Fix accept rate in speculative decoding metrics (#11572)

43f80884 · Qiaolin Yu · GitHub · 60b05032 · 43f80884
Unverified Commit 43f80884 authored Oct 13, 2025 by Qiaolin Yu Committed by GitHub Oct 13, 2025
Hide whitespace changes
Inline Side-by-side

Showing with 1 addition and 1 deletion

python/sglang/srt/managers/scheduler_metrics_mixin.py python/sglang/srt/managers/scheduler_metrics_mixin.py +1 -1

No files found.
--- a/python/sglang/srt/managers/scheduler_metrics_mixin.py
+++ b/python/sglang/srt/managers/scheduler_metrics_mixin.py
@@ -259,7 +259,7 @@ class SchedulerMetricsMixin:
            )
            # Calculate acceptance rate: accepted tokens / total draft tokens
            total_draft_tokens = self.spec_num_total_forward_ct * (
-                self.server_args.speculative_num_steps or 1
+                (self.server_args.speculative_num_steps or 0) + 1
            )
            spec_accept_rate = (
                self.spec_num_total_accepted_tokens / total_draft_tokens