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

Add profile-driven harness planner

Browse Source
This commit is contained in:
Gahow Wang
2026-05-12 21:28:44 +08:00
parent 63d6a111f4
commit 17e9681ca0
3 changed files with 976 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

164
tests/test_core_flow.py
View File

@@ -920,6 +920,170 @@ class CoreFlowTests(unittest.TestCase):
"topology_frontier_probe_for_slo_pressure",
)
def test_profile_driven_planner_scores_unmeasured_tp_frontier(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
tmp_path = Path(tmp)
study_path = _write_study_assets(
tmp_path,
engine_overrides={
"tunable_flags": [
"tensor-parallel-size",
"max-num-batched-tokens",
"enable-chunked-prefill",
],
"topology_constraints": {
"allowed_tensor_parallel_sizes": [1, 2, 4],
"allowed_tp_dp_products": [1, 2, 4],
},
},
)
result_path = tmp_path / "trial-0002.json"
result_path.write_text(
json.dumps(
{
"status": "completed",
"best_sampling_u": 0.5,
"best_request_rate": 2.0,
"best_pass_rate": 0.96,
"probes": [
{
"threshold": 0.75,
"feasible": False,
"payload": {
"request_count": 100,
"pass_rate": 0.6,
"request_rate": 3.0,
"early_stop_reason": "slo_pass_rate_unrecoverable",
"latency_summary": {
"failed_reason_counts": {"ttft_ms>4000.0": 35}
},
},
}
],
}
),
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-0002",
best_request_rate=2.0,
best_request_rate_per_gpu=1.0,
trials=[
TrialSummary(
trial_id="trial-0001",
status="completed",
best_request_rate=0.5,
best_request_rate_per_gpu=0.5,
config_patch={"env_patch": {}, "flag_patch": {}},
),
TrialSummary(
trial_id="trial-0002",
status="completed",
best_request_rate=2.0,
best_request_rate_per_gpu=1.0,
result_path=str(result_path),
config_patch={
"env_patch": {},
"flag_patch": {"tensor-parallel-size": 2},
},
),
],
),
)
plan = context["experiment_plan"]
self.assertEqual(plan["planner_version"], "profile-driven-v1")
self.assertEqual(plan["next_action"]["knob_family"], "topology")
self.assertEqual(
plan["next_action"]["config_patch"]["flag_patch"],
{"tensor-parallel-size": 4},
)
self.assertIn("ttft_prefill", context["bottleneck_hypotheses"][0]["name"])
self.assertFalse(context["harness_stop"]["should_stop"])
def test_profile_driven_planner_prefers_decode_concurrency_relief(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
tmp_path = Path(tmp)
study_path = _write_study_assets(
tmp_path,
trace_overrides={"request_mode": "decode_only"},
slo_overrides={
"ttft_rule": None,
"tpot_rule": {"kind": "fixed_ms", "threshold_ms": 20},
},
engine_overrides={
"base_flags": {
"host": "127.0.0.1",
"port": 8000,
"tensor-parallel-size": 4,
"max-num-seqs": 64,
},
"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_sampling_u": 0.25,
"best_request_rate": 1.0,
"best_pass_rate": 0.97,
"probes": [
{
"threshold": 0.5,
"feasible": False,
"payload": {
"request_count": 100,
"pass_rate": 0.5,
"request_rate": 2.0,
"early_stop_reason": "slo_pass_rate_unrecoverable",
"latency_summary": {
"failed_reason_counts": {"tpot_ms>20.0": 50}
},
},
}
],
}
),
encoding="utf-8",
)
study = load_study_spec(study_path)
context = build_harness_context(
study=study,
window_summary={},
state=StudyState(
study_id=study.study_id,
best_trial_id="trial-0001",
best_request_rate=1.0,
best_request_rate_per_gpu=0.25,
trials=[
TrialSummary(
trial_id="trial-0001",
status="completed",
best_request_rate=1.0,
best_request_rate_per_gpu=0.25,
result_path=str(result_path),
config_patch={"env_patch": {}, "flag_patch": {}},
)
],
),
)
plan = context["experiment_plan"]
self.assertEqual(plan["next_action"]["knob_family"], "max-num-seqs")
self.assertEqual(
plan["next_action"]["config_patch"]["flag_patch"],
{"max-num-seqs": 32},
)
def test_harness_stop_blocked_until_slo_driven_topology_frontier_is_measured(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
tmp_path = Path(tmp)