Install Nvidia GPU drivers on Ubuntu

This is a helper script to install Nvidia's GPU drivers on Ubuntu, supports AMD64 and ppc64el architechtures. This script was tested on ppc64el arch and works on 18.04 or newer Ubuntu releases. (Ubuntu 18.04/18.10 please make sure you have the latest linux-firmware package installed for GPU support).

For usage:

$ ./ubuntu-install-nvidia.sh -h | --help

Install drivers and cuda

$ ./ubuntu-install-nvidia.sh This script will download the latest cuda repo and key published by Nvidia, set up the repository, and install the drivers, cuda, cudann libraries and tools. After the installation is complete, reboot the system, run nvidia-smi command to list the GPUs.

 $ nvidia-smi 
Mon Mar 18 19:21:09 2019       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.39       Driver Version: 418.39       CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  On   | 00000004:04:00.0 Off |                    0 |
| N/A   29C    P0    37W / 300W |      0MiB / 16130MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   1  Tesla V100-SXM2...  On   | 00000004:05:00.0 Off |                    0 |
| N/A   32C    P0    37W / 300W |      0MiB / 16130MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   2  Tesla V100-SXM2...  On   | 00000035:03:00.0 Off |                    0 |
| N/A   30C    P0    35W / 300W |      0MiB / 16130MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   3  Tesla V100-SXM2...  On   | 00000035:04:00.0 Off |                    0 |
| N/A   32C    P0    39W / 300W |      0MiB / 16130MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

Install pytorch on Power9.

pytorch is not packaged for ppc64el in Ubuntu, you will need to build it from source as follows:

Install pre-requisites

 $ sudo apt install python3-pip libffi-dev libssl-dev cmake
 $ pip3 install certifi cffi numpy setuptools wheel pip pyyaml Collecting certifi

Build and install pytorch

 $ git clone --recursive https://github.com/pytorch/pytorch ~/pytorch
 $ cd pytorch
 $ python3 setup.py bdist_wheel
 $ pip3 install torch-1.1.0a0+670f509-cp36-cp36m-linux_ppc64le.whl

If you get pip3 errors on installing the wheel, do the following: $ python3 -m pip uninstall pip

Run pytorch examples.

• Download pytorch examples

 $ git clone https://github.com/pytorch/examples.git ~/pyt-examples
 $ cd ~/pyt-examples/word_language_model

• Add the following to main.py for parallel runs on GPU.

  parser.add_argument("--local_rank", type=int)

  # Set your device to local rank
  torch.cuda.set_device(args.local_rank) 

• Run the testcase as follows, and run $ watch nvidia-smi on a seperate terminal. Use either of the training command lines for the model, then run generate.py.

 $ python3 -m torch.distributed.launch --nproc_per_node=4 main.py --cuda --epochs 6
 $ python3 -m torch.distributed.launch --nproc_per_node=4 main.py --cuda --epochs 6 --tied
 $ python3 -m torch.distributed.launch --nproc_per_node=4 main.py --cuda --tied
 $ python3 -m torch.distributed.launch --nproc_per_node=4 main.py --cuda --emsize 1500 --nhid 1500 --dropout 0.65 --epochs 40 --tied
 $ python3 -m torch.distributed.launch --nproc_per_node=4 generate.py --cuda

Tensorflow on Power9

At this time no prebuilt packages exists for users to easily get started with tensorflow on Power9. Building tensorflow from sources fail, and I have not been able to get it to build following the recommended combinations of bazel and tensorflow.