Overshoot
← Back to all posts

How to Deploy Qwen3-VL-32B with vLLM

Younes El Hjouji

Detailed Deployment Instructions

1. Environment Setup

# Create a working directory (choose your preferred location)

# Create virtual environment with Python 3.12
python3.12 -m venv venv

# Activate virtual environment
source venv/bin/activate

# Verify Python version (should be 3.12.x)
python --version

# Upgrade pip
pip install --upgrade pip

# Install vLLM
pip install vllm

2. Launch Deployment

# Make sure you're in your working directory and activate virtual environment
source venv/bin/activate

# Recommended public-facing configuration
vllm serve Qwen/Qwen3-VL-32B-Instruct \
  --dtype bfloat16 \
  --max-model-len 8192 \
  --max-num-batched-tokens 8192 \
  --gpu-memory-utilization 0.75 \
  --port 8000

3. Verification

3.1. Check server is running:

curl http://localhost:8000/v1/models

3.2. Test text inference:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-VL-32B-Instruct",
    "prompt": "Hello, how are you?",
    "max_tokens": 50
  }'

3.3. Test video inference:

VIDEO_BASE64=$(base64 -w 0 /path/to/video.mp4)

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"Qwen/Qwen3-VL-32B-Instruct\",
    \"messages\": [{
      \"role\": \"user\",
      \"content\": [
        {
          \"type\": \"video_url\",
          \"video_url\": {
            \"url\": \"data:video/mp4;base64,\$VIDEO_BASE64\"
          }
        },
        {
          \"type\": \"text\",
          \"text\": \"Describe this video in detail.\"
        }
      ]
    }],
    \"max_tokens\": 150
  }"

Verified Results:

  • Deployment succeeded with the 8K / 75% GPU configuration.
  • Text inference returned in about 0.99 seconds for a short prompt.
  • A short video inference request completed in about 9.2 seconds and produced a detailed description.

4. Troubleshooting

Issue 1: Not enough free GPU memory

  • Solution: Lower --gpu-memory-utilization or reduce --max-model-len.

Issue 2: KV cache is too small for the requested context

  • Solution: Reduce --max-model-len until the requested context fits the available GPU memory.

Issue 3: Slow first load

  • Solution: This is expected. The initial run includes checkpoint loading, torch compilation, and CUDA graph capture.

5. Notes

  • This is the dense Qwen3-VL 32B model, so all parameters are active on every token.
  • The tested deployment used about 109 GB at the recommended 8K setting.
  • Native context is much larger, but 8K was the clean verified deployment configuration here.
  • This model supports text, image, and video inputs through vLLM.

6. References