PaddleOCR ARM/aarch64 support.

[mkd-enrique]

Hello fellow OCR developers/researchers!

I’m reaching out to ask whether there are any updates on a projected timeline for supporting aarch64/ARM architectures. I saw that the Paddle team recently announced support for the NVIDIA 50 Series, which is exciting, but I’m currently working on a server that uses an aarch64/ARM architecture with a GB10 (sm_121) GPU.

I’ve tested PaddleOCR on my own device and it works amazingly well. I was really hoping to use it in a project I’m currently developing, but due to the current incompatibility, that’s unfortunately not possible.

I’m very much looking forward to future support for aarch64 along with sm_120/sm_121 GPUs!

[FILALIHicham]

Hello @mkd-enrique !

I’m running into the exact same compatibility hurdles. Have you happened to find a stable workaround or a specific build configuration that works for you in the meantime?

Thanks in advance

[samilbeden]

Hi,

Same here — I’m hitting the same compatibility hurdles. So far, the only stable path I’ve managed is compiling and running the VL model in CPU-only mode; I couldn’t get a reliable GPU build working yet.

If you’ve found a specific flag set / build config that works (especially for GPU on ARM64), I’d really appreciate it if you can share it. I’ll also update you if I manage to get a stable GPU build in the meantime.

Thanks in advance,

[juliema1205]

I am in exactly the same situation. I am on a DGX spark, which is aarch64 architecture with a GB10 (sm_121) GPU. Actively seeking paddle team's future support for such a machine. If anybody have any build that work, love to try. Thanks.

[ximekon]

That's not a biggie. You can compile PaddlePaddle for the DGX Spark yourself. I just did it with help of Anthropic's Claude...

Step 1. Check this

uname -m          # is aarch64, right?
nvcc --version    # CUDA 13.0 (or 12.9+)
nvidia-smi        # GB10, Compute Capability 12.1
python3 --version # 3.10–3.12 recommended

Step 2. Dependencies

sudo apt-get update
sudo apt-get install -y \
    git cmake ninja-build \
    python3-dev python3-pip python3-venv \
    libopenblas-dev liblapack-dev \
    gfortran patchelf swig \
    wget curl unzip \
    libssl-dev zlib1g-dev

pip3 install cython wheel setuptools numpy protobuf

Step3: Check for CUDA 13.0 Toolkit for SBSA/aarch64
dpkg -l | grep cuda-toolkit

If not installed, install via apt on DGX OS:
sudo apt-get install cuda-toolkit-13-0

Step 4. Python Venv for Paddle Compiling
python3 -m venv paddle_env
source ~/paddle_env/bin/activate

Step 5: Install Build-Dependencies into Venv for Paddle Compiling

pip install --upgrade pip
pip install numpy protobuf cython wheel setuptools

Verify with:

which python3        # must show ~/paddle_env/bin/python3 — NOT /usr/bin/python3
python3 --version    # 3.12.3
pip --version        # pip xx.x from ~/paddle_env/...

Step 6. Clone PaddlePaddle Source

cd ~
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle

Current Stable-Branch (3.x)
git checkout develop # or git checkout v3.3.0

Submodules
git submodule update --init --recursive

Verify

ls ~/Paddle
git log --oneline -3

Step 7. Install Packaging
pip install -r ~/Paddle/python/requirements.txt

Step 8. This is nasty part A
many attempts, because "Eigen" gave errors during compilation.
This flag eventually this worked: DCMAKE_CUDA_FLAGS="-U__ARM_NEON -DEIGEN_DONT_VECTORIZE=1"

cd ~/Paddle/build

and also create a log file named "cmake_output.log" for which any AI is helpful to analyse in case anything goes south

cmake .. \
    -GNinja \
    -DCMAKE_BUILD_TYPE=Release \
    -DWITH_GPU=ON \
    -DWITH_TESTING=OFF \
    -DCUDA_ARCH_NAME=Manual \
    -DCUDA_ARCH_BIN="12.1" \
    -DWITH_ARM=ON \
    -DWITH_AVX=OFF \
    -DWITH_MKL=OFF \
    -DWITH_MKLDNN=OFF \
    -DWITH_TENSORRT=OFF \
    -DCMAKE_CUDA_FLAGS="-U__ARM_NEON -DEIGEN_DONT_VECTORIZE=1" \
    -DPYTHON_EXECUTABLE=$(which python3) \
    2>&1 | tee cmake_output.log

Step 9. This is nasty part II
in case everything went well, you can you start compiling:
ninja -j$(nproc) 2>&1 | tee build_output.log

This took about 40mins on my DGX

latest version of this walkthrough will be here: https://news.metaparadigma.de/dgx-spark-installing-paddlepaddle-ocr-on-nvidia-dgx-spark-5348/

[juliema1205]

Thank you, @ximekon! I followed the steps you posted and compiled a binary wheel that worked on my Spark. The configs you provided for step 8, in particular, are incredibly helpful. Thanks again.

And in case it helps, I made a the following minor tweaks (used Claude code too):
Step 2 — skipped pip3 install system-wide
Step 8 — build folder missing. Had to manually mkdir ~/Paddle/build first
Step 8 — cuDNN not installed: cmake requires it. Had to install libcudnn9-cuda-13 libcudnn9-dev-cuda-13 explicitly
Step 8 — NCCL not installed: Same situation. Had to install libnccl2 libnccl-dev before cmake would succeed.
Step 10 — --force-reinstall: pip install paddleocr pulls in a CPU-only paddlepaddle as a dependency. Without --force-reinstall your GPU wheel would be ignored.

[EndRollModel]

Thank you for providing this solution. I’m not very familiar with CMake and related tools. Although ChatGPT led me down a bit of a detour, I was eventually able to build it successfully.

Without this comment, I might have given up on using Paddle, and the project might not have been able to move forward.

[EvilPort2]

I have built and compiled a paddlepaddle GPU for NVIDIA GB10 Grace Blackwell. You can find the python wheel file here. Built and compiled from source here . https://drive.proton.me/urls/XWHH93QBXC#9PjuMilYe2mu

Do check it out and let me know if it helps.

(.venv_paddle) ➜  ~ neofetch
            .-/+oossssoo+/-.               evilport-ai-pc@gx10-d283 
        `:+ssssssssssssssssss+:`           ------------------------ 
      -+ssssssssssssssssssyyssss+-         OS: Ubuntu 24.04.4 LTS aarch64 
    .ossssssssssssssssssdMMMNysssso.       Host: GX10 5.36_GX10DGX 
   /ssssssssssshdmmNNmmyNMMMMhssssss/      Kernel: 6.17.0-1026-nvidia 
  +ssssssssshmydMMMMMMMNddddyssssssss+     Uptime: 2 hours, 14 mins 
 /sssssssshNMMMyhhyyyyhmNMMMNhssssssss/    Packages: 2693 (dpkg), 16 (snap) 
.ssssssssdMMMNhsssssssssshNMMMdssssssss.   Shell: zsh 5.9 
+sssshhhyNMMNyssssssssssssyNMMMysssssss+   Terminal: /dev/pts/1 
ossyNMMMNyMMhsssssssssssssshmmmhssssssso   CPU: (20) @ 2.808GHz 
ossyNMMMNyMMhsssssssssssssshmmmhssssssso   GPU: NVIDIA 000f:01:00.0 NVIDIA Corporation Device 2e12 
+sssshhhyNMMNyssssssssssssyNMMMysssssss+   Memory: 20763MiB / 124610MiB 
.ssssssssdMMMNhsssssssssshNMMMdssssssss.
 /sssssssshNMMMyhhyyyyhdNMMMNhssssssss/                            
  +sssssssssdmydMMMMMMMMddddyssssssss+                             
   /ssssssssssshdmNNNNmyNMMMMhssssss/
    .ossssssssssssssssssdMMMNysssso.
      -+sssssssssssssssssyyyssss+-
        `:+ssssssssssssssssss+:`
            .-/+oossssoo+/-.

(.venv_paddle) ➜  ~ uname -a
Linux gx10-d283 6.17.0-1026-nvidia #26-Ubuntu SMP PREEMPT_DYNAMIC Thu Jun 25 00:57:17 UTC 2026 aarch64 aarch64 aarch64 GNU/Linux
(.venv_paddle) ➜  ~ python --version
Python 3.12.3

after installation

(.venv_paddle) ➜  ~ paddlex --serve --pipeline PaddleOCR-VL.yaml
Creating model: ('PP-LCNet_x1_0_doc_ori', None, None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: `/home/evilport-ai-pc/.paddlex/official_models/PP-LCNet_x1_0_doc_ori`.
Creating model: ('UVDoc', None, None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: `/home/evilport-ai-pc/.paddlex/official_models/UVDoc`.
Creating model: ('PP-DocLayoutV2', None, None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: `/home/evilport-ai-pc/.paddlex/official_models/PP-DocLayoutV2`.
Creating model: ('PaddleOCR-VL-0.9B', None, None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: `/home/evilport-ai-pc/.paddlex/official_models/PaddleOCR-VL`.
Loading configuration file /home/evilport-ai-pc/.paddlex/official_models/PaddleOCR-VL/config.json
Loading weights file /home/evilport-ai-pc/.paddlex/official_models/PaddleOCR-VL/model.safetensors
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
use GQA - num_heads: 16- num_key_value_heads: 2
Loaded weights file from disk, setting weights to model.
All model checkpoint weights were used when initializing PaddleOCRVLForConditionalGeneration.

All the weights of PaddleOCRVLForConditionalGeneration were initialized from the model checkpoint at /home/evilport-ai-pc/.paddlex/official_models/PaddleOCR-VL.
If your task is similar to the task the model of the checkpoint was trained on, you can already use PaddleOCRVLForConditionalGeneration for predictions without further training.
Loading configuration file /home/evilport-ai-pc/.paddlex/official_models/PaddleOCR-VL/generation_config.json
INFO:     Started server process [276542]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8080 (Press CTRL+C to quit)

[yccccc12]

It works, thanks!