From 3e7cc27318c29ebbf3cde234f7bd47beb4a128b9 Mon Sep 17 00:00:00 2001 From: MayDomine <1583143678@qq.com> Date: Wed, 12 Nov 2025 11:29:33 +0800 Subject: [PATCH] At least tell the user that ngram verify is greedy! (#13039) --- python/sglang/srt/speculative/ngram_info.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/python/sglang/srt/speculative/ngram_info.py b/python/sglang/srt/speculative/ngram_info.py index 6be65770c..b0a5c8d99 100644 --- 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)