Step 1: Generate a virtual environment using Anaconda.
conda create -n tensorflow pip python=3.3
Step 2: Within that virtual environment on the login node,
install the tensorflow-gpu package. My VI is called
'tensorflow'.
conda install -n tensorflow anaconda tensorflow-gpu
Step 3: Make tf_gpu_test.py executable.
chmod +x tf_gpu_test.py
Step 4: Run the SLURM script.
sbatch submit.sbatch
| #!/bin/bash | |
| #SBATCH -p debug | |
| #SBATCH --output=j.out | |
| #SBATCH --error=j.err | |
| #SBATCH -t 00:01:00 | |
| #SBATCH -A your_acct_number_here | |
| #SBATCH -N 1 | |
| #SBATCH -n 1 # not sure if this matters for pure gpu runs | |
| #SBATCH -C tesla # or 'pascal' or whatever your gpu is called | |
| #SBATCH --gres=gpu:4 # or '--gres=gpu:2' if pascal above | |
| #SBATCH --qos=normal | |
| module load openmpi/1.10.2 # not really necessary for this run | |
| module load gcc/5.3.0 | |
| srun ./tf_gpu_test.py |
| #!/usr/bin/env python3 | |
| """ | |
| from the tensorflow documentation | |
| https://www.tensorflow.org/programmers_guide/using_gpu | |
| make executable with | |
| chmod +x tf_gpu_test.py | |
| """ | |
| import tensorflow as tf | |
| # Creates a graph. | |
| with tf.device('/gpu:0'): | |
| a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a') | |
| b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b') | |
| c = tf.matmul(a, b) | |
| # Creates a session with allow_soft_placement and log_device_placement set | |
| # to True. | |
| sess = tf.Session(config=tf.ConfigProto( | |
| allow_soft_placement=False, log_device_placement=True)) | |
| # Runs the op. | |
| print(sess.run(c)) |