From e6f6a353bf69b836812e0f6f3a34ead90b03357a Mon Sep 17 00:00:00 2001
From: Feng Yao <feng@thinkingmachines.ai>
Date: Wed, 1 Jul 2026 21:07:58 -0700
Subject: [PATCH] pad customized_info for mixed output batches (#29842)

---
 .../scheduler_components/output_streamer.py   | 20 +++-
 .../test_output_streamer_customized_info.py   | 95 +++++++++++++++++++
 2 files changed, 111 insertions(+), 4 deletions(-)
 create mode 100644 test/registered/unit/managers/test_output_streamer_customized_info.py

diff --git a/python/sglang/srt/managers/scheduler_components/output_streamer.py b/python/sglang/srt/managers/scheduler_components/output_streamer.py
index 1be8cc088..6b691d84e 100644
--- a/python/sglang/srt/managers/scheduler_components/output_streamer.py
+++ b/python/sglang/srt/managers/scheduler_components/output_streamer.py
@@ -499,11 +499,23 @@ class _GenerationStreamAccumulator:
                 req.indexer_topk if req.return_indexer_topk else None
             )
 
+        current_output_len = len(self.output_ids[-1])
         if req.customized_info is not None:
-            for k, v in req.customized_info.items():
-                if k not in self.customized_info:
-                    self.customized_info[k] = []
-                self.customized_info[k].append(v[send_token_offset : len(output_ids_)])
+            for key, req_values in req.customized_info.items():
+                if key not in self.customized_info:
+                    self.customized_info[key] = [
+                        [None] * len(prev_output_ids)
+                        for prev_output_ids in self.output_ids[:-1]
+                    ]
+                self.customized_info[key].append(
+                    [None] * current_output_len
+                    if req_values is None
+                    else req_values[send_token_offset : len(output_ids_)]
+                )
+
+        for per_request_values in self.customized_info.values():
+            if len(per_request_values) < len(self.output_ids):
+                per_request_values.append([None] * current_output_len)
 
     def to_payload(
         self, *, dp_rank: int, is_idle_batch: bool
diff --git a/test/registered/unit/managers/test_output_streamer_customized_info.py b/test/registered/unit/managers/test_output_streamer_customized_info.py
new file mode 100644
index 000000000..b0be57dcb
--- /dev/null
+++ b/test/registered/unit/managers/test_output_streamer_customized_info.py
@@ -0,0 +1,95 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.disaggregation.utils import DisaggregationMode
+from sglang.srt.managers.io_struct import unwrap_from_pickle
+from sglang.srt.managers.scheduler_components.output_streamer import (
+    _GenerationStreamAccumulator,
+)
+from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=1, suite="base-a-test-cpu")
+
+
+class _FakeReq:
+    def __init__(self, rid, output_ids, customized_info=None):
+        self.rid = rid
+        self.http_worker_ipc = None
+        self.finished_reason = None
+        self.finished_output = False
+        self.finished_len = None
+        self.stream = False
+        self.sampling_params = SimpleNamespace(
+            stream_interval=None,
+            skip_special_tokens=True,
+            spaces_between_special_tokens=True,
+            no_stop_trim=False,
+        )
+        self.output_ids = output_ids
+        self.output_ids_through_stop = output_ids
+        self.send_token_offset = 0
+        self.send_output_token_logprobs_offset = 0
+        self.send_decode_id_offset = 0
+        self.decoded_text = ""
+        self.origin_input_ids = []
+        self.reasoning_tokens = 0
+        self.cached_tokens = 0
+        self.retraction_count = 0
+        self.time_stats = None
+        self.mm_image_tokens = 0
+        self.mm_audio_tokens = 0
+        self.mm_video_tokens = 0
+        self.multimodal_inputs = None
+        self.customized_info = customized_info
+
+    def finished(self):
+        return False
+
+    def init_incremental_detokenize(self):
+        return self.output_ids_through_stop, 0
+
+    def check_match_stop_str_prefix(self):
+        return False
+
+
+class TestOutputStreamerCustomizedInfo(unittest.TestCase):
+    def test_customized_info_is_padded_for_mixed_batches(self):
+        accumulator = _GenerationStreamAccumulator(
+            return_logprob=False,
+            return_hidden_states=False,
+            return_routed_experts=False,
+            return_indexer_topk=False,
+            spec_algorithm=SpeculativeAlgorithm.NONE,
+            disaggregation_mode=DisaggregationMode.NULL,
+            default_stream_interval=1,
+            default_force_stream_interval=1,
+            get_cached_tokens_details=lambda req: None,
+        )
+
+        accumulator.accept(req=_FakeReq("r0", [10, 11]))
+        accumulator.accept(
+            req=_FakeReq(
+                "r1",
+                [20, 21, 22],
+                customized_info={"probe": [200, 201, 202]},
+            )
+        )
+        accumulator.accept(req=_FakeReq("r2", [30], customized_info={"other": [300]}))
+
+        payload = accumulator.to_payload(dp_rank=0, is_idle_batch=False)
+        customized_info = unwrap_from_pickle(payload.customized_info)
+
+        self.assertEqual(payload.output_ids, [[10, 11], [20, 21, 22], [30]])
+        self.assertEqual(
+            customized_info["probe"],
+            [[None, None], [200, 201, 202], [None]],
+        )
+        self.assertEqual(
+            customized_info["other"],
+            [[None, None], [None, None, None], [300]],
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()