File size: 13,906 Bytes
1d2de8a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 | """Resumable panel scoring, fit-only calibration and one-shot held-out reports."""
import gzip
import hashlib
import json
from collections import defaultdict
from pathlib import Path
import numpy as np
from scipy.optimize import minimize_scalar
from ._vendor.semantics import listed_probs, p_yes
from .api import TASKS
from .artifacts import ADAPTER_SHA, HEADS_SHA, digest, sha256
def load_panel(directory):
directory = Path(directory)
manifest = json.loads((directory / "manifest.json").read_text())
if (
manifest["adapter_sha256"] != ADAPTER_SHA
or manifest["heads_sha256"] != HEADS_SHA
or manifest["readout_mode"] != "four_collapsed"
):
raise ValueError("Panel belongs to a different checkpoint or answer semantics")
raw = gzip.decompress((directory / "jobs.json.gz").read_bytes())
if hashlib.sha256(raw).hexdigest() != manifest["jobs_sha256"]:
raise ValueError("Panel jobs checksum mismatch")
jobs = json.loads(raw)
if len({r["id"] for r in jobs}) != len(jobs):
raise ValueError("Duplicate panel branch IDs")
return jobs, manifest
def score_panel(model, panel, output):
"""Atomically persist each document so interruption never requires rescoring it."""
jobs, manifest = load_panel(panel)
output = Path(output)
output.mkdir(parents=True, exist_ok=True)
identity = {
"runtime": model.identity,
"panel_sha256": manifest["jobs_sha256"],
"panel_role": Path(panel).name.split("-")[0],
}
meta = output / "identity.json"
if meta.exists() and json.loads(meta.read_text()) != identity:
raise ValueError("Cannot resume with different model code, weights or panel")
meta.write_text(json.dumps(identity, indent=2))
documents = defaultdict(list)
for row in jobs:
documents[row["document_key"]].append(row)
for key, group in documents.items():
path = output / (key + ".json")
if path.exists():
record = json.loads(path.read_text())
body = {k: v for k, v in record.items() if k != "sha256"}
if (
record["sha256"] != digest(body)
or record["identity"] != digest(identity)
or [r["id"] for r in record["rows"]] != [r["id"] for r in group]
):
raise ValueError("Corrupt or mismatched resumed document")
continue
parts = group[0].get("parts") or [{"text": group[0]["doc"]}]
if any((r.get("parts") or [{"text": r["doc"]}]) != parts for r in group):
raise ValueError("Document key aliases different sources")
with model.prefill(parts) as state:
rows = []
for job in group:
result = model.engine.ask(state._data, job["block"], job["n"], job["head_key"])
rows.append({**result, **{k: job[k] for k in ("id", "task", "gold", "n", "question_id")}})
body = {
"identity": digest(identity),
"rows": rows,
"prefix_tokens": state.prefix_tokens,
"prefill_seconds": state._data["prefill_seconds"],
}
temp = path.with_suffix(".tmp")
temp.write_text(json.dumps({**body, "sha256": digest(body)}))
temp.replace(path)
print("Scored " + key + " " + str(len(rows)) + " branches", flush=True)
completed = {
"identity": digest(identity),
"documents": len(documents),
"branches": len(jobs),
"files": {key + ".json": sha256(output / (key + ".json")) for key in documents},
}
(output / "complete.json").write_text(json.dumps(completed, indent=2))
def read_scores(directory):
directory = Path(directory)
identity = json.loads((directory / "identity.json").read_text())
completed = json.loads((directory / "complete.json").read_text())
if completed["identity"] != digest(identity):
raise ValueError("Score identity mismatch")
rows = []
for name, checksum in completed["files"].items():
p = directory / name
if not p.resolve().is_relative_to(directory.resolve()) or sha256(p) != checksum:
raise ValueError("Score checksum mismatch")
record = json.loads(p.read_text())
rows.extend(record["rows"])
if len(rows) != completed["branches"]:
raise ValueError("Incomplete score set")
return rows, identity
def unit(row, temperature=1.0):
logits = row["letter_logits"]
task = row["task"]
gold = row["gold"]
if task in ("boolean", "entity", "multilabel"):
p = p_yes(logits, temperature)
return [1 - p, p], int(gold == 0)
width = row["n"] - 2 if row["head_key"].endswith("choiceR") else row["n"]
if not isinstance(gold, int) or not 0 <= gold < width:
return None, None
return listed_probs(logits, width, temperature).tolist(), gold
def fit_calibration(scores, output, *, panel_role):
if panel_role != "fit":
raise ValueError("Temperature fitting accepts fit panels only")
rows, identity = read_scores(scores)
if identity["panel_role"] != "fit":
raise ValueError("Scores were not generated from a fit panel")
output = Path(output)
if output.exists():
raise FileExistsError("Calibration artifacts are immutable")
temperatures, losses = {}, {}
for task in TASKS:
selected = [r for r in rows if r["task"] == task and unit(r)[0] is not None]
if not selected:
raise ValueError("No fit examples for " + task)
def loss(log_t, selected=selected):
t = float(np.exp(log_t))
return float(np.mean([-np.log(max(unit(r, t)[0][unit(r, t)[1]], 1e-300)) for r in selected]))
fit = minimize_scalar(loss, bounds=(np.log(0.05), np.log(20)), method="bounded")
temperatures[task] = float(np.exp(fit.x))
losses[task] = {"before": loss(0.0), "after": float(fit.fun), "units": len(selected)}
payload = {
"schema": "solomon-mlx-temperature-v1",
"runtime": identity["runtime"]["fingerprint"],
"temperatures": temperatures,
"fit_panel_sha256": identity["panel_sha256"],
"losses": losses,
"selection_role": "fit",
"heldout_used": False,
}
output.write_text(json.dumps({**payload, "sha256": digest(payload)}, indent=2))
return payload
def compare_rows(mlx_rows, cuda_rows, *, temperatures=None, reference_temperatures=None):
temperatures = temperatures or dict.fromkeys(TASKS, 1.0)
reference_temperatures = reference_temperatures or dict.fromkeys(TASKS, 1.0)
reference = {r["id"]: r for r in cuda_rows}
if len(reference) != len(cuda_rows) or set(reference) != {r["id"] for r in mlx_rows}:
raise ValueError("Comparison panels have different or duplicate branch IDs")
units = []
questions = defaultdict(list)
for row in mlx_rows:
other = {**row, "letter_logits": reference[row["id"]]["letter_logits"]}
p, gold = unit(row, temperatures[row["task"]])
q, _ = unit(other, reference_temperatures[row["task"]])
if p is None:
continue
left, right = int(np.argmax(p)), int(np.argmax(q))
item = {
"agreement": left == right,
"mlx_correct": left == gold,
"cuda_correct": right == gold,
"probability_drift": float(np.max(np.abs(np.asarray(p) - q))),
}
units.append(item)
questions[row["question_id"]].append(item)
if not units:
raise ValueError("No defined comparison targets")
agreement = float(np.mean([r["agreement"] for r in units]))
question_agreement = float(np.mean([all(x["agreement"] for x in r) for r in questions.values()]))
mlx_accuracy = float(np.mean([all(x["mlx_correct"] for x in r) for r in questions.values()]))
cuda_accuracy = float(np.mean([all(x["cuda_correct"] for x in r) for r in questions.values()]))
return {
"units": len(units),
"questions": len(questions),
"unit_decision_agreement": agreement,
"question_decision_agreement": float(
np.mean([all(x["agreement"] for x in r) for r in questions.values()])
),
"mlx_whole_question_accuracy": mlx_accuracy,
"cuda_whole_question_accuracy": cuda_accuracy,
"accuracy_degradation_percentage_points": 100 * (cuda_accuracy - mlx_accuracy),
"max_probability_drift": max(r["probability_drift"] for r in units),
"probability_comparison": {
"mlx_temperatures": temperatures,
"cuda_temperatures": reference_temperatures,
},
"mean_probability_drift": float(np.mean([r["probability_drift"] for r in units])),
"quality_gate_passed": agreement >= 0.999
and question_agreement >= 0.999
and cuda_accuracy - mlx_accuracy <= 0.0025,
}
def read_cuda_scores(directory, panel, reference_identity):
"""Reuse only scores bound to the exact pinned CUDA runtime and panel."""
jobs, manifest = load_panel(panel)
directory = Path(directory)
result = {}
for file in sorted(directory.glob("scores*.json.gz")):
payload = json.loads(gzip.decompress(file.read_bytes()))
identity = payload["identity"]
if not payload["complete"] or identity["runtime"] != reference_identity:
raise ValueError("Existing CUDA scores do not match the fresh reference runtime")
if identity["manifest"]["jobs_sha256"] != manifest["jobs_sha256"]:
raise ValueError("CUDA scores use another panel")
for key, row in payload["scores"].items():
if key in result:
raise ValueError("Duplicate CUDA score ID")
result[key] = row
if set(result) != {r["id"] for r in jobs}:
raise ValueError("CUDA score set is incomplete")
return [{**r, **result[r["id"]]} for r in jobs]
def select_calibration(fitted, dev_scores, output):
fitted, output = Path(fitted), Path(output)
if output.exists():
raise FileExistsError("Selected calibration is immutable")
fit = json.loads(fitted.read_text())
fit_payload = {k: v for k, v in fit.items() if k != "sha256"}
rows, identity = read_scores(dev_scores)
if (
fit["sha256"] != digest(fit_payload)
or identity["runtime"]["fingerprint"] != fit["runtime"]
or identity["panel_role"] != "dev"
):
raise ValueError("Calibration or development identity mismatch")
temperatures, selection = {}, {}
for task in TASKS:
selected = [r for r in rows if r["task"] == task and unit(r)[0] is not None]
if not selected:
raise ValueError("Missing development task " + task)
def loss(t, selected=selected):
values = [unit(r, t) for r in selected]
return float(np.mean([-np.log(max(p[g], 1e-300)) for p, g in values]))
original, candidate = loss(1.0), loss(fit["temperatures"][task])
temperatures[task] = fit["temperatures"][task] if candidate < original else 1.0
selection[task] = {"untempered_nll": original, "fit_temperature_nll": candidate}
payload = {
**fit_payload,
"temperatures": temperatures,
"selection_role": "dev_selected",
"fit_artifact_sha256": sha256(fitted),
"dev_panel_sha256": identity["panel_sha256"],
"development_selection": selection,
}
output.write_text(json.dumps({**payload, "sha256": digest(payload)}, indent=2))
return payload
def heldout_report(
scores,
cuda_directory,
panel,
calibration,
reference,
output,
*,
reference_binding="evaluations/cuda-acceptance/input/serving-binding.json",
):
"""Evaluate a frozen configuration once; an existing output cannot be replaced."""
output = Path(output)
if output.exists():
raise FileExistsError("Held-out report already exists; do not reuse it for selection")
rows, identity = read_scores(scores)
cal = json.loads(Path(calibration).read_text())
payload = {k: v for k, v in cal.items() if k != "sha256"}
if (
cal["sha256"] != digest(payload)
or cal["runtime"] != identity["runtime"]["fingerprint"]
or cal["selection_role"] != "dev_selected"
or identity["panel_role"] != "cert"
):
raise ValueError(
"Held-out evaluation requires frozen development-selected calibration and cert scores"
)
ref = json.loads(Path(reference).read_text())
cuda = read_cuda_scores(cuda_directory, panel, ref["identity"])
binding_path = Path(reference_binding)
source_manifest = json.loads((binding_path.parent / "manifest.json").read_text())
if sha256(binding_path) != source_manifest["files"][binding_path.name]:
raise ValueError("CUDA acceptance binding checksum mismatch")
binding = json.loads(binding_path.read_text())
for key in (
"adapter_sha256",
"trained_heads_sha256",
"model_sha256",
"numerics",
"placement",
"arithmetic",
):
if binding["runtime"][key] != ref["identity"][key]:
raise ValueError("CUDA calibration belongs to another reference runtime")
reference_temperatures = {task: binding["temperatures"]["models"][task]["temperature"] for task in TASKS}
report = {
**compare_rows(
rows, cuda, temperatures=cal["temperatures"], reference_temperatures=reference_temperatures
),
"cuda_calibration_binding_sha256": sha256(binding_path),
"runtime": identity["runtime"],
"panel_sha256": identity["panel_sha256"],
"calibration_sha256": sha256(calibration),
"reference_sha256": sha256(reference),
"scope": "text-only held-out panel",
"image_qualification": False,
}
output.write_text(json.dumps(report, indent=2))
return report
|