diff --git a/tests/v1/distributed/test_dbo.py b/tests/v1/distributed/test_dbo.py
index 866ae742bf3c..16f154d196ba 100644
--- a/tests/v1/distributed/test_dbo.py
+++ b/tests/v1/distributed/test_dbo.py
@@ -85,5 +85,4 @@ def test_dbo_dp_ep_gsm8k(all2all_backend: str, num_gpus_available):
         assert accuracy >= MIN_ACCURACY, (
             f"DBO+DP+EP accuracy too low ({all2all_backend}): "
             f"{accuracy:.3f} < {MIN_ACCURACY:.3f} "
-            f"(correct: {results['num_correct']}/{results['num_questions']})"
         )
diff --git a/vllm/v1/attention/backends/utils.py b/vllm/v1/attention/backends/utils.py
index 27f07218d9b2..8edfbb5140bc 100644
--- a/vllm/v1/attention/backends/utils.py
+++ b/vllm/v1/attention/backends/utils.py
@@ -166,9 +166,7 @@ def _make_metadata_with_slice(
     assert start_locs[first_req] <= first_tok < start_locs[first_req + 1], (
         "Token slice start outside of first request"
     )
-    assert start_locs[last_req] <= last_tok < start_locs[last_req + 1], (
-        "Token slice end outside of last request"
-    )
+    # NOTE: last token can be outside of the last request if we have CG padding.
 
     # If the "middle" request has tokens in both ubatches, we have to split it.
     # If ubatch_slice is the first ubatch then we will be splitting the last
diff --git a/vllm/v1/worker/dp_utils.py b/vllm/v1/worker/dp_utils.py
index c1509de821b0..6539d72d81cb 100644
--- a/vllm/v1/worker/dp_utils.py
+++ b/vllm/v1/worker/dp_utils.py
@@ -93,13 +93,16 @@ def _post_process_dp_padding(tensor: torch.Tensor, should_dp_pad: bool) -> torch
 
 # This just pads the second ubatch slice out to the total number of tokens
 # (num_tokens + padding) since we do `create_ubatch_slices` before applying DP padding.
-def _pad_out_ubatch_slice(ubatch_slices: UBatchSlices, num_total_tokens: int):
-    padded_second_ubatch_slice = slice(
+def _pad_out_ubatch_slice(
+    ubatch_slices: UBatchSlices, num_total_tokens: int
+) -> UBatchSlices:
+    padded_second_token_slice = slice(
         ubatch_slices[1].token_slice.start, num_total_tokens
     )
     ubatch_slices[1] = UBatchSlice(
-        padded_second_ubatch_slice, padded_second_ubatch_slice
+        ubatch_slices[1].request_slice, padded_second_token_slice
     )
+    return ubatch_slices
 
 
 def _synchronize_dp_ranks(