platform-detect

name

platform-detect

description

Detect the user's hardware platform and recommend optimal GAFIME engine configuration. Use when the user asks about hardware compatibility, which backend to use, GPU detection, VRAM availability, CUDA version, Apple Silicon support, or wants to know the best EngineConfig for their system. Also use when the user says things like "what GPU do I have", "will GAFIME work on my machine", "detect my hardware", or "configure for my system".

Platform Detection

Detect the user's compute hardware and generate an optimal GAFIME EngineConfig.

Instructions

1. Run the detection script:

   python .claude/skills/platform-detect/scripts/platform_detect.py
   

2. Read the JSON output. It reports:

   • OS and architecture
   • CUDA availability, GPU name, VRAM, compute capability
   • Apple Silicon / Metal availability
   • CPU core count and OpenMP support
   • Installed GAFIME native backends (CUDA when present, C++ core)
   • Installed GAFIME vendor payload packages (gafime-cuda, gafime-rocm)
   • Rust helper alias availability (from gafime import subfunctions)
   • Recommended backend and vram_budget_mb
   • Recommended discrete mode for the selected backend

3. Present findings to the user in a clear summary table.

4. Generate a ready-to-use Python snippet with the recommended EngineConfig:

   from gafime import GafimeEngine, EngineConfig, ComputeBudget
   
   config = EngineConfig(
       budget=ComputeBudget(
           vram_budget_mb=<detected_value>,
           keep_in_vram=True,
           max_discrete_candidates=100_000,
       ),
       backend="<detected_backend>",
       device_id=<detected_device>,
       enable_discrete_functions=True,
       discrete_mode="soft",
   )
   engine = GafimeEngine(config)
   

5. If no GPU payload is installed, recommend backend="core" and give the matching payload command if hardware is visible:

   • NVIDIA: pip install "gafime[cuda]"
   • AMD ROCm/HIP on Linux x86_64: pip install "gafime[rocm]"

Troubleshooting

• If the script fails with ModuleNotFoundError, GAFIME is not installed. Guide the user to pip install gafime.
• If NVIDIA hardware is detected but the CUDA payload is missing, suggest pip install "gafime[cuda]".
• If AMD ROCm/HIP hardware is detected on Linux x86_64 but the ROCm payload is missing, suggest pip install "gafime[rocm]".
• If AMD hardware is detected on Windows, recommend backend="core" for v0.4.7 because ROCm payload wheels are Linux-only in this release.
• On macOS, backend="auto" should prefer Metal and then C++ Core. CUDA is an invalid macOS backend.
• On Linux/Windows x86_64, backend="auto" should prefer installed CUDA payloads and then C++ Core. On Linux x86_64 it may also prefer an installed ROCm payload and then C++ Core.
• On Linux/Windows ARM64, recommend C++ Core. Current ARM wheels do not expose the CUDA backend.
• backend="gpu" is deprecated; recommend auto, cuda, metal, or core.
• On GPU backends, discrete feature engineering must use discrete_mode="soft". Hard mode is C++ Core only.
• If subfunctions is unavailable, rebuild or reinstall the native Rust helper extension.

Example

User says: "What hardware can GAFIME use on my machine?"

Actions: Run platform_detect.py, parse output.

Result: "You have an NVIDIA RTX 4060 with 8GB VRAM (CUDA 12.4, SM 8.9). I recommend using the CUDA backend with vram_budget_mb=6144 (leaving 2GB headroom). Here's your config: ..."