Files
aituner/runs/intervention-response-v2/analyze_pilot.py

652 lines
25 KiB
Python

#!/usr/bin/env python3
"""Analyze the uncensored 300-second phase-aware matched pilot."""
from __future__ import annotations
import argparse
import hashlib
import importlib.util
import json
import math
from collections import defaultdict
from pathlib import Path
from typing import Any, Iterable, Mapping
HERE = Path(__file__).resolve().parent
P1_PATH = HERE.parent / "intervention-response-v0" / "analyze_p1.py"
SCHEMA = "intervention-response-phase-aware-pilot-analysis-v2"
EXPECTED_ACTION_PAIRS = 9
EXPECTED_REPEAT_PAIRS = 12
MIN_EFFICACY_CLASS = 3
MAX_LAYER1_GAP_S = 1.0
def _load_p1():
spec = importlib.util.spec_from_file_location("intervention_response_p1", P1_PATH)
module = importlib.util.module_from_spec(spec)
assert spec.loader is not None
spec.loader.exec_module(module)
return module
P1 = _load_p1()
def sha256_file(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as source:
for chunk in iter(lambda: source.read(1 << 20), b""):
digest.update(chunk)
return digest.hexdigest()
def numeric(values: Iterable[float | int]) -> dict[str, Any]:
return P1.V0.numeric(values)
def _trial_record(
*,
run_root: Path,
session: Mapping[str, Any],
level: str,
result: Mapping[str, Any],
result_path: Path,
requests_path: Path,
state: dict[str, float],
outcome: dict[str, float],
telemetry_coverage: dict[str, float],
) -> dict[str, Any]:
return {
"trial_id": str(result_path.relative_to(run_root)),
"cell": str(result["cell"]),
"tp": int(result["tp"]),
"mns": int(result["mns"]),
"level": level,
"replicate": int(session["replicate"]),
"offered_rate_per_gpu": float(
result["selection"]["offered_req_s_per_gpu"]
),
"request_hash": str(result["selection"]["request_id_order_sha256"]),
"request_count": int(result["selection"]["count"]),
"result_sha256": sha256_file(result_path),
"requests_sha256": sha256_file(requests_path),
"full_pass_rate": float(result["pass_rate"]),
"full_feasible": bool(result["feasible"]),
"early_stopped": bool(result["early_stopped"]),
"state": state,
"outcome": outcome,
"telemetry_coverage": telemetry_coverage,
}
def telemetry_coverage(
records: list[dict[str, Any]], *, start_ns: int, end_ns: int
) -> dict[str, float]:
layer1 = [record for record in records if "step_index" in record]
timestamps = [int(record["submit_mono_ns"]) for record in layer1]
if timestamps != sorted(timestamps):
raise ValueError("Layer-1 timestamps are not monotonic")
selected = [timestamp for timestamp in timestamps if start_ns <= timestamp <= end_ns]
if not selected:
raise ValueError("Layer-1 coverage interval contains no records")
internal_gaps = [
(right - left) / 1e9
for left, right in zip(selected, selected[1:], strict=False)
]
coverage = {
"start_gap_s": (selected[0] - start_ns) / 1e9,
"end_gap_s": (end_ns - selected[-1]) / 1e9,
"max_internal_gap_s": max(internal_gaps, default=0.0),
}
if any(value > MAX_LAYER1_GAP_S for value in coverage.values()):
raise ValueError(f"Layer-1 coverage gap exceeds {MAX_LAYER1_GAP_S}s: {coverage}")
return coverage
def validate_result_against_manifest(
*,
result: Mapping[str, Any],
selection: Mapping[str, Any],
session: Mapping[str, Any],
level: str,
expected_duration_s: float,
) -> None:
identity = f"{session['session']}:{level}"
if int(result["mns"]) != int(session["mns"]) or int(result["tp"]) != 4:
raise ValueError(f"config mismatch: {identity}")
if bool(result["early_stopped"]):
raise ValueError(f"early-stopped measured result: {identity}")
if result.get("slo_early_stop_disabled") is not True:
raise ValueError(f"SLO early stop was enabled: {identity}")
if float(result["interval"]["elapsed_s"]) + 1e-9 < expected_duration_s:
raise ValueError(f"result does not cover full arrival window: {identity}")
if int(result["selection"]["count"]) != int(selection["selected_count"]):
raise ValueError(f"selection count mismatch: {identity}")
for result_key, manifest_key in (
("request_id_order_sha256", "request_id_order_sha256"),
("arrival_order_sha256", "arrival_order_sha256"),
("raw_length_order_sha256", "input_length_order_sha256"),
):
if result["selection"][result_key] != selection[manifest_key]:
raise ValueError(f"selection hash mismatch {result_key}: {identity}")
if int(result["observed_count"]) != int(selection["selected_count"]):
raise ValueError(f"request accounting mismatch: {identity}")
def load_interval_trials(
*,
run_root: Path,
manifest: Mapping[str, Any],
intervals_s: tuple[tuple[float, float], ...],
) -> tuple[dict[tuple[float, float], list[dict[str, Any]]], list[dict[str, Any]]]:
by_interval = {interval: [] for interval in intervals_s}
streams = []
duration_s = float(manifest["engine"]["duration_s"])
for session in manifest["sessions"]:
session_root = run_root / "sessions" / str(session["session"])
stream_paths = sorted((session_root / "opprof").glob("*.jsonl"))
if len(stream_paths) != 1:
raise ValueError(f"{session_root}: expected one Layer-1 stream")
stream_path = stream_paths[0]
stream = P1.load_jsonl(stream_path)
streams.append(
{
"session": str(session["session"]),
"path": str(stream_path.resolve()),
"sha256": sha256_file(stream_path),
"bytes": stream_path.stat().st_size,
}
)
repetition = manifest["repetitions"][str(session["replicate"])]
for level, selection in repetition["selections"].items():
result_path = session_root / level / "result.json"
requests_path = session_root / level / "requests.jsonl"
result = json.loads(result_path.read_text(encoding="utf-8"))
requests = P1.load_jsonl(requests_path)
validate_result_against_manifest(
result=result,
selection=selection,
session=session,
level=level,
expected_duration_s=duration_s,
)
start_ns = int(result["interval"]["start_mono_ns"])
for interval in intervals_s:
start_s, end_s = interval
interval_start_ns = start_ns + int(start_s * 1e9)
interval_end_ns = start_ns + int(end_s * 1e9)
coverage = telemetry_coverage(
stream, start_ns=interval_start_ns, end_ns=interval_end_ns
)
state = P1.V0.flatten_state(
P1.summarize_engine(
stream,
start_ns=interval_start_ns,
end_ns=interval_end_ns,
request_count=int(result["selection"]["count"]),
)
)
outcome = P1._prefix_outcome(result, requests, end_s)
by_interval[interval].append(
_trial_record(
run_root=run_root,
session=session,
level=level,
result=result,
result_path=result_path,
requests_path=requests_path,
state=state,
outcome=outcome,
telemetry_coverage=coverage,
)
)
return by_interval, streams
def analyze_window(
trials: list[dict[str, Any]], *, start_s: float, end_s: float, fraction: float
) -> dict[str, Any]:
actions, repeats = P1.build_pairs(trials)
response = P1.V0.response_statistics(actions, repeats)
response_qualifying = sorted(
feature for feature, item in response.items() if item["qualifies"]
)
labels = [int(pair["full_action_efficacy"]) for pair in actions]
cross_validation_possible = all(
set(
int(pair["full_action_efficacy"])
for pair in actions
if pair["group"]["replicate"] != held_out
)
== {0, 1}
for held_out in (1, 2, 3)
)
if cross_validation_possible:
outcome_cv = P1.one_feature_leave_repeat_out(
actions, delta_key="delta_outcome", features=P1.OUTCOME_FEATURES
)
telemetry_cv = P1.one_feature_leave_repeat_out(
actions, delta_key="delta_state", features=P1.V0.GATE_FEATURES
)
outcome_best = float(outcome_cv["best_balanced_accuracy"])
efficacy_qualifying = sorted(
feature
for feature, item in telemetry_cv["features"].items()
if item["balanced_accuracy"] >= P1.MIN_EFFICACY_BALANCED_ACCURACY
and item["balanced_accuracy"]
>= outcome_best + P1.MIN_EFFICACY_DELTA_OVER_OUTCOME
)
else:
unavailable = {
"status": "UNAVAILABLE",
"reason": "each leave-one-repetition-out train fold needs both classes",
}
outcome_cv = unavailable
telemetry_cv = unavailable
efficacy_qualifying = []
transitions = defaultdict(int)
for pair in actions:
transitions[pair["full_feasibility_transition"]] += 1
admitted = [float(trial["outcome"]["admitted_fraction"]) for trial in trials]
completed = [
float(trial["outcome"]["admitted_fraction"])
* float(trial["outcome"]["completed_over_admitted"])
for trial in trials
]
state_vectors = {
tuple(round(float(trial["state"][feature]), 12) for feature in P1.V0.ALL_FEATURES)
for trial in trials
}
per_cell_vectors: dict[str, set[tuple[float, ...]]] = defaultdict(set)
for trial in trials:
per_cell_vectors[str(trial["cell"])].add(
tuple(
round(float(trial["state"][feature]), 12)
for feature in P1.V0.ALL_FEATURES
)
)
ratio_features = (
"prefill_token_fraction",
"kv_usage_mean",
"kv_usage_max",
"graph_none_share",
"graph_full_share",
"graph_padding_fraction",
)
nonnegative_features = tuple(
feature
for feature in P1.V0.ALL_FEATURES
if feature != "kv_usage_end_minus_start"
)
action_metadata = [
{
"group": pair["group"],
"source": pair["source"],
"target": pair["target"],
"full_action_efficacy": pair["full_action_efficacy"],
"full_feasibility_transition": pair["full_feasibility_transition"],
"delta_state": pair["delta_state"],
"delta_outcome": pair["delta_outcome"],
}
for pair in actions
]
invariants = {
"expected_action_pair_count": len(actions) == EXPECTED_ACTION_PAIRS,
"expected_repeat_pair_count": len(repeats) == EXPECTED_REPEAT_PAIRS,
"finite_deltas": all(
math.isfinite(value)
for pair in [*actions, *repeats]
for key in ("delta_state", "delta_outcome")
for value in pair[key].values()
),
"all_results_uncensored": all(not trial["early_stopped"] for trial in trials),
"state_vectors_not_all_identical": len(state_vectors) > 1,
"per_cell_state_vectors_not_all_identical": all(
len(vectors) > 1 for vectors in per_cell_vectors.values()
),
"ratios_bounded": all(
0.0 <= float(trial["state"][feature]) <= 1.0
for trial in trials
for feature in ratio_features
),
"nonnegative_counters": all(
float(trial["state"][feature]) >= 0.0
for trial in trials
for feature in nonnegative_features
),
"layer1_boundary_and_internal_gaps_bounded": all(
value <= MAX_LAYER1_GAP_S
for trial in trials
for value in trial["telemetry_coverage"].values()
),
}
return {
"start_s": start_s,
"end_s": end_s,
"end_fraction": fraction,
"coverage_at_end": {
"admitted_fraction_of_total": numeric(admitted),
"completed_fraction_of_total": numeric(completed),
},
"action_pairs": len(actions),
"repeat_pairs": len(repeats),
"actions": action_metadata,
"trial_sanity": [
{
"trial_id": trial["trial_id"],
"cell": trial["cell"],
"level": trial["level"],
"replicate": trial["replicate"],
"admitted_fraction": trial["outcome"]["admitted_fraction"],
"completed_fraction": (
trial["outcome"]["admitted_fraction"]
* trial["outcome"]["completed_over_admitted"]
),
"telemetry_coverage": trial["telemetry_coverage"],
}
for trial in trials
],
"response_statistics": response,
"qualifying_response_features": response_qualifying,
"efficacy": {
"labels": numeric(labels),
"positive": sum(labels),
"negative": len(labels) - sum(labels),
"label_balance_sufficient": (
sum(labels) >= MIN_EFFICACY_CLASS
and len(labels) - sum(labels) >= MIN_EFFICACY_CLASS
),
"cross_validation_possible": cross_validation_possible,
"transitions": dict(sorted(transitions.items())),
"outcome_delta": outcome_cv,
"telemetry_delta": telemetry_cv,
"telemetry_qualifying_features": efficacy_qualifying,
},
"sanity": {
"trials": len(trials),
"invariants": invariants,
"red_flags": [name for name, passed in invariants.items() if not passed],
},
}
def stable_adjacent_features(windows: list[dict[str, Any]]) -> dict[str, list[str]]:
result = {}
for left, right in zip(windows, windows[1:], strict=False):
key = f"{left['end_fraction']:.2f}->{right['end_fraction']:.2f}"
result[key] = sorted(
set(left["qualifying_response_features"])
& set(right["qualifying_response_features"])
)
return result
def stable_adjacent_efficacy_features(
windows: list[dict[str, Any]],
) -> dict[str, list[str]]:
eligible = [window for window in windows if window["end_fraction"] >= 0.25]
result = {}
for left, right in zip(eligible, eligible[1:], strict=False):
key = f"{left['end_fraction']:.2f}->{right['end_fraction']:.2f}"
result[key] = sorted(
set(left["efficacy"]["telemetry_qualifying_features"])
& set(right["efficacy"]["telemetry_qualifying_features"])
)
return result
def consistent_load_regimes(
windows: list[dict[str, Any]], stable: dict[str, list[str]]
) -> dict[str, Any]:
by_end = {float(window["end_fraction"]): window for window in windows}
result = {}
for transition, features in stable.items():
_left_text, right_text = transition.split("->")
window = by_end[float(right_text)]
for feature in features:
deltas_by_level: dict[str, list[float]] = defaultdict(list)
all_deltas = []
for action in window["actions"]:
value = float(action["delta_state"][feature])
deltas_by_level[str(action["group"]["level"])].append(value)
all_deltas.append(value)
positive = sum(value > 1e-12 for value in all_deltas)
negative = sum(value < -1e-12 for value in all_deltas)
direction = 1 if positive >= negative else -1
consistent = []
for level, values in sorted(deltas_by_level.items()):
matching = sum(direction * value > 1e-12 for value in values)
nonzero = sum(abs(value) > 1e-12 for value in values)
if nonzero and matching / nonzero >= 2.0 / 3.0:
consistent.append(level)
result[f"{transition}:{feature}"] = {
"direction": direction,
"consistent_load_regimes": consistent,
"passes_two_regimes": len(consistent) >= 2,
}
return result
def mechanism_gate(
stable: Mapping[str, list[str]], load_consistency: Mapping[str, Mapping[str, Any]]
) -> dict[str, Any]:
by_transition = {}
for transition, features in stable.items():
qualifying = sorted(
feature
for feature in features
if load_consistency[f"{transition}:{feature}"]["passes_two_regimes"]
)
by_transition[transition] = qualifying
passing_transitions = sorted(
transition
for transition, features in by_transition.items()
if len(features) >= 2
)
return {
"minimum_features": 2,
"by_transition": by_transition,
"passing_transitions": passing_transitions,
"passes": bool(passing_transitions),
}
def controller_gate(
run_root: Path, manifest: Mapping[str, Any]
) -> dict[str, Any]:
path = run_root / "controller-state.json"
state = json.loads(path.read_text(encoding="utf-8"))
expected_sessions = {str(session["session"]) for session in manifest["sessions"]}
actual_sessions = set(state.get("sessions", {}))
session_invariants = [
passed
for session in state.get("sessions", {}).values()
for passed in session.get("validation", {}).get("invariants", {}).values()
]
invariants = {
"controller_complete": state.get("status") == "complete",
"completed_session_count": int(state.get("completed_sessions", -1))
== len(expected_sessions),
"exact_session_set": actual_sessions == expected_sessions,
"all_sessions_complete": all(
session.get("status") == "complete"
for session in state.get("sessions", {}).values()
),
"no_controller_failures": not state.get("failures"),
"under_h20_hour_cap": float(state.get("gpu_hours_total", math.inf))
<= float(manifest["budget"]["hard_cap_h20_hours"]),
"all_stream_validation_invariants_pass": bool(session_invariants)
and all(session_invariants),
}
return {
"path": str(path.resolve()),
"sha256": sha256_file(path),
"gpu_hours_total": float(state.get("gpu_hours_total", math.nan)),
"invariants": invariants,
"red_flags": [name for name, passed in invariants.items() if not passed],
}
def cumulative_coverage_gate(windows: list[dict[str, Any]]) -> dict[str, Any]:
trajectories: dict[str, list[tuple[float, float]]] = defaultdict(list)
trial_sets = []
for window in windows:
trial_sets.append({str(trial["trial_id"]) for trial in window["trial_sanity"]})
for trial in window["trial_sanity"]:
trajectories[str(trial["trial_id"])].append(
(
float(trial["admitted_fraction"]),
float(trial["completed_fraction"]),
)
)
invariants = {
"same_trials_at_every_checkpoint": bool(trial_sets)
and all(trial_set == trial_sets[0] for trial_set in trial_sets[1:]),
"admitted_fraction_monotonic": all(
all(right[0] + 1e-12 >= left[0] for left, right in zip(values, values[1:]))
for values in trajectories.values()
),
"completed_fraction_monotonic": all(
all(right[1] + 1e-12 >= left[1] for left, right in zip(values, values[1:]))
for values in trajectories.values()
),
}
return {
"invariants": invariants,
"red_flags": [name for name, passed in invariants.items() if not passed],
}
def audit(*, run_root: Path, manifest_path: Path, output_path: Path) -> dict[str, Any]:
manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
if manifest.get("schema") != "intervention-response-phase-aware-pilot-manifest-v2":
raise ValueError("unexpected phase-aware pilot manifest schema")
fractions = [float(value) for value in manifest["checkpoints"]["fractions"]]
seconds = [float(value) for value in manifest["checkpoints"]["seconds"]]
cumulative_intervals = tuple((0.0, end_s) for end_s in seconds)
quarter_intervals = ((0.0, 75.0), (75.0, 150.0), (150.0, 225.0), (225.0, 300.0))
intervals = tuple(dict.fromkeys([*cumulative_intervals, *quarter_intervals]))
trials_by_interval, streams = load_interval_trials(
run_root=run_root, manifest=manifest, intervals_s=intervals
)
cumulative = [
analyze_window(
trials_by_interval[interval],
start_s=interval[0],
end_s=interval[1],
fraction=fraction,
)
for fraction, interval in zip(fractions, cumulative_intervals, strict=True)
]
quarter_blocks = [
analyze_window(
trials_by_interval[interval],
start_s=interval[0],
end_s=interval[1],
fraction=interval[1] / 300.0,
)
for interval in quarter_intervals
]
stable = stable_adjacent_features(cumulative)
load_consistency = consistent_load_regimes(cumulative, stable)
mechanism = mechanism_gate(stable, load_consistency)
mechanism_features = sorted(
{
feature
for transition in mechanism["passing_transitions"]
for feature in mechanism["by_transition"][transition]
}
)
full = cumulative[-1]
efficacy_stable = stable_adjacent_efficacy_features(cumulative)
efficacy_candidates = sorted(
{feature for features in efficacy_stable.values() for feature in features}
)
efficacy_features = sorted(set(efficacy_candidates) & set(mechanism_features))
controller = controller_gate(run_root, manifest)
coverage = cumulative_coverage_gate(cumulative)
red_flags = sorted(
{
flag
for window in [*cumulative, *quarter_blocks]
for flag in window["sanity"]["red_flags"]
}
| set(controller["red_flags"])
| set(coverage["red_flags"])
)
if red_flags:
decision = "STOP_DATA_INVALID"
elif not mechanism["passes"]:
decision = "STOP_NO_PHASE_STABLE_RESPONSE"
elif not full["efficacy"]["label_balance_sufficient"]:
decision = "MECHANISM_ONLY_NO_LABEL_BALANCE"
elif not efficacy_features:
decision = "STOP_NO_INCREMENTAL_TUNING_SIGNAL"
else:
decision = "OPEN_E2E_POLICY_TEST"
payload = {
"schema": SCHEMA,
"status": "COMPLETE",
"decision": decision,
"claim_boundary": "Development mechanism pilot; not a held-out paper claim.",
"mechanism_features": mechanism_features,
"mechanism_gate": mechanism,
"stable_adjacent_features": stable,
"load_consistency": load_consistency,
"stable_incremental_efficacy_features": efficacy_features,
"stable_incremental_efficacy_candidates": efficacy_candidates,
"stable_adjacent_efficacy_features": efficacy_stable,
"cumulative": cumulative,
"quarter_blocks": quarter_blocks,
"provenance": {
"analysis_script": str(Path(__file__).resolve()),
"analysis_script_sha256": sha256_file(Path(__file__).resolve()),
"manifest": str(manifest_path.resolve()),
"manifest_sha256": sha256_file(manifest_path),
"run_root": str(run_root.resolve()),
"streams": streams,
},
"sanity": {
"streams": len(streams),
"stream_bytes": numeric(item["bytes"] for item in streams),
"controller": controller,
"cumulative_coverage": coverage,
"red_flags": red_flags,
},
}
output_path.parent.mkdir(parents=True, exist_ok=True)
output_path.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
return payload
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--run-root", type=Path, required=True)
parser.add_argument("--manifest", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
payload = audit(
run_root=args.run_root,
manifest_path=args.manifest,
output_path=args.output,
)
print(
json.dumps(
{
"decision": payload["decision"],
"mechanism_features": payload["mechanism_features"],
"stable_incremental_efficacy_features": payload[
"stable_incremental_efficacy_features"
],
"sanity": payload["sanity"],
},
indent=2,
sort_keys=True,
)
)
if __name__ == "__main__":
main()