Load the model format, and do the following: 1. Load the build_config if got an engine. 2. Load the parallel_config if got a checkpoint.
(self)
| 2615 | |
| 2616 | @model_validator(mode="after") |
| 2617 | def validate_model_format_misc(self): |
| 2618 | ''' |
| 2619 | Load the model format, and do the following: |
| 2620 | |
| 2621 | 1. Load the build_config if got an engine. |
| 2622 | 2. Load the parallel_config if got a checkpoint. |
| 2623 | ''' |
| 2624 | model_obj = _ModelWrapper(self.model) |
| 2625 | |
| 2626 | if model_obj.is_local_model and self.backend not in [ |
| 2627 | 'pytorch', '_autodeploy' |
| 2628 | ]: |
| 2629 | # Load parallel_config from the engine. |
| 2630 | model_format = get_model_format( |
| 2631 | self.model, trust_remote_code=self.trust_remote_code) |
| 2632 | |
| 2633 | if model_format is _ModelFormatKind.TLLM_ENGINE: |
| 2634 | if self.build_config is not None: |
| 2635 | logger.warning( |
| 2636 | "The build_config is ignored for model format of TLLM_ENGINE." |
| 2637 | ) |
| 2638 | self._load_config_from_engine(model_obj.model_dir) |
| 2639 | runtime_defaults = self._pretrained_config.runtime_defaults |
| 2640 | if runtime_defaults: |
| 2641 | self.kv_cache_config.fill_empty_fields_from_runtime_defaults( |
| 2642 | runtime_defaults) |
| 2643 | |
| 2644 | # Load parallel_config from the checkpoint. |
| 2645 | elif model_format is _ModelFormatKind.TLLM_CKPT: |
| 2646 | # We need to create a temporary instance to call _load_config_from_ckpt |
| 2647 | self._load_config_from_ckpt(model_obj.model_dir) |
| 2648 | else: |
| 2649 | model_format = _ModelFormatKind.HF |
| 2650 | |
| 2651 | # Store the model format in the values |
| 2652 | self._model_format = model_format |
| 2653 | return self |
| 2654 | |
| 2655 | @field_validator('calib_config', mode='before') |
| 2656 | @classmethod |
nothing calls this directly
no test coverage detected