| """Model-dependent defaults for the quantization form.""" | |
| import logging | |
| import gradio as gr | |
| from src.tools.model_size import get_model_size_b | |
| logger = logging.getLogger(__name__) | |
| def default_quant_method(model, oauth_token: gr.OAuthToken | None = None): | |
| """Use total parameters (including MoE experts), with Tuning as fallback.""" | |
| size = None | |
| if model and model.strip(): | |
| try: | |
| size = get_model_size_b(model.strip(), token=oauth_token.token if oauth_token else None) | |
| except Exception: | |
| logger.warning("Could not determine model size for the quantization default") | |
| return gr.update(value=( | |
| "RTN (Round-To-Nearest)" if size is not None and size > 100 | |
| else "Tuning (AutoRound, recommended)" | |
| )) | |