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

Add Stop-A: offered-L-C-A convergence early-stop for replay

Browse Source 
Phase 2 of the two-stop work. The L-C-A vector is a deterministic function of the
trace's offered metadata, so the convergence of prefix-vs-full L-C-A (the paper's
Fig. 9 curve) can be computed up front rather than monitored live, with identical
result and no per-request overhead.

- lca.find_convergence_prefix: earliest arrival-ordered prefix whose L and A family
  similarities reach tau and the slow C family reaches the stricter tau_c for
  stable_checks consecutive checkpoints. Self-similarity uses the raw log-feature
  vector (same window -> identical per-dim spread; RobustScaler is reserved for the
  cross-window Stop-C). If C never converges it reports the full set, which is the
  C-gate: no early stop on a cold/under-warmed cache. The checkpoint sims double as
  Phase 3 calibration data.
- spec.AdaptiveStopSpec (trace.adaptive_stop), disabled by default until the
  thresholds are calibrated, so existing studies are unaffected.
- worker._adaptive_replay_set truncates each probe's replay to the convergence
  prefix and records a certificate (converged, fraction, family similarity) into
  probe history and probe_details. Offered request_rate at the threshold is
  unchanged; only wall-clock replay shrinks.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Gahow Wang
2026-06-15 14:23:49 +08:00
parent 0f15bbc3f1
commit 51a9e4a007
4 changed files with 379 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

131
tests/test_core_flow.py
View File

@@ -30,6 +30,7 @@ from aituner.harness import (
from aituner.lca import (
build_study_workload_profile,
build_workload_profile,
find_convergence_prefix,
profile_similarity,
resolve_length_mode,
similarity_report,
@@ -38,6 +39,7 @@ from aituner.llm import _extract_response_text, build_prompt, parse_proposal_tex
from aituner.search import ThresholdProbe, binary_search_max_feasible
from aituner.slo import RequestOutcome, evaluate_request, summarize_evaluations
from aituner.spec import (
AdaptiveStopSpec,
ConfigPatch,
LLMEndpointSpec,
Proposal,
@@ -49,6 +51,7 @@ from aituner.spec import (
from aituner.store import StudyStore
from aituner.trace import load_trace_requests, summarize_window
from aituner.worker import (
_adaptive_replay_set,
_best_feasible_probe_record,
_latency_summary,
_run_one_request,
@@ -327,6 +330,134 @@ class CoreFlowTests(unittest.TestCase):
)["workload_lca_profile"]
self.assertNotIn("vector", legacy)
def _steady_requests(self, count: int, *, input_tokens: int = 100) -> list:
return [
TraceRequest(
row_id=f"r{i}",
arrival_s=float(i),
sampling_u=1.0,
body={},
prompt_tokens_hint=input_tokens,
completion_tokens_hint=16,
metadata={"hash_ids": None},
)
for i in range(count)
]
def _conv_window(self) -> WindowRecord:
return WindowRecord(
window_id="conv",
trace_path=Path("trace.jsonl"),
trace_type="chat",
window_start=0.0,
window_end=0.0,
source_payload={"block_size": 64},
)
def test_convergence_prefix_stops_early_on_stationary_trace(self) -> None:
requests = self._steady_requests(60)
point = find_convergence_prefix(
requests,
self._conv_window(),
gpu_count=1,
length_mode="total",
tau=0.9,
tau_c=0.9,
stable_checks=3,
max_checks=20,
min_fraction=0.1,
)
self.assertTrue(point.converged)
# A stationary workload should be trustworthy well before the full window.
self.assertLess(point.stop_index, len(requests))
self.assertLess(point.fraction, 1.0)
self.assertTrue(point.checks)
def test_convergence_prefix_waits_when_cache_warms_late(self) -> None:
window = self._conv_window()
# First half: no prefix reuse. Second half: every request reuses block 1,
# so the C dimension only stabilizes once the reuse regime is exercised.
requests = []
for i in range(30):
requests.append(
TraceRequest(
row_id=f"cold{i}",
arrival_s=float(i),
sampling_u=1.0,
body={},
prompt_tokens_hint=640,
completion_tokens_hint=16,
metadata={"hash_ids": [10_000 + i]},
)
)
for i in range(30):
requests.append(
TraceRequest(
row_id=f"warm{i}",
arrival_s=float(30 + i),
sampling_u=1.0,
body={},
prompt_tokens_hint=640,
completion_tokens_hint=16,
metadata={"hash_ids": [1, 2, 3, 4, 5]},
)
)
point = find_convergence_prefix(
requests,
window,
gpu_count=1,
length_mode="total",
tau=0.9,
tau_c=0.95,
stable_checks=2,
max_checks=20,
min_fraction=0.1,
)
# The C family similarity must be low while only the cold half is seen.
early = [c for c in point.checks if c["fraction"] <= 0.4]
self.assertTrue(early)
self.assertTrue(any(c["family_similarity"]["C"] < 0.9 for c in early))
def test_adaptive_replay_set_truncates_only_when_enabled(self) -> None:
from types import SimpleNamespace
requests = self._steady_requests(60)
window = self._conv_window()
enabled_study = SimpleNamespace(
trace=SimpleNamespace(
adaptive_stop=AdaptiveStopSpec(
enabled=True,
tau=0.9,
tau_c=0.9,
stable_checks=3,
max_checks=20,
min_fraction=0.1,
),
request_mode="chat",
),
hardware=SimpleNamespace(gpu_count=1),
)
replay, certificate = _adaptive_replay_set(
requests, study=enabled_study, window=window
)
self.assertIsNotNone(certificate)
self.assertTrue(certificate["enabled"])
self.assertEqual(len(replay), certificate["stop_index"])
self.assertLessEqual(len(replay), len(requests))
disabled_study = SimpleNamespace(
trace=SimpleNamespace(
adaptive_stop=AdaptiveStopSpec(enabled=False),
request_mode="chat",
),
hardware=SimpleNamespace(gpu_count=1),
)
passthrough, no_cert = _adaptive_replay_set(
requests, study=disabled_study, window=window
)
self.assertIsNone(no_cert)
self.assertEqual(len(passthrough), len(requests))
def test_lca_similarity_matrix_separates_different_profiles(self) -> None:
window = WindowRecord(
window_id="base",