From d48723b77d0ef2c220051fc53fa58beff9483b50 Mon Sep 17 00:00:00 2001
From: cklxx <1293822641@qq.com>
Date: Mon, 29 Dec 2025 20:37:07 +0800
Subject: [PATCH] Clamp logprob tokens with model vocab size (#14414)

---
 python/sglang/srt/managers/schedule_batch.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/python/sglang/srt/managers/schedule_batch.py b/python/sglang/srt/managers/schedule_batch.py
index 05125bd3a..e49ad99f6 100644
--- a/python/sglang/srt/managers/schedule_batch.py
+++ b/python/sglang/srt/managers/schedule_batch.py
@@ -1551,6 +1551,9 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             extend_input_logprob_token_ids = torch.tensor(
                 extend_input_logprob_token_ids
             )
+            # Clamp placeholder or out-of-range token IDs (e.g., multimodal hashes)
+            # so they stay within the vocab boundary before being sent to GPU.
+            extend_input_logprob_token_ids.clamp_(0, self.model_config.vocab_size - 1)
         else:
             extend_input_logprob_token_ids = None