Add metrics for speculative decoding (acceptance rate, average acceptance length) (#11441)

2025-10-13 11:24:27 -07:00
parent f4aa78801e
commit b6fb5d7666
9 changed files with 70 additions and 1 deletions
--- a/python/sglang/srt/managers/schedule_batch.py
+++ b/python/sglang/srt/managers/schedule_batch.py
@@ -608,6 +608,10 @@ class Req:
        # This is used to compute the average acceptance length per request.
        self.spec_verify_ct = 0

+        # The number of accepted tokens in speculative decoding for this request.
+        # This is used to compute the acceptance rate and average acceptance length per request.
+        self.spec_accepted_tokens = 0
+
        # For metrics
        self.metrics_collector = metrics_collector
        self.time_stats: TimeStats = TimeStats(disagg_mode=disagg_mode)