diff --git a/python/sglang/srt/distributed/parallel_state.py b/python/sglang/srt/distributed/parallel_state.py
index e8e14b9f3..447715230 100644
--- a/python/sglang/srt/distributed/parallel_state.py
+++ b/python/sglang/srt/distributed/parallel_state.py
@@ -1724,6 +1724,7 @@ def initialize_model_parallel(
     moe_data_model_parallel_size: int = 1,
     backend: Optional[str] = None,
     duplicate_tp_group: bool = False,
+    enable_symm_mem: bool = False,
 ) -> None:
     """
     Initialize model parallel groups.
@@ -1884,7 +1885,7 @@ def initialize_model_parallel(
             group_ranks,
             get_world_group().local_rank,
             backend,
-            use_pynccl=SYNC_TOKEN_IDS_ACROSS_TP,
+            use_pynccl=SYNC_TOKEN_IDS_ACROSS_TP or enable_symm_mem,
             use_mscclpp_allreduce=False,
             use_custom_allreduce=False,
             use_torch_symm_mem_allreduce=False,
diff --git a/python/sglang/srt/layers/linear.py b/python/sglang/srt/layers/linear.py
index 7af9eb004..74d9a7271 100644
--- a/python/sglang/srt/layers/linear.py
+++ b/python/sglang/srt/layers/linear.py
@@ -1503,9 +1503,13 @@ class RowParallelLinear(LinearBase):
         # Only fuse bias add into GEMM for rank 0 (this ensures that
         # bias will not get added more than once in TP>1 case)
         bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
-        with use_symmetric_memory(
-            get_tp_group(), disabled=not is_allocation_symmetric()
-        ):
+        if self.use_dp_attention_reduce:
+            symm_ctx = use_symmetric_memory(get_attention_tp_group())
+        else:
+            symm_ctx = use_symmetric_memory(
+                get_tp_group(), disabled=not is_allocation_symmetric()
+            )
+        with symm_ctx:
             output_parallel = self.quant_method.apply(self, input_parallel, bias=bias_)
 
         if self.reduce_results and self.tp_size > 1 and not skip_all_reduce:
diff --git a/python/sglang/srt/model_executor/model_runner.py b/python/sglang/srt/model_executor/model_runner.py
index 76d9ded64..e4948bc47 100644
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -1027,6 +1027,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
                 attention_context_model_parallel_size=self.attn_cp_size,
                 moe_data_model_parallel_size=self.moe_dp_size,
                 duplicate_tp_group=self.server_args.enable_pdmux,
+                enable_symm_mem=self.server_args.enable_symm_mem,
             )
             initialize_dp_attention(
                 server_args=self.server_args,
diff --git a/python/sglang/srt/models/llama.py b/python/sglang/srt/models/llama.py
index b8ad74015..447f57eaf 100644
--- a/python/sglang/srt/models/llama.py
+++ b/python/sglang/srt/models/llama.py
@@ -73,6 +73,7 @@ class LlamaMLP(nn.Module):
         reduce_results: bool = True,
         tp_rank: Optional[int] = None,
         tp_size: Optional[int] = None,
+        use_dp_attention_reduce: bool = False,
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
@@ -93,6 +94,7 @@ class LlamaMLP(nn.Module):
             reduce_results=reduce_results,
             tp_rank=tp_rank,
             tp_size=tp_size,
+            use_dp_attention_reduce=use_dp_attention_reduce,
         )
         if hidden_act != "silu":
             raise ValueError(