test_real_docs.py

from transformers import AutoModel, AutoTokenizer
import torch
import os
from PIL import Image
import time

# Force CPU for stability
device = "cpu"
print(f"Using device: {device}")

# Patch to avoid CUDA calls in custom code
torch.Tensor.cuda = lambda self, *args, **kwargs: self.to(device)
torch.nn.Module.cuda = lambda self, *args, **kwargs: self.to(device)

model_name = 'deepseek-ai/DeepSeek-OCR-2'

def test_docs():
    print(f"Loading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

    print(f"Loading model (may take a minute)...")
    # Load with default parameters that worked in test_minimal.py
    model = AutoModel.from_pretrained(
        model_name, 
        trust_remote_code=True, 
        use_safetensors=True
    )
    model = model.eval()
    
    # After loading, we monkeypatch bfloat16 for the inference logic
    torch.bfloat16 = torch.float32

    image_dir = "doc_images"
    output_dir = "ocr_results"
    os.makedirs(output_dir, exist_ok=True)

    images = sorted([f for f in os.listdir(image_dir) if f.endswith(".png")])
    
    for img_name in images:
        img_path = os.path.join(image_dir, img_name)
        print(f"\n--- Processing: {img_name} ---")
        
        # DeepSeek-OCR-2 needs specific ratios for its hardcoded query embeddings
        # base_size=1024 -> n_query=256 (supported)
        # image_size=768 -> n_query=144 (supported)
        
        prompt = "<image>\nFree OCR. "
        
        start_time = time.time()
        try:
            with torch.no_grad():
                res = model.infer(
                    tokenizer, 
                    prompt=prompt, 
                    image_file=img_path, 
                    output_path=output_dir,
                    base_size=1024, # Must be 1024 for 256 queries
                    image_size=768,  # Must be 768 for 144 queries
                    crop_mode=False,
                    eval_mode=True
                )
            
            elapsed = time.time() - start_time
            print(f"Done in {elapsed:.2f}s")
            
            result_file = os.path.join(output_dir, f"{img_name}.md")
            with open(result_file, "w") as f:
                f.write(res)
            
            print(f"Result saved to {result_file}")
            print("Preview (first 500 chars):")
            print("-" * 20)
            print(res[:500] + "...")
            print("-" * 20)
            
        except Exception as e:
            print(f"Inference failed for {img_name}: {e}")
            import traceback
            traceback.print_exc()

if __name__ == "__main__":
    test_docs()