Datasets:
File size: 61,967 Bytes
b48c9f8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 | #!/usr/bin/env python3
"""Run stepmold's example packs against one real model, in two arms, and write JSONL.
STEPMOLD_API_KEY=... python3 bench.py \\
--model 'openai:http://localhost:8000/v1#Qwen/Qwen2.5-7B-Instruct' --label qwen2.5-7b
Arm `stepmold` is `stepmold eval`, case by case: every gold case goes through
`stepmold.eval.evaluate` with the model behind a recorder that writes each request down.
Arm `plain` runs the same walker over a copy of the pack with the checks taken out, behind
an adapter that sends no grammar and keeps the first JSON object of each answer. This is
the "unverified arm" of stepmold's tests/production/test_longhorizon.py, pointed at a real
model. README.md lists every difference between the arms.
Which node and stage a request belongs to, and what stepmold decided about it, is read
from stepmold's own structured log events (`stepmold.log`), routed to the case-run that
emitted them by thread. Nothing in stepmold is patched.
Standard library only. Python 3.9 or newer.
"""
import argparse
import dataclasses
import datetime
import hashlib
import importlib
import importlib.util
import json
import logging
import os
import platform
import re
import subprocess
import sys
import threading
import time
import uuid
from concurrent.futures import ThreadPoolExecutor, as_completed
from types import SimpleNamespace
BENCH_VERSION = "0.1.0"
# `plain_schema` is opt-in: the plain arm plus the node's schema written into the answer
# prompt, the way stepmold's own json_object grammar mode writes it. See README.md.
ARMS = ("stepmold", "plain", "plain_schema")
DEFAULT_ARMS = "stepmold,plain"
STEPMOLD_GIT = "https://gitlab.com/ahkamboh/stepmold"
DEFAULT_STEPMOLD = os.path.join(os.path.expanduser("~"), "Documents", "GitHub", "stepmold")
# The host tools an example pack needs, relative to its own directory. A pack only names
# its actions (stepmold.tools); the host decides what those names run. For the shipped
# examples the bench is the host, so it says so here. `--tools PACK=SPEC` overrides.
PACK_TOOLS = {"refund_desk": "tools.py:registry"}
# The schema a plain-arm generate node is verified against: "is an object". The adapter
# always hands back an object, so verification never rejects anything in that arm. It is
# truthy on purpose: stepmold's codegen passes a truthy grammar to the answer call and
# None to the think call, which is how the adapter tells the two apart. It is never sent.
PLAIN_GRAMMAR = {"type": "object"}
# Request fields the bench will not take through --extra-body, and why.
REFUSED_EXTRA_BODY = {
"model": "the model comes from --model",
"messages": "the prompt is the pack's, rendered by stepmold",
"max_tokens": "every node sets its own budget; use --reasoning-reserve for headroom",
"seed": "stepmold varies the seed per re-sample and per gate draw; a fixed seed would "
"make every draw the same request",
"n": "stepmold reads one choice per request",
"stream": "the backend reads one JSON body per request",
}
# Constrained decoding belongs to the stepmold arm's --grammar-mode. Sent through
# --extra-body it would constrain the plain arm too.
CONSTRAINT_KEYS = ("response_format", "json_schema", "grammar", "guided_json", "guided_regex",
"guided_choice", "guided_grammar", "guided_decoding_backend",
"structured_outputs")
class BenchError(Exception):
"""A setup problem the operator has to fix. Printed, never a traceback."""
# --------------------------------------------------------------------------- stepmold
_SM = None
def load_stepmold(path=None):
"""Import stepmold once: from `path` when given, else whatever is installed.
Returns a namespace of the stepmold modules the bench uses. Bytecode writing is
turned off first, so running the bench leaves a stepmold checkout byte-for-byte as it
was.
"""
global _SM
if _SM is not None:
return _SM
sys.dont_write_bytecode = True
root = None
if path:
root = os.path.abspath(os.path.expanduser(path))
if not os.path.isfile(os.path.join(root, "stepmold", "__init__.py")):
raise BenchError("--stepmold %s: there is no stepmold/__init__.py there" % root)
if root not in sys.path:
sys.path.insert(0, root)
try:
package = importlib.import_module("stepmold")
except ImportError:
raise BenchError(
"stepmold is not importable. Pass --stepmold PATH to a checkout, or "
"pip install git+%s" % STEPMOLD_GIT
)
where = os.path.dirname(os.path.abspath(package.__file__))
if root and not where.startswith(root + os.sep):
raise BenchError(
"stepmold was already imported from %s, not from --stepmold %s" % (where, root)
)
names = ("errors", "eval", "graph", "verify", "codegen", "grammar", "pack", "cli",
"log", "model")
modules = {name: importlib.import_module("stepmold." + name) for name in names}
modules["openai"] = importlib.import_module("stepmold.backends.openai_compat")
_SM = SimpleNamespace(package=package, root=os.path.dirname(where), **modules)
return _SM
def stepmold_provenance(sm):
"""Which stepmold produced these numbers: version, commit, and where it came from."""
info = {"version": getattr(sm.package, "__version__", None), "path": sm.root,
"commit": None, "dirty": None, "source": "checkout"}
if os.path.isdir(os.path.join(sm.root, ".git")):
info["commit"] = _git(sm.root, "rev-parse", "HEAD")
status = _git(sm.root, "status", "--porcelain", "--untracked-files=no")
info["dirty"] = None if status is None else bool(status)
return info
info["source"] = "installed"
try:
from importlib import metadata
direct = json.loads(metadata.distribution("stepmold").read_text("direct_url.json")
or "{}")
info["commit"] = (direct.get("vcs_info") or {}).get("commit_id")
info["url"] = direct.get("url")
except Exception:
pass
return info
def _git(root, *args):
try:
out = subprocess.run(("git", "-C", root) + args, capture_output=True, text=True,
timeout=10)
except (OSError, subprocess.SubprocessError):
return None
return out.stdout.strip() if out.returncode == 0 else None
# ------------------------------------------------------------------- lenient parsing
def _refuse_constant(name):
raise ValueError("%s is not JSON" % name)
_DECODER = json.JSONDecoder(parse_constant=_refuse_constant)
def first_json_object(text):
"""The first JSON object in `text`, reading left to right, or None.
Lenient about where the object is (prose before it, a code fence, text after it),
strict about what it is: standard JSON, no NaN or Infinity. A `{` that does not
start a valid object is skipped and the search moves on.
This is the opposite end from stepmold's own `verify.extract_json`, which takes the
last object on purpose. The plain arm is the pipeline people write first.
"""
if not isinstance(text, str):
return None
index = text.find("{")
while index != -1:
try:
value, _ = _DECODER.raw_decode(text, index)
except (ValueError, RecursionError):
value = None
if isinstance(value, dict):
return value
index = text.find("{", index + 1)
return None
# ------------------------------------------------------------------------- recording
_LOCAL = threading.local()
def _current():
return getattr(_LOCAL, "recorder", None)
def _now():
return datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%dT%H:%M:%S.%fZ")
def _sha(text):
return hashlib.sha256(text.encode("utf-8")).hexdigest()[:16]
class Recorder:
"""What one case-run leaves behind: its requests, and what stepmold said about them.
Requests are written down by the model wrappers below. Everything else (which node
and stage a request served, rungs, gate draws, rejections, the path, the escalation
edges, the end node) comes from stepmold's own log events, delivered here by
`EventRouter` because the case-run's thread emitted them.
"""
def __init__(self, stamp, pack, sm, record_prompts=False, record_reasoning=False):
self.stamp = stamp
self.pack = pack # the ORIGINAL pack: grammars, asserts, node types
self.sm = sm
self.record_prompts = record_prompts
self.record_reasoning = record_reasoning
self.arm = stamp["arm"]
self.calls = []
self.path = []
self.escalations = []
self.gates = []
self.retries = 0
self.end_node = None
self.steps = None
self.generations = None
self.run_error = None
self._stage = None
self._node = None
self._corrected = False
self._rung = 0
self._draw = 0
self._failed = None
self._http_attempt = None
self._http_errors = 0
self._started = None
# ---------------------------------------------------------------- stepmold events
def on_event(self, name, fields):
node = fields.get("node")
if name == "node.think":
self._stage, self._node, self._corrected = "think", node, False
elif name == "node.emit":
self._stage, self._node = "answer", node
self._corrected = bool(fields.get("corrected"))
elif name == "node.retry":
self.retries += 1
self._rung = max(0, int(fields.get("attempt") or 1) - 1)
elif name == "node.rejected":
self._rejected(fields)
elif name == "node.sample":
self._draw = int(fields.get("draw") or 0)
self._rung = 0
elif name in ("node.agreed", "node.unsure"):
self.gates.append({
"node": node, "outcome": name.split(".")[1],
"agreed": fields.get("agreed"), "drawn": fields.get("of"),
"required": fields.get("required"), "asked": fields.get("asked"),
"distinct": fields.get("distinct"), "generations": fields.get("generations"),
})
elif name == "node.ok":
self.path.append(node)
self._reset()
elif name == "node.failed":
self.path.append(node)
self._failed = dict(fields)
self._reset()
elif name == "node.assert":
self.path.append(node)
if not fields.get("passed"):
# An assert node with an `on_fail` diverts here. One without raises
# instead and leaves `node.failed`, not this event.
self.escalations.append({
"node": node, "node_type": "assert", "edge": "on_fail",
"to": fields.get("to"), "error": "assert",
"reason": "expr was false or could not be evaluated: %s"
% fields.get("expr"),
})
elif name in ("edge.on_fail", "edge.on_unsure"):
failed = self._failed if (self._failed or {}).get("node") == node else {}
target = self.pack.nodes.get(node)
self.escalations.append({
"node": node, "node_type": getattr(target, "type", None),
"edge": name.split(".")[1], "to": fields.get("to"),
"error": failed.get("error"), "reason": failed.get("reason"),
})
elif name == "run.end":
self.end_node = fields.get("end_node")
self.path.append(self.end_node)
self.steps = fields.get("steps")
self.generations = fields.get("generations")
elif name == "run.error":
self.run_error = {"node": node, "error": fields.get("error"),
"reason": fields.get("reason")}
elif name == "backend.response":
self._http_attempt = fields.get("attempt")
elif name in ("backend.http_error", "backend.transport_error"):
self._http_errors += 1
def _reset(self):
self._stage = self._node = None
self._rung = self._draw = 0
self._corrected = False
def _rejected(self, fields):
"""stepmold refused the request just made. It is always the latest one."""
if not self.calls:
return
call = self.calls[-1]
call["accepted"] = False
call["reject_reason"] = fields.get("reason")
if fields.get("cause") == "empty_completion":
call["reject_kind"] = "empty"
elif not call.get("parsed_ok"):
call["reject_kind"] = "parse"
elif call.get("schema_ok") is False:
call["reject_kind"] = "schema"
else:
call["reject_kind"] = "assert"
# ---------------------------------------------------------------------- requests
def start_call(self, prompt, grammar_sent, max_tokens, temperature=None, seed=None):
# The stage normally comes from the event stepmold emits just before the call.
# Without one, a grammar on the call is the only other sign of an answer call.
stage = self._stage or ("answer" if grammar_sent else "think")
call = {"kind": "call"}
call.update(self.stamp)
call.update({
"seq": len(self.calls), "node": self._node, "stage": stage,
"attempt": self._rung + 1, "draw": self._draw + 1,
"corrected": bool(self._corrected) if stage == "answer" else False,
"grammar_sent": bool(grammar_sent), "max_tokens": max_tokens,
"temperature": temperature, "seed": seed,
"prompt_sha256": _sha(prompt or ""), "prompt_chars": len(prompt or ""),
"timestamp": _now(),
})
if self.record_prompts:
call["prompt"] = prompt
self._stage = None
self._http_attempt = None
self._http_errors = 0
self.calls.append(call)
return call
def end_call(self, call, started, text=None, response=None, error=None):
call["latency_ms"] = round((time.monotonic() - started) * 1000.0, 1)
if self._http_attempt is not None:
call["http_attempts"] = self._http_attempt
elif response is not None or self._http_errors:
call["http_attempts"] = self._http_errors + (1 if response is not None else 0)
else:
call["http_attempts"] = None
usage, choice, message = {}, {}, {}
if isinstance(response, dict):
usage = response.get("usage") or {}
choices = response.get("choices") or []
choice = choices[0] if choices and isinstance(choices[0], dict) else {}
message = choice.get("message") or {}
details = usage.get("completion_tokens_details") or {}
reasoning = message.get("reasoning") or message.get("reasoning_content")
call.update({
"prompt_tokens": usage.get("prompt_tokens"),
"completion_tokens": usage.get("completion_tokens"),
"reasoning_tokens": details.get("reasoning_tokens"),
"total_tokens": usage.get("total_tokens"),
"finish_reason": choice.get("finish_reason"),
"output": text,
"output_chars": len(text) if isinstance(text, str) else None,
"reasoning_chars": len(reasoning) if isinstance(reasoning, str) else None,
"parsed_ok": None, "schema_ok": None, "assert_ok": None, "accepted": None,
"reject_kind": None, "reject_reason": None, "error": None,
})
if self.record_reasoning:
call["reasoning"] = reasoning if isinstance(reasoning, str) else None
if error is not None:
call["error"] = "%s: %s" % (type(error).__name__, _clip(str(error), 500))
if call["stage"] == "answer" and self.arm == "stepmold":
# The first two of stepmold's three checks, run with stepmold's own functions
# on the text it is about to verify. The verdict that counts arrives as an
# event (`node.rejected`); these say which check a rejection came from.
value = None
parsed = False
if isinstance(text, str):
try:
value = self.sm.verify.extract_json(text)
parsed = isinstance(value, dict)
except Exception:
parsed = False
call["parsed_ok"] = parsed
if parsed:
call["schema_ok"] = self._schema_ok(call["node"], value)
def plain_answer(self, value, empty):
"""The plain adapter's verdict on the answer it just read. Nothing is rejected."""
if not self.calls:
return
call = self.calls[-1]
call["parsed_ok"] = value is not None
call["accepted"] = value is not None
if value is None:
call["reject_kind"] = "empty" if empty else "parse"
call["reject_reason"] = ("the reply was empty; nothing committed" if empty
else "no JSON object in the reply; nothing committed")
else:
# Diagnostic only: what stepmold's schema check would have said. The plain
# arm commits the object either way.
call["schema_ok"] = self._schema_ok(call.get("node"), value)
def _schema_ok(self, node_name, value):
node = self.pack.nodes.get(node_name)
if node is None:
return None
try:
self.sm.grammar.validate_against(node.grammar or {}, value)
except self.sm.grammar.ValidationError:
return False
except Exception:
return None
return True
def finalize(self):
"""Settle every verdict that no event contradicted."""
for call in self.calls:
if call.get("stage") != "answer" or self.arm != "stepmold":
continue
if call.get("accepted") is None and not call.get("error"):
call["accepted"] = True
node = self.pack.nodes.get(call.get("node"))
if call.get("parsed_ok") and call.get("schema_ok") and node is not None \
and node.assert_expr:
call["assert_ok"] = call.get("reject_kind") != "assert"
def _clip(text, limit):
text = " ".join(str(text).split())
return text if len(text) <= limit else text[:limit] + "..."
class EventRouter(logging.Handler):
"""Hands each stepmold log event to the recorder of the thread that emitted it."""
def __init__(self, event_attr, fields_attr):
logging.Handler.__init__(self, level=logging.DEBUG)
self.event_attr = event_attr
self.fields_attr = fields_attr
def emit(self, record):
recorder = _current()
if recorder is None:
return
name = getattr(record, self.event_attr, None)
if name is None:
return
try:
recorder.on_event(name, getattr(record, self.fields_attr, None) or {})
except Exception:
# Bookkeeping must never change the run it is watching.
pass
class capture_events(object):
"""Route stepmold's events to recorders while the block runs, then put logging back.
stepmold logs nothing unless asked (rule 1 of `stepmold.log`), so the package logger
is set to DEBUG here, which is what makes `node.think` and `node.emit` fire. It stops
propagating meanwhile, so a host that configured the root logger is not flooded.
"""
def __init__(self, sm):
self.sm = sm
self.logger = logging.getLogger(sm.log.ROOT)
self.handler = EventRouter(getattr(sm.log, "EVENT_ATTR", "stepmold_event"),
getattr(sm.log, "FIELDS_ATTR", "stepmold_fields"))
def __enter__(self):
self.saved = (self.logger.level, self.logger.propagate)
self.logger.addHandler(self.handler)
self.logger.setLevel(logging.DEBUG)
self.logger.propagate = False
return self
def __exit__(self, *exc):
self.logger.removeHandler(self.handler)
self.logger.setLevel(self.saved[0])
self.logger.propagate = self.saved[1]
return False
# --------------------------------------------------------------------------- models
_BACKEND = {}
def backend_class(sm):
"""stepmold's OpenAICompatModel with one change: each request is written down.
`generate` here is the parent's two lines (build the payload, post it, read the
content) with a recorder around them, so the request on the wire is byte-for-byte
the one stepmold would send. The signature keeps `sampling`, so stepmold still sends
its re-sample and gate-draw hints.
"""
if "class" in _BACKEND:
return _BACKEND["class"]
base = sm.openai.OpenAICompatModel
content_of = sm.openai._content
class BenchBackend(base):
def generate(self, prompt, grammar=None, max_tokens=512, sampling=None):
payload = self.build_payload(prompt, grammar, max_tokens, sampling=sampling)
recorder = _current()
call = None
if recorder is not None:
messages = payload.get("messages") or [{}]
call = recorder.start_call(
messages[0].get("content"),
grammar_sent=any(key in payload for key in CONSTRAINT_KEYS),
max_tokens=payload.get("max_tokens"),
temperature=payload.get("temperature"), seed=payload.get("seed"))
started = time.monotonic()
response = None
try:
response = self._post(payload)
text = content_of(response)
except BaseException as exc:
if call is not None:
recorder.end_call(call, started, response=response, error=exc)
raise
if call is not None:
recorder.end_call(call, started, text=text, response=response)
return text
_BACKEND["class"] = BenchBackend
return BenchBackend
class RecordingModel(object):
"""A transparent recorder around a model without a `sampling` keyword (FakeModel)."""
def __init__(self, inner):
self.inner = inner
def generate(self, prompt, grammar=None, max_tokens=512):
return _recorded(self.inner, prompt, grammar, max_tokens, None)
class RecordingSampledModel(RecordingModel):
"""The same, for a model that takes stepmold's per-call sampling hint."""
def generate(self, prompt, grammar=None, max_tokens=512, sampling=None):
return _recorded(self.inner, prompt, grammar, max_tokens, sampling)
def _recorded(inner, prompt, grammar, max_tokens, sampling):
recorder = _current()
call = None
if recorder is not None:
call = recorder.start_call(
prompt, grammar_sent=grammar is not None, max_tokens=max_tokens,
temperature=getattr(sampling, "temperature", None),
seed=getattr(sampling, "seed", None))
started = time.monotonic()
try:
if sampling is None:
text = inner.generate(prompt, grammar=grammar, max_tokens=max_tokens)
else:
text = inner.generate(prompt, grammar=grammar, max_tokens=max_tokens,
sampling=sampling)
except BaseException as exc:
if call is not None:
recorder.end_call(call, started, error=exc)
raise
if call is not None:
recorder.end_call(call, started, text=text)
return text
def recording(model, sm):
if sm.codegen.accepts_sampling(model):
return RecordingSampledModel(model)
return RecordingModel(model)
class PlainModel(object):
"""The plain arm's model: no grammar, one draw, and the first JSON object wins.
stepmold's walker calls this exactly as it calls any model. The think call arrives
with grammar None and its text goes back untouched (an empty one becomes empty
notes). The answer call arrives with the plain pack's permissive grammar, which is
dropped: the request goes out with grammar None. Its reply is reduced to its first
JSON object, re-serialised, so stepmold's verify (which checks only "is an object" in
the plain pack) accepts it. A reply with no JSON object becomes `{}`, which commits
nothing, so the walk goes on with whatever state it has.
No `sampling` keyword: the plain pack never re-samples and never draws twice, so
stepmold never has a hint to send.
With `schema_hint` (the plain_schema arm), the answer prompt also gets the node's
real schema appended, in the words stepmold's json_object grammar mode uses. The
text is built by stepmold's own backend (`build_payload` in json_object mode), so it
is exactly what stepmold would add, and nothing is sent but the text.
"""
def __init__(self, inner, sm, schema_hint=False):
self.inner = inner
self.sm = sm
self.schema_hint = schema_hint
def _hinted(self, prompt):
recorder = _current()
node = recorder.pack.nodes.get(recorder._node) if recorder is not None else None
if node is None or not node.grammar:
# Without the node there is no schema to show, and an unhinted call would
# quietly turn this arm into the plain one. Not a result: an infra error.
raise BenchError("plain_schema: could not tell which node this answer is for")
writer = _hint_writer(self.sm)
grammar = self.sm.grammar.schema_to_grammar(node.grammar)
return writer.build_payload(prompt, grammar, 1)["messages"][0]["content"]
def generate(self, prompt, grammar=None, max_tokens=512):
answer = grammar is not None
empty = False
if answer and self.schema_hint:
prompt = self._hinted(prompt)
try:
text = self.inner.generate(prompt, grammar=None, max_tokens=max_tokens)
except self.sm.errors.BackendError as exc:
if not getattr(exc, "empty_content", False):
raise
text, empty = None, True
if not answer:
return text if isinstance(text, str) else ""
value = first_json_object(text)
recorder = _current()
if recorder is not None:
recorder.plain_answer(value, empty)
return json.dumps(value) if value is not None else "{}"
_HINT = {}
def _hint_writer(sm):
"""An OpenAICompatModel in json_object mode, used only to write stepmold's schema text.
Constructing one performs no IO, and only `build_payload` is ever called on it.
"""
if "writer" not in _HINT:
_HINT["writer"] = sm.openai.OpenAICompatModel(
base_url="http://unused.invalid/v1", model="unused", api_key="",
grammar_mode="json_object")
return _HINT["writer"]
def plain_pack(pack):
"""The pack the plain arm walks: same graph, prompts and edges, no checks.
Every generate node loses its schema (replaced by PLAIN_GRAMMAR), its `assert:`, its
retries, its gate (`samples`/`agree`) and its `on_fail`/`on_unsure` edges, which
only the retry ladder and the gate can take. `two_stage`, both prompts, `max_tokens`
and `output` stay. Tool, assert and end nodes, the edges and `max_steps` are
untouched.
"""
nodes = {}
for name, node in pack.nodes.items():
if node.type == "generate":
node = dataclasses.replace(
node, grammar=dict(PLAIN_GRAMMAR), assert_expr=None, retries=0,
samples=None, agree=None, on_fail=None, on_unsure=None)
nodes[name] = node
return dataclasses.replace(pack, nodes=nodes)
# ----------------------------------------------------------------------- model specs
def parse_model_spec(spec, grammar_modes):
"""Read a model spec the way `stepmold.cli.resolve_model` does.
openai:<base_url>#<model>[#<grammar_mode>[#<reasoning_reserve>]]
fake:<path> a scripted FakeModel, resolved inside each pack
"""
scheme, _, rest = (spec or "").partition(":")
if scheme == "fake":
if not rest:
raise BenchError("fake: needs a script path, e.g. fake:fakes/script.json")
return SimpleNamespace(scheme="fake", spec=spec, model=spec, base_url=None,
grammar_mode=None, reasoning_reserve=None)
if scheme != "openai":
raise BenchError("--model: unknown scheme %r (known: openai, fake)" % scheme)
parts = rest.split("#")
base_url = parts[0].strip()
name = parts[1].strip() if len(parts) > 1 else ""
if not base_url or not name:
raise BenchError("--model: openai: needs a base url and a model name, e.g. "
"openai:http://localhost:8000/v1#Qwen/Qwen2.5-7B-Instruct")
mode = parts[2].strip() if len(parts) > 2 and parts[2].strip() else None
if mode is not None and mode not in grammar_modes:
raise BenchError("--model: unknown grammar mode %r (known: %s)"
% (mode, ", ".join(grammar_modes)))
reserve = None
if len(parts) > 3 and parts[3].strip():
try:
reserve = int(parts[3].strip())
except ValueError:
raise BenchError("--model: reasoning reserve must be an integer, got %r"
% parts[3].strip())
return SimpleNamespace(scheme="openai", spec=spec, model=name, base_url=base_url,
grammar_mode=mode, reasoning_reserve=reserve)
def _settle(flag_value, spec_value, default, what):
if flag_value is not None and spec_value is not None and flag_value != spec_value:
raise BenchError("--%s %r disagrees with the %s in --model (%r); give one"
% (what, flag_value, what.replace("-", " "), spec_value))
if flag_value is not None:
return flag_value
return spec_value if spec_value is not None else default
def parse_extra_body(text):
"""--extra-body as a dict, with `temperature` split out (see README)."""
if not text:
return {}, None
try:
body = json.loads(text)
except ValueError as exc:
raise BenchError("--extra-body is not valid JSON (%s)" % exc)
if not isinstance(body, dict):
raise BenchError("--extra-body must be a JSON object")
for key in sorted(body):
if key in REFUSED_EXTRA_BODY:
raise BenchError("--extra-body may not set %r: %s" % (key, REFUSED_EXTRA_BODY[key]))
if key in CONSTRAINT_KEYS:
raise BenchError(
"--extra-body may not set %r: constrained decoding is the stepmold arm's "
"--grammar-mode, and through --extra-body it would constrain the plain arm "
"too" % key)
temperature = body.pop("temperature", None)
if temperature is not None and (isinstance(temperature, bool)
or not isinstance(temperature, (int, float))):
raise BenchError("--extra-body temperature must be a number")
return body, temperature
# --------------------------------------------------------------------------- records
def case_key(record):
return (record.get("label"), record.get("arm"), record.get("pack"), record.get("case"),
record.get("run"))
def read_records(path):
"""Every well-formed record in a JSONL file. A torn last line is skipped."""
if not os.path.exists(path):
return
with open(path, "r", encoding="utf-8") as handle:
for line in handle:
line = line.strip()
if not line:
continue
try:
record = json.loads(line)
except ValueError:
continue
if isinstance(record, dict):
yield record
def _jsonable(value):
if isinstance(value, float) and (value != value or value in (float("inf"), float("-inf"))):
return repr(value)
if isinstance(value, dict):
return dict((str(key), _jsonable(item)) for key, item in value.items())
if isinstance(value, (list, tuple)):
return [_jsonable(item) for item in value]
return value
def dumps(record):
try:
return json.dumps(record, ensure_ascii=False, allow_nan=False, default=str)
except ValueError:
return json.dumps(_jsonable(record), ensure_ascii=False, allow_nan=False,
default=str)
class JsonlWriter(object):
"""Appends whole case-runs to the results file, one locked write each."""
def __init__(self, path, secret=None):
self.path = path
# Belt and braces: no record carries the key, and if one ever did it is cut out
# here. Too short a string would match ordinary text, so only a real key counts.
self.secret = secret if secret and len(secret) >= 8 else None
self.lock = threading.Lock()
parent = os.path.dirname(os.path.abspath(path))
os.makedirs(parent, exist_ok=True)
# A run killed mid-write can leave a torn last line. Start ours on a fresh one.
if os.path.exists(path) and os.path.getsize(path):
with open(path, "rb") as handle:
handle.seek(-1, os.SEEK_END)
torn = handle.read(1) != b"\n"
if torn:
with open(path, "a", encoding="utf-8") as handle:
handle.write("\n")
self.handle = open(path, "a", encoding="utf-8")
def write(self, records):
text = "".join(dumps(record) + "\n" for record in records)
if self.secret and self.secret in text:
text = text.replace(self.secret, "<redacted>")
with self.lock:
self.handle.write(text)
self.handle.flush()
try:
os.fsync(self.handle.fileno())
except OSError:
pass
def close(self):
self.handle.close()
def _sum(calls, key):
values = [call.get(key) for call in calls if isinstance(call.get(key), (int, float))]
return sum(values) if values else None
def pack_digest(path):
"""A fingerprint of a pack's files, so a result names exactly what it ran."""
digest = hashlib.sha256()
for folder, dirs, files in os.walk(path):
dirs[:] = sorted(d for d in dirs if d != "__pycache__")
for name in sorted(files):
if name.endswith(".pyc"):
continue
full = os.path.join(folder, name)
digest.update(os.path.relpath(full, path).encode("utf-8") + b"\0")
with open(full, "rb") as handle:
digest.update(handle.read())
digest.update(b"\0")
return digest.hexdigest()[:16]
def load_tools(spec):
"""Execute a host tools file afresh and return its registry.
Fresh every time on purpose: refund_desk's registry closes over an order store that
remembers refunds, and a second refund of the same order raises. Each case-run gets
its own store, so no case-run can see another's side effects.
"""
target, _, attribute = spec.partition(":")
full = os.path.abspath(target)
if not os.path.isfile(full):
raise BenchError("tools: no such file %s" % full)
module_spec = importlib.util.spec_from_file_location(
"_bench_tools_%s" % uuid.uuid4().hex[:12], full)
module = importlib.util.module_from_spec(module_spec)
module_spec.loader.exec_module(module)
names = [attribute] if attribute else ["registry", "REGISTRY"]
for name in names:
found = getattr(module, name, None)
if found is not None:
if callable(found) and not hasattr(found, "get"):
found = found()
return found
raise BenchError("tools: %s defines none of %s" % (full, ", ".join(names)))
# ------------------------------------------------------------------------------ bench
class Bench(object):
def __init__(self, args, sm, opener=None, sleeper=None, stream=None):
self.args = args
self.sm = sm
self.opener = opener
self.sleeper = sleeper
self.stream = stream or sys.stderr
self.print_lock = threading.Lock()
self.stop = threading.Event()
modes = tuple(sm.openai.GRAMMAR_MODES)
self.spec = parse_model_spec(args.model, modes)
self.grammar_mode = _settle(args.grammar_mode, self.spec.grammar_mode,
"response_format", "grammar-mode")
if self.grammar_mode not in modes:
raise BenchError("--grammar-mode %r is not one of %s"
% (self.grammar_mode, ", ".join(modes)))
self.reasoning_reserve = _settle(args.reasoning_reserve, self.spec.reasoning_reserve,
0, "reasoning-reserve")
self.extra_body, self.temperature = parse_extra_body(args.extra_body)
if self.temperature is None:
self.temperature = 0.0
self.arms = [arm.strip() for arm in args.arms.split(",") if arm.strip()]
unknown = [arm for arm in self.arms if arm not in ARMS]
if unknown or not self.arms:
raise BenchError("--arms: unknown arm(s) %s (known: %s)"
% (", ".join(unknown) or "none given", ", ".join(ARMS)))
if args.runs < 1 or args.concurrency < 1:
raise BenchError("--runs and --concurrency must be at least 1")
self.api_key = os.environ.get("STEPMOLD_API_KEY") or ""
default_label = "fake" if self.spec.scheme == "fake" else self.spec.model.split("/")[-1]
self.label = args.label or re.sub(r"[^A-Za-z0-9._-]+", "_", default_label).strip("_")
if not self.label:
raise BenchError("--label is required for this model spec")
self.out = args.out or os.path.join(
"results", re.sub(r"[^A-Za-z0-9._-]+", "_", self.label) + ".jsonl")
self.examples = self._examples_dir(args.examples)
self.packs = self._load_packs(args.packs, args.tools or [])
if self.spec.scheme == "openai":
# Built once now so a bad --extra-body or grammar mode fails before any request.
for arm in self.arms:
self._backend(arm)
self.jobs = self._plan(args.limit, args.runs)
# ---------------------------------------------------------------------- setup
def _examples_dir(self, given):
candidates = [given] if given else [os.path.join(self.sm.root, "examples")]
for candidate in candidates:
if candidate and os.path.isdir(candidate):
return os.path.abspath(candidate)
raise BenchError(
"no example packs at %s. pip install does not ship examples/; clone %s and "
"pass --stepmold (or --examples) pointing at the clone"
% (candidates[0], STEPMOLD_GIT))
def _load_packs(self, packs_arg, tools_args):
available = sorted(
name for name in os.listdir(self.examples)
if os.path.isfile(os.path.join(self.examples, name, "manifest.yaml")))
if packs_arg in (None, "", "all"):
chosen = available
else:
chosen = [name.strip() for name in packs_arg.split(",") if name.strip()]
missing = [name for name in chosen if name not in available]
if missing:
raise BenchError("--packs: no pack %s in %s (have: %s)"
% (", ".join(missing), self.examples, ", ".join(available)))
overrides = {}
for item in tools_args:
name, sep, spec = item.partition("=")
if not sep or not spec:
raise BenchError("--tools takes PACK=FILE.py[:ATTR], got %r" % item)
overrides[name.strip()] = spec.strip()
packs = {}
for name in chosen:
path = os.path.join(self.examples, name)
tools_spec = overrides.get(name)
if tools_spec is None and name in PACK_TOOLS:
target, _, attribute = PACK_TOOLS[name].partition(":")
tools_spec = os.path.join(path, target) + (":" + attribute if attribute else "")
try:
registry = load_tools(tools_spec) if tools_spec else None
pack = self.sm.pack.load_pack(path, tools=registry)
checker = getattr(self.sm.cli, "_check_output_shapes", None)
if checker is not None:
checker(pack)
except BenchError:
raise
except Exception as exc:
raise BenchError("pack %s does not load: %s: %s"
% (name, type(exc).__name__, exc))
if registry is None and any(n.type == "tool" for n in pack.nodes.values()):
raise BenchError("pack %s has tool nodes; pass --tools %s=FILE.py[:ATTR]"
% (name, name))
if not pack.evalset:
raise BenchError("pack %s has no evalset cases" % name)
packs[name] = SimpleNamespace(name=pack.name, dir=name, path=path, pack=pack,
plain=plain_pack(pack), tools=tools_spec,
digest=pack_digest(path))
return packs
def _plan(self, limit, runs):
jobs = []
for run in range(1, runs + 1):
for name in sorted(self.packs):
entry = self.packs[name]
cases = list(entry.pack.evalset)
if limit:
cases = cases[:limit]
seen = {}
for index, case in enumerate(cases, start=1):
label = case.name or "case %d" % index
seen[label] = seen.get(label, 0) + 1
if seen[label] > 1: # names are the resume key, so keep them unique
label = "%s #%d" % (label, index)
for arm in self.arms:
jobs.append(SimpleNamespace(arm=arm, pack=name, case=case,
case_name=label, case_index=index,
run=run, done=False))
return jobs
def setup(self):
"""What has to stay fixed for results to be resumed into the same label."""
return {
"model": self.spec.model, "base_url": self.spec.base_url,
"scheme": self.spec.scheme, "grammar_mode": self.grammar_mode,
"reasoning_reserve": self.reasoning_reserve, "temperature": self.temperature,
"extra_body": self.extra_body, "stepmold_version": self.provenance.get("version"),
"stepmold_commit": self.provenance.get("commit"),
}
def pack_digests(self):
return dict((entry.name, entry.digest) for entry in self.packs.values())
def _check_resume(self):
"""Refuse to mix results from two different setups under one label.
More packs, more runs or another --limit may be added to a label later. A
different model, grammar mode, extra body, reserve, temperature, stepmold commit,
or a pack whose files changed, may not.
"""
if self.args.allow_config_change:
return
setup, digests = self.setup(), self.pack_digests()
for record in read_records(self.out):
if record.get("kind") != "run" or record.get("label") != self.label:
continue
changed = [key for key in sorted(setup)
if (record.get("setup") or {}).get(key) != setup[key]]
changed += ["pack %s" % name for name, digest in sorted(digests.items())
if (record.get("pack_digests") or {}).get(name, digest) != digest]
if changed:
raise BenchError(
"%s already holds results for label %r made with a different setup "
"(%s differs). Use another --label or --out, or pass "
"--allow-config-change." % (self.out, self.label, ", ".join(changed)))
# ---------------------------------------------------------------------- models
def _backend(self, arm):
options = dict(
base_url=self.spec.base_url, model=self.spec.model, api_key=self.api_key,
grammar_mode=self.grammar_mode if arm == "stepmold" else "none",
reasoning_reserve=self.reasoning_reserve, temperature=self.temperature,
timeout=self.args.timeout, extra_body=dict(self.extra_body))
if self.args.http_retries is not None:
options["max_retries"] = self.args.http_retries
if self.opener is not None:
options["opener"] = self.opener
if self.sleeper is not None:
options["sleeper"] = self.sleeper
try:
return backend_class(self.sm)(**options)
except ValueError as exc:
raise BenchError(str(exc))
def _model(self, arm, shared):
if self.spec.scheme == "fake":
model = recording(shared, self.sm)
else:
model = self._backend(arm)
if arm == "stepmold":
return model
return PlainModel(model, self.sm, schema_hint=(arm == "plain_schema"))
# ------------------------------------------------------------------------- run
def run(self):
self.provenance = stepmold_provenance(self.sm)
self._check_resume()
done = set(case_key(record) for record in read_records(self.out)
if record.get("kind") == "case")
for job in self.jobs:
job.done = (self.label, job.arm, self.packs[job.pack].name, job.case_name,
job.run) in done
todo = sum(1 for job in self.jobs if not job.done)
self._say("%s: %d case-runs planned (%s x %d run%s), %d already in %s, %d to run"
% (self.label, len(self.jobs), "+".join(self.arms), self.args.runs,
"" if self.args.runs == 1 else "s", len(self.jobs) - todo, self.out,
todo))
if self.args.dry_run or not todo:
return 0
self.invocation = uuid.uuid4().hex[:12]
self.writer = JsonlWriter(self.out, secret=self.api_key)
self.writer.write([self._run_record()])
self.totals = {}
self.finished = 0
self.todo = todo
units = self._units()
try:
with capture_events(self.sm):
pool = ThreadPoolExecutor(max_workers=self.args.concurrency)
futures = [pool.submit(self._run_unit, unit) for unit in units]
try:
for future in as_completed(futures):
try:
future.result()
except Exception as exc: # one unit failing stops only that unit
self._say("bench: a worker stopped: %s: %s"
% (type(exc).__name__, _clip(str(exc), 300)))
except KeyboardInterrupt:
self.stop.set()
pool.shutdown(wait=False, cancel_futures=True)
self._say("interrupted: finished case-runs are saved in %s; run the "
"same command again to resume" % self.out)
return 130
pool.shutdown(wait=True)
finally:
self.writer.close()
self._report_totals()
return 0
def _units(self):
"""What one worker runs in order. Real models: one case-run each.
A scripted fake is one FakeModel per (arm, pack, run), consumed in evalset order
the way `stepmold eval` consumes it, so those case-runs stay together in order.
Finished case-runs in such a group are replayed and not written again.
"""
if self.spec.scheme != "fake":
return [[job] for job in self.jobs if not job.done]
groups = {}
order = []
for job in self.jobs:
key = (job.arm, job.pack, job.run)
if key not in groups:
groups[key] = []
order.append(key)
groups[key].append(job)
return [groups[key] for key in order if any(not job.done for job in groups[key])]
def _run_unit(self, unit):
shared = None
if self.spec.scheme == "fake":
entry = self.packs[unit[0].pack]
shared = self.sm.cli.resolve_model(self.spec.spec, entry.pack)
for job in unit:
if self.stop.is_set():
return
records = self._run_one(job, shared)
if not job.done:
self.writer.write(records)
self._progress(job, records[-1])
def _stamp(self, job):
return {
"invocation": self.invocation, "case_run_id": uuid.uuid4().hex[:16],
"label": self.label, "model": self.spec.model, "arm": job.arm,
"pack": self.packs[job.pack].name, "case": job.case_name,
"case_index": job.case_index, "run": job.run,
}
def _run_one(self, job, shared):
entry = self.packs[job.pack]
stamp = self._stamp(job)
recorder = Recorder(stamp, entry.pack, self.sm, self.args.record_prompts,
self.args.record_reasoning)
pack = entry.pack if job.arm == "stepmold" else entry.plain
started = time.monotonic()
result = None
failure = None
_LOCAL.recorder = recorder
try:
tools = load_tools(entry.tools) if entry.tools else None
model = self._model(job.arm, shared)
report = self.sm.eval.evaluate(pack, model, cases=[job.case], tools=tools)
result = report.cases[0]
except Exception as exc:
failure = "%s: %s" % (type(exc).__name__, _clip(str(exc), 500))
finally:
_LOCAL.recorder = None
wall_ms = (time.monotonic() - started) * 1000.0
try:
recorder.finalize()
if failure is None:
failure = self._infrastructure(result, recorder)
if failure is None:
return recorder.calls + [
self._case_record(stamp, job.case, result, recorder, wall_ms)]
except Exception as exc: # a bug here must cost this case-run, not the bench
failure = "bench error building the record: %s: %s" % (type(exc).__name__, exc)
record = {"kind": "infra_error"}
record.update(stamp)
record.update({"error": failure, "calls": len(recorder.calls),
"wall_ms": round(wall_ms, 1), "timestamp": _now()})
return recorder.calls + [record]
@staticmethod
def _infrastructure(result, recorder):
"""Why a case-run says nothing about the model, or None when it does.
A run that reached an end node, or stopped on one of stepmold's own run errors,
is a result. A backend that could not be reached or kept failing (stepmold's
`BackendError` escaping the run) is not, and neither is anything that is not a
stepmold error at all, except a scripted model running out of answers. Those get
an `infra_error` record instead of a `case` record, so a re-run retries them.
"""
if recorder.end_node is not None:
return None
error = (recorder.run_error or {}).get("error")
text = (result.error or "") if result is not None else ""
if error:
return (text or error) if error == "BackendError" else None
if text.startswith("ModelExhausted"):
return None
return text or "the run stopped without saying why"
def _case_record(self, stamp, case, result, recorder, wall_ms):
completed = recorder.end_node is not None
misses = [m for m in result.mismatches if m.field != "<ending>"]
total = len(case.expect)
if not completed:
outcome = "crashed"
elif result.passed:
outcome = "passed"
elif recorder.escalations:
outcome = "escalated"
else:
outcome = "silently_wrong"
calls = recorder.calls
answers = [call for call in calls if call.get("stage") == "answer"]
record = {"kind": "case"}
record.update(stamp)
record.update({
"outcome": outcome,
"passed": bool(result.passed),
"tier": result.tier,
"end_node": recorder.end_node,
"expected_end": case.end,
"end_ok": None if case.end is None else recorder.end_node == case.end,
"rescued_expected": bool(case.rescued),
"input": dict(case.input),
"output": dict(result.actual) if completed else None,
"expected": dict(case.expect),
"fields_total": total,
"fields_matched": total - len(misses) if completed else 0,
"mismatches": [{"field": m.field, "expected": m.expected, "actual": m.actual,
"node": m.node, "note": m.note} for m in result.mismatches],
"blame_node": result.node,
"error": result.error,
"escalated": bool(recorder.escalations),
"escalations": recorder.escalations,
"path": recorder.path,
"steps": recorder.steps,
"generations": recorder.generations if recorder.generations is not None
else len(answers),
"calls": len(calls),
"think_calls": len(calls) - len(answers),
"answer_calls": len(answers),
"rejected_calls": sum(1 for call in answers if call.get("accepted") is False),
"retries": recorder.retries,
"extra_draws": sum(max(0, (gate.get("drawn") or 1) - 1) for gate in recorder.gates),
"gates": recorder.gates,
"prompt_tokens": _sum(calls, "prompt_tokens"),
"completion_tokens": _sum(calls, "completion_tokens"),
"reasoning_tokens": _sum(calls, "reasoning_tokens"),
"grammar_mode": self.grammar_mode if stamp["arm"] == "stepmold" else "none",
"wall_ms": round(wall_ms, 1),
"timestamp": _now(),
})
return record
def _run_record(self):
bench_file = os.path.abspath(__file__)
with open(bench_file, "rb") as handle:
bench_sha = hashlib.sha256(handle.read()).hexdigest()[:16]
return {
"kind": "run", "invocation": self.invocation, "label": self.label,
"model": self.spec.model, "base_url": self.spec.base_url,
"model_spec": self.spec.spec, "arms": self.arms, "runs": self.args.runs,
"packs": sorted(entry.name for entry in self.packs.values()),
"limit": self.args.limit, "concurrency": self.args.concurrency,
"grammar_mode": self.grammar_mode, "reasoning_reserve": self.reasoning_reserve,
"temperature": self.temperature, "extra_body": self.extra_body,
"timeout_s": self.args.timeout, "http_retries": self.args.http_retries,
"api_key_set": bool(self.api_key), "setup": self.setup(),
"pack_digests": self.pack_digests(),
"stepmold": self.provenance, "bench": {"version": BENCH_VERSION,
"sha256": bench_sha},
"python": platform.python_version(), "platform": platform.platform(),
"started": _now(),
}
# --------------------------------------------------------------------- output
def _say(self, text):
with self.print_lock:
self.stream.write(text + "\n")
self.stream.flush()
def _progress(self, job, record):
with self.print_lock:
self.finished += 1
finished = self.finished
bucket = self.totals.setdefault(job.arm, {})
outcome = record.get("outcome") or record.get("kind")
bucket[outcome] = bucket.get(outcome, 0) + 1
status = {"passed": "PASS", "escalated": "ESCALATED", "silently_wrong": "WRONG",
"crashed": "CRASHED"}.get(record.get("outcome"), "INFRA ERROR")
line = "[%*d/%d] %-8s %-18s run %d case %-3d %-10s %2d calls %7.1fs" % (
len(str(self.todo)), finished, self.todo, job.arm, job.pack, job.run,
job.case_index, status, record.get("calls") or 0,
(record.get("wall_ms") or 0) / 1000.0)
if record.get("kind") == "infra_error":
line += " " + _clip(record.get("error"), 160)
self._say(line)
def _report_totals(self):
for arm in self.arms:
bucket = self.totals.get(arm, {})
count = sum(bucket.values())
if not count:
continue
self._say("%-8s %d case-runs: %s" % (arm, count, ", ".join(
"%d %s" % (bucket[key], key.replace("_", " "))
for key in ("passed", "escalated", "silently_wrong", "crashed", "infra_error")
if bucket.get(key))))
self._say("results: %s" % self.out)
# ------------------------------------------------------------------------------- CLI
def build_parser():
parser = argparse.ArgumentParser(
prog="bench.py",
description="Run stepmold's example packs against one model, in two arms "
"(stepmold, plain; plain_schema is opt-in), and write one JSONL record "
"per request and per case-run. The API key is read from "
"STEPMOLD_API_KEY.")
parser.add_argument("--model", required=True,
help="stepmold model spec: openai:<base_url>#<model>[#<grammar_mode>"
"[#<reasoning_reserve>]], or fake:<script> for offline runs")
parser.add_argument("--label", help="name for this model in the results "
"(default: the model name)")
parser.add_argument("--grammar-mode", dest="grammar_mode", default=None,
help="how the stepmold arm sends its grammar (default: the one in "
"--model, else response_format). The plain arm always uses none")
parser.add_argument("--reasoning-reserve", dest="reasoning_reserve", type=int,
default=None, help="tokens added to every request's max_tokens, for "
"models that think before answering (default 0)")
parser.add_argument("--arms", default=DEFAULT_ARMS,
help="comma list of stepmold, plain, plain_schema "
"(default: %s)" % DEFAULT_ARMS)
parser.add_argument("--runs", type=int, default=3, help="runs per case (default 3)")
parser.add_argument("--packs", default="all", help="all, or a comma list of pack names")
parser.add_argument("--limit", type=int, default=None,
help="only the first N gold cases of each pack")
parser.add_argument("--concurrency", type=int, default=8,
help="case-runs in flight at once (default 8)")
parser.add_argument("--extra-body", dest="extra_body", default=None,
help="JSON object merged into every request, e.g. "
"'{\"chat_template_kwargs\": {\"enable_thinking\": false}}'")
parser.add_argument("--out", help="results file (default: results/<label>.jsonl)")
parser.add_argument("--stepmold", default=None,
help="path to a stepmold checkout (default: %s if it exists, else "
"the installed stepmold)" % DEFAULT_STEPMOLD)
parser.add_argument("--examples", default=None,
help="directory of packs (default: <stepmold checkout>/examples)")
parser.add_argument("--tools", action="append", metavar="PACK=FILE.py[:ATTR]",
help="host tools for a pack (refund_desk's are found by default)")
parser.add_argument("--timeout", type=float, default=120.0,
help="seconds per HTTP request (default 120)")
parser.add_argument("--http-retries", dest="http_retries", type=int, default=None,
help="HTTP retries on 429/5xx (default: stepmold's, 2)")
parser.add_argument("--record-prompts", dest="record_prompts", action="store_true",
help="store each rendered prompt in its call record")
parser.add_argument("--record-reasoning", dest="record_reasoning", action="store_true",
help="store the model's reasoning text, when the server returns it")
parser.add_argument("--allow-config-change", dest="allow_config_change",
action="store_true",
help="resume into a file whose earlier runs used another setup")
parser.add_argument("--dry-run", dest="dry_run", action="store_true",
help="print the plan and what is already done, then stop")
return parser
def main(argv=None, opener=None, sleeper=None, stream=None):
args = build_parser().parse_args(argv)
path = args.stepmold
if path is None and os.path.isfile(os.path.join(DEFAULT_STEPMOLD, "stepmold",
"__init__.py")):
path = DEFAULT_STEPMOLD
try:
sm = load_stepmold(path)
bench = Bench(args, sm, opener=opener, sleeper=sleeper, stream=stream)
return bench.run()
except BenchError as exc:
(stream or sys.stderr).write("bench: %s\n" % exc)
return 2
if __name__ == "__main__":
sys.exit(main())
|