VLLM_SETUP_GUIDE.md

# RunPod vLLM Template Setup for Ultravox

## ✅ Use Pre-built vLLM (No Docker Building!)

This guide uses RunPod's **existing vLLM Docker image** - just configure and deploy.

---

## 🚀 Step-by-Step Setup (10 minutes)

### Step 1: Open RunPod Console

🔗 **Go to:** https://www.runpod.io/console/serverless

**Click:** "+ New Endpoint"

### Step 2: Select vLLM Template

**Search for:** "vLLM"

**Select:** "vLLM - Fast LLM Inference" (official RunPod template)

### Step 3: Configure Ultravox Model

**Endpoint Configuration:**

```
Name: ultravox-vllm

Container Image: runpod/worker-vllm:stable (pre-built!)

GPU Type: RTX 4090 (24GB VRAM)

Container Disk: 40 GB
```

**Environment Variables:**

Click "Add Environment Variable" and add these:

| Name | Value |
|------|-------|
| `MODEL_NAME` | `fixie-ai/ultravox-v0_2` |
| `HF_TOKEN` | `YOUR_HF_TOKEN_HERE` |
| `MAX_MODEL_LEN` | `4096` |
| `GPU_MEMORY_UTILIZATION` | `0.9` |
| `TRUST_REMOTE_CODE` | `true` |

**Scaling:**

```
Min Workers: 0
Max Workers: 3
Scale Down Delay: 600 (10 minutes)
```

### Step 4: Deploy

**Click:** "Deploy"

**Wait:** ~3-5 minutes for deployment

**Status:** Should show "Running" with green indicator

**Copy:** The **Endpoint ID** (looks like: `abc123def456`)

---

## 🧪 Test Your Endpoint

### Quick Test in RunPod Console

1. Go to your endpoint
2. Click "Requests" tab
3. Click "Send Test Request"
4. Use this payload:

```json
{
  "input": {
    "prompt": "Hello! How are you today?",
    "max_tokens": 100,
    "temperature": 0.7
  }
}
```

5. Click "Run"
6. Wait ~8-12 seconds (cold start)
7. Should return text response

### Test from Command Line

```bash
export RUNPOD_ENDPOINT_ID="your-endpoint-id-here"

python3 << 'EOF'
import runpod
import os

runpod.api_key = "YOUR_RUNPOD_API_KEY_HERE"
endpoint = runpod.Endpoint(os.getenv("RUNPOD_ENDPOINT_ID"))

result = endpoint.run_sync({
    "input": {
        "prompt": "What is artificial intelligence?",
        "max_tokens": 100
    }
}, timeout=60)

print("Response:", result)
EOF
```

---

## ⚙️ Configure Our Service

Once you have the **Endpoint ID**, update the config:

```bash
# SSH to server
ssh -p 33337 root@136.59.129.136

# Edit config
nano /workspace/ultravox-pipeline/config/runpod.yaml
```

**Add:**

```yaml
endpoints:
  ultravox:
    endpoint_id: "YOUR_ENDPOINT_ID_HERE"  # Paste it here
    gpu: "RTX_4090"
    min_workers: 0
    max_workers: 3
```

**Save and exit:** `Ctrl+X`, `Y`, `Enter`

---

## 🎯 Test from Service

```bash
cd /workspace/ultravox-pipeline/src/services/runpod_llm

# Set environment
export RUNPOD_API_KEY="YOUR_RUNPOD_API_KEY_HERE"
export RUNPOD_ENDPOINT_ID="your-endpoint-id"

# Run service
python3 service.py &

# Test
curl -X POST http://localhost:8105/runpod/inference \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ultravox",
    "input": {
      "text": "Hello, world!"
    },
    "parameters": {
      "max_tokens": 50
    }
  }'
```

---

## 📊 Expected Performance

### Cold Start (First Request)
- **Time:** 8-15 seconds
- **Why:** Downloading model from HF → Loading into VRAM
- **Frequency:** Once per idle period (after scale-down)

### Warm Inference
- **Time:** 0.3-0.8 seconds
- **Throughput:** ~30-50 tokens/second

### Costs
- **Price:** $0.34/hour when active
- **Testing:** ~$0.01 for 1 hour of testing
- **Production (2hr/day):** ~$20/month

---

## 🔧 Troubleshooting

### Model not loading
- Check `HF_TOKEN` is set correctly
- Verify model name: `fixie-ai/ultravox-v0_2`
- Check logs in RunPod console

### Out of memory
- Reduce `MAX_MODEL_LEN` to `2048`
- Set `GPU_MEMORY_UTILIZATION` to `0.8`

### Slow cold starts
- Pre-download model (advanced)
- Use network storage (costs extra)

### Connection timeout
- Increase timeout to 120 seconds
- Check endpoint is running (green status)

---

## ⚠️ Important Notes

**About vLLM + Ultravox:**
- vLLM is primarily for **text-only** models
- Ultravox is **multimodal** (audio + text)
- vLLM will work for **text input only**
- For **audio input**, you need custom handler (Docker)

**What works with vLLM:**
- ✅ Text → Text (LLM inference)
- ❌ Audio → Text (needs custom handler)
- ❌ Text → Audio (needs TTS integration)

**For full audio support:**
- Use the custom Docker image (build.sh)
- Or process audio client-side (convert to text first)

---

## 📝 Summary

1. ✅ No Docker building needed
2. ✅ Use RunPod's vLLM template
3. ✅ Set model name to `fixie-ai/ultravox-v0_2`
4. ✅ Add HF token for access
5. ⚠️ Text-only mode (no audio input/output)

For **full speech-to-speech**, build custom Docker image.
For **text-only testing**, vLLM is perfect!

---

**Ready?** Follow steps 1-4 above and paste your Endpoint ID here!