Clamp logprob tokens with model vocab size (#14414)

2025-12-29 20:37:07 +08:00
parent f3d73b0199
commit d48723b77d
1 changed files with 3 additions and 0 deletions
--- 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