| """Offline load/forward/generation/round-trip checks of the exported checkpoint.""" |
| import hashlib |
| import json |
| from pathlib import Path |
| import torch |
| from safetensors import safe_open |
| from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed |
|
|
| root = Path(__file__).resolve().parent |
| path = root/'model' |
| torch.set_num_threads(2) |
| tok = AutoTokenizer.from_pretrained(path, local_files_only=True, trust_remote_code=False) |
| model = AutoModelForCausalLM.from_pretrained(path, local_files_only=True, |
| trust_remote_code=False, torch_dtype=torch.float32).eval() |
| count = sum(p.numel() for p in model.parameters()) |
| assert count == 49999872 |
| assert len(tok) == 23258 |
| with safe_open(str(path/'model.safetensors'), framework='pt') as f: |
| tensor_count = len(f.keys()) |
| assert all(f.get_tensor(k).dtype == torch.float16 for k in f.keys()) |
| assert all(torch.isfinite(f.get_tensor(k)).all() for k in f.keys()) |
| samples = [] |
| for text in ['Однажды вечером', 'Once upon a time']: |
| assert tok.decode(tok.encode(text)) == text |
| set_seed(42) |
| inputs = tok(text, return_tensors='pt') |
| with torch.inference_mode(): |
| logits = model(**inputs).logits |
| assert torch.isfinite(logits).all() |
| generated = model.generate(**inputs, max_new_tokens=48, do_sample=True, |
| temperature=.8, top_p=.9, repetition_penalty=1.1, |
| pad_token_id=tok.pad_token_id, eos_token_id=tok.eos_token_id) |
| samples.append({'prompt':text,'output':tok.decode(generated[0],skip_special_tokens=True)}) |
| report = {'offline_transformers_load':'passed','trust_remote_code':False, |
| 'parameters':count,'vocabulary_size':len(tok),'tensor_count':tensor_count, |
| 'weights_dtype':'float16','finite_weights_and_logits':True, |
| 'russian_english_tokenizer_roundtrip':'passed', |
| 'generation_test':'execution only, NOT linguistic-quality certification', |
| 'generation_seed':42,'max_new_tokens':48,'temperature':.8,'top_p':.9, |
| 'repetition_penalty':1.1,'uncurated_samples':samples, |
| 'safetensors_bytes':(path/'model.safetensors').stat().st_size, |
| 'safetensors_sha256':hashlib.sha256((path/'model.safetensors').read_bytes()).hexdigest()} |
| (path/'verification.json').write_text(json.dumps(report,ensure_ascii=False,indent=2),encoding='utf-8') |
| print(json.dumps(report,ensure_ascii=False,indent=2)) |
|
|
