File size: 8,324 Bytes
be4fcab 4eac6b1 be4fcab 4eac6b1 be4fcab 18ed693 be4fcab 4eac6b1 be4fcab 4eac6b1 be4fcab 5644ac8 4eac6b1 be4fcab | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 | #!/usr/bin/env python3
"""Aggregate 42-task benchmark results into results.json + results_full42.csv.
Source of truth: the private `brywil/gguf-quant-lab` checkout's
`benchmarks/results_<label>_<suite>.json` files.
Scoring contract
----------------
- A task is **passed** only when every pytest item in it passed.
- A **void** is a harness limit (timeout, output-cap truncation, no completion).
Voids are excluded from capability: `valid = tasks - void`.
- The headline is `passed / valid`, not `passed / tasks`. Never count a void as a failure.
Usage
-----
python scripts/aggregate_results.py --lab ~/gguf-quant-lab --out .
Outputs
-------
results.json all full-42 and subset runs, with per-suite/total fields
results_full42.csv flat one-row-per-run table
A run is "full-42" when all four suites are present. Labels are grouped by stripping a
trailing `-rN` / `-repN` so repeated runs of the same arm compare together.
"""
from __future__ import annotations
import argparse
import csv
import glob
import json
import os
import re
from collections import defaultdict
SUITES = [
("opencode_tasks", "general", 25),
("opencode_tasks_hard", "hard", 5),
("opencode_tasks_frontier", "frontier", 9),
("opencode_tasks_tier4a", "tier4a", 3),
]
def server_timings(lab: str, label: str) -> dict:
"""Parse llama.cpp server logs for raw prefill/decode throughput.
Sums ``prompt eval time`` (prefill) and bare ``eval time`` (decode)
across every server log matching *label* (resumed runs produce 2-4
files). Returns ``{prefill_tok, prefill_ms, decode_tok, decode_ms}``
or an empty dict when no log yields timings.
"""
prompt_re = re.compile(r"prompt eval time =\s*([0-9.]+) ms /\s*(\d+) tokens")
decode_re = re.compile(r"(?<!prompt )eval time =\s*([0-9.]+) ms /\s*(\d+) tokens")
p_tok = p_ms = d_tok = d_ms = 0
found = False
for path in sorted(glob.glob(os.path.join(lab, "benchmarks", "results", f"{label}*server.log"))):
try:
text = open(path, errors="ignore").read()
except OSError:
continue
for m in prompt_re.finditer(text):
p_ms += float(m.group(1)); p_tok += int(m.group(2)); found = True
for m in decode_re.finditer(text):
d_ms += float(m.group(1)); d_tok += int(m.group(2)); found = True
if not found:
return {}
return {
"prefill_tok": p_tok, "prefill_ms": round(p_ms, 1),
"decode_tok": d_tok, "decode_ms": round(d_ms, 1),
"prefill_tok_s": round(p_tok / (p_ms / 1000), 1) if p_ms else None,
"decode_tok_s": round(d_tok / (d_ms / 1000), 1) if d_ms else None,
}
def model_path(lab: str, label: str) -> str:
"""Best-effort: the served GGUF basename from a local server log, if retained."""
for path in sorted(glob.glob(os.path.join(lab, "benchmarks", "results", f"{label}*server.log"))):
try:
text = open(path, errors="ignore").read(200_000)
except OSError:
continue
m = re.search(r"loading model '([^']+)'", text)
if m:
return os.path.basename(m.group(1))
return ""
def main() -> None:
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--lab", default=os.path.expanduser("~/gguf-quant-lab"),
help="checkout containing benchmarks/results_*.json")
ap.add_argument("--out", default=".", help="output directory")
args = ap.parse_args()
per_label: dict[str, dict] = defaultdict(dict)
for path in glob.glob(os.path.join(args.lab, "benchmarks", "results_*.json")):
base = os.path.basename(path)[len("results_"):-len(".json")]
for suffix, name, _ in SUITES:
if not base.endswith("_" + suffix):
continue
label = base[: -(len(suffix) + 1)]
try:
data = json.load(open(path))
except (OSError, json.JSONDecodeError):
continue
challenges = data.get("challenges", []) or []
tasks = len(challenges)
void = int(data.get("n_void", 0) or 0)
passed = sum(1 for c in challenges if c.get("passed"))
partial = (sum(float(c.get("score", 0) or 0) for c in challenges) / tasks) if tasks else 0.0
per_label[label][name] = {
"tasks": tasks,
"valid": tasks - void,
"passed": passed,
"void": void,
"partial": round(partial, 4),
"wall_s": round(float(data.get("wall_s", 0) or 0), 1),
"tokens_out": sum(int(c.get("tokens_out", 0) or 0) for c in challenges),
"source": os.path.basename(path),
}
break
runs, partial_runs = [], []
for label, suites in sorted(per_label.items()):
timings = server_timings(args.lab, label)
record = {
"label": label,
"group": re.sub(r"-(r|rep)\d+$", "", label),
"box": "thor" if (
label.lower().startswith("thor-") or
label.lower().startswith("thor_") or
label.lower().endswith("_thor") or
label.lower().endswith("-thor") or
"_thor_" in label.lower() or
"-thor-" in label.lower() or
"_thor-" in label.lower()
) else "gb10",
"model_path": model_path(args.lab, label),
"suites": suites,
"total_passed": sum(s["passed"] for s in suites.values()),
"total_valid": sum(s["valid"] for s in suites.values()),
"total_void": sum(s["void"] for s in suites.values()),
"total_wall_s": round(sum(s["wall_s"] for s in suites.values()), 1),
"total_tokens_out": sum(s["tokens_out"] for s in suites.values()),
**timings,
}
(runs if len(suites) == len(SUITES) else partial_runs).append(record)
os.makedirs(args.out, exist_ok=True)
payload = {
"generated_from": "brywil/gguf-quant-lab benchmarks/results_*.json",
"scoring": "passed/valid; valid = tasks - void; void = timeout/truncation/no completion (not a failure)",
"suites": {name: {"dir": suffix, "tasks": count} for suffix, name, count in SUITES},
"runs": runs,
"partial_runs": partial_runs,
}
with open(os.path.join(args.out, "results.json"), "w") as f:
json.dump(payload, f, indent=2)
csv_path = os.path.join(args.out, "results_full42.csv")
with open(csv_path, "w", newline="") as f:
writer = csv.writer(f)
writer.writerow(["label", "group", "box", "model_path",
"general", "hard", "frontier", "tier4a",
"valid_total", "void", "passed_total",
"wall_h", "tokens_out", "tokens_per_task",
"tok_s_eff", "tok_s_prefill", "tok_s_decode",
"prefill_ms", "decode_ms", "partial_mean"])
for r in sorted(runs, key=lambda x: (x["group"], x["label"])):
s = r["suites"]
writer.writerow([
r["label"], r["group"], r["box"], r["model_path"],
f"{s['general']['passed']}/{s['general']['valid']}",
f"{s['hard']['passed']}/{s['hard']['valid']}",
f"{s['frontier']['passed']}/{s['frontier']['valid']}",
f"{s['tier4a']['passed']}/{s['tier4a']['valid']}",
f"{r['total_passed']}/{r['total_valid']}", r["total_void"],
r["total_passed"], round(r["total_wall_s"] / 3600, 2),
r["total_tokens_out"],
round(r["total_tokens_out"] / max(1, r["total_valid"]), 1),
round(r["total_tokens_out"] / max(1, r["total_wall_s"]), 2),
r.get("prefill_tok_s"), r.get("decode_tok_s"),
r.get("prefill_ms"), r.get("decode_ms"),
round(sum(s[k]["partial"] * s[k]["tasks"] for k in s) / max(1, r["total_valid"]), 4),
])
print(f"full-42 runs: {len(runs)} | subset runs: {len(partial_runs)}")
print(f"wrote {os.path.join(args.out, 'results.json')} and results_full42.csv")
if __name__ == "__main__":
main()
|