SmolVLM — Easy-to-Run Vision-Language Model 🖼️💬

This project packages SmolVLM2 into a ready-to-use Docker image with both a web UI and an HTTP API for image–text-to-text tasks.

It’s designed for fast setup and offline-friendly deployment so you can:

• Pull the container
• Run it with one command
• Start sending images and prompts instantly

⚠️ Note: Maintained by an independent developer for ease-of-use.
Not production-hardened — you may need extra work for critical deployments.

✅ Offline Mode: Supported — models can be pre-baked into the image or mounted from local storage.
✅ Minimal GPU Requirements: Works on most modern NVIDIA GPUs with at least 4GB VRAM and CUDA support (Ampere or newer recommended). Lower VRAM cards may require using the 256M model variant.

🤝 Contributions Welcome: Bug fixes, feature requests, and PRs are appreciated.

---

🆘 Support & Issues

If you encounter problems or have suggestions:

• Open a GitHub Issue with details, logs, and reproduction steps
• Use Discussions for general questions

---

🚀 Quick Start

Standard run

docker run --gpus all -p 8888:8888 sensejworld/smolvlm:latest

MultiGPU setup

docker run --gpus all -e CUDA_VISIBLE_DEVICES=1 -p 8888:8888 sensejworld/smolvlm:latest

Then open: http://localhost:8888/ui

---

🌐 API Usage Example

curl -X POST "http://localhost:8888/ptt/convert" \
  -H "Accept: application/json" \
  -F "query=What is in this image?" \
  -F "image=@cat.jpg;type=image/jpeg"

Or without an image (uses demo image):

curl -X POST "http://localhost:8888/ptt/convert" \
  -F "query=Count cats in the image?"

---

📦 Docker Features

• Pinned dependencies for reproducibility
• GPU acceleration (CUDA)
• Preloaded model option for full offline usage
• Single container with UI + API
• Clean multi-stage build

---

🐳 Docker Hub

Find available images on Docker Hub — useful for:

• Selecting a specific version
• Checking the latest builds before pulling
• Matching CUDA/PyTorch versions to your needs

---

📜 Version History

v0.0.2 (Planned)

• Support picture Batch
• Support video

v0.0.1

• Initial release
• Basic SmolVLM2 inference via API & UI
• Demo image support
• Docker GPU support
• Added FastAPI-based /ptt/convert endpoint
• Integrated Gradio chat UI with image upload
• Shared inference worker between API and UI
• Offline model prefetch script (init_downloads.py) Run with:

docker run --gpus all -p 8888:8888 sensejworld/smolvlm:v0.0.1

---

🛠 Developer Notes

See notes.md for:

• Local Docker build/run instructions
• Conda setup for local dev
• FlashAttention verification commands
• API & UI testing commands

---

📜 License

This project is licensed under the MIT License.
Models by HuggingFaceTB — check their respective licenses.