File size: 12,393 Bytes
d0768a8 | 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 | #!/usr/bin/env python3
"""
Launch multiple preprocessing workers in parallel while preserving the ability
to resume from previous runs. Each worker operates on a disjoint shard of the
manifest and writes outputs to a temporary worker directory; once finished, the
results are merged back into the main output directory and the worker assets are
removed.
"""
from __future__ import annotations
import argparse
import json
import math
import os
import shutil
import subprocess
import sys
import tempfile
from pathlib import Path
from typing import Dict, Iterable, List, Optional, Sequence, Tuple
import time
MANIFEST_TRAIN = "train_manifest.jsonl"
MANIFEST_VAL = "val_manifest.jsonl"
STATS_FILE = "stats.json"
FEATURE_SUBDIRS = ["codes", "condition", "emo_vec", "text_ids"]
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Run IndexTTS2 preprocessing across multiple worker processes."
)
parser.add_argument("--manifest", type=Path, required=True, help="Source dataset manifest JSONL.")
parser.add_argument("--output-dir", type=Path, required=True, help="Destination directory.")
parser.add_argument("--tokenizer", type=Path, required=True, help="SentencePiece model path.")
parser.add_argument(
"--config",
type=Path,
default=Path("checkpoints/config_finetune.yaml"),
help="Model config YAML (default: checkpoints/config_finetune.yaml).",
)
parser.add_argument(
"--gpt-checkpoint",
type=Path,
default=Path("checkpoints/gpt.pth"),
help="Base GPT checkpoint (default: checkpoints/gpt.pth).",
)
parser.add_argument("--language", type=str, default="ja", help="Language hint for normaliser.")
parser.add_argument("--device", type=str, default="cuda", help="Device string (default: cuda).")
parser.add_argument("--batch-size", type=int, default=4, help="Batch size per worker.")
parser.add_argument("--workers", type=int, default=8, help="DataLoader worker threads per process.")
parser.add_argument(
"--num-processes",
type=int,
default=4,
help="Number of parallel preprocessing processes to launch.",
)
parser.add_argument(
"--launch-delay",
type=float,
default=0.0,
help="Seconds to sleep between launching each worker process (default: 0).",
)
parser.add_argument(
"--val-ratio",
type=float,
default=0.01,
help="Validation ratio passed to preprocess_data.py (default: 0.01).",
)
parser.add_argument(
"--skip-existing",
action="store_true",
help="Forward --skip-existing to each worker.",
)
parser.add_argument(
"--max-samples",
type=int,
default=0,
help="Maximum samples to process overall (0 = all remaining).",
)
parser.add_argument(
"--extra-args",
nargs=argparse.REMAINDER,
help="Additional arguments forwarded to preprocess_data.py after a '--'.",
)
parser.add_argument(
"--hf-cache-dir",
type=Path,
default=None,
help="Directory to use for Hugging Face caches/offline assets (default: ./hf_cache).",
)
return parser.parse_args()
def cache_has_required_assets(cache_dir: Path) -> bool:
required = [
cache_dir / "models--facebook--seamless-m4t-medium",
cache_dir / "models--amphion--MaskGCT",
]
return all(path.exists() for path in required)
def merge_manifest_shards(base_path: Path, pattern: str, target: Path) -> None:
shard_files = sorted(base_path.glob(pattern))
if not shard_files:
return
target.parent.mkdir(parents=True, exist_ok=True)
with target.open("a", encoding="utf-8") as dst:
for shard in shard_files:
with shard.open("r", encoding="utf-8") as src:
shutil.copyfileobj(src, dst)
shard.unlink()
def move_feature_tree(source_dir: Path, dest_dir: Path) -> None:
for sub in FEATURE_SUBDIRS:
src_sub = source_dir / sub
if not src_sub.exists():
continue
dst_sub = dest_dir / sub
dst_sub.mkdir(parents=True, exist_ok=True)
for path in src_sub.iterdir():
target_path = dst_sub / path.name
if target_path.exists():
# Already present; assume prior run completed this sample.
path.unlink()
else:
shutil.move(str(path), str(target_path))
def consolidate_previous_shards(output_dir: Path) -> None:
# Merge stray worker manifests if the previous run exited early.
merge_manifest_shards(output_dir, "train_manifest.worker_*.jsonl", output_dir / MANIFEST_TRAIN)
merge_manifest_shards(output_dir, "val_manifest.worker_*.jsonl", output_dir / MANIFEST_VAL)
for worker_dir in sorted(output_dir.glob("worker_*")):
merge_worker_results(worker_dir, output_dir)
def load_processed_ids(manifest_paths: Iterable[Path]) -> set[str]:
ids: set[str] = set()
for path in manifest_paths:
if not path.exists():
continue
with path.open("r", encoding="utf-8") as handle:
for line in handle:
if not line.strip():
continue
try:
record = json.loads(line)
except json.JSONDecodeError:
continue
sample_id = record.get("id")
if sample_id:
ids.add(sample_id)
return ids
def remaining_manifest_entries(
manifest: Path,
processed_ids: set[str],
max_samples: int,
) -> List[str]:
remaining: List[str] = []
with manifest.open("r", encoding="utf-8") as handle:
for line in handle:
if not line.strip():
continue
try:
record = json.loads(line)
except json.JSONDecodeError:
continue
sample_id = record.get("id")
if not sample_id or sample_id in processed_ids:
continue
remaining.append(line)
if max_samples and len(remaining) >= max_samples:
break
return remaining
def write_chunks(
lines: Sequence[str],
num_chunks: int,
work_dir: Path,
prefix: str,
) -> List[tuple[Path, int]]:
work_dir.mkdir(parents=True, exist_ok=True)
chunks: List[tuple[Path, int]] = []
if not lines:
return chunks
chunk_size = math.ceil(len(lines) / num_chunks)
for idx in range(num_chunks):
start = idx * chunk_size
if start >= len(lines):
break
end = min(len(lines), start + chunk_size)
chunk_path = work_dir / f"{prefix}_chunk_{idx:02d}.jsonl"
with chunk_path.open("w", encoding="utf-8") as handle:
handle.writelines(lines[start:end])
chunks.append((chunk_path, end - start))
return chunks
def launch_worker(
chunk_manifest: Path,
worker_output: Path,
args: argparse.Namespace,
hf_env: Dict[str, str],
) -> subprocess.Popen:
cmd = [
sys.executable,
"-u",
"tools/preprocess_data.py",
"--manifest",
str(chunk_manifest),
"--output-dir",
str(worker_output),
"--tokenizer",
str(args.tokenizer),
"--config",
str(args.config),
"--gpt-checkpoint",
str(args.gpt_checkpoint),
"--language",
args.language,
"--device",
args.device,
"--batch-size",
str(args.batch_size),
"--workers",
str(args.workers),
"--val-ratio",
str(args.val_ratio),
]
if args.skip_existing:
cmd.append("--skip-existing")
if args.extra_args:
cmd.append("--")
cmd.extend(args.extra_args)
env = os.environ.copy()
for key, value in hf_env.items():
env.setdefault(key, value)
return subprocess.Popen(cmd, env=env)
def append_and_remove(source: Path, destination: Path) -> None:
if not source.exists():
return
destination.parent.mkdir(parents=True, exist_ok=True)
with destination.open("a", encoding="utf-8") as dst, source.open("r", encoding="utf-8") as src:
shutil.copyfileobj(src, dst)
source.unlink()
def merge_worker_results(worker_dir: Path, main_output: Path) -> None:
append_and_remove(worker_dir / MANIFEST_TRAIN, main_output / MANIFEST_TRAIN)
append_and_remove(worker_dir / MANIFEST_VAL, main_output / MANIFEST_VAL)
move_feature_tree(worker_dir, main_output)
stats_path = worker_dir / STATS_FILE
if stats_path.exists():
stats_path.unlink()
shutil.rmtree(worker_dir)
def main() -> None:
args = parse_args()
output_dir = args.output_dir.expanduser().resolve()
output_dir.mkdir(parents=True, exist_ok=True)
hf_cache_dir = (
args.hf_cache_dir.expanduser().resolve()
if args.hf_cache_dir
else (Path.cwd() / "hf_cache")
)
hf_cache_dir.mkdir(parents=True, exist_ok=True)
hf_env = {
"HF_HOME": str(hf_cache_dir),
"HF_HUB_CACHE": str(hf_cache_dir),
"HF_DATASETS_CACHE": str(hf_cache_dir),
}
if cache_has_required_assets(hf_cache_dir):
hf_env["HF_HUB_OFFLINE"] = "1"
hf_env["TRANSFORMERS_OFFLINE"] = "1"
else:
print("[preprocess_multiproc] HF cache missing SeamlessM4T/MaskGCT weights; running in online mode to populate cache.")
consolidate_previous_shards(output_dir)
processed_ids = load_processed_ids(
[
output_dir / MANIFEST_TRAIN,
output_dir / MANIFEST_VAL,
]
)
manifest_path = args.manifest.expanduser().resolve()
remaining_lines = remaining_manifest_entries(manifest_path, processed_ids, args.max_samples)
if not remaining_lines:
print("No remaining samples. Nothing to do.")
return
num_processes = min(max(1, args.num_processes), len(remaining_lines))
chunk_root = manifest_path.parent.parent / "_preprocess_chunks"
chunk_root.mkdir(parents=True, exist_ok=True)
chunk_prefix = manifest_path.parent.name
chunks = write_chunks(remaining_lines, num_processes, chunk_root, chunk_prefix)
if not chunks:
print("Failed to create chunk manifests.")
return
total_samples = sum(count for _, count in chunks)
print(
f"[preprocess_multiproc] Remaining samples: {len(remaining_lines)} "
f"(assigned {total_samples} across {len(chunks)} workers)."
)
for idx, (chunk_path, count) in enumerate(chunks):
print(f"[preprocess_multiproc] Worker {idx:02d} -> {chunk_path} ({count} samples)")
processes: List[subprocess.Popen] = []
worker_dirs: List[Path] = []
try:
for idx, (chunk_path, count) in enumerate(chunks):
worker_dir = output_dir / f"worker_{idx:02d}"
worker_dir.mkdir(parents=True, exist_ok=True)
print(
f"[preprocess_multiproc] Launching worker {idx:02d} "
f"({count} samples) -> dir {worker_dir}"
)
proc = launch_worker(chunk_path, worker_dir, args, hf_env)
processes.append(proc)
worker_dirs.append(worker_dir)
if idx < len(chunks) - 1 and args.launch_delay > 0:
print(
f"[preprocess_multiproc] Sleeping {args.launch_delay:.2f}s before launching next worker..."
)
time.sleep(args.launch_delay)
return_codes = [proc.wait() for proc in processes]
if any(code != 0 for code in return_codes):
for idx, code in enumerate(return_codes):
if code != 0:
print(f"Worker {idx} exited with code {code}", file=sys.stderr)
raise RuntimeError("One or more workers failed. Check logs above.")
for worker_dir in worker_dirs:
merge_worker_results(worker_dir, output_dir)
finally:
for chunk_path, _ in chunks:
if chunk_path.exists():
chunk_path.unlink()
try:
if chunk_root.exists() and not any(chunk_root.iterdir()):
chunk_root.rmdir()
except OSError:
pass
print("All workers completed successfully.")
if __name__ == "__main__":
main()
|