🤖 Diffusion Chatbot

Flask server hosting the Qwen3-0.6B-diffusion-bd3lm-v0.1 model with real-time streaming inference. Watch diffusion language models generate text step-by-step!

✨ Features

• 🎯 Real-time Streaming: Watch the diffusion denoising process live
• 📡 Three API Endpoints: Simple generation, batch states, and SSE streaming
• ⚡ GPU Support: Automatic GPU detection with CPU fallback
• 🔄 Progressive Generation: See how different parts of text appear at different steps

📡 API Endpoints

1. Health Check

GET /health

2. Generate Text (Simple)

POST /generate
Content-Type: application/json

{
  "prompt": "Your question here",
  "max_new_tokens": 256
}

3. Generate with Real-time Streaming (SSE) ⭐

POST /generate_sse
Content-Type: application/json

{
  "prompt": "Your question here",
  "max_new_tokens": 100,
  "capture_interval": 10
}

💡 Example Usage

# Simple generation
curl -X POST https://YOUR_USERNAME-diffusion-chatbot.hf.space/generate \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Hello, how are you?", "max_new_tokens": 50}'

# Real-time streaming
curl -N -X POST https://YOUR_USERNAME-diffusion-chatbot.hf.space/generate_sse \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Write a poem", "max_new_tokens": 100, "capture_interval": 10}'

🔧 Technical Details

⚙️ Configuration

🧠 How It Works

Unlike traditional language models that generate text left-to-right, diffusion language models:

1. Start with all tokens masked
2. Iteratively denoise over multiple steps
3. Generate different parts of text at different steps
4. Create a unique "thought process" visualization

📝 Notes

• Model downloads automatically on first run (~1.5GB)
• First request may be slow as model loads
• GPU is optional - automatic CPU fallback
• Lower capture_interval = more frequent updates

🙏 Acknowledgments

• Model: dllm-hub
• Framework: dLLM
• Base: Qwen