from unsloth import FastVisionModel
from transformers import AutoModel
import torch

model, tokenizer = FastVisionModel.from_pretrained(
    "Tcyber/deepseek_ocr_ethiopian_legal_FULL",
    load_in_4bit=True,               # ~4GB weights, fits 16GB T4
    dtype=torch.float16,             # T4 has no bf16 kernels
    auto_model=AutoModel,
    trust_remote_code=True,
    unsloth_force_compile=False,     # avoid RAM spike
    use_gradient_checkpointing=False,
)
FastVisionModel.for_inference(model)

res = model.infer(tokenizer, prompt="<image>\nFree OCR. ", image_file="page.png",
                  output_path="out/", image_size=768, base_size=1024,
                  crop_mode=False, save_results=False)
print(res)