diff --git a/tests/kernels/moe/test_moe_layer.py b/tests/kernels/moe/test_moe_layer.py
index 01664d6a932..85619a91005 100644
--- a/tests/kernels/moe/test_moe_layer.py
+++ b/tests/kernels/moe/test_moe_layer.py
@@ -236,7 +236,6 @@ class MoETestConfig:
     use_gate: bool
     use_routed_input_transform: bool
     enable_eplb: bool = False
-    reduce_results: bool = False
     backend: str | None = None
     ep_size: int = 1
     dp_size: int = 1
@@ -295,7 +294,6 @@ def generate_valid_test_configs(
         use_shared_experts,
         use_gate,
         use_routed_input_transform,
-        reduce_results,
     ) in product(
         SHAPE_COMBOS,
         NUM_EXPERTS,
@@ -304,7 +302,6 @@ def generate_valid_test_configs(
         [False, True],  # shared
         [False, True],  # gate
         [False, True],  # routed input exform
-        [False, True],  # reduce results
     ):
         config = MoETestConfig(
             shape[0],  # m
@@ -318,7 +315,6 @@ def generate_valid_test_configs(
             use_gate,
             use_routed_input_transform,
             enable_eplb,
-            reduce_results,
             backend,
             ep_size,
             dp_size,
@@ -395,18 +391,7 @@ def is_valid_config(config: MoETestConfig) -> tuple[bool, str | None]:
         and config.backend.startswith("flashinfer_nvlink")
         and not current_platform.has_device_capability(90)
     ):
-        return False, "flashinfer_nvlink needs an H100+ GPUs"
-
-    # reduce_results incompatibilities
-    if config.reduce_results and config.use_shared_experts:
-        return False, "reduce_results=True is not compatible with shared_experts=True"
-
-    if config.reduce_results and config.quantization is not None:
-        return (
-            False,
-            "reduce_results=True only tested with unquantized data types in "
-            "order to limit number of tests run",
-        )
+        return False, "flashinfer_nvlink needs H100+ GPUs"
 
     # Backend-specific checks
     if config.backend is not None:
@@ -448,10 +433,6 @@ def is_valid_config(config: MoETestConfig) -> tuple[bool, str | None]:
     if config.enable_eplb and config.backend not in EPLB_SUPPORTED_BACKENDS:
         return False, f"EPLB not supported with {config.backend}."
 
-    world_size = config.tp_size * config.dp_size
-    if config.reduce_results and world_size == 1:
-        return False, "reduce_results=True only makes sense for multi-GPU tests"
-
     if (
         config.backend is not None
         and config.backend.startswith("flashinfer_nvlink")
@@ -846,7 +827,6 @@ def make_fused_moe_layer(
     tp_size: int,
     ep_size: int,
     dp_size: int,
-    reduce_results: bool,
     w1: torch.Tensor,
     w2: torch.Tensor,
     top_k: int,
@@ -874,7 +854,7 @@ def make_fused_moe_layer(
     routed_input_transform: torch.nn.Module | None = None,
     routed_output_transform: torch.nn.Module | None = None,
     pcp_size: int | None = 1,
-) -> tuple[Callable, FusedMoE]:
+) -> FusedMoE:
     quant_config, qw = make_quant_config(quantization, w1, w2, global_num_experts)
 
     kwargs = dict()
@@ -887,8 +867,10 @@ def make_fused_moe_layer(
     # Add gate and routed_input_transform if provided
     if gate is not None:
         kwargs["gate"] = gate
+
     if routed_input_transform is not None:
         kwargs["routed_input_transform"] = routed_input_transform
+        kwargs["routed_output_transform"] = routed_output_transform
 
     layer = builder(
         num_experts=global_num_experts,
@@ -896,7 +878,6 @@ def make_fused_moe_layer(
         hidden_size=hidden_size,
         intermediate_size=intermediate_size,
         params_dtype=in_dtype,
-        reduce_results=reduce_results,
         renormalize=renormalize,
         use_grouped_topk=use_grouped_topk,
         num_expert_group=num_expert_group,
@@ -936,36 +917,7 @@ def make_fused_moe_layer(
 
     layer.quant_method.process_weights_after_loading(layer)
 
-    def _moe(
-        hidden_states: torch.Tensor,
-        router_logits: torch.Tensor,
-    ) -> torch.Tensor:
-        if shared_experts is None:
-            final_shared_states = None
-            final_hidden_states = layer(hidden_states, router_logits)
-        else:
-            final_shared_states, final_hidden_states = layer(
-                hidden_states, router_logits
-            )
-
-        # Apply routed output transform if provided
-        # (e.g., latent space -> original space)
-        if routed_output_transform is not None:
-            final_hidden_states = routed_output_transform(final_hidden_states)
-
-        if shared_experts is not None:
-            assert not reduce_results
-            assert final_shared_states is not None
-            final_hidden_states += final_shared_states
-
-        if not reduce_results and layer.tp_size > 1:
-            final_hidden_states = layer.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
-
-        return final_hidden_states
-
-    return _moe, layer
+    return layer
 
 
 def make_fake_moe_layer(
@@ -999,7 +951,6 @@ def make_fake_moe_layer(
     tp_size: int = 1,
     dp_size: int = 1,
     ep_size: int = 1,
-    reduce_results: bool = False,
 ) -> Callable:
     activation = MoEActivation.from_str(activation)
 
@@ -1101,7 +1052,7 @@ def make_fake_moe_layer(
 
 
 def _test_body_regular(
-    moe_fn: Callable,
+    moe_layer: Callable,
     hidden_states: torch.Tensor,
     router_logits: torch.Tensor,
     vllm_config: VllmConfig,
@@ -1118,13 +1069,12 @@ def _test_body_regular(
         num_tokens=num_tokens,
         num_tokens_across_dp=num_tokens_across_dp,
     ):
-        output = moe_fn(hidden_states, router_logits)
+        output = moe_layer(hidden_states, router_logits)
 
     return baseline_output, output
 
 
 def _test_body_eplb(
-    moe_fn: Callable,
     moe_layer: FusedMoE,
     hidden_states: torch.Tensor,
     router_logits: torch.Tensor,
@@ -1145,7 +1095,6 @@ def _test_body_eplb(
     n: int,
     top_k: int,
     shared_experts,
-    reduce_results: bool,
     gate: torch.nn.Module | None,
     routed_input_transform: torch.nn.Module | None,
     routed_output_transform: torch.nn.Module | None,
@@ -1161,7 +1110,7 @@ def _test_body_eplb(
         num_tokens=num_tokens,
         num_tokens_across_dp=num_tokens_across_dp,
     ):
-        output_before = moe_fn(hidden_states, router_logits)
+        output_before = moe_layer(hidden_states, router_logits)
 
     # Create a fresh FusedMoE layer with enable_eplb=True
     # Delete the original layer's registration so the constructor can
@@ -1174,7 +1123,7 @@ def _test_body_eplb(
     # When using routed_input_transform, experts operate in latent space
     hidden_size_for_layer = k // 2 if routed_input_transform is not None else k
 
-    moe_fn, moe_layer = make_fused_moe_layer(
+    eplb_moe_layer = make_fused_moe_layer(
         quantization=quantization,
         use_ep=use_ep,
         hidden_size=hidden_size_for_layer,
@@ -1183,7 +1132,6 @@ def _test_body_eplb(
         tp_size=tp_size,
         ep_size=ep_size,
         dp_size=dp_size,
-        reduce_results=reduce_results,
         w1=w1,
         w2=w2,
         top_k=top_k,
@@ -1196,14 +1144,14 @@ def _test_body_eplb(
     )
 
     # Necessary?
-    if moe_layer._expert_map is not None:
-        moe_layer._expert_map = moe_layer._expert_map.to(device)
+    if eplb_moe_layer._expert_map is not None:
+        eplb_moe_layer._expert_map = eplb_moe_layer._expert_map.to(device)
 
     # All ranks must generate the same permutation
     initial_indices = torch.arange(num_experts, dtype=torch.long)
     shuffled_indices = initial_indices[torch.randperm(num_experts)]
 
-    expert_weights = [list(moe_layer.get_expert_weights())]
+    expert_weights = [list(eplb_moe_layer.get_expert_weights())]
 
     communicator = create_eplb_communicator(
         group_coordinator=get_eplb_group(),
@@ -1227,7 +1175,7 @@ def _test_body_eplb(
         num_experts, dtype=torch.int32, device=device
     )
 
-    moe_layer.set_eplb_state(
+    eplb_moe_layer.set_eplb_state(
         moe_layer_idx=0,
         expert_load_view=torch.zeros(
             (1, num_experts),
@@ -1244,7 +1192,7 @@ def _test_body_eplb(
         ),
     )
 
-    moe_layer.eplb_state.should_record_tensor = torch.ones(
+    eplb_moe_layer.eplb_state.should_record_tensor = torch.ones(
         (), dtype=torch.bool, device=device
     )
 
@@ -1255,7 +1203,7 @@ def _test_body_eplb(
         num_tokens=num_tokens,
         num_tokens_across_dp=num_tokens_across_dp,
     ):
-        output_after = moe_fn(hidden_states, router_logits)
+        output_after = eplb_moe_layer(hidden_states, router_logits)
 
     return output_before, output_after
 
@@ -1274,7 +1222,6 @@ def _run_one_config(
     num_experts: int,
     top_k: int,
     quantization: str | None,
-    reduce_results: bool,
     backend: str | None,
     test_body_fn: Callable,
     use_shared_experts: bool,
@@ -1341,7 +1288,6 @@ def _run_one_config(
         tp_size=tp_size,
         ep_size=ep_size,
         dp_size=dp_size,
-        reduce_results=reduce_results,
     )
 
     baseline_output = baseline_layer(hidden_states, router_logits)
@@ -1369,7 +1315,7 @@ def _run_one_config(
         hidden_size_for_layer = k // 2 if routed_input_transform is not None else k
 
         # Create initial MoE layer
-        moe_fn, moe_layer = make_fused_moe_layer(
+        moe_layer = make_fused_moe_layer(
             quantization=quantization,
             use_ep=use_ep,
             hidden_size=hidden_size_for_layer,
@@ -1378,7 +1324,6 @@ def _run_one_config(
             tp_size=tp_size,
             ep_size=ep_size,
             dp_size=dp_size,
-            reduce_results=reduce_results,
             w1=w1,
             w2=w2,
             top_k=top_k,
@@ -1402,7 +1347,6 @@ def _run_one_config(
 
         # Call the test body function with all necessary context
         expected, actual = test_body_fn(
-            moe_fn=moe_fn,
             moe_layer=moe_layer,
             hidden_states=hidden_states,
             router_logits=router_logits,
@@ -1423,7 +1367,6 @@ def _run_one_config(
             m=m,
             top_k=top_k,
             shared_experts=shared_experts,
-            reduce_results=reduce_results,
             gate=gate,
             routed_input_transform=routed_input_transform,
             routed_output_transform=routed_output_transform,
@@ -1520,7 +1463,6 @@ def test_moe_layer_no_parallel(
         test_config.num_experts,
         test_config.top_k,
         test_config.quantization,
-        test_config.reduce_results,
         test_config.backend,
         _test_body_regular,
         use_shared_experts=test_config.use_shared_experts,
@@ -1578,7 +1520,6 @@ def _parallel_worker(
                 test_config.num_experts,
                 test_config.top_k,
                 test_config.quantization,
-                test_config.reduce_results,
                 test_config.backend,
                 functools.partial(
                     _test_body_config, test_config=test_config, cpu_group=cpu_group
@@ -1597,7 +1538,7 @@ def _parallel_worker(
             failed = failed + 1
             if verbosity > 0:
                 traceback.print_exc()
-                print(f"\n{str(ex)}\nFAILED {ex.__class__}")
+                print(f"\n{str(ex)}\nFAILED")
             else:
                 print("F", end="")
         finally:
diff --git a/tests/kernels/moe/test_shared_fused_moe_routed_transform.py b/tests/kernels/moe/test_shared_fused_moe_routed_transform.py
index 89e23eb0d74..464754c9f1b 100644
--- a/tests/kernels/moe/test_shared_fused_moe_routed_transform.py
+++ b/tests/kernels/moe/test_shared_fused_moe_routed_transform.py
@@ -165,7 +165,6 @@ def test_routed_input_transform_inside_vs_outside(
             top_k=top_k,
             hidden_size=latent_size,
             intermediate_size=intermediate_size,
-            reduce_results=False,
             renormalize=True,
             params_dtype=dtype,
             tp_size=1,
@@ -183,7 +182,6 @@ def test_routed_input_transform_inside_vs_outside(
             top_k=top_k,
             hidden_size=latent_size,
             intermediate_size=intermediate_size,
-            reduce_results=False,
             renormalize=True,
             params_dtype=dtype,
             tp_size=1,
@@ -212,34 +210,20 @@ def test_routed_input_transform_inside_vs_outside(
         hidden_states = torch.randn(num_tokens, hidden_size, device="cuda", dtype=dtype)
         router_logits = torch.randn(num_tokens, num_experts, device="cuda", dtype=dtype)
 
-        # Clone inputs so any in-place modification by Method A
-        # cannot affect Method B's computation.
-        hidden_states_A = hidden_states.clone()
-        router_logits_A = router_logits.clone()
-
         with set_forward_context(None, vllm_config, num_tokens=num_tokens):
-            shared_out_A, routed_out_A = moe_with_transform(
-                hidden_states_A, router_logits_A
-            )
+            # Method A: combined output (shared + routed)
+            combined_A = moe_with_transform(hidden_states, router_logits)
 
+            # Method B: manually transform, get routed output, add shared
             transformed_hidden = routed_transform(hidden_states)
-            shared_out_B, routed_out_B = moe_without_transform(
-                transformed_hidden, router_logits
-            )
+            routed_out_B = moe_without_transform(transformed_hidden, router_logits)
+            shared_out_B = shared_experts(hidden_states)
+            combined_B = shared_out_B + routed_out_B
 
-        expected_shared_out = shared_experts(hidden_states)
-
-        _assert_close(
-            routed_out_A,
-            routed_out_B,
+        torch.testing.assert_close(
+            combined_A,
+            combined_B,
             atol=1e-3,
             rtol=1e-3,
-            label="Routed output: transform inside vs outside",
-        )
-        _assert_close(
-            shared_out_A,
-            expected_shared_out,
-            atol=1e-3,
-            rtol=1e-3,
-            label="Shared expert output",
+            msg="Combined output should match: transform inside vs outside",
         )
diff --git a/vllm/lora/layers/fused_moe.py b/vllm/lora/layers/fused_moe.py
index 835bffe58ca..d6eec675c6d 100644
--- a/vllm/lora/layers/fused_moe.py
+++ b/vllm/lora/layers/fused_moe.py
@@ -592,9 +592,6 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
     def forward(self, *args, **kwargs):
         return self.base_layer.forward(*args, **kwargs)
 
-    def maybe_all_reduce_tensor_model_parallel(self, *args, **kwargs):
-        return self.base_layer.maybe_all_reduce_tensor_model_parallel(*args, **kwargs)
-
     @property
     def quant_method(self):
         return self.base_layer.quant_method
diff --git a/vllm/model_executor/layers/fused_moe/fused_marlin_moe.py b/vllm/model_executor/layers/fused_moe/fused_marlin_moe.py
index 6c916cf3cb6..daae5b6bd16 100644
--- a/vllm/model_executor/layers/fused_moe/fused_marlin_moe.py
+++ b/vllm/model_executor/layers/fused_moe/fused_marlin_moe.py
@@ -716,7 +716,7 @@ class MarlinExperts(MarlinExpertsBase):
     ):
         assert self.w1_scale is not None
         assert self.w2_scale is not None
-        return fused_marlin_moe(
+        fused_marlin_moe(
             hidden_states=hidden_states,
             w1=w1,
             w2=w2,
diff --git a/vllm/model_executor/layers/fused_moe/layer.py b/vllm/model_executor/layers/fused_moe/layer.py
index 190a9cc3b5d..bf10bc9d5c4 100644
--- a/vllm/model_executor/layers/fused_moe/layer.py
+++ b/vllm/model_executor/layers/fused_moe/layer.py
@@ -230,11 +230,18 @@ class FusedMoE(PluggableLayer):
         hidden_size: Input hidden state size of the transformer
         intermediate_size: Intermediate size of the experts
         params_dtype: Data type for the parameters.
-        reduce_results: Whether to all_reduce on the output of the layer
         renormalize: Whether to renormalize the logits in the fused_moe kernel
         quant_config: Quantization configure.
         enable_eplb: Whether to enable expert parallelism load balancer.
         router_logits_dtype: Data type for router logits buffers.
+        routed_scaling_factor: A scaling factor that is applied to the topk_weights
+                               by the router or the output of the layer depending
+                               on the value of `apply_routed_scale_to_output`
+        apply_routed_scale_to_output: Determine whether or not `routed_scaling_factor`
+                                      is applied to the topk_weights or to the experts
+                                      output. It is applied to the experts output
+                                      instead of the topk_weights when this feature is
+                                      not supported by the router (or the experts).
     """
 
     # --8<-- [end:fused_moe]
@@ -246,7 +253,6 @@ class FusedMoE(PluggableLayer):
         hidden_size: int,
         intermediate_size: int,
         params_dtype: torch.dtype | None = None,
-        reduce_results: bool = False,
         renormalize: bool = True,
         use_grouped_topk: bool = False,
         num_expert_group: int | None = None,
@@ -274,12 +280,12 @@ class FusedMoE(PluggableLayer):
         gate: torch.nn.Module | None = None,
         shared_experts: torch.nn.Module | None = None,
         routed_input_transform: torch.nn.Module | None = None,
+        routed_output_transform: torch.nn.Module | None = None,
+        apply_routed_scale_to_output: bool = False,
         zero_expert_type: str | None = None,
     ):
         super().__init__()
 
-        self._routed_input_transform = routed_input_transform
-
         if params_dtype is None:
             params_dtype = torch.get_default_dtype()
         self.params_dtype = params_dtype
@@ -425,7 +431,6 @@ class FusedMoE(PluggableLayer):
 
         assert intermediate_size % self.tp_size == 0
         intermediate_size_per_partition = intermediate_size // self.tp_size
-        self.reduce_results = reduce_results
         self.renormalize = renormalize
 
         # TODO(bnell): these attributes are only used by monolithic kernels.
@@ -437,7 +442,14 @@ class FusedMoE(PluggableLayer):
         self.topk_group = topk_group
         self.custom_routing_function = custom_routing_function
         self.scoring_func = scoring_func
-        self.routed_scaling_factor = routed_scaling_factor
+        # When apply_routed_scale_to_output is True, we set the scaling factor
+        # to 1.0 so it ends up being a nop. Applying the scale will be handled
+        # by the runner in this case.
+        # The member variable must be set in the same way as the router since
+        # some quantization methods can access it.
+        self.routed_scaling_factor = (
+            routed_scaling_factor if not apply_routed_scale_to_output else 1.0
+        )
         self.e_score_correction_bias = e_score_correction_bias
         # TODO(bnell): end attributes
 
@@ -456,7 +468,7 @@ class FusedMoE(PluggableLayer):
             topk_group=topk_group,
             custom_routing_function=custom_routing_function,
             scoring_func=scoring_func,
-            routed_scaling_factor=routed_scaling_factor,
+            routed_scaling_factor=self.routed_scaling_factor,
             e_score_correction_bias=e_score_correction_bias,
             num_fused_shared_experts=self.num_fused_shared_experts,
             enable_eplb=enable_eplb,
@@ -578,12 +590,18 @@ class FusedMoE(PluggableLayer):
             layer_name=self.layer_name,
             moe_config=self.moe_config,
             router=self.router,
-            routed_input_transform=self._routed_input_transform,
             gate=gate,
             shared_experts=shared_experts,
             quant_method=self.quant_method,
-            reduce_results=self.reduce_results,
             enable_dbo=self.vllm_config.parallel_config.enable_dbo,
+            routed_input_transform=routed_input_transform,
+            routed_output_transform=routed_output_transform,
+            # When apply_routed_scale_to_output is True, we allow
+            # the scaling factor to be passed to the runner, otherwise
+            # we pass 1.0 so it ends up being a nop.
+            routed_scaling_factor=routed_scaling_factor
+            if apply_routed_scale_to_output
+            else 1.0,
         )
 
     # TODO(bnell): This method is provided as a hook so vllm/lora/layers/fused_moe.py
@@ -1514,32 +1532,11 @@ class FusedMoE(PluggableLayer):
         self.ensure_moe_quant_config_init()
         return self.quant_method.moe_quant_config
 
-    def must_reduce_shared_expert_outputs(self) -> bool:
-        """
-        The shared_experts are typically computed using the RowParallelLinear
-        layer. The result of this function is typically used as
-        the reduce_results argument to the module.
-        When just tensor-parallel is used, it is not required to reduce
-        the shared_experts results immediately. Instead we reduce at the
-        once at the end of the MoE op. (Refer to DeepSeekV2MoE module)
-        With EP and all2all kernels - this is no longer viable as all
-        GPU ranks in DP, produce the complete set of hidden_states.
-        Therefore it is required that we reduce the shared_experts output
-        early.
-        """
-        return self.runner.must_reduce_shared_expert_outputs()
-
-    def maybe_all_reduce_tensor_model_parallel(self, final_hidden_states: torch.Tensor):
-        """
-        Some combine kernels reduce across GPU ranks by default.
-        """
-        return self.runner.maybe_all_reduce_tensor_model_parallel(final_hidden_states)
-
     def forward(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
-    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+    ) -> torch.Tensor:
         return self.runner.forward(
             hidden_states,
             router_logits,
@@ -1613,7 +1610,6 @@ class FusedMoE(PluggableLayer):
             f"intermediate_size_per_partition={self.intermediate_size_per_partition}, "  # noqa: E501
             f"tp_size={self.tp_size},\n"
             f"ep_size={self.ep_size}, "
-            f"reduce_results={self.reduce_results}, "
         )
 
         return s
diff --git a/vllm/model_executor/layers/fused_moe/modular_kernel.py b/vllm/model_executor/layers/fused_moe/modular_kernel.py
index f2e6e2560e7..376fcdf5b65 100644
--- a/vllm/model_executor/layers/fused_moe/modular_kernel.py
+++ b/vllm/model_executor/layers/fused_moe/modular_kernel.py
@@ -1261,7 +1261,7 @@ class FusedMoEKernelModularImpl:
         topk_ids: torch.Tensor,
         apply_router_weight_on_input: bool,
         shared_experts_input: torch.Tensor | None,
-    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+    ) -> torch.Tensor:
         """
         The _finalize method is a wrapper around self.prepare_finalize.finalize
         that handles DBO, async and shared expert overlap.
diff --git a/vllm/model_executor/layers/fused_moe/runner/default_moe_runner.py b/vllm/model_executor/layers/fused_moe/runner/default_moe_runner.py
index 85c6563c084..8cd2fc65704 100644
--- a/vllm/model_executor/layers/fused_moe/runner/default_moe_runner.py
+++ b/vllm/model_executor/layers/fused_moe/runner/default_moe_runner.py
@@ -37,10 +37,6 @@ class DefaultMoERunner(MoERunnerBase):
     for different configurations (e.g., with/without shared experts, gates, etc.).
     """
 
-    @property
-    def reduce_results(self) -> bool:
-        return self._reduce_results
-
     @property
     def do_naive_dispatch_combine(self) -> bool:
         return (
diff --git a/vllm/model_executor/layers/fused_moe/runner/moe_runner.py b/vllm/model_executor/layers/fused_moe/runner/moe_runner.py
index 9ffbf3108f8..199ceab0659 100644
--- a/vllm/model_executor/layers/fused_moe/runner/moe_runner.py
+++ b/vllm/model_executor/layers/fused_moe/runner/moe_runner.py
@@ -26,18 +26,7 @@ class MoERunner(ABC):
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
-    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
-        raise NotImplementedError
-
-    @abstractmethod
-    def must_reduce_shared_expert_outputs(self) -> bool:
-        raise NotImplementedError
-
-    @abstractmethod
-    def maybe_all_reduce_tensor_model_parallel(
-        self,
-        final_hidden_states: torch.Tensor,
-    ):
+    ) -> torch.Tensor:
         raise NotImplementedError
 
     @abstractmethod
diff --git a/vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py b/vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py
index 692d45d3460..2d2d0bb4ebe 100644
--- a/vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py
+++ b/vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py
@@ -81,7 +81,9 @@ def _resolve_layer_name(layer_name: str | LayerName) -> str:
 
 # Note: _moe_forward and _moe_forward_shared should not contain any
 # implementation details, They should merely pass along control to
-# the runner's 'forward_dispatch' method.
+# the runner's '_forward_dispatch' method.
+# These functions should never be called directly since they do not
+# include all the functionality of the MoE layer.
 def _moe_forward(
     hidden_states: torch.Tensor,
     router_logits: torch.Tensor,
@@ -89,7 +91,7 @@ def _moe_forward(
     layer_name: _layer_name_type,
 ) -> torch.Tensor:
     layer = get_layer_from_name(_resolve_layer_name(layer_name))
-    return layer.runner.forward_dispatch(
+    return layer.runner._forward_dispatch(
         layer,
         hidden_states,
         router_logits,
@@ -113,7 +115,7 @@ def _moe_forward_shared(
     layer_name: _layer_name_type,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     layer = get_layer_from_name(_resolve_layer_name(layer_name))
-    return layer.runner.forward_dispatch(
+    return layer.runner._forward_dispatch(
         layer,
         hidden_states,
         router_logits,
@@ -143,7 +145,7 @@ def _moe_forward_shared_fake(
 direct_register_custom_op(
     op_name="moe_forward",
     op_func=_moe_forward,
-    mutates_args=["hidden_states"],  # is this still true?
+    mutates_args=["hidden_states"],
     fake_impl=_moe_forward_fake,
     tags=(torch.Tag.needs_fixed_stride_order,),
 )
@@ -157,6 +159,15 @@ direct_register_custom_op(
 )
 
 
+def _unpack(
+    result: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+) -> tuple[torch.Tensor | None, torch.Tensor]:
+    if isinstance(result, tuple):
+        return result
+    else:
+        return (None, result)
+
+
 class MoERunnerBase(MoERunner):
     """
     Abstract base class providing common functionality for MoE runner implementations.
@@ -174,7 +185,6 @@ class MoERunnerBase(MoERunner):
     allowing flexibility in the actual MoE computation implementation.
 
     Key abstract methods that subclasses must implement:
-    - reduce_results: Determines whether results should be reduced across ranks
     - _forward_impl: The core MoE computation logic specific to each runner type
     """
 
@@ -187,17 +197,23 @@ class MoERunnerBase(MoERunner):
         gate: torch.nn.Module | None,
         shared_experts: torch.nn.Module | None,
         quant_method: FusedMoEMethodBase,
-        reduce_results: bool,
         enable_dbo: bool,
+        routed_output_transform: torch.nn.Module | None = None,
+        routed_scaling_factor: float = 1.0,
     ):
         super().__init__()
         self.moe_config = moe_config
         self.router = router
         self.routed_input_transform = routed_input_transform
+        self.routed_output_transform = routed_output_transform
+        self.routed_scaling_factor = routed_scaling_factor
         self.gate = gate
         self.quant_method = quant_method
-        self._reduce_results = reduce_results
         self.enable_dbo = enable_dbo
+        self._fused_output_is_reduced = (
+            self.quant_method.moe_kernel is not None
+            and self.quant_method.moe_kernel.output_is_reduced()
+        )
 
         self._shared_experts: SharedExperts | None = None
         if shared_experts is not None:
@@ -209,7 +225,6 @@ class MoERunnerBase(MoERunner):
                 # called, i.e. by a MK or by the MoERunner.
                 # Once the MK can be created upfront, we can just pass in the proper
                 # flags derived from the quant_method's MK.
-                reduce_results=reduce_results,
                 quant_method=quant_method,
                 enable_dbo=enable_dbo,
             )
@@ -217,7 +232,7 @@ class MoERunnerBase(MoERunner):
         # Needed for string -> FusedMoE layer lookup in custom ops.
         self.layer_name = layer_name
 
-        self.forward_entry = self._select_forward()
+        self._forward_entry = self._select_forward()
 
     def _select_forward(self) -> Callable:
         if current_platform.is_tpu() or current_platform.is_cpu():
@@ -245,38 +260,6 @@ class MoERunnerBase(MoERunner):
     def is_internal_router(self) -> bool:
         return self.gate is not None
 
-    @property
-    @abstractmethod
-    def reduce_results(self) -> bool:
-        raise NotImplementedError
-
-    def must_reduce_shared_expert_outputs(self) -> bool:
-        """
-        The shared_experts are typically computed using the RowParallelLinear
-        layer. The result of this function is typically used as
-        the reduce_results argument to the module.
-        When just tensor-parallel is used, it is not required to reduce
-        the shared_experts results immediately. Instead we reduce at the
-        once at the end of the MoE op. (Refer to DeepSeekV2MoE module)
-        With EP and all2all kernels - this is no longer viable as all
-        GPU ranks in DP, produce the complete set of hidden_states.
-        Therefore it is required that we reduce the shared_experts output
-        early.
-        """
-        return (
-            self.quant_method.moe_kernel is not None
-            and self.quant_method.moe_kernel.output_is_reduced()
-        )
-
-    def maybe_all_reduce_tensor_model_parallel(self, final_hidden_states: torch.Tensor):
-        """
-        Some combine kernels reduce across GPU ranks by default.
-        """
-        if self.must_reduce_shared_expert_outputs():
-            return final_hidden_states
-        else:
-            return tensor_model_parallel_all_reduce(final_hidden_states)
-
     def apply_routed_input_transform(
         self, hidden_states: torch.Tensor
     ) -> tuple[torch.Tensor, torch.Tensor | None]:
@@ -286,10 +269,6 @@ class MoERunnerBase(MoERunner):
         is saved separately so shared experts get [S, hidden_size] while
         routed experts get the transformed [S, moe_latent_size].
 
-        TODO: For latent MoE bandwidth optimization, fc2_latent_proj could be
-        moved inside SharedFusedMoE to all-reduce on the smaller latent
-        dimension.
-
         Returns (possibly transformed) hidden states and the input for shared
         experts (or None if there are no shared experts).
         """
@@ -306,33 +285,79 @@ class MoERunnerBase(MoERunner):
             hidden_states if self._shared_experts is not None else None,
         )
 
-    def _maybe_reduce_output(
+    def apply_routed_output_transform(
         self,
-        states: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
-        trunc_sizes: list[int],
-    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
-        def trunc(x: torch.Tensor, trunc_size: int) -> torch.Tensor:
-            return x[..., :trunc_size]
+        fused_output: torch.Tensor,
+    ) -> torch.Tensor:
+        """Apply transform to routed expert output (e.g., latent to full dim).
 
-        def reduce_and_trunc(x: torch.Tensor, trunc_size: int) -> torch.Tensor:
-            return trunc(self.maybe_all_reduce_tensor_model_parallel(x), trunc_size)
+        Used by latent MoE models (e.g., NemotronH) where routed experts
+        operate in a compressed latent space and need projection back to
+        the full hidden dimension before combining with shared expert output.
+        """
+        if self.routed_output_transform is not None:
+            r = self.routed_output_transform(fused_output)
+            fused_output = r[0] if isinstance(r, tuple) else r
+        return fused_output
 
+    def _maybe_apply_routed_scale_to_output(
+        self,
+        shared_output: torch.Tensor | None,
+        fused_output: torch.Tensor,
+    ) -> tuple[torch.Tensor | None, torch.Tensor]:
+        """Apply routed_scaling_factor to the output with FP16 overflow
+        protection.
+
+        Scale the fused expert output by routed_scaling_factor. For FP16,
+        avoid overflow by dividing shared_output by the scale instead
+        (the decoder layer compensates with matching divisions).
+        """
+        if self.routed_scaling_factor != 1.0:
+            if fused_output.dtype != torch.float16:
+                fused_output *= self.routed_scaling_factor
+            elif shared_output is not None:
+                shared_output *= 1.0 / self.routed_scaling_factor
+        return shared_output, fused_output
+
+    def _maybe_reduce_shared_expert_output(
+        self,
+        shared_output: torch.Tensor | None,
+    ) -> torch.Tensor | None:
+        """All-reduce shared expert output when the combine kernel already
+        reduced fused output.
+
+        This is the "early" all-reduce path. When the combine kernel produces
+        already-reduced fused output, shared output must be reduced separately
+        to match.
+        """
+        if self._fused_output_is_reduced:
+            assert shared_output is not None
+            shared_output = tensor_model_parallel_all_reduce(shared_output)
+        return shared_output
+
+    def _maybe_reduce_final_output(
+        self,
+        states: torch.Tensor,
+        trunc_size: int,
+    ) -> torch.Tensor:
+        """Truncate padded dimensions and all-reduce the combined output.
+
+        This is the "late" all-reduce path. When neither fused nor shared
+        output was individually reduced, the combined sum is all-reduced
+        here. Skipped when sequence-parallel is active (SP handles its
+        own reduction) or when the early path already reduced both outputs.
+        """
+        # We don't need to reduce the final output if:
+        # - We are not running with TP or DP
+        # - The MK already reduced the fused output itself.
         if (
             not self.moe_config.is_sequence_parallel
-            and self.reduce_results
             and (self.moe_config.tp_size > 1 or self.moe_config.ep_size > 1)
+            and not self._fused_output_is_reduced
         ):
-            func = reduce_and_trunc
-        else:
-            func = trunc
+            states = tensor_model_parallel_all_reduce(states)
 
-        if isinstance(states, tuple):
-            return tuple(
-                [func(s, trunc_size) for s, trunc_size in zip(states, trunc_sizes)]
-            )
-        else:
-            assert len(trunc_sizes) == 1
-            return func(states, trunc_sizes[0])
+        return states[..., :trunc_size]
 
     def _encode_layer_name(self) -> str | LayerName:
         if _USE_LAYERNAME:
@@ -349,7 +374,15 @@ class MoERunnerBase(MoERunner):
         self,
         shared_experts_input: torch.Tensor | None,
         hidden_states: torch.Tensor,
-    ) -> tuple[torch.Tensor, list[int]]:
+    ) -> tuple[torch.Tensor, int]:
+        """Pad hidden_states to moe_config.hidden_dim and compute the
+        original dimension for later truncation.
+
+        For latent MoE, the routed hidden_states may be smaller than
+        hidden_dim. Padding ensures uniform tensor sizes through the
+        fused MoE kernel. The returned trunc_size is used by
+        _maybe_reduce_final_output to strip the padding from the result.
+        """
         shared_experts_hidden_dim = (
             shared_experts_input.shape[-1] if shared_experts_input is not None else 0
         )
@@ -365,10 +398,10 @@ class MoERunnerBase(MoERunner):
                 value=0.0,
             )
 
-        if self._shared_experts is not None:
-            orig_hidden_dims = [shared_experts_hidden_dim, transformed_hidden_dim]
+        if self.routed_output_transform is not None and shared_experts_hidden_dim > 0:
+            orig_hidden_dims = shared_experts_hidden_dim
         else:
-            orig_hidden_dims = [transformed_hidden_dim]
+            orig_hidden_dims = transformed_hidden_dim
 
         return hidden_states, orig_hidden_dims
 
@@ -388,6 +421,12 @@ class MoERunnerBase(MoERunner):
         router_logits: torch.Tensor,
         shared_experts_input: torch.Tensor | None,
     ) -> tuple[torch.Tensor | None, torch.Tensor]:
+        """Run expert routing and the fused MoE kernel via the quant method.
+
+        Orchestrates shared expert execution (before/after), expert selection
+        via the router, and the actual fused MoE computation. Returns
+        (shared_expert_output, fused_expert_output).
+        """
         # Run this before quant_method to avoid inplace issues.
         # TODO(bnell): probably not needed anymore since inplace is
         # disabled when shared experts are present.
@@ -428,6 +467,13 @@ class MoERunnerBase(MoERunner):
         )
 
     def _sequence_parallel_context(self):
+        """Return a context manager for sequence-parallel token
+        redistribution.
+
+        When sequence parallelism is active, returns a context that handles
+        local size tracking for proper token scatter/gather. Otherwise
+        returns a no-op context.
+        """
         ctx = get_forward_context()
         return (
             ctx.dp_metadata.sp_local_sizes(self.moe_config.sp_size)
@@ -448,22 +494,25 @@ class MoERunnerBase(MoERunner):
 
     def _maybe_add_zero_expert_output(
         self,
-        result: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
-    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+        result: torch.Tensor,
+    ) -> torch.Tensor:
+        """Add the zero expert's contribution to the final result.
+
+        When a ZeroExpertRouter is used, it computes a bias-like output
+        from the "zero expert" that is added to the combined routed+shared
+        expert output.
+        """
         if isinstance(self.router, ZeroExpertRouter):
             zero_expert_output = self.router.zero_expert_output
             assert zero_expert_output is not None
-            if isinstance(result, tuple):
-                result = (result[0], result[1] + zero_expert_output)
-            else:
-                result = result + zero_expert_output
+            result = result + zero_expert_output
         return result
 
     def forward(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
-    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+    ) -> torch.Tensor:
         """Invoke the fused moe layer.
 
         Input:
@@ -472,55 +521,88 @@ class MoERunnerBase(MoERunner):
 
         Output:
         - The new hidden_states.
-        or
-        - A tuple of (shared experts output, new hidden_states).
 
         Calling sequence
         - forward
-          - self.forward_entry (_moe_forward or _moe_forward_shared custom op)
-            - forward_dispatch
+          - self._forward_entry (_moe_forward or _moe_forward_shared custom op)
+            - _forward_dispatch
               - _forward_impl
 
         Note: The existence of _moe_forward and _moe_forward_shared custom ops are due
         to the following reasons:
         1. the chunking loop in ChunkingMoERunner._forward_impl cannot be compiled by
            torch.compile
-        2. pytorch cannot handle union types in custom op signatures so _moe_forward
-           and _moe_forward_shared must be split.
+        2. pytorch cannot handle union types in custom op signatures so
+           _moe_forward and _moe_forward_shared must be split.
 
         If ChunkingMoERunner._forward_impl can be implemented via torch.scan we can
         potentially get rid of _moe_forward and _moe_forward_shared and collapse the
         whole sequence into the 'forward' method.
         """
 
-        # Apply transform for routed experts (e.g., latent projection for latent MoE)
+        # Apply transform for routed experts (e.g., latent projection
+        # for latent MoE)
         hidden_states, shared_experts_input = self.apply_routed_input_transform(
             hidden_states
         )
 
-        hidden_states, og_hidden_dims = self._maybe_pad_hidden_states(
+        hidden_states, og_hidden_dim = self._maybe_pad_hidden_states(
             shared_experts_input,
             hidden_states,
         )
 
-        fused_output = self.forward_entry(
+        result = self._forward_entry(
             hidden_states,
             router_logits,
             shared_experts_input,
             self._encode_layer_name(),
         )
 
-        result = self._maybe_reduce_output(fused_output, og_hidden_dims)
+        #
+        # Note: there are two all-reduce points below. They are mutually
+        # exclusive, controlled by _fused_output_is_reduced
+        #  - When True: the combine kernel already reduced fused_output,
+        #    so we reduce shared_output here to match, then skip the
+        #    all-reduce in _maybe_reduce_final_output.
+        #  - When False: neither output is reduced yet, so we combine
+        #    them first and all-reduce the sum in _maybe_reduce_final_output.
+
+        # Extract outputs from result
+        shared_output, fused_output = _unpack(result)
+
+        # If combine kernel already reduced fused, reduce shared to match.
+        # See note above re: the two all-reduce points.
+        shared_output = self._maybe_reduce_shared_expert_output(shared_output)
+
+        shared_output, fused_output = self._maybe_apply_routed_scale_to_output(
+            shared_output, fused_output
+        )
+
+        # Apply output transform (e.g. latent -> full dim)
+        fused_output = self.apply_routed_output_transform(fused_output)
+
+        if shared_output is not None:
+            result = shared_output + fused_output
+        else:
+            result = fused_output
+
+        result = self._maybe_reduce_final_output(result, og_hidden_dim)
 
         return self._maybe_add_zero_expert_output(result)
 
-    def forward_dispatch(
+    def _forward_dispatch(
         self,
         layer: torch.nn.Module,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
         shared_experts_input: torch.Tensor | None,
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+        """Entry point called by the custom op to run the MoE computation.
+
+        Handles pre-dispatch setup (gate application, external shared expert
+        triggering, quant config init) then delegates to _forward_impl within
+        the sequence-parallel context.
+        """
         # TODO(bnell): this can be removed after MK migration is complete.
         layer.ensure_moe_quant_config_init()
 
@@ -549,4 +631,11 @@ class MoERunnerBase(MoERunner):
         router_logits: torch.Tensor,
         shared_experts_input: torch.Tensor | None,
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+        """Core MoE computation to be implemented by subclasses.
+
+        Performs expert routing, fused MoE kernel execution, and shared
+        expert computation. Returns a single tensor (fused output only)
+        or a tuple of (shared_output, fused_output) when shared experts
+        are present.
+        """
         raise NotImplementedError
diff --git a/vllm/model_executor/layers/fused_moe/runner/moe_runner_factory.py b/vllm/model_executor/layers/fused_moe/runner/moe_runner_factory.py
index 2143fa3ce08..feb4614d837 100644
--- a/vllm/model_executor/layers/fused_moe/runner/moe_runner_factory.py
+++ b/vllm/model_executor/layers/fused_moe/runner/moe_runner_factory.py
@@ -29,8 +29,9 @@ def create_moe_runner(
     gate: torch.nn.Module | None,
     shared_experts: SharedExperts | None,
     quant_method: FusedMoEMethodBase,
-    reduce_results: bool,
     enable_dbo: bool,
+    routed_output_transform: torch.nn.Module | None = None,
+    routed_scaling_factor: float = 1.0,
 ) -> MoERunner:
     return DefaultMoERunner(
         layer_name,
@@ -40,6 +41,7 @@ def create_moe_runner(
         gate,
         shared_experts,
         quant_method,
-        reduce_results,
         enable_dbo,
+        routed_output_transform=routed_output_transform,
+        routed_scaling_factor=routed_scaling_factor,
     )
diff --git a/vllm/model_executor/layers/fused_moe/runner/shared_experts.py b/vllm/model_executor/layers/fused_moe/runner/shared_experts.py
index 827a6e6bd3e..3a08d5d0fc2 100644
--- a/vllm/model_executor/layers/fused_moe/runner/shared_experts.py
+++ b/vllm/model_executor/layers/fused_moe/runner/shared_experts.py
@@ -5,10 +5,6 @@ from enum import IntEnum
 import torch
 
 import vllm.envs as envs
-from vllm.distributed import (
-    get_tensor_model_parallel_world_size,
-    tensor_model_parallel_all_reduce,
-)
 from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEConfig,
@@ -48,7 +44,6 @@ class SharedExperts:
         layer: torch.nn.Module,
         moe_config: FusedMoEConfig,
         quant_method: QuantizeMethodBase,
-        reduce_results: bool,
         enable_dbo: bool,
     ):
         from vllm.model_executor.layers.fused_moe.fused_moe_method_base import (
@@ -68,7 +63,6 @@ class SharedExperts:
         self._layer = layer
         self._moe_config = moe_config
         self._quant_method = quant_method
-        self._reduce_results = reduce_results
 
         # Allow disabling of the separate shared experts stream for
         # debug purposes.
@@ -139,18 +133,6 @@ class SharedExperts:
 
         return output
 
-    def _maybe_reduce_shared_out(self, shared_out: torch.Tensor) -> torch.Tensor:
-        # Reduce shared expert outputs if necessary, since the MLP
-        # should have been created with reduce_results=False.
-        if (
-            self._reduce_results
-            and self._quant_method.moe_kernel is not None
-            and self._quant_method.moe_kernel.output_is_reduced()
-            and get_tensor_model_parallel_world_size() > 1
-        ):
-            shared_out = tensor_model_parallel_all_reduce(shared_out)
-        return shared_out
-
     @property
     def _output_idx(self) -> int:
         return dbo_current_ubatch_id() if self.enable_dbo else 0
diff --git a/vllm/model_executor/layers/fused_moe/shared_fused_moe.py b/vllm/model_executor/layers/fused_moe/shared_fused_moe.py
index ed243e99263..9cfcb1baa9b 100644
--- a/vllm/model_executor/layers/fused_moe/shared_fused_moe.py
+++ b/vllm/model_executor/layers/fused_moe/shared_fused_moe.py
@@ -18,12 +18,8 @@ class SharedFusedMoE(FusedMoE):
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
-    ) -> tuple[torch.Tensor, torch.Tensor]:
-        result = super().forward(
+    ) -> torch.Tensor:
+        return super().forward(
             hidden_states=hidden_states,
             router_logits=router_logits,
         )
-        if self.shared_experts is None:
-            return None, result
-        else:
-            return result
diff --git a/vllm/model_executor/models/AXK1.py b/vllm/model_executor/models/AXK1.py
index f5ed4400fb6..d42fbed42ae 100644
--- a/vllm/model_executor/models/AXK1.py
+++ b/vllm/model_executor/models/AXK1.py
@@ -100,7 +100,7 @@ class AXK1MoE(nn.Module):
         self.tp_size = get_tensor_model_parallel_world_size()
         self.tp_rank = get_tensor_model_parallel_rank()
 
-        self.routed_scaling_factor = config.routed_scaling_factor
+        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
 
         self.ep_group = get_ep_group().device_group
         self.ep_rank = get_ep_group().rank_in_group
@@ -170,7 +170,6 @@ class AXK1MoE(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -180,9 +179,8 @@ class AXK1MoE(nn.Module):
             scoring_func=config.scoring_func,
             # we do scaling outside, set factor to 1.0 to avoid double mul
             # aiter applies routed_scaling_factor internally
-            routed_scaling_factor=1.0
-            if not self.is_rocm_aiter_moe_enabled
-            else self.routed_scaling_factor,
+            routed_scaling_factor=self.routed_scaling_factor,
+            apply_routed_scale_to_output=not self.is_rocm_aiter_moe_enabled,
             e_score_correction_bias=self.gate.e_score_correction_bias,
             enable_eplb=self.enable_eplb,
             num_redundant_experts=self.n_redundant_experts,
@@ -204,43 +202,20 @@ class AXK1MoE(nn.Module):
             hidden_states = sequence_parallel_chunk(hidden_states)
 
         if self.experts.is_internal_router:
-            # In this case, the gate/router runs inside the FusedMoE class
-            fused_moe_out = self.experts(
+            final_hidden_states = self.experts(
                 hidden_states=hidden_states, router_logits=hidden_states
             )
         else:
-            # router_logits: (num_tokens, n_experts)
             router_logits, _ = self.gate(hidden_states)
-            fused_moe_out = self.experts(
+            final_hidden_states = self.experts(
                 hidden_states=hidden_states, router_logits=router_logits
             )
 
-        shared_output, final_hidden_states = fused_moe_out
-        if self.shared_experts is None:
-            assert shared_output is None
-
-        # Fix FP16 overflow
-        # See AXK1DecoderLayer for more details.
-        if hidden_states.dtype != torch.float16:
-            if not self.is_rocm_aiter_moe_enabled:
-                final_hidden_states *= self.routed_scaling_factor
-        elif self.shared_experts is not None:
-            assert shared_output is not None
-            shared_output *= 1.0 / self.routed_scaling_factor
-
-        if self.shared_experts is not None:
-            assert shared_output is not None
-            final_hidden_states += shared_output
-
         if self.is_sequence_parallel:
             final_hidden_states = tensor_model_parallel_all_gather(
                 final_hidden_states, 0
             )
             final_hidden_states = final_hidden_states[:num_tokens]
-        elif self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
 
         return final_hidden_states.view(num_tokens, hidden_dim)
 
diff --git a/vllm/model_executor/models/afmoe.py b/vllm/model_executor/models/afmoe.py
index 22037336411..5bad52a0c49 100644
--- a/vllm/model_executor/models/afmoe.py
+++ b/vllm/model_executor/models/afmoe.py
@@ -131,7 +131,6 @@ class AfmoeMoE(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=self.route_norm if self.score_func == "sigmoid" else False,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -152,20 +151,10 @@ class AfmoeMoE(nn.Module):
 
         router_logits = self.gate(hidden_states.to(dtype=torch.float32))
 
-        fused_moe_out = self.experts(
+        final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
 
-        if self.shared_experts is not None:
-            shared_output, final_hidden_states = fused_moe_out
-            final_hidden_states = final_hidden_states + shared_output
-        else:
-            final_hidden_states = fused_moe_out
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
-
         return final_hidden_states.view(num_tokens, hidden_dim)
 
 
diff --git a/vllm/model_executor/models/aria.py b/vllm/model_executor/models/aria.py
index 7b891f8ee42..7a079c56540 100644
--- a/vllm/model_executor/models/aria.py
+++ b/vllm/model_executor/models/aria.py
@@ -283,7 +283,6 @@ class AriaTextMoELayer(nn.Module):
             hidden_size=config.hidden_size,
             intermediate_size=config.intermediate_size,
             quant_config=quant_config,
-            reduce_results=True,
             prefix=f"{prefix}.experts",
         )
 
@@ -301,12 +300,7 @@ class AriaTextMoELayer(nn.Module):
 
         router_output = torch.nn.functional.linear(hidden_states, self.router_weight)
 
-        sparse_expert_output = self.experts(hidden_states, router_output)
-
-        if self.shared_experts is not None:
-            return sparse_expert_output[0] + sparse_expert_output[1]
-        else:
-            return sparse_expert_output
+        return self.experts(hidden_states, router_output)
 
 
 class AriaTextDecoderLayer(LlamaDecoderLayer):
diff --git a/vllm/model_executor/models/bailing_moe.py b/vllm/model_executor/models/bailing_moe.py
index 7725dfa2a88..510d605f804 100644
--- a/vllm/model_executor/models/bailing_moe.py
+++ b/vllm/model_executor/models/bailing_moe.py
@@ -291,7 +291,6 @@ class BailingMoE(nn.Module):
             top_k=self.top_k,
             hidden_size=self.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=self.norm_expert_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -301,6 +300,7 @@ class BailingMoE(nn.Module):
             topk_group=self.topk_group,
             use_grouped_topk=self.use_grouped_topk,
             router_logits_dtype=self.router_dtype,
+            routed_scaling_factor=self.routed_scaling_factor,
         )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -314,21 +314,6 @@ class BailingMoE(nn.Module):
         final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-
-        if self.shared_experts is not None:
-            shared_output, final_hidden_states = final_hidden_states
-        else:
-            shared_output = None
-
-        final_hidden_states *= self.routed_scaling_factor
-
-        if shared_output is not None:
-            final_hidden_states = final_hidden_states + shared_output
-
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
         return final_hidden_states.view(num_tokens, hidden_size)
 
 
diff --git a/vllm/model_executor/models/bailing_moe_linear.py b/vllm/model_executor/models/bailing_moe_linear.py
index ecc5d63ced7..a63ad83f45b 100644
--- a/vllm/model_executor/models/bailing_moe_linear.py
+++ b/vllm/model_executor/models/bailing_moe_linear.py
@@ -358,7 +358,6 @@ class BailingMoeV25(nn.Module):
             top_k=self.top_k,
             hidden_size=self.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=self.norm_expert_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -368,6 +367,8 @@ class BailingMoeV25(nn.Module):
             topk_group=self.topk_group,
             use_grouped_topk=self.use_grouped_topk,
             router_logits_dtype=self.router_dtype,
+            routed_scaling_factor=self.routed_scaling_factor,
+            apply_routed_scale_to_output=True,
         )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -383,22 +384,6 @@ class BailingMoeV25(nn.Module):
             hidden_states=hidden_states, router_logits=router_logits
         )
 
-        # Handle tuple return from SharedFusedMoE
-        if self.shared_experts is not None:
-            shared_output, final_hidden_states = final_hidden_states
-        else:
-            shared_output = None
-
-        final_hidden_states *= self.routed_scaling_factor
-
-        if shared_output is not None:
-            final_hidden_states = final_hidden_states + shared_output
-
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
-
         return final_hidden_states.view(num_tokens, hidden_size)
 
 
diff --git a/vllm/model_executor/models/dbrx.py b/vllm/model_executor/models/dbrx.py
index ca6e6a49a98..a72f4e48716 100644
--- a/vllm/model_executor/models/dbrx.py
+++ b/vllm/model_executor/models/dbrx.py
@@ -85,7 +85,6 @@ class DbrxExperts(FusedMoE):
             hidden_size=config.d_model,
             intermediate_size=config.ffn_config.ffn_hidden_size,
             params_dtype=params_dtype,
-            reduce_results=True,
             renormalize=True,
             quant_config=quant_config,
             tp_size=get_tensor_model_parallel_world_size(),
diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py
index cd28fb0192f..1b01caded94 100644
--- a/vllm/model_executor/models/deepseek_v2.py
+++ b/vllm/model_executor/models/deepseek_v2.py
@@ -318,7 +318,6 @@ class DeepseekV2MoE(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -326,11 +325,9 @@ class DeepseekV2MoE(nn.Module):
             topk_group=getattr(config, "topk_group", 1),
             prefix=f"{prefix}.experts",
             scoring_func=getattr(config, "scoring_func", "softmax"),
-            # we do scaling outside, set factor to 1.0 to avoid double mul
             # aiter applies routed_scaling_factor internally
-            routed_scaling_factor=1.0
-            if not self.is_rocm_aiter_moe_enabled
-            else self.routed_scaling_factor,
+            routed_scaling_factor=self.routed_scaling_factor,
+            apply_routed_scale_to_output=not self.is_rocm_aiter_moe_enabled,
             e_score_correction_bias=self.gate.e_score_correction_bias,
             enable_eplb=self.enable_eplb,
             num_redundant_experts=self.n_redundant_experts,
@@ -363,43 +360,20 @@ class DeepseekV2MoE(nn.Module):
             hidden_states = sequence_parallel_chunk(hidden_states)
 
         if self.experts.is_internal_router:
-            # In this case, the gate/router runs inside the FusedMoE class
-            fused_moe_out = self.experts(
+            final_hidden_states = self.experts(
                 hidden_states=hidden_states, router_logits=hidden_states
             )
         else:
-            # router_logits: (num_tokens, n_experts)
             router_logits, _ = self.gate(hidden_states)
-            fused_moe_out = self.experts(
+            final_hidden_states = self.experts(
                 hidden_states=hidden_states, router_logits=router_logits
             )
 
-        shared_output, final_hidden_states = fused_moe_out
-        if self.shared_experts is None:
-            assert shared_output is None
-
-        # Fix FP16 overflow
-        # See DeepseekV2DecoderLayer for more details.
-        if hidden_states.dtype != torch.float16:
-            if not self.is_rocm_aiter_moe_enabled:
-                final_hidden_states *= self.routed_scaling_factor
-        elif self.shared_experts is not None:
-            assert shared_output is not None
-            shared_output *= 1.0 / self.routed_scaling_factor
-
-        if self.shared_experts is not None:
-            assert shared_output is not None
-            final_hidden_states += shared_output
-
         if self.is_sequence_parallel:
             final_hidden_states = tensor_model_parallel_all_gather(
                 final_hidden_states, 0
             )
             final_hidden_states = final_hidden_states[:num_tokens]
-        elif self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
 
         return final_hidden_states.view(num_tokens, hidden_dim)
 
diff --git a/vllm/model_executor/models/dots1.py b/vllm/model_executor/models/dots1.py
index 4e393145462..c176b736568 100644
--- a/vllm/model_executor/models/dots1.py
+++ b/vllm/model_executor/models/dots1.py
@@ -37,7 +37,6 @@ from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import (
     get_pp_group,
     get_tensor_model_parallel_world_size,
-    tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.attention import Attention
@@ -120,7 +119,6 @@ class Dots1MoE(nn.Module):
         prefix: str = "",
     ):
         super().__init__()
-        self.tp_size = get_tensor_model_parallel_world_size()
         self.routed_scaling_factor = config.routed_scaling_factor
         self.n_shared_experts = config.n_shared_experts
 
@@ -163,7 +161,6 @@ class Dots1MoE(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -171,9 +168,9 @@ class Dots1MoE(nn.Module):
             topk_group=config.topk_group,
             prefix=f"{prefix}.experts",
             scoring_func=config.scoring_func,
-            # we do scaling outside, set factor to 1.0 to avoid double mul
-            routed_scaling_factor=1.0,
             e_score_correction_bias=self.gate.e_score_correction_bias,
+            routed_scaling_factor=self.routed_scaling_factor,
+            apply_routed_scale_to_output=True,
         )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -182,16 +179,9 @@ class Dots1MoE(nn.Module):
 
         router_logits, _ = self.gate(hidden_states)
 
-        shared_out, routed_out = self.experts(
+        final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-        if self.shared_experts is not None:
-            final_hidden_states = (routed_out + shared_out) * self.routed_scaling_factor
-        else:
-            final_hidden_states = routed_out * self.routed_scaling_factor
-
-        if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states.view(num_tokens, hidden_dim)
 
 
diff --git a/vllm/model_executor/models/ernie45_moe.py b/vllm/model_executor/models/ernie45_moe.py
index f038cfb21f2..c92e230bcd2 100644
--- a/vllm/model_executor/models/ernie45_moe.py
+++ b/vllm/model_executor/models/ernie45_moe.py
@@ -194,7 +194,6 @@ class Ernie4_5_MoeMoE(nn.Module):
             top_k=config.moe_k,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=True,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -215,16 +214,6 @@ class Ernie4_5_MoeMoE(nn.Module):
             hidden_states=hidden_states, router_logits=router_logits
         )
 
-        if self.has_shared_experts:
-            final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
-        else:
-            final_hidden_states = final_hidden_states[1]
-
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
-
         return final_hidden_states.view(orig_shape)
 
 
diff --git a/vllm/model_executor/models/ernie45_vl_moe.py b/vllm/model_executor/models/ernie45_vl_moe.py
index 418fdcfa072..e4b7ac6fb00 100644
--- a/vllm/model_executor/models/ernie45_vl_moe.py
+++ b/vllm/model_executor/models/ernie45_vl_moe.py
@@ -263,7 +263,6 @@ class Ernie4_5_VLMoeMoE(nn.Module):
                 top_k=config.moe_k,
                 hidden_size=config.hidden_size,
                 intermediate_size=config.moe_intermediate_size[0],
-                reduce_results=False,
                 renormalize=True,
                 quant_config=quant_config,
                 e_score_correction_bias=self.e_score_correction_bias[0],
@@ -301,7 +300,6 @@ class Ernie4_5_VLMoeMoE(nn.Module):
                 top_k=config.moe_k,
                 hidden_size=config.hidden_size,
                 intermediate_size=config.moe_intermediate_size[1],
-                reduce_results=False,
                 renormalize=True,
                 quant_config=quant_config,
                 e_score_correction_bias=self.e_score_correction_bias[1],
@@ -342,9 +340,6 @@ class Ernie4_5_VLMoeMoE(nn.Module):
             visual_token_mask = visual_token_mask.repeat(1, self.hidden_size).bool()
             text_token_mask = ~visual_token_mask
             final_experts_hidden_states = torch.zeros_like(hidden_states)
-            final_shared_output = (
-                torch.zeros_like(hidden_states) if self.has_shared_experts else None
-            )
 
             text_hidden_states = hidden_states[text_token_mask].reshape(
                 -1, self.hidden_size
@@ -356,26 +351,20 @@ class Ernie4_5_VLMoeMoE(nn.Module):
             text_router_logits, _ = self.text_experts_gate(
                 text_hidden_states.to(dtype=torch.float32)
             )
-            text_shared_output, text_experts_output = self.text_experts(
+            text_output = self.text_experts(
                 hidden_states=text_hidden_states, router_logits=text_router_logits
             )
-            final_experts_hidden_states[text_token_mask] = text_experts_output.flatten()
-            if self.has_shared_experts:
-                final_shared_output[text_token_mask] = text_shared_output.flatten()
+            final_experts_hidden_states[text_token_mask] = text_output.flatten()
 
             vision_router_logits, _ = self.vision_experts_gate(
                 vision_hidden_states.to(dtype=torch.float32)
             )
-            vision_shared_output, vision_experts_output = self.vision_experts(
+            vision_output = self.vision_experts(
                 hidden_states=vision_hidden_states, router_logits=vision_router_logits
             )
-            final_experts_hidden_states[visual_token_mask] = (
-                vision_experts_output.flatten()
-            )
-            if self.has_shared_experts:
-                final_shared_output[visual_token_mask] = vision_shared_output.flatten()
+            final_experts_hidden_states[visual_token_mask] = vision_output.flatten()
 
-            final_hidden_states = (final_shared_output, final_experts_hidden_states)
+            final_hidden_states = final_experts_hidden_states
         else:
             # only text modal input
             text_router_logits, _ = self.text_experts_gate(
@@ -386,20 +375,6 @@ class Ernie4_5_VLMoeMoE(nn.Module):
                 hidden_states=hidden_states, router_logits=text_router_logits
             )
 
-        if self.has_shared_experts:
-            # for shared_experts model
-            final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
-        else:
-            # for not shared_experts model
-            final_hidden_states = final_hidden_states[1]
-
-        if self.tp_size > 1:
-            final_hidden_states = (
-                self.text_experts.maybe_all_reduce_tensor_model_parallel(
-                    final_hidden_states
-                )
-            )
-
         return final_hidden_states.view(orig_shape)
 
 
diff --git a/vllm/model_executor/models/exaone_moe.py b/vllm/model_executor/models/exaone_moe.py
index d7282edcf4f..a46cadf007e 100644
--- a/vllm/model_executor/models/exaone_moe.py
+++ b/vllm/model_executor/models/exaone_moe.py
@@ -31,6 +31,7 @@ from vllm.distributed import (
     get_tensor_model_parallel_world_size,
 )
 from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import ReplicatedLinear
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
@@ -116,12 +117,26 @@ class ExaoneMoe(nn.Module):
             self.physical_expert_start + self.n_local_physical_experts
         )
 
-        self.experts = FusedMoE(
+        if getattr(config, "num_shared_experts", 0) > 0:
+            intermediate_size = config.moe_intermediate_size * config.num_shared_experts
+            self.shared_experts = ExaoneMoeGatedMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+                prefix=f"{prefix}.shared_experts",
+            )
+        else:
+            self.shared_experts = None
+
+        self.experts = SharedFusedMoE(
+            shared_experts=self.shared_experts,
+            gate=self.gate,
             num_experts=self.n_routed_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -135,41 +150,16 @@ class ExaoneMoe(nn.Module):
             num_redundant_experts=self.n_redundant_experts,
         )
 
-        if getattr(config, "num_shared_experts", 0) > 0:
-            intermediate_size = config.moe_intermediate_size * config.num_shared_experts
-            self.shared_experts = ExaoneMoeGatedMLP(
-                hidden_size=config.hidden_size,
-                intermediate_size=intermediate_size,
-                hidden_act=config.hidden_act,
-                quant_config=quant_config,
-                reduce_results=self.experts.must_reduce_shared_expert_outputs(),
-                prefix=f"{prefix}.shared_experts",
-            )
-        else:
-            self.shared_experts = None
-
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         # NOTE: hidden_states can have either 1D or 2D shape.
         orig_shape = hidden_states.shape
         hidden_dim = hidden_states.shape[-1]
         hidden_states = hidden_states.view(-1, hidden_dim)
 
-        # router_logits: (num_tokens, n_experts)
-        router_logits, _ = self.gate(hidden_states)
-
         final_hidden_states = self.experts(
-            hidden_states=hidden_states, router_logits=router_logits
+            hidden_states=hidden_states, router_logits=hidden_states
         )
 
-        if self.shared_experts is not None:
-            shared_output = self.shared_experts(hidden_states)
-            final_hidden_states = final_hidden_states + shared_output
-
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
-                final_hidden_states
-            )
-
         return final_hidden_states.view(orig_shape)
 
 
diff --git a/vllm/model_executor/models/flex_olmo.py b/vllm/model_executor/models/flex_olmo.py
index 67be99a879f..1b2047eb231 100644
--- a/vllm/model_executor/models/flex_olmo.py
+++ b/vllm/model_executor/models/flex_olmo.py
@@ -76,7 +76,6 @@ class FlexOlmoMoE(nn.Module):
             top_k=hf_config.num_experts_per_tok,
             hidden_size=hf_config.hidden_size,
             intermediate_size=hf_config.intermediate_size,
-            reduce_results=True,
             renormalize=False,
             quant_config=None,
             tp_size=tp_size,
diff --git a/vllm/model_executor/models/gemma4.py b/vllm/model_executor/models/gemma4.py
index d166a9df38a..42762e36f81 100644
--- a/vllm/model_executor/models/gemma4.py
+++ b/vllm/model_executor/models/gemma4.py
@@ -349,7 +349,6 @@ class Gemma4MoE(nn.Module):
                 "moe_intermediate_size",
                 getattr(config, "expert_intermediate_size", None),
             ),
-            reduce_results=True,
             renormalize=True,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
diff --git a/vllm/model_executor/models/glm4_moe.py b/vllm/model_executor/models/glm4_moe.py
index d0e6cb6ada8..671e868da0a 100644
--- a/vllm/model_executor/models/glm4_moe.py
+++ b/vllm/model_executor/models/glm4_moe.py
@@ -184,7 +184,6 @@ class Glm4MoE(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -192,8 +191,8 @@ class Glm4MoE(nn.Module):
             topk_group=config.topk_group,
             prefix=f"{prefix}.experts",
             scoring_func="sigmoid",
-            # we do scaling outside, set factor to 1.0 to avoid double mul
-            routed_scaling_factor=1.0,
+            routed_scaling_factor=self.routed_scaling_factor,
+            apply_routed_scale_to_output=True,
             e_score_correction_bias=self.gate.e_score_correction_bias,
             enable_eplb=self.enable_eplb,
             num_redundant_experts=self.n_redundant_experts,
@@ -207,23 +206,9 @@ class Glm4MoE(nn.Module):
         # router_logits: (num_tokens, n_experts)
         router_logits = self.gate(hidden_states.to(dtype=torch.float32))
 
-        fused_moe_out = self.experts(
+        final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-
-        if self.shared_experts is not None:
-            shared_output, final_hidden_states = fused_moe_out
-            assert shared_output is not None
-            final_hidden_states = (
-                final_hidden_states * self.routed_scaling_factor + shared_output
-            )
-        else:
-            final_hidden_states = fused_moe_out * self.routed_scaling_factor
-
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
         return final_hidden_states.view(num_tokens, hidden_dim)
 
 
diff --git a/vllm/model_executor/models/gpt_oss.py b/vllm/model_executor/models/gpt_oss.py
index 4e4eb581842..b6edc344302 100644
--- a/vllm/model_executor/models/gpt_oss.py
+++ b/vllm/model_executor/models/gpt_oss.py
@@ -189,7 +189,6 @@ class MLPBlock(torch.nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.intermediate_size,
-            reduce_results=True,
             renormalize=True,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
diff --git a/vllm/model_executor/models/granitemoe.py b/vllm/model_executor/models/granitemoe.py
index 171b2e0ec5a..f57a8c942bb 100644
--- a/vllm/model_executor/models/granitemoe.py
+++ b/vllm/model_executor/models/granitemoe.py
@@ -104,7 +104,6 @@ class GraniteMoeMoE(nn.Module):
             hidden_size=hidden_size,
             intermediate_size=intermediate_size,
             params_dtype=params_dtype,
-            reduce_results=True,
             renormalize=True,
             quant_config=quant_config,
             tp_size=tp_size,
diff --git a/vllm/model_executor/models/grok1.py b/vllm/model_executor/models/grok1.py
index 0bd6a8f3d60..c9aa3d2068f 100644
--- a/vllm/model_executor/models/grok1.py
+++ b/vllm/model_executor/models/grok1.py
@@ -209,7 +209,6 @@ class Grok1MoE(nn.Module):
             hidden_size=hidden_size,
             intermediate_size=intermediate_size,
             params_dtype=params_dtype,
-            reduce_results=True,
             renormalize=renormalize,
             quant_config=quant_config,
             tp_size=tp_size,
diff --git a/vllm/model_executor/models/hunyuan_v1.py b/vllm/model_executor/models/hunyuan_v1.py
index a0130402c66..35d30006a66 100644
--- a/vllm/model_executor/models/hunyuan_v1.py
+++ b/vllm/model_executor/models/hunyuan_v1.py
@@ -39,7 +39,6 @@ from vllm.distributed import (
     get_ep_group,
     get_pp_group,
     get_tensor_model_parallel_world_size,
-    tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.attention import Attention
@@ -445,7 +444,6 @@ class HunYuanSparseMoeBlock(nn.Module):
             top_k=top_k,
             hidden_size=config.hidden_size,
             intermediate_size=intermediate_size,
-            reduce_results=False,
             renormalize=top_k > 1,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -464,11 +462,6 @@ class HunYuanSparseMoeBlock(nn.Module):
         final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-        if self.shared_mlp is not None:
-            final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
-
-        if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states.view(orig_shape)
 
diff --git a/vllm/model_executor/models/interns1_pro.py b/vllm/model_executor/models/interns1_pro.py
index 28331b8ef3e..9612ea57b2c 100644
--- a/vllm/model_executor/models/interns1_pro.py
+++ b/vllm/model_executor/models/interns1_pro.py
@@ -176,7 +176,6 @@ class InternS1ProMoeSparseMoeBlock(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=True,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
diff --git a/vllm/model_executor/models/jamba.py b/vllm/model_executor/models/jamba.py
index 980bcffb5f9..b4b3b6873db 100644
--- a/vllm/model_executor/models/jamba.py
+++ b/vllm/model_executor/models/jamba.py
@@ -90,7 +90,6 @@ class JambaMoE(nn.Module):
             self.intermediate_size,
             tp_size=tp_size,
             params_dtype=params_dtype,
-            reduce_results=True,
             renormalize=False,
             use_grouped_topk=False,
             quant_config=quant_config,
diff --git a/vllm/model_executor/models/kimi_linear.py b/vllm/model_executor/models/kimi_linear.py
index 4cd7b63c147..e586a3ac346 100644
--- a/vllm/model_executor/models/kimi_linear.py
+++ b/vllm/model_executor/models/kimi_linear.py
@@ -11,11 +11,10 @@ from vllm.config import CacheConfig, ModelConfig, ParallelConfig, VllmConfig
 from vllm.distributed import (
     get_pp_group,
     get_tensor_model_parallel_world_size,
-    tensor_model_parallel_all_reduce,
 )
 from vllm.logger import init_logger
 from vllm.model_executor.layers.activation import SiluAndMul
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import SharedFusedMoE
 from vllm.model_executor.layers.kda import KimiDeltaAttention
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
@@ -132,22 +131,6 @@ class KimiMoE(nn.Module):
 
         self.gate.e_score_correction_bias = nn.Parameter(torch.empty(num_experts))
 
-        self.experts = FusedMoE(
-            num_experts=num_experts,
-            top_k=config.num_experts_per_token,
-            hidden_size=hidden_size,
-            intermediate_size=moe_intermediate_size,
-            reduce_results=False,
-            renormalize=moe_renormalize,
-            quant_config=quant_config,
-            use_grouped_topk=config.use_grouped_topk,
-            num_expert_group=config.num_expert_group,
-            topk_group=config.topk_group,
-            prefix=f"{prefix}.experts",
-            scoring_func=config.moe_router_activation_func,
-            e_score_correction_bias=self.gate.e_score_correction_bias,
-        )
-
         if self.num_shared_experts is not None:
             intermediate_size = moe_intermediate_size * self.num_shared_experts
             self.shared_experts = KimiMLP(
@@ -158,22 +141,33 @@ class KimiMoE(nn.Module):
                 reduce_results=False,
                 prefix=f"{prefix}.shared_experts",
             )
+        else:
+            self.shared_experts = None
+
+        self.experts = SharedFusedMoE(
+            shared_experts=self.shared_experts,
+            num_experts=num_experts,
+            top_k=config.num_experts_per_token,
+            hidden_size=hidden_size,
+            intermediate_size=moe_intermediate_size,
+            renormalize=moe_renormalize,
+            quant_config=quant_config,
+            use_grouped_topk=config.use_grouped_topk,
+            num_expert_group=config.num_expert_group,
+            topk_group=config.topk_group,
+            prefix=f"{prefix}.experts",
+            scoring_func=config.moe_router_activation_func,
+            e_score_correction_bias=self.gate.e_score_correction_bias,
+            routed_scaling_factor=self.routed_scaling_factor,
+        )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         num_tokens, hidden_size = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_size)
-        if self.num_shared_experts is not None:
-            shared_output = self.shared_experts(hidden_states)
         router_logits, _ = self.gate(hidden_states)
-        final_hidden_states = (
-            self.experts(hidden_states=hidden_states, router_logits=router_logits)
-            * self.routed_scaling_factor
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states, router_logits=router_logits
         )
-        if shared_output is not None:
-            final_hidden_states = final_hidden_states + shared_output
-
-        if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states.view(num_tokens, hidden_size)
 
 
@@ -482,7 +476,7 @@ class KimiLinearModel(nn.Module):
         if self.config.is_moe:
             # Params for weights, fp8 weight scales, fp8 activation scales
             # (param_name, weight_name, expert_id, shard_id)
-            expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
                 self,
                 ckpt_gate_proj_name="w1",
                 ckpt_down_proj_name="w2",
diff --git a/vllm/model_executor/models/lfm2_moe.py b/vllm/model_executor/models/lfm2_moe.py
index d955b7127ad..4b49430c1fa 100644
--- a/vllm/model_executor/models/lfm2_moe.py
+++ b/vllm/model_executor/models/lfm2_moe.py
@@ -150,7 +150,6 @@ class Lfm2MoeSparseMoeBlock(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,  # needed for softmax score func
@@ -161,6 +160,7 @@ class Lfm2MoeSparseMoeBlock(nn.Module):
             num_redundant_experts=self.n_redundant_experts,
             scoring_func="sigmoid",
             e_score_correction_bias=self.gate.e_score_correction_bias,
+            routed_scaling_factor=self.routed_scaling_factor,
         )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -170,16 +170,10 @@ class Lfm2MoeSparseMoeBlock(nn.Module):
 
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
-        final_hidden_states = (
-            self.experts(hidden_states=hidden_states, router_logits=router_logits)
-            * self.routed_scaling_factor
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states, router_logits=router_logits
         )
 
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
-                final_hidden_states
-            )
-
         return final_hidden_states.view(orig_shape)
 
 
diff --git a/vllm/model_executor/models/llama4.py b/vllm/model_executor/models/llama4.py
index b84b4e2ae51..a1c0ac89605 100644
--- a/vllm/model_executor/models/llama4.py
+++ b/vllm/model_executor/models/llama4.py
@@ -135,7 +135,6 @@ class Llama4MoE(nn.Module):
             custom_routing_function=Llama4MoE.custom_routing_function,
             intermediate_size=intermediate_size_moe,
             apply_router_weight_on_input=True,
-            reduce_results=False,
             renormalize=False,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -151,19 +150,14 @@ class Llama4MoE(nn.Module):
 
         router_logits, _ = self.router(hidden_states)
 
-        shared_out, routed_out = self.experts(
+        experts_out = self.experts(
             hidden_states=hidden_states,
             router_logits=router_logits,
         )
-        experts_out = routed_out + shared_out
 
         if self.is_sequence_parallel:
             experts_out = tensor_model_parallel_all_gather(experts_out, 0)
             experts_out = experts_out[:num_tokens]
-        elif self.tp_size > 1:
-            experts_out = self.experts.maybe_all_reduce_tensor_model_parallel(
-                experts_out
-            )
 
         return experts_out
 
diff --git a/vllm/model_executor/models/longcat_flash.py b/vllm/model_executor/models/longcat_flash.py
index 375b0b69b1f..945fcb61509 100644
--- a/vllm/model_executor/models/longcat_flash.py
+++ b/vllm/model_executor/models/longcat_flash.py
@@ -300,7 +300,6 @@ class LongcatMoe(nn.Module):
             top_k=top_k,
             hidden_size=hidden_size,
             intermediate_size=intermediate_size,
-            reduce_results=True,
             params_dtype=params_dtype,
             renormalize=False,
             quant_config=quant_config,
diff --git a/vllm/model_executor/models/mimo_v2_flash.py b/vllm/model_executor/models/mimo_v2_flash.py
index 43475ed690c..0b466f16601 100644
--- a/vllm/model_executor/models/mimo_v2_flash.py
+++ b/vllm/model_executor/models/mimo_v2_flash.py
@@ -162,7 +162,6 @@ class MiMoV2MoE(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=True,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
diff --git a/vllm/model_executor/models/minimax_m2.py b/vllm/model_executor/models/minimax_m2.py
index 2a9e5f4e08a..84d8dda533f 100644
--- a/vllm/model_executor/models/minimax_m2.py
+++ b/vllm/model_executor/models/minimax_m2.py
@@ -36,7 +36,6 @@ from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import (
     get_pp_group,
     get_tensor_model_parallel_world_size,
-    tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.attention import Attention
 from vllm.model_executor.layers.fused_moe import FusedMoE
@@ -104,7 +103,6 @@ class MiniMaxM2MoE(nn.Module):
             e_score_correction_bias=self.e_score_correction_bias,
             hidden_size=config.hidden_size,
             intermediate_size=config.intermediate_size,
-            reduce_results=False,
             renormalize=True,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -134,9 +132,6 @@ class MiniMaxM2MoE(nn.Module):
         final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-        final_hidden_states = final_hidden_states
-        if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states.view(num_tokens, hidden_dim)
 
diff --git a/vllm/model_executor/models/minimax_text_01.py b/vllm/model_executor/models/minimax_text_01.py
index 21d74d8b058..67d7cb2d8bc 100644
--- a/vllm/model_executor/models/minimax_text_01.py
+++ b/vllm/model_executor/models/minimax_text_01.py
@@ -162,7 +162,6 @@ class MiniMaxText01MoE(nn.Module):
             hidden_size=self.hidden_size,
             intermediate_size=self.intermediate_size * self.tp_size,
             params_dtype=self.params_dtype,
-            reduce_results=True,
             renormalize=True,
             quant_config=self.quant_config,
             tp_size=self.tp_size,
diff --git a/vllm/model_executor/models/mixtral.py b/vllm/model_executor/models/mixtral.py
index 376fd7a1709..c182444f667 100644
--- a/vllm/model_executor/models/mixtral.py
+++ b/vllm/model_executor/models/mixtral.py
@@ -132,7 +132,6 @@ class MixtralMoE(nn.Module):
             hidden_size=hidden_size,
             intermediate_size=intermediate_size,
             params_dtype=params_dtype,
-            reduce_results=True,
             renormalize=True,
             quant_config=quant_config,
             tp_size=tp_size,
diff --git a/vllm/model_executor/models/nemotron_h.py b/vllm/model_executor/models/nemotron_h.py
index 8abbc808cce..fa068639648 100644
--- a/vllm/model_executor/models/nemotron_h.py
+++ b/vllm/model_executor/models/nemotron_h.py
@@ -216,7 +216,6 @@ class NemotronHMoE(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=self.moe_hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -231,6 +230,9 @@ class NemotronHMoE(nn.Module):
             num_redundant_experts=self.n_redundant_experts,
             is_sequence_parallel=self.is_sequence_parallel,
             routed_input_transform=self.fc1_latent_proj,
+            routed_output_transform=self.fc2_latent_proj,
+            routed_scaling_factor=self.routed_scaling_factor,
+            apply_routed_scale_to_output=True,
             router_logits_dtype=self.gate.out_dtype,
         )
 
@@ -244,38 +246,15 @@ class NemotronHMoE(nn.Module):
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
 
-        # SharedFusedMoE handles:
-        #   - shared experts (with original hidden_states)
-        #   - routed_input_transform (fc1_latent_proj) for latent MoE
-        #   - multistream parallelism between shared and routed experts
-        shared_output, final_hidden_states = self.experts(
+        final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
 
-        # Fix FP16 overflow
-        # See DeepseekV2DecoderLayer for more details.
-        if hidden_states.dtype != torch.float16:
-            final_hidden_states *= self.routed_scaling_factor
-        elif self.shared_experts is not None:
-            shared_output *= 1.0 / self.routed_scaling_factor
-
-        # TODO: See SharedFusedMoE.apply_routed_input_transform
-        # for bandwidth optimization
-        if self.use_latent_moe:
-            final_hidden_states, _ = self.fc2_latent_proj(final_hidden_states)
-
-        if self.shared_experts is not None:
-            final_hidden_states += shared_output
-
         if self.is_sequence_parallel:
             final_hidden_states = tensor_model_parallel_all_gather(
                 final_hidden_states, 0
             )
             final_hidden_states = final_hidden_states[:num_tokens]
-        elif self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
 
         return final_hidden_states.view(num_tokens, hidden_dim)
 
diff --git a/vllm/model_executor/models/olmoe.py b/vllm/model_executor/models/olmoe.py
index f0afe0e997c..fcde2e41afb 100644
--- a/vllm/model_executor/models/olmoe.py
+++ b/vllm/model_executor/models/olmoe.py
@@ -98,7 +98,6 @@ class OlmoeMoE(nn.Module):
             top_k=top_k,
             hidden_size=hidden_size,
             intermediate_size=intermediate_size,
-            reduce_results=True,
             renormalize=False,
             quant_config=quant_config,
             tp_size=tp_size,
diff --git a/vllm/model_executor/models/openpangu.py b/vllm/model_executor/models/openpangu.py
index 994ae82529a..7de84da5193 100644
--- a/vllm/model_executor/models/openpangu.py
+++ b/vllm/model_executor/models/openpangu.py
@@ -206,7 +206,6 @@ class OpenPanguMoE(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -214,8 +213,8 @@ class OpenPanguMoE(nn.Module):
             topk_group=1,
             prefix=f"{prefix}.experts",
             scoring_func="sigmoid",
-            # we do scaling outside, set factor to 1.0 to avoid double mul
-            routed_scaling_factor=1.0,
+            routed_scaling_factor=self.routed_scaling_factor,
+            apply_routed_scale_to_output=True,
             e_score_correction_bias=self.gate.e_score_correction_bias,
             enable_eplb=self.enable_eplb,
             num_redundant_experts=self.n_redundant_experts,
@@ -234,33 +233,15 @@ class OpenPanguMoE(nn.Module):
 
         router_logits, _ = self.gate(hidden_states)
 
-        fused_moe_out = self.experts(
+        final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
 
-        shared_output, final_hidden_states = fused_moe_out
-        if self.shared_experts is None:
-            assert shared_output is None
-
-        if hidden_states.dtype != torch.float16:
-            final_hidden_states *= self.routed_scaling_factor
-        elif self.shared_experts is not None:
-            assert shared_output is not None
-            shared_output *= 1.0 / self.routed_scaling_factor
-
-        if self.shared_experts is not None:
-            assert shared_output is not None
-            final_hidden_states += shared_output
-
         if self.is_sequence_parallel:
             final_hidden_states = tensor_model_parallel_all_gather(
                 final_hidden_states, 0
             )
             final_hidden_states = final_hidden_states[:num_tokens]
-        elif self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
 
         return final_hidden_states.view(num_tokens, hidden_dim)
 
diff --git a/vllm/model_executor/models/param2moe.py b/vllm/model_executor/models/param2moe.py
index 925f94e0600..fddd1a8f173 100644
--- a/vllm/model_executor/models/param2moe.py
+++ b/vllm/model_executor/models/param2moe.py
@@ -359,7 +359,6 @@ class Param2MoEMoEBlock(nn.Module):
             top_k=self.top_k,
             hidden_size=self.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=self.norm_expert_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -388,24 +387,11 @@ class Param2MoEMoEBlock(nn.Module):
             self.gate.weight.float(),
         ).to(hidden_states.dtype)
 
-        final_hidden = self.experts(
+        expert_output = self.experts(
             hidden_states=hidden_states,
             router_logits=router_logits,
         )
 
-        if self.shared_experts is not None:
-            shared_output, expert_output = final_hidden
-        else:
-            shared_output, expert_output = None, final_hidden
-
-        if shared_output is not None:
-            expert_output = expert_output + shared_output
-
-        if self.tp_size > 1:
-            expert_output = self.experts.maybe_all_reduce_tensor_model_parallel(
-                expert_output
-            )
-
         return expert_output.view(num_tokens, hidden_dim)
 
 
diff --git a/vllm/model_executor/models/phimoe.py b/vllm/model_executor/models/phimoe.py
index 0b55b7ec839..7d6083f202e 100644
--- a/vllm/model_executor/models/phimoe.py
+++ b/vllm/model_executor/models/phimoe.py
@@ -281,7 +281,6 @@ class PhiMoE(nn.Module):
             hidden_size=hidden_size,
             intermediate_size=intermediate_size,
             params_dtype=params_dtype,
-            reduce_results=True,
             renormalize=False,
             quant_config=quant_config,
             tp_size=tp_size,
diff --git a/vllm/model_executor/models/qwen2_moe.py b/vllm/model_executor/models/qwen2_moe.py
index 750835cccd9..b5d13e926d7 100644
--- a/vllm/model_executor/models/qwen2_moe.py
+++ b/vllm/model_executor/models/qwen2_moe.py
@@ -170,7 +170,6 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -187,12 +186,6 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
         final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-        if self.shared_expert is not None:
-            final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
-                final_hidden_states
-            )
 
         return final_hidden_states.view(orig_shape)
 
diff --git a/vllm/model_executor/models/qwen3_moe.py b/vllm/model_executor/models/qwen3_moe.py
index f2ce070be8b..f0f69d43537 100644
--- a/vllm/model_executor/models/qwen3_moe.py
+++ b/vllm/model_executor/models/qwen3_moe.py
@@ -212,7 +212,6 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -234,22 +233,15 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
 
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
-        shared_out, fused_out = self.experts(
+        final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-        final_hidden_states = (
-            shared_out + fused_out if shared_out is not None else fused_out
-        )
 
         if self.is_sequence_parallel:
             final_hidden_states = tensor_model_parallel_all_gather(
                 final_hidden_states, 0
             )
             final_hidden_states = final_hidden_states[:num_tokens]
-        elif self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
-                final_hidden_states
-            )
 
         # return to 1d if input is 1d
         return final_hidden_states.squeeze(0) if is_input_1d else final_hidden_states
diff --git a/vllm/model_executor/models/qwen3_next.py b/vllm/model_executor/models/qwen3_next.py
index 3bd026d9dc9..50d44dbbf63 100644
--- a/vllm/model_executor/models/qwen3_next.py
+++ b/vllm/model_executor/models/qwen3_next.py
@@ -153,7 +153,6 @@ class Qwen3NextSparseMoeBlock(nn.Module):
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=getattr(config, "norm_topk_prob", True),
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -183,18 +182,11 @@ class Qwen3NextSparseMoeBlock(nn.Module):
                 hidden_states=hidden_states, router_logits=router_logits
             )
 
-        if self.shared_expert is not None:
-            final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
-
         if self.is_sequence_parallel:
             final_hidden_states = tensor_model_parallel_all_gather(
                 final_hidden_states, 0
             )
             final_hidden_states = final_hidden_states[:num_tokens]
-        elif self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
-                final_hidden_states
-            )
 
         return final_hidden_states.view(orig_shape)
 
diff --git a/vllm/model_executor/models/sarvam.py b/vllm/model_executor/models/sarvam.py
index fa5ec44d7e7..3656fc921b2 100644
--- a/vllm/model_executor/models/sarvam.py
+++ b/vllm/model_executor/models/sarvam.py
@@ -341,7 +341,6 @@ class SarvamMLAMoE(nn.Module):
             top_k=self.top_k,
             hidden_size=self.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=self.norm_expert_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -370,20 +369,7 @@ class SarvamMLAMoE(nn.Module):
             router_logits=router_logits,
         )
 
-        if self.shared_experts is not None:
-            shared_output, expert_output = final_hidden
-        else:
-            shared_output, expert_output = None, final_hidden
-
-        if shared_output is not None:
-            expert_output = expert_output + shared_output
-
-        if self.tp_size > 1:
-            expert_output = self.experts.maybe_all_reduce_tensor_model_parallel(
-                expert_output
-            )
-
-        return expert_output.view(num_tokens, hidden_dim)
+        return final_hidden.view(num_tokens, hidden_dim)
 
 
 class SarvamMLABlock(nn.Module):
diff --git a/vllm/model_executor/models/step3_text.py b/vllm/model_executor/models/step3_text.py
index 18b689166a5..636a121c590 100644
--- a/vllm/model_executor/models/step3_text.py
+++ b/vllm/model_executor/models/step3_text.py
@@ -14,7 +14,6 @@ from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import (
     get_pp_group,
     get_tensor_model_parallel_world_size,
-    tensor_model_parallel_all_reduce,
 )
 from vllm.logger import init_logger
 from vllm.model_executor.layers.activation import SiluAndMul
@@ -71,7 +70,6 @@ class FusedMoEBlock(nn.Module):
             top_k=config.moe_top_k,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_expert_weight,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
@@ -94,8 +92,6 @@ class FusedMoEBlock(nn.Module):
         final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-        if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states.view(orig_shape)
 
diff --git a/vllm/model_executor/models/step3p5.py b/vllm/model_executor/models/step3p5.py
index bb4bf14a963..018f7895602 100644
--- a/vllm/model_executor/models/step3p5.py
+++ b/vllm/model_executor/models/step3p5.py
@@ -379,7 +379,6 @@ class FusedMoEBlock(nn.Module):
             top_k=config.moe_top_k,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_expert_weight,
             quant_config=quant_config,
             activation=activation,
@@ -397,30 +396,16 @@ class FusedMoEBlock(nn.Module):
         hidden_states = hidden_states.view(-1, hidden_dim)
 
         if self.experts.is_internal_router:
-            # In this case, the gate/router runs inside the FusedMoE class
-            fused_moe_out = self.experts(
+            final_hidden_states = self.experts(
                 hidden_states=hidden_states, router_logits=hidden_states
             )
         else:
-            # router_logits: (num_tokens, n_experts)
+            # TODO(bnell): this gate could be moved into the FusedMoE?
             router_logits, _ = self.gate(hidden_states)
-            fused_moe_out = self.experts(
+            final_hidden_states = self.experts(
                 hidden_states=hidden_states, router_logits=router_logits
             )
 
-        shared_output, final_hidden_states = fused_moe_out
-        if self.share_expert is None:
-            assert shared_output is None
-
-        if self.share_expert is not None:
-            assert shared_output is not None
-            final_hidden_states += shared_output
-
-        if self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-                final_hidden_states
-            )
-
         return final_hidden_states.view(num_tokens, hidden_dim)
 
 
diff --git a/vllm/model_executor/models/transformers/moe.py b/vllm/model_executor/models/transformers/moe.py
index 81d21abbd06..cf13958ef76 100644
--- a/vllm/model_executor/models/transformers/moe.py
+++ b/vllm/model_executor/models/transformers/moe.py
@@ -204,8 +204,6 @@ class MoEMixin(MixtureOfExperts):
         )
         assert intermediate_size is not None
 
-        # If there are shared experts, the results are
-        # reduced after mlp.forward() not inside FusedMoE
         num_shared_experts = getattr_iter(
             text_config,
             [
@@ -214,17 +212,6 @@ class MoEMixin(MixtureOfExperts):
             ],
             0,
         )
-        reduce_results = num_shared_experts == 0
-
-        def add_all_reduce(mlp: nn.Module):
-            """Adds an all-reduce to the output of `mlp.forward()`."""
-
-            class MLPWithAllReduce(mlp.__class__):
-                def forward(self, *args, **kwargs):
-                    output = super().forward(*args, **kwargs)
-                    return self.experts.maybe_all_reduce_tensor_model_parallel(output)
-
-            mlp.__class__ = MLPWithAllReduce
 
         # Unused kwargs since we use custom_routing_function:
         # - `scoring_func` and `e_score_correction_bias` only used for grouped
@@ -289,14 +276,11 @@ class MoEMixin(MixtureOfExperts):
                         if "bias" in experts_param_name:
                             has_bias = True
                             break
-                    # Double check there are no shared experts
-                    nonlocal reduce_results
-                    if reduce_results:
+                    # If the config does not specify num_shared_experts, but
+                    # the model has shared experts, we assume there is one.
+                    if self.num_shared_experts == 0:
                         for mlp_param_name, _ in mlp.named_parameters():
                             if "shared_expert" in mlp_param_name:
-                                reduce_results = False
-                                # If the config does not specify num_shared_experts, but
-                                # the model has shared experts, we assume there is one.
                                 self.num_shared_experts = 1
                                 break
                     # Replace experts module with FusedMoE
@@ -305,7 +289,6 @@ class MoEMixin(MixtureOfExperts):
                         top_k=top_k,
                         hidden_size=hidden_size,
                         intermediate_size=intermediate_size,
-                        reduce_results=reduce_results,
                         renormalize=renormalize,
                         # Hard coded because topk happens in Transformers
                         use_grouped_topk=False,
@@ -326,13 +309,6 @@ class MoEMixin(MixtureOfExperts):
                     self.moe_layers.append(fused_experts)
                     self.expert_weights.append(fused_experts.get_expert_weights())
                     self.num_moe_layers += 1
-                    # If results are not all-reduced in FusedMoE, ensure they
-                    # are all-reduced at the end of mlp.forward() if tensor
-                    # parallel or expert parallel is enabled
-                    if not reduce_results and (
-                        fused_experts.tp_size > 1 or fused_experts.ep_size > 1
-                    ):
-                        add_all_reduce(mlp)
                 else:
                     _recursive_replace(child_module, prefix=qual_name)