Adding information about vRAM requirements in model yaml files.

[R-Liebert]

Hi, when using Coral for fine tuning it would be nice to have a fast way of checking memory requirements for GPU clusters. I suggest putting a line about it in the config file for all models. As it's not written anywhere explicitly that the optimizer is AdamW, without going into the script for the trainer, this could ease the process of getting familiar with the package.

[saattrupdan]

You mean a simple comment about the GPU memory needed to finetuning a model, in each model config?

Ideally we'd rather have a fully fledged documentation site, but until then, would that help?

[R-Liebert]

Yes, definitely. Doesn't have to be anything complicated, but a comment saying f.ex. large-turbo can be trained on a single A10 with approx. vRAM requirements would be enough to be a reality check about where to start on projects. As you say, proper documentation is ideal, but it's very time consuming to make and we're in a field that changes rapidly so documentation gets outdated quickly.

…

On 26. des. 2024 at 10:39, Dan Saattrup Nielsen ***@***.***> wrote: You mean a simple comment about the GPU memory needed to finetuning a model, in each model config? Ideally we'd rather have a fully fledged documentation site, but until then, would that help? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[saattrupdan]

@R-Liebert Great, we'll get that included after the holidays 🙂