Nvidia Driver Update Within Docker for tf 1.13.1?

Archives DeepSpeech
1

Hello,

I am working with the new master branch in python3 which now uses the tf.dataset API for preprocessing and autotuning parallelization (thank you!). With the new pipeline, there are new dependencies required for tf 1.13.1 as well. All of the work I was doing prior was within an nvidia-docker container using nvidia driver version 384.145. The new tf package requires cuda 10 which requires nvidia driver version >= 410. Does anyone know if it is possible to install a different nvidia driver version for only my docker instance? I don’t necessarily want to change the driver version for my server, just the instance I’m running DeepSpeech.

Does anyone know if this is possible and where I could find out how to do it?

Thanks!

2

You can’t have a specific driver for just Docker :confused:

3

Ok. Thank you. I’ll see if I can push the new code through with 1.12 compatibility. So far,
1.12 = AUTOTUNE = multiprocessing.cpu_count()
1.13 = AUTOTUNE = tf.data.experimental.AUTOTUNE

1.13 = tf.nn.dropout(rate=FLAGS.learning_rate)
1.12 = tf.nn.dropout(keep_prob=1-FLAGS.learning_rate)

I think driver 410 is backward compat…that may also be an option.