vllm.model_executor.layers.quantization.torchao ¶

class TorchAOConfig(QuantizationConfig):
    """Config class for torchao."""

    def __init__(
        self,
        torchao_config,
        skip_modules: Optional[list[str]] = None,
        is_checkpoint_torchao_serialized: bool = False,
    ) -> None:
        """
        # TorchAO quantization relies on tensor subclasses. In order,
        # to enable proper caching this needs standalone compile
        if is_torch_equal_or_newer("2.8.0.dev"):
            os.environ["VLLM_TEST_STANDALONE_COMPILE"] = "1"
            logger.info(
                "Using TorchAO: Setting VLLM_TEST_STANDALONE_COMPILE=1")

        # TODO: remove after the torch dependency is updated to 2.8
        if is_torch_equal_or_newer(
                "2.7.0") and not is_torch_equal_or_newer("2.8.0.dev"):
            os.environ["VLLM_DISABLE_COMPILE_CACHE"] = "1"
            logger.info("Using TorchAO: Setting VLLM_DISABLE_COMPILE_CACHE=1")
        """
        super().__init__()
        self.torchao_config = torchao_config
        self.skip_modules = skip_modules or []
        self.is_checkpoint_torchao_serialized = is_checkpoint_torchao_serialized

    def __repr__(self) -> str:
        return (
            f"TorchAOConfig({self.torchao_config=}, {self.skip_modules=}, "
            f"{self.is_checkpoint_torchao_serialized=})"
        )

    def get_name(self) -> QuantizationMethods:
        return "torchao"

    def get_supported_act_dtypes(self) -> list[torch.dtype]:
        return [torch.float32, torch.float16, torch.bfloat16]

    @classmethod
    def get_min_capability(cls) -> int:
        return 75

    @staticmethod
    def get_config_filenames() -> list[str]:
        """torchao doesn't require additional config files, we use
        `config.json` from huggingface: `model_config.hf_config`
        """
        return []

    @classmethod
    def from_config(cls, config: dict[str, Any]) -> "TorchAOConfig":
        """Create the quant config from an hf model config"""
        try:
            from torchao.core.config import config_from_dict
        except ImportError as err:
            raise ImportError(
                "Please install torchao>=0.10.0 via "
                "`pip install torchao>=0.10.0` to use torchao quantization."
            ) from err

        quant_method = cls.get_from_keys_or(config, ["quant_method"], None)
        is_checkpoint_torchao_serialized = (
            quant_method is not None and "torchao" in quant_method
        )

        hf_config = cls.get_from_keys_or(config, ["quant_type"], None)
        assert hf_config is not None, "quant_type must be specified"
        assert len(hf_config) == 1 and "default" in hf_config, (
            "Expected only one key 'default' in quant_type dictionary"
        )
        quant_type = hf_config["default"]
        ao_config = config_from_dict(quant_type)

        # Adds skipped modules defined in "modules_to_not_convert"
        skip_modules = config.get("modules_to_not_convert", []) or []

        # Adds skipped modules defined in "module_fqn_to_config"
        _data = quant_type.get("_data", {})
        if not isinstance(_data, dict):
            _data = {}

        module_fqn = _data.get("module_fqn_to_config", {})
        if not isinstance(module_fqn, dict):
            module_fqn = {}

        for layer, layer_cfg in module_fqn.items():
            if layer_cfg is None:
                skip_modules.append(layer)

        return cls(ao_config, skip_modules, is_checkpoint_torchao_serialized)

    @classmethod
    def from_config_file(cls, config_file: str) -> "TorchAOConfig":
        """Initialize class from a config file. Example:
        ```
        config = Float8DynamicActivationFloat8WeightConfig(granularity=PerRow())
        fn = "torchao_config.json"

        with open(fn, "w") as f:
            f.write(json.dumps(config_to_dict(config)))
        ```
        """
        with open(config_file) as f:
            f.seek(0)
            f_read = f.read()
            config_dict = json.loads(f_read)

        hf_config = {"quant_type": {"default": config_dict}}
        return cls.from_config(hf_config)

    @classmethod
    def from_config_dict_json(cls, config_dict_json: str) -> "TorchAOConfig":
        """Iniitalize class from a config_dict json string, got from
        torchao_config_object = some AOBaseConfig object
        json.dumps(config_to_dict(torchao_config_object))
        """
        config_dict = json.loads(config_dict_json)
        hf_config = {"quant_type": {"default": config_dict}}
        return cls.from_config(hf_config)

    def get_quant_method(
        self, layer: torch.nn.Module, prefix: str
    ) -> Optional["QuantizeMethodBase"]:
        if not isinstance(layer, LinearBase):
            return None

        from torchao.quantization import ModuleFqnToConfig

        if should_skip(prefix, self.skip_modules):
            return UnquantizedLinearMethod()

        module_fqn = prefix
        if isinstance(self.torchao_config, ModuleFqnToConfig):
            module_fqn_to_config = self.torchao_config.module_fqn_to_config
            c = module_fqn_to_config.get(module_fqn) or module_fqn_to_config.get(
                "_default", None
            )
            if c is not None:
                current_torchao_config = TorchAOConfig(
                    c, self.skip_modules, self.is_checkpoint_torchao_serialized
                )
                return TorchAOLinearMethod(current_torchao_config)
            else:
                return UnquantizedLinearMethod()

        return TorchAOLinearMethod(self)

    def get_scaled_act_names(self) -> list[str]:
        return []

class TorchAOLinearMethod(LinearMethodBase):
    """Linear method for torchao.

    Args:
        quant_config: The torchao quantization config, a string that encodes
            the type of quantization and all relevant arguments.
    """

    def __init__(self, quant_config: TorchAOConfig):
        self.quant_config = quant_config

    def create_weights(
        self,
        layer: torch.nn.Module,
        input_size_per_partition: int,
        output_partition_sizes: list[int],
        input_size: int,
        output_size: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
        weight = Parameter(
            torch.empty(
                sum(output_partition_sizes),
                input_size_per_partition,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        if self.quant_config.is_checkpoint_torchao_serialized:
            weight = torchao_quantize_param_data(
                weight, self.quant_config.torchao_config
            )

        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})

        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        return F.linear(x, layer.weight, bias)

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        if self.quant_config.is_checkpoint_torchao_serialized:
            return

        # quantize the weight on the fly if the checkpoint is not already
        # quantized by torchao
        weight = torchao_quantize_param_data(
            layer.weight, self.quant_config.torchao_config
        )
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
        layer.register_parameter("weight", weight)