From c2ac37dbcc9a89d1186b26e991dcce4f2c586951 Mon Sep 17 00:00:00 2001 From: Liangsheng Yin Date: Fri, 29 May 2026 18:00:11 -0700 Subject: [PATCH] [Bug] ngram verify: keep `batch.seq_lens_sum` in sync after accept (#26753) --- python/sglang/srt/speculative/ngram_info.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/python/sglang/srt/speculative/ngram_info.py b/python/sglang/srt/speculative/ngram_info.py index 3269bc2f5..28dba86e1 100644 --- a/python/sglang/srt/speculative/ngram_info.py +++ b/python/sglang/srt/speculative/ngram_info.py @@ -465,6 +465,8 @@ class NgramVerifyInput(SpecInput): batch.seq_lens.add_(self.num_accept_tokens) batch.seq_lens_cpu.add_(num_accept_tokens_cpu) + # Keep seq_lens_sum in sync; attn backends size kv_indices from it. + batch.seq_lens_sum += int(num_accept_tokens_cpu.sum()) return logits_output, self.accept_tokens, num_correct_drafts