Mac M1 support?

[theidexisted]

I have tried the demo and it looks great!

Could you please kindly add mac M1 support?

[rcshou]

I second that. Would be great to have Mac Silicon support
thank you for the model!

[songtao-git]

I second it

[redmoe-moutain]

Thank you for your interest, but unfortunately we don't have enough bandwidth for this in the short term.
We welcome and look forward to contributions from the community to support this.
Thanks again!

[ygfrancois]

@theidexisted @rcshou @songtao-git
You can try to run it with CPU:

1. Change config: attn_implementation from "flash_attention_2" to "sdpa" or "eager"(sdpa recommended)

2. hf demo like(note that this is very slow, try a smaller image with not too much text on it):

import os
if "LOCAL_RANK" not in os.environ:
    os.environ["LOCAL_RANK"] = "0"

import torch
from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer
from qwen_vl_utils import process_vision_info
from dots_ocr.utils import dict_promptmode_to_prompt

def inference(image_path, prompt, model, processor):
    # image_path = "demo/demo_image1.jpg"
    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": image_path
                },
                {"type": "text", "text": prompt}
            ]
        }
    ]


    # Preparation for inference
    text = processor.apply_chat_template(
        messages, 
        tokenize=False, 
        add_generation_prompt=True
    )
    image_inputs, video_inputs = process_vision_info(messages)
    inputs = processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=True,
        return_tensors="pt",
    )

    # inputs = inputs.to("cuda")

    # Inference: Generation of the output
    generated_ids = model.generate(**inputs, max_new_tokens=24000)
    generated_ids_trimmed = [
        out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
    ]
    output_text = processor.batch_decode(
        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
    )
    print(output_text)



if __name__ == "__main__":
    # We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
    # model_path = "./weights/DotsOCR"
    model_path = "./weights/DotsOCR_cpu"
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        # attn_implementation="flash_attention_2",
        torch_dtype=torch.bfloat16,
        # device_map="auto",
        device_map="cpu", 
        trust_remote_code=True
    )
    processor = AutoProcessor.from_pretrained(model_path,  trust_remote_code=True)

    # image_path = "demo/demo_image1.jpg"
    image_path = "assets/showcase/tradition_zh.png"
    for prompt_mode, prompt in dict_promptmode_to_prompt.items():
        print(f"prompt: {prompt}")
        inference(image_path, prompt, model, processor)
    

[Radar-Lei]

+1

[NorthSeven-02]

Need Mac Silicon support + 1 pls

[pllopis]

In case it helps, here's a wrapper that does lots of hacky patching to get this to run on macOS: https://gist.github.com/pllopis/5faf2ecc66ae5d87e3460bf7950511a4

Don't forget to export PYTORCH_ENABLE_MPS_FALLBACK=1

Here's the corresponding pip freeze:

accelerate==1.9.0
aiofiles==24.1.0
annotated-types==0.7.0
anyio==4.10.0
audioop-lts==0.2.2
av==15.0.0
Brotli==1.1.0
certifi==2025.8.3
charset-normalizer==3.4.2
click==8.2.1
distro==1.9.0
-e git+https://github.com/rednote-hilab/dots.ocr@6a5e2ba95c4cbcffcbd11cad2eb4e42c4e23f952#egg=dots_ocr
fastapi==0.116.1
ffmpy==0.6.1
filelock==3.18.0
fsspec==2025.7.0
gradio==5.41.0
gradio_client==1.11.0
gradio_image_annotation==0.4.0
groovy==0.1.2
h11==0.16.0
hf-xet==1.1.7
httpcore==1.0.9
httpx==0.28.1
huggingface-hub==0.34.3
idna==3.10
Jinja2==3.1.6
jiter==0.10.0
markdown-it-py==3.0.0
MarkupSafe==3.0.2
mdurl==0.1.2
modelscope==1.28.2
mpmath==1.3.0
networkx==3.5
numpy==2.3.2
openai==1.99.1
orjson==3.11.1
packaging==25.0
pandas==2.3.1
pillow==11.3.0
psutil==7.0.0
pydantic==2.11.7
pydantic_core==2.33.2
pydub==0.25.1
Pygments==2.19.2
PyMuPDF==1.26.3
python-dateutil==2.9.0.post0
python-multipart==0.0.20
pytz==2025.2
PyYAML==6.0.2
qwen-vl-utils==0.0.11
regex==2025.7.34
requests==2.32.4
rich==14.1.0
ruff==0.12.7
safehttpx==0.1.6
safetensors==0.6.1
semantic-version==2.10.0
setuptools==80.9.0
shellingham==1.5.4
six==1.17.0
sniffio==1.3.1
starlette==0.47.2
sympy==1.14.0
tokenizers==0.21.4
tomlkit==0.13.3
torch==2.8.0
torchaudio==2.8.0
torchvision==0.23.0
tqdm==4.67.1
transformers==4.51.3
typer==0.16.0
typing-inspection==0.4.1
typing_extensions==4.14.1
tzdata==2025.2
urllib3==2.5.0
uvicorn==0.35.0
websockets==15.0.1

[theidexisted]

@pllopis

Thanks for such great work! It's a good start, I have tried and found:

• We need python 3.13
• The json parse logic need to be refined

[nissansz]

How to install and run on windows for cpu version?

[pllopis]

In case it helps, here's a wrapper that does lots of hacky patching to get this to run on macOS: https://gist.github.com/pllopis/5faf2ecc66ae5d87e3460bf7950511a4

Just to comment that I added all of this into a stand-alone repo here: https://github.com/pllopis/pdf2md/tree/main but it remains pretty much work-in-progress, in fact while I ported the M1/CPU support there I only tested that on CUDA for the moment, but should be little work to get the Apple Silicon to work, given that the code is mostly there (if it doesn't already work out of the box).