diff --git a/vllm/model_executor/layers/fused_moe/modular_kernel.py b/vllm/model_executor/layers/fused_moe/modular_kernel.py
index 0e55e827c20..f26e477ea2e 100644
--- a/vllm/model_executor/layers/fused_moe/modular_kernel.py
+++ b/vllm/model_executor/layers/fused_moe/modular_kernel.py
@@ -10,6 +10,7 @@ from typing import final
 import torch
 
 import vllm.envs as envs
+from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe.activation import (
     MoEActivation,
@@ -1132,6 +1133,17 @@ class FusedMoEKernelModularImpl:
         The _prepare method is a wrapper around self.prepare_finalize.prepare
         that handles DBO and async.
         """
+        # Skip cudagraph/DP padding tokens uniformly across all a2a backends:
+        # forcing padded rows' expert ids to -1 makes every prepare_finalize drop
+        # them (not dispatched / not computed by the experts). The V2 model runner
+        # marks them in forward_context.is_padding; it is None for runners that do
+        # not populate it, leaving topk_ids unchanged.
+        is_padding = get_forward_context().is_padding
+        if is_padding is not None:
+            n = topk_ids.shape[0]
+            # TODO: Properly support DBO (padding lives at the batch tail).
+            topk_ids = torch.where(is_padding[:n].unsqueeze(1), -1, topk_ids)
+
         if not self.prepare_finalize.supports_async():
             # We shouldn't be running an a2a kernel that doesn't
             # support async prepare/finalize
diff --git a/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py b/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py
index d0d487be692..129e3b5d5c2 100644
--- a/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py
+++ b/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py
@@ -289,12 +289,6 @@ class DeepEPV2PrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
             )
             a1 = a1 * topk_weights.to(a1.dtype)
 
-        is_padding = get_forward_context().is_padding
-        if is_padding is not None:
-            n = topk_ids.shape[0]
-            # TODO: Properly support DBO.
-            topk_ids = torch.where(is_padding[:n].unsqueeze(1), -1, topk_ids)
-
         if quant_config.is_block_quantized and not defer_input_quant:
             a1q, a1q_scale = moe_kernel_quantize_input(
                 a1,