ZoneTwelve/DeepSeek-OCR-AnyDevice

---

🌟 Github | 📥 Model Download | 📄 Paper Link | 📄 Arxiv Paper Link |

DeepSeek-OCR: Contexts Optical Compression

Explore the boundaries of visual-text compression.

Usage

Inference using Huggingface transformers on NVIDIA GPUs. Requirements tested on python 3.12.9 + CUDA11.8：

torch==2.6.0
transformers==4.46.3
tokenizers==0.20.3
einops
addict 
easydict
pip install flash-attn==2.7.3 --no-build-isolation

from transformers import AutoModel, AutoTokenizer
import torch
import os
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
model_name = 'deepseek-ai/DeepSeek-OCR'

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(model_name, _attn_implementation='flash_attention_2', trust_remote_code=True, use_safetensors=True)
model = model.eval().cuda().to(torch.bfloat16)

# prompt = "<image>\nFree OCR. "
prompt = "<image>\n<|grounding|>Convert the document to markdown. "
image_file = 'your_image.jpg'
output_path = 'your/output/dir'

# infer(self, tokenizer, prompt='', image_file='', output_path = ' ', base_size = 1024, image_size = 640, crop_mode = True, test_compress = False, save_results = False):

# Tiny: base_size = 512, image_size = 512, crop_mode = False
# Small: base_size = 640, image_size = 640, crop_mode = False
# Base: base_size = 1024, image_size = 1024, crop_mode = False
# Large: base_size = 1280, image_size = 1280, crop_mode = False

# Gundam: base_size = 1024, image_size = 640, crop_mode = True

res = model.infer(tokenizer, prompt=prompt, image_file=image_file, output_path = output_path, base_size = 1024, image_size = 640, crop_mode=True, save_results = True, test_compress = True)

vLLM

Refer to 🌟GitHub for guidance on model inference acceleration and PDF processing, etc.

[2025/10/23] 🚀🚀🚀 DeepSeek-OCR is now officially supported in upstream vLLM.

uv venv
source .venv/bin/activate
# Until v0.11.1 release, you need to install vLLM from nightly build
uv pip install -U vllm --pre --extra-index-url https://wheels.vllm.ai/nightly

from vllm import LLM, SamplingParams
from vllm.model_executor.models.deepseek_ocr import NGramPerReqLogitsProcessor
from PIL import Image

# Create model instance
llm = LLM(
    model="deepseek-ai/DeepSeek-OCR",
    enable_prefix_caching=False,
    mm_processor_cache_gb=0,
    logits_processors=[NGramPerReqLogitsProcessor]
)

# Prepare batched input with your image file
image_1 = Image.open("path/to/your/image_1.png").convert("RGB")
image_2 = Image.open("path/to/your/image_2.png").convert("RGB")
prompt = "<image>\nFree OCR."

model_input = [
    {
        "prompt": prompt,
        "multi_modal_data": {"image": image_1}
    },
    {
        "prompt": prompt,
        "multi_modal_data": {"image": image_2}
    }
]

sampling_param = SamplingParams(
            temperature=0.0,
            max_tokens=8192,
            # ngram logit processor args
            extra_args=dict(
                ngram_size=30,
                window_size=90,
                whitelist_token_ids={128821, 128822},  # whitelist: <td>, </td>
            ),
            skip_special_tokens=False,
        )
# Generate output
model_outputs = llm.generate(model_input, sampling_param)

# Print output
for output in model_outputs:
    print(output.outputs[0].text)

Visualizations

Acknowledgement

We would like to thank Vary, GOT-OCR2.0, MinerU, PaddleOCR, OneChart, Slow Perception for their valuable models and ideas.

We also appreciate the benchmarks: Fox, OminiDocBench.

---

🔍 Summary of Changes

1. Formatting & Clean-Up

• Removed extra spaces, blank lines, and inconsistent indentation.
• Fixed small style issues (like missing spaces in comments).
• Added missing newline at the end of the file.

---

2. Device and Dtype Handling

• Added automatic device detection:

  model_device = next(self.parameters()).device
  

• Added adaptive dtype logic:

  image_dtype = torch.bfloat16 if model_device.type == "cuda" else torch.float32
  

• Replaced all hardcoded .cuda() and .to(torch.bfloat16) with:

  .to(model_device)
  .to(image_dtype)
  

✅ Now works automatically on both GPU and CPU, without device mismatch errors.

---

3. Autocast and Inference Improvements

• Wrapped generation in a conditional autocast block:

  use_autocast = model_device.type == "cuda"
  if use_autocast:
      with torch.autocast("cuda", dtype=torch.bfloat16):
          with torch.no_grad():
              ...
  else:
      with torch.no_grad():
          ...
  

• Reduces memory usage and speeds up inference on GPU.

• Added torch.no_grad() for safer evaluation (no gradient tracking).

---

4. Image Preprocessing

• All image tensors now use:

  .to(image_dtype)
  

  instead of hardcoded torch.bfloat16.

• Improves flexibility and prevents dtype errors when running on CPU.

---

5. Generation Parameter Updates

• Adjusted text generation settings:

  do_sample = False
  num_beams = 1
  max_new_tokens = 4096  # was 8192
  min_new_tokens = 1
  repetition_penalty = 1.2
  pad_token_id = tokenizer.pad_token_id or tokenizer.eos_token_id
  

🧠 Results: Faster, more controlled generation; avoids repetitive or runaway outputs.

---

6. Safer Decoding

• Cleaned up decoding logic:

  input_length = input_ids.unsqueeze(0).to(model_device).shape[1]
  outputs = tokenizer.decode(output_ids[0, input_length:])
  

✅ Avoids CUDA-specific assumptions, consistent across devices.

---

7. Miscellaneous

• Added helpful comments for clarity.
• Improved readability around image transformation and saving results.
• Added extra blank lines for cleaner structure.

---

⚙️ Overall Impact

Category	Before	After
Device handling	Hardcoded .cuda()	Auto-detected and flexible
Dtype	Always bfloat16	Adaptive: bfloat16 (GPU) / float32 (CPU)
Inference	Could crash on CPU	Runs safely everywhere
Generation	Unbounded, repetitive	Tuned and stable
Readability	Mixed formatting	Clean and consistent

---

Citation

@article{wei2025deepseek,
  title={DeepSeek-OCR: Contexts Optical Compression},
  author={Wei, Haoran and Sun, Yaofeng and Li, Yukun},
  journal={arXiv preprint arXiv:2510.18234},
  year={2025}
}