GAMMA Engine Compatibility

Use before running workloads where engine choice affects correctness or logits access.

Rules

• game, comparison, and mind-meld require raw logits.
• Logits-capable engines: pytorch, pytorch_cuda, llamacpp, mlx, mlx_gpu, vllm.
• Wrapper engines without raw logits: openai, huggingface_inference, ollama.
• In GAMMA, vllm is CUDA-oriented; AMD uses PyTorch ROCm or CPU fallback.

Workflow

1. Detect hardware and available models.
2. Check whether the mode needs logits, attention, KV cache, or translation.
3. Validate model specs with the script when possible.
4. Pick the simplest engine that satisfies the mode.
5. Run a smoke command and record fallbacks.