You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: inference_perf/reportgen/base.py
+2-6Lines changed: 2 additions & 6 deletions
Original file line number
Diff line number
Diff line change
@@ -103,11 +103,7 @@ def calculate_slo_metrics(
103
103
ifttftisnotNone:
104
104
ttft_met= (ttft<=m.ttft_slo_sec)
105
105
else:
106
-
# If it's not streamable, we can't measure TTFT, so strictly speaking it's not "met"
107
-
# OR we treat non-streamable as failing the streaming SLO.
108
-
# Standard practice: if we can't measure it, it counts as failed or ignored.
109
-
# Here assuming failed if data is missing but SLO required.
110
-
ttft_met=False
106
+
ttft_met=True# If TTFT SLO is set but we don't have a TTFT value (non-streamable), we can choose to consider it as met or not. Here we consider it met to avoid penalizing non-streamable requests.
111
107
ttft_results.append(ttft_met)
112
108
113
109
# Check TPOT SLO (Only if streamable / tpot exists)
@@ -116,7 +112,7 @@ def calculate_slo_metrics(
116
112
iftpotisnotNone:
117
113
tpot_met= (tpot<=m.tpot_slo_sec)
118
114
else:
119
-
tpot_met=False
115
+
tpot_met=True# Similar logic as TTFT for non-streamable requests
0 commit comments