vllm.model_executor.layers.quantization.awq_marlin ¶

class AWQMarlinConfig(QuantizationConfig):
    """Config class for AWQ Marlin"""

    # num_bits -> type
    TYPE_MAP = {
        4: scalar_types.uint4,
    }

    def __init__(
        self,
        weight_bits: int,
        group_size: int,
        zero_point: bool,
        lm_head_quantized: bool,
        modules_to_not_convert: list[str] | None,
        full_config: dict[str, Any],
    ) -> None:
        super().__init__()
        self.pack_factor = 32 // weight_bits  # packed into int32
        self.group_size = group_size
        self.zero_point = zero_point
        self.lm_head_quantized = lm_head_quantized
        self.weight_bits = weight_bits
        self.modules_to_not_convert = modules_to_not_convert or []
        self.full_config = full_config

        if self.weight_bits not in self.TYPE_MAP:
            raise ValueError(
                f"Unsupported num_bits = {self.weight_bits}. "
                f"Supported num_bits = {self.TYPE_MAP.keys()}"
            )

        self.quant_type = self.TYPE_MAP[self.weight_bits]

        verify_marlin_supported(
            self.quant_type, group_size=self.group_size, has_zp=self.zero_point
        )

    def __repr__(self) -> str:
        return (
            f"AWQMarlinConfig(quant_type={self.quant_type}, "
            f"group_size={self.group_size}, "
            f"zero_point={self.zero_point}, "
            f"lm_head_quantized={self.lm_head_quantized}, "
            f"modules_to_not_convert={self.modules_to_not_convert})"
        )

    @classmethod
    def get_name(cls) -> "QuantizationMethods":
        return "awq_marlin"

    @classmethod
    def get_supported_act_dtypes(cls) -> list[torch.dtype]:
        return [torch.half, torch.bfloat16]

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

    @classmethod
    def get_config_filenames(cls) -> list[str]:
        return ["quantize_config.json"]

    @classmethod
    def from_config(cls, config: dict[str, Any]) -> "AWQMarlinConfig":
        weight_bits = cls.get_from_keys(config, ["bits"])
        group_size = cls.get_from_keys(config, ["group_size"])
        zero_point = cls.get_from_keys(config, ["zero_point"])
        lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
        modules_to_not_convert = cls.get_from_keys_or(
            config, ["modules_to_not_convert"], None
        )
        return cls(
            weight_bits,
            group_size,
            zero_point,
            lm_head_quantized,
            modules_to_not_convert,
            config,
        )

    @classmethod
    def override_quantization_method(
        cls, hf_quant_cfg, user_quant
    ) -> "QuantizationMethods | None":
        can_convert = cls.is_awq_marlin_compatible(hf_quant_cfg)
        is_valid_user_quant = (
            user_quant is None or user_quant == "marlin" or user_quant == "awq_marlin"
        )

        if can_convert and is_valid_user_quant:
            msg = (
                "The model is convertible to {} during runtime."
                " Using {} kernel.".format(cls.get_name(), cls.get_name())
            )
            logger.info(msg)
            return cls.get_name()

        if can_convert and user_quant == "awq":
            logger.info(
                "Detected that the model can run with awq_marlin"
                ", however you specified quantization=awq explicitly,"
                " so forcing awq. Use quantization=awq_marlin for"
                " faster inference"
            )
        return None

    def get_quant_method(
        self, layer: torch.nn.Module, prefix: str
    ) -> "QuantizeMethodBase | None":
        if isinstance(layer, LinearBase) or (
            isinstance(layer, ParallelLMHead) and self.lm_head_quantized
        ):
            if is_layer_skipped(
                prefix,
                self.modules_to_not_convert,
                self.packed_modules_mapping,
                skip_with_substr=True,
            ):
                return UnquantizedLinearMethod()
            # Check if the layer is supported by AWQMarlin.
            if not check_marlin_supports_layer(layer, self.group_size):
                logger.warning_once(
                    "Layer '%s' is not supported by AWQMarlin. Falling back to unoptimized AWQ kernels.",  # noqa: E501
                    prefix,
                )
                return AWQConfig.from_config(self.full_config).get_quant_method(
                    layer, prefix
                )
            quant_method = AWQMarlinLinearMethod(self)
            quant_method.input_dtype = get_marlin_input_dtype(prefix)
            return quant_method
        elif isinstance(layer, FusedMoE):
            from vllm.model_executor.layers.quantization.moe_wna16 import MoeWNA16Config

            if is_layer_skipped(
                prefix,
                getattr(self, "modules_to_not_convert", []),
                skip_with_substr=True,
            ):
                return UnquantizedFusedMoEMethod(layer.moe_config)
            if not check_moe_marlin_supports_layer(layer, self.group_size):
                logger.warning_once(
                    f"Layer '{prefix}' is not supported by AWQMoeMarlin. "
                    "Falling back to Moe WNA16 kernels."
                )
                return MoeWNA16Config.from_config(self.full_config).get_quant_method(
                    layer, prefix
                )
            moe_quant_method = AWQMarlinMoEMethod(self, layer.moe_config)
            moe_quant_method.input_dtype = get_marlin_input_dtype(prefix)
            return moe_quant_method
        return None

    @classmethod
    def is_awq_marlin_compatible(cls, quant_config: dict[str, Any]):
        # Extract data from quant config.
        quant_method = quant_config.get("quant_method", "").lower()
        num_bits = quant_config.get("bits")
        group_size = quant_config.get("group_size")
        zero_point = quant_config.get("zero_point")

        if not (current_platform.is_cuda_alike() or current_platform.is_cpu()):
            return False

        if quant_method != "awq":
            return False

        # If we cannot find the info needed in the config, cannot convert.
        if num_bits is None or group_size is None or zero_point is None:
            return False

        if num_bits not in cls.TYPE_MAP:
            return False

        return check_marlin_supported(
            quant_type=cls.TYPE_MAP[num_bits], group_size=group_size, has_zp=zero_point
        )

    def apply_vllm_mapper(self, hf_to_vllm_mapper: "WeightsMapper"):
        if self.modules_to_not_convert:
            self.modules_to_not_convert = hf_to_vllm_mapper.apply_list(
                self.modules_to_not_convert
            )

    def maybe_update_config(self, model_name: str, revision: str | None = None):
        if self.modules_to_not_convert:
            return

        unquant_dtypes = [torch.float16, torch.bfloat16, torch.float32]
        metadata = get_safetensors_params_metadata(model_name, revision=revision)
        layers = {param_name.rsplit(".", 1)[0] for param_name in metadata}
        quant_layers: set[str] = {
            param_name.rsplit(".", 1)[0]
            for param_name, info in metadata.items()
            if (dtype := info.get("dtype", None))
            and _SAFETENSORS_TO_TORCH_DTYPE[dtype] not in unquant_dtypes
        }
        self.modules_to_not_convert = list(layers - quant_layers)

class AWQMarlinLinearMethod(LinearMethodBase):
    """Linear method for AWQ Marlin.

    Uses choose_mp_linear_kernel to select the best available kernel
    (Conch, Exllama, or Marlin) for the current platform.

    Args:
        quant_config: The AWQ Marlin quantization config.
    """

    _kernel_backends_being_used: set[str] = set()

    def __init__(self, quant_config: AWQMarlinConfig) -> None:
        self.quant_config = quant_config
        self.quant_type = scalar_types.uint4
        self.input_dtype = None

        verify_marlin_supported(
            quant_type=self.quant_config.quant_type,
            group_size=self.quant_config.group_size,
            has_zp=self.quant_config.zero_point,
        )

    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,
    ) -> None:
        output_size_per_partition = sum(output_partition_sizes)
        weight_loader = extra_weight_attrs.get("weight_loader")

        if self.quant_config.group_size != -1:
            group_size = self.quant_config.group_size
        else:
            group_size = input_size

        mp_linear_kernel_config = MPLinearLayerConfig(
            full_weight_shape=(input_size, output_size),
            partition_weight_shape=(
                input_size_per_partition,
                output_size_per_partition,
            ),
            weight_type=self.quant_config.quant_type,
            act_type=params_dtype if self.input_dtype is None else self.input_dtype,
            group_size=self.quant_config.group_size,
            zero_points=self.quant_config.zero_point,
            has_g_idx=False,
        )

        kernel_type = choose_mp_linear_kernel(mp_linear_kernel_config)

        if kernel_type.__name__ not in self._kernel_backends_being_used:
            logger.info("Using %s for AWQMarlinLinearMethod", kernel_type.__name__)
            self._kernel_backends_being_used.add(kernel_type.__name__)

        # Weights are loaded in AWQ checkpoint format (packed along output dim).
        # Conversion to GPTQ-like format happens in process_weights_after_loading.
        qweight = PackedvLLMParameter(
            data=torch.empty(
                input_size_per_partition,
                output_size_per_partition // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            input_dim=0,
            output_dim=1,
            packed_dim=1,
            packed_factor=self.quant_config.pack_factor,
            weight_loader=weight_loader,
        )

        num_groups = input_size_per_partition // group_size

        qzeros = PackedvLLMParameter(
            data=torch.empty(
                num_groups,
                output_size_per_partition // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            input_dim=0,
            output_dim=1,
            packed_dim=1,
            packed_factor=self.quant_config.pack_factor,
            weight_loader=weight_loader,
        )

        scales = GroupQuantScaleParameter(
            data=torch.empty(
                num_groups,
                output_size_per_partition,
                dtype=params_dtype,
            ),
            input_dim=0,
            output_dim=1,
            weight_loader=weight_loader,
        )

        layer.register_parameter("qweight", qweight)
        layer.register_parameter("qzeros", qzeros)
        layer.register_parameter("scales", scales)

        self.kernel = kernel_type(
            mp_linear_kernel_config,
            w_q_param_name="qweight",
            w_s_param_name="scales",
            w_zp_param_name="qzeros",
        )

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        # AWQ checkpoints use a non-standard packing order and pack qweight
        # along the output dimension. Convert to the standard format
        # (GPTQ-like: standard bit order, qweight packed along input dim)
        # before handing off to the kernel.
        _convert_awq_to_standard_format(
            layer, "qweight", "qzeros", self.quant_config.quant_type.size_bits
        )
        self.kernel.process_weights_after_loading(layer)

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        return self.kernel.apply_weights(layer, x, bias)

class AWQMarlinMoEMethod(FusedMoEMethodBase):
    def __init__(
        self,
        quant_config: AWQMarlinConfig,
        moe: FusedMoEConfig,
    ):
        super().__init__(moe)
        self.quant_config = quant_config
        if self.quant_config.weight_bits != 4:
            raise ValueError("AWQMarlinMoEMethod only supports 4bit now.")
        self.quant_type = scalar_types.uint4
        self.input_dtype = None
        self.use_marlin = True

    def create_weights(
        self,
        layer: torch.nn.Module,
        num_experts: int,
        hidden_size: int,
        intermediate_size_per_partition: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
        layer.input_dtype = self.input_dtype
        extra_weight_attrs.update(
            {
                "is_transposed": True,
                "quant_method": FusedMoeWeightScaleSupported.GROUP.value,
            }
        )

        intermediate_size_full = extra_weight_attrs.pop(
            "intermediate_size_full", intermediate_size_per_partition
        )
        self.is_k_full = intermediate_size_per_partition == intermediate_size_full

        w13_qweight = Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                2 * intermediate_size_per_partition // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_qweight", w13_qweight)
        set_weight_attrs(w13_qweight, extra_weight_attrs)

        w2_qweight = Parameter(
            torch.empty(
                num_experts,
                intermediate_size_per_partition,
                hidden_size // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_qweight", w2_qweight)
        set_weight_attrs(w2_qweight, extra_weight_attrs)

        num_groups_w13 = hidden_size // self.quant_config.group_size
        num_groups_w2 = intermediate_size_per_partition // self.quant_config.group_size
        layer.num_groups_w13 = num_groups_w13
        layer.num_groups_w2 = num_groups_w2

        # WEIGHT_SCALES
        # Allocate 2 scales for w1 and w3 respectively.
        w13_scales = Parameter(
            torch.empty(
                num_experts,
                num_groups_w13,
                intermediate_size_per_partition * 2,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_scales", w13_scales)
        set_weight_attrs(w13_scales, extra_weight_attrs)

        w2_scales = Parameter(
            torch.empty(num_experts, num_groups_w2, hidden_size, dtype=params_dtype),
            requires_grad=False,
        )
        layer.register_parameter("w2_scales", w2_scales)
        set_weight_attrs(w2_scales, extra_weight_attrs)

        # WEIGHT_ZERO_POINT
        # Allocate 2 zero points for w1 and w3 respectively.
        w13_qzeros = Parameter(
            torch.empty(
                num_experts,
                num_groups_w13,
                2 * intermediate_size_per_partition // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_qzeros", w13_qzeros)
        set_weight_attrs(w13_qzeros, extra_weight_attrs)

        w2_qzeros = Parameter(
            torch.empty(
                num_experts,
                num_groups_w2,
                hidden_size // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_qzeros", w2_qzeros)
        set_weight_attrs(w2_qzeros, extra_weight_attrs)

        device = layer.w13_qweight.device
        layer.workspace = marlin_make_workspace_new(device, 4)

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        num_experts = layer.w13_qweight.shape[0]
        device = layer.w13_qweight.device
        is_a_8bit = self.input_dtype is not None and self.input_dtype.itemsize == 1

        if self.input_dtype == torch.float8_e4m3fn:
            ops.marlin_int4_fp8_preprocess(
                layer.w13_qweight.view(-1, layer.w13_qweight.size(2)),
                layer.w13_qzeros.view(-1, layer.w13_qzeros.size(2)),
                inplace=True,
            )
            ops.marlin_int4_fp8_preprocess(
                layer.w2_qweight.view(-1, layer.w2_qweight.size(2)),
                layer.w2_qzeros.view(-1, layer.w2_qzeros.size(2)),
                inplace=True,
            )
            layer.w13_scales.data = layer.w13_scales.data * 512
            layer.w2_scales.data = layer.w2_scales.data * 512

        layer.w13_g_idx_sort_indices = torch.nn.Parameter(
            torch.empty((num_experts, 0), dtype=torch.int32, device=device),
            requires_grad=False,
        )
        layer.w2_g_idx_sort_indices = torch.nn.Parameter(
            torch.empty((num_experts, 0), dtype=torch.int32, device=device),
            requires_grad=False,
        )

        marlin_w13_qweight = ops.awq_marlin_moe_repack(
            layer.w13_qweight,
            layer.w13_g_idx_sort_indices,
            size_k=layer.w13_qweight.shape[1],
            size_n=layer.w13_qweight.shape[2] * self.quant_config.pack_factor,
            num_bits=self.quant_config.weight_bits,
            is_a_8bit=is_a_8bit,
        )
        replace_parameter(layer, "w13_qweight", marlin_w13_qweight)

        marlin_w2_qweight = ops.awq_marlin_moe_repack(
            layer.w2_qweight,
            layer.w2_g_idx_sort_indices,
            size_k=layer.w2_qweight.shape[1],
            size_n=layer.w2_qweight.shape[2] * self.quant_config.pack_factor,
            num_bits=self.quant_config.weight_bits,
            is_a_8bit=is_a_8bit,
        )
        replace_parameter(layer, "w2_qweight", marlin_w2_qweight)

        # The modular kernel expects w13_weight and w2_weight,
        # but AWQ uses w13_qweight and w2_qweight
        # Alias for modular kernel
        layer.w13_weight = layer.w13_qweight
        # Alias for modular kernel
        layer.w2_weight = layer.w2_qweight

        # Why does this take the intermediate size for size_k?
        marlin_w13_scales = marlin_moe_permute_scales(
            s=layer.w13_scales,
            size_k=layer.intermediate_size_per_partition,
            size_n=layer.w13_scales.shape[2],
            group_size=self.quant_config.group_size,
            is_a_8bit=is_a_8bit,
        )
        if self.input_dtype == torch.int8 and layer.num_groups_w13 > 1:
            marlin_w13_scales, w13_input_global_scale = marlin_act_int8_process_scales(
                marlin_w13_scales
            )
            layer.register_parameter(
                "w13_input_global_scale",
                Parameter(w13_input_global_scale, requires_grad=False),
            )

        replace_parameter(layer, "w13_scales", marlin_w13_scales)

        marlin_w2_scales = marlin_moe_permute_scales(
            s=layer.w2_scales,
            size_k=layer.intermediate_size_per_partition,
            size_n=layer.w2_scales.shape[2],
            group_size=self.quant_config.group_size,
            is_a_8bit=is_a_8bit,
        )
        if self.input_dtype == torch.int8 and layer.num_groups_w2 > 1:
            marlin_w2_scales, w2_input_global_scale = marlin_act_int8_process_scales(
                marlin_w2_scales
            )
            layer.register_parameter(
                "w2_input_global_scale",
                Parameter(w2_input_global_scale, requires_grad=False),
            )

        replace_parameter(layer, "w2_scales", marlin_w2_scales)

        marlin_w13_zp = moe_awq_to_marlin_zero_points(
            layer.w13_qzeros,
            size_k=layer.w13_qzeros.shape[1],
            size_n=layer.w13_qzeros.shape[2] * self.quant_config.pack_factor,
            num_bits=self.quant_config.weight_bits,
            is_a_8bit=is_a_8bit,
        )
        replace_parameter(layer, "w13_qzeros", marlin_w13_zp)

        marlin_w2_zp = moe_awq_to_marlin_zero_points(
            layer.w2_qzeros,
            size_k=layer.w2_qzeros.shape[1],
            size_n=layer.w2_qzeros.shape[2] * self.quant_config.pack_factor,
            num_bits=self.quant_config.weight_bits,
            is_a_8bit=is_a_8bit,
        )
        replace_parameter(layer, "w2_qzeros", marlin_w2_zp)

        if hasattr(layer, "w13_bias") and layer.w13_bias is not None:
            layer.w13_bias.data = marlin_permute_bias(layer.w13_bias)

        if hasattr(layer, "w2_bias") and layer.w2_bias is not None:
            layer.w2_bias.data = marlin_permute_bias(layer.w2_bias)

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        from vllm.model_executor.layers.fused_moe.config import (
            awq_marlin_moe_quant_config,
        )

        return awq_marlin_moe_quant_config(
            w1_scale=layer.w13_scales,
            w2_scale=layer.w2_scales,
            weight_bits=self.quant_config.weight_bits,
            group_size=self.quant_config.group_size,
            w1_zp=getattr(layer, "w13_qzeros", None)
            if self.quant_config.zero_point
            else None,
            w2_zp=getattr(layer, "w2_qzeros", None)
            if self.quant_config.zero_point
            else None,
            w1_bias=getattr(layer, "w13_bias", None),
            w2_bias=getattr(layer, "w2_bias", None),
        )

    def select_gemm_impl(
        self,
        prepare_finalize,
        layer: torch.nn.Module,
    ):
        """
        Select the GEMM implementation for AWQ-Marlin MoE.
        Returns MarlinExperts configured for AWQ quantization.
        This is ONLY used when LoRA is enabled.
        Without LoRA, AWQ uses its own apply() method.
        """
        # Only use modular kernels when LoRA is enabled
        # Without LoRA, AWQ's own apply() method works fine and is more efficient
        if not self.moe.is_lora_enabled:
            raise NotImplementedError(
                "AWQ-Marlin uses its own apply() method when LoRA is not enabled. "
                "Modular kernels are only used for LoRA support."
            )

        from vllm.model_executor.layers.fused_moe import modular_kernel as mk
        from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
            BatchedMarlinExperts,
            MarlinExperts,
        )

        # Ensure quant config is initialized
        assert self.moe_quant_config is not None, (
            "moe_quant_config must be initialized before select_gemm_impl"
        )

        w13_g_idx = getattr(layer, "w13_g_idx", None)
        w2_g_idx = getattr(layer, "w2_g_idx", None)
        w13_g_idx_sort_indices = getattr(layer, "w13_g_idx_sort_indices", None)
        w2_g_idx_sort_indices = getattr(layer, "w2_g_idx_sort_indices", None)

        # Check if using batched expert format (for Expert Parallelism)
        if (
            prepare_finalize.activation_format
            == mk.FusedMoEActivationFormat.BatchedExperts
        ):
            # For batched format, use BatchedMarlinExperts
            max_num_tokens_per_rank = prepare_finalize.max_num_tokens_per_rank()
            assert max_num_tokens_per_rank is not None
            return BatchedMarlinExperts(
                max_num_tokens=max_num_tokens_per_rank,
                num_dispatchers=prepare_finalize.num_dispatchers(),
                moe_config=self.moe,
                quant_config=self.moe_quant_config,
                w13_g_idx=w13_g_idx,
                w2_g_idx=w2_g_idx,
                w13_g_idx_sort_indices=w13_g_idx_sort_indices,
                w2_g_idx_sort_indices=w2_g_idx_sort_indices,
                is_k_full=self.is_k_full,
            )
        else:
            # Standard Marlin experts for AWQ
            return MarlinExperts(
                moe_config=self.moe,
                quant_config=self.moe_quant_config,
                w13_g_idx=w13_g_idx,
                w2_g_idx=w2_g_idx,
                w13_g_idx_sort_indices=w13_g_idx_sort_indices,
                w2_g_idx_sort_indices=w2_g_idx_sort_indices,
                is_k_full=self.is_k_full,
            )

    def apply(
        self,
        layer: FusedMoE,
        x: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        shared_experts_input: torch.Tensor | None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        return fused_marlin_moe(
            x,
            layer.w13_qweight,
            layer.w2_qweight,
            getattr(layer, "w13_bias", None),
            getattr(layer, "w2_bias", None),
            layer.w13_scales,
            layer.w2_scales,
            topk_weights,
            topk_ids,
            input_global_scale1=getattr(layer, "w13_input_global_scale", None),
            input_global_scale2=getattr(layer, "w2_input_global_scale", None),
            quant_type_id=self.quant_type.id,
            apply_router_weight_on_input=layer.apply_router_weight_on_input,
            global_num_experts=layer.global_num_experts,
            expert_map=layer.expert_map,
            w1_zeros=layer.w13_qzeros,
            w2_zeros=layer.w2_qzeros,
            workspace=layer.workspace,
            input_dtype=self.input_dtype,
            inplace=not self.moe.disable_inplace,
        )