# Platform Detection Helper
## Overview
`detect_platform.py` provides unified hardware and OS detection for bash scripts, using the same `gpu_detection.py` module as `install.py` for consistency.
## Usage
```bash
# Run detection
python scripts/utils/detect_platform.py
# Output (JSON):
{
"os": "darwin",
"arch": "arm64",
"is_arm": true,
"is_x86": false,
"accelerator": "mps",
"has_cuda": false,
"has_rocm": false,
"has_mps": true,
"has_directml": false,
"cuda_version": null,
"rocm_version": null,
"directml_version": null,
"pytorch_index_url": "",
"needs_directml": false
}
```
## Supported Platforms
| Platform | Detection | PyTorch Index |
|----------|-----------|---------------|
| **Apple Silicon (M1/M2/M3)** | MPS via system_profiler | Default PyPI (MPS built-in) |
| **NVIDIA GPU (CUDA)** | nvcc version | cu121/cu118/cu102 |
| **AMD GPU (ROCm)** | rocminfo | rocm5.6 |
| **Windows DirectML** | torch-directml import | CPU + directml package |
| **CPU-only** | Fallback | CPU index |
## Integration with update_and_restart.sh
The script automatically:
1. Detects hardware platform (MPS/CUDA/ROCm/DirectML/CPU)
2. Selects optimal PyTorch index URL
3. Installs DirectML package if needed (Windows)
4. Falls back to basic detection if Python helper unavailable
## Benefits vs. Old Logic
**Old (Bash-only):**
- ❌ Only detected macOS vs. Linux with nvidia-smi
- ❌ Treated all macOS as CPU-only (performance loss on M-series)
- ❌ No ROCm, DirectML, or MPS support
**New (Python-based):**
- ✅ Detects MPS, CUDA, ROCm, DirectML, CPU
- ✅ Consistent with install.py logic
- ✅ Optimal PyTorch selection per platform
- ✅ Graceful fallback to old logic if detection fails
## Example Output (macOS M2)
```bash
▶ Installing dependencies (editable mode)...
ℹ Existing venv Python version: 3.13
ℹ Installing with venv pip (this may take 1-2 minutes)...
ℹ Apple Silicon MPS detected - using MPS-optimized PyTorch
```
## Example Output (Linux with NVIDIA)
```bash
▶ Installing dependencies (editable mode)...
ℹ CUDA detected (12.1) - using optimized PyTorch
Installing with: --extra-index-url https://download.pytorch.org/whl/cu121
```
## Maintenance
The detection logic is centralized in `src/mcp_memory_service/utils/gpu_detection.py`. Updates to that module automatically benefit both `install.py` and `update_and_restart.sh`.
