diff --git a/python/sglang/srt/layers/communicator.py b/python/sglang/srt/layers/communicator.py
index 2efc1e775..7a482f999 100644
--- a/python/sglang/srt/layers/communicator.py
+++ b/python/sglang/srt/layers/communicator.py
@@ -943,6 +943,23 @@ class CommunicateWithAllReduceAndLayerNormFn:
                 residual_input_mode=residual_input_mode,
             )
 
+        if (
+            (hidden_states_input_mode == ScatterMode.TP_ATTN_FULL)
+            and (
+                residual_input_mode in [ScatterMode.SCATTERED, ScatterMode.TP_ATTN_FULL]
+            )
+            and (hidden_states_output_mode == ScatterMode.TP_ATTN_FULL)
+            and (residual_output_mode == ScatterMode.TP_ATTN_FULL)
+            and context.attn_tp_size > 1
+        ):
+            # Used when the dense MLP is tensor-parallelized along the
+            # attention TP group (``moe_dense_tp_size > 1``): hidden states
+            # need an all-reduce inside the attention TP group before the
+            # next layernorm, while staying in TP_ATTN_FULL on both sides.
+            return (
+                CommunicateWithAllReduceAndLayerNormFn._tp_attn_all_reduce_and_layernorm
+            )
+
         raise NotImplementedError(
             f"{hidden_states_input_mode=} {residual_input_mode=} {hidden_states_output_mode=} {residual_output_mode=}"
         )
@@ -960,6 +977,25 @@ class CommunicateWithAllReduceAndLayerNormFn:
             hidden_states, residual = layernorm(hidden_states, residual)
         return hidden_states, residual
 
+    @staticmethod
+    def _tp_attn_all_reduce_and_layernorm(
+        hidden_states: torch.Tensor,
+        residual: torch.Tensor,
+        forward_batch: ForwardBatch,
+        layernorm: torch.nn.Module,
+        context: CommunicateContext,
+    ):
+        """All-reduce hidden states inside the attention TP group, then layernorm.
+
+        Used when the dense MLP shares the attention TP group
+        (``moe_dense_tp_size > 1``): both hidden states and residual stay in
+        ``TP_ATTN_FULL`` across the boundary.
+        """
+        hidden_states = get_attention_tp_group().all_reduce(hidden_states)
+        if hidden_states.shape[0] != 0:
+            hidden_states, residual = layernorm(hidden_states, residual)
+        return hidden_states, residual
+
     @staticmethod
     def _gather_hidden_states_and_residual(
         hidden_states: torch.Tensor,
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 247d113d8..76d9b440f 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -7199,10 +7199,11 @@ class ServerArgs:
         assert self.base_gpu_id >= 0, "base_gpu_id must be non-negative"
         assert self.gpu_id_step >= 1, "gpu_id_step must be positive"
 
-        assert self.moe_dense_tp_size in {
-            1,
+        assert self.moe_dense_tp_size in (
             None,
-        }, "moe_dense_tp_size only support 1 and None currently"
+            1,
+            self.tp_size,
+        ), "moe_dense_tp_size only supports None, 1, or tp_size currently"
 
         # Check served model name to not have colon as it is reserved for LoRA adapter syntax
         if not is_runai_obj_uri(self.served_model_name):
diff --git a/python/sglang/srt/utils/common.py b/python/sglang/srt/utils/common.py
index feb505d5d..f677b3ea5 100644
--- a/python/sglang/srt/utils/common.py
+++ b/python/sglang/srt/utils/common.py
@@ -3119,8 +3119,7 @@ def require_attn_tp_gather(server_args: ServerArgs):
 
     from sglang.srt.layers.moe.utils import get_moe_a2a_backend
 
-    assert server_args.moe_dense_tp_size in [1, None]
-    if not get_moe_a2a_backend().is_none() or server_args.moe_dense_tp_size == 1:
+    if not get_moe_a2a_backend().is_none() or server_args.moe_dense_tp_size is not None:
         if server_args.enable_dp_attention:
             return server_args.dp_size < server_args.tp_size
         else: