Tighten topology and auto-high validation
This commit is contained in:
@@ -1755,31 +1755,23 @@ def _parallel_size_can_vary(study: StudySpec) -> bool:
|
||||
effective_gpu_count = _effective_gpu_count(study)
|
||||
if effective_gpu_count <= 1:
|
||||
return False
|
||||
constraints = study.engine.topology_constraints
|
||||
if constraints is not None and constraints.allowed_tp_dp_products:
|
||||
legal_products = {
|
||||
item for item in constraints.allowed_tp_dp_products if item <= effective_gpu_count
|
||||
}
|
||||
return len(legal_products) > 1
|
||||
if constraints is not None:
|
||||
tp_values = (
|
||||
constraints.allowed_tensor_parallel_sizes
|
||||
if constraints.allowed_tensor_parallel_sizes
|
||||
else [1, 2, 4, 8]
|
||||
)
|
||||
dp_values = (
|
||||
constraints.allowed_data_parallel_sizes
|
||||
if constraints.allowed_data_parallel_sizes
|
||||
else [1]
|
||||
)
|
||||
products = {
|
||||
int(tp) * int(dp)
|
||||
for tp in tp_values
|
||||
for dp in dp_values
|
||||
if int(tp) > 0 and int(dp) > 0 and int(tp) * int(dp) <= effective_gpu_count
|
||||
}
|
||||
return len(products) > 1
|
||||
return True
|
||||
base = _normalized_topology_flags(study.engine.base_flags)
|
||||
legal = _legal_topology_points(
|
||||
study,
|
||||
current_tp=int(base["tensor-parallel-size"]),
|
||||
current_dp=int(base["data-parallel-size"]),
|
||||
current_ep=int(base["expert-parallel-size"]),
|
||||
current_enable_ep=bool(base["enable-expert-parallel"]),
|
||||
)
|
||||
signatures: set[str] = set()
|
||||
for point in legal:
|
||||
patch = _topology_patch(study, point)
|
||||
flags = {**study.engine.base_flags, **patch}
|
||||
normalized = _normalized_topology_flags(flags)
|
||||
if any(normalized.get(key) != point.get(key) for key in point):
|
||||
continue
|
||||
signatures.add(_config_signature({"env_patch": {}, "flag_patch": patch}))
|
||||
return len(signatures) > 1
|
||||
|
||||
|
||||
def _score_topology_candidate(
|
||||
|
||||
@@ -381,6 +381,9 @@ def resolve_auto_high_search(
|
||||
return search, evidence
|
||||
ceiling = min(float(policy.max_sampling_u), 1.0, float(trace_max_sampling_u))
|
||||
evidence["effective_ceiling"] = ceiling
|
||||
if ceiling < float(search.low):
|
||||
evidence["reason"] = "auto_high_ceiling_below_search_low"
|
||||
return search, evidence
|
||||
if abs(float(search.high) - ceiling) <= 1e-12:
|
||||
evidence["reason"] = "search_high_already_at_auto_high_ceiling"
|
||||
return search, evidence
|
||||
|
||||
@@ -309,6 +309,31 @@ class CoreFlowTests(unittest.TestCase):
|
||||
self.assertEqual(capped_by_trace.high, 0.7)
|
||||
self.assertEqual(trace_evidence["effective_ceiling"], 0.7)
|
||||
|
||||
low_above_ceiling = study.search.__class__.from_dict(
|
||||
{
|
||||
"low": 0.9,
|
||||
"high": 0.95,
|
||||
"tolerance": study.search.tolerance,
|
||||
"max_probes": study.search.max_probes,
|
||||
"sample_seed": study.search.sample_seed,
|
||||
"auto_high": {
|
||||
"enabled": True,
|
||||
"max_sampling_u": 0.8,
|
||||
"require_human_confirmation_beyond_trace": True,
|
||||
},
|
||||
}
|
||||
)
|
||||
unchanged, invalid_evidence = resolve_auto_high_search(
|
||||
search=low_above_ceiling,
|
||||
sampling_us=[0.1, 0.9],
|
||||
)
|
||||
self.assertEqual(unchanged.low, 0.9)
|
||||
self.assertEqual(unchanged.high, 0.95)
|
||||
self.assertEqual(
|
||||
invalid_evidence["reason"],
|
||||
"auto_high_ceiling_below_search_low",
|
||||
)
|
||||
|
||||
high_search = study.search.__class__.from_dict(
|
||||
{
|
||||
"low": 0.0,
|
||||
@@ -1531,6 +1556,89 @@ class CoreFlowTests(unittest.TestCase):
|
||||
"search_high_saturation_requires_parallel_size_evidence",
|
||||
)
|
||||
|
||||
def test_harness_stop_blocks_high_saturation_for_fixed_product_tp_dp_redistribution(
|
||||
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,
|
||||
},
|
||||
"tunable_flags": ["tensor-parallel-size", "data-parallel-size"],
|
||||
"topology_constraints": {
|
||||
"allowed_tensor_parallel_sizes": [1, 2, 4, 8],
|
||||
"allowed_data_parallel_sizes": [1, 2, 4, 8],
|
||||
"allowed_tp_dp_products": [8],
|
||||
"require_tp_dp_product_equals_gpu_count": True,
|
||||
},
|
||||
},
|
||||
)
|
||||
study = load_study_spec(study_path)
|
||||
result_path = tmp_path / "trial-0001.json"
|
||||
result_path.write_text(
|
||||
json.dumps(
|
||||
{
|
||||
"status": "completed",
|
||||
"best_sampling_u": 0.99609375,
|
||||
"best_request_rate": 8.0,
|
||||
"best_pass_rate": 1.0,
|
||||
"probes": [
|
||||
{
|
||||
"threshold": 0.99609375,
|
||||
"feasible": True,
|
||||
"payload": {
|
||||
"request_count": 10,
|
||||
"pass_rate": 1.0,
|
||||
"request_rate": 8.0,
|
||||
"early_stopped": False,
|
||||
"early_stop_reason": "",
|
||||
"latency_summary": {"failed_reason_counts": {}},
|
||||
},
|
||||
}
|
||||
],
|
||||
}
|
||||
),
|
||||
encoding="utf-8",
|
||||
)
|
||||
state = StudyState(
|
||||
study_id=study.study_id,
|
||||
best_trial_id="trial-0001",
|
||||
best_request_rate=8.0,
|
||||
best_request_rate_per_gpu=1.0,
|
||||
trials=[
|
||||
TrialSummary(
|
||||
trial_id="trial-0001",
|
||||
status="completed",
|
||||
best_request_rate=8.0,
|
||||
best_request_rate_per_gpu=1.0,
|
||||
result_path=str(result_path),
|
||||
config_patch={
|
||||
"env_patch": {},
|
||||
"flag_patch": {
|
||||
"tensor-parallel-size": 8,
|
||||
"data-parallel-size": 1,
|
||||
},
|
||||
},
|
||||
)
|
||||
],
|
||||
)
|
||||
context = build_harness_context(
|
||||
study=study,
|
||||
window_summary={"prompt_tokens_p95": 2048},
|
||||
state=state,
|
||||
)
|
||||
self.assertFalse(context["harness_stop"]["should_stop"])
|
||||
self.assertEqual(
|
||||
context["harness_stop"]["reason"],
|
||||
"search_high_saturation_requires_parallel_size_evidence",
|
||||
)
|
||||
|
||||
def test_harness_guided_first_tp_probe_for_latency_bottleneck(self) -> None:
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
tmp_path = Path(tmp)
|
||||
|
||||
Reference in New Issue
Block a user