diff --git a/python/sglang/srt/models/bailing_moe.py b/python/sglang/srt/models/bailing_moe.py
index faaadc90a..4d9a400a4 100644
--- a/python/sglang/srt/models/bailing_moe.py
+++ b/python/sglang/srt/models/bailing_moe.py
@@ -58,6 +58,7 @@ from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_deepep_mode,
     get_moe_a2a_backend,
+    should_use_dp_reduce_scatterv,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
@@ -386,6 +387,7 @@ class BailingMoESparseMoeBlock(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states.view(num_tokens, hidden_size)
diff --git a/python/sglang/srt/models/bailing_moe_linear.py b/python/sglang/srt/models/bailing_moe_linear.py
index 0833a68c2..d8750da8f 100644
--- a/python/sglang/srt/models/bailing_moe_linear.py
+++ b/python/sglang/srt/models/bailing_moe_linear.py
@@ -34,6 +34,7 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe import should_use_dp_reduce_scatterv
 from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
@@ -347,7 +348,12 @@ class BailingMoE(nn.Module):
         if self.num_shared_experts > 0:
             final_hidden_states = final_hidden_states + shared_output
 
-        if self.tp_size > 1 and not use_reduce_scatter and not should_allreduce_fusion:
+        if (
+            self.tp_size > 1
+            and not use_reduce_scatter
+            and not should_allreduce_fusion
+            and not should_use_dp_reduce_scatterv()
+        ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
 
diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index 6b23c2c88..9f2c82405 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -85,6 +85,7 @@ from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_moe_a2a_backend,
     get_moe_runner_backend,
+    should_use_dp_reduce_scatterv,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
@@ -655,6 +656,7 @@ class DeepseekV2MoE(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
@@ -744,6 +746,7 @@ class DeepseekV2MoE(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
diff --git a/python/sglang/srt/models/exaone_moe.py b/python/sglang/srt/models/exaone_moe.py
index ab31c0a59..18012448a 100755
--- a/python/sglang/srt/models/exaone_moe.py
+++ b/python/sglang/srt/models/exaone_moe.py
@@ -47,7 +47,7 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
-from sglang.srt.layers.moe import get_moe_a2a_backend
+from sglang.srt.layers.moe import get_moe_a2a_backend, should_use_dp_reduce_scatterv
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
@@ -300,7 +300,11 @@ class ExaoneMoESparseMoEBlock(nn.Module):
 
         if shared_output is not None:
             final_hidden_states = final_hidden_states + shared_output
-        if self.tp_size > 1 and not use_reduce_scatter:
+        if (
+            self.tp_size > 1
+            and not use_reduce_scatter
+            and not should_use_dp_reduce_scatterv()
+        ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states.view(num_tokens, hidden_dim)
diff --git a/python/sglang/srt/models/glm4_moe.py b/python/sglang/srt/models/glm4_moe.py
index 7e8dd9605..092b94b7e 100644
--- a/python/sglang/srt/models/glm4_moe.py
+++ b/python/sglang/srt/models/glm4_moe.py
@@ -61,6 +61,7 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_moe_a2a_backend,
+    should_use_dp_reduce_scatterv,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
@@ -598,6 +599,7 @@ class Glm4MoeSparseMoeBlock(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
@@ -632,6 +634,7 @@ class Glm4MoeSparseMoeBlock(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
diff --git a/python/sglang/srt/models/hunyuan_v3.py b/python/sglang/srt/models/hunyuan_v3.py
index b0f87dc32..f11827b98 100644
--- a/python/sglang/srt/models/hunyuan_v3.py
+++ b/python/sglang/srt/models/hunyuan_v3.py
@@ -34,6 +34,7 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe import should_use_dp_reduce_scatterv
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -191,10 +192,11 @@ class HYV3MoEFused(nn.Module):
                 hidden_states=hidden_states, topk_output=topk_output
             )
 
-        if self.ep_size > 1:
+        skip_post_reduce = should_use_dp_reduce_scatterv()
+        if self.ep_size > 1 and not skip_post_reduce:
             final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states)
 
-        if self.tp_size > 1:
+        if self.tp_size > 1 and not skip_post_reduce:
             final_hidden_states = moe_tensor_model_parallel_all_reduce(
                 final_hidden_states
             )
@@ -222,10 +224,11 @@ class HYV3MoEFused(nn.Module):
         current_stream.wait_stream(self.alt_stream)
         final_hidden_states = final_hidden_states + shared_output
 
-        if self.ep_size > 1:
+        skip_post_reduce = should_use_dp_reduce_scatterv()
+        if self.ep_size > 1 and not skip_post_reduce:
             final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states)
 
-        if self.tp_size > 1:
+        if self.tp_size > 1 and not skip_post_reduce:
             final_hidden_states = moe_tensor_model_parallel_all_reduce(
                 final_hidden_states
             )
diff --git a/python/sglang/srt/models/llada2.py b/python/sglang/srt/models/llada2.py
index 85c9e6a32..194abf32b 100644
--- a/python/sglang/srt/models/llada2.py
+++ b/python/sglang/srt/models/llada2.py
@@ -55,7 +55,11 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe import get_deepep_mode, get_moe_a2a_backend
+from sglang.srt.layers.moe import (
+    get_deepep_mode,
+    get_moe_a2a_backend,
+    should_use_dp_reduce_scatterv,
+)
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.token_dispatcher import DeepEPDispatcher
@@ -379,7 +383,11 @@ class LLaDA2MoeSparseMoeBlock(nn.Module):
         if self.num_shared_experts > 0:
             final_hidden_states = final_hidden_states + shared_output
 
-        if self.tp_size > 1 and not use_reduce_scatter:
+        if (
+            self.tp_size > 1
+            and not use_reduce_scatter
+            and not should_use_dp_reduce_scatterv()
+        ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states.view(num_tokens, hidden_size)
 
diff --git a/python/sglang/srt/models/llama4.py b/python/sglang/srt/models/llama4.py
index f0a237aa6..c2e4b12c9 100644
--- a/python/sglang/srt/models/llama4.py
+++ b/python/sglang/srt/models/llama4.py
@@ -39,6 +39,7 @@ from sglang.srt.layers.linear import (
     ReplicatedLinear,
     RowParallelLinear,
 )
+from sglang.srt.layers.moe import should_use_dp_reduce_scatterv
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -145,7 +146,11 @@ class Llama4MoE(nn.Module):
 
         out_aD = routed_out + shared_out
 
-        if self.tp_size > 1 and not use_reduce_scatter:
+        if (
+            self.tp_size > 1
+            and not use_reduce_scatter
+            and not should_use_dp_reduce_scatterv()
+        ):
             out_aD = tensor_model_parallel_all_reduce(out_aD)
 
         return out_aD
diff --git a/python/sglang/srt/models/mimo_v2_flash.py b/python/sglang/srt/models/mimo_v2_flash.py
index 9902b63d4..b6dd6c4b0 100644
--- a/python/sglang/srt/models/mimo_v2_flash.py
+++ b/python/sglang/srt/models/mimo_v2_flash.py
@@ -51,6 +51,7 @@ from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_moe_a2a_backend,
     get_moe_runner_backend,
+    should_use_dp_reduce_scatterv,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class
@@ -302,6 +303,7 @@ class MiMoV2MoE(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
diff --git a/python/sglang/srt/models/minimax_m2.py b/python/sglang/srt/models/minimax_m2.py
index b99c2471b..d6a38d082 100644
--- a/python/sglang/srt/models/minimax_m2.py
+++ b/python/sglang/srt/models/minimax_m2.py
@@ -61,6 +61,7 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_moe_a2a_backend,
+    should_use_dp_reduce_scatterv,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
@@ -556,6 +557,7 @@ class MiniMaxM2MoE(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
diff --git a/python/sglang/srt/models/sarvam_moe.py b/python/sglang/srt/models/sarvam_moe.py
index 9c4179362..faafd8c78 100644
--- a/python/sglang/srt/models/sarvam_moe.py
+++ b/python/sglang/srt/models/sarvam_moe.py
@@ -39,7 +39,10 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
-from sglang.srt.layers.moe import should_use_flashinfer_cutlass_moe_fp4_allgather
+from sglang.srt.layers.moe import (
+    should_use_dp_reduce_scatterv,
+    should_use_flashinfer_cutlass_moe_fp4_allgather,
+)
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
@@ -375,6 +378,7 @@ class SarvamMoESparseMoeBlock(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states.view(num_tokens, hidden_dim)
@@ -418,6 +422,7 @@ class SarvamMoESparseMoeBlock(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
diff --git a/python/sglang/srt/models/sdar_moe.py b/python/sglang/srt/models/sdar_moe.py
index a1b20f67c..55ccc04f2 100644
--- a/python/sglang/srt/models/sdar_moe.py
+++ b/python/sglang/srt/models/sdar_moe.py
@@ -34,6 +34,7 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_moe_a2a_backend,
+    should_use_dp_reduce_scatterv,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
@@ -160,12 +161,13 @@ class SDARMoeSparseMoeBlock(nn.Module):
         topk_output = self.topk(hidden_states, router_logits)
         out = self.experts(hidden_states, topk_output)  # (T, H)
 
-        # TP all-reduce (unless fused / reduce_scatter / fp4 allgather path)
+        # TP all-reduce (unless fused / reduce_scatter / fp4 allgather / dp reduce_scatterv path)
         if (
             self.tp_size > 1
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             out = tensor_model_parallel_all_reduce(out)
 
diff --git a/python/sglang/srt/models/step3p5.py b/python/sglang/srt/models/step3p5.py
index e1f13e604..bcb978e0d 100644
--- a/python/sglang/srt/models/step3p5.py
+++ b/python/sglang/srt/models/step3p5.py
@@ -31,6 +31,7 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_moe_a2a_backend,
+    should_use_dp_reduce_scatterv,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
@@ -237,6 +238,7 @@ class Step3p5MoEMLP(nn.Module):
             and not should_allreduce_fusion
             and not use_reduce_scatter
             and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+            and not should_use_dp_reduce_scatterv()
         ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)