At least tell the user that ngram verify is greedy! (#13039)

2025-11-12 11:29:33 +08:00
parent 8f01a12d43
commit 3e7cc27318
1 changed files with 3 additions and 0 deletions
--- a/python/sglang/srt/speculative/ngram_info.py
+++ b/python/sglang/srt/speculative/ngram_info.py
@@ -419,6 +419,9 @@ class NgramVerifyInput(SpecInput):
            self._greedy_verify(batch, logits_output)
        else:
            # NOTE: Compared with greedy_verify, the performance of _sampling_verify is relatively poor.
+            logger.warning(
+                "Currently Ngram speculative sampling forces to use greedy verification."
+            )
            self._greedy_verify(batch, logits_output)
            # self._sampling_verify(batch, logits_output, sampling_info)