My personal overlay for Gentoo on WSL.
- Pytorch
- Tensorflow
- Some NVIDIA SDKs, such as
cudnn
,nccl
,cusparselt
andtensorrt
. - and relative packages (
torchvision
,lightning
, ...)
pytorch
andtensorflow
are built with CUDA-12pytorch
is built with NCCL and Magma supporttensorflow
is built with NCCL and TensorRT support- In order to build tensorflow,
re2
,grpcio
,grpcio-tools
,abseil-cpp
,protobuf
,protobuf-python
andonnx
are built with C++17 here, whereas they are built with C++14 in offical overlay. See discussion in this pull request.
There is something strange with tensorflow-2.14 and tensorflow-2.15, see #62002 and #62075. This will not prevent tensorflow from running, and seems not to be significant. But tensorboard starts really slow with version 2.15.0 due to this error. So the current suggestion is to stick to version 2.13 (with keras 2) or update to version 2.16 (with keras 3).
With version 2.17, these warnings come back again.
Futhermore, I cannot build Tensorflow 2.18.0 with non-hermetic python. Maybe 2.17.1 is the last supported version.