sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl

sglang/srt/layers/quantization/mxfp4.py

@@ -22,6 +22,8 @@ from typing import TYPE_CHECKING, List, Optional
 import torch
 from torch.nn.parameter import Parameter
 
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
 from sglang.srt.layers.moe.utils import get_moe_runner_backend
 from sglang.srt.layers.quantization.base_config import (
     FusedMoEMethodBase,
@@ -59,8 +61,10 @@ if is_flashinfer_available():
 logger = logging.getLogger(__name__)
 
 if TYPE_CHECKING:
-    from sglang.srt.layers.moe.moe_runner import MoeRunnerConfig
-    from sglang.srt.layers.moe.topk import TopKOutput
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
 
 _is_hip = is_hip()
 
@@ -283,7 +287,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
         layer: torch.nn.Module,
         num_experts: int,
         hidden_size: int,
-        intermediate_size: int,
+        intermediate_size_per_partition: int,
         params_dtype: torch.dtype,
         with_bias: bool = False,
         **extra_weight_attrs,
@@ -296,26 +300,26 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
 
         # pad the intermediate size to be a multiple of 2 * mxfp4_block
         # for to hold non-uniform sharded tensor as well as swizzling
-        intermediate_size_per_partition_after_pad = intermediate_size
+        intermediate_size_per_partition_after_pad = intermediate_size_per_partition
         if _is_sm100_supported:
             if self.use_flashinfer:
                 intermediate_size_per_partition_after_pad = round_up(
-                    intermediate_size, 256
+                    intermediate_size_per_partition, 256
                 )
                 hidden_size = round_up(hidden_size, 256)
             else:
                 intermediate_size_per_partition_after_pad = round_up(
-                    intermediate_size, 64
+                    intermediate_size_per_partition, 64
                 )
         elif has_triton_kernels:
             # TODO: this is a hack to make
             # intermediate_size_per_partition_after_pad the same as the
             # per_rank_intermediate_size during weight loading
             intermediate_size_per_partition_after_pad = round_up(
-                intermediate_size, mxfp4_block
+                intermediate_size_per_partition, mxfp4_block
             )
 
-        self.intermediate_size = intermediate_size_per_partition_after_pad
+        self.intermediate_size_per_partition = intermediate_size_per_partition_after_pad
 
         self.hidden_size = hidden_size
         # Fused gate_up_proj (column parallel)
@@ -410,31 +414,35 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
             assert (
                 layer.w13_weight.dim() == 3
                 and layer.w13_weight.shape[0] == self.num_experts
-                and layer.w13_weight.shape[1] == self.intermediate_size * 2
+                and layer.w13_weight.shape[1]
+                == self.intermediate_size_per_partition * 2
                 and layer.w13_weight.shape[2] == self.hidden_size // 2
             )
             assert (
                 layer.w13_weight_scale.dim() == 3
                 and layer.w13_weight_scale.shape[0] == self.num_experts
-                and layer.w13_weight_scale.shape[1] == self.intermediate_size * 2
+                and layer.w13_weight_scale.shape[1]
+                == self.intermediate_size_per_partition * 2
                 and layer.w13_weight_scale.shape[2] == self.hidden_size // sf_block_size
             )
             assert (
                 layer.w2_weight.dim() == 3
                 and layer.w2_weight.shape[0] == self.num_experts
                 and layer.w2_weight.shape[1] == self.hidden_size
-                and layer.w2_weight.shape[2] == self.intermediate_size // 2
+                and layer.w2_weight.shape[2]
+                == self.intermediate_size_per_partition // 2
             )
             assert (
                 layer.w2_weight_scale.dim() == 3
                 and layer.w2_weight_scale.shape[1] == self.hidden_size
                 and layer.w2_weight_scale.shape[2]
-                == self.intermediate_size // sf_block_size
+                == self.intermediate_size_per_partition // sf_block_size
             )
             assert (
                 layer.w13_weight_bias.dim() == 2
                 and layer.w13_weight_bias.shape[0] == self.num_experts
-                and layer.w13_weight_bias.shape[1] == self.intermediate_size * 2
+                and layer.w13_weight_bias.shape[1]
+                == self.intermediate_size_per_partition * 2
             )
             assert (
                 layer.w2_weight_bias.dim() == 2
@@ -511,7 +519,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                 torch.stack(gemm1_scales_mxfp4_shuffled)
                 .reshape(
                     self.num_experts,
-                    2 * self.intermediate_size,
+                    2 * self.intermediate_size_per_partition,
                     self.hidden_size // sf_block_size,
                 )
                 .view(torch.float8_e4m3fn)
@@ -523,7 +531,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                 .reshape(
                     self.num_experts,
                     self.hidden_size,
-                    self.intermediate_size // sf_block_size,
+                    self.intermediate_size_per_partition // sf_block_size,
                 )
                 .view(torch.float8_e4m3fn)
             )
@@ -613,16 +621,26 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
 
         return tile_tokens_dim
 
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.runner = MoeRunner(MoeRunnerBackend.TRITON, moe_runner_config)
+
     def apply(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
 
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
         from sglang.srt.layers.moe.topk import TopKOutputChecker
 
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        moe_runner_config = self.moe_runner_config
+
         if self.use_flashinfer:
             # When bf16 mode is enabled, we don't need to quantize the input,
             # TRT-LLM automatically handles quantization in the kernel implementation and pipelines it with GEMM operations,
@@ -674,7 +692,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                 top_k,
                 None,  # n_group      # TODO: support n_group
                 None,  # topk_group   # TODO: support topk_group
-                self.intermediate_size,  # padded to multiple of 256
+                self.intermediate_size_per_partition,  # padded to multiple of 256
                 layer.moe_ep_rank * layer.num_local_experts,  # local_expert_offset
                 layer.num_local_experts,  # local num experts
                 None,
@@ -682,14 +700,14 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                 1,  # routing_method_type, renormalize
                 True,  # do finalize
             )[0]
-            return trtllm_gen_output
+            return StandardCombineInput(hidden_states=trtllm_gen_output)
 
         if self.use_triton_kernels:
             assert (
                 layer.moe_ep_size == 1
             ), "Expert parallel is not supported when using triton kernels"
             if self.with_bias:
-                return self.triton_kernel_moe_with_bias_forward(
+                output = self.triton_kernel_moe_with_bias_forward(
                     hidden_states=x,
                     w1=self.w13_weight_triton_tensor,
                     w1_pcg=self.w13_precision_config,
@@ -701,25 +719,22 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                     moe_runner_config=moe_runner_config,
                 )
             else:
-                return self.triton_kernel_moe_forward(
+                output = self.triton_kernel_moe_forward(
                     hidden_states=x,
                     w1=layer.w13_weight,
                     w2=layer.w2_weight,
                     topk_output=topk_output,
                     moe_runner_config=moe_runner_config,
                 )
+            return StandardCombineInput(hidden_states=output)
         else:
-            from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
-
-            return fused_experts(
-                hidden_states=x,
-                w1=layer.w13_weight,
-                w2=layer.w2_weight,
-                topk_output=topk_output,
-                moe_runner_config=moe_runner_config,
-                b1=layer.w13_weight_bias,
-                b2=layer.w2_weight_bias,
+            quant_info = TritonMoeQuantInfo(
+                w13_weight=layer.w13_weight,
+                w2_weight=layer.w2_weight,
+                w13_weight_bias=layer.w13_weight_bias,
+                w2_weight_bias=layer.w2_weight_bias,
             )
+            return self.runner.run(dispatch_output, quant_info)
 
 
 class Mxfp4DynamicQuantMoEMethod(FusedMoEMethodBase):
@@ -798,7 +813,7 @@ class Mxfp4DynamicQuantMoEMethod(FusedMoEMethodBase):
 
         return w, mx_scales
 
-    def process_weights_after_loading(self, layer: Module) -> None:
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
         w13, w13_mx_scales = self.mxfp4_quantize(layer.w13_weight.data)
         w2, w2_mx_scales = self.mxfp4_quantize(layer.w2_weight.data)
 
@@ -808,19 +823,27 @@ class Mxfp4DynamicQuantMoEMethod(FusedMoEMethodBase):
         layer.w2_weight = torch.nn.Parameter(w2, requires_grad=False)
         layer.w2_weight_scale = torch.nn.Parameter(w2_mx_scales, requires_grad=False)
 
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+
     def apply(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
         topk_weights, topk_ids, _ = topk_output
         if _is_hip:
             topk_weights = topk_weights.to(
                 torch.float32
             )  # aiter's moe_sorting requires topk_weights to be FP32
-        return fused_moe(
+        output = fused_moe(
             x,
             layer.w13_weight,
             layer.w2_weight,
@@ -831,8 +854,9 @@ class Mxfp4DynamicQuantMoEMethod(FusedMoEMethodBase):
             w2_scale=layer.w2_weight_scale,
             activation=(
                 ActivationType.Silu
-                if moe_runner_config.activation == "silu"
+                if self.moe_runner_config.activation == "silu"
                 else ActivationType.Gelu
             ),
             doweight_stage1=False,
         )
+        return StandardCombineInput(hidden_states=output)

sglang/srt/layers/quantization/quark/quark_moe.py

@@ -10,8 +10,17 @@ from aiter import ActivationType, QuantType, biased_grouped_topk
 from aiter.fused_moe import fused_moe
 from aiter.utility.fp4_utils import e8m0_shuffle
 
+from sglang.srt.layers.moe import MoeRunnerConfig
+from sglang.srt.layers.quantization.base_config import FusedMoEMethodBase
 from sglang.srt.utils import get_bool_env_var, mxfp_supported, set_weight_attrs
 
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+    from sglang.srt.layers.quantization.quark.quark import QuarkConfig
+
 logger = logging.getLogger(__name__)
 
 __all__ = ["QuarkMoEMethod", "QuarkW4A4MXFp4MoEMethod"]
@@ -19,31 +28,17 @@ __all__ = ["QuarkMoEMethod", "QuarkW4A4MXFp4MoEMethod"]
 OCP_MX_BLOCK_SIZE = 32
 
 if TYPE_CHECKING:
-    from sglang.srt.layers.moe.topk import TopKOutput
-
-
-class QuarkMoEMethod:
-    def __new__(cls, *args, **kwargs):
-        from sglang.srt.layers.quantization.base_config import FusedMoEMethodBase
-
-        if not hasattr(cls, "_initialized"):
-            original_init = cls.__init__
-            new_cls = type(
-                cls.__name__,
-                (FusedMoEMethodBase,),
-                {
-                    "__init__": original_init,
-                    **{k: v for k, v in cls.__dict__.items() if k != "__dict__"},
-                },
-            )
-            obj = super(new_cls, new_cls).__new__(new_cls)
-            obj.__init__(*args, **kwargs)
-            return obj
-        return super().__new__(cls)
+    from sglang.srt.layers.quantization import QuarkConfig
+
+
+class QuarkMoEMethod(FusedMoEMethodBase):
+
+    def __init__(self, quant_config: QuarkConfig):
+        self.quant_config = quant_config
 
     @staticmethod
     def get_moe_method(
-        quant_config: "QuarkConfig",  # type: ignore # noqa E501 # noqa F821
+        quant_config: QuarkConfig,  # type: ignore # noqa E501 # noqa F821
         module: torch.nn.Module,
         layer_name: str,
     ) -> "QuarkMoEMethod":
@@ -170,16 +165,25 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
         # layer.w2_weight_scale = torch.nn.Parameter(w2_weight_scale, requires_grad=False)
         layer.w2_weight_scale.data = w2_weight_scale.view(s0, s1, -1)
 
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+
     def apply(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+        moe_runner_config = self.moe_runner_config
         topk_weights, topk_ids, _ = topk_output
 
-        return fused_moe(
+        output = fused_moe(
             x,
             layer.w13_weight,
             layer.w2_weight,
@@ -195,3 +199,4 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
             ),
             doweight_stage1=False,
         )
+        return StandardCombineInput(hidden_states=output)

sglang/srt/layers/quantization/unquant.py

@@ -9,6 +9,8 @@ from torch.nn.parameter import Parameter
 
 from sglang.srt.custom_op import CustomOp
 from sglang.srt.layers.amx_utils import _amx_process_weight_after_loading
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
 from sglang.srt.layers.quantization.base_config import (
     FusedMoEMethodBase,
     LinearMethodBase,
@@ -24,8 +26,10 @@ from sglang.srt.utils import (
 )
 
 if TYPE_CHECKING:
-    from sglang.srt.layers.moe.moe_runner import MoeRunnerConfig
-    from sglang.srt.layers.moe.topk import TopKOutput
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
 
 has_triton_kernels = importlib.util.find_spec("triton_kernels") is not None
 
@@ -155,7 +159,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         layer: torch.nn.Module,
         num_experts: int,
         hidden_size: int,
-        intermediate_size: int,
+        intermediate_size_per_partition: int,
         params_dtype: torch.dtype,
         with_bias: bool = False,
         **extra_weight_attrs,
@@ -163,7 +167,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         self.with_bias = with_bias
 
         # Fused gate_up_proj (column parallel)
-        w13_weight_n, w13_weight_k = 2 * intermediate_size, hidden_size
+        w13_weight_n, w13_weight_k = 2 * intermediate_size_per_partition, hidden_size
         if self.use_triton_kernels:
             w13_weight_n, w13_weight_k = w13_weight_k, w13_weight_n
         w13_weight = torch.nn.Parameter(
@@ -175,7 +179,11 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
 
         if self.with_bias:
             w13_weight_bias = torch.nn.Parameter(
-                torch.empty(num_experts, 2 * intermediate_size, dtype=torch.float32),
+                torch.empty(
+                    num_experts,
+                    2 * intermediate_size_per_partition,
+                    dtype=torch.float32,
+                ),
                 requires_grad=False,
             )
             layer.register_parameter("w13_weight_bias", w13_weight_bias)
@@ -184,7 +192,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         # down_proj (row parallel)
         w2_weight_n, w2_weight_k = (
             hidden_size,
-            intermediate_size,
+            intermediate_size_per_partition,
         )
         if self.use_triton_kernels:
             w2_weight_n, w2_weight_k = w2_weight_k, w2_weight_n
@@ -222,33 +230,40 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
 
         return
 
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.runner = MoeRunner(MoeRunnerBackend.TRITON, moe_runner_config)
+
     def apply(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
 
         return self.forward(
-            x=x,
             layer=layer,
-            topk_output=topk_output,
-            moe_runner_config=moe_runner_config,
+            dispatch_output=dispatch_output,
         )
 
     def forward_cuda(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        moe_runner_config = self.moe_runner_config
 
         if self.use_triton_kernels:
             if self.with_bias:
                 assert self.triton_kernel_moe_with_bias_forward is not None
-                return self.triton_kernel_moe_with_bias_forward(
+                output = self.triton_kernel_moe_with_bias_forward(
                     hidden_states=x,
                     w1=layer.w13_weight,
                     w2=layer.w2_weight,
@@ -261,13 +276,14 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                 )
             else:
                 assert self.triton_kernel_moe_forward is not None
-                return self.triton_kernel_moe_forward(
+                output = self.triton_kernel_moe_forward(
                     hidden_states=x,
                     w1=layer.w13_weight,
                     w2=layer.w2_weight,
                     topk_output=topk_output,
                     moe_runner_config=moe_runner_config,
                 )
+            return StandardCombineInput(hidden_states=output)
         else:
             if _use_aiter:
                 assert not moe_runner_config.no_combine, "unsupported"
@@ -284,7 +300,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                     topk_weights = torch.ones_like(
                         topk_weights, dtype=torch.float32
                     )  # topk_weights must be FP32 (float32)
-                return fused_moe(
+                output = fused_moe(
                     x,
                     layer.w13_weight,
                     layer.w2_weight,
@@ -296,28 +312,30 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                         else ActivationType.Gelu
                     ),
                 )
+                return StandardCombineInput(hidden_states=output)
             else:
-                from sglang.srt.layers.moe.fused_moe_triton.fused_moe import (
-                    fused_experts,
-                )
 
-                return fused_experts(
-                    hidden_states=x,
-                    w1=layer.w13_weight,
-                    w2=layer.w2_weight,
-                    b1=getattr(layer, "w13_weight_bias", None),
+                quant_info = TritonMoeQuantInfo(
+                    w13_weight=layer.w13_weight,
+                    w2_weight=layer.w2_weight,
+                    b13=getattr(layer, "w13_weight_bias", None),
                     b2=getattr(layer, "w2_weight_bias", None),
-                    topk_output=topk_output,
-                    moe_runner_config=moe_runner_config,
                 )
+                return self.runner.run(dispatch_output, quant_info)
 
     def forward_cpu(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        moe_runner_config = self.moe_runner_config
+
         assert (
             moe_runner_config.activation == "silu"
         ), f"activation = {moe_runner_config.activation} is not supported."
@@ -332,7 +350,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             x, topk_weights = apply_topk_weights_cpu(
                 moe_runner_config.apply_router_weight_on_input, topk_weights, x
             )
-            return torch.ops.sgl_kernel.fused_experts_cpu(
+            output = torch.ops.sgl_kernel.fused_experts_cpu(
                 x,
                 layer.w13_weight,
                 layer.w2_weight,
@@ -348,33 +366,103 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                 None,  # a2_scale
                 True,  # is_vnni
             )
+            return StandardCombineInput(hidden_states=output)
         else:
             from sglang.srt.layers.moe.fused_moe_native import moe_forward_native
 
-            return moe_forward_native(
+            output = moe_forward_native(
                 layer,
                 x,
                 topk_output,
                 moe_runner_config,
             )
+            return StandardCombineInput(hidden_states=output)
 
     def forward_npu(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
-        from sglang.srt.layers.moe.fused_moe_native import moe_forward_native
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        import torch_npu
+
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_weights, topk_ids, _ = dispatch_output.topk_output
+
+        original_dtype = x.dtype
+        num_tokens = x.shape[0]
+        topk_weights = topk_weights.to(x.dtype)
+        topk_ids = topk_ids.to(torch.int32)
+        num_experts = layer.num_experts
+        top_k = layer.top_k
+        row_idx_len = num_tokens * top_k
+        row_idx = (
+            torch.arange(0, row_idx_len, dtype=torch.int32, device=topk_weights.device)
+            .view(top_k, -1)
+            .permute(1, 0)
+            .contiguous()
+        )
 
-        return moe_forward_native(
-            layer,
-            x,
-            topk_output,
-            moe_runner_config,
+        hidden_states, expanded_row_idx, expanded_expert_idx = (
+            torch_npu.npu_moe_init_routing(
+                x, row_idx=row_idx, expert_idx=topk_ids, active_num=num_tokens
+            )
+        )
+
+        expert_tokens = torch_npu.npu_moe_compute_expert_tokens(
+            expanded_expert_idx, num_experts
         )
 
-    def forward_tpu(self, *args, **kwargs) -> torch.Tensor:
+        expert_tokens = expert_tokens.to(torch.int64)
+        if layer.w13_weight.shape[-1] == layer.hidden_size:
+            w13 = layer.w13_weight.transpose(1, 2)
+            w2 = layer.w2_weight.transpose(1, 2)
+
+        # gmm1: gate_up_proj
+        hidden_states = torch_npu.npu_grouped_matmul(
+            x=[hidden_states],
+            weight=[w13],
+            split_item=2,
+            group_list_type=0,
+            group_type=0,
+            group_list=expert_tokens,
+            output_dtype=original_dtype,
+        )[0]
+
+        # act_fn:
+        if self.moe_runner_config.activation == "silu":
+            hidden_states = torch_npu.npu_swiglu(hidden_states)
+        else:
+            from sglang.srt.layers.activation import GeluAndMul
+
+            hidden_states = GeluAndMul()(hidden_states)
+
+        # gmm2: down_proj
+        hidden_states = torch_npu.npu_grouped_matmul(
+            x=[hidden_states],
+            weight=[w2],
+            split_item=2,
+            group_list_type=0,
+            group_type=0,
+            group_list=expert_tokens,
+            output_dtype=original_dtype,
+        )[0]
+
+        final_hidden_states = torch_npu.npu_moe_finalize_routing(
+            hidden_states,
+            skip1=None,
+            skip2=None,
+            bias=None,
+            scales=topk_weights,
+            expanded_src_to_dst_row=expanded_row_idx,
+            export_for_source_row=topk_ids,
+        )
+
+        return StandardCombineInput(hidden_states=final_hidden_states)
+
+    def forward_tpu(self, *args, **kwargs) -> CombineInput:
         raise NotImplementedError("The TPU backend currently does not support MoE.")
 
     forward_native = forward_cpu