minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Spec] Fix ngram metric off-by-1 in num_accepted_drafts_per_req_cpu (#24965)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-05-11 12:25:18 -07:00
committed by GitHub
parent 7407a62c1c
commit c7e53e68a2
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/speculative/ngram_worker.py
+3 -2
View File
@@ -328,8 +328,9 @@ class NGRAMWorker:
) )
req.time_stats.set_spec_verify_end_time(accepted_tokens=accepted) req.time_stats.set_spec_verify_end_time(accepted_tokens=accepted)
# Store accept_lens for per-request metrics # Store accept_lens (with bonus) for per-request metrics; downstream
accept_lens = verify_input.num_accepted_drafts # subtracts 1 to recover drafts-only counts.
accept_lens = verify_input.num_accepted_tokens
if batch.return_logprob: if batch.return_logprob:
add_output_logprobs_for_spec_v1(batch, verify_input, logits_output) add_output_logprobs_for_spec_v1(batch, verify_input, logits_output)
self._update_ngram_corpus(batch) self._update_ngram_corpus(batch)