model_executor/models/registry.py: the model registry

    @staticmethod
    def _get_cache_dir() -> Path:
        return Path(envs.VLLM_CACHE_ROOT) / "modelinfos"

    def _get_cache_filename(self) -> str:
        cls_name = f"{self.module_name}-{self.class_name}".replace(".", "-")
        return f"{cls_name}.json"

    @staticmethod
    def _get_modelinfo_module_hash(model_path: Path) -> str:
        if model_path.name == "__init__.py":
            # Package entry points often re-export classes implemented in
            # submodules, so include the package contents in the cache key.
            module_paths = sorted(model_path.parent.rglob("*.py"))
            root_path = model_path.parent
        else:
            module_paths = [model_path]
            root_path = model_path.parent

        hasher = safe_hash(b"", usedforsecurity=False)
        for path in module_paths:
            hasher.update(path.relative_to(root_path).as_posix().encode("utf-8"))
            hasher.update(b"\0")
            hasher.update(path.read_bytes())
            hasher.update(b"\0")
        return hasher.hexdigest()

    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
        try:
            try:
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
                logger.debug(
                    "Cached model info file for class %s.%s not found",
                    self.module_name,
                    self.class_name,
                )
                return None

            if mi_dict["hash"] != module_hash:
                logger.debug(
                    "Cached model info file for class %s.%s is stale",
                    self.module_name,
                    self.class_name,
                )
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
            logger.debug(
                "Cached model info for class %s.%s error. ",
                self.module_name,
                self.class_name,
            )
            return None

@lru_cache(maxsize=128)
def _try_load_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
) -> type[nn.Module] | None:
    from vllm.platforms import current_platform

    current_platform.verify_model_arch(model_arch)
    try:
        return model.load_model_cls()
    except Exception:
        logger.exception("Error in loading model architecture '%s'", model_arch)
        return None


@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
) -> _ModelInfo | None:
    try:
        return model.inspect_model_cls()
    except Exception:
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
        return None

    def resolve_model_cls(
        self,
        architectures: str | list[str],
        model_config: ModelConfig,
    ) -> tuple[type[nn.Module], str]:
        if isinstance(architectures, str):
            architectures = [architectures]
        if not architectures:
            raise ValueError("No model architectures are specified")

        # Require transformers impl
        if model_config.model_impl == "transformers":
            arch = self._try_resolve_transformers(architectures[0], model_config)
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
        elif model_config.model_impl == "terratorch":
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)

        # Fallback to transformers impl (after resolving convert_type)
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
            and getattr(model_config, "convert_type", "none") == "none"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

        for arch in architectures:
            normalized_arch = self._normalize_arch(arch, model_config)
            model_cls = self._try_load_model_cls(normalized_arch)
            if model_cls is not None:
                return (model_cls, arch)

        # Fallback to transformers impl (before resolving runner_type)
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

        return self._raise_for_unsupported(architectures)

    def _raise_for_unsupported(self, architectures: list[str]):
        all_supported_archs = self.get_supported_archs()

        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
                "to be inspected. Please check the logs for more details."
            )

        for arch in architectures:
            if arch in _PREVIOUSLY_SUPPORTED_MODELS:
                previous_version = _PREVIOUSLY_SUPPORTED_MODELS[arch]

                raise ValueError(
                    f"Model architecture {arch} was supported in vLLM until "
                    f"v{previous_version}, and is not supported anymore. "
                    "Please use an older version of vLLM if you want to "
                    "use this model architecture."
                )
            if arch in _OOT_SUPPORTED_MODELS:
                plugin_url = _OOT_SUPPORTED_MODELS[arch]

                raise ValueError(
                    f"Model architecture {arch} is not supported in-tree anymore. "
                    f"Please install the plugin at {plugin_url} if you want to "
                    "use this model architecture."
                )

        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}"
        )