File size: 4,718 Bytes
9d6c277 | 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 | """Score Exp 2 (RULER 6 tasks: 4K/32K/128K x 180) and compare runs with paired statistics.
Usage: python score_exp2.py # all comparisons whose result files exist
python score_exp2.py --by-task # also break RULER down by task
Each run folder ~/bench/exp2/<RUN>/ holds vllm bench serve results whose
generated_texts[i] belongs to JSONL row i (inputs were unshuffled).
"""
import argparse
import json
import os
import random
import re
from pathlib import Path
B = Path.home() / "bench"
EXP = B / os.environ.get("KVQ_EXP_OUT", "exp2") # results folder (exp2_gemma for Gemma)
DATA = {"ruler_4k": "ruler_4k_6t.jsonl", "ruler_32k": "ruler_32k_6t.jsonl", "ruler_128k": "ruler_128k_6t.jsonl",
"docvqa": "docvqa_1000.jsonl"}
# (treatment, baseline, what the difference measures)
PAIRS = [("BF16_rep2", "BF16", "noise floor (FA)"),
("BF16-FI_rep2", "BF16-FI", "noise floor (FI)"),
("BF16-FI", "BF16", "kernel effect"),
("FP8", "BF16", "FP8 KV effect"),
("FP4", "BF16-FI", "NVFP4 KV effect"),
("FP8", "BF16-TFA", "FP8 KV, same kernel")] # Gemma DocVQA: both on TRITON_FLASH_ATTN
def lev(a, b):
prev = list(range(len(b) + 1))
for i, ca in enumerate(a, 1):
cur = [i]
for j, cb in enumerate(b, 1):
cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (ca != cb)))
prev = cur
return prev[-1]
def anls(pred, answers):
p = pred.strip().lower()
best = 0.0
for g in answers:
g = g.strip().lower()
nl = lev(p, g) / max(len(p), len(g), 1)
best = max(best, 1 - nl if nl < 0.5 else 0.0)
return best
def score(ds, row, text):
if ds.startswith("ruler"):
return sum(a.lower() in text.lower() for a in row["answers"]) / len(row["answers"])
lines = [l for l in text.strip().split("\n") if l.strip()]
return anls(re.sub(r"[*`]", "", lines[-1]) if lines else "", row["answers"])
_cache = {}
def scores(run, ds):
key = (run, ds)
if key not in _cache:
f = EXP / run / f"{ds}.json"
if not f.exists():
_cache[key] = None
else:
rows = [json.loads(l) for l in open(B / os.environ.get("KVQ_DATA", "data") / DATA[ds])]
res = json.loads(f.read_text())
errs = sum(1 for e in res.get("errors", []) if e)
s = [score(ds, r, t) for r, t in zip(rows, res["generated_texts"])]
_cache[key] = (s, rows, errs)
return _cache[key]
def paired_ci(a, b, iters=10000, seed=0):
d = [x - y for x, y in zip(a, b)]
n = len(d)
rng = random.Random(seed)
means = sorted(sum(d[rng.randrange(n)] for _ in range(n)) / n for _ in range(iters))
return sum(d) / n, means[int(0.025 * iters)], means[int(0.975 * iters)]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--by-task", action="store_true")
a = ap.parse_args()
runs = sorted(p.name for p in EXP.iterdir() if p.is_dir())
print("Absolute scores (RULER: fraction of answers found; DocVQA: ANLS)")
print(f"{'run':<16}" + "".join(f"{d:>11}" for d in DATA))
for r in runs:
line = f"{r:<16}"
for d in DATA:
s = scores(r, d)
line += f"{'-':>11}" if s is None else f"{100 * sum(s[0]) / len(s[0]):>9.2f}" + (f"e{s[2]}" if s[2] else " ")
print(line)
print("\nPaired differences (treatment - baseline, points), 95% bootstrap CI, flips right->wrong / wrong->right")
for t, base, label in PAIRS:
for d in DATA:
st, sb = scores(t, d), scores(base, d)
if not st or not sb:
continue
diff, lo, hi = paired_ci(st[0], sb[0])
rw = sum(1 for x, y in zip(st[0], sb[0]) if y == 1 and x < 1)
wr = sum(1 for x, y in zip(st[0], sb[0]) if y < 1 and x == 1)
print(f"{label:<17} {t:>15} vs {base:<10} {d:<11} {100 * diff:+6.2f} [{100 * lo:+6.2f}, {100 * hi:+6.2f}] flips {rw}/{wr}")
if a.by_task:
print("\nRULER by task (mean over all lengths)")
tasks = sorted({row["task"] for row in scores(runs[0], "ruler_4k")[1]}) if scores(runs[0], "ruler_4k") else []
print(f"{'task':<18}" + "".join(f"{r:>16}" for r in runs))
for task in tasks:
line = f"{task:<18}"
for r in runs:
vals = []
for d in [k for k in DATA if k.startswith("ruler")]:
s = scores(r, d)
if s:
vals += [x for x, row in zip(s[0], s[1]) if row["task"] == task]
line += f"{100 * sum(vals) / len(vals):>16.2f}" if vals else f"{'-':>16}"
print(line)
if __name__ == "__main__":
main()
|