- Add offload_strategy to kwargs in _load_default_model and _load_model_by_name
- Fix parameter name: ram -> manual_ram_gb to match backend expectation
- Also pass load_in_4bit, load_in_8bit, and max_gpu_percent