quick start on a local machine with CPU instead of GPU?

[MoeBensu]

Hi :)

First of all, thank you very much for the useful tool.

Since it is stated in different sites in the README.md to have (powerful) GPUs. I am wondering, if it would be possible to configure the Marin NMT to run on CPU instead for translations purposes. I think it is generally possible in case of Marian NMT, but I want to ask if one can achieve with few changes in this repository. If not, would it still be feasible to attempt that and deploy Transins on a personal laptop (i.e. Mac Air 20XX)? I have to ask, since I still face a few of issues in the compilation step which in its turn takes quite few hours.

Thanks again :)

[jsteffen]

I've added a Dockerfile to the repo for running Marian with CPUs only. In order to use it with TransIns, you have to adapt some things:

• For each translation direction in docker-compose.yml, change the image to the CPU-only version. Also remove the deploy section so that no GPUs are requested for the Docker containers.

• In order to increase translation speed, make sure that Marian is using more than a single CPU core by adding cpu-threads: 4 to each Marian config.yml in src/main/resources/<lang-dir>. Please note that in my tests, I could not use more than 4 threads.

• If you plan on using the web interface, increase the number of polling retries to avoid timeouts. See here.
  You might want to change from 10 to 100.

I can't say if TransIns runs on a laptop. Feel free to try.

[MoeBensu]

Hi @jsteffen,

Thank you so much for the putting the Dockerfile together. I can confirm that I was able to run it and deploy it on a MacAir 2015 (4 Gigs RAM). It worked with very simple files containing one line of text. However, I ran into issues when I wanted to translate larger files. The reason was simply the machine resources were exhausted.

I was able though to find a solution and run it on a Slurm cluster node with Enroot. If you think, it could be helpful to have it as a quick start guide for others who happen to run it with container runtime other than docker as Enroot, I could gladly publish it as a pull request.

Thank you very much again!

[jsteffen]

Hi @MoeNeuron,
good to know that the CPU Dockerfile works.
You're welcome to do a pull request for your Slurm/enroot solution.