Evaluate vLLM 0.20 profiles against Frontier

This commit is contained in:
2026-07-16 23:59:10 +08:00
parent 008324e70c
commit 76107d3e87
33 changed files with 13973 additions and 14 deletions

View File

@@ -0,0 +1,229 @@
#!/usr/bin/env python3
"""Score a replacement-profile S2 sweep against the frozen real oracle."""
from __future__ import annotations
import argparse
import importlib.util
import json
import math
import sys
from pathlib import Path
from typing import Any
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
parser.add_argument("--shard-metrics", type=Path, action="append", required=True)
parser.add_argument("--ground-truth", type=Path, required=True)
parser.add_argument("--historical-metrics", type=Path, required=True)
parser.add_argument("--historical-analyzer", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
return parser.parse_args()
def load_module(path: Path):
spec = importlib.util.spec_from_file_location("simfid_s2_analyzer", path)
if spec is None or spec.loader is None:
raise ImportError(path)
module = importlib.util.module_from_spec(spec)
sys.modules[spec.name] = module
spec.loader.exec_module(module)
return module
def ranking(scores: dict[str, float]) -> list[dict[str, Any]]:
return [
{"rank": index + 1, "cell": cell, "score": score}
for index, (cell, score) in enumerate(
sorted(scores.items(), key=lambda item: (-item[1], item[0]))
)
]
def robust_loao(
analyzer: Any,
runs: list[dict[str, Any]],
mode: str,
reading: str,
real_scores: dict[str, float],
) -> dict[str, Any]:
"""Historical LOAO with an explicit null range for all-tied concordance."""
anchors = sorted(
{(row["cell_id"], int(row["probe_index"])) for row in runs},
key=lambda item: (analyzer.CELL_ORDER.index(item[0]), item[1]),
)
replicates = []
undefined = []
for cell, probe in anchors:
scores, detail = analyzer.cell_scores(
runs, mode, reading, removed=(cell, probe)
)
if scores is None:
undefined.append(
{"removed_cell_id": cell, "removed_probe_index": probe, **detail}
)
continue
metric = analyzer.rank_metrics(real_scores, scores)
replicates.append(
{
"removed_cell_id": cell,
"removed_probe_index": probe,
"top1_optimistic_regret": metric["top1"]["optimistic_regret"],
"top1_worst_case_regret": metric["top1"]["worst_case_regret"],
"top5_minimum_exact_five_overlap": metric["top5"]["minimum_exact_five_overlap"],
"top5_maximum_exact_five_overlap": metric["top5"]["maximum_exact_five_overlap"],
"top5_optimistic_regret": metric["top5"]["optimistic_regret"],
"top5_worst_case_regret": metric["top5"]["worst_case_regret"],
"tau_b": metric["kendall_tau_b"]["tau_b"],
"pairwise_exact_sign_accuracy": metric["pairwise_direction"]["exact_sign_accuracy"],
"pairwise_non_tied_concordance": metric["pairwise_direction"]["non_tied_concordance"],
"trap_reproduced": metric["named_interactions"]["trap_reproduced"],
"tp2_mns32_unique_global_best": metric["named_interactions"]["tp2_mns32_unique_global_best"],
}
)
scalar_keys = (
"top1_optimistic_regret",
"top1_worst_case_regret",
"top5_minimum_exact_five_overlap",
"top5_maximum_exact_five_overlap",
"top5_optimistic_regret",
"top5_worst_case_regret",
"tau_b",
"pairwise_exact_sign_accuracy",
"pairwise_non_tied_concordance",
)
ranges = {}
for key in scalar_keys:
values = [row[key] for row in replicates if row[key] is not None]
ranges[key] = {
"min": min(values) if values else None,
"max": max(values) if values else None,
}
return {
"replicate_count": len(anchors),
"defined_replicates": len(replicates),
"undefined_replicates": len(undefined),
"undefined": undefined,
"ranges": ranges,
"trap_reproduced_count": sum(row["trap_reproduced"] for row in replicates),
"tp2_mns32_unique_global_best_count": sum(
row["tp2_mns32_unique_global_best"] for row in replicates
),
"replicates": replicates,
"range_semantics": "deterministic LOAO sensitivity ranges, not confidence intervals",
}
def main() -> None:
args = parse_args()
analyzer = load_module(args.historical_analyzer)
ground = json.loads(args.ground_truth.read_text())
cells = {str(cell["cell_id"]): cell for cell in ground["cells"]}
if set(cells) != set(analyzer.CELL_ORDER):
raise ValueError("ground-truth cells differ from the frozen 3x4 surface")
result_rows: list[dict[str, Any]] = []
sources = []
for path in args.shard_metrics:
payload = json.loads(path.read_text())
if payload["status"] != "PASS":
raise ValueError(f"shard did not pass: {path}")
sources.append({"path": str(path.resolve()), "runs": len(payload["results"])})
result_rows.extend(payload["results"])
if len(result_rows) != 92:
raise ValueError(f"expected 92 replacement-profile probes, found {len(result_rows)}")
seen: set[tuple[str, int]] = set()
runs = []
for row in result_rows:
cell_id = str(row["cell"])
probe_index = int(row["probe_index"])
key = (cell_id, probe_index)
if key in seen:
raise ValueError(f"duplicate probe {key}")
seen.add(key)
real_probe = cells[cell_id]["probe_history"][probe_index]
if not math.isclose(
float(real_probe["sampling_u"]), float(row["sampling_u"]), rel_tol=0, abs_tol=1e-15
):
raise ValueError(f"sampling_u mismatch for {key}")
if int(real_probe["request_count"]) != int(row["selected_count"]):
raise ValueError(f"request count mismatch for {key}")
runs.append(
{
**row,
"cell_id": cell_id,
"mode": "vllm020-profile-only",
"probe_index": probe_index,
"sampling_u": float(row["sampling_u"]),
"request_count": int(row["selected_count"]),
"tensor_parallel_size": int(row["tensor_parallel_size"]),
"real_anchor": {
"feasible": bool(real_probe["feasible"]),
"pass_rate": float(real_probe["pass_rate"]),
"request_count": int(real_probe["request_count"]),
},
}
)
historical = json.loads(args.historical_metrics.read_text())
real_scores = {cell: float(value) for cell, value in historical["real_scores"].items()}
analyses = {}
for reading in analyzer.READINGS:
scores, detail = analyzer.cell_scores(runs, "vllm020-profile-only", reading)
if scores is None:
raise ValueError(f"undefined {reading} scores: {detail}")
analysis = {
"reading": reading,
"simulated_scores": scores,
"ranking": ranking(scores),
"cell_score_details": detail["cells"],
"metrics": analyzer.rank_metrics(real_scores, scores),
"loao": robust_loao(
analyzer, runs, "vllm020-profile-only", reading, real_scores
),
}
if reading == "SLO-gated":
analysis["false_feasibility"] = analyzer.false_feasibility(
runs, "vllm020-profile-only"
)
analyses[reading] = analysis
historical_modes = {}
for label, key in (
("historical-profile-only", "uncalibrated/SLO-gated"),
("historical-per-tp-calibration", "frozen-calibrated/SLO-gated"),
):
value = historical["analyses"][key]
historical_modes[label] = {
"simulated_scores": value["simulated_scores"],
"ranking": ranking(value["simulated_scores"]),
"metrics": value["metrics"],
"false_feasibility": value["false_feasibility"],
}
output = {
"schema": "frontier-qwen30-s2-profile-ablation.v1",
"scope": {
"cells": len(analyzer.CELL_ORDER),
"probes": len(runs),
"trace_horizon_seconds": 60,
"calibration_a_tp": 1.0,
},
"sources": {
"replacement_profile_shards": sources,
"ground_truth": str(args.ground_truth.resolve()),
"historical_metrics": str(args.historical_metrics.resolve()),
},
"real_scores": real_scores,
"historical_modes": historical_modes,
"vllm020_profile_only": analyses,
}
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(json.dumps(output, indent=2, sort_keys=True) + "\n")
print(args.output)
if __name__ == "__main__":
main()