Sebastian Husch Lee

Yes that sounds good to me!

If this is a quick change that doesn't have any adverse effects please go ahead and open a PR @MichelBartels !

Looking into this some more I'm not entirely sure if the original code is causing much of a slow down. I tried bumping up `top_k_per_sample=3` when initializing a `FARMReader` and...

Ahh okay, I didn't try timings for `top_k_per_sample` this high. And I would say if it doesn't affect timings for low `top_k_per_sample` and greatly improves timings for large `top_k_per_sample` then...

Okay I just redid the timings using the following code snippet ```python # Cell 1 from haystack.nodes import FARMReader from haystack.utils import clean_wiki_text, convert_files_to_docs, fetch_archive_from_http, print_answers from haystack.modeling.infer import Inferencer...

> And we would also need to ensure that the documentation explains when to use this feature, how to use this feature and what to expect from it. Where should...

@julian-risch Do we happen to have any GPU resources in our CI testing so I could add a test to make sure the changes to `use_amp` work as expected?

Hi @julian-risch, sorry about the additional doc updates. I was just making those as I was exploring the code. I'll put them into a separate PR next time. I have...

I just found some additional docs on using AMP with multiple GPUs. * For `torch.nn.DistributedDataParallel` it looks like the existing code should work (Link to [docs](https://pytorch.org/docs/stable/notes/amp_examples.html#distributeddataparallel-one-gpu-per-process) * For `torch.nn.DataParallel` it...

Hi, @julian-risch thanks for checking in. I don't think there are any new changes since the last review other than the merging of main. I hope to be able to...