File size: 13,985 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 | """Run matrix tasks concurrently across a fixed local GPU set."""
from __future__ import annotations
import argparse
from concurrent.futures import ThreadPoolExecutor, as_completed
from dataclasses import dataclass
import csv
import json
import os
from pathlib import Path
from queue import Queue
import re
import subprocess
import sys
import threading
from typing import Sequence
from dlb.conditional_matrix import (
CONDITIONAL_MATRIX_COLUMNS,
CONDITIONAL_MATRIX_SCHEMA,
ConditionalMatrixTask,
)
from dlb.matrix import MatrixTask, read_matrix
DEFAULT_GPUS = "0,1,2,3"
DEFAULT_MAX_JOBS = 4
DEFAULT_METRICS = "gen_ppl,entropy,self_bleu"
SAFE_LOG_COMPONENT = re.compile(r"[^A-Za-z0-9_.-]+")
@dataclass(frozen=True)
class TaskResult:
task: MatrixTask | ConditionalMatrixTask
gpu: str
exit_code: int
stdout_log: Path
stderr_log: Path
def parse_gpus(value: str) -> tuple[str, ...]:
"""Parse comma/space separated GPU IDs while rejecting shell syntax."""
parts = tuple(part for part in re.split(r"[,\s]+", value.strip()) if part)
if not parts:
raise ValueError("at least one GPU must be supplied")
allowed = re.compile(r"^[A-Za-z0-9_.:-]+$")
invalid = [part for part in parts if not allowed.fullmatch(part)]
if invalid:
raise ValueError("GPU IDs may contain only letters, numbers, '.', '_', ':', or '-'")
if len(set(parts)) != len(parts):
raise ValueError("GPU IDs must not be duplicated")
return parts
def active_gpus(gpus: Sequence[str], max_jobs: int) -> tuple[str, ...]:
if max_jobs <= 0:
raise ValueError("max jobs must be positive")
return tuple(gpus[index % len(gpus)] for index in range(max_jobs))
def _script(root: Path, name: str) -> str:
path = root / "scripts" / name
return str(path)
def command_for_task(
task: MatrixTask | ConditionalMatrixTask,
*,
root: Path,
stage: str,
metrics: str = DEFAULT_METRICS,
precision: str = "author",
) -> list[str]:
"""Return the argv for one matrix row and stage."""
conditional = getattr(task, "protocol", None) == "c64_zs_v1"
if stage in {"smoke", "generate"}:
script = "run_conditional_one.sh" if conditional else "run_one.sh"
command = [
"bash",
_script(root, script),
"--model",
task.model,
"--dataset",
task.dataset,
"--steps",
str(task.steps),
]
if conditional:
command.extend(["--seed", str(task.seed)])
else:
command.extend(["--num-samples", str(task.sample_count), "--seed", str(task.seed)])
if stage == "smoke":
smoke_root = root / "results" / ("conditional/smoke" if conditional else "smoke")
command.extend(["--results-root", str(smoke_root)])
return command
if stage == "evaluate":
eval_python = os.environ.get("DLB_EVAL_PYTHON")
if eval_python:
command = [eval_python]
else:
command = [
os.environ.get("DLB_CONDA", "conda"),
"run",
"-n",
os.environ.get("DLB_EVAL_ENV", "dlb-eval"),
"python",
]
return [
*command,
"-m",
"evaluation.conditional_evaluate" if conditional else "evaluation.evaluate",
"--root",
str(root),
"--samples",
str(Path(task.sample_dir) / "samples.jsonl"),
"--metrics",
metrics,
"--dataset",
task.dataset,
"--output",
task.metrics_path,
]
if stage == "benchmark":
return [
"bash",
_script(root, "benchmark_conditional_one.sh" if conditional else "benchmark_one.sh"),
"--model",
task.model,
"--dataset",
task.dataset,
"--steps",
str(task.steps),
"--seed",
str(task.seed),
"--precision",
precision,
]
raise ValueError(f"unsupported stage: {stage}")
def _safe_log_name(task_id: str) -> str:
return SAFE_LOG_COMPONENT.sub("_", task_id)
def _task_environment(root: Path, gpu: str) -> dict[str, str]:
env = os.environ.copy()
env["CUDA_VISIBLE_DEVICES"] = gpu
env["DLB_ROOT"] = str(root)
source_path = str(root / "src")
current = env.get("PYTHONPATH")
env["PYTHONPATH"] = source_path if not current else source_path + os.pathsep + current
return env
def _run_task(
task: MatrixTask,
*,
root: Path,
stage: str,
gpu: str,
metrics: str,
precision: str,
print_lock: threading.Lock,
) -> TaskResult:
logs = root / "results" / "logs"
logs.mkdir(parents=True, exist_ok=True)
name = _safe_log_name(task.task_id)
stdout_log = logs / f"4gpu-{stage}-{name}.out"
stderr_log = logs / f"4gpu-{stage}-{name}.err"
command = command_for_task(
task, root=root, stage=stage, metrics=metrics, precision=precision
)
with print_lock:
print(f"{stage.upper()} {task.task_id} gpu={gpu}", flush=True)
with stdout_log.open("w", encoding="utf-8") as stdout, stderr_log.open(
"w", encoding="utf-8"
) as stderr:
try:
completed = subprocess.run(
command,
cwd=root,
env=_task_environment(root, gpu),
stdout=stdout,
stderr=stderr,
check=False,
)
exit_code = completed.returncode
except OSError as error:
stderr.write(f"failed to start command: {error}\n")
exit_code = 127
return TaskResult(
task=task,
gpu=gpu,
exit_code=exit_code,
stdout_log=stdout_log,
stderr_log=stderr_log,
)
def _failure_path(root: Path, stage: str) -> Path:
names = {"generate": "generation", "evaluate": "evaluation"}
name = names.get(stage, stage)
return root / "results" / "logs" / f"{name}_4gpu_failures.tsv"
def _write_failures(root: Path, stage: str, results: Sequence[TaskResult]) -> None:
failures = [result for result in results if result.exit_code != 0]
path = _failure_path(root, stage)
if not failures:
path.unlink(missing_ok=True)
return
path.parent.mkdir(parents=True, exist_ok=True)
temporary = path.with_name(f".{path.name}.{os.getpid()}.partial")
try:
with temporary.open("w", encoding="utf-8", newline="") as handle:
writer = csv.DictWriter(
handle,
fieldnames=(
"task_id",
"model",
"dataset",
"steps",
"gpu",
"exit_code",
"stdout_log",
"stderr_log",
),
delimiter="\t",
lineterminator="\n",
)
writer.writeheader()
for result in failures:
writer.writerow(
{
"task_id": result.task.task_id,
"model": result.task.model,
"dataset": result.task.dataset,
"steps": result.task.steps,
"gpu": result.gpu,
"exit_code": result.exit_code,
"stdout_log": str(result.stdout_log),
"stderr_log": str(result.stderr_log),
}
)
os.replace(temporary, path)
except BaseException:
temporary.unlink(missing_ok=True)
raise
def dry_run(
tasks: Sequence[MatrixTask | ConditionalMatrixTask],
*,
root: Path,
stage: str,
gpus: Sequence[str],
metrics: str,
precision: str,
) -> int:
for index, task in enumerate(tasks):
gpu = gpus[index % len(gpus)]
command = command_for_task(
task, root=root, stage=stage, metrics=metrics, precision=precision
)
print(
json.dumps(
{
"stage": stage,
"task_id": task.task_id,
"gpu": gpu,
"cuda_visible_devices": gpu,
"command": command,
},
sort_keys=True,
)
)
return 0
def run_tasks(
tasks: Sequence[MatrixTask | ConditionalMatrixTask],
*,
root: Path,
stage: str,
gpus: Sequence[str],
metrics: str,
precision: str,
) -> int:
gpu_queue: Queue[str] = Queue()
for gpu in gpus:
gpu_queue.put(gpu)
print_lock = threading.Lock()
results: list[TaskResult] = []
def run_with_gpu(task: MatrixTask | ConditionalMatrixTask) -> TaskResult:
gpu = gpu_queue.get()
try:
return _run_task(
task,
root=root,
stage=stage,
gpu=gpu,
metrics=metrics,
precision=precision,
print_lock=print_lock,
)
finally:
gpu_queue.put(gpu)
with ThreadPoolExecutor(max_workers=len(gpus)) as executor:
futures = [executor.submit(run_with_gpu, task) for task in tasks]
for future in as_completed(futures):
result = future.result()
results.append(result)
if result.exit_code != 0:
with print_lock:
print(
f"FAILED {result.task.task_id} gpu={result.gpu} "
f"exit={result.exit_code}",
file=sys.stderr,
flush=True,
)
_write_failures(root, stage, results)
return 1 if any(result.exit_code != 0 for result in results) else 0
def _read_conditional_matrix(path: Path) -> list[ConditionalMatrixTask]:
if path.is_symlink() or not path.is_file():
raise ValueError(f"conditional matrix TSV is missing or unsafe: {path}")
with path.open("r", encoding="utf-8", newline="") as handle:
schema = handle.readline().rstrip("\r\n")
if schema != f"# schema={CONDITIONAL_MATRIX_SCHEMA}":
raise ValueError("conditional matrix TSV has an unsupported schema header")
reader = csv.DictReader(handle, delimiter="\t")
if tuple(reader.fieldnames or ()) != CONDITIONAL_MATRIX_COLUMNS:
raise ValueError("conditional matrix TSV columns do not match the canonical schema")
tasks: list[ConditionalMatrixTask] = []
for row_number, row in enumerate(reader, start=3):
if None in row or any(value is None for value in row.values()):
raise ValueError(f"conditional matrix TSV row {row_number} has missing fields")
try:
tasks.append(
ConditionalMatrixTask(
task_id=row["task_id"],
category=row["category"],
model=row["model"],
dataset=row["dataset"],
steps=int(row["steps"]),
sample_count=2048,
seed=int(row["seed"]),
environment=row["environment"],
adapter=row["adapter"],
source=row["source"],
provenance=row["provenance"],
protocol="c64_zs_v1",
conditioning_manifest=row["conditioning_manifest"],
conditioning_manifest_sha256=row["conditioning_manifest_sha256"],
sample_dir=row["sample_dir"],
metrics_path=row["metrics_path"],
timing_path=row["timing_path"],
)
)
except (KeyError, TypeError, ValueError) as error:
raise ValueError(f"invalid conditional matrix TSV row {row_number}") from error
return tasks
def read_any_matrix(path: Path) -> list[MatrixTask | ConditionalMatrixTask]:
with path.open("r", encoding="utf-8", newline="") as handle:
schema = handle.readline().rstrip("\r\n")
if schema == f"# schema={CONDITIONAL_MATRIX_SCHEMA}":
return _read_conditional_matrix(path)
return read_matrix(path)
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--root", type=Path, required=True)
parser.add_argument("--stage", choices=("smoke", "generate", "evaluate", "benchmark"), required=True)
parser.add_argument("--matrix", type=Path, required=True)
parser.add_argument("--gpus", default=os.environ.get("DLB_GPUS", DEFAULT_GPUS))
parser.add_argument("--max-jobs", type=int, default=int(os.environ.get("DLB_MAX_JOBS", DEFAULT_MAX_JOBS)))
parser.add_argument("--metrics", default=DEFAULT_METRICS)
parser.add_argument("--precision", choices=("author",), default="author")
parser.add_argument("--dry-run", action="store_true")
arguments = parser.parse_args(argv)
root = arguments.root.resolve()
try:
gpu_list = active_gpus(parse_gpus(arguments.gpus), arguments.max_jobs)
except ValueError as error:
parser.error(str(error))
tasks = read_any_matrix(arguments.matrix)
if arguments.dry_run:
return dry_run(
tasks,
root=root,
stage=arguments.stage,
gpus=gpu_list,
metrics=arguments.metrics,
precision=arguments.precision,
)
return run_tasks(
tasks,
root=root,
stage=arguments.stage,
gpus=gpu_list,
metrics=arguments.metrics,
precision=arguments.precision,
)
if __name__ == "__main__":
raise SystemExit(main())
|