← Back to all posts

How to Deploy InternVL3-2B 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
# Make sure you have Python 3.12 installed: sudo apt install python3.12 python3.12-venv
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 (this takes ~5-10 minutes)
pip install vllm

2. Launch Deployment

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

# Launch vLLM serve (full native context)
vllm serve OpenGVLab/InternVL3-2B \
  --trust-remote-code \
  --dtype bfloat16 \
  --max-model-len 16384 \
  --max-num-batched-tokens 16384 \
  --gpu-memory-utilization 0.9 \
  --port 8000

3. Verification

3.1. Check server is running:

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

Expected output:

{
  "object": "list",
  "data": [{
    "id": "OpenGVLab/InternVL3-2B",
    "object": "model",
    "created": 1770107841,
    "owned_by": "vllm",
    "root": "OpenGVLab/InternVL3-2B",
    "parent": null,
    "max_model_len": 16384
  }]
}

3.2. Test text inference:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "OpenGVLab/InternVL3-2B",
    "prompt": "Hello, how are you?",
    "max_tokens": 50
  }'

3.3. Test video inference:

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

# Time the request and measure latency
time curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"OpenGVLab/InternVL3-2B\",
    \"messages\": [{
      \"role\": \"user\",
      \"content\": [
        {
          \"type\": \"video_url\",
          \"video_url\": {
            \"url\": \"data:video/mp4;base64,\$VIDEO_BASE64\"
          }
        },
        {
          \"type\": \"text\",
          \"text\": \"Describe the user actions in this video in a sequential list.\"}
      ]
    }],
    \"max_tokens\": 1024
  }"

Example output:

Certainly. In the video, a man wearing a red jersey is standing on a soccer field.
He appears to be clapping his hands or gesturing with them while speaking, expressing his comments or opinions...

Video inference latency: 0.87s (real time) - Extremely fast!

4. Troubleshooting

Issue 1: Trust remote code required

  • Symptom: "ValueError: trust_remote_code is required for this model"
  • Solution: Add --trust-remote-code flag (already included in deployment command)

Issue 2: Out of memory

  • Symptom: CUDA out of memory
  • Solution: This model is small (2B params), only needs ~10GB. If OOM occurs, stop other models or reduce --gpu-memory-utilization to 0.5

Issue 3: Slow compilation on first run

  • Symptom: Server takes 2 minutes to fully start
  • Solution: This is normal. Torch compilation caches graphs for future use. Subsequent startups will be faster.

5. Notes

  • Context length: Model supports 16,384 tokens (session length)
  • Multimodal support: Supports both image and video inputs
  • Vision backbone: InternViT-300M-448px-V2.5
  • LLM backbone: Qwen2.5-1.5B
  • Flash Attention: Automatically enabled for vision encoder (AttentionBackendEnum.FLASH_ATTN)
  • Torch Compilation: Enabled for backbone with Inductor backend for better performance
  • Variable Visual Position Encoding (V2PE): Improves long-context understanding
  • Extremely fast: 0.87s video inference latency is exceptional for a 2B model

6. Model Specifications

  • Model size: 2B parameters
  • Architecture: InternVL3 (InternViT-300M-448px-V2.5 + Qwen2.5-1.5B)
  • Context window: 16,384 tokens
  • License: Apache-2.0
  • Quantization support: 8-bit quantization supported via load_in_8bit=True
  • Primary deployment: LMDeploy (lmdeploy>=0.7.3), but vLLM works great

7. References