paddlepaddle-gpu 3.2.0 (CUDA 12.9, Python 3.12, Linux x86_64): unmodified copy
This repository holds one file, an unmodified copy of a wheel published by the PaddlePaddle authors:
| file | paddlepaddle_gpu-3.2.0-cp312-cp312-linux_x86_64.whl |
| size | 1,889,100,386 bytes |
| sha256 | 127f7f1afbc7e268080f733718e3c667ebb06d1152dd32861a6201836c8ddba0 |
| published at | https://paddle-whl.bj.bcebos.com/stable/cu129/paddlepaddle-gpu/paddlepaddle_gpu-3.2.0-cp312-cp312-linux_x86_64.whl |
| project | https://github.com/PaddlePaddle/Paddle |
It is kept here only as a fallback for installing plantwaterer/glires, which uses
PaddleOCR. This build of PaddlePaddle is not on PyPI. glires's install.sh downloads the wheel from PaddlePaddle's own storage
first and uses this copy only if that fails; it checks the sha256 above whichever source the file came from. If you want
PaddlePaddle for anything else, get it from PaddlePaddle: https://www.paddlepaddle.org.cn/
We did not build or change this file and we are not affiliated with PaddlePaddle.
Licence
PaddlePaddle is licensed under Apache-2.0. LICENSE here is the licence file from inside the wheel
(paddlepaddle_gpu-3.2.0.dist-info/licenses/LICENSE), unchanged.
The wheel also contains compiled third-party libraries, which keep their own licences. The wheel itself carries no notices for them, so they are listed here, from the wheel's contents and from PaddlePaddle's build files at v3.2.0:
library in the wheel (paddle/libs/) |
what it is | licence |
|---|---|---|
libmklml_intel.so, libiomp5.so |
Intel MKL-ML 2019.0.5 and the Intel OpenMP runtime | Intel Simplified Software License (April 2018); text in third_party_licences/Intel-Simplified-Software-License.txt, copied from the archive PaddlePaddle's build downloads |
nvshmem_bootstrap_*.so.3, nvshmem_transport_*.so.3 |
NVIDIA NVSHMEM 3.2.5 | NVIDIA's licence for that release, with BSD-licensed portions; text in third_party_licences/NVSHMEM-3.2.5-License.txt, copied from the source archive PaddlePaddle's build downloads. Later NVSHMEM releases are Apache-2.0: https://github.com/NVIDIA/nvshmem/blob/devel/License.txt |
libdnnl.so.3 |
oneDNN | Apache-2.0: https://github.com/uxlfoundation/oneDNN/blob/main/LICENSE |
liblapack.so.3, libblas.so.3 |
Reference LAPACK 3.10.0 and BLAS | modified BSD: https://github.com/Reference-LAPACK/lapack/blob/master/LICENSE |
libgfortran.so.3, libquadmath.so.0 |
GCC runtime libraries | GPL-3.0 with the GCC Runtime Library Exception (https://www.gnu.org/licenses/gcc-exception-3.1.html); libquadmath is LGPL-2.1-or-later |
libwarpctc.so |
warp-ctc | Apache-2.0: https://github.com/baidu-research/warp-ctc/blob/master/LICENSE |
libwarprnnt.so |
warp-transducer | Apache-2.0: https://github.com/PaddlePaddle/warp-transducer/blob/master/LICENSE |
libflashattn.so, libflashattnv3.so, libflashmaskv2.so |
PaddlePaddle's flash-attention | BSD-3-Clause: https://github.com/PaddlePaddle/flash-attention/blob/main/LICENSE. These binaries also appear to have NVIDIA CUDA library code linked into them (they carry CUDA runtime messages); CUDA Toolkit licence: https://docs.nvidia.com/cuda/eula/index.html |
The remaining libraries in the wheel (libpaddle.so, libphi*.so, libcinnapi.so, libcommon.so) are PaddlePaddle's own.
The other NVIDIA CUDA libraries that PaddlePaddle needs are not in the wheel; pip or uv installs them as separate packages.