Linux Version For Training

[olivershetler]

Hi all,

We are in the process of attempting to reproduce your training environment. We have been running into issues with the environment on different operating systems. We are planning to use a Docker image to overcome this. However, we don't know what distribution and version of Linux you used for your training. Could you please share the Linux distribution(s) and version you used for training the model? Additionally, can you share the specs of the machine(s) that were used for training. We are particularly interested in the GPUs that were used, the amount of RAM, CUPs / number of threads, etc.

Thanks in advance,
Oliver Shetler & Brian Parbhu

[fwillett]

Hi Oliver & Brian,

A Docker image sounds useful. We trained on the Stanford computing cluster Sherlock, which I think uses "CentOS Linux release 7.9.2009". We used a variety of GPUs depending on what was available on the cluster at the time, including A100 and V100, usually with 4 CPU cores and 32GB of RAM for RNN training (more for running the language model). For example see this config file for RNN training.

Best,
Frank