Method from_pretrained

lmdeploy/pytorch/config.py:382–445

Instantiate one of the configuration classes of the library from a pretrained model configuration. Args: pretrained_model_name_or_path (str): the pretrained model path trust_remote_code (bool): Whether or not to allow for custom models define

(
        cls,
        pretrained_model_name_or_path: str,
        trust_remote_code: bool = False,
        dtype: str = 'auto',
        dist_config: DistConfig = None,
        hf_overrides: dict[str, Any] = None,
        is_draft_model: bool = False,
        spec_method: str = None,
        num_spec_tokens: int = 0,
        model_format: str = None,
        device_type: str = 'auto',
        block_size: int = 64,
    )

    @classmethod
    def from_pretrained(
        cls,
        pretrained_model_name_or_path: str,
        trust_remote_code: bool = False,
        dtype: str = 'auto',
        dist_config: DistConfig = None,
        hf_overrides: dict[str, Any] = None,
        is_draft_model: bool = False,
        spec_method: str = None,
        num_spec_tokens: int = 0,
        model_format: str = None,
        device_type: str = 'auto',
        block_size: int = 64,
    ):
        """Instantiate one of the configuration classes of the library from a
        pretrained model configuration.

        Args:
            pretrained_model_name_or_path (str): the pretrained model path
            trust_remote_code (bool):  Whether or not to allow for custom
                models defined on the Hub in their own modeling files.
            dtype (str): user specified data type for model weights and
                activations. Refer to `PyTorchEngineConfig` for details
            hf_overrides (dict[str, Any]): overrides for the HF config.
            model_format (str): the quantization format of the model.
        """
        from transformers import AutoConfig

        from lmdeploy.pytorch.transformers import config_from_pretrained
        hf_config = config_from_pretrained(pretrained_model_name_or_path, trust_remote_code=trust_remote_code)
        if getattr(hf_config, 'model_type', None) in ['phi3']:
            # phi3 + trust_remote_code leads to error when tp.
            hf_config = AutoConfig.from_pretrained(pretrained_model_name_or_path)

        # update quantization config
        hf_config = _patch_quantization_config(hf_config, model_format=model_format)

        model_config = cls.from_hf_config(
            hf_config,
            pretrained_model_name_or_path,
            dtype=dtype,
            dist_config=dist_config,
            is_draft_model=is_draft_model,
            spec_method=spec_method,
            num_spec_tokens=num_spec_tokens,
            device_type=device_type,
        )
        fp32_lm_head = False
        if hf_overrides is not None:
            logger.warning(f'Overriding HF config with {hf_overrides}')
            fp32_lm_head = hf_overrides.pop('fp32_lm_head', False)
            override_hf_config(model_config.hf_config, hf_overrides)

        # for fp32 head
        model_config.fp32_lm_head = fp32_lm_head
        model_config.tie_word_embeddings = getattr(hf_config, 'tie_word_embeddings', False)

        # for serialization of transformers modules