diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
index 56274f875f40..2e257c306ae9 100755
--- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
+++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
@@ -110,13 +110,11 @@ def __init__(
         assert len(
             self.initial_local_expert_ids) == self.expert_size_per_partition
 
-        max_num_tokens = model_config.max_num_tokens
         # The maximum number of tokens in MoE are multiplied by DP size when attention DP is enabled
-        if self.use_dp:
-            max_num_tokens *= model_config.mapping.world_size
-        self.moe_max_num_tokens = model_config.moe_max_num_tokens or max_num_tokens
+        moe_max_num_tokens = model_config.max_num_tokens * model_config.mapping.dp_size
+        self.moe_max_num_tokens = model_config.moe_max_num_tokens or moe_max_num_tokens
         # The auxiliary CUDA stream and CUDA events are only used when MoE chunking is applied
-        if self.moe_max_num_tokens < max_num_tokens:
+        if self.moe_max_num_tokens < moe_max_num_tokens:
             self.aux_stream = aux_stream_dict[
                 AuxStreamType.
                 MoeChunkingOverlap] if aux_stream_dict is not None else torch.cuda.Stream(
diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
index fcfdeb0fad28..68bf255a1422 100644
--- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
+++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
@@ -327,6 +327,18 @@ def __init__(
         apply_router_weight_on_input: bool = False,
         layer_idx: Optional[int] = None,
     ):
+        if model_config.moe_max_num_tokens is None:
+            moe_max_num_tokens = model_config.max_num_tokens * model_config.mapping.dp_size
+            # The default moe_max_num_tokens is calculated from the following formula:
+            # max_isl = 8196, max_batch_size = 1024, mtp = 0
+            # max_num_tokens = ((mtp+1)*max_batch_size+max_isl+128+63)//64*64 = 9344
+            # moe_max_num_tokens = max_num_tokens * 2 = 18688
+            # It can avoid OOM for 8k/1k cases.
+            default_moe_max_num_tokens = 18688
+            if moe_max_num_tokens > default_moe_max_num_tokens:
+                model_config._frozen = False
+                model_config.moe_max_num_tokens = default_moe_max_num_tokens
+                model_config._frozen = True
 
         super().__init__(
             routing_method=routing_method,
diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py
index 3249bac979b2..ed6f11993b20 100644
--- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py
+++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py
@@ -81,13 +81,9 @@ def __init__(
             self.num_experts)
         self.expert_size_per_partition = self.expert_end - self.expert_start
 
-        max_num_tokens = model_config.max_num_tokens
         # The maximum number of tokens in MoE are multiplied by DP size when attention DP is enabled
-        if self.use_dp:
-            max_num_tokens *= model_config.mapping.world_size
-        self.moe_max_num_tokens = (model_config.moe_max_num_tokens
-                                   if model_config.moe_max_num_tokens
-                                   is not None else max_num_tokens)
+        moe_max_num_tokens = model_config.max_num_tokens * model_config.mapping.dp_size
+        self.moe_max_num_tokens = model_config.moe_max_num_tokens or moe_max_num_tokens
 
         self._weights_created = False
         if not model_config.skip_create_weights_in_init:
diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
index c74c8966b68b..4eb9d77606b4 100755
--- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
+++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
@@ -150,12 +150,11 @@ def __init__(
         assert len(
             self.initial_local_expert_ids) == self.expert_size_per_partition
 
-        max_num_tokens = model_config.max_num_tokens
         # The maximum number of tokens in MoE are multiplied by DP size when attention DP is enabled
-        max_num_tokens *= model_config.mapping.world_size
-        self.moe_max_num_tokens = model_config.moe_max_num_tokens if model_config.moe_max_num_tokens is not None else max_num_tokens
+        moe_max_num_tokens = model_config.max_num_tokens * model_config.mapping.dp_size
+        self.moe_max_num_tokens = model_config.moe_max_num_tokens or moe_max_num_tokens
         # The auxiliary CUDA stream and CUDA events are only used when MoE chunking is applied
-        if self.moe_max_num_tokens < max_num_tokens:
+        if self.moe_max_num_tokens < moe_max_num_tokens:
             self.aux_stream = aux_stream_dict[
                 AuxStreamType.
                 MoeChunkingOverlap] if aux_stream_dict is not None else torch.cuda.Stream(
diff --git a/tensorrt_llm/mapping.py b/tensorrt_llm/mapping.py
index f78fe093f713..22824ea350d9 100644
--- a/tensorrt_llm/mapping.py
+++ b/tensorrt_llm/mapping.py
@@ -372,6 +372,10 @@ def node_rank(self):
     def local_rank(self):
         return self.rank % self.gpus_per_node
 
+    @property
+    def dp_size(self):
+        return self.tp_size if self.enable_attention_dp else 1
+
     def has_cp(self):
         return self.cp_size > 1