File size: 16,129 Bytes
52da7b7 | 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 | from __future__ import annotations
from dataclasses import dataclass
import time
from typing import Sequence
try: # pragma: no cover - exercised when NumPy is available in runtime envs.
import numpy as np
except Exception: # pragma: no cover
np = None # type: ignore[assignment]
try: # pragma: no cover - optional native ANN backend.
import faiss
except Exception: # pragma: no cover
faiss = None # type: ignore[assignment]
@dataclass(frozen=True)
class SparseSelection:
positions: list[int]
scores: list[float]
def _require_numpy() -> None:
if np is None:
raise RuntimeError("NumPy is required for the sparse-context kernel.")
def normalize_rows(matrix: object) -> object:
_require_numpy()
values = np.asarray(matrix, dtype=np.float32)
if values.ndim != 2:
raise ValueError("matrix must be rank-2")
norms = np.linalg.norm(values, axis=1, keepdims=True)
return values / np.maximum(norms, 1e-8)
class AnalyticalSparseAttention:
"""Content-dependent long-context selection from corpus-derived embeddings.
This is Reframr's analytical sparse-context kernel: it selects positions by
embedding geometry, then aggregates only the selected states. It does not
contain task-specific answer strings or prompt-pattern shortcuts.
"""
def __init__(self, embeddings: object, *, k_neighbors: int = 64) -> None:
_require_numpy()
self.embeddings = np.asarray(embeddings, dtype=np.float32)
if self.embeddings.ndim != 2:
raise ValueError("embeddings must be rank-2")
self.k_neighbors = max(1, int(k_neighbors))
self.normalized_embeddings = normalize_rows(self.embeddings)
self._context_token_ids: object | None = None
self._context_vectors: object | None = None
@property
def embedding_dim(self) -> int:
return int(self.embeddings.shape[1])
def select_positions(
self,
query_token_id: int,
context_token_ids: Sequence[int] | object,
*,
top_k: int | None = None,
) -> SparseSelection:
token_ids = self._coerce_token_ids(context_token_ids)
context_vectors = self.normalized_embeddings[token_ids]
return self._select_positions_from_vectors(
query_token_id,
token_ids,
context_vectors,
top_k=top_k,
)
def build_context_index(self, context_token_ids: Sequence[int] | object) -> None:
token_ids = self._coerce_token_ids(context_token_ids)
self._context_token_ids = token_ids
self._context_vectors = self.normalized_embeddings[token_ids]
def select_positions_cached(
self,
query_token_id: int,
*,
top_k: int | None = None,
) -> SparseSelection:
if self._context_token_ids is None or self._context_vectors is None:
raise RuntimeError("call build_context_index() before select_positions_cached()")
return self._select_positions_from_vectors(
query_token_id,
self._context_token_ids,
self._context_vectors,
top_k=top_k,
)
def _select_positions_from_vectors(
self,
query_token_id: int,
token_ids: object,
context_vectors: object,
*,
top_k: int | None = None,
) -> SparseSelection:
if token_ids.size == 0:
return SparseSelection(positions=[], scores=[])
query_id = int(query_token_id)
if query_id < 0 or query_id >= self.normalized_embeddings.shape[0]:
raise ValueError("query_token_id is outside the embedding table")
k = min(token_ids.size, max(1, int(top_k or self.k_neighbors)))
query_vector = self.normalized_embeddings[query_id]
scores = context_vectors @ query_vector
if k >= scores.size:
selected = np.argsort(scores)[::-1]
else:
selected = np.argpartition(scores, -k)[-k:]
selected = selected[np.argsort(scores[selected])[::-1]]
return SparseSelection(
positions=[int(index) for index in selected.tolist()],
scores=[float(scores[index]) for index in selected.tolist()],
)
def sparse_output(
self,
query_token_id: int,
context_token_ids: Sequence[int] | object,
context_states: object | None = None,
*,
top_k: int | None = None,
temperature: float = 1.0,
) -> object:
token_ids = self._coerce_token_ids(context_token_ids)
if context_states is None:
states = self.embeddings[token_ids]
else:
states = np.asarray(context_states, dtype=np.float32)
if states.ndim != 2 or states.shape[0] != token_ids.size:
raise ValueError("context_states must be rank-2 and match context length")
selection = self.select_positions(query_token_id, token_ids, top_k=top_k)
if not selection.positions:
return np.zeros(states.shape[1], dtype=np.float32)
selected_states = states[np.asarray(selection.positions, dtype=np.int64)]
scores = np.asarray(selection.scores, dtype=np.float32)
scaled = scores / max(float(temperature), 1e-6)
scaled -= float(scaled.max())
weights = np.exp(scaled)
weights /= max(float(weights.sum()), 1e-8)
return weights @ selected_states
def benchmark_selection(
self,
context_token_ids: Sequence[int] | object,
query_token_ids: Sequence[int] | object,
*,
top_k: int | None = None,
cache_context: bool = True,
) -> dict[str, object]:
token_ids = self._coerce_token_ids(context_token_ids)
queries = self._coerce_token_ids(query_token_ids)
build_started = time.perf_counter()
if cache_context:
self.build_context_index(token_ids)
build_elapsed = time.perf_counter() - build_started
started = time.perf_counter()
selected_total = 0
for query_id in queries.tolist():
if cache_context:
selection = self.select_positions_cached(int(query_id), top_k=top_k)
else:
selection = self.select_positions(int(query_id), token_ids, top_k=top_k)
selected_total += len(selection.positions)
elapsed = time.perf_counter() - started
return {
"context_tokens": int(token_ids.size),
"query_count": int(queries.size),
"top_k": min(int(top_k or self.k_neighbors), int(token_ids.size)) if token_ids.size else 0,
"selected_positions": int(selected_total),
"cache_context": bool(cache_context),
"index_build_seconds": build_elapsed,
"seconds": elapsed,
"queries_per_second": (float(queries.size) / elapsed) if elapsed > 0.0 else 0.0,
}
def _coerce_token_ids(self, token_ids: Sequence[int] | object) -> object:
ids = np.asarray(token_ids, dtype=np.int64)
if ids.ndim != 1:
raise ValueError("token ids must be rank-1")
if ids.size and (int(ids.min()) < 0 or int(ids.max()) >= self.embeddings.shape[0]):
raise ValueError("context token id is outside the embedding table")
return ids
def compare_selectors(
embeddings: object,
context_token_ids: Sequence[int] | object,
query_token_ids: Sequence[int] | object,
*,
top_k: int = 64,
hash_bits: int = 12,
probe_radius: int = 1,
seed: int = 2026,
) -> dict[str, object]:
_require_numpy()
exact = AnalyticalSparseAttention(embeddings, k_neighbors=top_k)
hashed = HashedSparseAttention(
embeddings,
k_neighbors=top_k,
hash_bits=hash_bits,
probe_radius=probe_radius,
seed=seed,
)
token_ids = exact._coerce_token_ids(context_token_ids)
queries = exact._coerce_token_ids(query_token_ids)
hashed.build_context_index(token_ids)
recalls: list[float] = []
for query_id in queries.tolist():
exact_positions = set(exact.select_positions(int(query_id), token_ids, top_k=top_k).positions)
hashed_positions = set(hashed.select_positions_cached(int(query_id), top_k=top_k).positions)
if not exact_positions:
recalls.append(1.0)
else:
recalls.append(len(exact_positions & hashed_positions) / len(exact_positions))
return {
"context_tokens": int(token_ids.size),
"query_count": int(queries.size),
"top_k": int(top_k),
"hash_bits": int(hash_bits),
"probe_radius": int(probe_radius),
"mean_recall_at_k": float(sum(recalls) / len(recalls)) if recalls else 0.0,
"min_recall_at_k": float(min(recalls)) if recalls else 0.0,
}
class HashedSparseAttention(AnalyticalSparseAttention):
"""Approximate sparse selector using deterministic random-hyperplane buckets.
It keeps the analytical embedding-geometry rule, but avoids scanning the full
context for every query. Buckets are built once from signs of fixed
hyperplane projections; each query scans only matching buckets, then reranks
the candidate set exactly by cosine similarity.
"""
def __init__(
self,
embeddings: object,
*,
k_neighbors: int = 64,
hash_bits: int = 12,
probe_radius: int = 1,
seed: int = 2026,
candidate_multiplier: int = 12,
) -> None:
super().__init__(embeddings, k_neighbors=k_neighbors)
self.hash_bits = max(1, int(hash_bits))
self.probe_radius = max(0, int(probe_radius))
self.candidate_multiplier = max(1, int(candidate_multiplier))
rng = np.random.default_rng(int(seed))
self.hyperplanes = rng.normal(
size=(self.embedding_dim, self.hash_bits)
).astype(np.float32)
self._bucket_positions: dict[int, list[int]] = {}
def build_context_index(self, context_token_ids: Sequence[int] | object) -> None:
token_ids = self._coerce_token_ids(context_token_ids)
self._context_token_ids = token_ids
self._context_vectors = self.normalized_embeddings[token_ids]
codes = self._codes_for_vectors(self._context_vectors)
buckets: dict[int, list[int]] = {}
for position, code in enumerate(codes.tolist()):
buckets.setdefault(int(code), []).append(position)
self._bucket_positions = buckets
def select_positions_cached(
self,
query_token_id: int,
*,
top_k: int | None = None,
) -> SparseSelection:
if self._context_token_ids is None or self._context_vectors is None:
raise RuntimeError("call build_context_index() before select_positions_cached()")
query_id = int(query_token_id)
if query_id < 0 or query_id >= self.normalized_embeddings.shape[0]:
raise ValueError("query_token_id is outside the embedding table")
k = min(self._context_token_ids.size, max(1, int(top_k or self.k_neighbors)))
candidate_positions = self._candidate_positions(query_id, k)
if len(candidate_positions) < k:
return super().select_positions_cached(query_id, top_k=top_k)
positions = np.asarray(candidate_positions, dtype=np.int64)
query_vector = self.normalized_embeddings[query_id]
scores = self._context_vectors[positions] @ query_vector
if k >= scores.size:
selected_local = np.argsort(scores)[::-1]
else:
selected_local = np.argpartition(scores, -k)[-k:]
selected_local = selected_local[np.argsort(scores[selected_local])[::-1]]
selected_positions = positions[selected_local]
return SparseSelection(
positions=[int(index) for index in selected_positions.tolist()],
scores=[float(scores[index]) for index in selected_local.tolist()],
)
def _candidate_positions(self, query_token_id: int, k: int) -> list[int]:
query_vector = self.normalized_embeddings[int(query_token_id)].reshape(1, -1)
query_code = int(self._codes_for_vectors(query_vector)[0])
candidate_limit = max(k, k * self.candidate_multiplier)
candidates: list[int] = []
seen: set[int] = set()
for code in self._probe_codes(query_code):
for position in self._bucket_positions.get(code, []):
if position in seen:
continue
seen.add(position)
candidates.append(position)
if len(candidates) >= candidate_limit:
return candidates
return candidates
def _codes_for_vectors(self, vectors: object) -> object:
projections = np.asarray(vectors, dtype=np.float32) @ self.hyperplanes
bits = projections >= 0.0
codes = np.zeros(bits.shape[0], dtype=np.int64)
for bit_index in range(self.hash_bits):
codes |= bits[:, bit_index].astype(np.int64) << bit_index
return codes
def _probe_codes(self, code: int) -> list[int]:
codes = [int(code)]
if self.probe_radius >= 1:
codes.extend(int(code) ^ (1 << bit) for bit in range(self.hash_bits))
if self.probe_radius >= 2:
for first in range(self.hash_bits):
for second in range(first + 1, self.hash_bits):
codes.append(int(code) ^ (1 << first) ^ (1 << second))
return codes
class FaissSparseAttention(AnalyticalSparseAttention):
"""Native FAISS-backed sparse selector over normalized embedding geometry."""
def __init__(
self,
embeddings: object,
*,
k_neighbors: int = 64,
approximate: bool = False,
hnsw_neighbors: int = 32,
ef_search: int = 64,
) -> None:
if faiss is None:
raise RuntimeError("faiss-cpu is not installed")
super().__init__(embeddings, k_neighbors=k_neighbors)
self.approximate = bool(approximate)
self.hnsw_neighbors = max(4, int(hnsw_neighbors))
self.ef_search = max(int(k_neighbors), int(ef_search))
self.index = self._new_index()
def _new_index(self) -> object:
if self.approximate:
index = faiss.IndexHNSWFlat(
self.embedding_dim,
self.hnsw_neighbors,
faiss.METRIC_INNER_PRODUCT,
)
index.hnsw.efSearch = self.ef_search
index.hnsw.efConstruction = max(self.ef_search, self.hnsw_neighbors * 2)
return index
return faiss.IndexFlatIP(self.embedding_dim)
def build_context_index(self, context_token_ids: Sequence[int] | object) -> None:
token_ids = self._coerce_token_ids(context_token_ids)
self._context_token_ids = token_ids
self._context_vectors = np.ascontiguousarray(
self.normalized_embeddings[token_ids],
dtype=np.float32,
)
self.index = self._new_index()
self.index.add(self._context_vectors)
def select_positions_cached(
self,
query_token_id: int,
*,
top_k: int | None = None,
) -> SparseSelection:
if self._context_token_ids is None or self._context_vectors is None:
raise RuntimeError("call build_context_index() before select_positions_cached()")
query_id = int(query_token_id)
if query_id < 0 or query_id >= self.normalized_embeddings.shape[0]:
raise ValueError("query_token_id is outside the embedding table")
k = min(self._context_token_ids.size, max(1, int(top_k or self.k_neighbors)))
query = np.ascontiguousarray(
self.normalized_embeddings[query_id].reshape(1, -1),
dtype=np.float32,
)
scores, indices = self.index.search(query, k)
valid = indices[0] >= 0
return SparseSelection(
positions=[int(index) for index in indices[0][valid].tolist()],
scores=[float(score) for score in scores[0][valid].tolist()],
)
|