File size: 2,310 Bytes
3bc7886 | 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 | """static_embed.py PASSAGES.parquet [...] - first-stage engram keys with sentence-transformers/static-retrieval-mrl-en-v1
(Apache-2.0): mean of per-token static vectors (no special tokens), Matryoshka-truncated to DIM, L2-normalised, float16 ->
emb_static/<name>.npy + .pid.npy. CPU only, ~100x cheaper than a transformer encoder, so the whole of Wikipedia indexes on the
spare cores while Beam keeps both GPUs; the cross-encoder re-rank recovers precision (10-19). Also importable: embed(texts)."""
import os, sys, time
import numpy as np
from tokenizers import Tokenizer
from safetensors.numpy import load_file
S = "/mnt/optane/hf-cache/hub/models--sentence-transformers--static-retrieval-mrl-en-v1/snapshots/f60985c706f192d45d218078e49e5a8b6f15283a/0_StaticEmbedding"
DIM = int(os.environ.get("DIM", 256))
TOK = Tokenizer.from_file(f"{S}/tokenizer.json"); TOK.no_padding(); TOK.no_truncation()
W = load_file(f"{S}/model.safetensors")["embedding.weight"][:, :DIM].astype(np.float32)
def embed(texts, maxlen=512):
enc = TOK.encode_batch(texts, add_special_tokens=False)
E = np.empty((len(texts), DIM), np.float32)
for i, e in enumerate(enc):
ids = e.ids[:maxlen]
E[i] = W[ids].mean(0) if ids else 0
E /= np.linalg.norm(E, axis=1, keepdims=True) + 1e-12
return E
if __name__ == "__main__":
import pyarrow.parquet as pq
out = os.environ.get("OUT", "/mnt/optane/engram/emb_static"); os.makedirs(out, exist_ok=True)
for src in sys.argv[1:]:
name = os.path.basename(src)[:-8]
if os.path.exists(f"{out}/{name}.npy"):
continue
t0 = time.time(); pf = pq.ParquetFile(src); n = pf.metadata.num_rows # streamed: ~8 GB is all Beam leaves free
E = np.lib.format.open_memmap(f"{out}/{name}.tmp.npy", "w+", np.float16, (n, DIM)); P = np.empty(n, np.int64); i = 0
for b in pf.iter_batches(batch_size=8192, columns=["pid", "text"]):
E[i:i + b.num_rows] = embed(b.column("text").to_pylist()); P[i:i + b.num_rows] = b.column("pid").to_numpy(); i += b.num_rows
E.flush(); del E
np.save(f"{out}/{name}.pid.npy", P); os.replace(f"{out}/{name}.tmp.npy", f"{out}/{name}.npy"); txt = P
print(f"{time.strftime('%H:%M:%S')} {name}: {len(txt)} passages in {time.time() - t0:.0f} s", flush=True)
|