gahow/aituner
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Continue gmu hill-climb after topology validation

Browse Source
This commit is contained in:
Gahow Wang
2026-06-24 19:09:35 +08:00
parent 8fa758797e
commit b075afe6f2
2 changed files with 187 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

124
tests/test_core_flow.py
View File

@@ -1800,6 +1800,130 @@ class CoreFlowTests(unittest.TestCase):
)
self.assertNotIn("gpu-memory-utilization", proposal.config_patch.flag_patch)
def test_harness_continues_gpu_mem_util_after_tied_same_topology_probe(self) -> None:
"""After adjacent topology validation, gpu-memory-utilization should hill-climb
on the incumbent topology even if an earlier gmu step tied the incumbent and
did not become state.best_trial_id."""
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": 50},
},
engine_overrides={
"tunable_flags": [
"tensor-parallel-size",
"data-parallel-size",
"gpu-memory-utilization",
],
"topology_constraints": {
"allowed_tensor_parallel_sizes": [1, 2, 4, 8],
"allowed_data_parallel_sizes": [1, 2],
"allowed_tp_dp_products": [1, 2, 4, 8],
},
},
)
study = load_study_spec(study_path)
result_path = tmp_path / "trial-0002.json"
result_path.write_text(
json.dumps(
{
"status": "completed",
"best_sampling_u": 0.75,
"best_request_rate": 6.5,
"best_pass_rate": 1.0,
"probes": [
{
"threshold": 0.75,
"feasible": True,
"payload": {
"request_count": 300,
"pass_rate": 1.0,
"request_rate": 6.5,
"latency_summary": {"failed_reason_counts": {}},
},
},
{
"threshold": 0.765625,
"feasible": False,
"payload": {
"request_count": 300,
"pass_rate": 0.6,
"request_rate": 6.7,
"early_stop_reason": "slo_pass_rate_unrecoverable",
"latency_summary": {
"failed_reason_counts": {"ttft_ms>4000.0": 80}
},
},
},
],
}
),
encoding="utf-8",
)
state = StudyState(
study_id=study.study_id,
best_trial_id="trial-0002",
best_request_rate=6.5,
best_request_rate_per_gpu=3.25,
trials=[
TrialSummary(
trial_id="trial-0001",
status="completed",
best_request_rate=2.2,
best_request_rate_per_gpu=2.2,
config_patch={"env_patch": {}, "flag_patch": {}},
),
TrialSummary(
trial_id="trial-0002",
status="completed",
best_request_rate=6.5,
best_request_rate_per_gpu=3.25,
result_path=str(result_path),
config_patch={
"env_patch": {},
"flag_patch": {"tensor-parallel-size": 2},
},
),
TrialSummary(
trial_id="trial-0003",
status="completed",
best_request_rate=8.4,
best_request_rate_per_gpu=2.1,
config_patch={
"env_patch": {},
"flag_patch": {"tensor-parallel-size": 4},
},
),
TrialSummary(
trial_id="trial-0004",
status="completed",
best_request_rate=6.5,
best_request_rate_per_gpu=3.25,
config_patch={
"env_patch": {},
"flag_patch": {
"tensor-parallel-size": 2,
"gpu-memory-utilization": 0.92,
},
},
),
],
)
context = build_harness_context(
study=study,
window_summary={"prompt_tokens_p95": 1500},
state=state,
)
proposal = build_harness_guided_proposal(context)
self.assertIsNotNone(proposal)
self.assertEqual(
proposal.config_patch.flag_patch,
{"tensor-parallel-size": 2, "gpu-memory-utilization": 0.94},
)
def test_harness_validates_unmeasured_tp_frontier_before_runtime_refinement(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
tmp_path = Path(tmp)