# Deployment and Optimization of DeepSeek-OCR
DeepSeek OCR is a powerful AI model that reads and understands text from images and documents with high accuracy. It works on GPUs using vLLM, allowing very fast processing of scanned PDFs, invoices.
#### Minimum Requirements for DeepSeek-OCR
Hardware Requirements
| Component | Minimum | Recommended |
| ----------------------- | -------------------------- | ---------------------------------- |
| GPU | NVIDIA A30 (16–24GB VRAM) | A100 / RTX 6000 Ada (40–80GB VRAM) |
| CUDA Compute Capability | ≥ 7.0 | ≥ 8.0 |
| VRAM | 16GB | 96GB+ |
| System RAM | 64GB | 128GB |
| Storage | 40GB free | 60GB+ |
#### **Software Requirements**
* Ubuntu **22.04 / 24.04**
* NVIDIA Driver **≥ 520**
* CUDA Toolkit **11.8**
* Python **3.10**
* PyTorch 2.6.0 (CU118)
* vLLM 0.8.5 (cu118 build)
* Conda environment (Miniconda)
#### Verify GPU
```
nvidia-smi
```
Checks if your GPU is properly detected by the OS and NVIDIA drivers.
> If GPU details appear : continue.\
> If not : install NVIDIA drivers first.
#### Install Miniconda
Download and install Conda, which isolates dependencies for DeepSeek-OCR.
```
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh
```
Downloads the Miniconda installer and saves it as `miniconda.sh`
```
bash miniconda.sh -b -p $HOME/miniconda
```
Installs Miniconda silently (`-b`) into your home directory
```
eval "$($HOME/miniconda/bin/conda shell.bash hook)"
```
Activates Conda so that `conda` command works in your terminal
#### Create Conda Environment
```
conda create -n deepseek-ocr python=3.10 -y
```
{% hint style="info" %}
If you get ToS messages
{% endhint %}
Run these:
```
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r
```
Re-create environment
```
conda create -n deepseek-ocr python=3.10 -y
conda activate deepseek-ocr
```
Creates an isolated environment to install DeepSeek OCR and its dependencies safely.
#### Install PyTorch (CUDA 11.8 build)
```
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu118
```
Installs GPU-enabled PyTorch 2.6.0 that matches CUDA 11.8.
Verify PyTorch GPU access
```
python - <<'PY'
import torch, sys
print("torch:", torch.__version__)
print("torch.cuda.runtime:", torch.version.cuda)
print("cuda available:", torch.cuda.is_available())
if torch.cuda.is_available():
print("device:", torch.cuda.get_device_name(0))
PY
```
Checks if PyTorch can detect and use your GPU properly.
#### Clone DeepSeek-OCR and install requirements
```
git clone https://github.com/deepseek-ai/DeepSeek-OCR.git
cd DeepSeek-OCR
pip install -r requirements.txt
```
Downloads the DeepSeek-OCR code and installs required Python libraries.
#### After Cloning DeepSeek-OCR You Will See These Files
#### Install CUDA Toolkit 11.8 (Required for FlashAttention)
Download installer
```
// Download and install CUDA 11.8 Toolkit
wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run
```
Install toolkit only (not the driver)
```
sudo sh cuda_11.8.0_520.61.05_linux.run --toolkit --silent --override
```
Installs CUDA 11.8 compiler tools required to build FlashAttention.
This will take about 5-10 minutes to download and install.
Verify installation
```
ls -la /usr/local/cuda-11.8/bin/nvcc
```
Set environment variables (put CUDA 11.8 FIRST in PATH)
```
export CUDA_HOME=/usr/local/cuda-11.8 export PATH=/usr/local/cuda-11.8/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64:$LD_LIBRARY_PATH
```
Verify the correct nvcc is now found which nvcc
```
nvcc --version
```
#### Install GCC-11 (required for CUDA 11.8 build tools)
```
sudo apt-get install -y gcc-11 g++-11
```
Set GCC-11 as the compiler for this session
```
export CC=/usr/bin/gcc-11
export CXX=/usr/bin/g++-11
```
```
// Verify
gcc-11 --version
g++-11 --version
```
```
pip install psutil
```
#### Install FlashAttention
```
pip install flash-attn --no-build-isolation
```
This should work now. The compilation will take 5-10 minutes.
#### Make CUDA & GCC Paths Permanent in Conda Env
```
// After successful installation
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
cat > $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh << 'EOF'
export CUDA_HOME=/usr/local/cuda-11.8
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH
export CC=/usr/bin/gcc-11
export CXX=/usr/bin/g++-11
EOF
```
Ensures correct CUDA & GCC versions auto-load every time you activate the environment.
#### Install vLLM 0.8.5 (CUDA 11.8 wheel)
> (If you already installed the correct vllm, this will just confirm/force it.)
```
// prefer the cu118 wheel index; this installs vllm 0.8.5 built for cu118
pip install --upgrade "vllm==0.8.5" --extra-index-url https://wheels.vllm.ai/cu118
// quick check
python -c "import vllm, sys; print('vllm', getattr(vllm,'__version__',None))"
```
Installs the version of vLLM that works correctly with CUDA 11.8 and PyTorch 2.6.
### Prepare input / output paths and a sample image
```
mkdir -p ~/deepseek_input ~/deepseek_output
```
Creates directories to store images and OCR output.
Now open the config file and set correct input/output paths
```
nano DeepSeek-OCR-vllm/config.py
```
Save and exit:
* **CTRL + O** Enter
* **CTRL + X** Exit
#### **Run DeepSeek-OCR**
```
time python run_dpsk_ocr_image.py
```
**Successfully ran DeepSeek-OCR model.**
