vllm-cpu-amxbf16 0.9.1__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -0,0 +1,1260 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import enum
+import importlib
+from enum import Enum
+from typing import Callable, Optional
+
+import torch
+from compressed_tensors import CompressionFormat
+from compressed_tensors.quantization import (ActivationOrdering,
+                                             QuantizationStrategy)
+
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.logger import init_logger
+from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
+                                                  FusedMoeWeightScaleSupported)
+from vllm.model_executor.layers.quantization.compressed_tensors.schemes.compressed_tensors_wNa16 import (  # noqa
+    WNA16_SUPPORTED_BITS, WNA16_SUPPORTED_TYPES_MAP)
+from vllm.model_executor.layers.quantization.utils import replace_parameter
+from vllm.model_executor.layers.quantization.utils.marlin_utils import (
+    check_moe_marlin_supports_layer, marlin_make_workspace_new,
+    marlin_moe_permute_scales)
+from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
+    prepare_moe_fp8_layer_for_marlin)
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    all_close_1d, normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize)
+from vllm.model_executor.utils import set_weight_attrs
+from vllm.platforms import current_platform
+from vllm.scalar_type import scalar_types
+
+has_pplx = importlib.util.find_spec("pplx_kernels") is not None
+
+if current_platform.is_cuda_alike():
+    from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
+        BatchedPrepareAndFinalize)
+    if has_pplx:
+        from vllm.model_executor.layers.fused_moe.pplx_prepare_finalize import (
+            PplxPrepareAndFinalize)
+
+logger = init_logger(__name__)
+
+
+class GPTQMarlinState(Enum):
+    REPACK = enum.auto()
+    READY = enum.auto()
+
+
+__all__ = [
+    "CompressedTensorsMoEMethod",
+    "CompressedTensorsW8A8Fp8MoEMethod",
+    "CompressedTensorsW8A8Fp8MoECutlassMethod",
+    "CompressedTensorsW8A8Int8MoEMethod",
+    "CompressedTensorsWNA16MarlinMoEMethod",
+    "CompressedTensorsWNA16MoEMethod",
+]
+
+
+class CompressedTensorsMoEMethod(FusedMoEMethodBase):
+
+    @staticmethod
+    def get_moe_method(
+        quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
+        layer: torch.nn.Module,
+    ) -> "CompressedTensorsMoEMethod":
+        # TODO: @dsikka: refactor this to use schemes as other kernels
+        # are supported + check if the layer is being ignored.
+        weight_quant = quant_config.target_scheme_map["Linear"].get("weights")
+        input_quant = quant_config.target_scheme_map["Linear"].get(
+            "input_activations")
+
+        if quant_config._is_wNa16_group_channel(weight_quant, input_quant):
+            # group_size=None means channelwise
+            group_size = weight_quant.group_size or -1
+            # Prefer to use the MarlinMoE kernel when it is supported.
+            if not check_moe_marlin_supports_layer(layer, group_size):
+                if (weight_quant.strategy in QuantizationStrategy.GROUP and
+                        weight_quant.actorder in (ActivationOrdering.GROUP,
+                                                  ActivationOrdering.DYNAMIC)):
+                    raise ValueError(
+                        "WNA16MoE is not supported with actorder=group/dynamic."
+                    )
+                logger.info_once("Using CompressedTensorsWNA16MoEMethod")
+                return CompressedTensorsWNA16MoEMethod(quant_config)
+            else:
+                logger.info_once("Using CompressedTensorsWNA16MarlinMoEMethod")
+                return CompressedTensorsWNA16MarlinMoEMethod(quant_config)
+        elif quant_config._is_fp8_w8a8_sm90(weight_quant, input_quant):
+            return CompressedTensorsW8A8Fp8MoECutlassMethod(quant_config)
+        elif quant_config._is_fp8_w8a8(weight_quant, input_quant):
+            return CompressedTensorsW8A8Fp8MoEMethod(quant_config)
+        elif quant_config._is_dynamic_token_w8a8(weight_quant, input_quant):
+            return CompressedTensorsW8A8Int8MoEMethod(quant_config)
+        else:
+            raise RuntimeError(
+                f"Unsupported FusedMoe scheme: {weight_quant}, {input_quant}")
+
+
+class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
+
+    def __init__(
+            self,
+            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
+    ):
+        self.quant_config = quant_config
+        self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
+            "weights")
+        self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
+            "input_activations")
+
+        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):
+            raise ValueError(
+                "For FP8 Fused MoE layers, we require per tensor "
+                "or channelwise, dynamic per token quantization. Found "
+                f"{self.weight_quant}, {self.input_quant}")
+
+        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.")
+
+        # For GPUs that lack FP8 hardware support, we can leverage the Marlin
+        # kernel for fast weight-only FP8 quantization
+        self.use_marlin = (not current_platform.has_device_capability(89)
+                           or envs.VLLM_TEST_FORCE_FP8_MARLIN)
+        # Disable marlin for rocm
+        if current_platform.is_rocm():
+            self.use_marlin = False
+        from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
+            is_rocm_aiter_moe_enabled)
+
+        self.rocm_aiter_moe_enabled = is_rocm_aiter_moe_enabled()
+
+    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
+
+        # WEIGHTS
+        w13_weight = torch.nn.Parameter(torch.empty(
+            num_experts,
+            2 * 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:
+            # Allocate 2 scales for w1 and w3 respectively.
+            # They are combined to a single scale after weight loading.
+            w13_weight_scale = torch.nn.Parameter(torch.ones(
+                num_experts, 2, 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,
+                2 * 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)
+
+        # 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: torch.nn.Module) -> None:
+        # Fp8 moe kernels require a single activation scale.
+        # We take the max of all the scales in case they differ.
+        if self.static_input_scales:
+            assert self.input_quant.strategy == QuantizationStrategy.TENSOR
+            if (layer.w13_input_scale is None or layer.w2_input_scale is None):
+                raise ValueError(
+                    "QuantConfig has static quantization, but found "
+                    "activation scales are None.")
+            if (not all_close_1d(layer.w13_input_scale)
+                    or not all_close_1d(layer.w2_input_scale)):
+                logger.warning_once(
+                    "Found input_scales that are not equal for "
+                    "fp8 MoE layer. Using the maximum across experts "
+                    "for each layer.")
+            layer.w13_input_scale = torch.nn.Parameter(
+                layer.w13_input_scale.max(), requires_grad=False)
+            layer.w2_input_scale = torch.nn.Parameter(
+                layer.w2_input_scale.max(), requires_grad=False)
+
+        if current_platform.is_fp8_fnuz():
+            # Normalize the weights and scales
+            w13_weight, w13_weight_scale, w13_input_scale = \
+                normalize_e4m3fn_to_e4m3fnuz(
+                    layer.w13_weight, layer.w13_weight_scale,
+                    layer.w13_input_scale)
+            w2_weight, w2_weight_scale, w2_input_scale = \
+                normalize_e4m3fn_to_e4m3fnuz(
+                    layer.w2_weight, layer.w2_weight_scale,
+                    layer.w2_input_scale)
+            # Reset the parameter
+            layer.w13_weight = torch.nn.Parameter(w13_weight,
+                                                  requires_grad=False)
+            layer.w13_weight_scale = torch.nn.Parameter(w13_weight_scale,
+                                                        requires_grad=False)
+            if w13_input_scale is not None:
+                layer.w13_input_scale = torch.nn.Parameter(w13_input_scale,
+                                                           requires_grad=False)
+            layer.w2_weight = torch.nn.Parameter(w2_weight,
+                                                 requires_grad=False)
+            layer.w2_weight_scale = torch.nn.Parameter(w2_weight_scale,
+                                                       requires_grad=False)
+            if w2_input_scale is not None:
+                layer.w2_input_scale = torch.nn.Parameter(w2_input_scale,
+                                                          requires_grad=False)
+
+        # For Per-TENSOR case, Fp8 moe kernel needs single weight scale
+        # for w13 per expert. Use max then dequant and requant each expert.
+        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
+            assert layer.w13_weight_scale is not None
+            shard_size = layer.intermediate_size_per_partition
+            max_w13_scales = layer.w13_weight_scale.max(dim=1).values
+            for expert_id in range(layer.local_num_experts):
+                start = 0
+                for shard_id in range(2):
+                    dq_weight = per_tensor_dequantize(
+                        layer.w13_weight[expert_id][start:start +
+                                                    shard_size, :],
+                        layer.w13_weight_scale[expert_id][shard_id])
+                    layer.w13_weight[expert_id][
+                        start:start + shard_size, :], _ = ops.scaled_fp8_quant(
+                            dq_weight, max_w13_scales[expert_id])
+                    start += shard_size
+            layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
+                                                        requires_grad=False)
+
+        # Property to determine if AITER is used
+        if self.rocm_aiter_moe_enabled:
+            from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa E501
+                rocm_aiter_fused_experts, shuffle_weights)
+
+            # reshaping weights is required for aiter moe kernel.
+            shuffled_w13, shuffled_w2 = shuffle_weights(
+                layer.w13_weight.data, layer.w2_weight.data)
+
+            layer.w13_weight = torch.nn.Parameter(shuffled_w13,
+                                                  requires_grad=False)
+            layer.w2_weight = torch.nn.Parameter(shuffled_w2,
+                                                 requires_grad=False)
+
+            self.rocm_aiter_fused_experts_func = rocm_aiter_fused_experts
+        else:
+            from vllm.model_executor.layers.fused_moe import fused_experts
+            self.fused_experts_func = fused_experts
+
+        if self.use_marlin:
+            prepare_moe_fp8_layer_for_marlin(layer, False)
+            # Activations not quantized for marlin.
+            del layer.w13_input_scale
+            del layer.w2_input_scale
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool = False,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        global_num_experts: int = -1,
+        expert_map: Optional[torch.Tensor] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        activation: str = "silu",
+    ) -> torch.Tensor:
+
+        topk_weights, topk_ids = FusedMoE.select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias)
+
+        if self.rocm_aiter_moe_enabled:
+            return self.rocm_aiter_fused_experts_func(
+                hidden_states=x,
+                w1=layer.w13_weight,
+                w2=layer.w2_weight,
+                topk_weights=topk_weights,
+                topk_ids=topk_ids,
+                activation=activation,
+                apply_router_weight_on_input=apply_router_weight_on_input,
+                use_fp8_w8a8=True,
+                per_channel_quant=self.weight_quant.strategy ==
+                QuantizationStrategy.CHANNEL,
+                w1_scale=layer.w13_weight_scale,
+                w2_scale=layer.w2_weight_scale,
+                a1_scale=layer.w13_input_scale,
+                a2_scale=layer.w2_input_scale)
+        if self.use_marlin:
+            assert activation == "silu", (
+                f"{activation} not supported for Marlin MoE.")
+            assert not apply_router_weight_on_input, (
+                "Apply router weight on input not supported for Marlin MoE.")
+            return torch.ops.vllm.fused_marlin_moe(
+                x,
+                layer.w13_weight,
+                layer.w2_weight,
+                layer.w13_weight_scale,
+                layer.w2_weight_scale,
+                router_logits,
+                topk_weights,
+                topk_ids,
+                quant_type_id=scalar_types.float8_e4m3fn.id,
+                global_num_experts=global_num_experts,
+                expert_map=expert_map)
+
+        return self.fused_experts_func(
+            hidden_states=x,
+            w1=layer.w13_weight,
+            w2=layer.w2_weight,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            inplace=True,
+            activation=activation,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            use_fp8_w8a8=True,
+            per_channel_quant=self.weight_quant.strategy ==
+            QuantizationStrategy.CHANNEL,
+            global_num_experts=global_num_experts,
+            expert_map=expert_map,
+            w1_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            a1_scale=layer.w13_input_scale,
+            a2_scale=layer.w2_input_scale)
+
+
+class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
+
+    def __init__(
+            self,
+            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
+    ):
+        self.quant_config = quant_config
+        self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
+            "weights")
+        self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
+            "input_activations")
+
+        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):
+            raise ValueError(
+                "For FP8 Fused MoE layers, we require per tensor "
+                "or channelwise, dynamic per token quantization. Found "
+                f"{self.weight_quant}, {self.input_quant}")
+
+        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.")
+
+        from vllm.model_executor.layers.fused_moe.cutlass_moe import (
+            cutlass_moe_fp8)
+        self.fused_experts = cutlass_moe_fp8  # type: ignore
+        self.disable_expert_map = False
+
+    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):
+
+        params_dtype = torch.float8_e4m3fn
+
+        # WEIGHTS
+        w13_weight = torch.nn.Parameter(torch.empty(
+            num_experts,
+            2 * 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:
+            # Allocate 2 scales for w1 and w3 respectively.
+            # They are combined to a single scale after weight loading.
+            w13_weight_scale = torch.nn.Parameter(torch.ones(
+                num_experts, 2, 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,
+                2 * 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)
+
+        # 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: torch.nn.Module) -> None:
+        # Fp8 moe kernels require a single activation scale.
+        # We take the max of all the scales in case they differ.
+        if self.static_input_scales:
+            assert self.input_quant.strategy == QuantizationStrategy.TENSOR
+            if (layer.w13_input_scale is None or layer.w2_input_scale is None):
+                raise ValueError(
+                    "QuantConfig has static quantization, but found "
+                    "activation scales are None.")
+            if (not all_close_1d(layer.w13_input_scale)
+                    or not all_close_1d(layer.w2_input_scale)):
+                logger.warning_once(
+                    "Found input_scales that are not equal for "
+                    "fp8 MoE layer. Using the maximum across experts "
+                    "for each layer.")
+            layer.w13_input_scale = torch.nn.Parameter(
+                layer.w13_input_scale.max(), requires_grad=False)
+            layer.w2_input_scale = torch.nn.Parameter(
+                layer.w2_input_scale.max(), requires_grad=False)
+
+        # For Per-TENSOR case, Fp8 moe kernel needs single weight scale
+        # for w13 per expert. Use max then dequant and requant each expert.
+        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
+            assert layer.w13_weight_scale is not None
+            shard_size = layer.intermediate_size_per_partition
+            max_w13_scales = layer.w13_weight_scale.max(dim=1).values
+            for expert_id in range(layer.local_num_experts):
+                start = 0
+                for shard_id in range(2):
+                    dq_weight = per_tensor_dequantize(
+                        layer.w13_weight[expert_id][start:start +
+                                                    shard_size, :],
+                        layer.w13_weight_scale[expert_id][shard_id])
+                    layer.w13_weight[expert_id][
+                        start:start + shard_size, :], _ = ops.scaled_fp8_quant(
+                            dq_weight, max_w13_scales[expert_id])
+                    start += shard_size
+            layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
+                                                        requires_grad=False)
+
+    def select_gemm_impl(self, prepare_finalize, moe):
+        from vllm.model_executor.layers.fused_moe.cutlass_moe import (
+            CutlassExpertsFp8)
+
+        assert moe is not None
+
+        max_experts_per_worker = (
+            (moe.num_experts + prepare_finalize.world_size - 1) //
+            prepare_finalize.world_size)
+        experts = CutlassExpertsFp8(
+            max_experts_per_worker, moe.in_dtype,
+            self.input_quant.strategy == QuantizationStrategy.TOKEN,
+            self.weight_quant.strategy == QuantizationStrategy.CHANNEL)
+
+        if has_pplx and isinstance(
+                prepare_finalize,
+            (BatchedPrepareAndFinalize, PplxPrepareAndFinalize)):
+            # no expert_map support in this case
+            self.disable_expert_map = True
+        return experts
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool = False,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        global_num_experts: int = -1,
+        expert_map: Optional[torch.Tensor] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        activation: str = "silu",
+    ) -> torch.Tensor:
+
+        topk_weights, topk_ids = FusedMoE.select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias,
+            indices_type=torch.uint32)
+
+        return self.fused_experts(
+            x,
+            layer.w13_weight,
+            layer.w2_weight,
+            topk_weights,
+            topk_ids,
+            activation=activation,
+            global_num_experts=global_num_experts,
+            expert_map=None if self.disable_expert_map else expert_map,
+            w1_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            a1_scale=layer.w13_input_scale,
+            a2_scale=layer.w2_input_scale,
+        )
+
+
+class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
+
+    def __init__(
+            self,
+            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
+    ):
+        self.quant_config = quant_config
+        self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
+            "weights")
+        self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
+            "input_activations")
+
+        per_channel = (
+            self.weight_quant.strategy == QuantizationStrategy.CHANNEL
+            and self.input_quant.strategy == QuantizationStrategy.TOKEN)
+        if not per_channel:
+            raise ValueError(
+                "For INT8 Fused MoE layers, we require channelwise, "
+                "dynamic per token quantization. Found "
+                f"{self.weight_quant}, {self.input_quant}")
+
+        self.static_input_scales = not self.input_quant.dynamic
+        if self.static_input_scales:
+            raise ValueError(
+                "For INT8 Fused MoE layers, we require channelwise, "
+                "dynamic per token quantization. Found static input scales.")
+
+    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):
+
+        params_dtype = torch.int8
+
+        # WEIGHTS
+        w13_weight = torch.nn.Parameter(torch.empty(
+            num_experts,
+            2 * 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
+        assert self.weight_quant.strategy == QuantizationStrategy.CHANNEL
+        w13_weight_scale = torch.nn.Parameter(torch.ones(
+            num_experts,
+            2 * 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)
+
+        # INPUT_SCALES
+        assert not self.static_input_scales
+        layer.w13_input_scale = None
+        layer.w2_input_scale = None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        pass
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool = False,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        global_num_experts: int = -1,
+        expert_map: Optional[torch.Tensor] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        activation: str = "silu",
+    ) -> torch.Tensor:
+        from vllm.model_executor.layers.fused_moe import fused_experts
+
+        topk_weights, topk_ids = FusedMoE.select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias)
+
+        return fused_experts(
+            hidden_states=x,
+            w1=layer.w13_weight,
+            w2=layer.w2_weight,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            inplace=True,
+            activation=activation,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            use_int8_w8a8=True,
+            per_channel_quant=True,
+            global_num_experts=global_num_experts,
+            expert_map=expert_map,
+            w1_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            a1_scale=layer.w13_input_scale,
+            a2_scale=layer.w2_input_scale)
+
+
+class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
+
+    def __init__(
+            self,
+            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
+    ):
+        self.quant_config = quant_config
+        # TODO: @dsikka: refactor this to use schemes as other kernels
+        # are supported + check if the layer is being ignored.
+        config = self.quant_config.target_scheme_map["Linear"].get("weights")
+        self.num_bits = config.num_bits
+        self.packed_factor = 32 // config.num_bits
+        self.strategy = config.strategy
+        self.group_size = config.group_size
+        self.actorder = config.actorder
+        assert config.symmetric, (
+            "Only symmetric quantization is supported for MoE")
+
+        if not (self.quant_config.quant_format
+                == CompressionFormat.pack_quantized.value
+                and self.num_bits in WNA16_SUPPORTED_BITS):
+            raise ValueError("For Fused MoE layers, only ",
+                             f"{CompressionFormat.pack_quantized.value} ",
+                             "is supported for the following bits: ",
+                             f"{WNA16_SUPPORTED_BITS}")
+        self.quant_type = WNA16_SUPPORTED_TYPES_MAP[self.num_bits]
+
+    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):
+
+        intermediate_size_full = extra_weight_attrs.pop(
+            "intermediate_size_full")
+
+        # Will transpose the loaded weight along the
+        # intermediate and hidden dim sizes. Will
+        # shard for TP along the transposed dims
+        extra_weight_attrs.update({
+            "is_transposed": True,
+            "quant_method": self.strategy
+        })
+        w13_weight = torch.nn.Parameter(torch.empty(
+            num_experts,
+            hidden_size // self.packed_factor,
+            2 * intermediate_size_per_partition,
+            dtype=torch.int32),
+                                        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,
+            intermediate_size_per_partition // self.packed_factor,
+            hidden_size,
+            dtype=torch.int32),
+                                       requires_grad=False)
+        layer.register_parameter("w2_weight_packed", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # In the case where we have actorder/g_idx,
+        # we do not partition the w2 scales
+        load_full_w2 = self.actorder and self.group_size != -1
+        w2_scales_size = (intermediate_size_full
+                          if load_full_w2 else intermediate_size_per_partition)
+
+        self.is_k_full = (not self.actorder) or (
+            intermediate_size_per_partition == intermediate_size_full)
+
+        if self.strategy == "channel":
+            num_groups_w2 = num_groups_w13 = 1
+            self.group_size = -1
+        else:
+            num_groups_w2 = w2_scales_size // self.group_size
+            num_groups_w13 = hidden_size // self.group_size
+
+        w13_scale = torch.nn.Parameter(torch.ones(
+            num_experts,
+            num_groups_w13,
+            2 * intermediate_size_per_partition,
+            dtype=params_dtype),
+                                       requires_grad=False)
+        layer.register_parameter("w13_weight_scale", w13_scale)
+        set_weight_attrs(w13_scale, extra_weight_attrs)
+
+        w2_scale = torch.nn.Parameter(torch.ones(num_experts,
+                                                 num_groups_w2,
+                                                 hidden_size,
+                                                 dtype=params_dtype),
+                                      requires_grad=False)
+        layer.register_parameter("w2_weight_scale", w2_scale)
+        set_weight_attrs(w2_scale, extra_weight_attrs)
+        set_weight_attrs(w2_scale, {"load_full_w2": load_full_w2})
+
+        w2_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
+                                             requires_grad=False)
+        layer.register_parameter("w2_weight_shape", w2_weight_shape)
+        set_weight_attrs(w2_weight_shape, extra_weight_attrs)
+        w13_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
+                                              requires_grad=False)
+
+        layer.register_parameter("w13_weight_shape", w13_weight_shape)
+        set_weight_attrs(w13_weight_shape, extra_weight_attrs)
+
+        w13_g_idx = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_g_idx", w13_g_idx)
+        set_weight_attrs(w13_g_idx, extra_weight_attrs)
+
+        w2_g_idx = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                intermediate_size_per_partition,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_g_idx", w2_g_idx)
+        set_weight_attrs(w2_g_idx, extra_weight_attrs)
+
+        w13_g_idx_sort_indices = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_g_idx_sort_indices",
+                                 w13_g_idx_sort_indices)
+        set_weight_attrs(w13_g_idx_sort_indices, extra_weight_attrs)
+
+        w2_g_idx_sort_indices = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                intermediate_size_per_partition,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_g_idx_sort_indices",
+                                 w2_g_idx_sort_indices)
+        set_weight_attrs(w2_g_idx_sort_indices, extra_weight_attrs)
+
+        layer.a13_scale = None
+        layer.a2_scale = None
+        layer.marlin_state = GPTQMarlinState.REPACK
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        num_experts = layer.w13_weight_g_idx.shape[0]
+        device = layer.w13_weight_g_idx.device
+
+        # when running models with grouped act order,
+        # resort to g_idx values provided in checkpoint
+        if self.actorder == "group":
+            w13_g_idx_sort_indices = torch.empty_like(layer.w13_weight_g_idx)
+            w2_g_idx_sort_indices = torch.empty_like(layer.w2_weight_g_idx)
+            w13_sorted_g_idx = torch.empty_like(layer.w13_weight_g_idx)
+            w2_sorted_g_idx = torch.empty_like(layer.w2_weight_g_idx)
+
+            for e in range(num_experts):
+                w13_g_idx_sort_indices[e] = torch.argsort(
+                    layer.w13_weight_g_idx[e]).to(torch.int32)
+                w2_g_idx_sort_indices[e] = torch.argsort(
+                    layer.w2_weight_g_idx[e]).to(torch.int32)
+                w13_sorted_g_idx[e] = layer.w13_weight_g_idx[e][
+                    w13_g_idx_sort_indices[e]]
+                w2_sorted_g_idx[e] = layer.w2_weight_g_idx[e][
+                    w2_g_idx_sort_indices[e]]
+
+            replace_parameter(layer, "w13_weight_g_idx", w13_sorted_g_idx)
+            replace_parameter(layer, "w2_weight_g_idx", w2_sorted_g_idx)
+            replace_parameter(layer, "w13_g_idx_sort_indices",
+                              w13_g_idx_sort_indices)
+            replace_parameter(layer, "w2_g_idx_sort_indices",
+                              w2_g_idx_sort_indices)
+
+        else:
+            layer.w13_weight_g_idx = torch.nn.Parameter(
+                torch.empty((num_experts, 0), dtype=torch.int32,
+                            device=device),
+                requires_grad=False,
+            )
+            layer.w2_weight_g_idx = torch.nn.Parameter(
+                torch.empty((num_experts, 0), dtype=torch.int32,
+                            device=device),
+                requires_grad=False,
+            )
+            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.gptq_marlin_moe_repack(
+            layer.w13_weight_packed,
+            layer.w13_g_idx_sort_indices,
+            layer.w13_weight_packed.shape[1] * self.packed_factor,
+            layer.w13_weight_packed.shape[2],
+            self.num_bits,
+        )
+        replace_parameter(layer, "w13_weight_packed", marlin_w13_qweight)
+        marlin_w2_qweight = ops.gptq_marlin_moe_repack(
+            layer.w2_weight_packed,
+            layer.w2_g_idx_sort_indices,
+            layer.w2_weight_packed.shape[1] * self.packed_factor,
+            layer.w2_weight_packed.shape[2],
+            self.num_bits,
+        )
+        replace_parameter(layer, "w2_weight_packed", marlin_w2_qweight)
+        # Repack scales
+        marlin_w13_scales = marlin_moe_permute_scales(
+            s=layer.w13_weight_scale,
+            size_k=layer.w13_weight_packed.shape[2],
+            size_n=layer.w13_weight_scale.shape[2],
+            group_size=self.group_size,
+        )
+        replace_parameter(layer, "w13_weight_scale", marlin_w13_scales)
+        marlin_w2_scales = marlin_moe_permute_scales(
+            s=layer.w2_weight_scale,
+            size_k=layer.w2_weight_scale.shape[1] *
+            (self.group_size if self.group_size != -1 else self.packed_factor),
+            size_n=layer.w2_weight_scale.shape[2],
+            group_size=self.group_size,
+        )
+        replace_parameter(layer, "w2_weight_scale", marlin_w2_scales)
+
+        layer.workspace = marlin_make_workspace_new(device, 4)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool = False,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        global_num_experts: int = -1,
+        expert_map: Optional[torch.Tensor] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        activation: str = "silu",
+    ) -> torch.Tensor:
+        assert activation == "silu", (
+            f"{activation} not supported for Marlin MoE.")
+        assert not apply_router_weight_on_input, (
+            "Apply router weight on input not supported for Marlin MoE.")
+
+        topk_weights, topk_ids = FusedMoE.select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias)
+
+        return torch.ops.vllm.fused_marlin_moe(
+            x,
+            layer.w13_weight_packed,
+            layer.w2_weight_packed,
+            layer.w13_weight_scale,
+            layer.w2_weight_scale,
+            router_logits,
+            topk_weights,
+            topk_ids,
+            quant_type_id=self.quant_type.id,
+            global_num_experts=global_num_experts,
+            expert_map=expert_map,
+            g_idx1=layer.w13_weight_g_idx,
+            g_idx2=layer.w2_weight_g_idx,
+            sort_indices1=layer.w13_g_idx_sort_indices,
+            sort_indices2=layer.w2_g_idx_sort_indices,
+            workspace=layer.workspace,
+            is_k_full=self.is_k_full)
+
+
+class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
+
+    def __init__(
+            self,
+            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
+    ):
+        self.quant_config = quant_config
+        # TODO: @dsikka: refactor this to use schemes as other kernels
+        # are supported + check if the layer is being ignored.
+        config = self.quant_config.target_scheme_map["Linear"].get("weights")
+        self.num_bits = config.num_bits
+        self.packed_factor = 32 // config.num_bits
+        self.strategy = config.strategy
+        # channelwise is not supported by this kernel
+        assert config.strategy == "group"
+        self.group_size = config.group_size
+        # grouped actorder isn't supported by this kernel
+        assert config.actorder != "group"
+        assert config.symmetric, (
+            "Only symmetric quantization is supported for MoE")
+
+        if not (self.quant_config.quant_format
+                == CompressionFormat.pack_quantized.value
+                and self.num_bits in WNA16_SUPPORTED_BITS):
+            raise ValueError("For Fused MoE layers, only ",
+                             f"{CompressionFormat.pack_quantized.value} ",
+                             "is supported for the following bits: ",
+                             f"{WNA16_SUPPORTED_BITS}")
+
+    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):
+
+        # Will transpose the loaded weight along the
+        # intermediate and hidden dim sizes. Will
+        # shard for TP along the transposed dims
+        extra_weight_attrs.update({
+            "is_transposed": True,
+            "quant_method": self.strategy
+        })
+        w13_weight = torch.nn.Parameter(torch.empty(
+            num_experts,
+            hidden_size // self.packed_factor,
+            2 * intermediate_size_per_partition,
+            dtype=torch.int32),
+                                        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,
+            intermediate_size_per_partition // self.packed_factor,
+            hidden_size,
+            dtype=torch.int32),
+                                       requires_grad=False)
+        layer.register_parameter("w2_weight_packed", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        w2_scales_size = intermediate_size_per_partition
+
+        if self.strategy == "channel":
+            num_groups_w2 = num_groups_w13 = 1
+            self.group_size = -1
+        else:
+            num_groups_w2 = w2_scales_size // self.group_size
+            num_groups_w13 = hidden_size // self.group_size
+
+        w13_scale = torch.nn.Parameter(torch.ones(
+            num_experts,
+            num_groups_w13,
+            2 * intermediate_size_per_partition,
+            dtype=params_dtype),
+                                       requires_grad=False)
+        layer.register_parameter("w13_weight_scale", w13_scale)
+        set_weight_attrs(w13_scale, extra_weight_attrs)
+
+        w2_scale = torch.nn.Parameter(torch.ones(num_experts,
+                                                 num_groups_w2,
+                                                 hidden_size,
+                                                 dtype=params_dtype),
+                                      requires_grad=False)
+        layer.register_parameter("w2_weight_scale", w2_scale)
+        set_weight_attrs(w2_scale, extra_weight_attrs)
+        set_weight_attrs(w2_scale, {"load_full_w2": False})
+
+        w2_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
+                                             requires_grad=False)
+        layer.register_parameter("w2_weight_shape", w2_weight_shape)
+        set_weight_attrs(w2_weight_shape, extra_weight_attrs)
+        w13_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
+                                              requires_grad=False)
+
+        layer.register_parameter("w13_weight_shape", w13_weight_shape)
+        set_weight_attrs(w13_weight_shape, extra_weight_attrs)
+
+        w13_g_idx = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_g_idx", w13_g_idx)
+        set_weight_attrs(w13_g_idx, extra_weight_attrs)
+
+        w2_g_idx = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                intermediate_size_per_partition,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_g_idx", w2_g_idx)
+        set_weight_attrs(w2_g_idx, extra_weight_attrs)
+
+        w13_g_idx_sort_indices = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_g_idx_sort_indices",
+                                 w13_g_idx_sort_indices)
+        set_weight_attrs(w13_g_idx_sort_indices, extra_weight_attrs)
+
+        w2_g_idx_sort_indices = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                intermediate_size_per_partition,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_g_idx_sort_indices",
+                                 w2_g_idx_sort_indices)
+        set_weight_attrs(w2_g_idx_sort_indices, extra_weight_attrs)
+
+        layer.a13_scale = None
+        layer.a2_scale = None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        # Reconfigure packed weights and scales to match moe_wna16 format
+        layer.w13_weight_packed = torch.nn.Parameter(
+            layer.w13_weight_packed.transpose(1, 2).contiguous().view(
+                torch.uint8),
+            requires_grad=False)
+        layer.w2_weight_packed = torch.nn.Parameter(
+            layer.w2_weight_packed.transpose(1,
+                                             2).contiguous().view(torch.uint8),
+            requires_grad=False)
+        layer.w13_weight_scale = torch.nn.Parameter(
+            layer.w13_weight_scale.transpose(1, 2).contiguous(),
+            requires_grad=False)
+        layer.w2_weight_scale = torch.nn.Parameter(
+            layer.w2_weight_scale.transpose(1, 2).contiguous(),
+            requires_grad=False)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool = False,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        global_num_experts: int = -1,
+        expert_map: Optional[torch.Tensor] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        activation: str = "silu",
+    ) -> torch.Tensor:
+        from vllm.model_executor.layers.fused_moe import fused_experts
+
+        topk_weights, topk_ids = FusedMoE.select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias)
+
+        return fused_experts(
+            x,
+            layer.w13_weight_packed,
+            layer.w2_weight_packed,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            inplace=True,
+            activation=activation,
+            use_int4_w4a16=self.num_bits == 4,
+            use_int8_w8a16=self.num_bits == 8,
+            global_num_experts=global_num_experts,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            expert_map=expert_map,
+            w1_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            w1_zp=None,
+            w2_zp=None,
+            block_shape=[0, self.group_size])