Ignore non-SLO probe bookkeeping in bottleneck diagnosis
This commit is contained in:
@@ -341,6 +341,8 @@ def _active_bottleneck(probe: dict[str, Any] | None) -> str:
|
||||
and not str(k).startswith("tpot")
|
||||
and not str(k).startswith("probe_elapsed_s>")
|
||||
)
|
||||
if ttft_count == 0 and tpot_count == 0 and request_failed_count == 0:
|
||||
return "none_obvious"
|
||||
if ttft_count >= max(tpot_count, request_failed_count):
|
||||
return "ttft_prefill"
|
||||
if tpot_count >= max(ttft_count, request_failed_count):
|
||||
|
||||
Reference in New Issue
Block a user