Processing massive documents with standard OCR models is a recipe for disaster. As output sequences lengthen, the growing KV cache causes memory consumption to skyrocket and generation speeds to plummet.
This bottleneck prevents end-to-end models from handling long-form content efficiently, creating significant hurdles for advanced RAG pipelines and document automation.
The Bottleneck: KV Cache Explosion
Modern end-to-end models, such as DeepSeek OCR, utilize LLMs as decoders to leverage language priors. While powerful, they suffer from a critical efficiency decline during long-horizon tasks.
- Increased Latency: Generation slows down significantly as more tokens are processed.
- Memory Pressure: The accumulated KV cache drives up GPU memory requirements.
- Scalability Limits: Handling dozens of pages becomes computationally unfeasible on standard hardware.
This stands in stark contrast to humans, who exhibit no such decline in efficiency during long-duration copying tasks.
The Solution: Reference Sliding Window Attention (R-SWA)
Baidu’s groundbreaking approach, detailed in their recent technical report, introduces Unlimited OCR. This model is designed to emulate human parsing working memory by replacing all decoder attention layers with Reference Sliding Window Attention (R-SWA).
By implementing R-SWA, the model reduces computation costs while maintaining a constant KV cache throughout the entire decoding process. This allows for high-speed transcription of massive documents in a single forward pass.
| Feature | Standard LLM OCR | Unlimited OCR (R-SWA) |
|---|---|---|
| KV Cache Growth | Linear/Cumulative | Constant |
| Processing Speed | Degrades with length | Consistent efficiency |
| Max Context Capacity | Limited by memory | Up to 32K+ context |
The combination of high-compression encoders and constant-cache decoders means you can transcribe dozens of pages without the usual performance penalty.
Implementation Guide
Inference Transformers
Inference using Huggingface transformers on NVIDIA GPUs. Requirements tested on python 3.12.3 + CUDA12.9：
torch==2.10.0
torchvision==0.25.0
transformers==4.57.1
Pillow==12.1.1
matplotlib==3.10.8
einops==0.8.2
addict==2.4.0
easydict==1.13
pymupdf==1.27.2.2
psutil==7.2.2
import os
import torch
from transformers import AutoModel, AutoTokenizer
model_name = 'baidu/Unlimited-OCR'
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(
model_name,
trust_remote_code=True,
use_safetensors=True,
torch_dtype=torch.bfloat16,
)
model = model.eval().cuda()
# ── Single image supports two configs: gundam or base ──
# gundam: base_size=1024, image_size=640, crop_mode=True
# base: base_size=1024, image_size=1024, crop_mode=False
model.infer(
tokenizer,
prompt='<image>document parsing.',
image_file='your_image.jpg',
output_path='your/output/dir',
base_size=1024, image_size=640, crop_mode=True,
max_length=32768,
no_repeat_ngram_size=35, ngram_window=128,
save_results=True,
)
# ── Multi page / PDF only uses base (image_size=1024) ──
model.infer_multi(
tokenizer,
prompt='<image>Multi page parsing.',
image_files=['page1.png', 'page2.png', 'page3.png'],
output_path='your/output/dir',
image_size=1024,
max_length=32768,
no_repeat_ngram_size=35, ngram_window=1024,
save_results=True,
)
# ── PDF (convert pages to images, then multi-page parsing) ──
import tempfile, fitz # PyMuPDF
def pdf_to_images(pdf_path, dpi=300):
doc = fitz.open(pdf_path)
tmp_dir = tempfile.mkdtemp(prefix='pdf_ocr_')
mat = fitz.Matrix(dpi / 72, dpi / 72)
paths = []
for i, page in enumerate(doc):
out = os.path.join(tmp_dir, f'page_{i+1:04d}.png')
page.get_pixmap(matrix=mat).save(out)
paths.append(out)
doc.close()
return paths
model.infer_multi(
tokenizer,
prompt='<image>Multi page parsing.',
image_files=pdf_to_images('your_doc.pdf', dpi=300),
output_path='your/output/dir',
image_size=1024,
max_length=32768,
no_repeat_ngram_size=35, ngram_window=1024,
save_results=True,
)
SGLang
Set up the environment (uv-managed virtualenv). Install the local SGLang wheel first, then pin kernels==0.9.0 and install PyMuPDF for PDF-to-image conversion:
uv venv --python 3.12 source .venv/bin/activate uv pip install wheel/sglang-0.0.0.dev11416+g92e8bb79e-py3-none-any.whl uv pip install kernels==0.11.7 uv pip install pymupdf==1.27.2.2
Start the SGLang server:
python -m sglang.launch_server \
--model baidu/Unlimited-OCR \
--served-model-name Unlimited-OCR \
--attention-backend fa3 \
--page-size 1 \
--mem-fraction-static 0.8 \
--context-length 32768 \
--enable-custom-logit-processor \
--disable-overlap-schedule \
--skip-server-warmup \
--host 0.0.0.0 \
--port 10000
Send streaming requests to the OpenAI-compatible API:
import base64
import json
import os
import tempfile
import fitz
import requests
from sglang.srt.sampling.custom_logit_processor import DeepseekOCRNoRepeatNGramLogitProcessor
server_url = "http://127.0.0.1:10000"
session = requests.Session()
session.trust_env = False
def pdf_to_images(pdf_path, dpi=300):
doc = fitz.open(pdf_path)
tmp_dir = tempfile.mkdtemp(prefix="pdf_ocr_")
mat = fitz.Matrix(dpi / 72, dpi / 72)
image_paths = []
for i, page in enumerate(doc):
image_path = os.path.join(tmp_dir, f"page_{i + 1:04d}.png")
page.get_pixmap(matrix=mat).save(image_path)
image_paths.append(image_path)
doc.close()
return image_paths
def encode_image(image_path):
ext = os.path.splitext(image_path)[1].lower()
mime = "image/jpeg" if ext in (".jpg", ".jpeg") else f"image/{ext.lstrip('.')}"
with open(image_path, "rb") as f:
data = base64.b64encode(f.read()).decode("utf-8")
return {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{data}"}}
def build_content(prompt, image_paths):
return [{"type": "text", "text": prompt}] + [encode_image(path) for path in image_paths]
def generate(prompt, image_paths, image_mode, ngram_window):
payload = {
"model": "Unlimited-OCR",
"messages": [{"role": "user", "content": build_content(prompt, image_paths)}],
"temperature": 0,
"skip_special_tokens": False,
"images_config": {"image_mode": image_mode},
"custom_logit_processor": DeepseekOCRNoRepeatNGramLogitProcessor.to_str(),
"custom_params": {
"ngram_size": 35,
"window_size": ngram_window,
},
"stream": True,
}
response = session.post(
f"{server_url}/v1/chat/completions",
headers={"Content-Type": "application/json"},
data=json.dumps(payload),
timeout=1200,
stream=True,
)
response.raise_for_status()
chunks = []
for line in response.iter_lines(chunk_size=1, decode_unicode=True):
if not line or not line.startswith("data: "):
continue
data = line[len("data: "):]
if data == "[DONE]":
break
event = json.loads(data)
delta = event["choices"][0].get("delta", {}).get("content", "")
if delta:
print(delta, end="", flush=True)
chunks.append(delta)
print()
return "".join(chunks)
# Single image supports two configs: gundam or base. Example below uses gundam.
generate("document parsing.", ["your_image.jpg"], image_mode="gundam", ngram_window=128)
# Multi image (base only)
generate("Multi page parsing.", ["page1.png", "page2.png"], image_mode="base", ngram_window=1024)
# PDF (base only)
generate("Multi page parsing.", pdf_to_images("your_doc.pdf", dpi=300), image_mode="base", ngram_window=1024)
For batch inference, infer.py starts the SGLang server automatically and sends concurrent requests for an image directory or PDF:
# Image directory
python infer.py \
--image_dir ./examples/images \
--output_dir ./outputs \
--concurrency 8 \
--image_mode gundam
# PDF pages
python infer.py \
--pdf ./examples/document.pdf \
--output_dir ./outputs \
--concurrency 8 \
--image_mode gundam
Useful options:
--model_dir baidu/Unlimited-OCR # Local path or Hugging Face model ID --gpu 0 # CUDA_VISIBLE_DEVICES value --server_log ./log/sglang_server.log
Ready to revolutionize your document parsing? You can explore the full implementation and source code on the official Unlimited OCR GitHub repository.
