Spaces:
Running on Zero
Running on Zero
File size: 11,799 Bytes
3bd5145 | 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 | import os
import io
import json
import fcntl
import errno
from typing import Dict, Tuple, Optional, Any, List
import numpy as np
import torch
DEFAULT_META_FILENAME = "shards.meta.json"
DEFAULT_BIN_PATTERN = "data-{shard_id:05d}.bin"
DEFAULT_IDX_PATTERN = "data-{shard_id:05d}.idx"
DEFAULT_IDX_BIN_PATTERN = "data-{shard_id:05d}.idx.bin"
def _get_rank_and_world_size() -> Tuple[int, int]:
try:
# Accelerate / DDP envs
rank = int(os.environ.get("RANK", "0"))
world_size = int(os.environ.get("WORLD_SIZE", "1"))
except Exception:
rank, world_size = 0, 1
return rank, world_size
def _ensure_dir(path: str):
os.makedirs(path, exist_ok=True)
class FileLock:
def __init__(self, lock_path: str):
self.lock_path = lock_path
_ensure_dir(os.path.dirname(lock_path))
# Create lock file if not exists
self.fd = os.open(lock_path, os.O_CREAT | os.O_RDWR)
def acquire(self):
while True:
try:
fcntl.flock(self.fd, fcntl.LOCK_EX)
break
except OSError as e:
if e.errno != errno.EINTR:
raise
def release(self):
fcntl.flock(self.fd, fcntl.LOCK_UN)
def __enter__(self):
self.acquire()
return self
def __exit__(self, exc_type, exc, tb):
self.release()
class ShardedBinIdxWriter:
"""
Multi-shard cache writer. Each sample is serialized as a single NPZ blob and
appended to a shard's .bin file, with an entry in the shard's .idx text file:
"data_id\toffset\tlength\n".
Concurrency safety:
- Use a per-shard lock file to serialize append operations across processes.
- Meta file creation is idempotent and verified for consistency.
Monotonicity:
- Enforces per-shard strictly increasing data_id on append. If violated, raises ValueError.
"""
def __init__(self, cache_dir: str, num_shards: int = 16, meta_filename: str = DEFAULT_META_FILENAME):
self.cache_dir = cache_dir
self.num_shards = int(num_shards)
self.meta_filename = meta_filename
_ensure_dir(self.cache_dir)
self.meta_path = os.path.join(self.cache_dir, self.meta_filename)
self._init_or_validate_meta()
def _init_or_validate_meta(self):
rank, _ = _get_rank_and_world_size()
if not os.path.exists(self.meta_path):
# Create atomically using a lock on the meta file path
lock = FileLock(self.meta_path + ".lock")
with lock:
if not os.path.exists(self.meta_path):
meta = {
"version": 1,
"record_format": "npz",
"num_shards": self.num_shards,
"bin_pattern": DEFAULT_BIN_PATTERN,
"idx_pattern": DEFAULT_IDX_PATTERN,
}
with open(self.meta_path, "w") as f:
json.dump(meta, f)
# Validate
with open(self.meta_path, "r") as f:
meta = json.load(f)
if meta.get("num_shards") != self.num_shards:
raise ValueError(f"num_shards mismatch: existing {meta.get('num_shards')} vs new {self.num_shards}")
def _shard_paths(self, shard_id: int) -> Tuple[str, str, str, str]:
bin_path = os.path.join(self.cache_dir, DEFAULT_BIN_PATTERN.format(shard_id=shard_id))
idx_path = os.path.join(self.cache_dir, DEFAULT_IDX_PATTERN.format(shard_id=shard_id))
idx_bin_path = os.path.join(self.cache_dir, DEFAULT_IDX_BIN_PATTERN.format(shard_id=shard_id))
lock_path = idx_path + ".lock"
return bin_path, idx_path, idx_bin_path, lock_path
@staticmethod
def _to_numpy_inputs(sample: Dict[str, Any]) -> Dict[str, Any]:
numpy_inputs: Dict[str, Any] = {}
for key, value in sample.items():
if isinstance(value, torch.Tensor):
if value.dtype == torch.bfloat16:
numpy_inputs[key] = value.detach().cpu().half().numpy()
else:
numpy_inputs[key] = value.detach().cpu().numpy()
else:
numpy_inputs[key] = value
return numpy_inputs
def write_sample(self, data_id: int, sample: Dict[str, Any], compress: bool = True):
shard_id = int(data_id) % self.num_shards
bin_path, idx_path, idx_bin_path, lock_path = self._shard_paths(shard_id)
_ensure_dir(os.path.dirname(bin_path))
_ensure_dir(os.path.dirname(idx_path))
numpy_inputs = self._to_numpy_inputs(sample)
# Serialize to NPZ bytes first
buffer = io.BytesIO()
# Use compressed to reduce disk
if compress:
np.savez_compressed(buffer, **numpy_inputs)
else:
np.savez(buffer, **numpy_inputs)
blob = buffer.getvalue()
# Serialize append operations under a single lock
lock = FileLock(lock_path)
with lock:
# Enforce per-shard strictly increasing data_id by checking the last record
last_id = None
if os.path.exists(idx_bin_path):
file_size = os.path.getsize(idx_bin_path)
if file_size % (8 * 3) != 0:
raise ValueError(f"Corrupted idx.bin file: {idx_bin_path}")
if file_size >= (8 * 3):
with open(idx_bin_path, "rb") as bif:
bif.seek(file_size - (8 * 3))
tail = np.fromfile(bif, dtype=np.int64, count=3)
if tail.size == 3:
last_id = int(tail[0])
if last_id is not None and int(data_id) <= last_id:
raise ValueError(
f"data_id for shard {shard_id} must be strictly increasing; last={last_id}, got={int(data_id)}"
)
# Append NPZ bytes to the .bin shard.
with open(bin_path, "ab") as bf:
offset = bf.tell()
bf.write(blob)
length = len(blob)
# Append textual idx for human readability/debug
with open(idx_path, "a") as inf:
inf.write(f"{int(data_id)}\t{int(offset)}\t{int(length)}\n")
# Append binary idx for O(1) random access at read time.
with open(idx_bin_path, "ab") as bif:
np.array([int(data_id), int(offset), int(length)], dtype=np.int64).tofile(bif)
class ShardedBinIdxReader:
"""
Reader for multi-shard bin+idx cache written by ShardedBinIdxWriter.
Loads per-shard index into memory on first access for O(1) lookup.
"""
def __init__(self, cache_dir: str, meta_filename: str = DEFAULT_META_FILENAME):
self.cache_dir = cache_dir
self.meta_path = os.path.join(cache_dir, meta_filename)
if not os.path.exists(self.meta_path):
raise FileNotFoundError(f"Meta file not found: {self.meta_path}")
with open(self.meta_path, "r") as f:
meta = json.load(f)
self.num_shards = int(meta["num_shards"])
self.record_format = meta.get("record_format", "npz")
if self.record_format != "npz":
raise ValueError(f"Unsupported record_format: {self.record_format}")
# Cached per-shard structures: (mm, sort_idx, sorted_keys)
self._shard_arrays: Dict[int, Tuple[np.memmap, np.ndarray, np.ndarray]] = {}
def _shard_paths(self, shard_id: int) -> Tuple[str, str, str]:
bin_path = os.path.join(self.cache_dir, DEFAULT_BIN_PATTERN.format(shard_id=shard_id))
idx_path = os.path.join(self.cache_dir, DEFAULT_IDX_PATTERN.format(shard_id=shard_id))
idx_bin_path = os.path.join(self.cache_dir, DEFAULT_IDX_BIN_PATTERN.format(shard_id=shard_id))
return bin_path, idx_path, idx_bin_path
def _ensure_binary_idx(self, idx_path: str, idx_bin_path: str):
if os.path.exists(idx_bin_path):
return
if not os.path.exists(idx_path):
# Nothing to build
return
# Build binary idx from text idx (one-time cost)
rows: List[List[int]] = []
with open(idx_path, "r") as f:
for line in f:
line = line.strip()
if not line:
continue
parts = line.split("\t")
if len(parts) != 3:
continue
try:
did = int(parts[0])
off = int(parts[1])
leng = int(parts[2])
rows.append([did, off, leng])
except Exception:
continue
if len(rows) == 0:
return
arr = np.asarray(rows, dtype=np.int64)
with open(idx_bin_path, "wb") as bif:
arr.tofile(bif)
def _load_idx_if_needed(self, shard_id: int):
if shard_id in self._shard_arrays:
return
_, idx_path, idx_bin_path = self._shard_paths(shard_id)
self._ensure_binary_idx(idx_path, idx_bin_path)
if not os.path.exists(idx_bin_path):
# Empty shard
self._shard_arrays[shard_id] = (np.memmap(idx_bin_path, mode='w+', dtype=np.int64, shape=(0,)), np.array([], dtype=np.int64), np.array([], dtype=np.int64))
return
file_size = os.path.getsize(idx_bin_path)
if file_size == 0:
self._shard_arrays[shard_id] = (np.memmap(idx_bin_path, mode='r', dtype=np.int64, shape=(0,)), np.array([], dtype=np.int64), np.array([], dtype=np.int64))
return
if file_size % (8 * 3) != 0:
raise ValueError(f"Corrupted idx.bin file: {idx_bin_path}")
num_rows = file_size // (8 * 3)
mm = np.memmap(idx_bin_path, mode='r', dtype=np.int64, shape=(num_rows * 3,))
mm = mm.reshape(num_rows, 3)
keys = mm[:, 0]
# Use stable sort; build sorted view
sort_idx = np.argsort(keys, kind='mergesort')
sorted_keys = keys[sort_idx]
self._shard_arrays[shard_id] = (mm, sort_idx, sorted_keys)
def has(self, data_id: int) -> bool:
shard_id = int(data_id) % self.num_shards
self._load_idx_if_needed(shard_id)
mm, sort_idx, sorted_keys = self._shard_arrays.get(shard_id, (None, None, None))
if mm is None or sorted_keys is None or len(sorted_keys) == 0:
return False
did = int(data_id)
pos = np.searchsorted(sorted_keys, did, side='right') - 1
if pos < 0 or pos >= len(sorted_keys):
return False
return sorted_keys[pos] == did
def get(self, data_id: int) -> Optional[Dict[str, Any]]:
shard_id = int(data_id) % self.num_shards
bin_path, _, _ = self._shard_paths(shard_id)
self._load_idx_if_needed(shard_id)
mm, sort_idx, sorted_keys = self._shard_arrays.get(shard_id, (None, None, None))
if mm is None or len(sorted_keys) == 0:
return None
did = int(data_id)
pos = np.searchsorted(sorted_keys, did, side='right') - 1
if pos < 0 or pos >= len(sorted_keys) or sorted_keys[pos] != did:
return None
row = int(sort_idx[pos])
offset = int(mm[row, 1])
length = int(mm[row, 2])
if not os.path.exists(bin_path):
return None
with open(bin_path, "rb") as bf:
bf.seek(offset)
blob = bf.read(length)
buffer = io.BytesIO(blob)
npz = np.load(buffer)
loaded: Dict[str, Any] = {}
for k, v in npz.items():
if hasattr(v, "dtype"):
loaded[k] = torch.from_numpy(np.array(v))
else:
loaded[k] = v
return loaded
|