From c47f0e7cdde48ddc718e3c6ee8bc87bebee2e8ff Mon Sep 17 00:00:00 2001
From: Jun Liu <jun.c.liu@rakuten.com>
Date: Tue, 26 May 2026 23:01:10 +0900
Subject: [PATCH] [PD] Fix top logprobs crash in prefill path (#26299)

---
 python/sglang/srt/disaggregation/prefill.py   | 11 ++++++++
 .../test_disaggregation_basic.py              | 26 +++++++++++++++++++
 2 files changed, 37 insertions(+)

diff --git a/python/sglang/srt/disaggregation/prefill.py b/python/sglang/srt/disaggregation/prefill.py
index bb07f4012..93dee0cc2 100644
--- a/python/sglang/srt/disaggregation/prefill.py
+++ b/python/sglang/srt/disaggregation/prefill.py
@@ -510,6 +510,17 @@ class SchedulerDisaggregationPrefillMixin:
                 logits_output.input_token_logprobs = tuple(
                     logits_output.input_token_logprobs.tolist()
                 )
+            if logits_output.next_token_top_logprobs_val:
+                logits_output.next_token_top_logprobs_val = [
+                    v.tolist() for v in logits_output.next_token_top_logprobs_val
+                ]
+                logits_output.next_token_top_logprobs_idx = [
+                    x.tolist() for x in logits_output.next_token_top_logprobs_idx
+                ]
+            if logits_output.next_token_token_ids_logprobs_val:
+                logits_output.next_token_token_ids_logprobs_val = [
+                    v.tolist() for v in logits_output.next_token_token_ids_logprobs_val
+                ]
 
         for i, (req, next_token_id) in enumerate(
             zip(batch.reqs, next_token_ids, strict=True)
diff --git a/test/registered/disaggregation/test_disaggregation_basic.py b/test/registered/disaggregation/test_disaggregation_basic.py
index f2ee9de56..0cdbf0340 100644
--- a/test/registered/disaggregation/test_disaggregation_basic.py
+++ b/test/registered/disaggregation/test_disaggregation_basic.py
@@ -72,6 +72,32 @@ class TestDisaggregationAccuracy(PauseResumeInPlaceMixin, PDDisaggregationServer
             len(input_logprobs) > 0
         ), f"input_logprobs should have at least one token, but got {len(input_logprobs)}"
 
+    def test_chat_completion_top_logprobs(self):
+        client = openai.Client(api_key="empty", base_url=f"{self.lb_url}/v1")
+        response = client.chat.completions.create(
+            model="dummy",
+            messages=[
+                {"role": "system", "content": "You are a helpful AI assistant."},
+                {"role": "user", "content": "What is the capital of France?"},
+            ],
+            temperature=0,
+            max_tokens=8,
+            logprobs=True,
+            top_logprobs=5,
+        )
+
+        self.assertIsNotNone(response.choices[0].logprobs)
+        content_logprobs = response.choices[0].logprobs.content
+        self.assertGreater(len(content_logprobs), 0)
+
+        first_top_logprobs = next(
+            (item.top_logprobs for item in content_logprobs if item.top_logprobs),
+            None,
+        )
+        self.assertIsNotNone(first_top_logprobs)
+        self.assertGreater(len(first_top_logprobs), 0)
+        self.assertIsInstance(first_top_logprobs[0].token, str)
+
     def test_structured_output(self):
         json_schema = json.dumps(
             {