Stop harness when feasible probe reaches search high

This commit is contained in:
2026-05-06 17:59:09 +08:00
parent 5d96689ea6
commit f653af09a8
2 changed files with 64 additions and 14 deletions

View File

@@ -600,6 +600,68 @@ class CoreFlowTests(unittest.TestCase):
self.assertIsNotNone(proposal)
self.assertTrue(proposal.should_stop)
def test_harness_stop_allows_feasible_high_probe_with_some_failures(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
tmp_path = Path(tmp)
study_path = _write_study_assets(tmp_path)
study = load_study_spec(study_path)
result_path = tmp_path / "trial-0004.json"
result_path.write_text(
json.dumps(
{
"status": "completed",
"best_sampling_u": 0.99609375,
"best_request_rate": 1.77,
"best_pass_rate": 0.968,
"probes": [
{
"threshold": 0.99609375,
"feasible": True,
"payload": {
"request_count": 1063,
"pass_rate": 0.968,
"request_rate": 1.77,
"early_stopped": False,
"early_stop_reason": "",
"latency_summary": {
"failed_reason_counts": {
"tpot_ms>50.0": 34,
}
},
},
}
],
}
),
encoding="utf-8",
)
state = StudyState(
study_id=study.study_id,
best_trial_id="trial-0004",
best_request_rate=1.77,
best_request_rate_per_gpu=0.4425,
trials=[
TrialSummary(
trial_id="trial-0004",
status="completed",
best_request_rate=1.77,
best_request_rate_per_gpu=0.4425,
result_path=str(result_path),
config_patch={
"env_patch": {},
"flag_patch": {"tensor-parallel-size": 4},
},
)
],
)
context = build_harness_context(
study=study,
window_summary={"prompt_tokens_p95": 2048},
state=state,
)
self.assertTrue(context["harness_stop"]["should_stop"])
self.assertEqual(context["harness_stop"]["reason"], "search_high_saturated_by_incumbent")
def test_harness_guided_first_tp_probe_for_latency_bottleneck(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
tmp_path = Path(tmp)