---
license: mit
language:
- en
- zh
base_model:
- deepseek-ai/DeepSeek-OCR
base_model_relation: quantized
pipeline_tag: image-text-to-text
library_name: transformers
---
# Note: I made a mistake with the initial upload, and some tensors were initially not compressed, so I temporarily made the model private while I regenerated and uploaded the fixed version. If you downloaded the earlier version of the weights, please redownload them. The fixed version of the weights is ~200 MB smaller than before. 
## Update: Found and fixed another minor error, which saves a further ~0.16 MB in disk space.

For more information (including how to compress models yourself), check out https://huggingface.co/DFloat11 and https://github.com/LeanModels/DFloat11

Feel free to request for other models for compression as well (for either the `diffusers` library, ComfyUI, or any other model), although models that use architectures which are unfamiliar to me might be more difficult.

### How to Use

#### `transformers`

1. Install the DFloat11 pip package *(installs the CUDA kernel automatically; requires a CUDA-compatible GPU and PyTorch installed)*:

    ```bash
    pip install dfloat11[cuda12]
    # or if you have CUDA version 11:
    # pip install dfloat11[cuda11]
    pip install transformers==4.46.3
    
2. To use the DFloat11 model, run the following example code in Python:
    ```python
    from transformers import AutoModel, AutoTokenizer
    from dfloat11 import DFloat11Model
    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().to(torch.bfloat16) # Casting the model to BF16 before calling `.eval()` appears to cause slightly different bounding box results
    DFloat11Model.from_pretrained("mingyi456/DeepSeek-OCR-DF11", device = "cpu", bfloat16_model = model)
    model = model.cuda()
    
    # 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)

    ```

### Compression Details

This is the `pattern_dict` for compression:

```python
pattern_dict={
	r"lm_head": [],
	r"model\.embed_tokens": [],
	r"model\.layers\.0": [
		"self_attn.q_proj",
		"self_attn.k_proj",
		"self_attn.v_proj",
		"self_attn.o_proj",
		"mlp.gate_proj",
		"mlp.up_proj",
		"mlp.down_proj"
	],
	
	r"model\.layers\.[1-9]\d*": [
		"self_attn.q_proj",
		"self_attn.k_proj",
		"self_attn.v_proj",
		"self_attn.o_proj",

		"mlp.experts.0.gate_proj",
		"mlp.experts.0.up_proj",
		"mlp.experts.0.down_proj",
		"mlp.experts.1.gate_proj",
		"mlp.experts.1.up_proj",
		"mlp.experts.1.down_proj",
		"mlp.experts.2.gate_proj",
		"mlp.experts.2.up_proj",
		"mlp.experts.2.down_proj",
		"mlp.experts.3.gate_proj",
		"mlp.experts.3.up_proj",
		"mlp.experts.3.down_proj",
		"mlp.experts.4.gate_proj",
		"mlp.experts.4.up_proj",
		"mlp.experts.4.down_proj",
		"mlp.experts.5.gate_proj",
		"mlp.experts.5.up_proj",
		"mlp.experts.5.down_proj",
		"mlp.experts.6.gate_proj",
		"mlp.experts.6.up_proj",
		"mlp.experts.6.down_proj",
		"mlp.experts.7.gate_proj",
		"mlp.experts.7.up_proj",
		"mlp.experts.7.down_proj",
		"mlp.experts.8.gate_proj",
		"mlp.experts.8.up_proj",
		"mlp.experts.8.down_proj",
		"mlp.experts.9.gate_proj",
		"mlp.experts.9.up_proj",
		"mlp.experts.9.down_proj",
		"mlp.experts.10.gate_proj",
		"mlp.experts.10.up_proj",
		"mlp.experts.10.down_proj",
		"mlp.experts.11.gate_proj",
		"mlp.experts.11.up_proj",
		"mlp.experts.11.down_proj",
		"mlp.experts.12.gate_proj",
		"mlp.experts.12.up_proj",
		"mlp.experts.12.down_proj",
		"mlp.experts.13.gate_proj",
		"mlp.experts.13.up_proj",
		"mlp.experts.13.down_proj",
		"mlp.experts.14.gate_proj",
		"mlp.experts.14.up_proj",
		"mlp.experts.14.down_proj",
		"mlp.experts.15.gate_proj",
		"mlp.experts.15.up_proj",
		"mlp.experts.15.down_proj",
		"mlp.experts.16.gate_proj",
		"mlp.experts.16.up_proj",
		"mlp.experts.16.down_proj",
		"mlp.experts.17.gate_proj",
		"mlp.experts.17.up_proj",
		"mlp.experts.17.down_proj",
		"mlp.experts.18.gate_proj",
		"mlp.experts.18.up_proj",
		"mlp.experts.18.down_proj",
		"mlp.experts.19.gate_proj",
		"mlp.experts.19.up_proj",
		"mlp.experts.19.down_proj",
		"mlp.experts.20.gate_proj",
		"mlp.experts.20.up_proj",
		"mlp.experts.20.down_proj",
		"mlp.experts.21.gate_proj",
		"mlp.experts.21.up_proj",
		"mlp.experts.21.down_proj",
		"mlp.experts.22.gate_proj",
		"mlp.experts.22.up_proj",
		"mlp.experts.22.down_proj",
		"mlp.experts.23.gate_proj",
		"mlp.experts.23.up_proj",
		"mlp.experts.23.down_proj",
		"mlp.experts.24.gate_proj",
		"mlp.experts.24.up_proj",
		"mlp.experts.24.down_proj",
		"mlp.experts.25.gate_proj",
		"mlp.experts.25.up_proj",
		"mlp.experts.25.down_proj",
		"mlp.experts.26.gate_proj",
		"mlp.experts.26.up_proj",
		"mlp.experts.26.down_proj",
		"mlp.experts.27.gate_proj",
		"mlp.experts.27.up_proj",
		"mlp.experts.27.down_proj",
		"mlp.experts.28.gate_proj",
		"mlp.experts.28.up_proj",
		"mlp.experts.28.down_proj",
		"mlp.experts.29.gate_proj",
		"mlp.experts.29.up_proj",
		"mlp.experts.29.down_proj",
		"mlp.experts.30.gate_proj",
		"mlp.experts.30.up_proj",
		"mlp.experts.30.down_proj",
		"mlp.experts.31.gate_proj",
		"mlp.experts.31.up_proj",
		"mlp.experts.31.down_proj",
		"mlp.experts.32.gate_proj",
		"mlp.experts.32.up_proj",
		"mlp.experts.32.down_proj",
		"mlp.experts.33.gate_proj",
		"mlp.experts.33.up_proj",
		"mlp.experts.33.down_proj",
		"mlp.experts.34.gate_proj",
		"mlp.experts.34.up_proj",
		"mlp.experts.34.down_proj",
		"mlp.experts.35.gate_proj",
		"mlp.experts.35.up_proj",
		"mlp.experts.35.down_proj",
		"mlp.experts.36.gate_proj",
		"mlp.experts.36.up_proj",
		"mlp.experts.36.down_proj",
		"mlp.experts.37.gate_proj",
		"mlp.experts.37.up_proj",
		"mlp.experts.37.down_proj",
		"mlp.experts.38.gate_proj",
		"mlp.experts.38.up_proj",
		"mlp.experts.38.down_proj",
		"mlp.experts.39.gate_proj",
		"mlp.experts.39.up_proj",
		"mlp.experts.39.down_proj",
		"mlp.experts.40.gate_proj",
		"mlp.experts.40.up_proj",
		"mlp.experts.40.down_proj",
		"mlp.experts.41.gate_proj",
		"mlp.experts.41.up_proj",
		"mlp.experts.41.down_proj",
		"mlp.experts.42.gate_proj",
		"mlp.experts.42.up_proj",
		"mlp.experts.42.down_proj",
		"mlp.experts.43.gate_proj",
		"mlp.experts.43.up_proj",
		"mlp.experts.43.down_proj",
		"mlp.experts.44.gate_proj",
		"mlp.experts.44.up_proj",
		"mlp.experts.44.down_proj",
		"mlp.experts.45.gate_proj",
		"mlp.experts.45.up_proj",
		"mlp.experts.45.down_proj",
		"mlp.experts.46.gate_proj",
		"mlp.experts.46.up_proj",
		"mlp.experts.46.down_proj",
		"mlp.experts.47.gate_proj",
		"mlp.experts.47.up_proj",
		"mlp.experts.47.down_proj",
		"mlp.experts.48.gate_proj",
		"mlp.experts.48.up_proj",
		"mlp.experts.48.down_proj",
		"mlp.experts.49.gate_proj",
		"mlp.experts.49.up_proj",
		"mlp.experts.49.down_proj",
		"mlp.experts.50.gate_proj",
		"mlp.experts.50.up_proj",
		"mlp.experts.50.down_proj",
		"mlp.experts.51.gate_proj",
		"mlp.experts.51.up_proj",
		"mlp.experts.51.down_proj",
		"mlp.experts.52.gate_proj",
		"mlp.experts.52.up_proj",
		"mlp.experts.52.down_proj",
		"mlp.experts.53.gate_proj",
		"mlp.experts.53.up_proj",
		"mlp.experts.53.down_proj",
		"mlp.experts.54.gate_proj",
		"mlp.experts.54.up_proj",
		"mlp.experts.54.down_proj",
		"mlp.experts.55.gate_proj",
		"mlp.experts.55.up_proj",
		"mlp.experts.55.down_proj",
		"mlp.experts.56.gate_proj",
		"mlp.experts.56.up_proj",
		"mlp.experts.56.down_proj",
		"mlp.experts.57.gate_proj",
		"mlp.experts.57.up_proj",
		"mlp.experts.57.down_proj",
		"mlp.experts.58.gate_proj",
		"mlp.experts.58.up_proj",
		"mlp.experts.58.down_proj",
		"mlp.experts.59.gate_proj",
		"mlp.experts.59.up_proj",
		"mlp.experts.59.down_proj",
		"mlp.experts.60.gate_proj",
		"mlp.experts.60.up_proj",
		"mlp.experts.60.down_proj",
		"mlp.experts.61.gate_proj",
		"mlp.experts.61.up_proj",
		"mlp.experts.61.down_proj",
		"mlp.experts.62.gate_proj",
		"mlp.experts.62.up_proj",
		"mlp.experts.62.down_proj",
		"mlp.experts.63.gate_proj",
		"mlp.experts.63.up_proj",
		"mlp.experts.63.down_proj",

		"mlp.shared_experts.gate_proj",
		"mlp.shared_experts.up_proj",
		"mlp.shared_experts.down_proj",
	],
	
	r"model\.sam_model\.blocks\.\d+": (
		"attn.qkv",
		"attn.proj",
		"mlp.lin1",
		"mlp.lin2",
	),
	
	r"model\.vision_model\.embeddings\.position_embedding": [],
	
	r"model\.vision_model\.transformer\.layers\.\d+": (
		"self_attn.qkv_proj",
		"self_attn.out_proj",
		"mlp.fc1",
		"mlp.fc2",
	),
	
	r"model\.projector\.layers": []
}
```