File size: 19,808 Bytes
784f572 | 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 | """Render and execute registry-bound, concrete sampler timing commands."""
from __future__ import annotations
import argparse
import json
import math
import os
from pathlib import Path
import re
import subprocess
from typing import Iterable, Literal
import uuid
from dlb.adapters.base import AdapterError, BaseTeacherAdapter
from dlb.checkpoints import load_checkpoint_manifest
from dlb.command import ADAPTERS
from dlb.io import atomic_json_write, sha256_file
from dlb.registry import load_registry, step_grid_for_model
from dlb.runner import RunRequest, _resolve_request
from dlb.timing import _validate_metadata
def _run_dir(root: Path, request: RunRequest) -> Path:
result_root = (
root / "results/conditional"
if request.generation_mode == "conditional_prefix"
else root / "results"
)
return (
result_root
/ "samples"
/ request.dataset_id
/ request.model_id
/ f"steps_{request.step_count}"
)
def _output_path(output_root: Path, request: RunRequest) -> Path:
return (
output_root
/ request.dataset_id
/ request.model_id
/ f"steps_{request.step_count}"
/ "timing.json"
)
def _request(
model: str,
dataset: str,
steps: int,
seed: int,
generation_mode: Literal["unconditional", "conditional_prefix"] = "unconditional",
) -> RunRequest:
return RunRequest(
run_id=f"benchmark-{model}-{dataset}-steps-{steps}",
model_id=model,
dataset_id=dataset,
step_count=steps,
seed=seed,
sample_count=2048 if generation_mode == "conditional_prefix" else 1,
generation_mode=generation_mode,
)
def render_benchmark_matrix(
*,
root: Path,
models: Iterable[str] | None,
datasets: Iterable[str],
steps: int,
seed: int,
precision: str,
output_root: Path,
dry_run: bool,
generation_mode: Literal["unconditional", "conditional_prefix"] = "unconditional",
) -> list[dict[str, object]]:
"""Render every selected cell, preserving explicit unsupported records."""
root = root.resolve()
registry = load_registry(root / "configs/experiments.yaml")
selected_models = tuple(registry.models) if models is None else tuple(models)
records: list[dict[str, object]] = []
for model_id in selected_models:
for dataset_id in datasets:
model = registry.models.get(model_id)
support = model.datasets.get(dataset_id) if model is not None else None
if model is None or support is None or support.status != "supported":
records.append(
{
"status": "unsupported",
"model": model_id,
"dataset": dataset_id,
"reason": getattr(support, "reason", None) or "unknown model/dataset cell",
}
)
continue
allowed_steps = step_grid_for_model(registry, model_id)
if steps not in allowed_steps:
joined_steps = ",".join(str(step) for step in allowed_steps)
records.append(
{
"status": "error",
"model": model_id,
"dataset": dataset_id,
"reason": (
f"invalid step count {steps} for {model_id}/{dataset_id}; "
f"allowed: {joined_steps}"
),
}
)
continue
adapter = ADAPTERS[model_id]
request = _request(model_id, dataset_id, steps, seed, generation_mode)
if generation_mode == "conditional_prefix":
try:
request, _ = _resolve_request(request, root, adapter)
except (OSError, ValueError) as error:
records.append(
{
"status": "error",
"model": model_id,
"dataset": dataset_id,
"reason": str(error),
}
)
continue
output = _output_path(output_root.resolve(), request)
dry_attempt = "0" * 32
staged_output = output.with_name(
f".{output.name}.{dry_attempt}.staged.json"
)
metadata_path = output.with_name(
f".benchmark_metadata.{dry_attempt}.json"
)
try:
command = adapter.render_benchmark_command(
request,
_run_dir(root, request),
output=staged_output,
metadata_path=metadata_path,
precision=precision,
dry_run=dry_run,
)
except (AdapterError, OSError, ValueError) as error:
records.append(
{
"status": "error",
"model": model_id,
"dataset": dataset_id,
"reason": str(error),
}
)
continue
records.append(
{
"status": "supported",
"model": model_id,
"dataset": dataset_id,
"steps": steps,
"environment": model.environment,
"hook": adapter.benchmark_hook(request),
**adapter.author_precision_policy(request),
"batch_size": 1,
"warmups": 5,
"repeats": 32,
"output": str(output),
"staged_output": str(staged_output),
"metadata_path": str(metadata_path),
"generation_mode": generation_mode,
"command": command,
}
)
return records
def _precision_policy_binding(
root: Path, request: RunRequest, adapter: BaseTeacherAdapter
) -> dict[str, object]:
"""Bind an audited static policy to exact run, checkpoint and config bytes."""
binding: dict[str, object] = {
"model": request.model_id,
"dataset": request.dataset_id,
"upstream": adapter.upstream,
"source_commit": request.source_sha256,
"experiment_config_sha256": request.config_sha256,
"checkpoint_sha256": request.checkpoint_sha256,
"checkpoint_lock_id": request.checkpoint_lock_id,
"checkpoint_selection": request.checkpoint_selection,
"adapter_identity": request.adapter_identity,
}
if request.generation_mode == "conditional_prefix":
binding.update(
{
"generation_mode": request.generation_mode,
"conditioning_manifest": request.conditioning_manifest,
"conditioning_manifest_sha256": request.conditioning_manifest_sha256,
"conditioning_config_sha256": request.conditioning_config_sha256,
"prefix_length": request.prefix_length,
"evaluation_continuation_length": request.evaluation_continuation_length,
"prompt_count": request.prompt_count,
"diversity_prompt_count": request.diversity_prompt_count,
"completions_per_diversity_prompt": request.completions_per_diversity_prompt,
"completion_schedule": request.completion_schedule,
}
)
selection = request.checkpoint_selection
resource_id = selection.get("resource") if isinstance(selection, dict) else None
if isinstance(resource_id, str):
manifest = load_checkpoint_manifest(root / "artifacts/checkpoints.yaml")
resource = manifest.resources.get(resource_id)
if resource is None:
raise ValueError(f"precision policy references unknown checkpoint {resource_id}")
binding.update(
{
"checkpoint_resource": resource_id,
"checkpoint_repository": getattr(resource.source, "repo_id", None),
"checkpoint_revision": getattr(resource.source, "revision", None),
}
)
if request.model_id in {"flm", "fmlm"}:
config_path = root / "checkpoints" / resource.destination / "config.json"
if config_path.is_symlink() or not config_path.is_file():
raise ValueError(
"FLM/FMLM precision policy requires the downloaded checkpoint config.json"
)
try:
config = json.loads(config_path.read_text(encoding="utf-8"))
except (OSError, UnicodeDecodeError, json.JSONDecodeError) as error:
raise ValueError("FLM/FMLM checkpoint config.json is invalid") from error
if not isinstance(config, dict):
raise ValueError("FLM/FMLM checkpoint config.json must be a mapping")
dtype = config.get("torch_dtype")
if not isinstance(dtype, str) or not dtype:
raise ValueError(
"FLM/FMLM checkpoint config does not declare torch_dtype"
)
binding.update(
{
"checkpoint_config_path": str(config_path.resolve()),
"checkpoint_config_sha256": sha256_file(config_path),
"checkpoint_config_torch_dtype": dtype,
"checkpoint_config_architectures": config.get("architectures"),
"checkpoint_config_auto_map": config.get("auto_map"),
}
)
return binding
def _metadata(
root: Path, request: RunRequest, adapter: BaseTeacherAdapter, attempt_id: str
) -> dict[str, object]:
binding = _precision_policy_binding(root, request, adapter)
policy: dict[str, object] = adapter.author_precision_policy(request)
if policy.get("precision") == "resolved-from-checkpoint-config-at-execution":
policy["precision"] = "bf16-mixed-static-author-policy"
return {
"seed": request.seed,
"sample_count": request.sample_count,
"dataset": request.dataset_id,
"model": request.model_id,
"steps": request.step_count,
"environment": request.environment,
"source_commit": request.source_sha256,
"config_sha256": request.config_sha256,
"checkpoint_sha256": request.checkpoint_sha256,
"checkpoint_lock_id": request.checkpoint_lock_id,
"checkpoint_selection": request.checkpoint_selection,
"checkpoint_teacher_family": request.checkpoint_teacher_family,
"adapter_identity": request.adapter_identity,
"requested_precision": "author",
"attempt_id": attempt_id,
"precision_policy_binding": binding,
**policy,
**(
{
"generation_mode": request.generation_mode,
"conditioning_manifest": request.conditioning_manifest,
"conditioning_manifest_sha256": request.conditioning_manifest_sha256,
"conditioning_config_sha256": request.conditioning_config_sha256,
"prefix_length": request.prefix_length,
"evaluation_continuation_length": request.evaluation_continuation_length,
"prompt_count": request.prompt_count,
"diversity_prompt_count": request.diversity_prompt_count,
"completions_per_diversity_prompt": request.completions_per_diversity_prompt,
"completion_schedule": request.completion_schedule,
}
if request.generation_mode == "conditional_prefix"
else {}
),
}
def _load_staged_timing(
path: Path, expected_metadata: dict[str, object], attempt_id: str
) -> None:
if path.is_symlink() or not path.is_file():
raise RuntimeError("benchmark sampler completed without a fresh staged timing result")
try:
payload = json.loads(path.read_text(encoding="utf-8"))
timing = payload["timing"]
metadata = payload["metadata"]
except (OSError, UnicodeDecodeError, json.JSONDecodeError, KeyError, TypeError) as error:
raise RuntimeError("fresh staged timing result is invalid") from error
if payload.get("schema") != "dlb-generation-timing-v1":
raise RuntimeError("fresh staged timing result has the wrong schema")
try:
_validate_metadata(metadata)
except (TypeError, ValueError) as error:
raise RuntimeError(f"fresh staged timing metadata is invalid: {error}") from error
if metadata.get("attempt_id") != attempt_id:
raise RuntimeError("fresh staged timing attempt provenance does not match")
for key, expected in expected_metadata.items():
if metadata.get(key) != expected:
raise RuntimeError(f"fresh staged timing provenance differs at {key}")
if not isinstance(timing, dict) or any(
timing.get(key) != expected
for key, expected in {
"mode": "primary_latency",
"warmups": 5,
"repeats": 32,
"batch_size": 1,
"num_timed_samples": 32,
}.items()
):
raise RuntimeError("fresh staged timing protocol is invalid")
raw = timing.get("raw_durations_seconds")
seconds = timing.get("seconds_per_sample")
if (
not isinstance(raw, list)
or len(raw) != 32
or any(type(value) not in {int, float} or not math.isfinite(value) or value < 0 for value in raw)
or type(seconds) not in {int, float}
or not math.isfinite(seconds)
or seconds < 0
):
raise RuntimeError("fresh staged timing durations are invalid")
def run_timing_attempt(
command: list[str],
*,
cwd: Path,
final_output: Path,
staged_output: Path,
expected_metadata: dict[str, object],
attempt_id: str,
) -> Path:
"""Publish a staged result only after the complete sampler process succeeds."""
if re.fullmatch(r"[0-9a-f]{32}", attempt_id) is None:
raise ValueError("attempt_id must be 32 lowercase hexadecimal characters")
final_output = final_output.absolute()
staged_output = staged_output.absolute()
expected_stage = final_output.with_name(
f".{final_output.name}.{attempt_id}.staged.json"
)
if staged_output != expected_stage:
raise ValueError("staged timing path does not match this attempt")
final_output.parent.mkdir(parents=True, exist_ok=True)
if staged_output.exists() or staged_output.is_symlink():
raise RuntimeError("staged timing attempt path already exists")
superseded = final_output.with_name(
f".{final_output.name}.{attempt_id}.superseded.json"
)
if superseded.exists() or superseded.is_symlink():
raise RuntimeError("superseded timing path already exists")
if final_output.exists() or final_output.is_symlink():
if final_output.is_symlink() or not final_output.is_file():
raise RuntimeError("existing timing result is unsafe")
os.replace(final_output, superseded)
try:
completed = subprocess.run(command, cwd=cwd, check=False)
if completed.returncode:
raise RuntimeError(
f"benchmark sampler exited with status {completed.returncode}"
)
_load_staged_timing(staged_output, expected_metadata, attempt_id)
os.replace(staged_output, final_output)
except BaseException:
staged_output.unlink(missing_ok=True)
raise
return final_output
def execute_one(
*,
root: Path,
model: str,
dataset: str,
steps: int,
seed: int,
precision: str,
generation_mode: Literal["unconditional", "conditional_prefix"] = "unconditional",
) -> Path:
"""Resolve provenance, then execute one real sampler wrapper in this environment."""
root = root.resolve()
request = _request(model, dataset, steps, seed, generation_mode)
adapter: BaseTeacherAdapter = ADAPTERS[model]
resolved, adapter = _resolve_request(request, root, adapter)
timing_root = (
root / "results/conditional/timing"
if generation_mode == "conditional_prefix"
else root / "results/timing"
)
output = _output_path(timing_root, resolved)
attempt_id = uuid.uuid4().hex
staged_output = output.with_name(f".{output.name}.{attempt_id}.staged.json")
metadata_path = output.with_name(f".benchmark_metadata.{attempt_id}.json")
metadata = _metadata(root, resolved, adapter, attempt_id)
atomic_json_write(metadata_path, metadata)
try:
command = adapter.render_benchmark_command(
resolved,
_run_dir(root, resolved),
output=staged_output,
metadata_path=metadata_path,
precision=precision,
dry_run=False,
)
return run_timing_attempt(
command,
cwd=root,
final_output=output,
staged_output=staged_output,
expected_metadata=metadata,
attempt_id=attempt_id,
)
finally:
metadata_path.unlink(missing_ok=True)
def _csv(value: str) -> tuple[str, ...]:
items = tuple(value.split(","))
if not items or any(not item for item in items):
raise argparse.ArgumentTypeError("comma-separated values must not be empty")
return items
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--root", type=Path, required=True)
parser.add_argument("--models", type=_csv, required=True)
parser.add_argument("--datasets", type=_csv, required=True)
parser.add_argument("--steps", type=int, required=True)
parser.add_argument("--seed", type=int, required=True)
parser.add_argument("--precision", choices=("author",), required=True)
parser.add_argument(
"--generation-mode",
choices=("unconditional", "conditional_prefix"),
default="unconditional",
)
parser.add_argument("--dry-run", action="store_true")
arguments = parser.parse_args(argv)
if arguments.dry_run:
records = render_benchmark_matrix(
root=arguments.root,
models=arguments.models,
datasets=arguments.datasets,
steps=arguments.steps,
seed=arguments.seed,
precision=arguments.precision,
output_root=(
arguments.root / "results/conditional/timing"
if arguments.generation_mode == "conditional_prefix"
else arguments.root / "results/timing"
),
dry_run=True,
generation_mode=arguments.generation_mode,
)
for record in records:
print(json.dumps(record, sort_keys=True))
return 1 if any(record["status"] == "error" for record in records) else 0
if len(arguments.models) != 1 or len(arguments.datasets) != 1:
parser.error("execution requires exactly one model and one dataset")
try:
output = execute_one(
root=arguments.root,
model=arguments.models[0],
dataset=arguments.datasets[0],
steps=arguments.steps,
seed=arguments.seed,
precision=arguments.precision,
generation_mode=arguments.generation_mode,
)
except (KeyError, OSError, RuntimeError, ValueError) as error:
parser.error(str(error))
print(json.dumps({"status": "succeeded", "output": str(output)}))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|