File size: 10,753 Bytes
0da1f00 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 | #!/usr/bin/env python3
"""Build the REPORT.md tables from this run's results.json files.
Reads the original Ray run plus the serial reruns, preferring a rerun when one
exists for a scenario index. Every harness family writes a different schema, so
each extractor below is keyed off what the harness actually emits rather than a
shared shape.
python3 make_tables.py # writes tables/*.md
"""
from __future__ import annotations
import json
import os
import glob
import statistics
HERE = os.path.dirname(os.path.abspath(__file__))
ORIG = os.path.join(HERE, "validate", "20260923-065922")
RERUN = os.path.join(HERE, "rerun")
OUT = os.path.join(HERE, "tables")
# Paper Table 7/8 "Speedup" column, arXiv 2605.23215 (H100 SXM5, locked clocks).
# Keyed by our scenario index so the mapping is explicit and auditable.
PAPER = {
0: ("Llama-3.1", 1.04), 1: ("Mixtral", 0.97), 3: ("BitNet 1.58b", 1.12),
4: ("GPT-OSS (MXFP4)", 1.02), 5: ("EAGLE-3", 0.98), 6: ("Gemma-4", 1.00),
7: ("Mamba", 1.05), 8: ("Mamba2", 0.97), 9: ("RWKV-7", 1.18),
10: ("GLA", 1.85), 11: ("RetNet", 1.86), 12: ("Qwen-3-Next", 1.24),
13: ("Kimi-Linear", 1.20), 14: ("TTT-E2E", 1.10), 15: ("Jamba", 1.02),
16: ("FLUX.1-Dev", 1.01), 17: ("HunyuanVideo-1.5", 0.97), 18: ("SDXL", 1.17),
19: ("SAM3.1", 1.05), 20: ("Whisper", 0.95), 21: ("CosyVoice3", 2.13),
22: ("Qwen2-VL", 0.91), 23: ("Qwen3-VL", 1.39), 25: ("Qwen-2.5-Omni", 2.02),
26: ("SigLIP-2", 0.93), 27: ("DINOv3", 0.99), 28: ("SwinV2", 1.17),
29: ("MobileNetV4", 1.15), 30: ("ConvNeXtV2", 0.99), 31: ("EfficientNetV2", 1.05),
32: ("YOLOv10", 1.06), 33: ("RTDetrV2", 1.08), 34: ("3DGS", 0.99),
35: ("InstantNGP", 0.97), 36: ("PointTransformerV3", 1.00), 37: ("OpenFold3", 1.03),
38: ("Pi0", 3.48), 39: ("DP3", 1.42), 40: ("DLRMv2", 1.06), 41: ("LightGCN", 1.03),
42: ("BGE-M3", 1.06), 43: ("ColBERTv2", 3.08), 44: ("LLaDA", 1.07),
45: ("Oasis", 1.29), 46: ("V-JEPA 2", 1.01),
}
# 24 (Qwen3-VL-235B-FP8) has no distinct paper row; the paper's single Qwen3-VL
# row is mapped to 23 (the 8B checkpoint).
def load_tasks() -> dict:
tasks = {}
with open(os.path.join(ORIG, "run.jsonl")) as fh:
for line in fh:
rec = json.loads(line)
if rec.get("event") == "task_finished":
tasks[rec["result"]["index"]] = rec["result"]
return tasks
def rerun_paths() -> dict:
out = {}
for path in glob.glob(os.path.join(RERUN, "*", "results.json")):
out[int(os.path.basename(os.path.dirname(path))[:3])] = path
return out
def walk(obj, path=""):
if isinstance(obj, dict):
for key, val in obj.items():
yield from walk(val, f"{path}.{key}")
elif isinstance(obj, list):
for i, val in enumerate(obj):
yield from walk(val, f"{path}[{i}]")
else:
yield path, obj
THROUGHPUT_KEYS = (
"images_per_second", "videos_per_second", "utterances_per_second",
"throughput_req_s", "points_per_second", "samples_per_second",
)
def throughput_ratios(data: dict) -> list[tuple[str, float]]:
"""(scenario, fastkernels/reference) for every throughput scenario."""
rows = []
for key in ("scenarios", "throughput_scenarios"):
for entry in data.get(key) or []:
if not isinstance(entry, dict):
continue
for field in ("speedup", "speedup_vs_sglang"):
if isinstance(entry.get(field), (int, float)):
rows.append((str(entry.get("scenario")), float(entry[field])))
break
if rows:
return rows
ours = data.get("fastkernels") or data.get("ours")
ref = data.get("reference") or data.get("vllm_omni") or data.get("timm")
if isinstance(ours, dict) and isinstance(ref, dict):
ours_t = {t["name"]: t for t in ours.get("throughput", []) if isinstance(t, dict)}
ref_t = {t["name"]: t for t in ref.get("throughput", []) if isinstance(t, dict)}
for name, o in ours_t.items():
r = ref_t.get(name)
if not r:
continue
for field in THROUGHPUT_KEYS:
if o.get(field) and r.get(field):
rows.append((name, o[field] / r[field]))
break
return rows
def latency_speedups(data: dict) -> list[tuple[str, float]]:
rows = []
for entry in data.get("latency_scenarios") or []:
if not isinstance(entry, dict):
continue
for field in ("speedup", "speedup_vs_sglang"):
if isinstance(entry.get(field), (int, float)):
rows.append((str(entry.get("scenario")), float(entry[field])))
break
return rows
VERDICT_LEAVES = ("overall_pass", "pass", "passed", "correctness_pass")
def verdicts(data: dict) -> list[tuple[str, bool]]:
out = []
for path, val in walk(data):
leaf = path.split(".")[-1].split("[")[0]
if isinstance(val, bool) and leaf in VERDICT_LEAVES:
out.append((path, val))
return out
def similarity_metrics(data: dict) -> list[tuple[str, float]]:
"""Alignment-ish scalars, so the report can show what was actually checked."""
out = []
for path, val in walk(data):
leaf = path.split(".")[-1].split("[")[0]
if not isinstance(val, (int, float)) or isinstance(val, bool):
continue
if any(tag in leaf for tag in ("cosine", "cos", "match_rate", "psnr", "exact_match")):
out.append((path.lstrip("."), float(val)))
return out
def fmt(x, nd=3):
return "—" if x is None else f"{x:.{nd}f}"
def main() -> None:
os.makedirs(OUT, exist_ok=True)
tasks = load_tasks()
reruns = rerun_paths()
records = {}
for idx in sorted(tasks):
task = tasks[idx]
path = reruns.get(idx) or os.path.join(
ORIG, os.path.basename(task["run_dir"]), "results.json"
)
if not os.path.isfile(path):
records[idx] = {"task": task, "data": None, "source": "missing"}
continue
with open(path) as fh:
data = json.load(fh)
records[idx] = {
"task": task,
"data": data,
"source": "rerun" if idx in reruns else "orig",
}
# ---- coverage ----
lines = ["| # | model | harness | tp | source | results.json | internal verdict |",
"| --- | --- | --- | --- | --- | --- | --- |"]
for idx, rec in records.items():
task, data = rec["task"], rec["data"]
if data is None:
verdict = "—"
have = "no"
else:
have = "yes"
vs = verdicts(data)
bad = [p for p, v in vs if v is False]
verdict = "**FAIL** " + ", ".join(b.lstrip(".") for b in bad) if bad else (
"pass" if vs else "no boolean field")
lines.append(
f"| {idx} | `{task['name']}` | `{task['harness']}` | {task['tp']} | "
f"{rec['source']} | {have} | {verdict} |"
)
open(os.path.join(OUT, "coverage.md"), "w").write("\n".join(lines) + "\n")
# ---- throughput ----
lines = ["| # | model | scenario | fastkernels / reference |",
"| --- | --- | --- | --- |"]
means = {}
for idx, rec in records.items():
if rec["data"] is None:
continue
rows = throughput_ratios(rec["data"])
if rows:
means[idx] = statistics.fmean(r for _, r in rows)
for scen, ratio in rows:
lines.append(f"| {idx} | `{rec['task']['name']}` | {scen} | {ratio:.3f}x |")
open(os.path.join(OUT, "throughput.md"), "w").write("\n".join(lines) + "\n")
# ---- latency ----
lines = ["| # | model | scenario | speedup |", "| --- | --- | --- | --- |"]
for idx, rec in records.items():
if rec["data"] is None:
continue
for scen, ratio in latency_speedups(rec["data"]):
lines.append(f"| {idx} | `{rec['task']['name']}` | {scen} | {ratio:.3f}x |")
open(os.path.join(OUT, "latency.md"), "w").write("\n".join(lines) + "\n")
# ---- alignment ----
lines = ["| # | model | metric | value |", "| --- | --- | --- | --- |"]
for idx, rec in records.items():
if rec["data"] is None:
continue
for path, val in similarity_metrics(rec["data"]):
if "latencies" in path:
continue
lines.append(f"| {idx} | `{rec['task']['name']}` | `{path}` | {val:.6f} |")
open(os.path.join(OUT, "alignment.md"), "w").write("\n".join(lines) + "\n")
# ---- paper comparison ----
lines = ["| # | architecture | paper (H100, locked) | this run (H200, unlocked) | delta |",
"| --- | --- | --- | --- | --- |"]
deltas = []
for idx, (arch, paper_val) in sorted(PAPER.items()):
ours = means.get(idx)
if ours is None:
lines.append(f"| {idx} | {arch} | {paper_val:.2f}x | — | — |")
continue
delta = (ours - paper_val) / paper_val * 100.0
deltas.append((arch, delta, paper_val, ours))
lines.append(
f"| {idx} | {arch} | {paper_val:.2f}x | {ours:.3f}x | {delta:+.1f}% |"
)
absd = [abs(d) for _, d, _, _ in deltas]
flipped = [(a, p, o) for a, _, p, o in deltas if (p - 1) * (o - 1) < 0]
lines += [
"",
f"Comparable architectures: **{len(deltas)}**. "
f"Exact matches: **{sum(1 for x in absd if x == 0)}**. "
f"Median \\|delta\\|: **{statistics.median(absd):.1f}%**. "
f"Max \\|delta\\|: **{max(absd):.1f}%**.",
"",
f"Within 2%: {sum(1 for x in absd if x <= 2)}/{len(absd)} · "
f"within 5%: {sum(1 for x in absd if x <= 5)}/{len(absd)} · "
f"within 10%: {sum(1 for x in absd if x <= 10)}/{len(absd)} · "
f"within 20%: {sum(1 for x in absd if x <= 20)}/{len(absd)}",
"",
f"Direction reversals (one side above 1x, the other below): **{len(flipped)}**",
"",
"| architecture | paper | this run |",
"| --- | --- | --- |",
]
for arch, paper_val, ours in flipped:
lines.append(f"| {arch} | {paper_val:.2f}x | {ours:.3f}x |")
open(os.path.join(OUT, "paper_comparison.md"), "w").write("\n".join(lines) + "\n")
print(f"wrote {OUT}/coverage.md, throughput.md, latency.md, alignment.md, paper_comparison.md")
print(f" scenarios with results: {sum(1 for r in records.values() if r['data'])}/{len(records)}")
print(f" comparable vs paper : {len(deltas)}")
print(f" median |delta| : {statistics.median(absd):.1f}%")
print(f" direction reversals : {len(flipped)}")
if __name__ == "__main__":
main()
|