search

yaroslavvb / tensorflow-community-wheels

Place to upload links to TensorFlow wheels
452 stars 35 forks source link

Ubuntu TensorFlow-GPU 1.8.0 with CUDA 9.1, cuDNN 7.1, Python 3.6.5, AVX, AVX2, FMA, and SSE4.2 #65

Open cedrickchee opened 6 years ago

cedrickchee commented 6 years ago

Hi everyone,

I am aware that many of us are eagerly waiting for TensorFlow's team to release binaries that support CUDA 9.1. I have built a TensorFlow 1.8.0 wheel with the following support:

Build specs:

Repo Path Compiler CUDA/cuDNN SIMD Compute Capability Notes
1.8.0/py36/GPU/cuda91_cudnn71_avx2_fma_sse4 GCC 5.4.0 9.1.85.3/7.1.3 x86_64 3.7 Python 3.6.5. Compiler optimization flags: -march=native. jemalloc enabled. No XLA, OpenCL, MPI support. NCCL 1.3.4. CXX flag: D_GLIBCXX_USE_CXX11_ABI=0

Install using pip:

pip install https://github.com/cedrickchee/tensorflow-community-builds/releases/download/1.8.0_py36_gpu_cuda91_cudnn71_avx2_fma_sse4/tensorflow-1.8.0-cp36-cp36m-linux_x86_64.whl

More info is in my repo. Enjoy :slightly_smiling_face:

cipri-tom commented 6 years ago

Sorry if I'm a noob, but you have to install CUDA 9.1 separately, right ? So far I've been using some docker images, like the official tensorflow ones, which have CUDA 9.0. Do you know of any images with 9.1 ?

cedrickchee commented 6 years ago

Yes, you are right on the separate installation of CUDA 9.1. I don't use Docker for my use case. So, I am not so aware of Docker images with 9.1.

cipri-tom commented 6 years ago

ok. Thank you!

And thank you for the wheel !

MuTaTeD commented 6 years ago

Hi, It appears that cudnn 7.1.3 is no longer available on the nVidia site for cuda 9.1. only thing available is 7.1.4 for 9.2. Can you update and provide new wheel or maybe link to cudnn 7.1.3...

Thanks

cedrickchee commented 6 years ago

Link to cuDNN 7.1.3 for CUDA 9.1: http://developer.download.nvidia.com/compute/redist/cudnn/v7.1.3/cudnn-9.1-linux-x64-v7.1.tgz

precabal commented 6 years ago

newbie here: will this wheel work if I don't have cuda and want the wheel just for the cpu optimizations?

cipri-tom commented 6 years ago

@precabal I think it will, as long as you don't use hard device assignments i.e. tf.device("/device:GPU:0"). Try it and let us know 😃

precabal commented 6 years ago

thanks @cipri-tom, I installed it successfully but when trying to run an example I'm getting an error saying ImportError: libcublas.so.9.1: cannot open shared object file: No such file or directory. This happens when running import tensorflow as tf. After a naïve search in google it seems to me that it is a cuda related library that is not loading, so I'm stuck there... any help?

cipri-tom commented 6 years ago

ah... so then I was wrong and you cannot use it without valid CUDA installed :( . Sorry. I thought TF only loads CUDA when needed

precabal commented 6 years ago

no problem. it was worth trying it out. thanks for your reply!