# Dual GPU Setup Summary

## What We Built

A secondary llama-swap container optimized for your **AMD RX 6800** GPU using ROCm.

### Architecture

```
Primary GPU (NVIDIA GTX 1660)     Secondary GPU (AMD RX 6800)
         ↓                                    ↓
   llama-swap (CUDA)                  llama-swap-amd (ROCm)
   Port: 8090                         Port: 8091
         ↓                                    ↓
   NVIDIA models                       AMD models
   - llama3.1                         - llama3.1-amd
   - darkidol                         - darkidol-amd
   - vision (MiniCPM)                 - moondream-amd
```

## Files Created

1. **Dockerfile.llamaswap-rocm** - Custom multi-stage build:
   - Stage 1: Builds llama.cpp with ROCm from source
   - Stage 2: Builds llama-swap from source
   - Stage 3: Runtime image with both binaries

2. **llama-swap-rocm-config.yaml** - Model configuration for AMD GPU

3. **docker-compose.yml** - Updated with `llama-swap-amd` service

4. **bot/utils/gpu_router.py** - Load balancing utility

5. **bot/globals.py** - Updated with `LLAMA_AMD_URL`

6. **setup-dual-gpu.sh** - Setup verification script

7. **DUAL_GPU_SETUP.md** - Comprehensive documentation

8. **DUAL_GPU_QUICK_REF.md** - Quick reference guide

## Why Custom Build?

- llama.cpp doesn't publish ROCm Docker images (yet)
- llama-swap doesn't provide ROCm variants
- Building from source ensures latest ROCm compatibility
- Full control over compilation flags and optimization

## Build Time

The initial build takes 15-30 minutes depending on your system:
- llama.cpp compilation: ~10-20 minutes
- llama-swap compilation: ~1-2 minutes
- Image layering: ~2-5 minutes

Subsequent builds are much faster due to Docker layer caching.

## Next Steps

Once the build completes:

```bash
# 1. Start both GPU services
docker compose up -d llama-swap llama-swap-amd

# 2. Verify both are running
docker compose ps

# 3. Test NVIDIA GPU
curl http://localhost:8090/health

# 4. Test AMD GPU
curl http://localhost:8091/health

# 5. Monitor logs
docker compose logs -f llama-swap-amd

# 6. Test model loading on AMD
curl -X POST http://localhost:8091/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama3.1-amd",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 50
  }'
```

## Device Access

The AMD container has access to:
- `/dev/kfd` - AMD GPU kernel driver
- `/dev/dri` - Direct Rendering Infrastructure
- Groups: `video`, `render`

## Environment Variables

RX 6800 specific settings:
```yaml
HSA_OVERRIDE_GFX_VERSION=10.3.0  # Navi 21 (gfx1030) compatibility
ROCM_PATH=/opt/rocm
HIP_VISIBLE_DEVICES=0            # Use first AMD GPU
```

## Bot Integration

Your bot now has two endpoints available:

```python
import globals

# NVIDIA GPU (primary)
nvidia_url = globals.LLAMA_URL  # http://llama-swap:8080

# AMD GPU (secondary)
amd_url = globals.LLAMA_AMD_URL  # http://llama-swap-amd:8080
```

Use the `gpu_router` utility for automatic load balancing:

```python
from bot.utils.gpu_router import get_llama_url_with_load_balancing

# Round-robin between GPUs
url, model = get_llama_url_with_load_balancing(task_type="text")

# Prefer AMD for vision
url, model = get_llama_url_with_load_balancing(
    task_type="vision",
    prefer_amd=True
)
```

## Troubleshooting

If the AMD container fails to start:

1. **Check build logs:**
   ```bash
   docker compose build --no-cache llama-swap-amd
   ```

2. **Verify GPU access:**
   ```bash
   ls -l /dev/kfd /dev/dri
   ```

3. **Check container logs:**
   ```bash
   docker compose logs llama-swap-amd
   ```

4. **Test GPU from host:**
   ```bash
   lspci | grep -i amd
   # Should show: Radeon RX 6800
   ```

## Performance Notes

**RX 6800 Specs:**
- VRAM: 16GB
- Architecture: RDNA 2 (Navi 21)
- Compute: gfx1030

**Recommended Models:**
- Q4_K_M quantization: 5-6GB per model
- Can load 2-3 models simultaneously
- Good for: Llama 3.1 8B, DarkIdol 8B, Moondream2

## Future Improvements

1. **Automatic failover:** Route to AMD if NVIDIA is busy
2. **Health monitoring:** Track GPU utilization
3. **Dynamic routing:** Use least-busy GPU
4. **VRAM monitoring:** Alert before OOM
5. **Model preloading:** Keep common models loaded

## Resources

- [ROCm Documentation](https://rocmdocs.amd.com/)
- [llama.cpp ROCm Build](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md#rocm)
- [llama-swap GitHub](https://github.com/mostlygeek/llama-swap)
- [Full Setup Guide](./DUAL_GPU_SETUP.md)
- [Quick Reference](./DUAL_GPU_QUICK_REF.md)