Function should_use_mlx

optillm/inference.py:234–269

Determine if a model should use MLX instead of PyTorch

(model_id: str)

    return platform.system() == "Darwin" and platform.machine() == "arm64"

def should_use_mlx(model_id: str) -> bool:
    """Determine if a model should use MLX instead of PyTorch"""
    if not MLX_AVAILABLE or not is_apple_silicon():
        return False
    
    # Models that should use MLX
    mlx_patterns = [
        "mlx-community/",
        "mlx-",
        "-mlx-"
    ]
    
    # Known problematic models that should prefer MLX on Apple Silicon
    problematic_models = [
        "Qwen/Qwen3-",
        "google/gemma-3-",
        "google/gemma3-"
    ]
    
    model_lower = model_id.lower()
    
    # Direct MLX model detection
    for pattern in mlx_patterns:
        if pattern.lower() in model_lower:
            return True
    
    # Problematic model detection
    for pattern in problematic_models:
        if pattern.lower() in model_lower:
            logger.warning(f"Model {model_id} detected as potentially problematic with MPS backend")
            suggested_mlx = suggest_mlx_alternative(model_id)
            logger.warning(f"Consider using MLX model: {suggested_mlx}")
            # Don't auto-switch, but recommend
            return False
    
    return False

def suggest_mlx_alternative(model_id: str) -> str:
    """Suggest MLX alternative for a given model"""