File size: 16,369 Bytes
d38c1d3 | 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 | """
Memory-mapped CSR (Compressed Sparse Row) storage for co-occurrence graphs.
Replaces Python dict-of-dicts with three numpy.memmap arrays:
indptr.bin — row pointers (V+1 int32)
indices.bin — column indices (nnz int32), sorted within each row
data.bin — edge weights (nnz float32)
43M edges: ~346MB on disk, ~50MB RSS (OS pages in only accessed rows).
Python dicts for the same: ~4GB. 11x reduction.
Write-Ahead Log (WAL) buffers real-time edge updates (teach/RLHF)
without rebuilding the CSR. Reads merge CSR + WAL transparently.
"""
import os
import math
import threading
import numpy as np
try:
from scipy.sparse import csr_matrix as scipy_csr, coo_matrix as scipy_coo
from scipy.sparse.linalg import norm as sparse_norm_scipy
HAS_SCIPY = True
except ImportError:
HAS_SCIPY = False
class MMapCSR:
"""Memory-mapped CSR matrix backed by three flat files.
Zero-copy row access via numpy views into mmap'd arrays.
OS page cache manages hot/cold rows automatically.
"""
def __init__(self, path: str, readonly: bool = True):
"""Open an existing CSR from disk.
Args:
path: directory containing indptr.bin, indices.bin, data.bin
readonly: if True, open read-only (no writes to mmap)
"""
self._path = path
mode = 'r' if readonly else 'r+'
self._indptr = np.memmap(
os.path.join(path, 'indptr.bin'),
dtype=np.int32, mode=mode,
)
self._indices = np.memmap(
os.path.join(path, 'indices.bin'),
dtype=np.int32, mode=mode,
)
self._data = np.memmap(
os.path.join(path, 'data.bin'),
dtype=np.float32, mode=mode,
)
self.num_rows = len(self._indptr) - 1
self.nnz = len(self._indices)
# Pre-compute per-row norms for fast cosine
self._row_norms = None # lazy
# scipy view (lazy)
self._scipy_mat = None
@property
def scipy_matrix(self):
"""scipy.sparse.csr_matrix view over the same mmap'd data. No copy."""
if self._scipy_mat is None:
if not HAS_SCIPY:
raise ImportError("scipy required for spmv. pip install scipy")
self._scipy_mat = scipy_csr(
(self._data, self._indices, self._indptr),
shape=(self.num_rows, self.num_rows),
)
return self._scipy_mat
def get_row(self, row_idx: int):
"""Get (col_indices, weights) for a row. Zero-copy numpy views.
Returns:
(np.ndarray[int32], np.ndarray[float32]) — column indices and weights
"""
if row_idx < 0 or row_idx >= self.num_rows:
return np.array([], dtype=np.int32), np.array([], dtype=np.float32)
start = int(self._indptr[row_idx])
end = int(self._indptr[row_idx + 1])
if start == end:
return np.array([], dtype=np.int32), np.array([], dtype=np.float32)
return self._indices[start:end], self._data[start:end]
def get_row_dict(self, row_idx: int) -> dict:
"""Get row as {col_idx: weight} dict. Compatibility shim for convergence."""
cols, vals = self.get_row(row_idx)
if len(cols) == 0:
return {}
return dict(zip(cols.astype(int), vals.astype(float)))
def row_dot(self, row_idx: int, query: dict) -> float:
"""Dot product between a CSR row and a sparse query dict."""
cols, vals = self.get_row(row_idx)
if len(cols) == 0:
return 0.0
total = 0.0
for i in range(len(cols)):
c = int(cols[i])
if c in query:
total += float(vals[i]) * query[c]
return total
def row_norm(self, row_idx: int) -> float:
"""L2 norm of a row."""
_, vals = self.get_row(row_idx)
if len(vals) == 0:
return 0.0
return float(np.sqrt(np.dot(vals, vals)))
def row_cosine(self, row_idx: int, query: dict, q_norm: float = 0.0) -> float:
"""Cosine similarity between a CSR row and a sparse query dict."""
dot = self.row_dot(row_idx, query)
if dot == 0.0:
return 0.0
rn = self.row_norm(row_idx)
if rn == 0.0:
return 0.0
if q_norm == 0.0:
q_norm = math.sqrt(sum(v * v for v in query.values()))
if q_norm == 0.0:
return 0.0
return dot / (rn * q_norm)
def spmv(self, query_vec: np.ndarray) -> np.ndarray:
"""Sparse matrix-vector multiply: CSR @ query_vec.
Uses scipy BLAS if available, else manual row iteration.
"""
if HAS_SCIPY:
return np.asarray(self.scipy_matrix @ query_vec).ravel()
# Fallback: manual
result = np.zeros(self.num_rows, dtype=np.float32)
for i in range(self.num_rows):
start = int(self._indptr[i])
end = int(self._indptr[i + 1])
if start == end:
continue
cols = self._indices[start:end]
vals = self._data[start:end]
result[i] = np.dot(vals, query_vec[cols])
return result
@property
def row_norms_vec(self):
"""Precomputed L2 norms for all rows. Lazy, cached."""
if self._row_norms is None:
if HAS_SCIPY:
mat = self.scipy_matrix
# Per-row L2 norm
self._row_norms = np.sqrt(
np.asarray(mat.multiply(mat).sum(axis=1)).ravel()
)
else:
norms = np.zeros(self.num_rows, dtype=np.float32)
for i in range(self.num_rows):
_, vals = self.get_row(i)
if len(vals) > 0:
norms[i] = float(np.sqrt(np.dot(vals, vals)))
self._row_norms = norms
return self._row_norms
def stats(self) -> dict:
"""Return size stats."""
return {
'num_rows': self.num_rows,
'nnz': self.nnz,
'avg_degree': self.nnz / max(self.num_rows, 1),
'disk_mb': (
os.path.getsize(os.path.join(self._path, 'indptr.bin')) +
os.path.getsize(os.path.join(self._path, 'indices.bin')) +
os.path.getsize(os.path.join(self._path, 'data.bin'))
) / (1024 * 1024),
}
class CSRWriteAheadLog:
"""Buffers real-time edge updates without rebuilding CSR.
teach() writes to WAL + LMDB. ask() reads CSR + WAL overlay merged.
When WAL exceeds max_entries, signal for background CSR rebuild.
"""
def __init__(self, max_entries: int = 100_000):
self._log = {} # word_idx → {neighbor_idx: weight}
self._count = 0
self._max = max_entries
self._lock = threading.Lock()
@property
def needs_flush(self) -> bool:
return self._count >= self._max
@property
def entry_count(self) -> int:
return self._count
def add_edge(self, word_a: int, word_b: int, weight: float):
"""Buffer a co-occurrence edge update."""
with self._lock:
if word_a not in self._log:
self._log[word_a] = {}
if word_b not in self._log:
self._log[word_b] = {}
old_a = self._log[word_a].get(word_b, 0)
old_b = self._log[word_b].get(word_a, 0)
self._log[word_a][word_b] = old_a + weight
self._log[word_b][word_a] = old_b + weight
self._count += 2
def get_overlay(self, word_idx: int) -> dict:
"""Get pending WAL updates for one word. Returns {} if none."""
with self._lock:
return dict(self._log.get(word_idx, {}))
def merge_read(self, csr: MMapCSR, word_idx: int) -> dict:
"""Read CSR row + WAL overlay, merged into one dict.
This is the primary read path for convergence.
"""
# Base from CSR (zero-copy read, converted to dict)
if word_idx < csr.num_rows:
base = csr.get_row_dict(word_idx)
else:
# New word beyond CSR — lives entirely in WAL
base = {}
# Overlay from WAL
overlay = self.get_overlay(word_idx)
if not overlay:
# Self-connection if empty
if not base:
return {word_idx: 1.0}
return base
# Merge: WAL adds to base
merged = dict(base)
for k, v in overlay.items():
merged[k] = merged.get(k, 0) + v
if not merged:
merged[word_idx] = 1.0
return merged
def clear(self):
"""Reset WAL after successful CSR rebuild."""
with self._lock:
self._log.clear()
self._count = 0
self._scipy_cache = None
def snapshot(self) -> dict:
"""Get a frozen copy of all WAL entries (for CSR rebuild)."""
with self._lock:
return {k: dict(v) for k, v in self._log.items()}
# --- Persistence (LMDB-backed, crash-safe) ---
_EDGE_FMT = '<iif' # word_a, word_b, weight — 12 bytes per edge
_EDGE_SIZE = 12
def persist_to_lmdb(self, env, wal_db):
"""Write all WAL edges to LMDB wal_edges sub-db. Atomic transaction."""
import struct
snapshot = self.snapshot()
if not snapshot:
return 0
# Flatten to unique directed edges (a→b, not both a→b and b→a)
seen = set()
edges = []
for word_a, neighbors in snapshot.items():
for word_b, weight in neighbors.items():
key = (min(word_a, word_b), max(word_a, word_b))
if key not in seen:
seen.add(key)
# Clamp to float32 range to prevent overflow
clamped = max(-1e30, min(1e30, weight))
edges.append((word_a, word_b, clamped))
with env.begin(write=True) as txn:
# Clear old WAL entries
txn.drop(wal_db, delete=False)
# Write all edges as sequential keys
for i, (a, b, w) in enumerate(edges):
txn.put(
struct.pack('<i', i),
struct.pack(self._EDGE_FMT, a, b, w),
db=wal_db,
)
txn.put(b'__count__', struct.pack('<i', len(edges)), db=wal_db)
return len(edges)
def load_from_lmdb(self, env, wal_db):
"""Load persisted WAL edges from LMDB into memory. Called at startup."""
import struct
loaded = 0
with env.begin(db=wal_db) as txn:
count_raw = txn.get(b'__count__', db=wal_db)
if not count_raw:
return 0
count = struct.unpack('<i', count_raw)[0]
for i in range(count):
val = txn.get(struct.pack('<i', i), db=wal_db)
if val and len(val) == self._EDGE_SIZE:
a, b, w = struct.unpack(self._EDGE_FMT, val)
# Add directly to log without double-counting
if a not in self._log:
self._log[a] = {}
if b not in self._log:
self._log[b] = {}
self._log[a][b] = self._log[a].get(b, 0) + w
self._log[b][a] = self._log[b].get(a, 0) + w
self._count += 2
loaded += 1
return loaded
# --- Background flush thread ---
_flush_thread = None
_flush_stop = None
_flush_env = None
_flush_db = None
_last_persisted_count = 0
def start_background_flush(self, env, wal_db, interval_sec: float = 5.0):
"""Start background thread that persists WAL to LMDB every interval_sec."""
import time as _time
self._flush_env = env
self._flush_db = wal_db
self._flush_stop = threading.Event()
self._last_persisted_count = self._count
def _flush_loop():
while not self._flush_stop.is_set():
self._flush_stop.wait(interval_sec)
if self._flush_stop.is_set():
break
if self._count != self._last_persisted_count:
self.persist_to_lmdb(env, wal_db)
self._last_persisted_count = self._count
self._flush_thread = threading.Thread(target=_flush_loop, daemon=True)
self._flush_thread.start()
def stop_background_flush(self):
"""Stop background flush and do a final persist."""
if self._flush_stop:
self._flush_stop.set()
if self._flush_thread:
self._flush_thread.join(timeout=5)
# Final flush
if self._flush_env and self._flush_db and self._count > self._last_persisted_count:
self.persist_to_lmdb(self._flush_env, self._flush_db)
_scipy_cache = None
_scipy_cache_count = -1
def effective_scipy_matrix(self, csr: 'MMapCSR'):
"""Get CSR + WAL merged as a single scipy sparse matrix. Cached.
Rebuilds only when WAL has changed since last call.
"""
if not HAS_SCIPY:
raise ImportError("scipy required")
if self._scipy_cache is not None and self._scipy_cache_count == self._count:
return self._scipy_cache
base = csr.scipy_matrix
if self._count == 0:
self._scipy_cache = base
self._scipy_cache_count = 0
return base
# Build COO from WAL entries
rows, cols, vals = [], [], []
with self._lock:
for word_idx, neighbors in self._log.items():
for neighbor_idx, weight in neighbors.items():
rows.append(word_idx)
cols.append(neighbor_idx)
vals.append(weight)
count_now = self._count
if not rows:
self._scipy_cache = base
self._scipy_cache_count = count_now
return base
# Expand shape if WAL has new words beyond CSR
max_idx = max(max(rows), max(cols), base.shape[0] - 1)
shape = (max_idx + 1, max_idx + 1)
if shape != base.shape:
base.resize(shape) # in-place, returns None
wal_sparse = scipy_coo(
(np.array(vals, dtype=np.float32),
(np.array(rows, dtype=np.int32), np.array(cols, dtype=np.int32))),
shape=shape,
).tocsr()
self._scipy_cache = base + wal_sparse
self._scipy_cache_count = count_now
return self._scipy_cache
def build_csr_from_dicts(cooc: dict, num_words: int, output_path: str):
"""Build CSR files from a dict-of-dicts co-occurrence graph.
Args:
cooc: {word_idx: {neighbor_idx: weight, ...}, ...}
num_words: total vocabulary size (V)
output_path: directory to write indptr.bin, indices.bin, data.bin
"""
os.makedirs(output_path, exist_ok=True)
# Pass 1: count edges per row
row_counts = np.zeros(num_words, dtype=np.int32)
for row_idx in range(num_words):
if row_idx in cooc:
row_counts[row_idx] = len(cooc[row_idx])
# Build indptr
indptr = np.zeros(num_words + 1, dtype=np.int32)
np.cumsum(row_counts, out=indptr[1:])
nnz = int(indptr[-1])
# Pass 2: fill indices and data, sorted within rows
indices = np.zeros(nnz, dtype=np.int32)
data = np.zeros(nnz, dtype=np.float32)
for row_idx in range(num_words):
if row_idx not in cooc or not cooc[row_idx]:
continue
start = int(indptr[row_idx])
edges = sorted(cooc[row_idx].items(), key=lambda x: x[0])
for j, (col, weight) in enumerate(edges):
indices[start + j] = col
data[start + j] = weight
# Write files
indptr_path = os.path.join(output_path, 'indptr.bin')
indices_path = os.path.join(output_path, 'indices.bin')
data_path = os.path.join(output_path, 'data.bin')
indptr.tofile(indptr_path)
indices.tofile(indices_path)
data.tofile(data_path)
print(f"CSR built: {num_words:,} rows, {nnz:,} edges, "
f"{(os.path.getsize(indptr_path) + os.path.getsize(indices_path) + os.path.getsize(data_path)) / 1024 / 1024:.1f} MB")
return indptr_path, indices_path, data_path
|