verify_adapter_files.py

"""Verify a downloaded LoRA adapter directory without loading torch.

Parses the safetensors header directly, so this runs on the laptop, in CI, or on
a judge's machine with nothing installed but Python. Confirms the file is intact,
reports its SHA-256, and counts the LoRA tensors and parameters.

    python scripts/verify_adapter_files.py models/adapters/<run>/checkpoint-1590
"""
from __future__ import annotations

import argparse
import hashlib
import json
import struct
import sys
from pathlib import Path

EXPECTED_SHA256_PREFIX = "3b625d63ed4f28640a6fb14072486305"
EXPECTED_TENSORS = 696
EXPECTED_PARAMS = 37_152_768


def read_header(path: Path) -> dict:
    """safetensors layout: 8-byte little-endian header length, then JSON."""
    with open(path, "rb") as fh:
        (n,) = struct.unpack("<Q", fh.read(8))
        return json.loads(fh.read(n))


def main() -> int:
    ap = argparse.ArgumentParser(description="Verify a LoRA adapter directory")
    ap.add_argument("adapter_dir", help="Directory holding adapter_model.safetensors")
    ap.add_argument("--expect-sha256", default=EXPECTED_SHA256_PREFIX,
                    help="Expected SHA-256 prefix ('' to skip the check)")
    args = ap.parse_args()

    d = Path(args.adapter_dir)
    weights = d / "adapter_model.safetensors"
    config = d / "adapter_config.json"
    if not weights.exists():
        print(f"FAIL: {weights} not found")
        return 1

    header = read_header(weights)
    tensors = {k: v for k, v in header.items() if k != "__metadata__"}
    params = 0
    dtypes = set()
    for meta in tensors.values():
        n = 1
        for dim in meta["shape"]:
            n *= dim
        params += n
        dtypes.add(meta["dtype"])

    digest = hashlib.sha256(weights.read_bytes()).hexdigest()

    print(f"adapter dir : {d}")
    print(f"weights     : {weights.name}  {weights.stat().st_size:,} bytes")
    print(f"sha256      : {digest}")
    print(f"tensors     : {len(tensors):,}")
    print(f"parameters  : {params:,}")
    print(f"dtypes      : {sorted(dtypes)}")

    if config.exists():
        cfg = json.loads(config.read_text(encoding="utf-8"))
        print(f"base model  : {cfg.get('base_model_name_or_path')}")
        print(f"lora        : r={cfg.get('r')} alpha={cfg.get('lora_alpha')} "
              f"dropout={cfg.get('lora_dropout')}")
        print(f"targets     : {', '.join(sorted(cfg.get('target_modules') or []))}")
    else:
        print("WARN: adapter_config.json missing -- PEFT cannot load this directory")

    ok = True
    if args.expect_sha256 and not digest.startswith(args.expect_sha256):
        print(f"FAIL: sha256 does not start with {args.expect_sha256}")
        ok = False
    if len(tensors) != EXPECTED_TENSORS:
        print(f"WARN: expected {EXPECTED_TENSORS} tensors, found {len(tensors)}")
    if params != EXPECTED_PARAMS:
        print(f"WARN: expected {EXPECTED_PARAMS:,} parameters, found {params:,}")

    print("OK" if ok else "MISMATCH")
    return 0 if ok else 1


if __name__ == "__main__":
    sys.exit(main())