File size: 13,288 Bytes
067ccf9 4afb49d 067ccf9 4afb49d 067ccf9 4afb49d 067ccf9 4afb49d 067ccf9 4afb49d 067ccf9 4afb49d 067ccf9 4afb49d 067ccf9 4afb49d 067ccf9 | 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 | """HF Job (a100-large): zero-shot intfloat/multilingual-e5-large embeddings for all 24.2M normalised records.
Spec (matches the validated zero-shot evaluation):
* sentence-transformers, model loaded in float16 on CUDA, prompt "query: " on every text, normalize_embeddings=True
* variant "full": raw_name.strip() + " | " + raw_address.strip() (just the name when the address is empty), max_seq_length=80
* variant "name": raw_name.strip(), max_seq_length=32
* batch_size 512; sentence-transformers length-sorts inside each encode() call and restores the original order.
Each table is fed to encode() in contiguous chunks of CHUNK rows (for progress logging and bounded host memory);
rows are never reordered.
* output per table: emb/e5_{variant}/{table}.npy (float16, shape (n_rows, 1024), parquet row order)
emb/e5_{variant}/{table}_ids.parquet (entity_id, same order)
* order: variant "full" for s1,s2,s3,t1,t2,t3, then variant "name"; each table is committed to the Hub
(Arushhh/amlc-work, dataset) as soon as it is written, in a background thread so the GPU keeps working.
Job splitting (env): TABLES=s1,s2,s3 restricts the job to those tables (default: all six, in the order above);
JOB_TAG names this job's manifest emb/e5_manifest_{JOB_TAG}.json (default emb/e5_manifest.json);
SKIP_EXISTING=1 (default) skips a variant/table whose .npy (exact expected byte size) and _ids.parquet are already on the Hub.
In-job integrity checks per table: shape/dtype via np.load(mmap), non-finite count, min/max L2 norm over all rows,
and an order check (256 random rows re-encoded on their own must have cosine > 0.99 with the stored rows).
"""
import os, sys, time, json, shutil, threading
from concurrent.futures import ThreadPoolExecutor
import numpy as np
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from huggingface_hub import HfApi, snapshot_download, CommitOperationAdd
REPO = 'Arushhh/amlc-work'
MODEL = 'intfloat/multilingual-e5-large'
PROMPT = 'query: '
BATCH = 512
CHUNK = int(os.environ.get('CHUNK', 262144))
DIM = 1024
ALL_TABLES = ['s1', 's2', 's3', 't1', 't2', 't3']
TABLES = [t for t in ALL_TABLES if t in os.environ.get('TABLES', ','.join(ALL_TABLES)).split(',')]
JOB_TAG = os.environ.get('JOB_TAG', '')
MANIFEST = f'emb/e5_manifest_{JOB_TAG}.json' if JOB_TAG else 'emb/e5_manifest.json'
SKIP_EXISTING = os.environ.get('SKIP_EXISTING', '1') == '1'
VARIANTS = [('full', 80), ('name', 32)]
EXPECTED_ROWS = {'s1': 2206821, 's2': 5034616, 's3': 5285603, 't1': 1732544, 't2': 4887273, 't3': 5082316}
WORK = os.environ.get('WORKDIR', '/workspace/e5')
TOKEN = os.environ['HF_TOKEN']
T0 = time.time()
def log(*a):
print(f"[{time.strftime('%H:%M:%S', time.gmtime())} +{time.time() - T0:7.0f}s]", *a, flush=True)
def build_texts(names, addrs, variant):
out = []
n_empty_name = n_empty_addr = 0
for n, a in zip(names, addrs):
n = (n or '').strip()
if not n:
n_empty_name += 1
if variant == 'name':
out.append(n)
continue
a = (a or '').strip()
if a:
out.append(n + ' | ' + a)
else:
n_empty_addr += 1
out.append(n)
return out, n_empty_name, n_empty_addr
def main():
os.makedirs(WORK, exist_ok=True)
api = HfApi(token=TOKEN)
log('python', sys.version.split()[0], 'torch', torch.__version__, 'cuda', torch.version.cuda)
import sentence_transformers, transformers, huggingface_hub
log('sentence-transformers', sentence_transformers.__version__, 'transformers', transformers.__version__,
'huggingface_hub', huggingface_hub.__version__)
assert torch.cuda.is_available(), 'CUDA not available'
gpu = torch.cuda.get_device_name(0)
log('GPU:', gpu, f'{torch.cuda.get_device_properties(0).total_memory / 2**30:.1f} GiB', 'CPUs:', os.cpu_count())
du = shutil.disk_usage(WORK)
log(f'disk at {WORK}: free {du.free / 2**30:.0f} GiB of {du.total / 2**30:.0f} GiB')
log(f'TABLES={TABLES} JOB_TAG={JOB_TAG!r} manifest={MANIFEST} SKIP_EXISTING={SKIP_EXISTING}')
assert TABLES, 'no tables selected'
on_hub = {}
if SKIP_EXISTING:
try:
for f in api.list_repo_tree(REPO, path_in_repo='emb', recursive=True, repo_type='dataset'):
if getattr(f, 'size', None) is not None:
on_hub[f.path] = f.size
except Exception as ex: # noqa: BLE001 (no emb/ folder yet)
log('list_repo_tree emb/:', repr(ex))
skip = set()
for variant, _ in VARIANTS:
for table in TABLES:
npy_p, ids_p = f'emb/e5_{variant}/{table}.npy', f'emb/e5_{variant}/{table}_ids.parquet'
if on_hub.get(npy_p) == 128 + EXPECTED_ROWS[table] * DIM * 2 and ids_p in on_hub:
skip.add((variant, table))
log(f'SKIP {variant}/{table}: already on the Hub ({npy_p} {on_hub[npy_p]} bytes, {ids_p} present)')
t = time.time()
snapshot_download(REPO, repo_type='dataset', allow_patterns=['norm/*'], local_dir=f'{WORK}/hub', token=TOKEN)
log(f'downloaded norm/* in {time.time() - t:.0f}s:',
{f: os.path.getsize(f'{WORK}/hub/norm/{f}') for f in sorted(os.listdir(f'{WORK}/hub/norm'))})
from sentence_transformers import SentenceTransformer
t = time.time()
model = SentenceTransformer(MODEL, device='cuda', model_kwargs={'torch_dtype': torch.float16})
model.eval()
pdt = next(model.parameters()).dtype
if pdt != torch.float16:
log('params loaded as', pdt, '-> casting with .half()')
model.half()
pdt = next(model.parameters()).dtype
assert pdt == torch.float16, pdt
am = getattr(model[0].auto_model.config, '_attn_implementation', '?')
log(f'MODEL LOADED {MODEL} in {time.time() - t:.0f}s: param dtype {pdt}, attn {am}, modules',
[type(m).__name__ for m in model], 'dim', model.get_sentence_embedding_dimension())
assert model.get_sentence_embedding_dimension() == DIM
pool = ThreadPoolExecutor(max_workers=1)
futures = []
manifest = {'model': MODEL, 'prompt': PROMPT, 'normalize_embeddings': True, 'dtype': 'float16', 'batch_size': BATCH,
'chunk_rows': CHUNK, 'job_tables': TABLES, 'skipped_already_on_hub': sorted(f'{v}/{t}' for v, t in skip), 'gpu': gpu, 'torch': torch.__version__,
'sentence_transformers': sentence_transformers.__version__, 'transformers': transformers.__version__,
'tables': {}}
mlock = threading.Lock()
def upload(variant, table, npy, ids, manifest_bytes):
dest = f'emb/e5_{variant}'
for attempt in range(1, 6):
try:
t = time.time()
ops = [CommitOperationAdd(path_in_repo=f'{dest}/{table}.npy', path_or_fileobj=npy),
CommitOperationAdd(path_in_repo=f'{dest}/{table}_ids.parquet', path_or_fileobj=ids),
CommitOperationAdd(path_in_repo=MANIFEST, path_or_fileobj=manifest_bytes)]
api.create_commit(repo_id=REPO, repo_type='dataset', operations=ops,
commit_message=f'e5-large zero-shot {variant}/{table}')
sz = os.path.getsize(npy)
log(f'UPLOADED {dest}/{table}.npy ({sz} bytes) + {table}_ids.parquet in {time.time() - t:.0f}s '
f'({sz / 2**20 / max(time.time() - t, 1e-6):.0f} MiB/s)')
os.remove(npy)
return sz
except Exception as ex: # noqa: BLE001
log(f'upload {dest}/{table} attempt {attempt} FAILED: {ex!r}')
time.sleep(30 * attempt)
raise RuntimeError(f'upload of {dest}/{table} failed 5 times')
grand_n = 0
grand_t = 0.0
rng = np.random.default_rng(20260925)
for variant, max_len in VARIANTS:
model.max_seq_length = max_len
log(f'===== variant {variant}: max_seq_length={model.max_seq_length}, prompt={PROMPT!r}, batch={BATCH}')
os.makedirs(f'{WORK}/out/e5_{variant}', exist_ok=True)
var_n = 0
var_t = 0.0
for table in TABLES:
if (variant, table) in skip:
continue
tb = pq.read_table(f'{WORK}/hub/norm/{table}.parquet', columns=['entity_id', 'raw_name', 'raw_address'])
n = tb.num_rows
if n != EXPECTED_ROWS[table]:
log(f'WARNING {table}: {n} rows, expected {EXPECTED_ROWS[table]}')
texts, e_name, e_addr = build_texts(tb.column('raw_name').to_pylist(), tb.column('raw_address').to_pylist(),
variant)
log(f'{variant}/{table}: {n} rows, empty names {e_name}, empty addresses {e_addr}; examples: {texts[:2]}')
npy = f'{WORK}/out/e5_{variant}/{table}.npy'
ids = f'{WORK}/out/e5_{variant}/{table}_ids.parquet'
pq.write_table(pa.table({'entity_id': tb.column('entity_id')}), ids)
ids_arr = tb.column('entity_id')
del tb
out = np.lib.format.open_memmap(npy, mode='w+', dtype=np.float16, shape=(n, DIM))
t = time.time()
for s in range(0, n, CHUNK):
tc = time.time()
chunk = texts[s:s + CHUNK]
e = model.encode(chunk, batch_size=BATCH, prompt=PROMPT, normalize_embeddings=True,
convert_to_numpy=True, show_progress_bar=False)
assert e.shape == (len(chunk), DIM), e.shape
out[s:s + len(chunk)] = e.astype(np.float16, copy=False)
done = s + len(chunk)
el = time.time() - t
log(f' {variant}/{table} rows {done}/{n}: chunk {len(chunk) / (time.time() - tc):.0f} rec/s, '
f'table {done / el:.0f} rec/s, ETA table {(n - done) / (done / el) / 60:.1f} min')
enc_s = time.time() - t
out.flush()
del out
# ---- integrity checks
arr = np.load(npy, mmap_mode='r')
assert arr.shape == (n, DIM) and arr.dtype == np.float16, (arr.shape, arr.dtype)
hdr = int(arr.offset)
nmin, nmax, nonfin = 9.0, 0.0, 0
for s in range(0, n, 1 << 20):
blk = np.asarray(arr[s:s + (1 << 20)], dtype=np.float32)
nonfin += int((~np.isfinite(blk)).sum())
nr = np.linalg.norm(blk, axis=1)
nmin = min(nmin, float(nr.min()))
nmax = max(nmax, float(nr.max()))
idx = np.sort(rng.choice(n, size=min(256, n), replace=False))
chk = model.encode([texts[i] for i in idx], batch_size=BATCH, prompt=PROMPT, normalize_embeddings=True,
convert_to_numpy=True, show_progress_bar=False).astype(np.float32)
cos = (np.asarray(arr[idx], dtype=np.float32) * chk).sum(1)
ids_back = pq.read_table(ids).column('entity_id')
ids_ok = ids_back.equals(ids_arr) if hasattr(ids_back, 'equals') else None
del arr
fsize = os.path.getsize(npy)
rps = n / enc_s
var_n += n
var_t += enc_s
grand_n += n
grand_t += enc_s
log(f'TABLE DONE {variant}/{table}: n={n} encode {enc_s:.0f}s = {rps:.0f} rec/s | variant so far '
f'{var_n} rows {var_n / var_t:.0f} rec/s | ALL so far {grand_n} rows {grand_n / grand_t:.0f} rec/s | '
f'norm min {nmin:.4f} max {nmax:.4f} nonfinite {nonfin} | order check cos min {cos.min():.5f} '
f'mean {cos.mean():.5f} | ids equal {ids_ok} | file {fsize} bytes (expected {128 + n * DIM * 2})')
if cos.min() < 0.99 or nonfin > 0 or not ids_ok or fsize != hdr + n * DIM * 2:
log(f'FATAL integrity failure on {variant}/{table}; aborting')
sys.exit(3)
with mlock:
manifest['tables'][f'{variant}/{table}'] = dict(
rows=n, max_seq_length=max_len, encode_seconds=round(enc_s, 1), rec_per_s=round(rps, 1),
npy_bytes=fsize, norm_min=nmin, norm_max=nmax, nonfinite=nonfin,
order_check_cos_min=float(cos.min()), empty_names=e_name, empty_addresses=e_addr)
mb = json.dumps(manifest, indent=1).encode()
# surface any earlier upload failure before queueing more work
for f in futures:
if f.done() and f.exception() is not None:
raise f.exception()
futures.append(pool.submit(upload, variant, table, npy, ids, mb))
del texts, ids_arr
log(f'===== variant {variant} encoded: {var_n} rows in {var_t:.0f}s = {var_n / max(var_t, 1e-9):.0f} rec/s')
log('all encoding done; waiting for uploads')
for f in futures:
f.result()
pool.shutdown()
log(f'ALL DONE: {grand_n} row-encodings in {grand_t:.0f}s encode time ({grand_n / max(grand_t, 1e-9):.0f} rec/s), '
f'wall {time.time() - T0:.0f}s')
log(f'GPU {torch.cuda.get_device_name(0)} max_memory_allocated '
f'{torch.cuda.max_memory_allocated() / 2**30:.2f} GiB ({torch.cuda.max_memory_allocated()} bytes)')
print(json.dumps(manifest, indent=1), flush=True)
if __name__ == '__main__':
main()
|