SCOPD / evaluation /scripts /merge_adapter.py
enmingzhangzz's picture
Publish evaluation-and-model-card
a47010e verified
Raw History Blame Contribute Delete
3.3 kB
"""Merge a released adapter locally for evaluation; no training is performed."""
from __future__ import annotations
import argparse
import hashlib
import json
from pathlib import Path
import sys
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--adapter', type=Path, required=True)
parser.add_argument('--base-model', help='Override with the matching pinned local Hub snapshot')
parser.add_argument('--base-revision', help='Defaults to the adapter provenance revision')
parser.add_argument('--output', type=Path, required=True)
parser.add_argument('--device', choices=['cpu', 'cuda'], default='cpu')
args = parser.parse_args()
if args.output.exists() and any(args.output.iterdir()):
raise FileExistsError('Refusing to overwrite an existing model')
provenance = json.loads((args.adapter / 'provenance.json').read_text())
args.base_model = args.base_model or provenance['base_model']
args.base_revision = args.base_revision or provenance['base_revision']
digest = hashlib.sha256((args.adapter / 'adapter_model.safetensors').read_bytes()).hexdigest()
if digest != provenance['adapter_sha256']:
raise ValueError('Adapter SHA256 mismatch')
if args.base_revision != provenance['base_revision']:
raise ValueError('Base revision differs from adapter provenance')
qwen3 = provenance['base_model'] == 'Qwen/Qwen3-VL-4B-Instruct'
if qwen3:
from scopd_eval.qwen3_runtime import activate
else:
from scopd_eval.runtime import activate
activate()
from huggingface_hub import snapshot_download
import torch
from peft import PeftModel
from transformers import AutoProcessor
if qwen3:
from transformers import Qwen3VLForConditionalGeneration as Model
else:
from transformers import Qwen2_5_VLForConditionalGeneration as Model
base = Path(args.base_model)
if not base.is_dir():
if args.base_model != provenance['base_model']:
raise ValueError('Base model differs from adapter provenance')
base = Path(snapshot_download(args.base_model, revision=args.base_revision,
allow_patterns=['*.json', '*.safetensors', '*.txt', '*.jinja']))
elif base.name != args.base_revision:
raise ValueError('Local base must be the pinned Hub snapshot directory named by its revision')
model = Model.from_pretrained(
base, torch_dtype=torch.bfloat16, attn_implementation='eager',
device_map={'': args.device}, local_files_only=True)
model = PeftModel.from_pretrained(model, args.adapter, is_trainable=False)
model = model.merge_and_unload(safe_merge=True)
model.save_pretrained(args.output, safe_serialization=True, max_shard_size='5GB')
AutoProcessor.from_pretrained(base, local_files_only=True).save_pretrained(args.output)
(args.output / 'merge_manifest.json').write_text(json.dumps({
'adapter_sha256': digest, 'base_model': provenance['base_model'],
'base_revision': args.base_revision, 'safe_merge': True,
'dtype': 'bfloat16'}, indent=2) + '\n')
print(f'Merged model ready: {args.output}')
if __name__ == '__main__':
main()