b.zip
# Example extract layout: ./qwen2vl2b/ (veg.serialized.bin, weight_sharing_model_1_of_1.serialized.bin, tokenizer.json, ...)

The zip is about 1.8 GB; extracted size is about 2.6 GB. Reserve disk space accordingly.

Run the script

• Enter the sample directory and set ADSP (same as the install page)

  Device

  cd qai-appbuilder/samples/Multimodal/qwen_vl
  export ADSP_LIBRARY_PATH=$(python3 -c "import os, qai_appbuilder; print(os.path.join(os.path.dirname(qai_appbuilder.__file__), 'libs'))")

• Start the Gradio web UI (default 0.0.0.0:7861)

  Device

  python3 qwen_vl.py --model qwen2 --path ./qwen2vl2b --host 0.0.0.0 --port 7861

Open http://<device-ip>:7861, upload an image, and enter a question.

• Optional headless check: after loading the model, call Inference once. On AIRbox Q900, with sample image vlm_demo.png and prompt Describe this image in one sentence., a response similar to the following was observed:

A webpage with a form for uploading files.

warning

run_inference.py --model qwen_vl does not replace model download. The entrypoint requires --path to the model directory and has no --chipset 9075 argument.

tip

If TBB is installed, some setups set LD_PRELOAD to libtbb.so per the upstream README; on Q900 a one-shot inference also succeeded without it.