minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Fix accept rate in speculative decoding metrics (#13212)

Browse Source
This commit is contained in:
neo
2025-11-14 00:54:50 +08:00
committed by GitHub
parent 67e9d287ee
commit c8b7516fe8
1 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/scheduler_metrics_mixin.py
+5 -2
View File
@@ -274,9 +274,12 @@ class SchedulerMetricsMixin:
self.spec_num_accepted_tokens / self.spec_num_forward_ct self.spec_num_accepted_tokens / self.spec_num_forward_ct
) )
# Calculate acceptance rate: accepted tokens / total draft tokens # Calculate acceptance rate: accepted tokens / total draft tokens
total_draft_tokens = self.spec_num_forward_ct * ( draft_tokens_fallback = (self.server_args.speculative_num_steps or 0) + 1
(self.server_args.speculative_num_steps or 0) + 1 num_draft_tokens = (
self.server_args.speculative_num_draft_tokens or draft_tokens_fallback
) )
total_draft_tokens = self.spec_num_forward_ct * num_draft_tokens
spec_accept_rate = ( spec_accept_rate = (
self.spec_num_accepted_tokens / total_draft_tokens self.spec_num_accepted_tokens / total_draft_tokens
if total_draft_tokens > 0 if total_draft_tokens > 0