Development Verification
- Install the wheel file built from the source code.
1 2 3 4 5 6 7 8 9 10 11 12
python3 -m pip install \ numpy==2.4.6 \ protobuf==7.35.0 \ opt_einsum==3.3.0 \ networkx==3.6.1 \ Pillow==12.2.0 \ httpx==0.28.1 \ safetensors==0.7.0 \ typing_extensions==4.15.0 python3 -m pip install --no-deps \ python/dist/paddlepaddle_gpu-3.3.0.dev20260319-cp311-cp311-linux_aarch64.whl
- Check the import and CUDA compilation capabilities.
1 2 3 4 5 6 7 8
python3 - <<'PY' import sysconfig import paddle print("soabi", sysconfig.get_config_var("SOABI")) print("paddle_version", paddle.__version__) print("compiled_with_cuda", paddle.device.is_compiled_with_cuda()) PY
The expected output is as follows.1 2 3
soabi cpython-311-aarch64-linux-gnu paddle_version 3.3.0.dev20260319 compiled_with_cuda True
- Run GPU tensor verification.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
python3 - <<'PY' import ctypes import paddle print("load_libcuda", ctypes.CDLL("libcuda.so.1")._name) print("paddle_version", paddle.__version__) print("compiled_with_cuda", paddle.device.is_compiled_with_cuda()) print("device_count", paddle.device.cuda.device_count()) paddle.set_device("gpu:0") x = paddle.arange(6, dtype="float32").reshape([2, 3]) y = paddle.ones([3, 2], dtype="float32") z = paddle.matmul(x, y) + 1 paddle.device.synchronize() print("place", z.place) pr
The expected output is as follows.1 2 3 4 5 6
load_libcuda libcuda.so.1 paddle_version 3.3.0.dev20260319 compiled_with_cuda True device_count 1 place Place(gpu:0) result [[4.0, 4.0], [13.0, 13.0]]
The pass criteria are as follows.
- The wheel file name contains the target Python ABI: cp311-cp311-linux_aarch64.
- Paddle is successfully imported.
- compiled_with_cuda is True.
- The value of device_count is greater than 0.
- The GPU tensor calculation result is correct.
Parent topic: Development Guide