[Bug] ngram verify: keep batch.seq_lens_sum in sync after accept (#26753)
This commit is contained in:
@@ -465,6 +465,8 @@ class NgramVerifyInput(SpecInput):
|
|||||||
|
|
||||||
batch.seq_lens.add_(self.num_accept_tokens)
|
batch.seq_lens.add_(self.num_accept_tokens)
|
||||||
batch.seq_lens_cpu.add_(num_accept_tokens_cpu)
|
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
|
return logits_output, self.accept_tokens, num_correct_drafts
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user