File size: 10,520 Bytes
06f9561 | 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 | from __future__ import annotations
import fcntl
import hashlib
import os
import sqlite3
import threading
import time
from collections import defaultdict, deque
from pathlib import Path
from typing import Any, Callable, Iterator
import numpy as np
from torch.utils.data import ConcatDataset
from psi.data.seed.omni_sqlite_cache import local_sqlite_path
_TRUE_VALUES = {"1", "true", "yes", "y", "on"}
_CACHE_VERSION = "psi0-video-motion-36d-f32-v1"
_PREFETCH_LOCK = threading.Lock()
_PREFETCH_THREAD: threading.Thread | None = None
_PREFETCH_CURRENT_KEY: str | None = None
_PREFETCH_PENDING: tuple[Any, str] | None = None
_PREFETCHED_KEYS: deque[str] = deque(maxlen=16)
def video_motion_local_cache_enabled() -> bool:
return os.environ.get("PSI_VIDEO_MOTION_LOCAL_CACHE", "0").strip().lower() in _TRUE_VALUES
def _motion_cache_root() -> Path:
value = os.environ.get("PSI_VIDEO_MOTION_LOCAL_CACHE_DIR", "").strip()
return Path(value or "/tmp/psi0_video_motion_arrays")
def _motion_cache_path(motion_dir: Path) -> Path:
# The source motion trees are immutable training inputs. The job-local
# cache root is cleared by the launcher, so hashing path + format avoids
# three remote stat calls on every cache hit.
canonical = os.path.realpath(os.fspath(motion_dir))
digest = hashlib.sha1(f"{_CACHE_VERSION}\0{canonical}".encode("utf-8")).hexdigest()
return _motion_cache_root() / digest[:2] / f"{digest}.npy"
def _load_array(path: Path) -> np.ndarray:
array = np.load(path, allow_pickle=False)
if array.dtype != np.float32 or array.ndim != 2 or array.shape[1] != 36:
raise ValueError(
f"Invalid staged video motion array at {path}: "
f"shape={array.shape} dtype={array.dtype}"
)
return array
def load_or_build_video_motion_array(
motion_dir: Path,
builder: Callable[[Path], np.ndarray],
) -> tuple[np.ndarray, bool]:
"""Return an exact float32 array from node-local storage.
Advisory per-file locks let ranks and DataLoader workers share one local
copy without modifying any source data. The boolean reports whether this
call built the local artifact.
"""
motion_dir = Path(motion_dir)
if not video_motion_local_cache_enabled():
return builder(motion_dir), False
cache_path = _motion_cache_path(motion_dir)
if cache_path.exists():
try:
return _load_array(cache_path), False
except (OSError, ValueError):
pass
cache_path.parent.mkdir(parents=True, exist_ok=True)
lock_path = cache_path.with_suffix(cache_path.suffix + ".lock")
with lock_path.open("a+b") as lock_file:
fcntl.flock(lock_file.fileno(), fcntl.LOCK_EX)
try:
if cache_path.exists():
try:
return _load_array(cache_path), False
except (OSError, ValueError):
cache_path.unlink(missing_ok=True)
array = np.asarray(builder(motion_dir), dtype=np.float32)
if array.ndim != 2 or array.shape[1] != 36:
raise ValueError(
f"Unexpected video motion array at {motion_dir}: shape={array.shape}"
)
tmp_path = cache_path.with_name(
f".{cache_path.name}.{os.getpid()}.{threading.get_ident()}.tmp"
)
try:
with tmp_path.open("wb") as f:
np.save(f, array, allow_pickle=False)
os.replace(tmp_path, cache_path)
finally:
tmp_path.unlink(missing_ok=True)
return array, True
finally:
fcntl.flock(lock_file.fileno(), fcntl.LOCK_UN)
def _iter_datasets(dataset: Any, visited: set[int] | None = None) -> Iterator[Any]:
if visited is None:
visited = set()
identity = id(dataset)
if identity in visited:
return
visited.add(identity)
if callable(getattr(dataset, "prefetch_omni_motion_source_keys", None)):
yield dataset
return
if isinstance(dataset, ConcatDataset):
for child in dataset.datasets:
yield from _iter_datasets(child, visited)
return
children = getattr(dataset, "datasets", None)
if isinstance(children, (list, tuple)):
for child in children:
yield from _iter_datasets(child, visited)
wrapped = getattr(dataset, "dataset", None)
if wrapped is not None:
yield from _iter_datasets(wrapped, visited)
raw_dataset = getattr(dataset, "raw_dataset", None)
if raw_dataset is not None:
yield from _iter_datasets(raw_dataset, visited)
def _prefetch_slot_count() -> int:
try:
return max(int(os.environ.get("PSI_VIDEO_MOTION_PREFETCH_SLOTS", "0")), 0)
except ValueError:
return 0
class _PrefetchSlot:
def __init__(self) -> None:
self.file = None
def __enter__(self):
slot_count = _prefetch_slot_count()
lock_dir_value = os.environ.get("PSI_VIDEO_MOTION_PREFETCH_LOCK_DIR", "").strip()
if slot_count <= 0 or not lock_dir_value:
return self
lock_dir = Path(lock_dir_value)
lock_dir.mkdir(parents=True, exist_ok=True)
first_slot = os.getpid() % slot_count
while self.file is None:
for offset in range(slot_count):
slot = (first_slot + offset) % slot_count
candidate = (lock_dir / f"slot.{slot:03d}.lock").open("a+b")
try:
fcntl.flock(candidate.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
except BlockingIOError:
candidate.close()
continue
self.file = candidate
break
if self.file is None:
time.sleep(0.02)
return self
def __exit__(self, exc_type, exc, traceback) -> None:
del exc_type, exc, traceback
if self.file is not None:
try:
fcntl.flock(self.file.fileno(), fcntl.LOCK_UN)
finally:
self.file.close()
self.file = None
def _stage_motion_shard(dataset: Any, shard_key: str) -> None:
leaves = {
str(leaf.omni_oss_source_group()): leaf
for leaf in _iter_datasets(dataset)
if callable(getattr(leaf, "omni_oss_source_group", None))
}
if not leaves:
return
shard_path = Path(shard_key)
index_path = shard_path.parent / "index.rank00.sqlite"
sqlite_path = local_sqlite_path(index_path)
connection = sqlite3.connect(
f"file:{sqlite_path}?mode=ro&immutable=1",
uri=True,
timeout=120.0,
)
try:
row = connection.execute(
"SELECT shard_id FROM shards WHERE shard=?",
(shard_path.name,),
).fetchone()
if row is None:
raise KeyError(f"OSS shard is absent from its index: {shard_path}")
grouped_keys: dict[str, list[str]] = defaultdict(list)
for source_group, source_key in connection.execute(
"SELECT source_group, source_key FROM samples "
"WHERE shard_id=? ORDER BY sample_idx",
(int(row[0]),),
):
source_group = str(source_group)
if source_group in leaves:
grouped_keys[source_group].append(str(source_key))
finally:
connection.close()
started = time.perf_counter()
source_keys = 0
source_motions = 0
built = 0
with _PrefetchSlot():
for source_group, keys in grouped_keys.items():
counts = leaves[source_group].prefetch_omni_motion_source_keys(keys)
source_keys += int(counts[0])
source_motions += int(counts[1])
built += int(counts[2])
if os.environ.get("PSI_VIDEO_MOTION_PREFETCH_LOG", "0").strip().lower() in _TRUE_VALUES:
print(
f"[VIDEO MOTION PREFETCH] pid={os.getpid()} shard={shard_path.name} "
f"keys={source_keys:,} motions={source_motions:,} built={built:,} "
f"elapsed={time.perf_counter() - started:.3f}s",
flush=True,
)
def _prefetch_worker(dataset: Any, shard_key: str) -> None:
global _PREFETCH_THREAD, _PREFETCH_CURRENT_KEY, _PREFETCH_PENDING
current_dataset = dataset
current_key: str | None = shard_key
while current_key is not None:
try:
_stage_motion_shard(current_dataset, current_key)
except Exception as exc:
if os.environ.get("PSI_VIDEO_MOTION_PREFETCH_LOG", "0").strip().lower() in _TRUE_VALUES:
print(
f"[VIDEO MOTION PREFETCH] pid={os.getpid()} failed "
f"{Path(current_key).name}: {exc}",
flush=True,
)
with _PREFETCH_LOCK:
_PREFETCHED_KEYS.append(current_key)
pending = _PREFETCH_PENDING
_PREFETCH_PENDING = None
if pending is None:
_PREFETCH_THREAD = None
_PREFETCH_CURRENT_KEY = None
current_key = None
else:
current_dataset, current_key = pending
_PREFETCH_CURRENT_KEY = current_key
def prefetch_video_motion_for_omni_shard(dataset: Any, shard_path: str | Path | None) -> None:
"""Stage only the next video shard's motion arrays, not full batches."""
global _PREFETCH_THREAD, _PREFETCH_CURRENT_KEY, _PREFETCH_PENDING
if not video_motion_local_cache_enabled():
return
if os.environ.get("PSI_VIDEO_MOTION_PREFETCH", "0").strip().lower() not in _TRUE_VALUES:
return
if shard_path is None:
return
if not any(True for _ in _iter_datasets(dataset)):
return
shard_key = os.path.realpath(os.fspath(shard_path))
with _PREFETCH_LOCK:
if (
shard_key == _PREFETCH_CURRENT_KEY
or shard_key in _PREFETCHED_KEYS
or (_PREFETCH_PENDING is not None and shard_key == _PREFETCH_PENDING[1])
):
return
thread = _PREFETCH_THREAD
if thread is not None and thread.is_alive():
_PREFETCH_PENDING = (dataset, shard_key)
return
thread = threading.Thread(
target=_prefetch_worker,
args=(dataset, shard_key),
name="video-motion-shard-prefetch",
daemon=True,
)
_PREFETCH_THREAD = thread
_PREFETCH_CURRENT_KEY = shard_key
thread.start()
|