paddle-dgx β PaddlePaddle GPU wheels for NVIDIA DGX Spark (GB10)
Prebuilt paddlepaddle-gpu wheels for the NVIDIA DGX Spark (GB10 Grace
Blackwell: aarch64 + sm_121, CUDA 13) β a platform PaddlePaddle upstream ships
no wheels for and has no plans to support (Paddle#76215).
Source of truth for the build recipe and patches: github.com/devskale/paddle-dgx (this repo mirrors the release wheels for one-line pip installs).
Install
# v3.4.0 (latest)
pip install https://huggingface.co/skaledev/paddle-dgx/resolve/main/paddlepaddle_gpu-3.4.0.dev20260909-cp312-cp312-linux_aarch64.whl
pip install paddleocr # does NOT pull a CPU paddlepaddle
# v3.3.0
pip install https://huggingface.co/skaledev/paddle-dgx/resolve/main/paddlepaddle_gpu-3.3.0.dev20251226-cp312-cp312-linux_aarch64.whl
Python 3.12, Ubuntu 24.04 arm64, CUDA 13, driver 580.x.
Wheels
| Wheel | Paddle tag | Verified |
|---|---|---|
paddlepaddle_gpu-3.4.0.dev20260909-cp312-cp312-linux_aarch64.whl |
v3.4.0 | run_check() green on GPU, PaddleOCRVL pipeline verified (needs the flashattn-arm patch to build) |
paddlepaddle_gpu-3.3.0.dev20260319-cp312-cp312-linux_aarch64.whl |
v3.3.1 | run_check() green on GPU |
paddlepaddle_gpu-3.3.0.dev20251226-cp312-cp312-linux_aarch64.whl |
v3.3.0 | run_check() green on GPU, PP-OCRv6 line OCR on real scans |
sha256 sidecar files are next to each wheel. Verify:
sha256sum -c paddlepaddle_gpu-*.whl.sha256
Smoke test
python -c "import paddle; print(paddle.is_compiled_with_cuda())" # True
License
Wheels built from PaddlePaddle/Paddle (Apache-2.0).
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support