From b684b0b72f81cbd251c40db717a992ff3c8a210e Mon Sep 17 00:00:00 2001
From: Qiaolin Yu <liin1211@outlook.com>
Date: Thu, 2 Apr 2026 01:55:16 -0700
Subject: [PATCH] Fix spec v2 + logprob when max_num_token is set (#20799)

---
 .../scheduler_output_processor_mixin.py       | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/python/sglang/srt/managers/scheduler_output_processor_mixin.py b/python/sglang/srt/managers/scheduler_output_processor_mixin.py
index ef31d34d6..e3b8dc87e 100644
--- a/python/sglang/srt/managers/scheduler_output_processor_mixin.py
+++ b/python/sglang/srt/managers/scheduler_output_processor_mixin.py
@@ -1034,6 +1034,8 @@ class SchedulerOutputProcessorMixin:
                         and not req.input_logprob_sent
                         # Decode server does not send input logprobs
                         and self.disaggregation_mode != DisaggregationMode.DECODE
+                        # Only send when input logprobs have been computed (after prefill)
+                        and req.input_token_logprobs_val is not None
                     ):
                         input_token_logprobs_val.append(req.input_token_logprobs_val)
                         input_token_logprobs_idx.append(req.input_token_logprobs_idx)
@@ -1055,39 +1057,38 @@ class SchedulerOutputProcessorMixin:
                         input_token_ids_logprobs_idx.append([])
 
                     if req.return_logprob:
+                        logprob_end = max(len(output_ids_), 1)
                         output_token_logprobs_val.append(
                             req.output_token_logprobs_val[
-                                send_output_token_logprobs_offset:
+                                send_output_token_logprobs_offset:logprob_end
                             ]
                         )
                         output_token_logprobs_idx.append(
                             req.output_token_logprobs_idx[
-                                send_output_token_logprobs_offset:
+                                send_output_token_logprobs_offset:logprob_end
                             ]
                         )
                         output_top_logprobs_val.append(
                             req.output_top_logprobs_val[
-                                send_output_token_logprobs_offset:
+                                send_output_token_logprobs_offset:logprob_end
                             ]
                         )
                         output_top_logprobs_idx.append(
                             req.output_top_logprobs_idx[
-                                send_output_token_logprobs_offset:
+                                send_output_token_logprobs_offset:logprob_end
                             ]
                         )
                         output_token_ids_logprobs_val.append(
                             req.output_token_ids_logprobs_val[
-                                send_output_token_logprobs_offset:
+                                send_output_token_logprobs_offset:logprob_end
                             ]
                         )
                         output_token_ids_logprobs_idx.append(
                             req.output_token_ids_logprobs_idx[
-                                send_output_token_logprobs_offset:
+                                send_output_token_logprobs_offset:logprob_end
                             ]
                         )
-                        req.send_output_token_logprobs_offset = len(
-                            req.output_token_logprobs_val
-                        )
+                        req.send_output_token_logprobs_offset = logprob_end
                     else:
                         output_token_logprobs_val.append([])
                         output_token_logprobs_idx.append([])