diff --git a/tests/models/test_deepseek_v41_replay_batch.py b/tests/models/test_deepseek_v41_replay_batch.py
index 5a3625dcbf..42d6d4d725 100644
--- a/tests/models/test_deepseek_v41_replay_batch.py
+++ b/tests/models/test_deepseek_v41_replay_batch.py
@@ -63,6 +63,7 @@ def state(monkeypatch):
     cfg.scheduler_config.max_num_seqs = 16
     cfg.scheduler_config.max_num_batched_tokens = 1024
     cfg.parallel_config.data_parallel_size = 1
+    cfg.aux_output_config.enable_return_routed_experts = False
     cfg.compilation_config.fast_moe_cold_start = False
     cfg.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
     layers = DecoderReplayLayers(WINDOW, MagicMock(), MagicMock(), set(), "swa_first")
@@ -195,6 +196,27 @@ def test_prompt_logprobs_requests_keep_their_rows(state):
     assert build.batch.max_query_len == 300
 
 
+@pytest.mark.parametrize("cg_mode", [CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE])
+def test_aux_output_requests_keep_every_routing_row(state, cg_mode):
+    """R3 needs every prefill row even when final logits only need the window."""
+    state.vllm_config.aux_output_config.enable_return_routed_experts = True
+    for index in range(3):
+        state.add_request(
+            index,
+            SimpleNamespace(
+                replay_start=0,
+                sampling_params=SimpleNamespace(prompt_logprobs=None),
+            ),
+        )
+    batch = _input_batch(QUERY_LENS, SEQ_LENS, PREFILLING)
+    state.decoder_replay_layers.trim_threshold = WINDOW + 1
+    replay, _ = _prepare(state, batch, cg_mode)
+    if cg_mode == CUDAGraphMode.NONE:
+        assert replay is None
+    else:
+        torch.testing.assert_close(replay.rows, torch.arange(401, device=DEVICE))
+
+
 def test_replay_batch_keeps_adaptive_verification_query_bound(state):
     """Adaptive verification bounds the decodes' device-side query lengths
     above their CPU lengths; the replay batch keeps that bound."""
diff --git a/vllm/models/deepseek_v41/nvidia/model_state.py b/vllm/models/deepseek_v41/nvidia/model_state.py
index e8bbaef64d..4e6c459d1e 100644
--- a/vllm/models/deepseek_v41/nvidia/model_state.py
+++ b/vllm/models/deepseek_v41/nvidia/model_state.py
@@ -273,7 +273,7 @@ class DeepseekV41ModelState(DefaultModelState):
             params = new_req_data.sampling_params
             self._keeps_rows_np[req_index] = (
                 params is not None and params.prompt_logprobs is not None
-            )
+            ) or self.vllm_config.aux_output_config.enable_return_routed_experts
 
     def prepare_inputs(
         self, input_batch: InputBatch, req_states: RequestState