diff --git a/python/sglang/srt/layers/communicator.py b/python/sglang/srt/layers/communicator.py
index 56d3fff29..591ce2d7f 100644
--- a/python/sglang/srt/layers/communicator.py
+++ b/python/sglang/srt/layers/communicator.py
@@ -1304,8 +1304,12 @@ class CommunicateSummableTensorPairFn:
 
         # DP scatter (if DP attention is enabled)
         if context.attn_dp_size > 1:
+            if get_tensor_model_parallel_world_size() == get_attention_dp_size():
+                group = get_tp_group()
+            else:
+                group = get_attention_tp_group()
             hidden_states_output, global_hidden_states = (
-                get_local_dp_buffer(),
+                get_local_dp_buffer(group),
                 hidden_states,
             )
             dp_scatter(hidden_states_output, global_hidden_states, forward_batch)
diff --git a/python/sglang/srt/models/deepseek_v4.py b/python/sglang/srt/models/deepseek_v4.py
index 1535a81db..4ab6d00e8 100644
--- a/python/sglang/srt/models/deepseek_v4.py
+++ b/python/sglang/srt/models/deepseek_v4.py
@@ -26,7 +26,11 @@ from sglang.jit_kernel.deepseek_v4 import (
     fused_rope_inplace,
 )
 from sglang.srt.configs.deepseek_v4 import DeepSeekV4Config
-from sglang.srt.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from sglang.srt.distributed import (
+    get_pp_group,
+    get_tensor_model_parallel_world_size,
+    get_tp_group,
+)
 from sglang.srt.environ import envs
 from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
 from sglang.srt.layers.attention.dsv4.compressor import Compressor
@@ -885,7 +889,10 @@ class DeepseekV4DecoderLayer(nn.Module):
             input_ids = input_ids[cp_rank::cp_size].contiguous()
             input_ids_global = input_ids
         elif _use_tp_moe_gather:
-            hidden_states, local_hidden_states = get_global_dp_buffer(), hidden_states
+            hidden_states, local_hidden_states = (
+                get_global_dp_buffer(get_tp_group()),
+                hidden_states,
+            )
             dp_gather_partial(hidden_states, local_hidden_states, forward_batch)
         _a2a_scatter_chunks: Optional[List[torch.Tensor]] = None
         if _use_tp_attn_a2a_scatter:
@@ -901,7 +908,10 @@ class DeepseekV4DecoderLayer(nn.Module):
             input_ids_global=input_ids_global,
         )
         if _use_tp_moe_gather:
-            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,
+            )
             dp_scatter(hidden_states, global_hidden_states, forward_batch)
         if _use_tp_attn_a2a_scatter:
             assert _a2a_scatter_chunks is not None