[Bug] ngram verify: keep batch.seq_lens_sum in sync after accept (#26753)

This commit is contained in:
Liangsheng Yin
2026-05-29 18:00:11 -07:00
committed by GitHub
parent 95cd2fd29f
commit c2ac37dbcc
@@ -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