--- license: mit language: - en pipeline_tag: image-text-to-text library_name: transformers tags: - text-generation-inference - OCR - VLM new_version: prithivMLmods/DeepSeek-OCR-Latest-BF16 --- > [!warning] This version is experimental. Please refer to the newer versions pinned above to avoid any complexities.šŸ‘†šŸ‘†šŸ‘† > [!IMPORTANT] > This is a copy of the model weights from the [https://huggingface.co/deepseek-ai/DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR) model. These weights cannot be used for other purposes. If you wish to do so, please visit the original model page. Previously, inference with the model [https://huggingface.co/deepseek-ai/DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR) ran smoothly on `transformers==4.46.3`. However, running it on the latest versions of `transformers` caused compatibility issues with `LlamaAttention`. We have addressed and fixed the problem, and the model now runs smoothly with the latest `transformers (v4.57.1)` or any compatible version. This page, which includes the model weights and corrected configuration, fixed the issue and allowed `transformers` inference to run smoothly. > [!note] Last updated: 3:00 PM (IST), DEC 04, 2025. ``` transformers: 4.57.1 torch: 2.6.0+cu124 (or) the latest version (i.e., torch 2.9.0) cuda: 12.4 device: NVIDIA H200 MIG 3g.71gb ``` ``` Currently supported up to `transformers==4.57.2`. Support for Transformers v5 will be added soon. ``` ## Quick Start with Transformers > [!note] Try it now on Google Colab: https://huggingface.co/strangervisionhf/deepseek-ocr-latest-transformers/blob/main/deepseek_ocr_demo.ipynb #### Install the required packages ``` gradio torch torchvision transformers==4.57.1 accelerate matplotlib einops addict easydict ``` ### Run [app.py] ```py import gradio as gr import torch import requests from transformers import AutoModel, AutoTokenizer import spaces from typing import Iterable import os import tempfile from PIL import Image, ImageDraw import re from gradio.themes import Soft from gradio.themes.utils import colors, fonts, sizes from docling_core.types.doc import DoclingDocument, DocTagsDocument device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print("CUDA_VISIBLE_DEVICES=", os.environ.get("CUDA_VISIBLE_DEVICES")) print("torch.__version__ =", torch.__version__) print("torch.version.cuda =", torch.version.cuda) print("cuda available:", torch.cuda.is_available()) print("cuda device count:", torch.cuda.device_count()) if torch.cuda.is_available(): print("current device:", torch.cuda.current_device()) print("device name:", torch.cuda.get_device_name(torch.cuda.current_device())) print("Using device:", device) colors.steel_blue = colors.Color( name="steel_blue", c50="#EBF3F8", c100="#D3E5F0", c200="#A8CCE1", c300="#7DB3D2", c400="#529AC3", c500="#4682B4", c600="#3E72A0", c700="#36638C", c800="#2E5378", c900="#264364", c950="#1E3450", ) class SteelBlueTheme(Soft): def __init__( self, *, primary_hue: colors.Color | str = colors.gray, secondary_hue: colors.Color | str = colors.steel_blue, neutral_hue: colors.Color | str = colors.slate, text_size: sizes.Size | str = sizes.text_lg, font: fonts.Font | str | Iterable[fonts.Font | str] = ( fonts.GoogleFont("Outfit"), "Arial", "sans-serif", ), font_mono: fonts.Font | str | Iterable[fonts.Font | str] = ( fonts.GoogleFont("IBM Plex Mono"), "ui-monospace", "monospace", ), ): super().__init__( primary_hue=primary_hue, secondary_hue=secondary_hue, neutral_hue=neutral_hue, text_size=text_size, font=font, font_mono=font_mono, ) super().set( background_fill_primary="*primary_50", background_fill_primary_dark="*primary_900", body_background_fill="linear-gradient(135deg, *primary_200, *primary_100)", body_background_fill_dark="linear-gradient(135deg, *primary_900, *primary_800)", button_primary_text_color="white", button_primary_text_color_hover="white", button_primary_background_fill="linear-gradient(90deg, *secondary_500, *secondary_600)", button_primary_background_fill_hover="linear-gradient(90deg, *secondary_600, *secondary_700)", button_primary_background_fill_dark="linear-gradient(90deg, *secondary_600, *secondary_700)", button_primary_background_fill_hover_dark="linear-gradient(90deg, *secondary_500, *secondary_600)", slider_color="*secondary_500", slider_color_dark="*secondary_600", block_title_text_weight="600", block_border_width="3px", block_shadow="*shadow_drop_lg", button_primary_shadow="*shadow_drop_lg", button_large_padding="11px", color_accent_soft="*primary_100", block_label_background_fill="*primary_200", ) steel_blue_theme = SteelBlueTheme() css = """ #main-title h1 { font-size: 2.3em !important; } #output-title h2 { font-size: 2.1em !important; } """ print("Determining device...") device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(f"āœ… Using device: {device}") print("Loading model and tokenizer...") model_name = "strangervisionhf/deepseek-ocr-latest-transformers" 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, ).to(device).eval() # Move to device and set to eval mode if device.type == 'cuda': model = model.to(torch.bfloat16) print("āœ… Model loaded successfully to device and in eval mode.") def find_result_image(path): for filename in os.listdir(path): if "grounding" in filename or "result" in filename: try: image_path = os.path.join(path, filename) return Image.open(image_path) except Exception as e: print(f"Error opening result image {filename}: {e}") return None @spaces.GPU def process_ocr_task(image, model_size, task_type, ref_text): """ Processes an image with DeepSeek-OCR. The model is already on the correct device. """ if image is None: return "Please upload an image first.", None print("āœ… Model is already on the designated device.") with tempfile.TemporaryDirectory() as output_path: # Build the prompt if task_type == "Free OCR": prompt = "\nFree OCR." elif task_type == "Convert to Markdown": prompt = "\n<|grounding|>Convert the document to markdown." elif task_type == "Parse Figure": prompt = "\nParse the figure." elif task_type == "Locate Object by Reference": if not ref_text or ref_text.strip() == "": raise gr.Error("For the 'Locate' task, you must provide the reference text to find!") prompt = f"\nLocate <|ref|>{ref_text.strip()}<|/ref|> in the image." else: prompt = "\nFree OCR." temp_image_path = os.path.join(output_path, "temp_image.png") image.save(temp_image_path) size_configs = { "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 (Recommended)": {"base_size": 1024, "image_size": 640, "crop_mode": True}, } config = size_configs.get(model_size, size_configs["Gundam (Recommended)"]) print(f"šŸƒ Running inference with prompt: {prompt}") text_result = model.infer( tokenizer, prompt=prompt, image_file=temp_image_path, output_path=output_path, base_size=config["base_size"], image_size=config["image_size"], crop_mode=config["crop_mode"], save_results=True, test_compress=True, eval_mode=True, ) print(f"====\nšŸ“„ Text Result: {text_result}\n====") result_image_pil = None pattern = re.compile(r"<\|det\|>\[\[(\d+),\s*(\d+),\s*(\d+),\s*(\d+)\]\]<\|/det\|>") matches = list(pattern.finditer(text_result)) if matches: print(f"āœ… Found {len(matches)} bounding box(es). Drawing on the original image.") image_with_bboxes = image.copy() draw = ImageDraw.Draw(image_with_bboxes) w, h = image.size for match in matches: coords_norm = [int(c) for c in match.groups()] x1_norm, y1_norm, x2_norm, y2_norm = coords_norm x1 = int(x1_norm / 1000 * w) y1 = int(y1_norm / 1000 * h) x2 = int(x2_norm / 1000 * w) y2 = int(y2_norm / 1000 * h) draw.rectangle([x1, y1, x2, y2], outline="red", width=3) result_image_pil = image_with_bboxes else: print("āš ļø No bounding box coordinates found in text result. Falling back to search for a result image file.") result_image_pil = find_result_image(output_path) return text_result, result_image_pil with gr.Blocks() as demo: gr.Markdown("# **DeepSeek OCR [exp]**", elem_id="main-title") with gr.Row(): with gr.Column(scale=1): image_input = gr.Image(type="pil", label="Upload Image", sources=["upload", "clipboard"]) model_size = gr.Dropdown(choices=["Tiny", "Small", "Base", "Large", "Gundam (Recommended)"], value="Large", label="Resolution Size") task_type = gr.Dropdown(choices=["Free OCR", "Convert to Markdown", "Parse Figure", "Locate Object by Reference"], value="Convert to Markdown", label="Task Type") ref_text_input = gr.Textbox(label="Reference Text (for Locate task)", placeholder="e.g., the teacher, 20-10, a red car...", visible=False) submit_btn = gr.Button("Process Image", variant="primary") with gr.Column(scale=2): output_text = gr.Textbox(label="Output (OCR)", lines=8, show_copy_button=True) output_image = gr.Image(label="Layout Detection (If Any)", type="pil") with gr.Accordion("Note", open=False): gr.Markdown("Inference using Huggingface transformers on NVIDIA GPUs. This app is running with transformers version 4.57.2 and torch version 2.9.0.") def toggle_ref_text_visibility(task): return gr.Textbox(visible=True) if task == "Locate Object by Reference" else gr.Textbox(visible=False) task_type.change(fn=toggle_ref_text_visibility, inputs=task_type, outputs=ref_text_input) submit_btn.click(fn=process_ocr_task, inputs=[image_input, model_size, task_type, ref_text_input], outputs=[output_text, output_image]) if __name__ == "__main__": demo.queue(max_size=20).launch(css=css, theme=steel_blue_theme, share=True) ``` --- ## **Implementation Example** ![Screenshot 2025-10-29 at 19-35-59 DeepSeek OCR Experimental - a Hugging Face Space by prithivMLmods](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/yOY1-MPP2qlhaC-_nVr3z.png) ![Screenshot 2025-10-29 at 19-34-45 DeepSeek OCR Experimental - a Hugging Face Space by prithivMLmods](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/x0RvOH7ZfNXk1K1h4AjTX.png) ![Screenshot 2025-10-29 at 19-32-13 Markdown Live Preview](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/a3YbzPEOH6mibBdhGb2s2.png)