"""Repeatable, same-pipeline CPU comparison of the published student and teacher.""" import hashlib, json, os, platform, statistics, sys, time from pathlib import Path import numpy as np import torch, transformers, numba HERE = Path(__file__).resolve().parent ROOT = HERE.parent sys.path[:0] = [str(ROOT / 'upstream'), str(ROOT / 'student8n256')] import eval.span_f1 as upstream from fast_decode import decode_fast from train.labels import load_labels torch.set_num_threads(8) torch.set_num_interop_threads(1) upstream.decode = lambda lp, T, s, e: decode_fast(np.ascontiguousarray(lp), s, e).tolist() teacher = upstream.Predictor(str(ROOT / 'model'), device='cpu', max_len=128, stride=16) student = upstream.Predictor(str(HERE / 'hf_release'), device='cpu', max_len=128, stride=16) _, tags, _, _ = load_labels() assert teacher.tags[:13] == tags[:13] and student.model.config.num_labels == 13 student.tags = student.tags[:13] student.T = student.T[:13, :13] student.s0 = np.ascontiguousarray(student.s0[:13]) student.e0 = np.ascontiguousarray(student.e0[:13]) sample = '안녕하세요. 고객 김철수의 연락처는 010-1234-5678이고 이메일은 minsu.kim@example.com입니다. 서울시 강남구 테헤란로 123으로 서류를 보내 주세요. 상담 내용을 확인한 뒤 담당자가 다시 연락드리겠습니다.\n' def document(n): return (sample * (n // len(sample) + 1))[:n] def run(p, text): spans = p.predict(text, bs=16) spans = [s for s in spans if s['label'] in ('PERSON', 'ADDRESS', 'ORGANIZATION')] return json.dumps({'spans': spans}, ensure_ascii=False).encode('utf8') for p in (teacher, student): for _ in range(3): run(p, document(1000)) out = {'method': 'same upstream Predictor, PyTorch FP32, filtering to 3 labels after decoding; includes UTF-8 JSON serialization; model load excluded', 'platform': platform.platform(), 'cpu': platform.processor(), 'affinity': sorted(os.sched_getaffinity(0)), 'torch': torch.__version__, 'transformers': transformers.__version__, 'numpy': np.__version__, 'numba': numba.__version__, 'threads': 8, 'interop_threads': 1, 'window': 128, 'overlap': 16, 'batch': 16, 'skip_ids': False, 'warmup_per_model': 3, 'repetitions': 3, 'rows': []} target = HERE / 'hf_release' / 'eval' / 'speed_same_cpu.json' for n in (100, 1000, 5000, 10000, 50000, 100000, 200000): text = document(n) row = {'chars': n, 'sha256_utf8': hashlib.sha256(text.encode()).hexdigest(), 'teacher': [], 'student': []} for i in range(3): for name, model in ((('teacher', teacher), ('student', student)) if i % 2 == 0 else (('student', student), ('teacher', teacher))): start = time.perf_counter(); payload = run(model, text); elapsed = time.perf_counter() - start row[name].append(elapsed) print(n, i + 1, name, round(elapsed, 6), flush=True) for name in ('teacher', 'student'): row[name + '_median'] = statistics.median(row[name]) out['rows'].append(row) target.write_text(json.dumps(out, ensure_ascii=False, indent=2), encoding='utf8')