# GLaDOS TTS - ROCm Deployment (AMD GPU)

This guide covers running the Wyoming TTS server with AMD GPU acceleration via ROCm.

## Prerequisites

- Linux host with an AMD GPU supported by ROCm
- ROCm kernel drivers installed (`amdgpu` module loaded)
- Docker with `amdgpu` device support

Verify ROCm is available on the host:

```bash
ls /dev/kfd /dev/dri
rocm-smi  # if ROCm tools are installed
```

## Running with docker-compose

```bash
docker compose up glados-tts-rocm
```

This uses the `Dockerfile.rocm` image with:
- PyTorch ROCm backend (device mapped to `cuda` internally)
- Half-precision (`--half`) enabled for ~2x speedup
- Model pre-loaded at startup (`--preload`)
- MIOpen convolution solver caching for faster repeated inference

## Environment Variables

| Variable | Default | Description |
|----------|---------|-------------|
| `DEVICE` | `rocm` | Set to `rocm` for AMD GPU |
| `HALF` | `true` | Use float16 inference |
| `PRELOAD` | `true` | Load model at startup |
| `MODEL_DIR` | `/data` | Model storage directory |
| `URI` | `tcp://0.0.0.0:10200` | Wyoming server URI |

ROCm-specific optimizations are applied automatically:
- `MIOPEN_FIND_MODE=1` - cache convolution solver solutions
- `MIOPEN_DEBUG_CONV_GEMM=0` - default GEMM mode

## Device Mapping

The server maps `rocm` to `cuda` internally since PyTorch ROCm uses the CUDA API compatibility layer. No code changes are needed.

## Troubleshooting

**No GPU detected in container logs**:
```bash
# Verify host has ROCm devices
ls -la /dev/kfd /dev/dri

# Check if amdgpu module is loaded
lsmod | grep amdgpu
```

**Out of memory**: Reduce `HALF=false` or ensure sufficient GPU VRAM (model requires ~4GB).