Model capability protocols (SupportsPP, SupportsLoRA, ...)

def _check_vllm_model_init(model: type[object] | object) -> bool:
    model_init = getattr(model, "__init__")  # noqa: B009
    return supports_kw(model_init, "vllm_config")


def _check_vllm_model_embed_input_ids(model: type[object] | object) -> bool:
    model_embed_input_ids = getattr(model, "embed_input_ids", None)
    if not callable(model_embed_input_ids):
        logger.warning(
            "The model (%s) is missing the `embed_input_ids` method.",
            model,
        )
        return False

    return True


def _check_vllm_model_forward(model: type[object] | object) -> bool:
    model_forward = getattr(model, "forward", None)
    if not callable(model_forward):
        return False

    vllm_kws = ("input_ids", "positions")
    missing_kws = tuple(kw for kw in vllm_kws if not supports_kw(model_forward, kw))

    if missing_kws and (isinstance(model, type) and issubclass(model, nn.Module)):
        logger.warning(
            "The model (%s) is missing "
            "vLLM-specific keywords from its `forward` method: %s",
            model,
            missing_kws,
        )

    return len(missing_kws) == 0


@overload
def is_vllm_model(model: type[object]) -> TypeIs[type[VllmModel]]: ...


@overload
def is_vllm_model(model: object) -> TypeIs[VllmModel]: ...


def is_vllm_model(
    model: type[object] | object,
) -> TypeIs[type[VllmModel]] | TypeIs[VllmModel]:
    return (
        _check_vllm_model_init(model)
        and _check_vllm_model_embed_input_ids(model)
        and _check_vllm_model_forward(model)
    )

@runtime_checkable
class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
    """The interface required for all multi-modal models."""

    supports_multimodal: ClassVar[Literal[True]] = True
    """
    A flag that indicates this model supports multi-modal inputs.

    Note:
        There is no need to redefine this flag if this class is in the
        MRO of your model class.
    """

    supports_multimodal_raw_input_only: ClassVar[bool] = False

def supports_multimodal(
    model: type[object] | object,
) -> TypeIs[type[SupportsMultiModal]] | TypeIs[SupportsMultiModal]:
    return getattr(model, "supports_multimodal", False)

def supports_eagle3(
    model: type[object] | object,
) -> TypeIs[type[SupportsEagle3]] | TypeIs[SupportsEagle3]:
    return isinstance(model, SupportsEagle3)

def supports_lora(
    model: type[object] | object,
) -> TypeIs[type[SupportsLoRA]] | TypeIs[SupportsLoRA]:
    result = _supports_lora(model)

    if not result:
        lora_attrs = (
            "packed_modules_mapping",
            "embedding_modules",
        )
        missing_attrs = tuple(attr for attr in lora_attrs if not hasattr(model, attr))

        if getattr(model, "supports_lora", False):
            if missing_attrs:
                logger.warning(
                    "The model (%s) sets `supports_lora=True`, "
                    "but is missing LoRA-specific attributes: %s",
                    model,
                    missing_attrs,
                )
        else:
            if not missing_attrs:
                logger.warning(
                    "The model (%s) contains all LoRA-specific attributes, "
                    "but does not set `supports_lora=True`.",
                    model,
                )

    return result


def _supports_lora(model: type[object] | object) -> bool:
    if isinstance(model, type):
        return isinstance(model, _SupportsLoRAType)

    return isinstance(model, SupportsLoRA)

def supports_pp(
    model: type[object] | object,
) -> bool | TypeIs[type[SupportsPP]] | TypeIs[SupportsPP]:
    supports_attributes = _supports_pp_attributes(model)
    supports_inspect = _supports_pp_inspect(model)

    if supports_attributes and not supports_inspect:
        logger.warning(
            "The model (%s) sets `supports_pp=True`, but does not accept "
            "`intermediate_tensors` in its `forward` method",
            model,
        )

    if not supports_attributes:
        pp_attrs = ("make_empty_intermediate_tensors",)
        missing_attrs = tuple(attr for attr in pp_attrs if not hasattr(model, attr))

        if getattr(model, "supports_pp", False):
            if missing_attrs:
                logger.warning(
                    "The model (%s) sets `supports_pp=True`, "
                    "but is missing PP-specific attributes: %s",
                    model,
                    missing_attrs,
                )
        else:
            if not missing_attrs:
                logger.warning(
                    "The model (%s) contains all PP-specific attributes, "
                    "but does not set `supports_pp=True`.",
                    model,
                )

    return supports_attributes and supports_inspect


def _supports_pp_attributes(model: type[object] | object) -> bool:
    if isinstance(model, type):
        return SupportsPP in model.__mro__ or isinstance(model, _SupportsPPType)

    return isinstance(model, SupportsPP)


def _supports_pp_inspect(model: type[object] | object) -> bool:
    model_forward = getattr(model, "forward", None)
    if not callable(model_forward):
        return False

    return supports_kw(model_forward, "intermediate_tensors")

    @staticmethod
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
        return _ModelInfo(
            architecture=model.__name__,
            is_text_generation_model=is_text_generation_model(model),
            is_pooling_model=is_pooling_model(model),
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
            attn_type=get_attn_type(model),
            score_type=get_score_type(model),
            supports_multimodal=supports_multimodal(model),
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
            requires_raw_input_tokens=requires_raw_input_tokens(model),
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
            supports_pp=supports_pp(model),
            supports_lora=supports_lora(model),
            supports_hc_sp=supports_hc_sp(model),
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
            is_hybrid=is_hybrid(model),
            supports_replayssm=supports_replayssm(model),
            supports_transcription=supports_transcription(model),
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
            supports_kv_cache_grouping=supports_kv_cache_grouping(model),
            has_noops=has_noops(model),
            supported_video_pruning_methods=getattr(
                model, "supported_video_pruning_methods", ()
            ),
            supports_mm_device_do_normalize=getattr(
                model, "supports_mm_device_do_normalize", False
            ),
        )