Fix harness SLO early-stop diagnosis
This commit is contained in:
@@ -607,6 +607,7 @@ def _active_bottleneck(probe: dict[str, Any] | None) -> str:
|
||||
if not str(k).startswith("ttft")
|
||||
and not str(k).startswith("tpot")
|
||||
and not str(k).startswith("probe_elapsed_s>")
|
||||
and str(k) != "slo_pass_rate_unrecoverable"
|
||||
)
|
||||
if ttft_count == 0 and tpot_count == 0 and request_failed_count == 0:
|
||||
return "none_obvious"
|
||||
|
||||
@@ -1084,6 +1084,97 @@ class CoreFlowTests(unittest.TestCase):
|
||||
{"max-num-seqs": 32},
|
||||
)
|
||||
|
||||
def test_slo_unrecoverable_does_not_mask_latency_bottleneck(self) -> None:
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
tmp_path = Path(tmp)
|
||||
study_path = _write_study_assets(
|
||||
tmp_path,
|
||||
slo_overrides={
|
||||
"ttft_rule": {"kind": "fixed_ms", "threshold_ms": 4000},
|
||||
"tpot_rule": {"kind": "fixed_ms", "threshold_ms": 25},
|
||||
},
|
||||
engine_overrides={
|
||||
"tunable_flags": [
|
||||
"tensor-parallel-size",
|
||||
"max-num-seqs",
|
||||
],
|
||||
"topology_constraints": {
|
||||
"allowed_tensor_parallel_sizes": [1, 2, 4],
|
||||
"allowed_tp_dp_products": [1, 2, 4],
|
||||
},
|
||||
},
|
||||
)
|
||||
result_path = tmp_path / "trial-0001.json"
|
||||
result_path.write_text(
|
||||
json.dumps(
|
||||
{
|
||||
"status": "completed",
|
||||
"best_request_rate": 0.065,
|
||||
"best_request_rate_per_gpu": 0.065,
|
||||
"best_pass_rate": 1.0,
|
||||
"probes": [
|
||||
{
|
||||
"threshold": 0.015625,
|
||||
"feasible": False,
|
||||
"payload": {
|
||||
"request_count": 290,
|
||||
"pass_rate": 0.041,
|
||||
"request_rate": 0.483,
|
||||
"early_stop_reason": "slo_pass_rate_unrecoverable",
|
||||
"latency_summary": {
|
||||
"failed_reason_counts": {
|
||||
"ttft_ms>4000.0": 2,
|
||||
"tpot_ms>25.0": 14,
|
||||
"slo_pass_rate_unrecoverable": 263,
|
||||
}
|
||||
},
|
||||
},
|
||||
},
|
||||
{
|
||||
"threshold": 0.001953125,
|
||||
"feasible": True,
|
||||
"payload": {
|
||||
"request_count": 39,
|
||||
"pass_rate": 1.0,
|
||||
"request_rate": 0.065,
|
||||
"latency_summary": {"failed_reason_counts": {}},
|
||||
},
|
||||
},
|
||||
],
|
||||
}
|
||||
),
|
||||
encoding="utf-8",
|
||||
)
|
||||
study = load_study_spec(study_path)
|
||||
context = build_harness_context(
|
||||
study=study,
|
||||
window_summary={"prompt_tokens_p95": 7628, "prompt_tail_ratio_p95_p50": 3.8},
|
||||
state=StudyState(
|
||||
study_id=study.study_id,
|
||||
best_trial_id="trial-0001",
|
||||
best_request_rate=0.065,
|
||||
best_request_rate_per_gpu=0.065,
|
||||
trials=[
|
||||
TrialSummary(
|
||||
trial_id="trial-0001",
|
||||
status="completed",
|
||||
best_request_rate=0.065,
|
||||
best_request_rate_per_gpu=0.065,
|
||||
best_pass_rate=1.0,
|
||||
result_path=str(result_path),
|
||||
config_patch={"env_patch": {}, "flag_patch": {}},
|
||||
)
|
||||
],
|
||||
),
|
||||
)
|
||||
self.assertNotEqual(
|
||||
context["bottleneck_hypotheses"][0]["name"],
|
||||
"admission_or_queueing",
|
||||
)
|
||||
proposal = build_harness_guided_proposal(context)
|
||||
self.assertIsNotNone(proposal)
|
||||
self.assertEqual(proposal.config_patch.flag_patch, {"tensor-parallel-size": 2})
|
||||
|
||||
def test_harness_stop_blocked_until_slo_driven_topology_frontier_is_measured(self) -> None:
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
tmp_path = Path(tmp)
|
||||
|
||||
Reference in New Issue
Block a user