Download source/tools/quantization_cases.py from yunusemrejr/Franok: direct link, hf CLI and curl.
- Browser
- Download file 2.59 kB
-
https://huggingface.co/yunusemrejr/Franok/resolve/main/source/tools/quantization_cases.py
- Command line
-
hf download hf://yunusemrejr/Franok/source/tools/quantization_cases.py
-
curl -L -o quantization_cases.py https://huggingface.co/yunusemrejr/Franok/resolve/main/source/tools/quantization_cases.py
2.59 kB
| """Build bilingual validation prefixes and full-precision distributions for Q4/Q8 checks.""" | |
| import argparse | |
| import hashlib | |
| import json | |
| from pathlib import Path | |
| import numpy as np | |
| import torch | |
| from safetensors.torch import load_file | |
| from franok.model import Config, Franok | |
| from training.third.batches import Conversations | |
| def main(): | |
| p = argparse.ArgumentParser() | |
| p.add_argument("--run", type=Path, required=True) | |
| p.add_argument("--out", type=Path, required=True) | |
| p.add_argument("--per-language", type=int, default=48) | |
| p.add_argument("--threads", type=int, default=2) | |
| a = p.parse_args() | |
| torch.set_num_threads(a.threads) | |
| config = Config(**json.loads((a.run / "config.json").read_text())) | |
| model = Franok(config).eval() | |
| model.load_state_dict(load_file(str(a.run / "model.safetensors"))) | |
| data = Conversations("val") | |
| g = np.random.default_rng(97521) | |
| cases = [] | |
| seen = set() | |
| for lang in ("en", "tr"): | |
| for i in range(a.per_language): | |
| band = (64, 128, 256, 512)[i % 4] | |
| for _ in range(2000): | |
| pre, ans, _, name = data.sample(g, lang, band, False) | |
| if len(pre) > 210 or len(ans) < 3: | |
| continue | |
| digest = hashlib.sha256( | |
| np.asarray(pre, dtype="<u2").tobytes() | |
| ).hexdigest() | |
| if digest not in seen: | |
| seen.add(digest) | |
| break | |
| else: | |
| raise RuntimeError("Could not draw an independent prefix") | |
| ans = ans[:16] | |
| ids = pre.astype(int).tolist() + ans.astype(int).tolist() | |
| out = model(torch.tensor([ids[:-1]]))[0, len(pre) - 1 :].log_softmax(-1) | |
| v, k = out.topk(64) | |
| gold = out[torch.arange(len(ans)), torch.tensor(ans.astype(int))] | |
| cases.append( | |
| dict( | |
| lang=lang, | |
| source=name, | |
| group=("dev" if (i // 4) % 2 == 0 else "confirm"), | |
| ids=ids, | |
| at=len(pre), | |
| key=digest, | |
| gold=gold.tolist(), | |
| top=k.tolist(), | |
| probs=v.exp().tolist(), | |
| ) | |
| ) | |
| a.out.parent.mkdir(parents=True, exist_ok=True) | |
| a.out.write_text(json.dumps(cases, separators=(",", ":"))) | |
| print( | |
| len(cases), | |
| "prefixes", | |
| sum(len(r["gold"]) for r in cases), | |
| "reference positions", | |
| ) | |
| if __name__ == "__main__": | |
| main() | |