[diffusion] nightly: keep nightly regression baselines on a comparable methodology (#42226)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
This commit is contained in:
Mick
2026-10-05 15:19:21 +08:00
committed by GitHub
co-authored by Mick Qian
parent 438d9d2074
commit dd2ffdb680
2 changed files with 104 additions and 1 deletions
@@ -296,6 +296,13 @@ def generate_dashboard(
Returns the markdown string.
"""
history_count = len(history)
history = [
run
for run in history
if run.get("methodology") == current.get("methodology")
and run.get("warmup_requests") == current.get("warmup_requests")
]
lines: list[str] = []
lines.append("# SGLang-Diffusion Nightly Performance Dashboard\n")
ts = current.get("timestamp", datetime.now(timezone.utc).isoformat())
@@ -311,6 +318,13 @@ def generate_dashboard(
"First request = the first request after the server reports ready. "
"Server perf dumps are telemetry only.*\n"
)
excluded_runs = history_count - len(history)
if excluded_runs:
lines.append(
f"*Excluded {excluded_runs} historical run(s) from baselines and trends "
"because their measurement methodology or warmup count differs. "
"Missing metadata is not treated as matching explicit metadata.*\n"
)
current_cases = _extract_case_results(current)
current_records = _extract_case_records(current)