diff --git a/python/sglang/srt/layers/communicator.py b/python/sglang/srt/layers/communicator.py
index 7de4337f9..30f068595 100644
--- a/python/sglang/srt/layers/communicator.py
+++ b/python/sglang/srt/layers/communicator.py
@@ -46,6 +46,7 @@ from sglang.srt.layers.dp_attention import (
     get_attention_cp_rank,
     get_attention_cp_size,
     get_attention_dp_size,
+    get_attention_tp_group,
     get_attention_tp_rank,
     get_attention_tp_size,
     get_dp_global_num_tokens,
@@ -855,7 +856,7 @@ class CommunicateSimpleFn:
             return tuple(gathered_hidden_states)
 
         hidden_states, local_hidden_states = (
-            get_local_dp_buffer(),
+            get_local_dp_buffer(get_attention_tp_group()),
             hidden_states,
         )
         attn_tp_all_gather_into_tensor(
@@ -965,7 +966,7 @@ class CommunicateWithAllReduceAndLayerNormFn:
 
         if residual_input_mode == ScatterMode.SCATTERED and context.attn_tp_size > 1:
             residual, local_residual = (
-                get_local_dp_buffer(),
+                get_local_dp_buffer(get_attention_tp_group()),
                 residual,
             )
             attn_tp_all_gather_into_tensor(residual, local_residual)
@@ -982,7 +983,7 @@ class CommunicateWithAllReduceAndLayerNormFn:
                 hidden_states += residual
 
             hidden_states, local_hidden_states = (
-                get_global_dp_buffer(),
+                get_global_dp_buffer(get_tp_group()),
                 hidden_states,
             )
             dp_gather_partial(hidden_states, local_hidden_states, forward_batch)
@@ -1210,8 +1211,12 @@ class CommunicateSummableTensorPairFn:
         context: CommunicateContext,
         allow_reduce_scatter: bool = False,
     ):
+        if get_tensor_model_parallel_world_size() == get_attention_dp_size():
+            group = get_tp_group()
+        else:
+            group = get_attention_tp_group()
         hidden_states, global_hidden_states = (
-            get_local_dp_buffer(),
+            get_local_dp_buffer(group),
             hidden_states,
         )
         if should_use_dp_reduce_scatterv():
@@ -1237,7 +1242,7 @@ class CommunicateSummableTensorPairFn:
         hidden_states += residual
         residual = None
         hidden_states, local_hidden_states = (
-            get_local_dp_buffer(),
+            get_local_dp_buffer(get_attention_tp_group()),
             hidden_states,
         )
         attn_tp_all_gather_into_tensor(
diff --git a/python/sglang/srt/layers/communicator_nsa_cp.py b/python/sglang/srt/layers/communicator_nsa_cp.py
index 243b18e00..2508929f1 100644
--- a/python/sglang/srt/layers/communicator_nsa_cp.py
+++ b/python/sglang/srt/layers/communicator_nsa_cp.py
@@ -34,6 +34,7 @@ from sglang.srt.layers.communicator import (
 from sglang.srt.layers.dp_attention import (
     attn_cp_all_gather_into_tensor,
     attn_cp_reduce_scatter_tensor,
+    get_attention_cp_group,
     get_local_dp_buffer,
 )
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
@@ -154,7 +155,7 @@ class NSACPCommunicateWithAllReduceAndLayerNormFn(
         if nsa_use_prefill_cp(forward_batch):
             assert context.attn_dp_size == 1
             hidden_states, local_hidden_states = (
-                get_local_dp_buffer(),
+                get_local_dp_buffer(get_attention_cp_group()),
                 hidden_states,
             )
             attn_cp_all_gather_into_tensor(
diff --git a/python/sglang/srt/layers/dp_attention.py b/python/sglang/srt/layers/dp_attention.py
index b8d761784..89d238928 100644
--- a/python/sglang/srt/layers/dp_attention.py
+++ b/python/sglang/srt/layers/dp_attention.py
@@ -126,8 +126,8 @@ class _DpGatheredBufferWrapper:
         cls._global_num_tokens = global_num_tokens
 
     @classmethod
-    def get_global_dp_buffer(cls) -> torch.Tensor:
-        with use_symmetric_memory(get_tp_group(), disabled=not cls._dp_max_padding):
+    def get_global_dp_buffer(cls, group: GroupCoordinator) -> torch.Tensor:
+        with use_symmetric_memory(group, disabled=not cls._dp_max_padding):
             buffer = torch.empty(
                 (cls._global_dp_buffer_len, cls._hidden_size),
                 dtype=cls._dtype,
@@ -136,8 +136,8 @@ class _DpGatheredBufferWrapper:
         return buffer
 
     @classmethod
-    def get_local_dp_buffer(cls) -> torch.Tensor:
-        with use_symmetric_memory(get_tp_group(), disabled=not cls._dp_max_padding):
+    def get_local_dp_buffer(cls, group: GroupCoordinator) -> torch.Tensor:
+        with use_symmetric_memory(group, disabled=not cls._dp_max_padding):
             buffer = torch.empty(
                 (cls._local_dp_buffer_len, cls._hidden_size),
                 dtype=cls._dtype,
@@ -193,12 +193,12 @@ def set_dp_buffer_len(
     )
 
 
-def get_global_dp_buffer() -> torch.Tensor:
-    return _DpGatheredBufferWrapper.get_global_dp_buffer()
+def get_global_dp_buffer(group: GroupCoordinator) -> torch.Tensor:
+    return _DpGatheredBufferWrapper.get_global_dp_buffer(group=group)
 
 
-def get_local_dp_buffer() -> torch.Tensor:
-    return _DpGatheredBufferWrapper.get_local_dp_buffer()
+def get_local_dp_buffer(group: GroupCoordinator) -> torch.Tensor:
+    return _DpGatheredBufferWrapper.get_local_dp_buffer(group=group)
 
 
 def get_global_dp_buffer_len() -> int:
diff --git a/python/sglang/srt/layers/moe/token_dispatcher/standard.py b/python/sglang/srt/layers/moe/token_dispatcher/standard.py
index 7658c28d4..caf99571d 100644
--- a/python/sglang/srt/layers/moe/token_dispatcher/standard.py
+++ b/python/sglang/srt/layers/moe/token_dispatcher/standard.py
@@ -220,7 +220,10 @@ class StandardDispatcher(BaseDispatcher):
     def combine(self, combine_input: StandardCombineInput) -> torch.Tensor:
         (hidden_states,) = combine_input
         if should_use_flashinfer_cutlass_moe_fp4_allgather():
-            hidden_states, global_hidden_states = get_local_dp_buffer(), hidden_states
+            hidden_states, global_hidden_states = (
+                get_local_dp_buffer(get_tp_group()),
+                hidden_states,
+            )
             get_tp_group().reduce_scatterv(
                 global_hidden_states,
                 output=hidden_states,