FinConRank / infer.py
yoohoo233's picture
Upload 35 files
cf814c5 verified
Raw History Blame Contribute Delete
2.21 kB
"""Offline pointwise FinConRank scorer; no generation and no fusion."""
import argparse
import json
from pathlib import Path
import torch
from peft import PeftModel
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from financial_constraints import extract_constraints
from scoring import score_batch
def load(base, adapter):
processor = AutoProcessor.from_pretrained(base, local_files_only=True)
processor.image_processor.max_pixels = 602112
model = Qwen2VLForConditionalGeneration.from_pretrained(
base, torch_dtype=torch.float16, local_files_only=True, attn_implementation='sdpa')
model = PeftModel.from_pretrained(model, adapter, local_files_only=True).to('cuda').eval()
tokens = [processor.tokenizer.convert_tokens_to_ids(t) for t in ('True', 'False')]
assert all(isinstance(t, int) for t in tokens) and tokens[0] != tokens[1]
return processor, model, tokens
def main():
p = argparse.ArgumentParser(description=__doc__)
p.add_argument('--base-model', required=True, help='Explicit existing local Qwen base snapshot')
p.add_argument('--adapter', required=True)
p.add_argument('--query', required=True)
p.add_argument('--image', nargs='+', required=True)
p.add_argument('--constraints-json', help='Optional explicit constraints JSON; otherwise deterministic query-only parser')
p.add_argument('--output', required=True)
args = p.parse_args()
constraints = json.loads(Path(args.constraints_json).read_text()) if args.constraints_json else extract_constraints(args.query)
processor, model, tokens = load(args.base_model, args.adapter)
torch.cuda.reset_peak_memory_stats()
# Batch size one is deliberate; original-vs-copy validation uses the same setting.
scores = [score_batch(processor, model, tokens, args.query, constraints, [Path(image)])[0] for image in args.image]
Path(args.output).write_text(json.dumps({'scores': scores, 'constraints': constraints, 'token_ids': tokens,
'peak_allocated_bytes': torch.cuda.max_memory_allocated(), 'max_pixels': 602112,
'dtype': 'float16', 'batch_size': 1}, ensure_ascii=False, indent=2) + '\n')
if __name__ == '__main__':
main()