Use normalized full config signatures

This commit is contained in:
2026-06-26 21:28:10 +08:00
parent 7f50b8b8ea
commit 48911b658b
3 changed files with 233 additions and 29 deletions

View File

@@ -25,6 +25,7 @@ from aituner.http_client import (
)
from aituner.job import append_job, build_trial_job
from aituner.harness import (
_effective_config_signature,
build_harness_context,
build_harness_guided_proposal,
build_harness_stop_proposal,
@@ -358,6 +359,36 @@ class CoreFlowTests(unittest.TestCase):
"search_high_lowered_to_trace_ceiling",
)
def test_effective_config_signature_treats_noop_patch_as_baseline(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
study_path = _write_study_assets(
Path(tmp),
engine_overrides={
"base_flags": {
"host": "127.0.0.1",
"port": 8000,
"tensor-parallel-size": 8,
"data-parallel-size": 1,
"gpu-memory-utilization": 0.5,
"max-num-seqs": 8,
},
},
)
study = load_study_spec(study_path)
baseline = _effective_config_signature(study, {"env_patch": {}, "flag_patch": {}})
noop_tp = _effective_config_signature(
study,
{"env_patch": {}, "flag_patch": {"tensor-parallel-size": 8}},
)
changed_tp = _effective_config_signature(
study,
{"env_patch": {}, "flag_patch": {"tensor-parallel-size": 4}},
)
self.assertEqual(baseline, noop_tp)
self.assertNotEqual(baseline, changed_tp)
def test_lca_workload_profile_uses_standard_10d_features(self) -> None:
window = WindowRecord(
window_id="w1",
@@ -1639,6 +1670,153 @@ class CoreFlowTests(unittest.TestCase):
"search_high_saturation_requires_parallel_size_evidence",
)
def test_harness_does_not_repropose_noop_topology_equivalent_to_baseline(
self,
) -> None:
with tempfile.TemporaryDirectory() as tmp:
tmp_path = Path(tmp)
study_path = _write_study_assets(
tmp_path,
engine_overrides={
"base_flags": {
"host": "127.0.0.1",
"port": 8000,
"tensor-parallel-size": 8,
"data-parallel-size": 1,
"gpu-memory-utilization": 0.5,
"max-num-seqs": 8,
},
"tunable_flags": ["tensor-parallel-size", "max-num-seqs"],
"topology_constraints": {
"allowed_tensor_parallel_sizes": [1, 2, 4, 8],
"allowed_tp_dp_products": [1, 2, 4, 8],
},
},
)
study = load_study_spec(study_path)
trial1_result = tmp_path / "trial-0001.json"
trial1_result.write_text(
json.dumps(
{
"status": "completed",
"best_sampling_u": 0.935616858887,
"best_request_rate": 8.0,
"best_pass_rate": 1.0,
"probes": [
{
"threshold": 0.935616858887,
"feasible": True,
"payload": {
"request_count": 480,
"pass_rate": 1.0,
"request_rate": 8.0,
"early_stopped": False,
"early_stop_reason": "",
"latency_summary": {"failed_reason_counts": {}},
},
}
],
}
),
encoding="utf-8",
)
trial2_result = tmp_path / "trial-0002.json"
trial2_result.write_text(
json.dumps(
{
"status": "completed",
"best_sampling_u": 0.810867944369,
"best_request_rate": 6.95,
"best_pass_rate": 0.9784,
"probes": [
{
"threshold": 0.873242401628,
"feasible": False,
"payload": {
"request_count": 450,
"pass_rate": 0.7844,
"request_rate": 7.5,
"early_stopped": True,
"early_stop_reason": "slo_pass_rate_unrecoverable",
"latency_summary": {
"failed_reason_counts": {
"ttft_ms>2000.0": 42,
"slo_pass_rate_unrecoverable": 49,
}
},
},
},
{
"threshold": 0.810867944369,
"feasible": True,
"payload": {
"request_count": 417,
"pass_rate": 0.9784,
"request_rate": 6.95,
"early_stopped": False,
"early_stop_reason": "",
"latency_summary": {
"failed_reason_counts": {"ttft_ms>2000.0": 9}
},
},
},
],
}
),
encoding="utf-8",
)
state = StudyState(
study_id=study.study_id,
best_trial_id="trial-0002",
best_parallel_size=4,
best_sampling_u=0.810867944369,
best_request_rate=6.95,
best_request_rate_per_gpu=1.7375,
next_trial_index=3,
trials=[
TrialSummary(
trial_id="trial-0001",
status="completed",
parallel_size=8,
best_request_rate=8.0,
best_request_rate_per_gpu=1.0,
result_path=str(trial1_result),
config_patch={"env_patch": {}, "flag_patch": {}},
),
TrialSummary(
trial_id="trial-0002",
status="completed",
parallel_size=4,
best_request_rate=6.95,
best_request_rate_per_gpu=1.7375,
result_path=str(trial2_result),
config_patch={
"env_patch": {},
"flag_patch": {"tensor-parallel-size": 4},
},
),
],
)
context = build_harness_context(
study=study,
window_summary={"prompt_tokens_p95": 2048},
state=state,
)
actions = context["experiment_plan"]["candidate_actions"]
self.assertFalse(
any(
action.get("config_patch", {}).get("flag_patch")
== {"tensor-parallel-size": 8}
for action in actions
)
)
proposal = build_harness_guided_proposal(context)
self.assertTrue(
proposal is None
or proposal.config_patch.flag_patch != {"tensor-parallel-size": 8}
)
def test_harness_guided_first_tp_probe_for_latency_bottleneck(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
tmp_path = Path(tmp)