"""Thesis-B recognizer (Perfilev, HSE 2026; released QLoRA adapters AntoniusPerf/runic-ocr-qwen-vl-lora) on our images. Reproduces the inference of RUNIC-OCR/src/runic_ocr_experiments.py::_infer_qwen: system + user prompt, processor from the adapter folder with max_pixels, 4-bit NF4 base (fp16 compute) when bitsandbytes is available, greedy decoding. python perfilev_infer.py --base --adapter --max_pixels 147456 --input x.jsonl --out y.jsonl """ import argparse import json import os import torch from PIL import Image SYS_PROMPT = ("You are an expert runologist OCR system. You read runic inscriptions " "and output ONLY their transliteration in the Rundata convention, " "with no explanation.") USR_PROMPT = "Transliterate the runic inscription in this image." def main(): ap = argparse.ArgumentParser() ap.add_argument("--base", required=True) ap.add_argument("--adapter", required=True) ap.add_argument("--max_pixels", type=int, default=384 * 384) ap.add_argument("--input", required=True) ap.add_argument("--image_root", default="") ap.add_argument("--out", required=True) ap.add_argument("--max_new_tokens", type=int, default=64) ap.add_argument("--limit", type=int, default=0) args = ap.parse_args() print("ARGS", vars(args), flush=True) rows = [json.loads(l) for l in open(args.input) if l.strip()] if args.limit: rows = rows[: args.limit] done = {json.loads(l)["id"] for l in open(args.out)} if os.path.exists(args.out) else set() todo = [r for r in rows if r["id"] not in done] print(f"rows {len(rows)} todo {len(todo)}", flush=True) if not todo: return from peft import PeftModel from transformers import AutoModelForImageTextToText, AutoProcessor proc = AutoProcessor.from_pretrained(args.adapter, max_pixels=args.max_pixels) try: from transformers import BitsAndBytesConfig import bitsandbytes # noqa: F401 q = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.float16) print("base in 4-bit NF4", flush=True) except Exception as e: # noqa: BLE001 q = None print("bitsandbytes unavailable, base in fp16:", e, flush=True) base = AutoModelForImageTextToText.from_pretrained(args.base, quantization_config=q, dtype=torch.float16, device_map="cuda") model = PeftModel.from_pretrained(base, args.adapter).eval() fout = open(args.out, "a") for k, r in enumerate(todo): img = Image.open(os.path.join(args.image_root, r["image"])).convert("RGB") msgs = [{"role": "system", "content": SYS_PROMPT}, {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": USR_PROMPT}]}] text = proc.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True) inp = proc(text=[text], images=[img], return_tensors="pt").to("cuda") with torch.no_grad(): out = model.generate(**inp, max_new_tokens=args.max_new_tokens, do_sample=False) dec = proc.batch_decode(out[:, inp["input_ids"].shape[1]:], skip_special_tokens=True)[0].strip() fout.write(json.dumps(dict(r, pred=dec), ensure_ascii=False) + "\n") fout.flush() if k % 50 == 0: print(f"{k}/{len(todo)} {dec[:60]!r}", flush=True) print("DONE", flush=True) if __name__ == "__main__": main()