Audit graph-aligned Frontier surface

This commit is contained in:
2026-07-18 00:09:06 +08:00
parent e2cd43808d
commit e6f3e4a690

View File

@@ -0,0 +1,272 @@
#!/usr/bin/env python3
"""Compare a graph-aligned Frontier Qwen30 surface with audited real vLLM data.
The input surface may have been produced by independent TP jobs. Therefore
this script deliberately reads the per-cell ``result.json`` files rather than
the runner's root manifest, which is only a convenience artifact and can be
overwritten by concurrent dispatch.
"""
from __future__ import annotations
import argparse
import hashlib
import json
import math
from itertools import combinations
from pathlib import Path
from typing import Any
CONFIGS = tuple(f"tp{tp}_mns{mns}" for tp in (1, 2, 4) for mns in (8, 16, 32, 64))
METRICS = ("ttft_ms", "tpot_ms", "e2e_ms")
STATISTICS = ("mean", "p90")
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
parser.add_argument("--sim-root", type=Path, required=True)
parser.add_argument("--real-audit", type=Path, required=True)
parser.add_argument("--json-output", type=Path, required=True)
parser.add_argument("--markdown-output", type=Path, required=True)
return parser.parse_args()
def read_json(path: Path) -> dict[str, Any]:
with path.open(encoding="utf-8") as source:
payload = json.load(source)
if not isinstance(payload, dict):
raise ValueError(f"expected JSON object: {path}")
return payload
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(value: Any, label: str) -> float:
if not isinstance(value, (int, float)) or isinstance(value, bool):
raise ValueError(f"{label} is not numeric: {value!r}")
result = float(value)
if not math.isfinite(result) or result < 0:
raise ValueError(f"{label} is invalid: {value!r}")
return result
def parse_config(name: str) -> tuple[int, int]:
try:
tp_part, mns_part = name.split("_", 1)
return int(tp_part.removeprefix("tp")), int(mns_part.removeprefix("mns"))
except ValueError as error:
raise ValueError(f"invalid config name: {name}") from error
def load_sim_cell(root: Path, name: str) -> dict[str, Any]:
tp, mns = parse_config(name)
path = root / "runs" / name / f"tp{tp}" / "result.json"
result = read_json(path)
if result.get("status") != "completed":
raise ValueError(f"{path}: status is not completed")
config = result.get("config")
if config != {"tp": tp, "mns": mns, "name": name}:
raise ValueError(f"{path}: config drift: {config!r}")
if result.get("trace_label") != f"tp{tp}":
raise ValueError(f"{path}: trace label does not match TP")
if result.get("request_count") != 129:
raise ValueError(f"{path}: expected 129 requests")
score = result.get("score")
if not isinstance(score, dict):
raise ValueError(f"{path}: missing score")
metrics: dict[str, dict[str, float]] = {}
for metric in METRICS:
prefix = metric.removesuffix("_ms")
metrics[metric] = {
statistic: numeric(score.get(f"{prefix}_{statistic}_ms"), f"{path}:{metric}:{statistic}")
for statistic in STATISTICS
}
return {
"config": name,
"tp": tp,
"mns": mns,
"trace_label": result["trace_label"],
"trace_sha256": result.get("trace_sha256"),
"offered_request_rate": numeric(result.get("offered_request_rate"), f"{path}:rate"),
"offered_request_rate_per_gpu": numeric(
result.get("offered_request_rate_per_gpu"), f"{path}:rate_per_gpu"
),
"request_count": result["request_count"],
"elapsed_seconds": numeric(result.get("elapsed_seconds"), f"{path}:elapsed"),
"result_path": str(path.resolve()),
"result_sha256": sha256_file(path),
"metrics": metrics,
}
def load_real_cell(audit: dict[str, Any], name: str) -> dict[str, dict[str, float]]:
configs = audit.get("configs")
if not isinstance(configs, dict) or name not in configs:
raise ValueError(f"real audit lacks {name}")
result = configs[name]
metrics = result.get("metrics")
if not isinstance(metrics, dict):
raise ValueError(f"real audit {name} lacks metrics")
parsed: dict[str, dict[str, float]] = {}
for metric in METRICS:
values = metrics.get(metric)
if not isinstance(values, dict):
raise ValueError(f"real audit {name} lacks {metric}")
parsed[metric] = {
"mean": numeric(values.get("pooled_mean_ms"), f"real:{name}:{metric}:mean"),
"p90": numeric(values.get("pooled_p90_ms"), f"real:{name}:{metric}:p90"),
}
return parsed
def ranking(values: dict[str, float]) -> list[str]:
return [name for name, _ in sorted(values.items(), key=lambda item: (item[1], item[0]))]
def pairwise_agreement(sim: dict[str, float], real: dict[str, float]) -> dict[str, int]:
concordant = discordant = ties = 0
for left, right in combinations(sorted(sim), 2):
sim_delta = sim[left] - sim[right]
real_delta = real[left] - real[right]
if math.isclose(sim_delta, 0.0, abs_tol=1e-9) or math.isclose(
real_delta, 0.0, abs_tol=1e-9
):
ties += 1
elif (sim_delta > 0) == (real_delta > 0):
concordant += 1
else:
discordant += 1
return {
"concordant_pairs": concordant,
"discordant_pairs": discordant,
"tied_pairs": ties,
"informative_pairs": concordant + discordant,
}
def comparison_summary(cells: dict[str, dict[str, Any]], real_audit: dict[str, Any]) -> dict[str, Any]:
summaries: dict[str, Any] = {}
for metric in METRICS:
for statistic in STATISTICS:
sim_values = {
name: cells[name]["metrics"][metric][f"sim_{statistic}_ms"]
for name in CONFIGS
}
real_values = {
name: load_real_cell(real_audit, name)[metric][statistic] for name in CONFIGS
}
sim_ranking = ranking(sim_values)
real_ranking = ranking(real_values)
pairwise = pairwise_agreement(sim_values, real_values)
summaries[f"{metric}:{statistic}"] = {
"sim_winner": sim_ranking[0],
"real_winner": real_ranking[0],
"winner_match": sim_ranking[0] == real_ranking[0],
"sim_ranking": sim_ranking,
"real_ranking": real_ranking,
**pairwise,
"pairwise_agreement_fraction": (
pairwise["concordant_pairs"] / pairwise["informative_pairs"]
if pairwise["informative_pairs"]
else None
),
}
return summaries
def render_markdown(payload: dict[str, Any]) -> str:
lines = [
"# Frontier piecewise graph-profile vs. real vLLM",
"",
"Each cell uses its TP-normalized, 129-request trace. Frontier values are one deterministic simulation; "
"real values pool three fresh-server trials (387 requests/cell).",
"",
"| Config | TTFT sim / real mean (ms) | TPOT sim / real mean (ms) | E2E sim / real mean (ms) |",
"|---|---:|---:|---:|",
]
for name in CONFIGS:
cell = payload["cells"][name]
entries = []
for metric in METRICS:
values = cell["metrics"][metric]
entries.append(f"{values['sim_mean_ms']:.1f} / {values['real_mean_ms']:.1f}")
lines.append(f"| {name} | " + " | ".join(entries) + " |")
lines.extend(["", "## Selection agreement", "", "| Target | Frontier winner | Real winner | Match | Pairwise agreement |", "|---|---|---|---:|---:|"])
for target, summary in payload["selection"].items():
agreement = summary["pairwise_agreement_fraction"]
agreement_text = "n/a" if agreement is None else f"{agreement:.1%}"
lines.append(
f"| {target} | {summary['sim_winner']} | {summary['real_winner']} | "
f"{'yes' if summary['winner_match'] else 'no'} | {agreement_text} "
f"({summary['concordant_pairs']}/{summary['informative_pairs']}) |"
)
lines.append("")
return "\n".join(lines)
def main() -> None:
args = parse_args()
sim_root = args.sim_root.resolve()
real_path = args.real_audit.resolve()
real_audit = read_json(real_path)
if set(real_audit.get("configs", {})) != set(CONFIGS):
raise ValueError("real audit config set does not match the 12-cell surface")
cells: dict[str, dict[str, Any]] = {}
for name in CONFIGS:
sim = load_sim_cell(sim_root, name)
real = load_real_cell(real_audit, name)
metrics = {}
for metric in METRICS:
metrics[metric] = {
f"sim_{statistic}_ms": sim["metrics"][metric][statistic]
for statistic in STATISTICS
}
metrics[metric].update(
{
f"real_{statistic}_ms": real[metric][statistic]
for statistic in STATISTICS
}
)
metrics[metric].update(
{
f"{statistic}_ratio_sim_over_real": sim["metrics"][metric][statistic]
/ real[metric][statistic]
for statistic in STATISTICS
}
)
cells[name] = {key: value for key, value in sim.items() if key != "metrics"}
cells[name]["metrics"] = metrics
payload = {
"schema": "frontier-qwen30-piecewise-graph-comparison-v1",
"contract": {
"configs": list(CONFIGS),
"request_count_per_cell": 129,
"real_trials_per_cell": 3,
"trace_policy": "TP-normalized arrivals with full 16-token prefix blocks only",
"graph_semantics": "Frontier piecewise; CUDA_EVENT for prefill/mixed and KERNEL_ONLY for captured decode",
},
"sim_root": str(sim_root),
"real_audit": str(real_path),
"real_audit_sha256": sha256_file(real_path),
"cells": cells,
"selection": comparison_summary(cells, real_audit),
}
args.json_output.parent.mkdir(parents=True, exist_ok=True)
args.markdown_output.parent.mkdir(parents=True, exist_ok=True)
args.json_output.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n", encoding="utf-8")
args.markdown_output.write_text(render_markdown(payload), encoding="utf-8")
print(json.dumps(payload["selection"], indent=2, sort_keys=True))
if __name__ == "__main__":
main()